☰

Machine Learning in Mental Health:

A Systematic Review of the HCI
Literature to Support the
Development of Effective and
Implementable ML Systems
ANJA THIEME and DANIELLE BELGRAVE, Microsoft Research
GAVIN DOHERTY, Trinity College Dublin

ACM Trans. Comput.-Hum. Interact., Vol. 27, No. 5, Article 34, Publication date: August 2020.
DOI: https://doi.org/10.1145/3398069

High prevalence of mental illness and the need for effective mental health care, combined with recent advances in AI,
has led to an increase in explorations of how the field of machine learning (ML) can assist in the detection, diagnosis
and treatment of mental health problems. ML techniques can potentially offer new routes for learning patterns of
human behavior; identifying mental health symptoms and risk factors; developing predictions about disease
progression; and personalizing and optimizing therapies. Despite the potential opportunities for using ML within
mental health, this is an emerging research area, and the development of effective ML-enabled applications that are
implementable in practice is bound up with an array of complex, interwoven challenges. Aiming to guide future
research and identify new directions for advancing development in this important domain, this article presents an
introduction to, and a systematic review of, current ML work regarding psycho-socially based mental health
conditions from the computing and HCI literature. A quantitative synthesis and qualitative narrative review of 54
papers that were included in the analysis surfaced common trends, gaps, and challenges in this space. Discussing our
findings, we (i) reflect on the current state-of-the-art of ML work for mental health, (ii) provide concrete suggestions
for a stronger integration of human-centered and multi-disciplinary approaches in research and development, and
(iii) invite more consideration of the potentially far-reaching personal, social, and ethical implications that ML models
and interventions can have, if they are to find widespread, successful adoption in real-world mental health contexts.

CCS Concepts: • Human-centered computing → Human computer interaction (HCI);

Additional Key Words and Phrases: Mental health, mental illness, machine learning, systematic review, AI
applications, ethics, society + AI, interpretability, interaction design, health care, real-world interventions

ACM Reference format:

Anja Thieme, Danielle Belgrave, and Gavin Doherty. 2020. Machine Learning in Mental Health: A Systematic Review
of the HCI Literature to Support the Development of Effective and Implementable ML Systems. ACM Trans. Comput.-
Hum. Interact. 27, 5, Article 34 (August 2020), 53 pages. https://doi.org/10.1145/3398069

1 INTRODUCTION
Increases in the occurrence and global burden of mental illness have made the prevention and
treatment of mental disorders a public health priority [90, 91, 204, 207]. A 2017 US report
showed that an estimated 46.6 million adults have been affected by a mental illness. This
equates to nearly 20% of the US population alone [169]. Responding to the need for more
effective mental health services, the role of digital technology for improving access,
engagement, and outcomes of therapeutic treatment is increasing in importance and has led to
a wide range of health technologies and applications (e.g., [41, 156, 187]). These include
mobile apps and wearable devices to assist symptom monitoring and health risk assessments
[12, 48], computerized treatments [49, 157, 171], and mental health peer or community
support [99, 137, 150]. These systems as well as people's everyday technology interactions
and the information that is accumulated in electronic health care records (EHR) increasingly
provide a wealth of personal health and behavioral data [65, 116, 124]. Eyre et al. [56] even
suggest that the field of “mental health captures arguably the largest amount of data of any
medical specialty” (p. 21). Growth in data availability alongside improvements to computing
power has led to a surge in research and applications of machine learning (ML) technologies
[25, 186]. The field of ML extends statistical and computational methods to construct more
robust systems with an ability to automatically learn from data [173]. These techniques have
been applied successfully in the domains of gaming and recommender systems and show
promise in helping to understand large-scale health data. By offering new routes to improving
our understanding of human behaviors and predicting or optimizing outcomes [85, 173], ML
approaches are increasingly being explored for mental health (e.g., [40, 42, 83, 166]).

In recent years, reviews of the literature and research surveys that focus on applications of ML
for mental health have started to emerge in the medical and clinical psychology domain.
Existing research assesses the accuracy, reliability, and effectiveness of algorithms [100, 158],
as well as opportunities and challenges for their adoption in practice [25, 124]. Much of the
work addresses algorithm use in the area of neuroscience, specifically in neuroimaging
research (e.g., [10, 173, 181, 206]). Other works study algorithmic performance in predicting
the outcomes of clinical interventions (e.g., pharmacological treatments) for specific mental
health conditions (e.g., [100]), or discuss approaches to identify key behavioral markers for
clinical states from mobile mental health sensing data [124, 158]. To provide a better
overview of the different application areas of ML in the mental health domain, Shatte et al.
[173] recently conducted a scoping review to map the key concepts underpinning this field
from 300 literature records. The authors identified four main application domains, with the
majority of studies investigating the following: (i) detection and diagnosis of mental health
conditions, and others addressing (ii) prognosis, treatment, and support, (iii) public health, or
(iv) research and administration. They conclude that, by generating new insights into mental
health and well-being, these works demonstrate the potential of ML to improve the efficiency
of clinical and research practices.

The impact of ML in mental health will be strongly mediated by the design of systems which
employ ML, which motivates us to examine recent research in computing and HCI addressing
this topic. Complementing research perspectives from medical science and clinical psychology,
our article presents a systematic review of the ACM Guide to Computing Literature to derive a
deeper understanding of the current landscape of ML applications for mental health from an
HCI and computing science perspective. In this regard, our work builds on a recent review by
Sanches et al. [170], which mapped the design space of technologies for supporting affective
health as reported in HCI; and identified that most innovation has occurred in the areas of
automated diagnosis, and self-tracking. As researchers who are actively working at the
intersection of HCI, ML and mental health, we are excited about the prospective benefits that
ML techniques could bring to mental health. Simultaneously, from the outset of the review, we
were also aware that the development of effective and implementable ML systems is bound up
with an array of complex, interwoven socio-technical challenges. In this regard, our review is
likely shaped by both our cautious optimism that ML approaches can be usefully and
successfully applied in this domain; and a strong human-centered perspective on technology
development as well as commitment to creating responsible AI applications that seek to
improve societal outcomes. As a result, we take, at times, a slightly more critical view on
research that proposes potentially impactful real-world interventions yet remains solely
centered on technical innovation. Aiming to move the field forward in achieving many of its
ambitious goals for real-world impact, we invite the community to engage more actively and
critically with many of the complex challenges involved in order to realize successful use of ML
in mental health.

These challenges include generating large-scale, high-quality datasets, which are

representative of the diversity of the population, and gaining access to such datasets with the
purpose of developing more robust and fairer ML models (cf. [20]). Mental health, in
particular, affects a broad spectrum of people—spanning different demographics (age, gender,
ethnicity), geographic locations, and socio-economic statuses—which calls for the inclusion of
a wide range of people for this diversity to be reflected in the dataset to mitigate risks of bias
[66, 73, 76, 80]. However, data collection is costly and particularly complicated where
information is deeply personal as well as sensitive due to the stigma that is often associated
with mental conditions [29, 53, 116, 190]. Subsequently, this raises the question of whether
or not people should trust ML applications with the collection and processing of their personal
data, and to what extent and by what mechanisms people should agree to the collection of such
personal data.

These challenges are further exacerbated by forms of error, uncertainty, and bias, which are
an obstacle for the ready deployment of “state-of-the-art” ML algorithms into real-world
intelligent systems (cf. [190]). Even in cases where good accuracy can be achieved, there is
always the challenge of generalization, whereby models that are trained with high accuracy in
one scenario may not transfer to scenarios outside of the environment of the training dataset
[106]. This may introduce various sources of bias in the model, for example, demographic
disparity due to under-representation of certain groups in the training data [19, 66, 73]. Such
disparities may become magnified in sensitive domains such as mental health. This brings into
question the ethical implications of deploying a ML algorithm into an actionable health
diagnosis or treatment recommendation. This needs an interdisciplinary approach to model
interpretability, where clinical, HCI, and other domain experts support the understanding of
uncertainty, accuracy, and potential biases in ML outputs.

Finally, if ML applications are to find widespread adoption and success in real-world mental
health contexts, it is also crucial to consider potentially far reaching personal, societal, and
economic implications that the introduction of ML interventions can have. This includes
ethical questions about responsibility and accountability for ML-directed decision making
[15]; risks of potentially fallible ML outputs and biases; malicious uses of ML (see related
works in domains of criminal justice, loan decisions [161] or automated facial analysis [24],
and adversarial attacks in image processing [74] and speech recognition [37]); or digital
exclusion due to lack of knowledge, access or other barriers to technology use [70].

To provide a knowledge base to inform future research, our analysis of the computing
literature presents a quantitative synthesis and qualitative narrative overview [69] of ML
applications in mental health. Our aims are to (i) provide a comprehensive introduction to this
important and evolving area of research, (ii) highlight existing trends and gaps to guide future
work and encourage a stronger involvement by the HCI community, and (iii) sensitize the
community to many of the complex technical, societal and ethical challenges that are bound
up with the development of ML applications, if they were to be effective and implementable in
health care practice. In this regard, our literature review was guided by six main questions:

• What types of ML models and applications are currently being developed for mental
health?

• What motivates the use of ML in the reported works and what aspects of mental health
do they target?

• What types and scale of data is used for ML analysis and how is access to mental health
data achieved?
 • What techniques were applied (and challenges encountered) in developing and
evaluating ML models?

• What key learnings are reported in the literature and to what extent do they apply to real-
world contexts?

• To what extent do the papers describe ethical challenges or implications?

First, we describe our systematic review methodology; followed by the findings that were
extracted from the review corpus papers. The article concludes with a critical discussion of the
findings and provides a set of concrete suggestions for steps forward in developing ML
applications and systems that are useful, ethical and implementable in supporting mental
health.

2 METHODOLOGY
To structure the identification and selection of relevant articles for our review, we followed the
PRIMSA literature review guidelines [104, 123].

2.1 Record Identification

Relevant papers were identified by searching the electronic database of the Association for
Computing Machinery (ACM) Guide to the Computing Literature, which is the most
comprehensive bibliographic database in the field of computing and HCI research. It
integrates full text-articles of conference proceedings, journals, magazines, books, and
abstracts of key publishers including ACM, IEEE, Springer, Elsevier, John Wiley & Sons, and
Kluwer. The final corpus presented here resulted from a search conducted on the 15th of
November 2019. It included the search terms “mental health” AND “ML” (see full query
syntax1), which identified 122 records.

2.2 Record Selection

The titles and abstracts of the 122 records were independently screened by two researchers to
determine their fit with regards to addressing an application of “ML” in the context of “mental
health.” Papers were eligible for inclusion if they reported an application of ML for
understanding, detecting, diagnosing, treating affective mental health problems or conditions
(e.g., stress, depression, anxiety), psycho-social functioning (e.g., general mental health or
well-being [188]), and practices to support mental health more broadly (e.g., mental health
care providers).

Papers were excluded if they described topics of neuroscience, neurobiology, or neurological

conditions—including cell structure, cortex, and (f)MRI research (n = 22 [3, 54, 59, 72, 92,
94, 95, 96, 98, 103, 108, 115, 117, 131, 146, 180, 183, 185, 195, 199, 214, 220]), and in
one case epilepsy [5]. We also excluded neurodevelopmental disorders such as autism or
ADHD [60, 175] that present primarily as behavioral conditions. Although they both can
affect a person's ability to socialize and communicate with others, we focused our review on
psychosocially oriented mental health conditions that, instead, are primarily caused or
influenced by life experiences, as well as maladjusted cognitive and behavioral processes.
Among the most common psychosocial conditions are mood, anxiety, eating, personality, and
substance abuse conditions as well as schizophrenia. This selection criterion is consistent with
the mental health literature and other systematic reviews on affective mental health [29, 170].
Further excluded were papers that described ML research outside a specific focus on mental
health (n = 16 [21, 26, 32, 43, 75, 81, 88, 114, 118, 120, 160, 167, 200, 208, 215, 221]); or
that otherwise did not fit thematically (n = 20 [11, 14, 18, 29, 30, 36, 55, 87, 101, 107, 110,
113, 138, 143, 147, 174, 178, 194, 197, 198]). Examples include: a workshop on digital
biomarkers [55], a study of the effectiveness of eye-movements [198], or encryption methods
for protecting the privacy of databases [18]; as well as review or overview papers that did not
directly report an application of ML for mental health [29, 30, 194]. Seven records could not
be accessed (e.g., [142]). Based on a review of title and abstract, we identified 38 records as
eligible, excluded 42 records, and noted uncertainty or disagreements in the classification of
42 records, which required full-text screening. Following full-text review, another 26 papers
were excluded, leaving a final corpus of 54 articles for inclusion in the systematic review
(Figure 1).

Fig. 1. Procedural flowchart following the PRIMSA guidelines.

2.3 Data Extraction

To assist the systematic extraction of data from the papers, we created a data extraction sheet
(see Table 1 for an excerpt). It includes columns for characterizing the papers by authors,
affiliations, title, abstract, publication type and year, and individual columns to describe the
following: the type of ML application, its motivation, main data source, and target users. We
further recorded information about data access, data subjects, data scale, and data processing
steps, the ML algorithms used, and approaches to their evaluation. For each paper, we also
summarized the main research insights that were reported and listed any descriptions of ML-
specific data challenges (e.g., data quality, bias, fairness, uncertainty/error, algorithmic
interpretability). Finally, we noted if the works include topics such as real-world application,
study or design challenges, and discuss ethical issues. The extraction sheet was pilot tested on
ten randomly selected papers that fulfilled the inclusion criteria. It was first developed and
completed by one of the review authors, and then checked by another author. Each paper was
analyzed using this template. Once data extraction was completed, we added additional
columns to aid synthesis across papers. This included among others: the papers’ main
contribution type (Figure 2, right), target mental health behavior or condition (Figure 3, left),
and the category of ML algorithm used. The findings provide a quantitative and narrative
summary of the corpus with detailed examples of relevant publications. This approach has
been chosen to reflect both the breadth and depth of the trends and challenges reported; as
well as to help identify any gaps in the literature.

Fig. 2. Left: Graph showing an increasing trend in the number of ML mental

health publications over time. Right: Frequency distribution of the different
research contribution types of the papers in the review corpus.
 Fig. 3. Left: Distribution of the types of mental health behaviors or conditions
that were the target of support across all review papers. Right: Frequency
distribution of the main data domains that were used in the respective papers to
extract insights for mental health.

Table 1. Datasheet Excerpt of all 54 Papers Including Data Domain, Purpose and Description
of the ML Application or Approach, Its Motivation, and Target Mental Health Symptom or
Condition

ML
Data Application/Approach Motivation
Reference Domain Purpose (What) (Why)

Sensors (21)

Chang et al. Audio Detecting Development of an To assist in th

[31] symptoms/ automatic mental-health early diagnosi
condition monitor based on the and longitudin
human voice. Initial step: monitoring of
developing categorization mental illness
of voice utterances for symptoms in
analysis of mental health everyday spee
symptoms. conversation.

Broek et al. Audio Detecting Development of a speech- To support

[23] symptoms/ based stress indicator. efficient
condition Comparison of controlled treatment of
storytelling study (ST) PTSD, which
with an ecologically valid requires
reliving (RL) study. objective
understanding
patients’
emotional
distress.
 ML
Data Application/Approach Motivation
Reference Domain Purpose (What) (Why)
Salekin et al. Audio Detecting Development of a weakly To objectively
[168] symptoms/ supervised learning and
condition framework for detecting unobtrusively
social anxiety and detect speaker
depression from long high in social
audio clips that includes a anxiety or
novel feature modeling depression
technique (NN2Vec). symptoms tha
do not require
extensive
equipment or
clinical trainin

Mitra et al. Audio Detecting Development of a To assist

[122] symptoms/ depression-level accurate
condition recognizer based on a set diagnosis of
of acoustic features in depressive
spoken audio. symptoms.

Frogner et Accelerometer Detecting Development of multiple To accurately

al. [62] symptoms/ ML models to detect detect
condition presence and level of depression fro
depression from motor very easy to
activity recordings. obtain motor
activity.

Mallol- Body (skin Detecting Development of a multi- To aid non-

Ragolta et al. conductance) symptoms/ modal approach to intrusive
[112] condition estimate changes in PTSD measures of
symptom severity based PTSD sympto
on self-reports and skin severity throu
conductance physiology. skin conducta
responses;
reducing need
for self-report

Rabbi et al. Multiple Detecting Development + study of To continuous

[153] (audio + symptoms/ multi-modal mobile monitor a
activity) condition sensing system to person's ment
simultaneously assess well-being via
 ML
Data Application/Approach Motivation
Reference Domain Purpose (What) (Why)
mental and physical mobile sensin
health from passive that is easy, lo
sensing of everyday cost, secure +
speech in naturalistic protects priva
conditions.

Gjoreski et Multiple Detecting Development of a method To assist men

al. [67] (body) symptoms/ for continuous detection health and we
condition of stressful events from a being self-
commercial wrist worn managing by
device. developing a
stress-detectio
application as
part of a mobi
app.

DeMasi and Mobile phone Detecting Modeling the relationship To explore if

Recht [44] (GPS) symptoms/ between user mental well-
condition characteristics and being can be
algorithmic predictions of inferred from
peoples’ daily mental smartphone
well-being from behavioral dat
smartphone GPS data. and
automatically
tracked over
time.

Zakaria et al. Mobile Detecting Proposed development of To automatica

[217] phone/laptop symptoms/ a stress monitoring and non-
(Wi-Fi) condition system that is driven by intrusively det
indoor localization signs of excess
technology to predict stress from
excessive stress. mobile phone
without the ne
for installing a
app.

Zakaria et al. Mobile Detecting Development of To help detect

[218] phone/laptop symptoms/ StressMon, a stress and individuals’
(Wi-Fi) condition depression detection stress and
 ML
Data Application/Approach Motivation
Reference Domain Purpose (What) (Why)
system that leverages depression ea
location data from a and overcome
university Wi-Fi system need for app u
to better understand
physical social
interactions.

Cao et al. Mobile phone Detecting Development of an To assist

[27] (acceleration) symptoms/ architecture for modeling unobtrusive
condition mobile phone typing detection of
dynamics for inferring psychiatric
mood states in bipolar diseases in
patients (based on a late patient's daily
fusion strategy for data lives.
integration).

Quisel et al. Mobile phone Detecting Testing pre-existing To identify

[152] (multiple) symptoms/ classifier of varied self- effective (leas
condition reported mental health disruptive) tim
and nervous system window for
conditions (multi-task passively
trained CNN model) for collected mob
different data collection health data wi
time windows. high accuracy

Spathis et al. Mobile phone Detecting Development of ML To accurately

[176] (multiple) symptoms/ models to predict mood predict mood
condition from passive mobile from passive
phone sensing data and data for ment
personality trait health
questionnaire responses. assessment to
avoid frequen
experience
sampling
(burden).

Morshed et Mobile phone Detecting Development of ML To develop a

al. [128] (multiple) symptoms/ models to predict mood passive metho
condition instabilities from passive to model moo
sensing/multi-modal states at scale
 ML
Data Application/Approach Motivation
Reference Domain Purpose (What) (Why)
data in situated
communities.

Wang et al. Mobile phone Understanding Development of the To unobtrusiv

[201] (multiple) mental health StudentLife smartphone capture stude
app that incorporates life via objecti
sensing + EMA to assess smartphone d
college student mental to understand
health, academic mental health
performance + behavioral education
trends. outcomes.

Nosakhare Mobile phone Understanding Development of To provide too

and Picard (multiple) + mental health framework to map multi- for objective d
[135] activity modal behavioral analysis to hel
observational data to individuals
meaningful feature monitor their
representations, and to well-being usi
uncover behavior real-world
patterns predictive of measurement
stress/well-being.

Doryab et al. Mobile phone Understanding/ Development of a method To assist

[50] (multiple) predicting risks to infer the progression of prediction,
a primary health prevention an
parameter and applying general self-
parameter ranking to see management
which behavioral data has episodes of
the highest “impact” on mania and
health. depression of
people with
bipolar.

Alam et al. Multiple Understanding/ Development of a cloud- To effectively

[6] (body) predicting risks based system architecture predict (norm
for collecting and atypical, and
processing real-time suicidal) ment
body-sensor data as well states of patie
as additional patient with mental
information for assessing health conditi
 ML
Data Application/Approach Motivation
Reference Domain Purpose (What) (Why)
suicide risks. to monitor
suicide risk.

Hirsch et al. Audio Improving Design considerations in To effectively

[77] (counselling treatment developing ML system to assess therapi
session) automatically assess performance t
motivational Interviewing aid their skills
(MI) skills of development
psychotherapists from retention for
audio recordings of better patient
counselling session. outcomes.

Hirsch et al. Audio Improving User study of a ML To effectively

[78] (counselling treatment system to automatically assess therapi
session) assess the motivational performance t
Interviewing (MI) skills aid their skills
of psychotherapists development
directly from the audio retention for
recording of a counselling better patient
session. outcomes.

Text (16)

Chancellor Social media: Understanding Development of To understand

[28] Reddit mental health statistical methods to deviant
content identify “helpful” vs. behaviors on
“unhelpful” online mental online mental
health/wellness health
comments. communities.

Saha and De Social media: Understanding Development of a ML To study the

Choudhury Reddit mental health classifier for inferring expression of
[165] content expressions of stress from stress in socia
social media posts and media in colle
time series analysis to affected by gu
examine temporal violence
patterns (before/after) incidents.
gun violence.

Kavuluru et Social media: Understanding Development of To assist hum

al. [89] Reddit mental health identifiers of “helpful” moderators w
 ML
Data Application/Approach Motivation
Reference Domain Purpose (What) (Why)
content comments posted within review online
the Reddit community: posts through
Suicide Watch (SW), indicating
using varied text-mining and/or
techniques. prioritizing
useful/helpful
comments.

Park et al. Social media: Understanding Application of methods of To inform the

[141] Reddit mental health text mining, qualitative future design
content analysis and data mental health
visualization to compare related online
discussion topics in three communities
different online mental patient
health communities on education
Reddit. programs.

Nguyen et al. Social media: Understanding Application of text- To improve

[133] Live Journal mental health mining to better understanding
content understand linguistic mental illness
features and topics
related to mental health
discussed within online
communities on the Live
Journal platform.

Fatima et al. Social media: Detecting Development of three ML To make use o

[57] Live Journal symptoms/ models for classifying user-generate
condition depressive posts, content to
communities and the identify
degree of depression depression an
from online social media characterize it
(Live Journaling posts). degree of
severity.

Gaur et al. Social media: Detecting Development of multi- To cost-

[64] Reddit symptoms/ class classification effectively offe
condition algorithm that analysis actionable
mental health subreddit information to
posts and quantifies their clinicians abo
 ML
Data Application/Approach Motivation
Reference Domain Purpose (What) (Why)
relationship to DSM-5 patients’ ment
categories. health for web
based
intervention.

Joshi et al. Social media: Detecting Development of a model To help early

[86] Twitter symptoms/ to identify different types diagnosis of
condition of mental health mental illness
conditions from peoples’ facilitate help
social media tweets. seeking from
professional
counselors (in
India).

Yazdavar et Social media: Detecting Development of a To unobtrusiv

al. [211] Twitter symptoms/ statistical model for monitor clinic
condition monitoring different depressive
symptoms of depression symptoms in
by modeling user- social media.
generated content in
social media tweets over
time.

Chen et al. Social media: Detecting Development of a model To earlier

[33] Twitter symptoms/ that includes measures of identify and
condition eight basic emotions and better monito
temporal data as features people with, o
in prediction self- risk of
reported diagnosis of depression, fr
depression on Twitter. Twitter.

Ernala et al. Social media: Detecting Empirical study to assess To obtain

[53] Twitter + symptoms/ internal and external clinically valid
Facebook condition predictive validity of diagnostic
different social media- information fr
derived proxy diagnostic sensitive patie
signals for schizophrenia. populations.

Diedrich et Stories + Detecting Development of an ML To determine

al. [45] psychiatric symptoms/ model to determine mental health
reports record condition schizophrenia from problems
 ML
Data Application/Approach Motivation
Reference Domain Purpose (What) (Why)

written text narratives; through text

and use of clustering classification
techniques to extract key achieve more
diagnostic categories accurate
from psychiatric reports. diagnostic
classification
systems.

Nobles et al. Messages Understanding/ Development of a model To identify su

[134] (SMS) predicting risks that identifies periods of clues in text
suicidality. Report on communicatio
collection + analysis of as indicators o
text messages of heightened
individuals with a history suicide risk fo
of suicidal behaviors. more effective
prevention.

Pestian et al. Suicide notes Understanding/ Development of a To provide

[145] predicting risks classifier for predicting emergency
suicide through natural departments
language processing of with an
written suicide notes. evidence-base
risk assessme
tool for
predicting
repeated suici
attempts.

Adamou et Medical notes Understanding/ Application of text- To increase

al. [2] (from Health predicting risks mining techniques of accuracy of
record) medical notes to improve predictive mo
accuracy of a predictive in efforts to
model of suicide risk provide a tool
within 3 or 6 months at that could
point of referral to mental support clinic
health services. assessment of
suicide risk.

Wilbourne et Messages Improving Use of ML tools to aid To evaluate an

al. [205] (chat app) treatment supporters of text-based, improve the
 ML
Data Application/Approach Motivation
Reference Domain Purpose (What) (Why)
technology-enabled quality of the
mental health responses tha
intervention to assess the Silby coaches
quality of their coaching provide.
in real-time.

Structured
Data (10)

Galiatsatos Questionnaire Understanding Development of Bayesian To better

et al. [63] (from Health mental health models to better understand th
record) understand the most kinds of factor
significant psychological that affect
symptoms in mental mental health
health patients with patients who
depression. have thoughts
death or suicid

Feng et al. Questionnaire Detecting Development of a To more

[61] (from Health symptoms/ classifier to distinguish conveniently,
record) condition bipolar I from bipolar II efficiently, and
patients using only a accurately
small number of features. distinguish
between bipol
and II
assessment.

Srividya et Questionnaire Detecting Application of clustering To identify

al. [179] symptoms/ for data labelling and individuals wh
condition subsequent development are mentally
of a classifier to distressed to
determine the mental support early
health state of a person as detection, and
mentally stressed, neutral thereby, to
or happy. benefit society

Spathis et al. Questionnaire Detecting Development of multi- To provide an

[177] symptoms/ task encoder-decoder effective, read
condition RNN that learns patterns to-use tool for
from different users to early diagnosi
predict their mood from a mood issues a
 ML
Data Application/Approach Motivation
Reference Domain Purpose (What) (Why)

limited number of self- scale via mobi

reports mental health
apps.

Ojeme and Health record Detecting Development of a class- To provide

Mbogho symptoms/ bridge multi-dimensional reliable and
[136] condition Bayes network clinician
classification approach to interpretable
simultaneously identify diagnostic
depression and physical results for
illness. detection of
depression +
physical illnes
in Nigeria.

Yang and Questionnaire Understanding Application of 5 ML To improve

Bath [209] mental health models and their understanding
combinations to better underlying
predict and understand pathophysiolo
factors of depression in of depression
older people. developing
appropriate
interventions.

Panagiotako- Questionnaire Understanding/ Development of an To provide

poulos et al. predicting risks application for archiving medical staff
[139] and retrieving patient applications th
health records. Data make use of
analysis to find multi-parame
associations in context contextual dat
data and to predict collected over
patient stress in a given longer-term
context. periods.

Patterson Health record Understanding/ Application of artificial To develop +

and Cloud predicting risks neural networks (ANNs) deploy
[144] for predicting re- systematic ris
hospitalization of severely assessment
mentally ill outpatients. decision supp
tool to guide
 ML
Data Application/Approach Motivation
Reference Domain Purpose (What) (Why)
intervention;
reducing rates
costs of
rehospitalizat

Tran et al. Health record Understanding/ Development of a To improve ea

[192] predicting risks framework to detection of
automatically predict suicide and
low-, moderate-, and prevention.
high-risk of suicide given
mental health history,
risk assessment and
clinical intervention data.

Aguilar-Ruiz Questionnaire Assessing Development of To provide

et al. [4] patient– knowledge model for insight that
clinician describing the would enable
relationship relationship between doctors to bet
(psychiatric) patients and communicate
their doctors in a hospital with their
context. patients to
increase patie
satisfaction.

Multi-modal
system use
(7)

Jain and Chatbot, web Detecting Development of a To help early

Agarwal media activity symptoms/ methodological diagnosis of
[82] + wearables condition framework for creating mental illness
an electronic health facilitate help
portfolio based on daily seeking, share
computer interactions for health
psychiatric symptom progression, a
diagnosis + prognosis. optimize
treatments.

Tavabi [184] Embodied Detecting Proposed development of To augment

agent (video, symptoms/ multi-modal ML methods clinical resour
audio, text) condition for augmenting embodied in diagnosis a
 ML
Data Application/Approach Motivation
Reference Domain Purpose (What) (Why)

interactive agents with treatment of

emotional intelligence patients throu
and assist in mental automatic
health assessment. behavior
analysis.

Zhou et al. Interaction + Detecting Development of a To develop

[222] video data, symptoms/ multimodal signal system effective,
questionnaires condition that analysis a person's physically
social media stream and noninvasive,
images of a close-up low-cost
video (i.e. from mobile) approach to
to monitor and predict assess mental
mental health states. health via
pervasive
multimodal
sensors.

Rastogi et al. Multi-modal Detecting Development of a CBT- To study signs

[154] robot symptoms/ based, multi-modal, depression fro
Interactions condition humanoid robot “unobtrusive”
interaction framework for multi-modal
depression detection. communicatio
with social rob

Ray et al. Multi-modal Detecting Development of a novel To use

[155] human/agent symptoms/ ML framework with behavioral cue
interaction condition attention mechanisms at to predict
data several layers to identify depression
+ extract important severity to
features from multi- address
modal data to predict subjectivity
level of depression. problems of
existing
diagnostic tes

Tsiakas et al. Audio-visual + Optimizing Development of dialogue To create

[193] structured health system that models adaptive
data screening optimal transitions in a dialogue to aid
screening process for effective
 ML
Data Application/Approach Motivation
Reference Domain Purpose (What) (Why)
anxiety and depression symptom
based on user response to screening and
questions + emotions. provide referr
to relevant
treatment
resources.

Paredes et Phone app use Improving User study of a smart- To help peopl
al. [140] data; user treatment phone application that cope better wi
traits + self- uses ML for personalized stress at scale
reports recommendations of beyond what
constructive stress coping individual or
behaviors and services. group therapi
can provide
today.

3 FINDINGS
The final review corpus includes papers published between the years of 2000 and 2019.
Publications increased in recent years (Figure 2, left), with nearly two-third of all papers
published in the last three years.

Of the 54 papers, 33 are conference publications (7 abstracts, 5 short- and 21 full-length

proceeding papers), 14 are journal articles, and 7 symposium or workshop papers.
Furthermore, Figure 2 (right) shows how the vast majority of the papers primarily describe
the development of a ML model based on specific data as their main research contribution (n
= 40). Seven papers are proposals of specific concepts [28, 82, 154], data methods [31],
models [184], or systems [193, 217], and three apply existing ML algorithms to better
understand [209] and assess mental health [201], or improve the communication of mental
health providers [205]. Furthermore, few papers describe the conduct of empirical studies of
an end-to-end ML system [78, 140] or assess the quality of ML predictions [53]. One paper
specifically discusses design implications for user-centric, deployable ML systems [77].

3.1 Types of ML Applications, Their Data and Mental

Health Focus
This section describes what mental health behaviors or conditions were targeted, what types of
data was used to extract mental health-related insights, and the types of ML applications and
models that were developed.

3.1.1 Target Mental Health Behaviors or Conditions. The works reviewed can broadly be
grouped into two main application areas: (i) the majority of papers that come under assisting
understanding, detection and diagnosis of mental health status, (n = 49); and (ii) a small
portion of papers assess patient–clinician relationships (n = 1) or seek to improve treatment
(n = 4).

Of all 54 papers, a large proportion described a focus on supporting people with mental health
behaviors or conditions of depression (n = 13) and suicide (n = 6). Some works addressed
stress (n = 5), bipolar disorder (n = 3), mood (n = 3), PTSD (n = 3), anxiety (n = 2), substance
abuse (n = 2), or schizophrenia (n = 1). A number of papers (n = 6) targeted multiple mental
health conditions (i.e., schizophrenia and mania [45]); and others focused more broadly on
mental illness (n = 5), or mental health (n = 5). See Figure 3 (left) for details.

3.1.2 Main Data Domains for ML in Mental Health. We identified four main types of data
(Figure 3, right) that were used to extract mental health-related insights through ML: (i)
sensors, (ii) text, (iii) structured data, and (iv) multi-modal technology interactions.

Sensor-based ML approaches were most common (n = 21). Here, the majority of papers
reported uses of mobile phone sensors for data collection (n = 9) or analyzed audio signals (n
= 6 [23, 31, 77, 78, 122, 168]). The second largest data source was text (n = 16), which was
primarily extracted from social media (n = 11); and, in a few instances, from SMS [134] or text
messaging [205]; and from clinical [2, 45] or suicide notes [145]. Papers that analyzed
structured data (n = 10) included the evaluation of questionnaires (n = 7) and health records
(n = 3). Several papers (n = 7) described complex multi-modal systems, or frameworks that
built on everyday technology [82, 140, 193, 222], robot [154], or human/virtual agent [155,
184] interactions.

3.1.3 Types of ML Applications for Mental Health. Next, we describe, the types of ML
applications or models that have been developed in each of the main application areas of (i)
understanding, detecting and diagnosis of mental health symptoms or outcomes; and (ii)
assessing patient–clinician relationships and improving mental health treatment.

3.1.3.1 Understanding, Detecting and Diagnosis of Mental Health Status. A large proportion
of the papers described uses of ML to assist in the detection or diagnosis of mental health
symptoms or conditions (n = 27). Many of these works focus on the (early) detection (and
monitoring) of depression or its symptoms (n = 10) [31, 33, 44, 57, 62, 122, 136 154, 211,
222], most often through the analysis of acoustic features of speech [31, 122] or Twitter
tweets [33, 86, 211]. Other examples include the detection of mood states from mobile
sensing data [128, 176], or phone typing dynamics [27] as well as stress assessments from
location [218], biometrical and accelerometer data [67]. This is complemented by recent
trends in analyzing human-robot [154] or agent interactions [155] to help assess peoples’
mental health status. Furthermore, text analysis was performed to detect and automatically
extract diagnostic information from written narratives or psychiatric records [45], while
questionnaire data was studied to help differentiate between mental health states or
diagnosis such as patients who experience bipolar I or bipolar II [61].

Aside from mental health detection and diagnosis, a significant proportion of the papers
described approaches to understanding and predicting mental health risks (n = 8).
Predominantly this included efforts to predict future suicide risks from either sensor data [6],
health records [2, 192], or text [145, 134]. Examples include: the analysis of written suicide
notes [145]; of suicidality periods from the SMS messages of individuals with a history of
suicidal behaviors [134]; and of suicide risk at time of a person's referral to mental health
services [192], and subsequent periods [2]. Outside of suicide prediction, individual papers
sought to help predict: episodes of mania or depression in people who experience bipolar
conditions [50]; risks of re-hospitalization of outpatients with severe mental health
difficulties [144]; and experiences of patient stress [139].

In the context of social media analysis, a number of papers (n = 5) further aimed to better
understand the linguistic characteristics of mental health-related content shared in online
communities; focusing primarily on Reddit2 posts, and in one case, data from Live Journal3
[133]. Here, text-mining approaches were used [89] or proposed [28] to identify helpful and
unhelpful comments in online mental health communities to assist human moderators to
prioritize their responses to comments [28, 89]. Saha and De Choudhury [165] further
developed a classifier for inferring expressions of stress from Reddit posts by college students
before and after incidences of gun violence; while others extracted linguistic features and
topics in mental health communities to learn more about themes discussed online [133, 141].

Outside of these three main categories, more isolated investigations included: the application
of ML to gain more insight into what factors (e.g., psychological symptoms, contextual
influences) may impact a person's mental health the most [63, 209] and their relation to
mental health outcomes [135, 201]. Further, Tsiakas et al. [193] described a prototype system
that engages the user in dialogue with a female avatar that asks a series of questions to screen
for symptoms of depression and anxiety. Proposed as an adaptive system, the screening
questions are optimized and encouragements offered based on the users responses and their
emotional state. Table 1 provides summaries of all paper records, their purpose and targets;
and illustrates how the use of different data domains (e.g., sensor, text) is distributed across
the corpus.

Next, we expand on the small number of papers that did not focus on assessing mental health,
and instead explored how ML could help assess patient–clinician relationships and improve
mental health treatment.

3.1.3.2 Assessing Patient–Clinician Relationships and Improving Mental Health Treatment.

Although much of the research that focused on mental health assessment has the motivation
to provide effective tools to aid clinicians and other care providers in their work (e.g., [2, 28,
61, 86, 89, 122, 134, 145]), several papers (n = 5) described investigations of how ML
techniques can be leveraged to assess the patient–clinician relationship and improve the
content or delivery of mental health treatment.

For example, Aguilar-Ruiz et al. [4] developed a knowledge model from questionnaire data
about psychiatric hospital patients’ experiences of their relationship with their doctors to help
improve doctor communication. The remaining papers either sought to help identify what may

be the optimal treatment intervention for a particular individual [140] or help improve the
communication skills of mental health care professionals (MHPs) as part of talk-based
psychotherapy interventions [77, 78, 205]. For instance, Paredes et al. [140] applied ML in a
mobile phone app to help recommend personalized coping strategies for stress management.
Their system learned from users’ engagement with different stress interventions to predict
which intervention—out of a given set—may be correlated with stress reduction for a
particular person, which becomes the basis for personalized intervention recommendations.

In contrast, the other three papers focus on ways to improve the treatment itself by assisting
MHPs to improve their professional communications. Hirsch et al. [77, 78] describe the
design of an assessment and training tool for counsellors that uses speech and language
processing to automatically generate evaluations of the motivational interviewing (MI) skills
of a therapists from the audio of a face-to-face counseling session. They present the results as
an interactive visual dashboard that highlights strengths and weaknesses in the counsellors’
communication. Finally, Wilbourne et al. [205] use ML tools to aid human coaches of a text
chat-based app called Silby4 to assess the quality and help improve their coaching response in
real-time. However, the article does not report any system details or research findings.

In summary, the vast majority of papers described ML approaches to support the following: (i)
the detection and diagnosis of mental health symptoms or conditions, (ii) predictions of
mental health risks, or (iii) understanding of mental health-related behaviors (e.g., on online
communities). Explorations of how ML could be leveraged outside of mental health
assessment to support, e.g., (iv) mental health treatment or (v) health professionals remain
scarce.

3.2 Motivations for Applying ML to Mental Health

The following interconnected themes summarize motivations for applying ML to the domain
of mental health.

3.2.1 Easy, Timely, Unobtrusive Access to more Objective, Scalable Mental Health Data. The
use of social media [64, 165, 211], sensors [23, 27, 128, 168], and other technology
interaction data [154, 222] has been described as allowing for the “non-burdensome,”
“unobtrusive,” or “passive” assessment of peoples” mental health. These systems were
suggested to enable “honest sharing of mental health concerns” (p. 754) [64] and to provide
“natural data” as it is “generated by individuals in the normal course of their lives” (p. 10655)
[133]. Sensor data was particularly valued for enabling the automatic, longer-term tracking of
a person's mental health-related behaviors [44]. Social media content was claimed to present
a “true reflection” (p. 358) [86] and “an unbiased collection of individuals’ language usages
and behaviours” (p. 1652) [33]. Further, such data was reported to be easy-to-access and
retrieve; to offer a route to timely information for timely interventions; and to allow for data
collection to be realized at scale [33, 165, 168, 222]. The analysis of data that is generated as
part of peoples’ everyday technology interactions and digital content creation was also
reported to help identify objective markers [23, 168, 201] and systematic tools for capturing
[61, 135, 184, 192] mental health behaviors, or assessing the skills of health professionals
[78]. This argument was mostly justified through descriptions of the disadvantages of
traditional questionnaires, interviews, self-report and survey tools with regards to: sampling
biases, subjective reporting biases, risks of incomplete information, or underrepresentation
[64, 78, 128, 153, 155, 211].

Despite much enthusiasm for easier, timely, and purportedly less biased data capture; one
paper questioned the validity of developing diagnostic models for mental health conditions
based on proxy data (e.g., a person's participation in a mental health community) rather than
clinically validated diagnostic information [53].

3.2.2 Time and Cost Savings: Reducing Burden on Participant or Patient Effort and Clinician
Time. ML approaches have also been described to potentially provide advantages in terms of
saving time and costs. They can reduce efforts demanded of study participants or patients
(e.g., to self-report) [176] and provide alternatives to clinical assessments [64, 128], which
can be confined to health care professionals and specialized clinics; and thus, be expensive in
terms of clinician time [23, 63, 122, 222]. In contrast, the collection of mental health proxy
data from public social media is described as inexpensive, as it can be gathered with low effort
and does not require any direct engagement with individuals [53].

3.2.3 Towards more “Accurate” and Reliable Mental Health Practices and Clinical Decision
Making. Social media and sensor data has mostly been analyzed to help support, or speed up
early detection, diagnosis, and treatment of peoples’ mental health [6, 23, 31, 33, 86, 89,
193, 211]. However, for structured data and text analysis outside of social media, there was a
stronger emphasis on the need to advance existing health care practices by developing more
“accurate,” “reliable,” and “evidence-based” clinical assessment tools. For example, where
text was analyzed to better understand or predict (acute) suicide risk, the need for novel, data-
driven tools was argued by foregrounding the insufficiency of existing clinical approaches;
suggesting that “traditional methodologies deployed in assessing suicide have not lived up to
promise” (p.1) [2]. As a consequence, clinicians “are often left to manage suicidal patients by
clinical judgment alone” (p.96) [145], and “are not able to reliably predict when someone is at
greatest risk” (p.1) [134]. Similarly, papers that analyzed health records and questionnaire
responses were often motivated to develop “automated,” more “reliable,” “less labor-intense,”
and “interpretable” diagnostic or risk assessment tools needed to improve existing diagnostic
or clinical decision making practices [61, 63, 112, 136, 139, 144, 179, 192]. For example,
Tran et al. [192] made explicitly the point that, through their experiments in detecting suicide
risk patterns from patient history, they demonstrated how their “proposed framework
outperforms risk assessment instruments by medical practitioners” (p.1410).

In summary, key motivations for the use of ML for mental health include the following: (i) the
possibilities afforded by access to behavioral data which is collected continuously and non-
invasively; (ii) advantages of timely and automated data processing for efficiency and cost
savings; as well as (iii) claims that data-driven assessments provide more objective, accurate
and reliable assessments that can improve (clinical) practices and decision making. Thus, the
literature often argues that novel models have advantages over existing approaches.

3.3 Data Scale, Subjects and Access in Mental Health

In this section, we outline how mental health data has been accessed or collected in the works
reported; including details on the scale and from whom the data was sourced.

3.3.1 Source and Scale of Mental Health Data. ML algorithms build mathematical models
based on training data to make predictions or decisions without being explicitly programmed
[93]. The papers in our corpus are split between those that collect data for this purpose (n =
29) and those that make use of existing data (n = 23). Existing data is provided through
previously generated datasets (n = 14, plus 1 hybrid) and health records (n = 6, plus 2 hybrids).
In both categories, we identified three hybrid papers that described both a study of data
collection and existing data use [45, 168, 222]. These have therefore been added to both
category counts (see Figure 4). Very few records (n = 5) did not describe any data access or
processing [28, 77, 82, 154, 205].

Fig. 4. Left: Proportion of papers reporting data collection or access to existing

data. Right: Types of data subjects included in data collection or retrieved from
existing data sources.

For the 29 records that collect data (see Figure 4, right), eight described recruitment or
sampling of “patients” and “people with specific mental health conditions.” The remaining
records primarily captured data from individuals who were described as “normal users,”
“healthy subjects,” “students”, or “older adults” [44, 67, 139, 140, 153, 168, 179, 193, 201,
217, 218, 222], or for whom data was sampled from public social media (n = 8, plus 1 record
that also includes a diagnostic sample [53]). One record further collected audio data from
mental health professionals (MHPs) [78]. Table 2 provides an overview of the numbers of
people, including those included as “control” groups, that were studied in each data
experiment. Due to their data sampling approach (described below), many of the social media
papers did not specify any “user” numbers; and instead report the “total number of posts and
comments” that were analyzed (e.g., 3000 [89], 4026 [57], 5000 [133], 7410 [141], 113,337
[165]). Contrary to this, data that is accessed as part of existing datasets and health records
predominantly included information about “patients,” or “people with a mental health
condition” (n = 18); and to a lesser extent individuals without a clinical mental health
diagnosis (n = 5) such as: mobile phone users [176, 177], students and workers [128, 135],
social media users [64].

Table 2. Overview of the Number Individuals Who Were Included in the Respective Studies
Data Collection Data Access

Participants Participants
without a without a
Patients/People Clinical Patients/People Clinical
with a Mental Mental with a Mental Mental
Health Health Health Health
Condition Diagnosis MHPs Condition Diagnosis

10 Patients with Use of clinical 21 5 Patients with Use of clinical

bipolar condition screening tools: Counsellors symptoms of screening
[50] 7 Older adults [78] depression [222]— tools:
24 Patients with [153] Study 2 805
PTSD [23] 20 Participants 20 Individuals (12 Participants
25 People with (reported people with bipolar (48 students,
MDD [31] interest in condition, 8 people 757
26 Students with stress as control group) information
suicidal history management) [27] workers) [128]
[134] [140] 55 Individuals (23 16,952 Older
59 Individuals (31 26 Healthy people with adults (2191
people with adults [67] unipolar or bipolar with and 14751
schizophrenia, 16 33 Students depression, 32 without
people with mania, [44] people without symptoms of
9 people as control 48 Students symptoms of depression)
group) [45]— [201] depression) [62] [209] No use
Study 1 105 Students 66 Participants of of clinical
66 Notes (33 notes [168]—Study 1 EASE dataset [112] screening
of people who 108 Students 79 Psychiatry tools:
committed suicide [217, 218] No reports [45]— 224 College
+ 33 notes of use of clinical Study 2 students [135]
people who screening tools: 84 Patients [122] 566 Mobile
simulated a suicide 7 Participants 91 Patient records phone users
note) [145] (under- [63] [177]
90 Patients in a defined) [193] 130 Patients [2] 7,261 Users of
psychiatric 10 Participants 142 Individuals a commercial
hospital [4] (high scores on from DAIC-WOZ wellness
143 Individuals stress scale) database [168]— platform [152]
(88 patients with [139] Study 2 17,251 Mobile
schizophrenia + 55 27 Participants 196 Outpatient phone users
individuals as [222] - Study 1 mental health [176]
control group) 200 Twitter records [144] No numbers of
[53]—Study users (100 197 Patients with research
dataset 4 mental disorder bipolar condition individuals
Data Collection Data Access

Participants Participants
without a without a
Patients/People Clinical Patients/People Clinical
with a Mental Mental with a Mental Mental
Health Health Health Health
Condition Diagnosis MHPs Condition Diagnosis
+ 100 random) [61] reported for
[86] 201 Patients from [64]
585 Twitter various reference
users with self- datasets [6]
reported 275 Individuals
diagnosis of from E-DAIC
depression [33] dataset [155]
656 1,090 Hospital
Participants patients with
(300 students + symptoms of
353 working depression and
professionals) comorbid
[179] conditions [136]
1,965 Twitter 7,746 Patient
users (1,426 EMRs [192]
who self-report No numbers of
schizophrenia + research
539 individuals individuals
as control reported for [184]
group) [53]—
Study datasets
1–3
4,000 Twitter
users (2,000
who self-report
symptoms of
depression +
2,000 people as
control group)
[211] No
numbers of
research
individuals
reported for
[57, 89, 133,
 Data Collection Data Access

Participants Participants
without a without a
Patients/People Clinical Patients/People Clinical
with a Mental Mental with a Mental Mental
Health Health Health Health
Condition Diagnosis MHPs Condition Diagnosis

141, 165]

Table 2 further outlines the number of people that were included in the respective studies. It
shows that, outside of the analysis of health records (e.g., [136, 192]) and specific existing
large-scale datasets [152, 209], the number of “patients” or “people with a mental condition”
included was generally quite low, especially when considering that advanced ML approaches
require a lot of data. Next, we outline how the papers approached the collection and access to
data, and how they conceptualized it as mental health data.

3.3.2 Data Collection and Related Conceptualizations of Mental Health. Only a small
proportion of the papers (n = 8) recruited either psychiatric hospital patients [4], or people
with a diagnosed mental health condition [23, 31, 45, 50, 53, 134, 145]. For example,
Pestian et al. [145] described the process by which three MHPs conducted linguistic
annotations of notes written by people who “committed suicide” and compared those to people
who “simulated” writing a suicide note as “controls.” The authors however do not mention
how they obtained access to real suicide notes. In contrast, Nobles et al. [134] actively
recruited individuals with a history of past suicidal thoughts and behaviors. In a lab study,
participants downloaded and labelled all outgoing SMS messages to identify events of
attempted suicide, suicidal ideation, or depression. Psycholinguistic features and word
occurrences in the SMS texts were then analyzed to identify cues of heightened suicide risks.

A significant proportion of the papers (n = 21) described data capture studies that involved
people who may not have a mental health problem, or diagnosable mental illness. Thus, to
define and extract mental health specific behaviors (e.g., from sensor and interaction data) a
number of approaches were applied. Most commonly (n = 8), the researchers used (i)
questionnaires or standardized clinical scales 5 to screen for specific mental health symptoms
and their severity within a study population [44, 67, 140, 153, 168, 201, 217, 218]. For
assessments of symptoms of depression, this commonly included the CES-D [153], BDI [44],
and PHQ [140, 201, 217, 218]. For symptoms of anxiety, reported instruments encompassed
the STAI-Y [67], SIAS, and SPS [168]; and for symptoms of stress the PSS [201, 217, 218]. In
a few instances, the researchers further employed (ii) experimental scenarios to induce and
control for specific experiences in study participants such as stress [67], anxiety [168] and
emotional states [222]. For example, Salekin et al. [168] approached their data collection by
using various scales to assess the social anxiety of university students about public speaking,
and divided them into a low and high anxiety group. Later, participants had to quickly prepare
a 3-minute speech and present it in front of a large video camera. Audio of their speech was
then analyzed to detect “socially anxious speakers.” In addition, (iii) ecological momentary
assessments 6 (EMA) were regularly applied to evaluate users’ experiences and support data
labelling [44, 67, 128, 139, 140, 176, 218].

Of the thirteen papers that described data studies involving individuals for whom no clinical
screening tools were used to assess their mental health status, nine presented an analysis of
social media content. These works extracted data from public posts, mostly using specific
Reddit or Twitter APIs [33, 89, 141]. Only in one instance, there was direct engagement with
social media users to recruit individuals with clinically assessed schizophrenia from inpatient
and outpatient psychiatric departments [53]. For mental health diagnosis or the detection of
specific mental health states, these works primarily prospected for different types of
“diagnostic signals” in online social behaviors that can be grouped into the following: (i)
affiliation behaviors, (ii) self-report, and (iii) external validation (see framework by [53]).
Here, most papers (n = 6) focused on affiliation behaviors, whereby membership to an online
mental health community [57, 64, 89, 141, 133, 165], or engagement with mental health
content (e.g., using hashtags of #anxiety, #depression or #stress [86]) are treated as a proxy
for diagnostic information. The remaining papers (n = 3) identified users with a diagnosis of
depression through public self-report of a mental illness diagnosis [33, 211]; for example by
pooling Twitter posts of people who stated “I was/have been diagnosed with depression” [33].
Across these examples, we found no evidence of external validation of assessed diagnostic
signals through, e.g., clinical appraisals or clinical scales, with exception of Ernala et al. [53].
The authors [ibid] contribute an empirical study that assesses and compares the internal and
external predictive validity of different social media-derived proxy diagnostic signals for
mental illness diagnosis of schizophrenia (see further Section 3.5.1). In other works, expert
assessment was used to help validate proxy signals, or guide data analysis [64, 89, 165, 211].
For example, Yazdavar et al. [211] developed with psychologists a lexicon with 1,620
depression-related symptoms (categorized based on clinical PHQ-9 symptoms of depression
[97]) to guide their analysis.

Finally, Hirsch et al. [78] described automatically extracting insights about the MI skills of
counsellors from audio signals. Initially, session recordings were labelled using an established
MI Skills Code (MISC) [121]. This was then combined with speech signal processing to
generate an MI quality score (composed of measures of: empathy, MI spirit, reflection-to-
question ratio, and others—as informed by the MI Treatment Integrity Scale [129]).

Thus, across all data collection papers, we found a range of different approaches for capturing,
processing, and labelling data to help isolate indicators of mental health, or facets in the
communication skills of MHPs. Although many papers targeted detection and diagnosis of
mental health conditions, outside of recruiting patient populations and explicitly applying
clinical measures, there was rarely any (external) diagnostic validation of the assessed
phenomena.

3.3.3 Access to Pre-existing Mental Health Data as an Alternative to Data Collection. Fifteen
papers reported utilizing pre-existing datasets to train predictive models or develop new ML
approaches [6, 27, 62, 64, 112, 122, 128, 135, 152, 155, 168, 176, 177, 184, 209]. This
included the use of various resources of multi-modal data such as the Analysis Interview
Corpus (DAIC-WOZ) [196, 184] and its extended E-DAIC dataset [155]. These datasets
contain audio-video recordings of clinical or AI agent-conducted interviews with people who
experience psychological distress conditions, such as anxiety, depression, and PTSD. Other
examples include the AVEC 2013 audio-visual dataset for studies of depression [122, 168],
and the EASE dataset of people undergoing trauma treatment, e.g., for PTSD [112]. The
BiAffect mobile phone and Depresjon dataset were used to access acceleration data of people
with depression [62] and bipolar conditions [27], while the English Longitudinal Study of
Ageing (ELSA) provided psychological and mental health data on older adults as indicators of
depression [209]. Finally, a few papers reported on the re-use of previously collected user
data in the context of a commercial wellness platform [152], for social media analysis [64],
and mood or well-being research (e.g., Emotion Sense [176, 177], SNAPSHOT7 [135], and
StudentLife8 [128]).

Alongside existing datasets, a number of papers (n = 8) accessed (electronic) health records

and other clinical notes, recordings, or reports for their analysis. These records can provide an
important data resource as they can document a wealth of information about demographics
and care delivery such as admission dates, types and frequency of interventions, and the
results of clinical assessments. However, the papers provided few, if any details on how access
to health record data was negotiated. Zhou et al. [222] for example only mentioned having
been provided with video and audio chat content of patients with symptoms of depression by a
psychiatrist; while other papers [136, 144, 192] only stated the type of hospital, health
department or services from which data was received. Often, patient information was solicited
from a hospital [2] or other mental health service, institute or psychiatry department [45, 61,
63] that at least one or more of the paper authors were affiliated with. This further suggests
that this kind of data access and analysis may be primarily led and conducted by health
organizations or requires close collaboration with these institutions and care providers.

3.4 Types of ML Techniques Used and Model Evaluation

Approaches
Next, we outline the ML techniques used in the papers, and how generated ML models were
evaluated.

3.4.1 Machine Learning Tasks and Techniques: Primarily Classification and Supervised
Learning. A number of different ML techniques can such as classification, regression,
association, and clustering can be applied to common tasks such as identifying correlations
and pattern recognition in high-dimensional datasets to achieve more simplified, human-
interpretable formats [22, 136]. Building on the approach by Shatte et al. [173], we grouped
the papers in our corpus into four ML-algorithm categories: (i) supervised, (ii) unsupervised,
(iii) semi-supervised learning; and (v) novel techniques (see Appendix A1 for an overview).

The vast majority of the papers in our corpus (n = 37) used supervised learning, and most
often described the application of one or more of these techniques: Support Vector Machines,
Random Forest, Decision Trees, k-Nearest neighbors, supervised LDA, Lasso, and Logistic
Regression [23, 33, 44, 45, 50, 53, 57, 61, 62, 67, 89, 112, 122, 128, 133, 135, 139, 145,
155, 165, 176, 179, 184, 192, 193, 209, 218, 222]. For supervised learning, data is labelled
and then used to train a model that then can predict the label for new data. Here, the dataset
contains both the inputs and desired outputs. In our corpus, supervised learning was primarily
applied for classification tasks, whereby a set of previously classified training instances is used
to build a model that can predict, for example, a binary class label (e.g., presence or absence of
a symptom), or a limited set of class labels (e.g., mental health condition) of unseen instances.

Unsupervised learning uses mathematical techniques to cluster data to provide new insights.
Here, the dataset only contains inputs, but no desired output labels. To discover patterns and
help structure the data, clustering methods respond to the presence or absence of
commonalities in each piece of data. Across all papers, only two specifically applied clustering
to distinguish language use in online communications [141], and extract diagnostics from
psychiatric reports [45]. However, most often, data clustering was applied as an initial step to
classification (described above) to aid the selection of features or identify labels for developing
supervised learning classifiers [2, 63, 64, 89, 122, 139, 155, 176, 179].

Only one of the papers explicitly described the use of semi-supervised learning techniques
[211] that combine labelled and unlabeled data in their model; and few papers (n = 6)
reported the use of novel methods. Novel methods included the application of custom ML
models to create multi-dimensional classifiers [136, 152] or to forecast mental well-being
from sparse self-report data [177]; deep learning (DL) methods [86, 134]; and reinforcement
learning (RL) to help create personalized recommendations for a stress-management
interventions [140]. The remaining papers either described proposals or concepts that did not
apply any ML [28, 31, 77, 82]; or applied existing classifiers to newly collected data [201].

Finally, the analysis of natural language (NLP), speech and text, presents a specialized area
of ML that mostly utilizes unsupervised techniques. Various works applied lexicon- and other
text-mining approaches (e.g., [71]) to help extract keywords (i.e., depression), topics, or
linguistic features from text to learn high quality features from human speech or text to
develop different classification models, or determine its semantic polarity [2, 33, 45, 64, 89,
133, 134, 141, 145, 155, 165, 211]. A small number of works (e.g., [23, 78, 122]) also
analyzed acoustic, para-linguistic features in speech such as estimates of prosody, pitch, or
speech rate.

Thus, in keeping with the majority of the papers’ focus on mental health assessment, the works
primarily applied supervised ML techniques to investigate if, and how well, certain mental
health behaviors, states, or conditions could be classified through newly developed data
models. Most unsupervised learning techniques were applied to support data labelling and
feature selection for classification. This is in keeping with clinical systematic review findings
by Lee et al. [100] and Shatte et al. [173]. Other routes to leveraging ML techniques e.g., for
enabling personalization, however, remain under-explored.

3.4.2 Performance Evaluation of ML Models: Common Techniques and Performance

Measures. As described, labelled classification tasks were most prominent in our corpus. The
performance of developed classification models is typically evaluated by their ability to
generalize classifications or predictions to new cases, meaning how accurately a classifier
predicts the correct class labels for new data, for which the desired output is known. To
achieve this, part of the available training data is typically “held-back” (not included in
training the model), and instead used to test how well the model performs on that held-out
data (cf. [149]). For this, the papers reported evaluation techniques of Leave-one-user-out
(LOSO) or k-fold cross-validation [2, 4, 6, 23, 27, 33, 44, 50, 53, 57, 61, 62, 63, 64, 67, 78,
86, 89, 122, 128, 133, 134, 136, 139, 140, 144, 145, 152, 153, 165, 168, 176, 177, 179,
192, 209, 211, 218, 222].

To report the performance of developed classifiers, the majority of papers reported measures
of accuracy, precision, recall, and F1-scores [4, 6, 27, 33, 45, 53 57, 61, 64, 89, 135, 134,
153, 165, 168, 179, 192, 222]. In a few instances Log Loss was used that considers the
uncertainty of a prediction based on how much it varies from the actual label [209]. The
measure of precision indicates how useful a prediction is (low false positive rate) and recall
how complete it is (low false negative rate). Accuracy is the measure of how many samples or
individuals are correctly classified out of the total number classified, and the F1 score is a
calculated weighted harmonic mean of the classifiers precision and recall (Figure 5). For
imbalanced datasets with unequal error costs, the area under the ROC curve (AUC) metric
was often used (cf. [53, 61, 139, 152, 176, 218]) and described as a more appropriate
evaluation technique. In a few instances of regression tasks9 [27, 112, 122, 128, 155, 193],
metrics of mean error (e.g., MSE, MAE, RMSE, SMAPE) were applied [50, 62, 112, 155,179]
to reveal any unexpected values, sensitivities towards outliers, and risks of over- or under-
estimating false predictions [51]. Individual works also applied more specific metrics to
evaluate multi-dimensional classification (i.e., using Hamming score, Hamming Loss, Exact-
match [136]); or the confidence [139], coherence [64], and completeness [45] of clustering
outcomes (e.g., using WCSS, Dunn Index, DB Index, or Silhouette Index to assess similarity
within, and separation between, clusters [179]).

Fig. 5. Definitions of the five most commonly used performance measures for
classification accuracy/error.

In summary, developed ML models were commonly evaluated using aggregate metrics such as
accuracy, AUC, and mean square error. While these metrics present established performance
measures, as aggregate measures, they can hide varying model performance or biases across
different population groups (cf. [73]). This also emphasizes the need to ensure that existing
datasets capture the complexity of the real world (e.g., to not under-represent certain groups);
especially given that the papers in our corpus generally assess the generalizability of achieved
models by using parts of their training data.

3.5 Research Insights and ML Specific Challenges

Next, we summarize the three main types of contributions that were reported from the
research and developed models and provide a brief overview of commonly reported ML-
specific challenges.

3.5.1 Research Insights. Although the majority of the papers described motivations to help
detect or diagnose mental health problems to impact health management practices (see
Section 3.1), the vast majority of the papers (n = 42) primarily focused on the technical or
algorithmic development of (initial) ML models. Here, success of newly developed models is
primarily reported through performance measures of the “accuracy of the (best) classifier” or
“robustness of clustering” [2, 4, 6, 23, 27, 33, 44, 45, 50, 57, 61, 62, 63, 64, 67, 86, 89,
112, 122, 128, 133, 134, 135, 136, 139, 141, 144, 145, 152, 153, 155, 165, 168, 176, 177,
179, 184, 192, 209, 211, 218, 222]. To further demonstrate how newly developed models
“outperformed existing ones” (with few exceptions [44, 135, 136]), the performance metrics
are often compared with default or baseline models, and other state-of-the art approaches [2,
27, 33, 62, 122, 134, 135, 152, 165, 168, 176, 177, 211, 218]. In addition to performance
reports, a number of these papers foregrounded methodological contributions such as: new
approaches to data labelling [168, 211] and feature extraction [112]; the inclusion of time-
dependent features [33, 165, 192, 211, 218]; improvements to data representations [177] and
data integration [27]; and strategies to optimize data collection (periods) [152, 128, 218].
Building on these results, authors often concluded how their work presented a “proof-of-
concept” that showed “the potential” of using a particular technology [27, 50, 153], data
source [57, 62, 64, 112, 128, 134, 165, 176, 218], or algorithm [44] for understanding,
detecting or inferring a relationship with mental health.

As a second contribution type, a number of studies sought to advance our understanding of

mental health. To this end, they extracted the “importance” of identified features and their
“relations” with mental health [50, 63, 89, 135]; or complemented their ML analysis through
the reporting of qualitative findings [139, 133], visualizations [57, 141], and other user
information [201] to contextualize and aid the interpretation of ML outputs in relation to
mental health. Especially for understanding online mental health communications, the works
illustrated how the identification of discussion topics can inform the design of online mental
health interventions [64, 141], assist moderators of online communities in prioritizing their
responses [28, 89]; and informing education and intervention strategies [165, 133, 141].
Outside of online social media, Yang and Bath [209] for example calculated what features
derived from questionnaire data were particularly related to symptoms of depression in older
age. Of the top 80 identified influential features, they found nine key factors, including
“loneliness” and “quality of life.” Aiming to advance our understanding beyond individual
factors that can impact peoples’ behaviors or mental health, Nosakhare and Picard [135]
studied what combinations of health behaviors may lead to certain health outcomes. To this
end, they analyzed stress experience patterns from multi-modal data and extracted “behavior
combinations” that had the highest probability of co-occurring. Although innovative in
approach, the insights achieved by these works were often preliminary and require further
research to substantiate.

Thirdly, of all 54 papers that we analyzed, few reported empirical research findings of ML
models or deployed ML interventions [53, 78, 140]. This includes work by Ernala et al. [53]
who assessed the clinical validity of ML models that were developed based on “proxy”
diagnostic information sourced from social media. The authors found that the predictive
models that were based on this proxy data had strong internal validity but performed poorly
when tested on the social media data of people who had a clinically diagnosis (poor external
validity). ML models built on affiliation behaviors alone (e.g., being a follower of a Twitter
account that focused on schizophrenia) were reported to have the poorest performance. Their
study also revealed that the inclusion of clinical judgment to appraise self-reported mental
illness on social media showed the best performance among the three tested proxy signals
[ibid]. This work therefore contributes to important discourse about construct validity of
captured data and the importance of involving clinical expertise and assessments for
developing accurate and reliable ML-supported diagnosis.

Furthermore, Parades et al. [140] conducted a 4-week study of a mobile app to explore how
ML could be utilized to personalize a stress-management intervention. Their experimental
study design varied app recommendations to be either driven by the “ML” or “randomly”
selected; and whether the user “can” or “cannot” self-select the recommended intervention
among other options. The results showed how both ML conditions had the greatest and
statistically significant positive impact on stress reduction. Yet, their findings also showed how
the ML algorithm reduced the diversity of the intervention recommendations over time. To
avoid boredom and attrition; the authors suggest “adding diversity” as an objective to the ML
algorithm.

Finally, Hirsch et al. [78] reported the findings of a study with 21 counsellors evaluating an
interactive user interface that visualizes the output of a system that automatically assesses
their MI skills from audio. They evaluated how counsellors responded to the concept of
automated skills assessment, how the system may fit within or disrupt their clinical practice,
and what concerns they may have. Results indicated difficulties for counsellors to understand
some of the global measures (i.e., how “MI spirit” was derived from the data); as well as
perceptions of system-derived data as being “objective” and “hard to contest.” More
experienced counsellors were also more likely to question the accuracy and calculation of
system feedback; and there was a desire for “actionable” feedback to help improve their skills.

Despite these works reflecting single instances in our corpus, they help initiate important
discussions of the role, acceptance, and broader implications of positioning ML systems within
peoples’ work practices, and lives. They also begin to show how deploying and studying ML
systems in real-world mental health care contexts is needed to inform and test the design of
useful and effective ML-enabled interventions.

3.5.2 Frequently Reported Data and Modeling Challenges. One of the most frequently
described data challenges has been the capture of accurate, reliable mental health data (n =
22) due to “noise, ambiguity, or redundancies” in the data [28, 31, 112, 136, 153, 165, 168,
176, 177, 211, 222]; and difficulties to identify “robust labels” for “subjective, non-discrete
human experiences” [64, 67, 89, 135, 153, 155, 168]. This challenge was particularly
pronounced where information to help (clinically) validate assessed phenomena was missing
[31, 53, 133, 165, 168], and also for research aimed at identifying data models that are
transferable to other (real-world) data contexts [23, 45, 67, 218]. In terms of ambiguity in
data, Rabbi et al. [153] and Hirsch et al. [78] describe difficulties to isolate, in audio signals,
the speaker of interest from the environment (e.g., speech emitted from a television). For text
analysis, ambiguous terms like “depression” were described as a challenge as it also describes
the “economy,” a “historic era,” and is semantically difficult to separate from expressions of
transient sadness: “I am depressed, I have an exam tomorrow” [211]. Informal language (e.g.,
word repeats “yayayay”), abbreviations (e.g., “ikr” for “I know right”), colloquialisms and
improper sentence structure (e.g., “hehe thanks”) [134] further add complexity; alongside
elaborate lexical variations that people deliberately develop to undermine communication
bans online (i.e., changing “thighgap” to “thyghgappp” in eating disorder contexts [28]).

A large proportion of the papers (n = 23) also acknowledged varied limitations of the
dataset(s) they worked with, primarily suggesting a need for “larger datasets” (e.g., [61, 82,
89, 128, 145, 155, 177]) to compensate for missing or sparse data, and to be mindful of noise
and errors in data recording. Also acknowledged was a limited generalizability of the
established results due to restrictions in the study sample [2, 45, 61, 89, 134, 136, 141, 152,
211, 168, 222] and uncertainty about other unknown confounding variables [57, 165, 168].
For example, Yazdavar et al. [211] acknowledged that their focus on social media data meant
they would only capture people who generated ample content online and were open to
expressing symptoms of depression publicly. In addition, a number of papers mentioned
concerns about biased, missing, or incomplete data [6, 45, 45, 53, 78, 128, 141, 176, 177].
Risks of potential biases were most elaborated by Ernala et al. [53]. The authors conducted an
error analysis that revealed how statistical data distributions can be drastically different
between the social media proxy datasets that they analyzed, and actual patient datasets, which
foregrounded “population and sampling biases.” Additional linguistic analysis also showed
how patients with a clinical diagnosis of schizophrenia, in contrast to the Twitter users
studied, largely had private Facebook accounts and did not exhibit disclosures about their
schizophrenia experiences and support seeking behaviors on their social media. This brings
into question to what extent proxy diagnostic data can, indeed, provide clinically grounded
“diagnosis” information about a person [ibid].

Furthermore, outside of data processing challenges specific to the respective ML tasks and
techniques applied in each work, a number of records (n = 8) explicitly outlined difficulties
with integrating varied, often multi-modal data sources [27, 122, 128, 139, 154, 155, 192,
222]. For example, Tran et al. [192] described the complexity of working with temporal
medical databases that host multiple time-indexed records for each patient that can include:
sparse and irregular episode data; heterogeneity in patient records; distribution shifts (i.e.,
new record keeping or treatment procedures), and many other types of information (i.e.,
demographics).

Finally, some of the records (n = 9) acknowledged limitations with regards to the ML

modeling approach that was chosen by advising caution regarding the use of retrospective
data for predicting future behavior [144], acknowledging that current classifiers were
designed to detect presence, duration or frequency of symptoms, but not symptom severity
[211], and proposing the use of more “personalized approaches and individualized models” to
more accurately assess experiences of mood [50] or stress [67], and support efficiency in
detecting specific mental health conditions [133, 152]. Some works [165, 201] were mindful
in their reports about difficulties with speculation about the cause and effect of achieved,
often correlation-based results that do not permit any causal claims [128]; whereas others
described the potential implications of errors in model predictions [50, 53, 78, 134, 218]
(see further Section 3.6.4), or mentioned needs for data security through secure data storage
and handling [6, 82, 153, 201]. See Appendix A2 for further detail.

In summary, the works described a number of common data modeling challenges. Primarily,
these included the following: (i) difficulties to robustly measure and label peoples’ mental
health as a complex, multi-faceted, and dynamic concept from often noisy or ambiguous data,
(ii) technical challenges in generating low-dimensional features that reduce (initially perhaps
richer, diverse) data sources into a small number of quantifiable categories suitable for
modeling, (iii) choices in model selection and training algorithms, (iv) acknowledgement of
needs for “more data” to increase model accuracy and generalizability, and (v) to reduce risks
of errors or biases.

3.6 Ethical and Research Issues in Real-World

Applications of ML

This final section describes the extent to which developed ML models (i) were envisioned for,
and used within real-world mental health contexts, (ii) followed user-centered methods in
their design or study, and (iii) described any design challenges and ethical issues regarding the
research or deployment of ML-systems.

3.6.1 Real-World Use (and Potential Implications) of Developed ML-Models or Applications.

As described previously, only two records described user studies of ML-applications [78,
140]. Despite few examples of ML-enabled systems in-use, a substantial amount of the
records included speculative descriptions, proposals, and claims how developed ML models
may come to impact on clinical or everyday health management practices [6, 27, 33, 64, 89,
112, 128, 134, 136, 141, 152, 165, 168, 176, 177, 179, 211, 209, 218, 222]. For example,
Zhou et al. [222], who suggest the development of multi-modal sensor systems to
unobtrusively assess mental health from everyday technology interactions, described the
potential impact of their work as follows: “We expect that the outcome of this research will be
an effective tool for assessing the affective states of individuals on a large scale. It can be
used as an enabling component for developing new mental health solutions, including
identifying the onset and severity of mental health problems in individuals and may prove to
be of use to clinicians, for self-awareness, and for support from family and friends” (p. 1402).
Similarly, Salekin et al. [168] suggested: “The ability to identify symptomatic individuals
from their audio data represents an objective indicator of symptom severity that can
complement health-care providers’ other assessment modalities and inform treatment”; and
claimed their “framework is a scalable complement to health-care providers’ self-report,
interview, and other assessment modalities” (p. 21f).

Thus, despite strong motivation for developing ML approaches that can make a real difference
in this important domain, only a very small number of works sought to introduce developed
data tools and insights into real-world settings. This means that the actual impact of achieved
ML models in terms of effectiveness and relevance for mental health; or use and acceptance by
laypeople, remain—so far—mostly speculative.

3.6.2 Multi-disciplinary Research Teams and Engagement with User-Centered Design. In

keeping with the review findings by Shatte et al. [173], we found that the majority of the
papers were authored by multi-disciplinary teams (n = 29). This included experts from health
and social sciences (i.e., medicine, psychology, psychiatry, behavioral, and educational
sciences), engineering (i.e., computing science, data science, intelligent systems), and
occasionally, arts and design [77, 78, 140]. Of the remaining works, a substantial proportion
was authored by experts in computing (n = 22), and in few instances by experts in health (n =
1), psychology (n = 1), or social work (n = 1). Despite a predominantly multi-disciplinary set-
up within the research teams, there was however little reporting of user-centered design
processes. Notably, the work of Hirsch et al. [77, 78] presents the strongest example of
research that followed both a participatory, iterative design process and presented a pilot
study to evaluate their ML systems with prospective users. Mostly, user involvement was only
sought in the collection of real-world user data (e.g., [4, 50, 136]), and discussed in terms of
pragmatic challenges (e.g., requirements of keeping technology charged and used; and users’
compliant with data collection [44, 152, 153, 176, 201]; software compatibility issues in data
extraction from varied devices [44, 134]; and other technology infrastructure challenges
[218]). Only a few studies described the active involvement of target-users, MHPs, or other
domain experts in data labelling (e.g., [64, 134, 136, 211]) and for validating ML model
results [53, 218]. For example, Chang et al. [31] adapted contextual inquiry [17] as a method
to capture tacit diagnostic knowledge of MHPs in categorizing voice utterances of people
suffering from major depressive disorder (MDD). Zakaria et al. [218] conducted semi-
structured interviews and collected survey data during their data collection study to “verify”
primary causes of student stress. This includes information about how the students were
managing their stress experiences, and insights about their work meeting dates, duration, and
location. Using such data primarily as “ground truth” to validate their models, there is limited
reporting of the interview findings in the article.

The general restriction of user involvement to data collection and labelling suggests a gap in
user-centric dialogue and more collaborative involvement with MHPs and people with lived
experiences of mental health (problems) that could support a deeper engagement with
important mental health needs, and also aid with the challenges involved in appropriately
addressing these needs through ML-enabled insights or applications.

3.6.3 Designing Interpretable and Trustworthy ML Models and Applications for Mental
Health. A key challenge for the use of ML-enabled outputs and systems within real-world
mental health context is to ensure that non-ML experts are able evaluate the performance of
ML models and decide whether to trust their outputs. However, only a few papers mentioned
the need for future work to develop front-end interfaces for MHPs to present and interact with
the ML outputs [192, 141], support clinician understanding of how certain data features
influence model decisions [155], to “explain the reasoning” behind ML predictions [192, 177],
and study the acceptance of proposed data tools by MHPs [82]. Spathis et al. [177] for
example explain: “Although the scope of model interpretability is very wide, including
causality, informativeness, and transparency, at least post-hoc interpretations and
visualizations are needed to qualitatively evaluate what a model has learned. This is
especially relevant in clinical setups where clinicians can only rely on interpretable models
to make informed decisions”. Furthermore, Tran et al. [192] described “transparency in
modeling decisions and interpretability in results” (p. 1411) as a key modeling consideration
and presented earning trust from clinicians for deploying their modeling solution in their daily
work-flow as the main challenge: “We anticipate that the initial resistance will be significant
as the implication of taking the advice from the machine is profound for professionals” (p.
1417).

Only one research project [77, 78] explicitly engaged with the design challenges of creating an
interactive interface for presenting model outputs that are human interpretable. Here, user
evaluation findings showed how more experienced MHPs were more likely to question the
accuracy and calculation of system feedback and expressed a desire to be able to inspect and
potentially dispute ML outputs that seem unreasonable. Further, there had been a tendency,
especially of trainee MHPs, to uncritically accept system generated outputs as “objective
measures,” even when trainees acknowledged that they did not fully understand how feedback
was derived, or what it precisely meant. Here, their willingness to trust the ML system was
bound up with the perceived “legibility” of the output rather than its statistical accuracy [78].
Thus, the authors concluded that designers, especially when developing systems that can have
potentially adverse impact on human welfare, carry the responsibility to create mechanisms
that enable users to contest system outputs and suggest developing reasonably accurate
models first, before deploying them in a health context.

Thus, outside of understanding and addressing ML model development challenges (Section

3.5.2), there is a need for more study of how model outputs are interacted with and become
interpreted by laypeople—who may become the end-users or beneficiaries of ML-enabled
solutions. Existing works further emphasize the importance for interface design to support an
appropriate level of understanding and trust in the models.

3.6.4 Considerations of Ethics. Our final theme captures the extent to which the papers
described or addressed ethical issues or procedures in their research. Echoing recent reports
by Sanches et al. [170], we found that a significant proportion of the papers (n = 26) did not
include any mention of ethics despite their focus on a sensitive area of health care research
(Table 3). Of the remaining papers, a significant proportion (n = 15) primarily reported
approvals or exemption form ethical review processes. Next, we expand on additional ethical
considerations that were communicated, and how they relate to core ethical health care
principles of (i) autonomy, (ii) beneficence and non-maleficence, and (iii) justice.

Table 3. Types and Frequency of Ethical Issues or Approaches that Were Described or
Addressed in the Papers

Detail/Steps Taken Paper/Author(s)

No mention of N/A Adamou et al. [2]; Arguilar-Ruiz et al.

ethics/ethical [4]; Alam et al. [6]; Chang et al. [31];
concerns Chen et al. [33]; DeMasi and Recht [44];
Diederich et al. [45]; Fatima et al. [57];
Frogner et al. [62]; Galiatsatos et al.
[63]; Gjoreski et al. [67]; Joshi et al.
[86]; Kavuluru et al. [89]; Mallol-
Ragolta et al. [112]; Nguyen et al. [133];
Nosakhare and Picard [135];
Panagiotakopoulos et al. [139]; Patterson
and Cloud [144]; Pestian et al. [145];
Rastogi et al. [154]; Ray et al. [155];
Spathis et al. [177]; Tavabi [184]; Tran et
al. [192]; Tsiakas et al. [193]; Wilbourne
et al. [205]

Reports of ethical Institutional/Regional Doryab et al. [50]; Ernala et al. [53];

 Detail/Steps Taken Paper/Author(s)
approval/review IRB Nobles et al. [134]; Paredes et al. [140];
exemption Salekin et al. [168]; Wang et al. [201];
Yazdevar et al. [211]; Zakaria et al. [217]

Re-use of data (e.g., that Feng et al. [61], Gaur et al. [64];
previously received or was Morshed et al. [128]; Quisel et al. [152];
exempt from ethical Spathis et al. [176]
approval)

Statement of having Ojeme and Mbogho [136], Srividya et al.

“ethical clearance” [179]

Statement of study and Park et al. [141]

data being exempt from
ethics review

Privacy protection Public data access + user Park et al. [141]; Saha and De Choudhury
anonymization [165]; Yazdevar et al. [211]

No recording of person Cao et al. [27]; Mitra et al.; Morshed et al.

identifiable data (e.g., [128]; Rabbi et al. [153]; Salekin et al.
text, speech, low-level [168]; Zakaria et al. [218]
interactions)

Confidential treatment/no Salekin et al. [168]; Wang et al. [201];

(public) sharing of data Zakaria et al. [218]; Zhou et al. [222]

Consent and user Informed consent prior to Broek et al. [23]; Feng et al. [61];
control over data study for primary data Srividya et al. [179]; Yang and Bath
use; ability to collection [209]; Zhou et al. [222]
contest ML
Need for users to choose Jain and Argawal [82]
data source used for
diagnostic assessments

Ability to contest system Hirsch et al. [77, 78]; Nobles et al. [134]
feedback

Study planning Study risk Nobles et al. [134]; Salekin et al. [168];
and conduct assessment/planning Zakaria et al. [218]
with, or supervision by
MHP (e.g., licensed
clinical psychologist,
practicing psychiatrists)
 Detail/Steps Taken Paper/Author(s)

Study coordination by Broek et al. [23]; Nobles et al. [134];

person trained with Salekin et al. [168]
relevant expertise

Post-study mood Nobles et al. [134]

assessment to
identify/help mitigate any
induced negative
experiences

Avoidance of mental Paredes et al. [140]; Zakaria et al. [218]

illness screening or
specific data instruments
to avoid harm

Broader Broader impact of Ernala et al. [53]; Hirsch et al. [77, 78];
implications and interventions (e.g., on Zakaria et al. [218]
guidelines health work-practices,
patient well-being)

Justice/fairness Zakaria et al. [218]

Lack of ethical Chancellor [28]; Morshed et al. [128];

guidelines/data Zakaria et al. [218]
regulations

3.6.4.1 Autonomy (Including User Consent and Human Agency in ML-Informed Decision-
Making Processes) A large amount of the papers addressed the value of autonomy through the
application of privacy protecting measures to respect, and ensure confidential treatment of,
peoples’ personal information (n = 11) [27, 122, 128, 141, 153, 165, 168, 201, 211, 218,
222]. For example, with sensor-based data capture, authors often chose to only record or
process higher level data such as the number or duration of specific phone interactions [27,
128], or audio features from human speech [122, 153, 201] rather than any typed or spoken
words to preserve users’ privacy. Here, Rabbi et al.’s [153] described how such measures not
only enabled data capture in a realistic user environment but were also perceived as user-
friendly: “Although the recorded features do not allow reconstruction of audio afterwards,
they enabled us to infer when human voice was present and whether there was conversation.
(…) it is worth mentioning that during the study we learned that the privacy sensitive audio
data collection was very well accepted by users” (p. 387).

Similarly, for social media data, some of the authors acknowledged the analysis of potentially
sensitive behavioral health information. They justify their data use by reporting to have pooled
“publicly” available, so called observational data, whose data collection did not involve any
interaction or intervention with subjects [141, 165, 211]. The argument is thus that such usage
does not require explicit user consent. For example, Saha and De Choudhury [165] described
how no direct contact was made with users who posted in the subreddits they analyzed and
that it was deemed impractical to gain informed consent from thousands of people. The
authors acknowledged that “therefore individuals may be unaware of the implications of
social media content, with regards to their ability to signal underlying psychological risk” (p.
23). Outside of social media studies, proposals for the need for users to take control over their
data use for diagnostic assessments were rare [82]. Few papers explicitly mentioned user
consent processes for primary data collection (n = 5) [23, 61, 179, 209, 222]. Among those
that did not explicitly mention consent were studies that reported patient interviews in a
psychiatric hospital [4]; the analysis of mental health records [45]; or audio and video
recorded conversations between patients with symptoms of depression and their psychiatrists
[222]. Here, arguably, requirements for consent are balanced with protection of anonymity,
feasibility constraints, and the potential benefits to the public that may arise from a better
understanding, or detection of mental health status (a perspective that may be informed by
public health ethics [35]).

Nevertheless, there may be a need for more explicit dialogue and efforts to nurture a clearer
understanding for those from whom data is being collected and analyzed as to what
constitutes the purpose of the data analysis, and what risks and benefits sharing their data
may entail for the person. This could be crucial for supporting people's autonomy and their
ability to make well-informed choices.

Finally, the concept of autonomy also needs to be considered where ML model results are used
as part of interventions that could drive or automate (clinical) assessments and decision-
making processes. In Section 3.6.3, we described findings by Hirsch et al. [78] that showed a
tendency by MHPs to perceive ML system evaluations as “more objective,” and to be over-
trusting of ML outputs—irrespective of a clear understanding of how results were derived, nor
if they were accurate. This over-reliance however can have strong negative implications if
model predictions are wrong, and difficult for people to scrutinize or contest. Nobles et al.
[134] exemplified this through perhaps an extreme example that raises awareness how—in the
context of a false ML alert of high suicide risk—peoples’ autonomy could be claimed by health
care services. Reflecting on questions of care responsibility, and how system outputs may
become evaluated by MHPs, and compared with human judgement if the person denies the
result, the authors [ibid] write: “The field would need to answer questions related to
mandated reporting and involuntary hospitalization. For example, would a clinician be
legally and ethically mandated to intervene as they would if a patient endorsed active suicide
intent in person? What is the most appropriate action for someone who denies having
suicidal thoughts, plans, or intent but whose text messages indicate elevated risk?” (p. 7).
Again, this emphasizes the need to better assist laypeople in evaluating the capabilities and
limitations of ML models to help counteract tendencies to uncritically accept machine-
generated insights (cf. [78]).

3.6.4.2 Beneficence and Non-maleficence. All papers were motivated in their work to
positively contribute to mental health and peoples’ welfare. The principle of beneficence
however does not only entail encouraging human flourishing and well-being by doing the right
thing, but also suggests to “do it well” [13]. This means that ML-enabled mental health
interventions should be designed to maximize benefits and minimize harm (cf. [170]).

Most explicit considerations of non-maleficence were apparent in a few works (n = 5) that

described active approaches to avoid “harm for study participants” as part of data collection
efforts. This includes the joint planning and assessing of risks involved in data collection
studies together with MHPs [134, 168, 218] and the presence of a trained psychologist or
therapist during research activities to safeguard participants who may experience distress [23,
134, 168]. In addition, some researchers made explicit choices to not screen for the presence
of any mental illness [140], or to omit critical clinical questions such as “item 9” on the PHQ
scale that assesses suicidal thoughts [218]. Regarding the latter, the researchers
acknowledged that a non-clinical research team may lack the necessary training to handle any
definite answers to this question [ibid].

Outside of user study reports, there was a lack of critical engagement with the potential
implications of introducing generated ML outputs into real-world mental health or care
practices. Although the papers described excitement with how achieved ML models and
related insights might come to benefit people, there was little reflection on how people might
respond to systems that identify, or “diagnose” them with a mental health problem; and alert
them, or others, of specific “risks.” For example, Zakaria et al. [218] describe the possibilities
of applying their ML system as an intervention as follows: “In real-world operation, students
who are concurrently depressed and severely stressed and frequently depressed but not
severely stressed are those that StressMon detects as ‘red-flags’ so that interventions can
take place as early as possible” (p. 13). The authors however also report a misclassification
rate of 18.20% in stress detection, which meant that most of their participants were identified
by the system as “severely stressed” at some point in the study; and for 9 (out of 55) students
there were several instances of depression misclassification, including one student who
remained completely undetected by their model [ibid]. What this and similar works therefore
fail to acknowledge or discuss is how proposed classifiers may come to be sensibly
implemented in practice, and what the risks and implications might be of such interventions;
especially when model predictions are likely false at least some of the times.

In another instance, Chen et al. [33] described the implications of their work on predicting
Twitter users with symptoms of depression from self-report diagnosis posts in this way: “After
learning the traces and patterns of depressed users from these features, the trained
classifiers can be easily applied for detecting Twitter users with depression who did not post
about their conditions and users who are at risk of depression” (p. 1660). Related to this
instance, open questions remain about the extent to which individuals may appreciate or reject
the idea of “depression detection” from their Twitter uses, and what harm could arise if
communications of such a proxy diagnosis are not carefully scaffolded and appropriate
safeguards in place to support the person. There is also concern about potential uses of such
technologies to deliberately identify and target individuals who may be more vulnerable (e.g.,
with advertising).

Thus, it is important for researchers to have awareness and recognize potential risks of harm
that may come from how developed ML insights or systems may be applied and appropriated
in practice. In our corpus, Hirsch et al. [77, 78] are among the very few that considered the
broader use and potential negative implications of ML predictions within a specific health care
context. They described, e.g., risks of supervisors of mental health counselors, who are being
assessed and “judged by a machine,” to potentially rely too heavily on ML recommendation in
evaluations of job performance. The authors also warned about risks of increasing financial
and organizational pressures to “rationalize” mental health care through ML technology; as
well as counsellor concerns about workplace surveillance and decisions to “fire” someone
based on automated skill assessments. As a result, trainee therapists could potentially start
adapting their practices to improve machine scores rather than their counselling skills, which,
ultimately, could be detrimental rather than helpful to patient care. These examples
foreground the importance of a more critical engagement with the broader ethical, societal and
workplace challenges that can be bound up with new ML systems.

3.6.4.3 Justice. Finally, the ethical principal of justice focuses on the fair distribution of
benefits, risks, and costs [170] and is often treated synonymously with fairness [39]. In the
context of ML research, this can include the study of what constitutes a fair distribution of
resources in the design and evaluation of algorithmic systems; removal of bias from the ML
learning process (see [53, 78, 141], Section 3.5.2); or the perceived fairness of a decision-
making process [102]. Only one paper [218] explicitly mentions “justice,” and describes it as
requiring fair participation: “Fairness is true for StressMon, as its data collection is not
influenced by factors such as the socioeconomic status or technical experience of the user.
Instead, StressMon leverages Wi-Fi, which is readily available in public spaces (e.g., offices,
campuses and shopping malls) and commodity devices (e.g., laptops and mobile phones)” (p.
23). Here, it is argued that fairness is ensured since the resource provided—an infrastructure
system to monitor stress and depression—is available to all people through their devices.
What's missing in such arguments, is the acknowledgement that not all people may have
access to, or continuously carry, laptops or mobile phones (e.g., due to the digital divide [70]).
4 DISCUSSION
This systematic review provides an introduction to the emerging area of research and
development of ML in mental health. We now discuss existing approaches and future
directions based on three key trends and associated challenges that we identified through this
review: (i) identifying important health care needs to inform ML development, (ii) evaluating
the effectiveness of ML-interventions, and (iii) understanding the broader implications of new
ML systems through deeper study within real-world contexts.

4.1 Identifying Key Health Care Needs and Problem

Definitions for ML

The findings of our review show a recent growth in ML research in the domain of mental
health, with many of the works seeking to explore how ML could be leveraged for “social good”
by helping to address the significant personal and economic burden that is caused by mental
illness. In line with recent reports by Shatte et al.’s [173], the vast majority of this research
described approaches to the detection and diagnosis of mental health behaviors or conditions.
Fewer works explored how ML approaches can support our understanding of mental health
(e.g., [28, 89, 133, 141, 165]) or be leveraged in treatment (cf. [4, 77, 78, 140]). This raises
the question how meaningful research questions and problem scenarios for ML are commonly
identified, and how best to support such choices to maximize ML utility in the mental health
domain.

Here, one assumption might be that the general need for access to large-scale, high-quality
mental health data required for ML modeling plays a moderating role in the types of research
questions and ML applications that are being developed. In the health care domain in general,
and for mental health specifically (e.g., [134]), there is an emphasis on the challenges and
costs that are involved in gaining access to, and collecting data both at scale, and of sufficient
diversity. As shown in our analysis (Table 2), and especially for data collection studies, the
numbers of participants that represented patients or people with a mental health condition
was often low —especially when considering the data demands of advanced ML techniques.
Larger numbers were achieved in analysis of health records, yet their access is often restricted
to, and requires collaboration with, health organizations. As a result, there is a risk that the
expense of data collection may limit study design, forcing researchers to use readily available
data (e.g., social media, public databases). Such data, in turn, may be suboptimal for exploring
a particular research question outside of the original data context. Similarly, the availability of
clinical outcome measures to assess mental health through clinically validated scales and
screening questionnaires (Section 3.3.2) may also contribute to explanations of the
prevalence of algorithmic modeling to assist particularly in mental health symptom detection
and diagnosis.

We believe, however, there is a lot more scope for other, perhaps more important and
innovative uses of ML if we were to ask: how ML can meaningfully augment existing health
care practices, or help make certain processes easier or more effective for mental health
service users. Finding the most beneficial (as opposed to the most obvious) applications of
ML will require creative exploration of the design space coupled with an understanding of the
real problems faced by potential users and mental health services on a day-to-day basis.
Next, we (i) expand on proposals to identify key mental health care needs and broaden the
focus of ML and (ii) suggest more active, yet careful approaches in negotiating data access to
lift constraints.

4.1.1 Wider Opportunities for ML: Moving Beyond Mental Health Detection and Categorial
Diagnosis. A key motivation of the majority of review papers was the development of ML
models to help achieve more effective tools or approaches to aid mental health assessment and
monitoring. As a new and evolving area of research, there are however a lot more
opportunities for ML to expand the scope of what is currently possible.

Understanding Mental Health Status and Discriminating between Disease

Categories. Thus far, few studies have sought to advance our understanding of mental
health by extracting the importance of identified (behavioral) features, their combinations,
and relations with mental health [50, 57, 63, 89, 133, 135, 139, 141, 165, 201, 209]. The
vast majority of papers described ML classification tasks aimed at identifying whether a
particular individual belongs to a particular diagnostic category, e.g., “depressed” or “not
depressed.” However, looking at a mental illness, like depression, as one broad category may
not take the variability of depression symptoms into account, and how the illness manifests
[189]. Furthermore, in the medical domain, and for everyday psychiatric practice, it is often
argued that the more challenging question is often not detecting the presence of mental health
conditions and whether a person is in need of treatment, but the differential diagnoses that
discriminate between multiple likely illness categories, and to identify optimal treatments
[25]. Here, ML approaches such as multi-class prediction or multi-task learning may be well
suited to explore differences across mental illness subtypes or treatment groups. ML
techniques may also assist in identifying yet-to-be-discovered mental illness dimensions and
support recent clinical efforts that seek to supplement discrete definitions through a more
continuous, dimensional symptom system [ibid].

Personalizing and Optimizing Mental Health Treatment. In our review corpus, one
paper explored how ML could enable personalized recommendations for stress treatment
[140]. There is ample scope for future work to study how ML could be applied to allow, e.g.,
for a more effective tailoring of interventions to each persons’ unique mental health and
support needs; and assist in the development of more effective mental health treatments. The
ability to potentially predict treatment effectiveness on an individual level presents a
particular benefit of ML approaches over traditional clinical and statistical methods, whose
aim often is to identify treatment options that explain the benefits and variance for the
“majority of a clinical group,” and formally test for “group effects” (cf. [25]).

ML approaches also have potential for enabling more targeted adjustments to treatment
through advancing our understanding of what types of interventions, or their form of delivery,
may work most effectively for particular people [11, 172, 194]. Albeit still scarce, studies are
starting to emerge that propose uses of ML to provide just-in-time adaptive interventions
(JITAIs) (e.g., [84, 130]). Often motivated to create more engaging, responsive and adaptive
treatments based on information about the person or their environment, JITAIs utilize
algorithms to optimize interventions for each person based on proximal outcomes [172]. For
example, Jeong and Breazeal [84] employed ML to assess a persons’ emotional state
(analyzing their facial expressions and SMS) and used this information to tailor what positive
psychology intervention the person would receive.

For digital mental health services specifically, such as online cognitive-behavioral therapy
(iCBT) interventions (e.g., SilverCloud Health10 [49, 127, 157]) or mobile mental health apps
(e.g., IntelliCare11 [125]), there is further great potential for the analysis of log event data
[128, 194]. For example, ML could be used to discover usage patterns in log data that can help
predict future user behaviors or mental health states [194]. This may include predictions of
users’ risk of drop-out from treatment, or risk of rapid declines in mental health through
which more timely and bespoke interventions could be enabled. Other approaches, such as
association analysis, can further help uncover what features in a digital behavioral health
intervention often occur together [ibid] and help derive opportunities for personalization and
to optimize treatment. This has recently been exemplified by Chikersal et al. [34], who used
association rule mining (ARM) to learn what about the communications of therapeutic
supporters who guide patients through an iCBT program for depression and anxiety is linked
with better improvements in patient mental health. Specifically, the authors analyzed how
specific linguistic strategies in support messages to patients correlated with better patient
outcomes dependent on the patients’ specific context (e.g., their current mental health,
treatment week, level of engagement with iCBT). The research showed how certain support
strategies (e.g., use of more positive words, or words referencing social behaviors) were “more”
or “less” important depending on how actively users engaged with the treatment. This, in turn,
can help human supporters of iCBT interventions to better tailor their communications to
each clients’ circumstances. Explorations of ML use for assisting the communication skills and
work practices of MHPs have also been evident in a small number of papers in our review
corpus in the context of face-to-face therapy [77, 78] and for improving coaching via a text-
chat app [205].
Supporting Positive and Preventative Approaches to Mental Health. Lastly, we
want to note that the vast majority of paper records focused on symptoms and conditions
indictive of mental health difficulties. This leaves scope for uses of ML in supporting
preventative approaches (outside of acute risk detection) and assisting in positive mental
health outcomes (e.g., resilience, self-determination, personal growth). Under-exploration in
these areas may partly be reflective of a lesser understanding of “positive mental health”
concepts [188], and a lack of available data [173]. This underrepresentation may also be
partly due to the search methodology applied in this review, which did not include terms like
“mental well-being,” “psychological well-being,” “subjective well-being” or related constructs.

4.1.2 Data Access Challenges: Identify Tradeoffs for Data Sharing that People are Willing to
Make. Bound up with challenges in identifying important health and care needs are
requirements for access to relevant, large-scale, high-quality data to allow for effective ML
modeling. This can be particularly difficult in the mental health domain due to ethical and
privacy challenges involved in (i) recruiting individuals who may be more vulnerable to
research [132] and (ii) the time-consuming and effortful nature of data acquisition that often
requires multi-disciplinary partnerships with health care providers and do not scale easily
[53].

Improving Informed Consent Processes and Users Trust in Data Applications.

For many of the social media studies papers, the pooling of “publicly” available data [e.g., 141,
165, 211] has often been described as not requiring explicit user consent. Recently, there is
however increasing debate on whether the use of public data to predict, e.g., mental health
states, may border on medical diagnosis and should be considered as human subjects research
[29]. This is echoed in user research that suggests that social media users often do not have
awareness that their online content is used for research, and express concerns about such use
“without their consent” [58]. Describing how people attitudes to data use are highly
contextual, Fiesler and Proferes [58] found that Twitter users “felt less comfortable” about
uses of their entire Twitter history (rather than individual tweets), and where content had
more personal significance or sensitivity. They also described ideas of data consent or
permission as stemming from the underlying importance of respect for the user and the need
for data uses (for research or ML applications) to align with user expectations. Although
obtaining consent at scale presents a practical challenge [29, 165], there are increasingly
proposals for how users could, at least, be informed about the use of their data and be given
opportunities to opt-in or opt-out (e.g., by tweeting that their tweet is included in research)
[58]. The feasibility of such approaches will require future testing.

This example, and the need for access to rich, personal data for developing effective ML
models and interventions, also raise the question how to ensure that people generally agree to,
and can trust researchers and data applications with the collection and processing of their
sensitive information? Likely, this requires careful tradeoffs between data needs for
algorithmic purposes and how related data practices are justifiable in terms of benefit or
potentially harm to the person (cf. [119]). For example, while sensitive data such as a person's
gender, age, or clinical diagnosis can aid in differentiating health-behavior patterns and
groups, and enable testing for diversity in a dataset [66, 76, 80]; we have to consider how
comfortable people may feel about providing such data. For this, individuals need to be better
supported in assessing “the potential benefits of data sharing” and “how potential risks are
mitigated through safeguards or outweighed by the potential benefits” (e.g., effectiveness of
interventions). This will enable them to make more informed choices about data uses; and, in
turn, aid their trust in, and acceptance of, data applications.

This might be achieved by (i) making processes of how we seek consent more comprehensive
and usable (in line with GDPR regulations), (ii) explaining more clearly the benefits of data
use to the person and the mechanisms employed to protect their data (taking active steps to
mitigate risks), and (iii) ensuring that people have more control and actual choice(s) about
whether their data is being used for specific ML purposes or not.

Need to Develop Responsible Approaches for Data Sharing and Data Donation.
Difficulties in gaining access to mental health data have also led to proposals to build and
leverage shared infrastructures and data repositories for conducting data research [53].
Creating benchmark datasets [73] and having better methods for data sharing can support the
replication of research findings and improve scientific quality [124]. For example, systems
such as the Clinical Record Interactive Search (CRIS) enable researchers to access large-scale
electronic mental health record data from the UK. To ensure responsible use, applications for
data access are reviewed and monitored by a committee for compliance with ethical and legal
requirements [116]. Similar initiatives exist in the US, e.g., through the Connected and Open
Research Ethics (CORE) program that manages shared health care resources and helps
navigate many of the complex, ethical, and practical challenges involved in collecting sensitive
health care data [191]. Other data collection efforts include crowdsourcing and data donation
programs such as PatientsLikeMe12 and OurDataHelps,13 where people can choose to share
data and information about their health for data science and research purposes. Research
charities are also playing an emerging role by matching researchers and their research
questions to datasets [116] and providing funding for mental health research (e.g., MQ
charity14).

4.2 Evidencing the (Real-World) Effectiveness of ML-

Interventions

With few exceptions [77, 78, 140], the papers in our corpus primarily assessed the
effectiveness of newly developed ML models based on their predictive performance—
measured in terms of accuracy and errors (cf. [149]) and comparison with (state-of-the-art)
baselines—on held-out data. Yet, this often provides little insight as to how reliably a model
may perform in the real world; or how it would find useful adoption within health care
services. As such, these papers predominantly provide proof-of-concept studies that
necessitate continued research and development to further improve (classification) accuracy
[173]. Further, there is little exploration of how developed ML approaches would be perceived
by, and come to actually benefit, their proposed users (e.g., clinicians, patients, online
community moderators).

4.2.1 Beyond Accuracy in Model Performance: Risks of Overclaiming and Premature

Generalization. As is perhaps less surprising in a review of the computing and HCI literature
on “ML applications” in mental health, we found that the majority of papers focused on the
technical or algorithmic development of initial ML models (Section 3.5.1). As such, they
predominantly report their technical contributions through new data methods and accuracy
metrics (Section 3.4.2) and discuss key data modeling challenges (Section 3.5.2). At the same
time, many of these technical papers also include speculative descriptions, proposals, and
claims as to how their new ML models may come to be used to impact clinical or everyday
health management practices (Section 3.6.1). Despite great enthusiasm for how ML
approaches could be transformative to the mental health domain, it is important to not to
prematurely overclaim anticipated (clinical) benefits or generalize too soon from initial
proof-of-concepts. Next, we discuss the importance to (i) acknowledge how the conduct and
impact of research and technological development is assessed and shaped by different
scientific disciplines and (ii) be cautious in making clinical or diagnostic claims where datasets
did not include much, or any, “clinically validated” data (cf. [64, 128]), and where achieved
ML model results were not evaluated or studied in actual health care contexts.

Acknowledging and Addressing Disciplinary Differences in the Conduct and

Evaluation of Research. In computing, research is typically exploratory in nature and
seeks to “find” an answer to a question or problem. In contrast, clinical research tends to be
hypothesis-driven and involve the design of studies to test and “confirm” an answer to a
question [124]. Furthermore, while computer scientists often focus on proof-of-concepts (e.g.,
“Does it work at all?”), clinical scientists value generalizability (e.g., “Does it work for all
populations at all circumstances?”). Often in a quest to identify “novel solutions,” computing
scientists can also have a higher tolerance for risks than clinical researchers, who value
internal validity and confidence in research results [ibid]. Naturally, these disciplinary
differences are reflected in the types of data sources that are used for ML analysis (e.g., clinical
vs. general population/proxy data) as well as the methods that are employed to evaluate the
“success” of the research or development output. This variability complicates the comparison
of findings across studies [124]. It also means that for ML research that seeks to inform
clinical diagnosis and decision-making, it is imperative that algorithmic models are built on
(clinically) valid data [53], perform robustly and reliably outside their training or test
environment (and without discriminating against sub-groups), and assessed for their practical
usefulness and the value they might bring to real-world health care practices (e.g., reduced
clinician time, improved patient outcomes).

Mental Health Constructs and Clinical Validity of ML Results. A significant

proportion of the data collection studies (21 out of 29) did not include any patient data or
external assessments by clinicians and were often conducted as part of lab or pilot studies that
used frequent EMA's to gather “ground truth” data (e.g., a person's mood captured by an
Affect Grid [176]). Data collected in this way can differ significantly from standardized clinical
screening or assessment methods that are administered by trained MHPs. For example, in
social media analysis, there has been an increase in criticism [29, 53, 165] of the use of self-
disclosed, sentence-based labelling such as “I was diagnosed with…” [33, 105, 211, 219] as a
mechanism for “diagnostic” ground truth, as this does not conform with clinical assessment
tools such as the DSM [8]. The DSM provides a written manual for making accurate
psychiatric diagnosis that is based on 60 years of empirical results [29]. Concerns about a lack
of clinical grounding, theoretical contextualization, and psychometric validity were
particularly prominent in the paper by Ernala et al. [53]. Their study compared different
approaches to diagnosing social media users with “schizophrenia” and found poor external
validity where ML models that were based on “proxy information” were tested on people who
had a clinical diagnosis. Additionally, Chancellor et al. [30] also raised concern that many
“mental health status observations” tend to be based on single units of observation (e.g., an
online post) without additional context about an individual or any methodological
substantiation of how a single moment of distress may relate to the presence of a mental
health condition. Many social media studies further imply experimental rigor by including
“control” groups. However, these are often selected as a random sample of online service users
(e.g., [33, 211]), without any (formal) validation that these were individuals who did not have
specific mental health symptoms (e.g., [30]). Outside of social media data, Saeb et al. [164]
also called for caution in the interpretation of ML outputs following their review of studies that
used smartphones and wearable sensors to predict clinical outcomes based on a publicly
available dataset. Having replicated the approaches taken, they found that almost half of the
examined studies used a popular cross-validation method (record-wise cross-validation) that
significantly overestimates the algorithms prediction accuracy.

Thus, for developing effective and implementable ML systems for mental health, and as ML
models advance in technical development and accuracy, more research is needed to (i) test the
validity of the mental health constructs that are assessed (e.g., diagnostic validity) and (ii)
ensure that ML outputs are transferable and their prediction robust for use in “practice”
(reliability). Furthermore, as ML model insights are intended for use and become incorporated
in real-world mental health interventions, future studies have to start assessing (iii) their
practical use, value for, and acceptance by, key stakeholders (cf. [124]); as well as (iv) their
actual effectiveness for improving (promised) mental health outcomes, and reducing costs. In
this regard, it is recommended to involve MHPs and the individuals targeted by ML
predictions throughout the research design and development process. Clincial experts, i.e., can
provide key insights into construct validity, assessments of ground truth and biases, as well as
important context information that can help in the interpretation of data findings, improve
rigor, and manage deployment risks and tradeoffs [29, 53].

4.2.2 Avoid Dehumanization and Undermining the Value of Other Data Methods or Clinical
Expertise. To evaluate the potential usefulness of new ML approaches, it is important to
examine how existing work positions itself and its contributions to mental health research and
practice. Section 3.2 described key motivations for the use of ML for mental health to include
(i) unobtrusive or continuous data access, (ii) automatic data processing for efficiency and cost
savings, and (iii) claims that data-derived assessments provide objective, more accurate and
reliable information to help improve existing (clinical) tools and decision-making practices
(e.g., [61, 192]). The latter argument in particular was often substantiated through an
emphasis on the disadvantages and insufficiencies of traditional questionnaires and self-
report tools (e.g., [153, 211]) as well as clinical approaches (e.g., [2, 63, 82, 134, 136, 139,
144, 145, 179]). Together, these arguments suggest a potential superiority of, and possibilities
for, new data tools to “outperform” existing approaches [192]. Next, we discuss reasons for
why it may be advisable to exercise caution in positioning ML contributions in this way.

Overcoming Methodological Limitations and Improving Insights by Combining

the Strengths of Different Data Methods. Different research and data analysis methods
contribute different types of insights and have their own strengths and limitations. For
example, validated clinical tools present instruments that have been extensively tested
psychometrically to ensure results are both accurate and consistent. The accuracy and
reliability of ML models is, inevitably, limited by the quality of the data used in their training.
ML models are also prone to error, uncertainty, and bias (cf. [190]). Even where ML models
perform with minimal error, challenges remain for their generalization to contexts outside the
specific training environment [106]. Taking these and other described data limitations (see
Section 3.5.2) into consideration, and outside of much evidence of real-world evaluations of
the effectiveness of enabled ML predictions, researchers working in this space need to be
careful with any claims that data-derived assessments indeed provide more “objective,
accurate, and reliable” information. We believe, in these early stages of research and
development, that it is important to set and communicate appropriate expectations of what
ML outputs, to-date, can realistically achieve and what their limitations are. This is
particularly important for setting up successful research collaborations and productive ML
development partnerships with health care providers. Here, a more open dialogue about the
potential and challenges of achieving robust ML models is important for nurturing empathy
and trust. This can pave the way for health care providers to better comprehend what is
required of them, for example, to ensure “good-quality data capture” as well as developing
their understanding of the unique data analytics capabilities that new ML approaches afford.
For example, a key strength of ML methods is their capability to mathematically identify, e.g.,
the most relevant variables in a dataset based on an outcome of interest. In contrast,
conventional statistical methods typically rely on the investigators—their assumptions and
expertise—to specify the variables that are relevant for a particular analysis [100, 194].
Similarly, while studies such as randomized controlled trials (RCTs) can have advantages in
helping to control for, and reduce certain sources of bias when assessing the effectiveness of an
intervention, they reveal little insight as to why or how certain factors contributed (more or
less) to an outcome (cf. [194]). All this suggests the need to better understand what different
research and data methods can explain and contribute to knowledge generation, and how
they could best come to complement (rather than compete with) each other.

Empowering MHPs through Data Insights and Supporting Their Agency as

Health Care Experts. Much of the reported ML work is motivated to help develop new
tools and methods to assist mental health care, which is often provided through MHPs.
Therefore, it is important to be careful in the positioning of new data methods or systems to
not unnecessarily undermine the important role of health or care providers. This can risk
reducing their willingness to support the development as well as adoption and acceptance of
ML approaches into their work practices. Instead, future work should explore how to design
ML-interventions such that they can become valuable tools to assist clinicians in their
information needs and decision-making processes rather than attempting to replace or
outperform them in their clinical expertise.

Avoiding Stigmatization and Dehumanization. In their recent review of ML

approaches used for mental health predictions in social media, Chancellor et al. [30] critiqued
how humans were represented in data research. In various studies, the authors found that
individuals who may not have a mental health condition were often described as “normal” or
“neurotypical.” Such terminology however risks stigmatizing people who have a mental health
condition by othering them and their experiences. Further, the authors described trends in
computationally focused work to treat individuals as “data points” for machine training and
optimization. At an extreme, humans become the “objects” of analysis and represented
through their online “accounts” and “blogs.” In dividing the person from the data, unique
details of their experiences are abstracted away to identify large-scale patterns or phenomena.
Such simplifications are at odds with the complexities and subtleties of people's lives and felt
mental health experiences. While HCI research tends to place the human and their needs—as
the “subject” of interest—at the center of technology design processes, the area of ML—
drawing on statistics, computer science and optimization research—views the abstracted
model or data point as the “object” of study. Yet, within human-centered research, such
objectifications can imply a stronger interest in machine analysis than the people that the
research suggests it is interested to help. As a consequence, this can potentially cloud the
responsibilities and ethical priorities of the researchers [30]. Thus, it is important that
researchers are mindful in their reporting practices to avoid stigmatizing language that can
harm people and communities; and diminish objectification as people and their individual
experiences are being transformed into compressed mathematical representations.

4.3 Understanding Opportunities and Risks of ML-Systems

in Context

As evident in this review, the field of ML in mental health presents an emerging area that, so
far, has mostly contributed to the discovery and development of basic (multi-disciplinary)
research insights, with very few initial investigations of potential ML interventions (cf. [77, 78,
140]). The field of Implementation Science often describes this as the initial stage in what
presents a complex, multifaceted process of moving important research innovation into actual
work flows, and for sustaining and scaling-up effective health care interventions [68]. To
endure on this early journey toward achieving real-world impact, researchers in HCI and ML
will need to (i) continue basic research, (ii) expand development and initial testing of new ML-
interventions, and (iii) start moving towards a rigorous analysis of the effectiveness of these
interventions. To further maximize the usefulness of potential ML interventions, it is
paramount to (iv) more actively include MHPs and people with lived experiences of mental
health in research and development processes (e.g., through observational studies, interviews,
focus groups, and collaborative partnerships), and to develop and study new ML techniques in
real-world settings. Although the majority of review papers presented contributions by multi-
disciplinary research and engineering teams, there was little reporting of user-centered
research methods and explicit dialogue with MHPs and other potential beneficiaries (outside
of data access and labelling efforts) to inform ML research. A stronger collaboration with, and
closer involvement of, key stakeholders will be crucial to identify important health care needs
and scenarios for ML development. Simultaneously, the study of new ML systems will likely
foreground new challenges (e.g., adoption into work practices, ethical issues) that will need to
be considered and addressed if ML-enabled interventions are to succeed in the real world (cf.
[20, 79, 173]).

There remain many challenges in order to move from proof-of-concept explorations toward
the design and study of ML tools and interventions that are useful in broader populations
[124]. To help move toward this goal, our review foregrounded the following two areas of
research and development that require further consideration: (i) the need to better support
laypeople's understanding of ML outputs and (ii) to recognize and appropriately respond to
broader practical and ethical implications that can be bound up with the use of these
interventions in real-world mental health care contexts.
4.3.1 Design to Support Appropriate Understanding and Use of ML-Outputs by Laypeople. A
key challenge in the design of ML-enabled systems is the generation of outputs that are
interpretable and (clinically) useful to mental health care providers or target recipients. To
address these challenges, work in this area often includes methods which support
“understanding of the model” (e.g., [106, 149]). They include the following: extracting (and
visualizing) model outputs and properties; estimating the influence of training examples; or
learning local approximations to explain individual predictions of complex models post-hoc.
Beyond this more data-driven understanding of methods, interpretability is mostly
understood in terms of end-users being able to simulate, trust, or debug model decisions [1,
78, 149], and designing interactions with intelligent systems that can aid human
understanding and decision making (cf. [20]). In our corpus only one paper [78] engaged
explicitly with this topic of ML intelligibility. The authors [ibid] described the tendency of
participants to perceive and uncritically accept ML-generated outputs as “factual information,”
even when they acknowledged that they did not fully understand how feedback was derived, or
what it precisely meant. Their willingness to trust the ML system was found to be bound up
with the perceived “legibility” of the output—the extent to which the system seemed to “make
sense” to the user—rather than its statistical accuracy. This demonstrates the need for more
research investigating how an appropriate interpretation of ML outputs by laypeople can be
supported (see also recent work by [212]).

To support laypeople's understanding of how specific (behavior) data and model results relate
to a health outcome, technical or mathematical explanations of model accuracy or uncertainty
however might be limited. For example, how should MHPs interpret the significance of a
prediction that indicates e.g., an 83% risk of suicide; and how can they meaningfully
differentiate this from a 78% or 88% risk prediction? To enable and support laypeople to
appropriately make use of ML outputs, user interface design and interactive visualizations or
simulations can play a key role in generating comprehensive mappings for users, and help
them assess, inspect, and cross-validate ML outputs in line with their own assessments of a
situation. This is needed to better enable laypeople to calibrate their understanding of a
system's capabilities and limitations to reduce risks of over-reliance on potentially over-
confident predictions [109, 111]. To support scrutiny and encourage more careful
interpretations of ML interferences, this suggests the need for (i) stronger efforts in supporting
peoples’ awareness of the probabilistic (rather than deterministic) nature of many ML models,
and their likely proneness to errors, (ii) the provision of relevant additional context
information and evidence that can help users to affirm, or contest ML outputs [77], and (iii)
the inclusion of opportunities for user input and a strengthening of their role as data
controllers [7] through encouragements to ask questions, to inspect any conclusions that seem
unreasonable; and to facilitate the recording of any disagreements with a system (cf. [77]), or
even correct identified errors. For early examples of HCI approaches to assist the
interpretation and use of ML-enabled interventions, see recent work in health care more
broadly, such as personalized fitness apps [52] or clinical support tools in critical surgery
decisions [210].

4.3.2 Recognize and Respond to Broader Practical and Ethical Implications of ML-System
Use. ML systems are increasingly becoming “real” [79] and embedded in high-stakes domains
like mental health care [77, 78], where they can have significant implications for people's
lives. Thus, we need to give close consideration to the practical challenges and broader, often
un-anticipated ethical risks that can be bound up with the design and deployment of ML-
interventions for mental health, and pro-actively work to mitigate risks associated with their
use in practice.

Across the review papers, we found generally little discussion of ethical issues outside reports
of formal research approvals, user study considerations, and the adoption of risk-averse and
privacy protecting data management practices. Few papers engaged with the broader
implications of using developed ML models within health care practice; mostly when reporting
errors in model predictions [53, 134, 218]. This included discussions of: how ML systems
may implicate the relationship between different stakeholders (e.g., patients, clinicians,
supervisors) [78]; how ML algorithms misclassified or did not at all detect certain individuals
[218], and how a false-alarm of a high suicide risk alert, could—at an extreme—lead to a
person's involuntary hospitalization [134]. Even in less extreme cases, the false identification
of a mental health condition could have severe implications for a person's self-esteem,
reputation or employment (particularly for people working, i.e., in the police force, as pilots)
[29, 189]. This raises the question of who is responsible and accountable for errors and for
making choices if and how individuals should be alerted to their own mental health status
[29]. Within the broader literature on HCI in digital mental health, researchers have started to
discuss the challenges involved in making people aware of machine-detected problems in ways
that are sensible, and respond carefully to peoples’ expectations, needs, or troubles (e.g., [170,
216]). Such challenges are particularly prevalent in contexts where behavioral analysis is done
outside of explicit user awareness (e.g., mental health inferences drawn from a person's social
media). For example, Young and Garett [216] outlined a first working protocol that suggests
when, and which stakeholders should intervene (or not), in the case that people were found to
express suicidal intentions on social media. This, and similar works [29], acknowledge the
need for researchers to have a process in place for supporting people who are identified as “at
risk,” including who to contact, and what information to share to address (psychological)
concerns.

Thus, to help realize the potential of ML to truly benefit people, and find acceptance by them,
requires researchers to (i) engage in more open discourse about the opportunities as well as
ethical difficulties bound up with the use of ML for specific mental health contexts and (ii)
extend efforts to collaborate more closely with health care users and/or providers throughout
ML system design and evaluation processes [29, 30].
4.4 Limitations

The corpus included in this review is by no means complete, and new work constantly
emerging. We acknowledge that the implications of this work are limited by our search
methodology that was restricted to broad terms (“mental health” and “ML”) as well as our
record selection criteria. As such, our work excludes important research and ML development
for neurological and neurodevelopmental conditions and may under-represent other mental
health related works that focus, for example, on preventative and more well-being-centric
approaches. The article also has limitations due its focus on the computing and HCI literature
through the ACM Guide. This was a deliberate decision to provide a report that focused on the
current landscape of computing research where ML has become applied in the context of
mental health. It enabled the inclusion of more in-depth descriptions of existing research and
development, as well as rich discussions of current trends and important challenges with
regards to data access, conceptualization and modeling of mental health behaviors, and
broader ethical and real-world implementation and use considerations. To reduce risks of bias
in data collection (i) identified paper records were independently screened by two researchers
and disagreements resolved through full-text review and discussion and (ii) a data extraction
sheet was used to systematically elicit key information from each paper. Care was also taken in
reporting the findings to balance the accounts of different approaches and findings; with
reports and interpretations continually reviewed and re-evaluated by all members of the
research team.

5 CONCLUSION
Recent years have witnessed an increase in excitement and exploratory research on potential
applications of ML for mental health. Our review has offered an overview of this area of
research and highlighted current trends and challenges. Aiming to shape the future direction
of work, we have discussed current approaches and potential steps toward achieving ML
systems that are effective and implementable for mental health care.

Specifically, we have examined how constraints and requirements for access to large-scale,
high-quality data can pose challenges to study design and urge researchers to extend efforts to
gain more in-depth understanding of the specific needs or challenges that are faced by MHPS
and people with lived mental health experiences. Deeper and more creative explorations of the
design space can meaningfully inform future research questions and problem scenarios for ML
to ensure the domain can truly benefit from novel data tools. This may extend beyond more
obvious ML applications for mental health. Bound-up with data access is the need to better
assist people in assessing potential benefits of data sharing and how potential risks are
mitigated or outweighed by potential benefits (e.g., effectiveness of interventions), such that
they can make more informed choices about data uses and to aid their trust in, and acceptance
of, data applications.

Furthermore, since the field of ML in mental health is still in its infancy, we have urged for
more caution in presentations of ML development to avoid premature claims on the potential
usefulness and real-world impact of new models. This is especially important considering the
complexity and difficulties involved in generating robust as well as technically and clinically
reliable ML outputs. So far, the majority of models are rarely tested for use in clinical
environments, leaving gaps in assessments of their practicality, acceptance, and effectiveness
for improving mental health-related outcomes, or services.

In addition, while it was often argued in the literature that novel ML models have advantages
over existing research and clinical methods, we suggested to look at these as complementary
approaches to knowledge generation. Furthermore, we proposed that there is a lot more scope
for future research to also extend explorations of how ML-interventions can become valuable
tools to address the needs not only of mental health care recipients, but to support the
practices of mental health care experts. In applying ML approaches to the capture and
assessment of rich human needs and experiences, researchers should also be mindful to not
translate and abstract away too much from the individual person and their unique context in
data analysis, interpretation, and representation.

Finally, we argued that helping the field achieve its many ambitious visions for ML in mental
health requires continued efforts in conducting basic, multi-disciplinary research in deep
collaboration with health partners, developing and testing new ML-interventions, and
studying their effectiveness within real-world use contexts. This includes a key focus on the
challenges of designing new ML-enabled systems that are sufficiently interpretable and
(clinically) useful to its target users or recipients. It also requires that research and
development efforts recognize and carefully respond to the broader practical and ethical
implications that the use of ML systems could have for people, health care, and society.

A APPENDIX
A.1 Examples of Common ML Models or Techniques in
Each ML Algorithm Category
 Semi-
Supervised Unsupervised supervised Novel Methods

ML Support Vector k-means clustering Semi- Deep learning

models/ Machines (SVM) Hierarchical supervised ML (DL)
techniques k-Nearest clustering Self-training Active learning,
neighbors (k- Unsupervised Mixture i.e.,
NN) Hidden Markov models Reinforcement
Naïve Bayes Models (HMM) Co-training + Learning (RL)
(NB) Latent Dirichlet multi-view Custom-ML
Regression Allocation (LDA) learning methods
analysis, e.g., Unsupervised Graph-based
Logistic Neural networks methods
Regression (LR), (NN)
Lasso Association rule
Supervised techniques
Latent Dirichlet
Allocation
Decision Trees
(DT)
Random Forests
(RF)
Supervised
Hidden Markov
Models (HMM)
Supervised
Neural networks
(NN)

A.2 Frequency and Types of Specific ML/Data-Challenges

and Limitation Described

Category Subcategory Detail Paper/Author(s)

Capturing Need for ground No clear definition + Gaur et al. [64]; Gjoreski
accurate/ truth, robust reliable measure of et al. [67]; Nosakhare and
reliable labels, and subjective non-discrete Picard [135]; Rabbi et al.
data validation experiences [153]

Challenges in generating Kavuluru et al. [89]; Ray

Category Subcategory Detail Paper/Author(s)
low-dimensional, et al. [155]; Salekin et al.
meaningful data labels [168]

Lack of clinical Chang et al. [31]; Ernala

validation/information to et al. [53]; Nguyen et al.
infer mental health [133]; Saha and De
Choudhury [165]; Salekin
et al. [168]

Ecological validity: Broek et al. [23]; Diedrich

Transferability of data et al. [45: Study 1];
(differences of lab- vs. real- Gjoreski et al. [67];
world data) Zakaria et al. [218]

Category Subcategory Detail Paper/Author(s)

Noisy/ ambiguous Ambiguous words/lexical Chancellor [28];

signals variations Nobles et al. [134];
Saha and De
Choudhury [165];
Yazdavar et al. [211]

Ambiguity in signals (e.g., Chang et al. [31];

for audio: robust speaker Mallol-Ragolta et al.
detection; distinguish [112]; Rabbi et al.
personal speaking style [153]; Salekin et al.
from symptoms) [168]; Spathis et al.
[176, 177]; Zhou et al.
[222]

Managing irrelevant, Ojeme and Mbogho

redundant information [136]

Dataset Restrictions due to Too small or restricted Adamou et al. [2];

limitations data study sample/need for Diederich et al. [45];
subjects/scale/study larger (more diverse) Feng et al. [61],
context datasets Kavuluru et al. [89];
Morshed et al. [128],
Nobles et al. [134];
Ojeme and Mbogho
[136]; Parades et al.
 Category Subcategory Detail Paper/Author(s)
[140]; Park et al.
[141], Pestian et al.
[145]; Quisel et al.
[152]; Ray et al. [155];
Salekin et al. [168];
Spathis et al. [177],
Yazdavar et al. [211];
Zhou et al. [222]

Unknown confounding Fatima et al. [57];

variables + limitations of Saha and De
study context Choudhury [165];
Salekin et al. [168]

Reference dataset not Alam et al. [6]

explicitly designed for
mental health-related
analysis

Biased, missing, General acknowledgement Ernala et al. [53];

incomplete data of biases inherent to model Hirsch et al. [78]; Park
design and datasset used et al. [141]
for training

Difficulties due to missing Alam et al. [6]; Spathis

data values/sparse data et al. [176, 177]

Need for inclusion of other Diedrich et al. [45];

information (e.g., Pestian et al. [145];
biological and genetic data, Mallol-Ragolta et al.
fMRI, video, facial [112]; Morshed et al.
expressions, social media [128]
data)

Data Continuous data Identifying optimal time- Frogner et al. [62];

processing segments/features for data Mallol-Ragolta et al.
analysis [112]

Data integration Modeling multi-modal data Cao et al. [27]; Mitra

challenges of different signals, et al. [122]; Morshed
durations, densities; data et al. [128],
fusion challenges Panagiotakopoulos et
al. [139]; Ray et al.
 Category Subcategory Detail Paper/Author(s)
[155]; Rastogi et al.
[154]; Tran et al.
[192]; Zhou et al.
[222]

Complex mapping of multi- Ojeme and Mbogho

label classification [134]

Limitations Modeling approach Detection of presence, Yazdavar et al. [211]

of ML chosen duration + frequency of
modeling/ symptoms (not severity)
implications
Use of retrospective data Patterson and Cloud
for predicting future [144]
behavior

Focus on population rather Doryab et al. [50];

than individual Gjoreski et al. [67];
Nguyen et al. [133];
Quisel et al. [152]

Claims Limited ability to make Morshed et al. [128];

causal claims Saha and De
Choudhury [164];
Wang et al. [201]

Errors Errors in Doryab et al. [50];

classifications/predictions Ernala et al. [53];
and fallibility of ML Hirsch et al. [78];
Systems Nobles et al. [134],
Zakaria et al. [218]

Other Need for data Secure storage and Alam et al. [6]; Rabbi
security handling of data/need for et al. [153]; Jain and
secure models Agarwal [82]; Wang et
al. [201]
REFERENCES
[1] Ashraf Abdul, Jo Vermeulen, Danding Wang, Brian Y. Lim, and Mohan Kankanhalli.
2018. Trends and trajectories for explainable, accountable and intelligible systems: An
HCI research agenda. In Proceedings of the 2018 CHI Conference on Human Factors
in Computing Systems (CHI’18). ACM, Paper 582, 18 pages. DOI:
https://doi.org/10.1145/3173574.3174156 Navigate to

[2] Marios Adamou, Grigoris Antoniou, Elissavet Greasidou, Vincenzo Lagani, Paulos
Charonyktakis, and Ioannis Tsamardinos. 2018. Mining free-text medical notes for
suicide risk assessment. In Proceedings of the 10th Hellenic Conference on Artificial
Intelligence (SETN’18). ACM, Paper 47, 8 pages. DOI:
https://doi.org/10.1145/3200947.3201020 Navigate to

[3] Ehsan Adeli, Kim-Han Thung, Le An, Guorong Wu, Feng Shi, Tao Wang, and
Dinggang Shen. 2018. Semi-supervised discriminative classification robust to sample-
outliers and feature-noises. IEEE Transactions on Pattern Analysis and Machine
Intelligence 41, 2 (2018), 515–522. DOI:
https://doi.org/10.1109/TPAMI.2018.2794470 Navigate to

[4] Jesus S. Aguilar-Ruiz, Raquel Costa, and Federico Divina. 2004. Knowledge discovery
from doctor–patient relationship. In Proceedings of the 2004 ACM symposium on
Applied computing (SAC’04). ACM, 280–284. DOI:
https://doi.org/10.1145/967900.967960 Navigate to

[5] Malik Anas Ahmad, Nadeem Ahmad Khan, and Waqas Majeed. 2014. Computer-
assisted analysis system of electroencephalogram for diagnosing epilepsy. In
Proceedings of the 22nd International Conference on Pattern Recognition (ICPR’14).
IEEE Computer Society, Washington, DC, 3386–3391. DOI:
https://doi.org/10.1109/ICPR.2014.583 Navigate to

[6] Md. Golam Rabiul Alam, Eung Jun Cho, Eui-Nam Huh, and Choong Seon Hong. 2014.
Cloud based mental state monitoring system for suicide risk reconnaissance using
wearable bio-sensors. In Proceedings of the 8th International Conference on
Ubiquitous Information Management and Communication (ICUIMC’14). ACM, Paper
56, 6 pages. DOI: https://doi.org/10.1145/2557977.2558020 Navigate to

[7] Saleema Amershi, Maya Cakmak, William Bradley Knox, and Todd Kulesza. 2014.
Power to the people: The role of humans in interactive machine learning. AI Magazine
35, 4 (2014), 105–120. DOI: https://doi.org/10.1609/aimag.v35i4.2513
Navigate to

[8] American Psychiatry Association. 2019. Diagnostic and Statistical Manual of Mental
Disorders (DSM–5). Last retrieved 7 th July 2019 from
 https://www.psychiatry.org/psychiatrists/practice/dsm. Navigate to

[9] Elena M. Andresen, Judith A. Malmgren, William B. Carter, and Donald L. Patrick.
1994. Screening for depression in well older adults: Evaluation of a short form of the
CES-D. American Journal of Preventive Medicine 10, 2 (1994), 77–84.
Navigate to

[10] Mohammad R. Arbabshirani, Sergey Plis, Jing Sui, and Vince D. Calhoun. 2017.
Single subject prediction of brain disorders in neuroimaging: Promises and pitfalls.
Neuroimage 145, PT B (2017), 137–165. DOI:
https://doi.org/10.1016/j.neuroimage.2016.02.079 Navigate to

[11] Greg Barish, Hilary Aralis, Eric Elbogen, and Patricia Lester. 2019. A mobile app for
patients and those who care about them: A case study for veterans with PTSD + Anger.
In Proceedings of the 13th EAI International Conference on Pervasive Computing
Technologies for Healthcare (PervasiveHealth’19). ACM, 1–10. DOI:
https://doi.org/10.1145/3329189.3329248 Navigate to

[12] Jakob E. Bardram, Mads Frost, Károly Szántó, Maria Faurholt-Jepsen, Maj Vinberg,
and Lars Vedel Kessing. 2013. Designing mobile health technology for bipolar
disorder: A field trial of the monarca system. In Proceedings of the SIGCHI Conference
on Human Factors in Computing Systems (CHI’13). ACM, 2627–2636. DOI:
https://doi.org/10.1145/2470654.2481364 Navigate to

[13] Reid Bates. 2004. A critical analysis of evaluation practice: The Kirkpatrick model
and the principle of beneficence. Evaluation and Program Planning 27, 3 (2004),
341–347. DOI: https://doi.org/10.1016/j.evalprogplan.2004.04.011
Navigate to

[14] Matthew J. Bauman, Kate S. Boxer, Tzu-Yun Lin, Erika Salomon, Hareem Naveed,
Lauren Haynes, Joe Walsh, Jen Helsby, Steve Yoder, Robert Sullivan, Chris Schneweis,
and Rayid Ghani. 2018. Reducing incarceration through prioritized interventions. In
Proceedings of the 1st ACM SIGCAS Conference on Computing and Sustainable
Societies (COMPASS’18). ACM, Paper 6, 8 pages. DOI:
https://doi.org/10.1145/3209811.3209869 Navigate to

[15] Victoria Bellotti, and Keith Edwards. 2001. Intelligibility and accountability: Human
considerations in context-aware systems. Human-Computer Interaction 16, 2–4
(2001), 193–212. DOI: https://doi.org/10.1207/S15327051HCI16234_05
Navigate to

[16] Charles C. Benight, Kotaro Shoji, Lori E. James, Edward E. Waldrep, Douglas L.
Delahanty, and Roman Cieslak. 2015. Trauma Coping Self-Efficacy: A context specific
self-efficacy measure for traumatic stress. Psychological Trauma: Theory, Research,
Practice, and Policy 7, 6 (2015), 591. Navigate to
[17] Hugh Beyer, and Karen Holtzblatt. 1997. Contextual design: Defining customer-
centered systems. In Principles of Contextual Inquiry (Chapter 3). Elsevier, 41–66.
Navigate to

[18] Vincent Bindschaedler, Paul Grubbs, David Cash, Thomas Ristenpart, and Vitaly
Shmatikov. 2018. The tao of inference in privacy-protected databases. Proceedings of
the VLDB Endowment 11, 11 (2018), 1715–1728. DOI:
https://doi.org/10.14778/3236187.3236217 Navigate to

[19] Sarah Bird, Krishnaram Kenthapadi, Emre Kiciman, and Margaret Mitchell. 2019.
Fairness-aware machine learning: Practical challenges and lessons learned. In
Proceedings of the 12th ACM International Conference on Web Search and Data
Mining (WSDM’19). ACM, 834–835. DOI:
https://doi.org/10.1145/3289600.3291383 Navigate to

[20] Daniel Bone, Chi-Chun Lee, Theodora Chaspari, James Gibson, and Shrikanth
Narayanan. 2017. Signal processing and machine learning for mental health research
and clinical applications [perspectives]. IEEE Signal Processing Magazine 34, 5
(2017), 196–195. DOI: https://doi.org/10.1109/MSP.2017.2718581
Navigate to

[21] Katia Bourahmoune and Toshiyuki Amagasa. 2019. AI-powered posture training:
Application of machine learning in sitting posture recognition using the lifechair smart
cushion. In Proceedings of the 28th International Joint Conference on Artificial
Intelligence (IJCAI’19). Sarit Kraus (Ed.), AAAI Press, 5808–5814. Navigate to

[22] Leo Breiman and Adele Cutler. 2007. Random Forests-Classification Description.
Department of Statistics, University of California, Berkeley. Navigate to

[23] Egon L. Broek, Frans Sluis, and Ton Dijkstra. 2013. Cross-validation of bimodal
health-related stress assessment. Personal Ubiquitous Computing 17, 2 (2013), 215–
227. DOI: http://dx.doi.org/10.1007/s00779-011-0468-z Navigate to

[24] Joy Buolamwini, and Timnit Gebru. 2018. Gender shades: Intersectional accuracy
disparities in commercial gender classification. In Proceedings of the Conference on
Fairness, Accountability and Transparency (FAT’18). 77–91. Navigate to

[25] Danilo Bzdok, and Andreas Meyer-Lindenberg. 2018. Machine learning for precision
psychiatry: Opportunities and challenges. Biological Psychiatry: Cognitive
Neuroscience and Neuroimaging 3, 3 (2018), 223–230. DOI:
https://doi.org/10.1016/j.bpsc.2017.11.007 Navigate to

[26] Longbing Cao, Philip S. Yu, Chengqi Zhang, and Huaifeng Zhang. 2008. Data
Mining for Business Applications ( 1st ed .). Springer Publishing Company,
Incorporated. Navigate to
[27] Bokai Cao, Lei Zheng, Chenwei Zhang, Philip S. Yu, Andrea Piscitello, John Zulueta,
Olu Ajilore, Kelly Ryan, and Alex D. Leow. 2017. DeepMood: Modeling mobile phone
typing dynamics for mood detection. In Proceedings of the 23rd ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining (KDD’17). ACM,
747–755. DOI: https://doi.org/10.1145/3097983.3098086 Navigate to

[28] Stevie Chancellor. 2018. Computational methods to understand deviant mental

wellness communities. In Proceedings of the Extended Abstracts CHI 2018. ACM,
Paper DC05. https://doi.org/10.1145/3170427.3173021 Navigate to

[29] Stevie Chancellor, Michael L. Birnbaum, Eric D. Caine, Vincent M. B. Silenzio, and
Munmun De Choudhury. 2019. A taxonomy of ethical tensions in inferring mental
health states from social media. In Proceedings of the Conference on Fairness,
Accountability, and Transparency (FAT’19). ACM, 79–88. DOI:
https://doi.org/10.1145/3287560.3287587 Navigate to

[30] Stevie Chancellor, Eric P. S. Baumer, and Munmun De Choudhury. 2019. Who is the
“human” in human-centered machine learning: The case of predicting mental health
from social media. Proceedings of the ACM on Human-Computer Interaction 3,
CSCW, Article 147 (November 2019) , 32 pages. DOI:
https://doi.org/10.1145/3359249 Navigate to

[31] Keng-hao Chang, Matthew K. Chan, and John Canny. 2011. AnalyzeThis: Unobtrusive
mental health monitoring by voice. In Proceedings of the CHI’11 Extended Abstracts
on Human Factors in Computing Systems (CHI EA’11). ACM, 1951–1956. DOI:
https://doi.org/10.1145/1979742.1979859 Navigate to

[32] Ritu Chauhan, and Harleen Kaur. 2017. A feature-based selection technique for
reduction of large scale data. International Journal of Data Analysis Techniques and
Strategies 9, 3 (2017), 207–221. DOI:
https://doi.org/10.1504/IJDATS.2017.086630 Navigate to

[33] Xuetong Chen, Martin D. Sykora, Thomas W. Jackson, and Suzanne Elayan. 2018.
What about mood swings: Identifying depression on twitter with temporal measures of
emotions. In Proceedings of the Companion the Web Conference 2018 (WWW’18).
1653–1660. DOI: https://doi.org/10.1145/3184558.3191624 Navigate to

[34] Prerna Chikersal, Danielle Belgrave, Gavin Doherty, Angel Enrique, Jorge E.
Palacios, Derek Richards, and Anja Thieme. 2020. Understanding client support
strategies to improve clinical outcomes in an online mental health intervention. In
Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems
(CHI’20). ACM. DOI: http://dx.doi.org/10.1145/3313831.3376341
Navigate to

[35] James F. Childress, Ruth R. Faden, Ruth D. Gaare, Lawrence O. Gostin, Jeffrey Kahn,
 Richard J. Bonnie, Nancy E. Kass, Anna C. Mastroianni, Jonathan D. Moreno, and
Phillip Nieburg. 2002. Public health ethics: Mapping the terrain. The Journal of Law,
Medicine & Ethics 30, 2 (2002), 170–178. DOI: https://doi.org/10.1111/j.1748-
720X.2002.tb00384.x Navigate to

[36] Pietro Cipresso, Silvia Serino, Yuri Ostrovsky, and Justin T. Baker. 2018. Pervasive
computing paradigms for mental health. In Proceedings of the 7th International
Conference, Mindcare 2018 ( 1st ed .). Springer Publishing Company, Incorporated.
Navigate to

[37] Moustapha M. Cisse, Yossi Adi, Natalia Neverova, and Joseph Keshet. 2017. Houdini:
Fooling deep structured visual and speech recognition mossdels with adversarial
examples. In Proceedings of the s Neural Information Processing Systems (NIPS’17).
6977–6987. Navigate to

[38] Sheldon Cohen, Tom Kamarck, and Robin Mermelstein. 1983. A global measure of
perceived stress. Journal of Health and Social Behavior 24, 4 (1983), 385–396. DOI:
https://www.jstor.org/stable/2136404 Navigate to

[39] Jason A. Colquitt, and Jessica B. Rodell. 2015. Measuring justice and fairness. In
Oxford Handbook of Justice in the Workplace. Vol. 187, Oxford University Press, 202.
Navigate to

[40] Glen Coppersmith, Kim Ngo, Ryan Leary, and Anthony Wood. 2016. Exploratory
analysis of social media prior to a suicide attempt. In Proceedings of the 3rd Workshop
on Computational Lingusitics and Clinical Psychology, 106–117. DOI:
https://www.aclweb.org/anthology/W16-0311 Navigate to

[41] David Coyle, Anja Thieme, Conor Linehan, Madeline Balaam, Jayne Wallace, and
Siân Lindley. 2014. Emotional wellbeing. International Journal of Human Computer
Studies 8, 72 (2014), 627−628. DOI:
http://dx.doi.org/10.1016/j.ijhcs.2014.05.008 Navigate to

[42] Munmun De Choudhury, Emre Kiciman, Mark Dredze, Glen Coppersmith, and
Mrinal Kumar. 2016. Discovering shifts to suicidal ideation from mental health content
in social media. In Proceedings of the 2016 CHI Conference on Human Factors in
Computing Systems (CHI’16). ACM, 2098–2110. DOI:
https://doi.org/10.1145/2858036.2858207 Navigate to

[43] Drew DeHaas, Jesse Craig, Colin Rickert, Margaret J. Eppstein, Paul Haake, and
Kirsten Stor. 2007. Feature selection and classification in noisy epistatic problems
using a hybrid evolutionary approach. In Proceedings of the 9th Annual Conference on
Genetic and Evolutionary Computation (GECCO’07). ACM, 1872–1872. DOI:
https://doi.org/10.1145/1276958.1277331 Navigate to

[44] Orianna DeMasi and Benjamin Recht. 2017. A step towards quantifying when an
 algorithm can and cannot predict an individual's wellbeing. In Proceedings of the 2017
ACM International Joint Conference on Pervasive and Ubiquitous Computing and
Proceedings of the 2017 ACM International Symposium on Wearable Computers.
763–771. DOI: https://doi.org/10.1145/3123024.3125609 Navigate to

[45] Joachim Diederich, Aqeel Al-Ajmi, and Peter Yellowlees. 2007. Ex-ray: Data mining
and mental health. Applied Soft Computing 7, 3 (2007), 923–928. DOI:
http://dx.doi.org/10.1016/j.asoc.2006.04.007 Navigate to

[46] Ed Diener, Derrick Wirtz, William Tov, Chu Kim-Prieto, Dong-won Choi, Shigehiro
Oishi, and Robert Biswas-Diener. 2010. New well-being measures: Short scales to
assess flourishing and positive and negative feelings. Social Indicators Research 97, 2
(2010), 143–156. DOI: https://doi.org/10.1007/s11205-009-9493-y
Navigate to

[47] Loretta Dipietro, Carl J. Caspersen, Adrian M. Ostfeld, and Ethan R. Nadel. 1993. A
survey for assessing physical activity among older adults. Medicine & Science in Sports
& Exercise 25, 5 (1993), 628–642. DOI: http://dx.doi.org/10.1249/00005768-
199305000-00016 Navigate to

[48] Kevin Doherty, José Marcano-Belisario, Martin Cohn, Nikolaos Mastellos, Cecily
Morrison, Josip Car, and Gavin Doherty. 2019. Engagement with mental health
screening on mobile devices: Results from an antenatal feasibility study. In
Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems
(CHI’19). ACM, Paper 186, 15 pages. DOI:
https://doi.org/10.1145/3290605.3300416 Navigate to

[49] Gavin Doherty, David Coyle, and John Sharry. 2012. Engagement with online mental
health interventions: An exploratory clinical study of a treatment for depression. In
Proceedings of the SIGCHI Conference on Human Factors in Computing Systems
(CHI’12). ACM, 1421–1430. DOI: https://doi.org/10.1145/2207676.2208602
Navigate to

[50] Afsaneh Doryab, Mads Frost, et al. 2015. Impact factor analysis: Combining
prediction with parameter ranking to reveal the impact of behavior on health outcome.
Personal Ubiquitous Computing 19, 2 (2015), 355–365. DOI:
http://dx.doi.org/10.1007/s00779-014-0826-8 Navigate to

[51] Georgios Drakos. 2018. How to select the right evaluation metric for machine
learning models: Part 1 regression metrics. Towards Data Science. Last retrieved 6th
of July 2019 from https://towardsdatascience.com/how-to-select-the-right-
evaluation-metric-for-machine-learning-models-part-1-regrression-
metrics-3606e25beae0. Navigate to

[52] Malin Eiband, Hanna Schneider, Mark Bilandzic, Julian Fazekas-Con, Mareike Haug,
 and Heinrich Hussmann. 2018. Bringing Transparency Design into Practice. In
Proceedings of the 23rd International Conference on Intelligent User Interfaces
(IUI’18). ACM, 211–223. DOI: https://doi.org/10.1145/3172944.3172961
Navigate to

[53] Sindhu Kiranmai Ernala, Michael L. Birnbaum, Kristin A. Candan, Asra F. Rizvi,
William A. Sterling, John M. Kane, and Munmun De Choudhury. 2019. Methodological
gaps in predicting mental health states from social media: Triangulating diagnostic
signals. In Proceedings of the 2019 CHI Conference on Human Factors in Computing
Systems (CHI’19). ACM, Paper 134, 16 pages. DOI:
https://doi.org/10.1145/3290605.3300364 Navigate to

[54] Emad N. Eskandar and Barry J. Richmond. 1991. Decoding of neuronal signals in
visual pattern recognition. In Proceedings of the 4th International Conference on
Neural Information Processing Systems (NIPS’91). J. E. Moody, S. J. Hanson, and R.
P. Lippmann (Eds.), Morgan Kaufmann Publishers Inc., San Francisco, CA, 356–363.
Navigate to

[55] Deborah Estrin, J. P. Pollak, and Tauhidur Rahman. 2017. MOBISYS MobiSys:
Mobile systems, applications, and services. In Proceedings of the 1st Workshop on
Digital Biomarkers. ACM. https://dl.acm.org/citation.cfm?id=3089341.
Navigate to

[56] Harris A. Eyre, Ajeet B. Singh, and Charles Reynolds III . 2016. Tech giants enter
mental health. World Psychiatry 15, 1 (2016), 21–22. DOI:
https://dx.doi.org/10.1002%2Fwps.20297 Navigate to

[57] Iram Fatima, Hamid Mukhtar, Hafiz Farooq Ahmad, and Kashif Rajpoot. 2018.
Analysis of user-generated content from online social communities to characterise and
predict depression degree. Journal of Information Science 44, 5 (October 2018), 683–
695. DOI: https://doi.org/10.1177/0165551517740835 Navigate to

[58] Casey Fiesler, and Nicholas Proferes. 2018. “Participant” perceptions of Twitter
research ethics. Social Media + Society 4, 1 (2018), 1–14. DOI:
https://doi.org/10.1177%2F2056305118763366 Navigate to

[59] Thomas Filk and Albrecht von Müller. 2008. Evolutionary learning of small
networks. Complex 13, 3 (2008), 43–54. DOI:
http://dx.doi.org/10.1002/cplx.v13:3 Navigate to

[60] David Feil-Seifer and Maja J. Matarić. 2012. Distance-based computational models
for facilitating robot Interaction with children. Journal of Human-Robot Interaction 1,
1 (2012), 55–77. https://doi.org/10.5898/JHRI.1.1.Feil-Seifer. Navigate to

[61] Chaonan Feng, Huimin Gao, Xuefeng B. Ling, Jun Ji, and Yantao Ma. 2018. Shorten
bipolarity checklist for the differentiation of subtypes of bipolar disorder using
 machine learning. In Proceedings of the 2018 6th International Conference on
Bioinformatics and Computational Biology (ICBCB’18). ACM, 162–166. DOI:
https://doi.org/10.1145/3194480.3194508 Navigate to

[62] Joakim Ihle Frogner, Farzan Majeed Noori, Pål Halvorsen, Steven Alexander Hicks,
Enrique Garcia-Ceja, Jim Torresen, and Michael Alexander Riegler. 2019. One-
dimensional convolutional neural networks on motor activity measurements in
detection of depression. In Proceedings of the 4th International Workshop on
Multimedia for Personal Health & Health Care (HealthMedia’19). ACM, 9–15. DOI:
https://doi.org/10.1145/3347444.3356238 Navigate to

[63] Dimitrios Galiatsatos, Georgia Konstantopoulou, George Anastassopoulos, Marina

Nerantzaki, Konstantinos Assimakopoulos, and Dimitrios Lymberopoulos. 2015.
Classification of the most significant psychological symptoms in mental patients with
depression using bayesian network. In Proceedings of the 16th International
Conference on Engineering Applications of Neural Networks (INNS) (EANN’15).
Lazaros Iliadis and Chrisina Jane (Eds.), ACM, Paper 15, 8 pages. DOI:
https://doi.org/10.1145/2797143.2797159 Navigate to

[64] Manas Gaur, Ugur Kursuncu, Amanuel Alambo, Amit Sheth, Raminta Daniulaityte,
Krishnaprasad Thirunarayan, and Jyotishman Pathak. 2018. “Let me tell you about
your mental health!”: Contextualized classification of reddit posts to DSM-5 for web-
based intervention. In Proceedings of the 27th ACM International Conference on
Information and Knowledge Management (CIKM’18). ACM, 753–762. DOI:
https://doi.org/10.1145/3269206.3271732 Navigate to

[65] Marzyeh Ghassemi, Tristan Naumann, Peter Schulam, Andrew L. Beam, and Rajesh
Ranganath. 2018. Opportunities in machine learning for healthcare.
arXiv:1806.00388. DOI: https://arxiv.org/abs/1806.00388 Navigate to

[66] Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan,
Hanna Wallach, Hal Daumeé III , and Kate Crawford. 2018. Datasheets for datasets.
arXiv:1803.09010. DOI: https://arxiv.org/abs/1803.09010 Navigate to

[67] Martin Gjoreski, Hristijan Gjoreski, Mitja Luštrek, and Matjaž Gams. 2016.
Continuous stress detection using a wrist device: In laboratory and real life. In
Proceedings of the 2016 ACM International Joint Conference on Pervasive and
Ubiquitous Computing: AdjunctAdjunct (UbiComp’16). 1185–1193. DOI:
https://doi.org/10.1145/2968219.2968306 Navigate to

[68] Russell E. Glasgow, Cynthia Vinson, David Chambers, Muin J. Khoury, Robert M.
Kaplan, and Christine Hunter. 2012. National Institutes of Health approaches to
dissemination and implementation science: Current and future directions. American
Journal of Public Health 102, 7 (2012), 1274–1281. DOI:
https://doi.org/10.2105/AJPH.2012.300755 Navigate to
[69] Bart N. Green, Claire D. Johnson, and Alan Adams. 2006. Writing narrative
literature reviews for peer-reviewed journals: Secrets of the trade. Journal of
Chiropractic Medicine 5, 3 (2006), 101–117. DOI:
https://dx.doi.org/10.1016%2FS0899-3467(07)60142-6 Navigate to

[70] Ben Greer, Dan Robotham, Sara Simblett, Hannah Curtis, Helena Griffiths, and Til
Wykes. 2019. Digital exclusion among mental health service users: Qualitative
investigation. Journal of Medical Internet Research 21, 1 (2019), e11696. DOI:
https://doi.org/10.2196/11696 Navigate to

[71] Thomas L. Griffiths, and Mark Steyvers. 2004. Finding scientific topics. Proceedings
of the National academy of Sciences 101, 1 (2004), 5228–5235. DOI:
https://doi.org/10.1073/pnas.0307752101 Navigate to

[72] Tayfun Gürel, Luc De Raedt, and Stefan Rotter. 2007. Ranking neurons for mining
structure-activity relations in biological neural networks: NeuronRank.
Neurocomputing 70, 10–12 (2007), 1897–1901. DOI:
http://dx.doi.org/10.1016/j.neucom.2006.10.064 Navigate to

[73] Anhong Guo, Ece Kamar, Jennifer Wortman Vaughan, Hanna Wallach, and Meredith
Ringel Morris. 2020. Toward fairness in AI for people with disabilities: A research
roadmap. SIGACCESS Access. Comput. 125, Article 2 (October 2019), 1 page. DOI:
https://doi.org/10.1145/3386296.3386298 Navigate to

[74] Chuan Guo, Mayank Rana, Moustapha Cisse, and Laurens Van Der Maaten. 2017.
Countering adversarial images using input transformations. arXiv:1711.00117. DOI:
https://arxiv.org/abs/1711.00117 Navigate to

[75] Kiwan Han, Jeonghun Ku, Kwanguk Kim, Hee Jeong Jang, Junyoung Park, Jae-Jin
Kim, Chan Hyung Kim, Min-Hyung Choi, In Young Kim, and Sun I. Kim. 2009. Virtual
reality prototype for measurement of expression characteristics in emotional
situations. Computers in Biology and Medicine 39, 2 (2009), 173–179.
http://dx.doi.org/10.1016/j.compbiomed.2008.12.002 Navigate to

[76] Hoda Heidari, Claudio Ferrari, Krishna P. Gummadi, and Andreas Krause. 2018.
Fairness behind a veil of ignorance: A welfare analysis for automated decision making.
arXiv:1806.04959. DOI: https://arxiv.org/abs/1806.04959 Navigate to

[77] Tad Hirsch, Kritzia Merced, Shrikanth Narayanan, Zac E. Imel, and David C. Atkins.
2017. Designing contestability: Interaction design, machine learning, and mental
health. In Proceedings of the 2017 Conference on Designing Interactive Systems
(DIS’17). ACM, 95–99. DOI: https://doi.org/10.1145/3064663.3064703
Navigate to

[78] Tad Hirsch, Christina Soma, Kritzia Merced, Patty Kuo, Aaron Dembe, Derek D.
 Caperton, David C. Atkins, and Zac E. Imel. 2018. “It's hard to argue with a computer”:
Investigating Psychotherapists’ Attitudes towards Automated Evaluation. In
Proceedings of the 2018 Designing Interactive Systems Conference (DIS’18). ACM,
559–571. DOI: https://doi.org/10.1145/3196709.3196776 Navigate to

[79] Kristina Höök. 2000. Steps to take before intelligent user interfaces become real.
Interacting with Computers 12, 4 (2000), 409–426.
http://dx.doi.org/10.1016/S0953-5438(99)00006-5 Navigate to

[80] Ayanna Howard, Cha Zhang, and Eric Horvitz. 2017. Addressing bias in machine
learning algorithms: A pilot study on emotion recognition for intelligent systems. In
Proceedings of the IEEE Workshop on Advanced Robotics and its Social Impacts
(ARSO’17). IEEE, 1–7. DOI: https://doi.org/10.1109/ARSO.2017.8025197
Navigate to

[81] Vignesh Jagadeesh, S. Karthikeyan, and B. S. Manjunath. 2010. Spatio-temporal

optical flow statistics (STOFS) for activity classification. In Proceedings of the 7th
Indian Conference on Computer Vision, Graphics and Image Processing (ICVGIP’10).
ACM, 178–182. DOI: http://dx.doi.org/10.1145/1924559.1924583
Navigate to

[82] Vidhi Jain and Prakhar Agarwal. 2017. Symptomatic diagnosis and prognosis of
psychiatric disorders through personal gadgets. In Proceedings of the 2017 CHI
Conference Extended Abstracts on Human Factors in Computing Systems (CHI
EA’17). ACM, 118–123. DOI: https://doi.org/10.1145/3027063.3048417
Navigate to

[83] Natasha Jaques, Sara Taylor, Ehimwenma Nosakhare, Akane Sano, and Rosalind
Picard. 2016. Multi-task learning for predicting health, stress, and happiness. In
Proceedings of the NIPS Workshop on Machine Learning for Healthcare. Last
Retrieved 14th September 2019 from
https://pdfs.semanticscholar.org/b228/7a406985980515d5cc63e9b37fb17
c5186f8.pdf. Navigate to

[84] Sooyeon Jeong and Cynthia Lynn Breazeal. 2016. Improving smartphone users’
affect and wellbeing with personalized positive psychology interventions. In
Proceedings of the 4th International Conference on Human Agent Interaction
(HAI’16). ACM, New York, NY, 131–137. DOI:
https://doi.org/10.1145/2974804.2974831 Navigate to

[85] Michael I. Jordan, and Tom M. Mitchell. 2015. Machine learning: Trends,
perspectives, and prospects. Science 349, 6245 (2015), 255–260. DOI:
https://doi.org/10.1126/science.aaa8415 Navigate to

[86] Deepali J. Joshi, Mohit Makhija, Yash Nabar, Ninad Nehete, and Manasi S.
 Patwardhan. 2018. Mental health analysis using deep learning for feature extraction.
In Proceedings of the ACM India Joint International Conference on Data Science and
Management of Data (CoDS-COMAD’18). 356–359. DOI:
https://doi.org/10.1145/3152494.3167990 Navigate to

[87] K. A. Kasmiran, A. Y. Zomaya, A. A. Mazari, and R. J. Garsia. 2010. SVM-enabled

prognostic method for clinical decision making: The use of CD4 T-cell level and HIV-1
viral load for guiding treatment initiation and alteration. In Proceedings of the 2010
IEEE 23rd International Symposium on Computer-Based Medical Systems
(CBMS’10). IEEE Computer Society, 19–25. DOI:
http://dx.doi.org/10.1109/CBMS.2010.6042686 Navigate to

[88] Pavleen Kaur, Ravinder Kumar, and Munish Kumar. 2019. A healthcare monitoring
system using random forest and internet of things (IoT). Multimedia Tools
Applications 78, 14 (July 2019), 19905–19916. DOI:
https://doi.org/10.1007/s11042-019-7327-8 Navigate to

[89] Ramakanth Kavuluru, María Ramos-Morales, Tara Holaday, Amanda G. Williams,

Laura Haye, and Julie Cerel. 2016. Classification of Helpful Comments on Online
Suicide Watch Forums. In Proceedings of the 7th ACM International Conference on
Bioinformatics, Computational Biology, and Health Informatics (BCB’16). ACM, 32–
40. DOI: https://doi.org/10.1145/2975167.2975170 Navigate to

[90] Ronald C. Kessler et al. 2007. Lifetime prevalence and age-of-onset distributions of
mental disorders in the World Health Organization's World Mental Health Survey
Initiative. World Psychiatry 6, 3 (2007), 168. DOI: s Navigate to

[91] Ronald C. Kessler et al. 2008. Individual and societal effects of mental disorders on
earnings in the United States: Results from the national comorbidity survey
replication. American Journal of Psychiatry 165, 6 (2008), 703–711. DOI:
https://doi.org/10.1176/appi.ajp.2008.08010126 Navigate to

[92] Alex V. Kotlar and Thomas S. Wingo. 2018. Tutorial: Rapidly Identifying Disease-
associated Rare Variants using Annotation and Machine Learning at Whole-genome
Scale Online. In Proceedings of the 2018 ACM International Conference on
Bioinformatics, Computational Biology, and Health Informatics (BCB’18). ACM, 558–
558. DOI: https://doi.org/10.1145/3233547.3233666 Navigate to

[93] John R. Koza, Forrest H. Bennett, David Andre, and Martin A. Keane. 1996.
Automated design of both the topology and sizing of analog electrical circuits using
genetic programming. In Proceedings of the Artificial Intelligence in Design’96.
Springer, 151–170. DOI: https://doi.org/10.1007/978-94-009-0279-4_9
Navigate to

[94] Jens Kremkow, Arvind Kumar, Stefan Rotter, and Ad Aertsen. 2007. Emergence of
 population synchrony in a layered network of the cat visual cortex. Neurocomputing
70, 10–12 (2007), 2069–2073. DOI:
http://dx.doi.org/10.1016/j.neucom.2006.10.130 Navigate to

[95] Nikolaus Kriegeskorte, Jerzy Bodurka, and Peter Bandettini. 2008. Artifactual time-
course correlations in echo-planar fMRI with implications for studies of brain function.
International Journal of Imaging Systems and Technology 18, 5–6 (2008), 345–349.
DOI: http://dx.doi.org/10.1002/ima.v18:5/6 Navigate to

[96] Birgit Kriener, Tom Tetzlaff, Ad Aertsen, Markus Diesmann, and Stefan Rotter.
2008. Correlations and population dynamics in cortical networks. Neural Computing
20, 9 (2008), 2185–2226. DOI: http://dx.doi.org/10.1162/neco.2008.02-07-
474 Navigate to

[97] Kurt Kroenke, and Robert L. Spitzer. 2002. The PHQ-9: A new depression diagnostic
and severity measure. Psychiatric Annals 32, 9 (2002), 509–515. DOI:
https://doi.org/10.3928/0048-5713-20020901-06 Navigate to

[98] Arvind Kumar, Sven Schrader, Ad Aertsen, and Stefan Rotter. 2008. The high-
conductance state of cortical networks. Neural Computing 20, 1 (2008), 1–43. DOI:
http://dx.doi.org/10.1162/neco.2008.20.1.1 Navigate to

[99] Reeva Lederman, John Gleeson, Greg Wadley, Simon D'alfonso, Simon Rice, Olga
Santesteban-Echarri, and Mario Alvarez-Jimenez. 2019. Support for carers of young
people with mental illness: Design and trial of a technology-mediated therapy. ACM
Transactions on Computer-Human Interaction 26, 1, Article 4 (2019) , 33 pages. DOI:
https://doi.org/10.1145/3301421 Navigate to

[100] Yena Lee, Renee-Marie Ragguett, Rodrigo B. Mansur, Justin J. Boutilier, Joshua D.
Rosenblat, Alisson Trevizol, Elisa Brietzke, et al. 2018. Applications of machine
learning algorithms to predict therapeutic outcomes in depression: A meta-analysis
and systematic review. Journal of Affective Disorders 241, 519–532. DOI:
https://doi.org/10.1016/j.jad.2018.08.073 Navigate to

[101] Yugyung Lee, Saranya Krishnamoorthy, and Deendayal Dinakarpandian. 2013. A

semantic framework for intelligent matchmaking for clinical trial eligibility criteria.
ACM Transactions on Intelligent Systems and Technology 4, 4, Article 71 (2013) , 32
pages. DOI: https://doi.org/10.1145/2508037.2508052 Navigate to

[102] Min Kyung Lee, Anuraag Jain, Hea Jin Cha, Shashank Ojha, and Daniel Kusbit.
2019. Procedural justice in algorithmic fairness: Leveraging transparency and outcome
control for fair algorithmic mediation. Proceedings of the ACM on Human-Computer
Interaction 3, CSCW, Article 182 (2019) , 26 pages. DOI:
https://doi.org/10.1145/3359284 Navigate to

[103] Sidney R. Lehky. 2004. Bayesian estimation of stimulus responses in Poisson spike
 trains. Neural Computing 16, 7 (2004), 1325–1343. DOI:
http://dx.doi.org/10.1162/089976604323057407 Navigate to

[104] Alessandro Liberati, Douglas G. Altman, Jennifer Tetzlaff, Cynthia Mulrow, Peter C.
Gøtzsche, John P. A. Ioannidis, Mike Clarke, Pl. J. Devereaux, Jos Kleijnen, and David
Moher. 2009. The PRISMA statement for reporting systematic reviews and meta-
analyses of studies that evaluate health care interventions: explanation and
elaboration. PloS Medicine 6, 7 (2009), e1000100. DOI:
https://doi.org/10.1371/journal.pmed.1000100 Navigate to

[105] Huijie Lin, Jia Jia, Quan Guo, Yuanyuan Xue, Qi Li, Jie Huang, Lianhong Cai, and
Ling Feng. 2014. User-level psychological stress detection from social media using
deep neural network. In Proceedings of the 22nd ACM international conference on
Multimedia (MM’14). ACM, 507–516. DOI:
https://doi.org/10.1145/2647868.2654945 Navigate to

[106] Zachary C. Lipton. 2018. The mythos of model interpretability. Queue 16, 3 (May-
June 2018), 31–57. DOI: https://doi.org/10.1145/3236386.3241340
Navigate to

[107] Fannie Liu. 2019. Expressive biosignals: Authentic social cues for social connection.
In Proceedings of the Extended Abstracts of the 2019 CHI Conference on Human
Factors in Computing Systems (CHI EA’19). ACM, Paper DC12, 5 pages. DOI:
https://doi.org/10.1145/3290607.3299081 Navigate to

[108] Zengjian Liu, Buzhou Tang, Xiaolong Wang, and Qingcai Chen. 2017. De-
identification of clinical notes via recurrent neural network and conditional random
field. Journal of Biomedical Informatics 75, S (2017), S34–S42. DOI:
https://doi.org/10.1016/j.jbi.2017.05.023 Navigate to

[109] Joseph B. Lyons, Garrett G. Sadler, Kolina Koltai, Henri Battiste, Nhut T. Ho,
Lauren C. Hoffmann, David Smith, Walter Johnson, and Robert Shively. 2017. Shaping
trust through transparent design: Theoretical and experimental guidelines. In
Proceedings of the Advances in Human Factors in Robots and Unmanned Systems.
Springer, 127–136. DOI: https://doi.org/10.1007/978-3-319-41959-6_11
Navigate to

[110] Anmol Madan, Manuel Cebrian, David Lazer, and Alex Pentland. 2010. Social
sensing for epidemiological behavior change. In Proceedings of the 12th ACM
international conference on Ubiquitous computing (UbiComp’10). ACM, 291–300.
DOI: https://doi.org/10.1145/1864349.1864394 Navigate to

[111] Maria Madsen, and Shirley Gregor. 2000. Measuring human-computer trust. In
Proceedings of the 11th Australasian Conference on Information Systems. Vol. 53, 6–
8. Navigate to
[112] Adria Mallol-Ragolta, Svati Dhamija, and Terrance E. Boult. 2018. A multimodal
approach for predicting changes in PTSD symptom severity. In Proceedings of the 20th
ACM International Conference on Multimodal Interaction (ICMI’18). ACM, 324–333.
DOI: https://doi.org/10.1145/3242969.3242981 Navigate to

[113] Mirza Mansoor Baig, Hamid Gholamhosseini, Aasia A. Moqeem, Farhaan Mirza,
and Maria Lindén. 2017. A systematic review of wearable patient monitoring systems—
Current challenges and opportunities for clinical adoption. Journal of Medical
Systems 41, 7 (2017), 1–9. DOI: https://doi.org/10.1007/s10916-017-0760-1
Navigate to

[114] Martin Maritsch, Caterina Bérubé, Mathias Kraus, Vera Lehmann, Thomas Züger,
Stefan Feuerriegel, Tobias Kowatsch, and Felix Wortmann. 2019. Improving heart rate
variability measurements from consumer smartwatches with machine learning. In
Adjunct Proceedings of the 2019 ACM International Joint Conference on Pervasive
and Ubiquitous Computing and Proceedings of the 2019 ACM International
Symposium on Wearable Computers (UbiComp/ISWC’19 Adjunct). ACM, 934–938.
DOI: https://doi.org/10.1145/3341162.3346276 Navigate to

[115] Maja J. Matarić. 2019. Human-machine and human-robot interaction for long-term
user engagement and behavior change. In Proceedings of the 25th Annual
International Conference on Mobile Computing and Networking (MobiCom’19).
ACM, Paper 56, 2 pages. DOI: https://doi.org/10.1145/3300061.3300141
Navigate to

[116] Andrew M. McIntosh, Robert Stewart, Ann John, Daniel J. Smith, Katrina Davis,
Cathie Sudlow, Aiden Corvin, et al. 2016. Data science for mental health: A UK
perspective on a global challenge. The Lancet Psychiatry 3, 10 (2016), 993–998. DOI:
https://doi.org/10.1016/S2215-0366(16)30089-X Navigate to

[117] Quinten McNamara, Alejandro De La Vega, and Tal Yarkoni. 2017. Developing a
comprehensive framework for multimodal feature extraction. In Proceedings of the
23rd ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining (KDD’17). ACM, 1567–1574. DOI:
https://doi.org/10.1145/3097983.3098075 Navigate to

[118] Abhinav Mehrotra and Mirco Musolesi. 2017. Designing effective movement digital
biomarkers for unobtrusive emotional state mobile monitoring. In Proceedings of the
1st Workshop on Digital Biomarkers (DigitalBiomarkers’17). ACM, 3–8. DOI:
https://doi.org/10.1145/3089341.3089342 Navigate to

[119] Susan Michie, Lucy Yardley, Robert West, Kevin Patrick, and Felix Greaves. 2017.
Developing and evaluating digital interventions to promote behavior change in health
and health care: Recommendations resulting from an international workshop. Journal
of Medical Internet Research 19, 6 (2017), e232. DOI:
 https://doi.org/10.2196/jmir.7126 Navigate to

[120] Gatis Mikelsons, Abhinav Mehrotra, Mirco Musolesi, and Nigel Shadbolt. 2019.
Evaluating Machine Learning Algorithms for Prediction of the Adverse Valence Index
Based on the Photographic Affect Meter. In Proceedings of the 5th ACM Workshop on
Mobile Systems for Computational Social Science (MCSS’19). ACM, New York, NY, 5–
10. DOI: https://doi.org/10.1145/3325426.3329948 Navigate to

[121] William R. Miller, Theresa B. Moyers, Denise Ernst, and Paul Amrhein. 2003.
Manual for the Motivational Interviewing Skill Code (MISC), Version 2.1. Center on
Alcoholism, Substance Abuse and Addictions, University of New Mexico. Last retrieved
14th of September from https://casaa.unm.edu/download/misc.pdf.
Navigate to

[122] Vikramjit Mitra, Elizabeth Shriberg, Mitchell McLaren, Andreas Kathol, Colleen
Richey, Dimitra Vergyri, and Martin Graciarena. 2014. The SRI AVEC-2014 evaluation
system. In Proceedings of the 4th International Workshop on Audio/Visual Emotion
Challenge (AVEC’14). ACM, 93–101. DOI:
https://doi.org/10.1145/2661806.2661818 Navigate to

[123] David Moher, Alessandro Liberati, Jennifer Tetzlaff, and Douglas G. Altman. 2009.
Preferred reporting items for systematic reviews and meta-analyses: The PRISMA
statement. Annals of Internal Medicine 151, 4 (2009), 264–269. DOI:
https://doi.org/10.7326/0003-4819-151-4-200908180-00135
Navigate to

[124] David C. Mohr, Mi Zhang, and Stephen M. Schueller. 2017. Personal sensing:
understanding mental health using ubiquitous sensors and machine learning. Annual
Review of Clinical Psychology 13, 23–47. DOI: https://doi.org/10.1146/annurev-
clinpsy-032816-044949 Navigate to

[125] David C. Mohr, Kathryn Noth Tomasino, Emily G. Lattie, Hannah L. Palac, Mary J.
Kwasny, Kenneth Weingardt, Chris J. Karr, et al. 2017. IntelliCare: An eclectic, skills-
based app suite for the treatment of depression and anxiety. Journal of Medical
Internet Research 19, 1 (2017), e10. DOI: https://doi.org/10.2196/jmir.6645
Navigate to

[126] Stuart A. Montgomery, and M. A. R. I. E. Åsberg. 1979. A new depression scale

designed to be sensitive to change. The British Journal of Psychiatry 134, 4 (1979),
382–389. DOI: https://doi.org/10.1192/bjp.134.4.382 Navigate to

[127] Cecily Morrison, and Gavin Doherty. 2014. Analyzing engagement in a web-based
intervention platform through visualizing log-data. Journal of Medical Internet
Research 16, 11 (2014), e252. DOI: https://doi.org/10.2196/jmir.3575
Navigate to
[128] Mehrab Bin Morshed, Koustuv Saha, Richard Li, Sidney K. D'Mello, Munmun De
Choudhury, Gregory D. Abowd, and Thomas Plötz. 2019. Prediction of mood instability
with passive sensing. Proceedings of the ACM on Interactive, Mobile, Wearable and
Ubiquitous Technologies 3, 3, Article 75 (2019), 21 pages. DOI:
https://doi.org/10.1145/3351233 Navigate to

[129] T. B. Moyers, T. Martin, J. K. Manuel, W. R. Miller, and D. Ernst. 2010. Revised

global scales: Motivational interviewing treatment integrity 3.1. 1 (MITI 3.1. 1).
Unpublished manuscript, University of New Mexico, Albuquerque, NM.
Navigate to

[130] Inbal Nahum-Shani, Shawna N. Smith, Bonnie J. Spring, Linda M. Collins, Katie
Witkiewitz, Ambuj Tewari, and Susan A. Murphy. 2017. Just-in-time adaptive
interventions (JITAIs) in mobile health: Key components and design principles for
ongoing health behavior support. Annals of Behavioral Medicine 52, 6 (2017), 446–
462. DOI: https://doi.org/10.1007/s12160-016-9830-8 Navigate to

[131] Hiroyuki Nakahara, Shun-Ichi Amari, and Barry J. Richmond. 2006. A Comparison
of Descriptive Models of a Single Spike Train by Information-Geometric Measure.
Neural Computing 18, 3 (2006), 545–568. DOI:
http://dx.doi.org/10.1162/089976606775623289 Navigate to

[132] Lisa P. Nathan, Anja Thieme, Deborah Tatar, and Stacy Branham. 2017.
Disruptions, dilemmas and paradoxes: Ethical matter(s) in design research.
Interacting with Computers 29, 1 (2017), 1–9. DOI:
https://doi.org/10.1093/iwc/iww034 Navigate to

[133] Thin Nguyen, Bridianne O'Dea, Mark Larsen, Dinh Phung, Svetha Venkatesh, and
Helen Christensen. 2017. Using linguistic and topic analysis to classify sub-groups of
online depression communities. Multimedia Tools and Applications 76, 8 (2017),
10653–10676. DOI: https://doi.org/10.1007/s11042-015-3128-x
Navigate to

[134] Alicia L. Nobles, Jeffrey J. Glenn, Kamran Kowsari, Bethany A. Teachman, and
Laura E. Barnes. 2018. Identification of imminent suicide risk among young adults
using text messages. In Proceedings of the 2018 CHI Conference on Human Factors in
Computing Systems (CHI’18). ACM, Paper 413, 11 pages. DOI:
https://doi.org/10.1145/3173574.3173987 Navigate to

[135] Ehimwenma Nosakhare and Rosalind Picard. 2019. Probabilistic latent variable
modeling for assessing behavioral influences on well-being. In Proceedings of the 25th
ACM SIGKDD International Conference on Knowledge Discovery & Data Mining
(KDD’19). ACM, 2718–2726. DOI: https://doi.org/10.1145/3292500.3330738
Navigate to
[136] Blessing Ojeme, and Audrey Mbogho. 2016. Selecting learning algorithms for
simultaneous identification of depression and comorbid disorders. Procedia Computer
Science 96, 1294–1303. DOI: https://doi.org/10.1016/j.procs.2016.08.174
Navigate to

[137] Kathleen O'Leary, Stephen M. Schueller, Jacob O. Wobbrock, and Wanda Pratt.
2018. “Suddenly, we got to become therapists for each other”: Designing peer support
chats for mental health. In Proceedings of the 2018 CHI Conference on Human
Factors in Computing Systems (CHI’18). ACM, Paper 331, 14 pages. DOI:
https://doi.org/10.1145/3173574.3173905 Navigate to

[138] Sharon Oviatt, Björn Schuller, Philip R. Cohen, Daniel Sonntag, Gerasimos
Potamianos, and Antonio Krüger (Eds.). 2018. The Handbook of Multimodal-
Multisensor Interfaces. ACM and Morgan & Claypool, xvii–xix. DOI:
https://doi.org/10.1145/3107990.3107991 Navigate to

[139] Theodor Chris Panagiotakopoulos, Dimitrios Panagiotis Lyras, Miltos Livaditis,

Kyriakos N. Sgarbas, George C. Anastassopoulos, and Dimitrios K. Lymberopoulos.
2010. A contextual data mining approach toward assisting the treatment of anxiety
disorders. IEEE Transactions on Information Technology in Biomedicine 14, 3 (2010),
567–581. DOI: https://doi.org/10.1109/TITB.2009.2038905 Navigate to

[140] Pablo Paredes, Ran Gilad-Bachrach, Mary Czerwinski, Asta Roseway, Kael Rowan,
and Javier Hernandez. 2014. PopTherapy: Coping with stress through pop-culture. In
Proceedings of the 8th International Conference on Pervasive Computing
Technologies for Healthcare (PervasiveHealth’14). ICST (Institute for Computer
Sciences, Social-Informatics and Telecommunications Engineering), 109–117. DOI:
http://dx.doi.org/10.4108/icst.pervasivehealth.2014.255070
Navigate to

[141] Albert Park, Mike Conway, and Annie T. Chen. 2018. Examining thematic similarity,
difference, and membership in three online mental health communities from Reddit: A
text mining and visualization approach. Computers in Human Behavior 78, 98–112.
DOI: https://doi.org/10.1016/j.chb.2017.09.001 Navigate to

[142] Ives Cavalcante Passos, Benson Mwangi, and Flavio Kapczinski. 2019. Personalized
Psychiatry: Big Data Analytics in Mental Health ( 1st ed .). Springer Publishing
Company, Incorporated. Navigate to

[143] Mensah Kwabena Patrick. 2015. Textual prediction of attitudes towards mental
health. International Journal of Knowledge Engineering and Data Mining 3, 3/4
(2015), 274–285. DOI: http://dx.doi.org/10.1504/IJKEDM.2015.074076
Navigate to

[144] David A. Patterson and Richard N. Cloud. 2000. The application of artificial neural
 networks for outcome prediction in a cohort of severely mentally ill outpatients.
Journal of Technology and Human Services 16, 2/3 (2000), 47–61. DOI:
http://dx.doi.org/10.1300/J017v16n02_05 Navigate to

[145] John P. Pestian, Pawel Matykiewicz, and Jacqueline Grupp-Phelan. 2008. Using
natural language processing to classify suicide notes. In Proceedings of the Workshop
on Current Trends in Biomedical Natural Language Processing (BioNLP’08). ACM,
Stroudsburg, PA, 96–97. Navigate to

[146] Lawrence Pfeffer, David Ide, Craig Stewart, and Dietmar Plenz. 2004. A life support
system for stimulation of and recording from rodent neuron networks grown on multi-
electrode arrays. In Proceedings of the 17th IEEE Symposium on Computer-Based
Medical Systems (CBMS’04). IEEE Computer Society, 473–478. DOI:
http://dx.doi.org/10.1109/CBMS.2004.5 Navigate to

[147] André Pimenta, Sergio Gonçalves, Davide Carneiro, Florentino Fde-riverola, José
Neves, and Paulo Novais. 2015. Mental workload management as a tool in e-learning
scenarios. In Proceedings of the 5th International Conference on Pervasive and
Embedded Computing and Communication Systems (PECCS’15). César Benavente-
Peces, Olivier Paillet, and Andreas Ahrens (Eds.), SCITEPRESS—Science and
Technology Publications, 25–32. DOI:
https://doi.org/10.5220/0005237700250032 Navigate to

[148] John P. Pollak, Phil Adams, and Geri Gay. 2011. PAM: A photographic affect meter
for frequent, in situ measurement of affect. In Proceedings of the SIGCHI Conference
on Human Factors in Computing Systems (CHI’11). ACM, 725–734. DOI:
https://doi.org/10.1145/1978942.1979047 Navigate to

[149] Forough Poursabzi-Sangdeh, Daniel G. Goldstein, Jake M. Hofman, Jennifer

Wortman Vaughan, and Hanna Wallach. 2018. Manipulating and measuring model
interpretability. arXiv:1802.07810. DOI: https://arxiv.org/abs/1802.07810
Navigate to

[150] Yada Pruksachatkun, Sachin R. Pendse, and Amit Sharma. 2019. Moments of
change: Analyzing peer-based cognitive support in online mental health forums. In
Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems
(CHI’19). ACM, Paper 64, 1–13. DOI: https://doi.org/10.1145/3290605.3300294
Navigate to

[151] Alessandro Puiatti, Steven Mudda, Silvia Giordano, and Oscar Mayora. 2011.
Smartphone-centred wearable sensors network for monitoring patients with bipolar
disorder. In Proceedings of the 2011 Annual International Conference of the IEEE
Engineering in Medicine and Biology Society. IEEE, 3644–3647. DOI:
https://doi.org/10.1109/IEMBS.2011.6090613 Navigate to
[152] Thomas Quisel, Wei-Nchih Lee, and Luca Foschini. 2017. Observation time vs.
performance in digital phenotyping. In Proceedings of the 1st Workshop on Digital
Biomarkers (DigitalBiomarkers’17). ACM, 33–36. DOI:
https://doi.org/10.1145/3089341.3089347 Navigate to

[153] Mashfiqui Rabbi, Shahid Ali, Tanzeem Choudhury, and Ethan Berke. 2011. Passive
and in-situ assessment of mental and physical well-being using mobile sensors. In
Proceedings of the 13th international conference on Ubiquitous computing
(UbiComp’11). ACM, 385–394. DOI: https://doi.org/10.1145/2030112.2030164
Navigate to

[154] Neelesh Rastogi, Fazel Keshtkar, and Md Suruz Miah. 2018. A multi-modal human
robot interaction framework based on cognitive behavioral therapy model. In
Proceedings of the Workshop on Human-Habitat for Health (H3): Human-Habitat
Multimodal Interaction for Promoting Health and Well-Being in the Internet of
Things Era (H3’18). ACM, Paper 2, 6 pages. DOI:
https://doi.org/10.1145/3279963.3279968 Navigate to

[155] Anupama Ray, Siddharth Kumar, Rutvik Reddy, Prerana Mukherjee, and Ritu Garg.
2019. Multi-level attention network using text, audio and video for depression
prediction. In Proceedings of the 9th International on Audio/Visual Emotion
Challenge and Workshop (AVEC’19). ACM, 81–88. DOI:
https://doi.org/10.1145/3347320.3357697 Navigate to

[156] Stefan Rennick-Egglestone, Sarah Knowles, Gill Toms, Penny Bee, Karina Lovell,
and Peter Bower. 2016. Health technologies sin the wild’: Experiences of engagement
with computerised CBT. In Proceedings of the 2016 CHI Conference on Human
Factors in Computing Systems (CHI’16). ACM, 2124–2135. DOI:
https://doi.org/10.1145/2858036.2858128 Navigate to

[157] Derek Richards, Ladislav Timulak, Emma O'Brien, Claire Hayes, Noemi Vigano,
John Sharry, and Gavin Doherty. 2015. A randomized controlled trial of an internet-
delivered treatment: Its potential as a low-intensity community intervention for adults
with symptoms of depression. Behaviour Research and Therapy 75, 20–31. DOI:
https://doi.org/10.1016/j.brat.2015.10.005 Navigate to

[158] Darius A. Rohani, Maria Faurholt-Jepsen, Lars Vedel Kessing, and Jakob E.
Bardram. 2018. Correlations between objective behavioral features collected from
mobile and wearable devices and depressive mood symptoms in patients with affective
disorders: Systematic review. JMIR mHealth and uHealth 6, 8 (2018), e165. DOI:
https://doi.org/10.2196/mhealth.9691 Navigate to

[159] Anne K. Rosenstiel, and Francis J. Keefe. 1983. The use of coping strategies in
chronic low back pain patients: Relationship to patient characteristics and current
adjustment. Pain 17, 1 (1983), 33–44. DOI: https://doi.org/10.1016/0304-
 3959(83)90125-2 Navigate to

[160] Sushmita Roy, Terran Lane, and Margaret Werner-Washburne. 2009. Learning
structurally consistent undirected probabilistic graphical models. In Proceedings of the
26th Annual International Conference on Machine Learning (ICML’09). ACM, New
York, NY, 905–912. DOI: http://dx.doi.org/10.1145/1553374.1553490
Navigate to

[161] Cynthia Rudin. 2019. Stop explaining black box machine learning models for high
stakes decisions and use interpretable models instead. Nature Machine Intelligence 1
(2019), 206–215. DOI: https://doi.org/10.1038/s42256-019-0048-x
Navigate to

[162] James A. Russell. 1980. A circumplex model of affect. Journal of Personality and
Social Psychology 39, 6 (1980), 1161. http://dx.doi.org/10.1037/h0077714
Navigate to

[163] Daniel W. Russell. 1996. UCLA loneliness scale (version 3): Reliability, validity, and
factor structure. Journal of Personality Assessment 66, 1 (1996), 20–40. DOI:
https://doi.org/10.1207/s15327752jpa6601_2 Navigate to

[164] Sohrab Saeb, Luca Lonini, Arun Jayaraman, David C. Mohr, and Konrad P.
Kording. 2016. Voodoo machine learning for clinical predictions. Biorxiv, 059774.
DOI: https://doi.org/10.1101/059774 Navigate to

[165] Koustuv Saha and Munmun De Choudhury. 2017. Modeling stress with social media
around incidents of gun violence on college campuses. Proceedings of the ACM on
Human-Computer Interaction 1, CSCW, Article 92 (2017) , 27 pages. DOI:
https://doi.org/10.1145/3134727 Navigate to

[166] Koustuv Saha, Benjamin Sugar, John Torous, Bruno Abrahao, Emre Kıcıman, and
Munmun De Choudhury. 2019. A social media study on the effects of psychiatric
medication use. In Proceedings of the International AAAI Conference on Web and
Social Media 13, 1 (2019), 440–451. DOI:
https://wvvw.aaai.org/ojs/index.php/ICWSM/article/view/3242
Navigate to

[167] Koustuv Saha, Sang Chan Kim, Manikanta D. Reddy, Albert J. Carter, Eva Sharma,
Oliver L. Haimson, and Munmun De Choudhury. 2019. The language of LGBTQ±
minority stress experiences on social media. Proceedings of the ACM on Human-
Computer Interaction 3, CSCW, Article 89 (November 2019) , 22 pages. DOI:
https://doi.org/10.1145/3361108 Navigate to

[168] Asif Salekin, Jeremy W. Eberle, Jeffrey J. Glenn, Bethany A. Teachman, and John
A. Stankovic. 2018. A weakly supervised learning framework for detecting social
anxiety and depression. Proceedings of the ACM on Interactive Mobile Wearable
 Ubiquitous Technology 2, 2, Article 81 (2018) , 26 pages. DOI:
https://doi.org/10.1145/3214284 Navigate to

[169] SAMHS (Substance Abuse and Mental Health Services Administration). 2018. Key
substance use and mental health indicators in the United States: Results from the 2017
National Survey on Drug Use and Health (HHS Publication No. SMA 18-5068,
NSUDH Series H-53). Center for Behavioral Health Statistics and Quality, Substance
Abuse and Mental Health Services Administration, Rockville, MD. Retrieved from
https://www.samhsa.gov/data/sites/default/files/cbhsq-
reports/NSDUHFFR2017/NSDUHFFR2017.pdf. Navigate to

[170] Pedro Sanches, Axel Janson, Pavel Karpashevich, Camille Nadal, Chengcheng Qu,
Claudia Daudén Roquet, Muhammad Umair, Charles Windlin, Gavin Doherty, Kristina
Höök, and Corina Sas. 2019. HCI and affective health: Taking stock of a decade of
studies and charting future research directions. In Proceedings of the 2019 CHI
Conference on Human Factors in Computing Systems (CHI’19). ACM, Paper 245, 17
pages. DOI: https://doi.org/10.1145/3290605.3300475 Navigate to

[171] Jessica Schroeder, Chelsey Wilkes, Kael Rowan, Arturo Toledo, Ann Paradiso, Mary
Czerwinski, Gloria Mark, and Marsha M. Linehan. 2018. Pocket Skills: A
conversational mobile web app to support dialectical behavioral therapy. In
Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems
(CHI’18). ACM, Paper 398, 15 pages. DOI:
https://doi.org/10.1145/3173574.3173972 Navigate to

[172] Stephen M. Schueller, Adrian Aguilera, and David C. Mohr. 2017. Ecological
momentary interventions for depression and anxiety. Depression and Anxiety 34, 6
(2017), 540–545. DOI: https://doi.org/10.1002/da.22649 Navigate to

[173] Adrian B. R. Shatte, Delyse M. Hutchinson, and Samantha J. Teague. 2019. Machine
learning in mental health: A scoping review of methods and applications.
Psychological Medicine 49, 9 (2019), 1426–1448. DOI:
https://doi.org/10.1017/S0033291719000151 Navigate to

[174] Hakim Sidahmed, Elena Prokofyeva, and Matthew B. Blaschko. 2016. Discovering
predictors of mental health service utilization with k-support regularized logistic
regression. Information Sciences 329, C (2016), 937–949. DOI:
https://doi.org/10.1016/j.ins.2015.03.069 Navigate to

[175] Insu Song, Denise Dillon, Tze Jui Goh, and Min Sung. 2011. A health social network
recommender system. In Proceedings of the 14th international conference on Agents
in Principle, Agents in Practice (PRIMA’11). David Kinny, Jane Yung-jen Hsu, Guido
Governatori, and Aditya K. Ghose (Eds.), Springer, 361–372. DOI:
http://dx.doi.org/10.1007/978-3-642-25044-6_29 Navigate to
[176] Dimitris Spathis, Sandra Servia-Rodriguez, Katayoun Farrahi, Cecilia Mascolo, and
Jason Rentfrow. 2019. Passive mobile sensing and psychological traits for large scale
mood prediction. In Proceedings of the 13th EAI International Conference on
Pervasive Computing Technologies for Healthcare (PervasiveHealth’19). ACM, 272–
281. DOI: https://doi.org/10.1145/3329189.3329213 Navigate to

[177] Dimitris Spathis, Sandra Servia-Rodriguez, Katayoun Farrahi, Cecilia Mascolo, and
Jason Rentfrow. 2019. Sequence multi-task learning to forecast mental wellbeing from
sparse self-reported data. In Proceedings of the 25th ACM SIGKDD International
Conference on Knowledge Discovery & Data Mining (KDD’19). ACM, 2886–2894.
DOI: https://doi.org/10.1145/3292500.3330730 Navigate to

[178] B. Sri Nandhini and J. I. Sheeba. 2015. cyberbullying detection and classification
using information retrieval algorithm. In Proceedings of the 2015 International
Conference on Advanced Research in Computer Science Engineering and Technology
(ICARCSET’15). ACM, Paper 20, 5 pages. DOI:
http://dx.doi.org/10.1145/2743065.2743085 Navigate to

[179] M. Srividya, S. Mohanavalli, and N. Bhalaji. 2018. Behavioral modeling for mental
health using machine learning algorithms. Journal of Medical Systems 42, 5 (2018),
88. DOI: https://doi.org/10.1007/s10916-018-0934-5 Navigate to

[180] Benjamin Staude, Stefan Rotter, and Sonja Grün. 2008. Can spike coordination be
differentiated from rate covariation?. Neural Computing 20, 8 (2008), 1973–1999.
DOI: http://dx.doi.org/10.1162/neco.2008.06-07-550 Navigate to

[181] Klaas E. Stephan, Florian Schlagenhauf, Quentin J. M. Huys, Sudhir Raman,

Eduardo A. Aponte, Kay Henning Brodersen, Lionel Rigoux, et al. 2017. Computational
neuroimaging strategies for single patient predictions. Neuroimage 145, Pt. B (2017),
180–199. DOI: https://doi.org/10.1016/j.neuroimage.2016.06.038
Navigate to

[182] Shelley E. Taylor, William T. Welch, Heejung S. Kim, and David K. Sherman. 2007.
Cultural differences in the impact of social support on psychological and biological
stress responses. Psychological Science 18, 9 (2007), 831–837. DOI:
https://doi.org/10.1111/j.1467-9280.2007.01987.x Navigate to

[183] Teewoon Tan, Ling Guan, and John Burne. 1999. A real-time image analysis system
for computer-assisted diagnosis of neurological disorders. Real-Time Imaging 5, 4
(1999), 253–269. DOI: http://dx.doi.org/10.1006/rtim.1998.0139
Navigate to

[184] Leili Tavabi. 2019. Multimodal machine learning for interactive mental health
therapy. In Proceedings of the 2019 International Conference on Multimodal
Interaction (ICMI’19). Wen Gao, Helen Mei Ling Meng, Matthew Turk, Susan R.
 Fussell, Björn Schuller, Yale Song, and Kai Yu (Eds.), ACM, 453–456. DOI:
https://doi.org/10.1145/3340555.3356095 Navigate to

[185] Tom Tetzlaff, Stefan Rotter, Eran Stark, Moshe Abeles, Ad Aertsen, and Markus
Diesmann. 2008. Dependence of neuronal correlations on filter characteristics and
marginal spike train statistics. Neural Computing 20, 9 (2008), 2133–2184.
http://dx.doi.org/10.1162/neco.2008.05-07-525 Navigate to

[186] Oliver Theobald. 2017. Machine Learning for Absolute Beginners. A Plain English
Introduction. Scatterplot Press. Navigate to

[187] Anja Thieme, John McCarthy, Paula Johnson, Stephanie Phillips, Jayne Wallace,
Siân Lindley, Karim Ladha, Daniel Jackson, Diana Nowacka, Ashur Rafiev, Cassim
Ladha, Thomas Nappey, Mathew Kipling, Peter Wright, Thomas D. Meyer, and Patrick
Olivier. 2016. Challenges for designing new technology for health and wellbeing in a
complex mental healthcare context. In Proceedings of the 2016 CHI Conference on
Human Factors in Computing Systems (CHI’16). ACM, 2136–2149. DOI:
https://doi.org/10.1145/2858036.2858182 Navigate to

[188] Anja Thieme, Jayne Wallace, Thomas D. Meyer, and Patrick Olivier. 2015.
Designing for mental wellbeing: Towards a more holistic approach in the treatment
and prevention of mental illness. In Proceedings of the 2015 British HCI Conference
(British HCI’15). ACM, New York, NY, 1–10. DOI:
http://dx.doi.org/10.1145/2783446.2783586 Navigate to

[189] Anja Thieme, Danielle Belgrave, Akane Sano, and Gavin Doherty. 2020. Reflections
on mental health assessment and ethics for machine learning applications.
Interactions 27, 2 (2020), 6–7. DOI: https://doi.org/10.1145/3381342
Navigate to

[190] Graham Thornicroft, Diana Rose, Aliya Kassam, and Norman Sartorius. 2007.
Stigma: Ignorance, prejudice or discrimination? The British Journal of Psychiatry
190, 3 (2007), 192–193. DOI: https://doi.org/10.1192/bjp.bp.106.025791
Navigate to

[191] John Torous, and Camille Nebeker. 2017. Navigating ethics in the digital age:
Introducing connected and open research ethics (CORE), a tool for researchers and
institutional review boards. Journal of Medical Internet Research 19, 2 (2017), e38.
DOI: https://doi.org/10.2196/jmir.6793 Navigate to

[192] Truyen Tran, Dinh Phung, Wei Luo, Richard Harvey, Michael Berk, and Svetha
Venkatesh. 2013. An integrated framework for suicide risk prediction. In Proceedings
of the 19th ACM SIGKDD International Conference on Knowledge Discovery and
Data Mining (KDD’13). Rayid Ghani, Ted E. Senator, Paul Bradley, Rajesh Parekh, and
Jingrui He (Eds.), ACM, 1410–1418. DOI:
 https://doi.org/10.1145/2487575.2488196 Navigate to

[193] Konstantinos Tsiakas, Lynette Watts, Cyril Lutterodt, Theodoros Giannakopoulos,

Alexandros Papangelis, Robert Gatchel, Vangelis Karkaletsis, and Fillia Makedon.
2015. A multimodal adaptive dialogue manager for depressive and anxiety disorder
screening: A wizard-of-oz experiment. In Proceedings of the 8th ACM International
Conference on Pervasive Technologies Related to Assistive Environments (PETRA’15).
ACM, Paper 82, 4 pages. DOI: https://doi.org/10.1145/2769493.2769572
Navigate to

[194] Robin Turkington, Maurice Mulvenna, Raymond Bond, Siobhan O'Neill, and Cherie
Armour. 2018. The application of user event log data for mental health and wellbeing
analysis. In Proceedings of the 32nd International BCS Human Computer Interaction
Conference (HCI’18). BCS Learning and Development Ltd., Swindon, UK, Paper 4, 14
pages. DOI: https://doi.org/10.14236/ewic/HCI2018.4 Navigate to

[195] Jessica A. Turner, K. C. Anderson, and R. M. Siegel. 2003. Cell responsiveness in

macaque superior temporal polysensory area measured by temporal discriminants.
Neural Computing 15, 9 (2003), 2067–2090.
http://dx.doi.org/10.1162/089976603322297296 Navigate to

[196] Michel Valstar, Jonathan Gratch, Björn Schuller, Fabien Ringeval, Denis Lalanne,
Mercedes Torres Torres, Stefan Scherer, Giota Stratou, Roddy Cowie, and Maja Pantic.
2016. AVEC 2016: Depression, mood, and emotion recognition workshop and
challenge. In Proceedings of the 6th International Workshop on Audio/Visual
Emotion Challenge (AVEC’16). ACM, 3–10. DOI:
https://doi.org/10.1145/2988257.2988258 Navigate to

[197] Jose Juan Dominguez Veiga and Tomas E. Ward. 2016. Data collection
requirements for mobile connected health: An end user development approach. In
Proceedings of the 1st International Workshop on Mobile Development (Mobile!
2016). ACM, 23–30. DOI: https://doi.org/10.1145/3001854.3001856
Navigate to

[198] Mélodie Vidal, Andreas Bulling, and Hans Gellersen. 2011. Analysing EOG signal
features for the discrimination of eye movements with wearable devices. In
Proceedings of the 1st International Workshop on Pervasive Eye Tracking and Mobile
Eye-Based Interaction (PETMEI’11). ACM, 15–20. DOI:
http://dx.doi.org/10.1145/2029956.2029962 Navigate to

[199] Nicole Voges, Ad Aertsen, and Stefan Rotter. 2007. Statistical analysis of spatially
embedded networks: From grid to random node positions. Neurocomputing 70, 10–12
(2007), 1833–1837. DOI: http://dx.doi.org/10.1016/j.neucom.2006.10.126
Navigate to
[200] Tuong Manh Vu, Charlotte Probst, Joshua M. Epstein, Alan Brennan, Mark Strong,
and Robin C. Purshouse. 2019. Toward inverse generative social science using multi-
objective genetic programming. In Proceedings of the Genetic and Evolutionary
Computation Conference (GECCO’19). Manuel López-Ibáñez (Ed.), ACM, 1356–1363.
DOI: https://doi.org/10.1145/3321707.3321840 Navigate to

[201] Rui Wang, Fanglin Chen, Zhenyu Chen, Tianxing Li, Gabriella Harari, Stefanie
Tignor, Xia Zhou, Dror Ben-Zeev, and Andrew T. Campbell. 2014. StudentLife:
Assessing mental health, academic performance and behavioral trends of college
students using smartphones. In Proceedings of the 2014 ACM International Joint
Conference on Pervasive and Ubiquitous Computing (UbiComp’14). ACM, 3–14. DOI:
https://doi.org/10.1145/2632048.2632054 Navigate to

[202] David Watson, Lee Anna Clark, and Auke Tellegen. 1988. Development and
validation of brief measures of positive and negative affect: The PANAS scales. Journal
of Personality and Social Psychology 54, 6 (1988), 1063. Navigate to

[203] Frank W. Weathers, Brett T. Litz, Debra S. Herman, Jennifer A. Huska, and
Terence M. Keane. 1993. The PTSD checklist (PCL): Reliability, validity, and diagnostic
utility. In Proceedings of the Annual Convention of the International Society for
Traumatic Stress Studies. Vol. 462. Navigate to

[204] Harvey A. Whiteford, Louisa Degenhardt, Jürgen Rehm, Amanda J. Baxter, Alize J.
Ferrari, Holly E. Erskine, Fiona J. Charlson, et al. 2013. Global burden of disease
attributable to mental and substance use disorders: Findings from the Global Burden
of Disease Study 2010. The Lancet 382, 9904 (2013), 1575–1586. DOI:
https://doi.org/10.1016/S0140-6736(13)61611-6 Navigate to

[205] Paula Wilbourne, Geralyn Dexter, and David Shoup. 2018. Research driven: Sibly
and the transformation of mental health and wellness. In Proceedings of the 12th EAI
International Conference on Pervasive Computing Technologies for Healthcare
(PervasiveHealth’18). ACM, 389–391. DOI:
https://doi.org/10.1145/3240925.3240932 Navigate to

[206] Choong-Wan Woo, Luke J. Chang, Martin A. Lindquist, and Tor D. Wager. 2017.
Building better biomarkers: Brain models in translational neuroimaging. Nature
Neuroscience 20, 3 (2017), 365–377. DOI: https://doi.org/10.1038/nn.4478
Navigate to

[207] World Health Organization (WHO). 2018. Fact sheets: Depression. Last retrieved
11th June 2019 from http://www.who.int/mediacentre/factsheets/fs369/en/.
Navigate to

[208] Ali Yadollahi, Ameneh Gholipour Shahraki, and Osmar R. Zaiane. 2017. Current
state of text sentiment analysis from opinion to emotion mining. ACM Computing
 Surveys 50, 2, Article 25 (2017) , 33 pages. DOI: https://doi.org/10.1145/3057270
Navigate to

[209] Hui Yang and Peter A. Bath. 2019. Automatic prediction of depression in older age.
In Proceedings of the 3rd International Conference on Medical and Health
Informatics 2019 (ICMHI’19). ACM, 36–44. DOI:
https://doi.org/10.1145/3340037.3340042 Navigate to

[210] Qian Yang, Aaron Steinfeld, and John Zimmerman. 2019. Unremarkable AI: Fitting
intelligent decision support into critical, clinical decision-making processes. In
Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems
(CHI’19). ACM, Paper 238, 11 pages. DOI:
https://doi.org/10.1145/3290605.3300468 Navigate to

[211] Amir Hossein Yazdavar, Hussein S. Al-Olimat, Monireh Ebrahimi, Goonmeet Bajaj,
Tanvi Banerjee, Krishnaprasad Thirunarayan, Jyotishman Pathak, and Amit Sheth.
2017. Semi-supervised approach to monitoring clinical depressive symptoms in social
media. In Proceedings of the 2017 IEEE/ACM International Conference on Advances
in Social Networks Analysis and Mining 2017 (ASONAM’17). Jana Diesner, Elena
Ferrari, and Guandong Xu (Eds.), ACM, 1191–1198. DOI:
https://doi.org/10.1145/3110025.3123028 Navigate to

[212] Ming Yin, Jennifer Wortman Vaughan, and Hanna Wallach. 2019. Understanding
the effect of accuracy on trust in machine learning models. In Proceedings of the 2019
CHI Conference on Human Factors in Computing Systems (CHI’19). ACM, Paper 279,
12 pages. DOI: https://doi.org/10.1145/3290605.3300509 Navigate to

[213] Robert C. Young, Jeffery T. Biggs, Veronika E. Ziegler, and Dolores A. Meyer. 1978.
A rating scale for mania: Reliability, validity and sensitivity. The British Journal of
Psychiatry 133, 5 (1978), 429–435. DOI: https://doi.org/10.1192/bjp.133.5.429
Navigate to

[214] Yakun Yu, Qian Wang, Hao Hu, Shanshan Su, and Zhen Wang. 2018. Multi-atlas
based early prediction of post-traumatic stress disorder. In Proceedings of the 2nd
International Symposium on Image Computing and Digital Medicine (ISICDM’2018).
ACM, 69–72. DOI: https://doi.org/10.1145/3285996.3286012 Navigate to

[215] Liu Yue, Zhang Chunhong, Tian Chujie, Zhao Xiaomeng, Zhang Ruizhi, and Ji Yang.
2018. Application of data mining for young children education using emotion
information. In Proceedings of the 2018 International Conference on Data Science
and Information Technology (DSIT’18). ACM, 96–104. DOI:
https://doi.org/10.1145/3239283.3239321 Navigate to

[216] Sean D. Young, and Renee Garett. 2018. Ethical issues in addressing social media
posts about suicidal intentions during an online study among youth: Case study. JMIR
 Mental Health 5, 2 (2018), e33. Navigate to

[217] Camellia Zakaria, Youngki Lee, and Rajesh Balan. 2019. Passive detection of
perceived stress using location-driven sensing technologies at scale (demo). In
Proceedings of the 17th Annual International Conference on Mobile Systems,
Applications, and Services (MobiSys’19). ACM, 667–668. DOI:
https://doi.org/10.1145/3307334.3328574 Navigate to

[218] Camellia Zakaria, Rajesh Balan, and Youngki Lee. 2019. StressMon: Scalable
detection of perceived stress and depression using passive sensing of changes in work
routines and group interactions. Proceedings of the ACM on Human-Computer
Interaction 3, CSCW, Article 37 (November 2019) , 29 pages. DOI:
https://doi.org/10.1145/3359139 Navigate to

[219] Liang Zhao, Jia Jia, and Ling Feng. 2015. Teenagers’ stress detection based on time-
sensitive micro-blog comment/response actions. In Proceedings of the IFIP
International Conference on Artificial Intelligence in Theory and Practice. Springer,
26–38. DOI: https://doi.org/10.1007/978-3-319-25261-2_3 Navigate to

[220] Andrey Zhdanov, Talma Hendler, Leslie Ungerleider, and Nathan Intrator. 2007.
Inferring functional brain states using temporal evolution of regularized classifiers.
Intelligent Neuroscience 2007, Article 52069 . DOI:
https://doi.org/10.1155/2007/52609 Navigate to

[221] Charles Zheng, Rakesh Achanta, and Yuval Benjamini. 2018. Extrapolating expected
accuracies for large multi-class problems. Journal of Machine Learning Research 19, 1
(January 2018), 2609–2638. Navigate to

[222] Dawei Zhou, Jiebo Luo, Vincent Silenzio, Yun Zhou, Jile Hu, Glenn Currier, and
Henry Kautz. 2015. Tackling mental health by integrating unobtrusive multimodal
sensing. In Proceedings of the 29th AAAI Conference on Artificial Intelligence
(AAAI’15). AAAI Press, 1401–1408. DOI:
https://www.aaai.org/ocs/index.php/AAAI/AAAI15/paper/view/9546/933
4b. Navigate to

Footnotes
1Full Query Syntax used: "query": {(+"machine learning" +"mental health")} "filter":
{"publicationYear":{"gte":1990}}, {owners.owner=GUIDE}.
2Reddit (www.reddit.com) is a website that offers a collection of forums, where users
can share content or comment on other peoples’ posts. The service consists of more
 than one million communities, called ‘subreddits’, and has more than 330 million
monthly active users.
3Live Journal (www.livejournal.com) is a social networking service with
approximately 30 million monthly visitors. Users have a profile page, can maintain a
personal blog, connect and communicate with others, and form an online community
in the form of a collective blog [133].
4 https://www.sibly.co/.
5Examples of assessment scales used include, for depression: PHQ-8 and PHQ-9 [97],
Epidemiological Study Depression Scale [9], and MADRS depression rating system
[126]; for mania the Mania Rating Scale (MRS) [213]; for mood HAMD mood scores
(Hamilton rating scale for depression) [50]; for affect: PANAS for positive and
negative affect [202], Photographic Affect meter (PAM) [148]; for stress: Coping
Stress Questionnaire (CSQ) [159]; Trauma-Focused Coping Self-Efficacy (CSE-T) [16];
Perceived Stress Scale (PSS) [38]; PTSD severity checklist (PCL) [203]; for mental
wellbeing: eight-item flourishing scale [46] and SF-36 Mental Health Score
(www.optum.com/sf36); for levels of social inclusion and connectedness: 20-item
UCLA loneliness scale [163]; and for physical activity: Yale Physical Activity Survey
(YPAS) [47].
6EMAs are often short questions designed to capture in-situ real time information about a
person's experience [128]. Examples of EMA's used include the following: Experience
Sampling Method (ESM) based on two-dimensional Circumplex model of emotion
[162]; PAM picture library to assess mood [151]; EMAs built on single item stress
survey [182]; Stress Monitoring Test (SMT) [139].
7 https://snapshot.media.mit.edu/info/.
8 https://studentlife.cs.dartmouth.edu/.
9Like classification, regression is a predictive modelling task. Although classification
predicts a class label for a given observation, regression instead predicts a continuous
quantity (e.g., amount, sizes, ranges, time series).
10 https://www.silvercloudhealth.com/.
11 https://intellicare.cbits.northwestern.edu/.
12 https://www.patientslikeme.com/.
13 https://ourdatahelps.org/.
14 https://www.mqmentalhealth.org/.

The research of Gavin Doherty is funded in part by Science Foundation Ireland Grant no.
13/RC/2106 to the Adapt Centre.

Authors’ addresses: A. Thieme (corresponding author) and D. Belgrave, Healthcare

Intelligence, Microsoft Research, 21 Station Road, CB1 2FB Cambridge, Cambridgeshire, UK;
emails: anthie@microsoft.com, dabelgra@microsoft.com; G. Doherty, School of
Computer Science and Statistics, Trinity College Dublin, College Green, Dublin 2, IRL; email:
gavin.doherty@tcd.ie.

Permission to make digital or hard copies of part or all of this work for personal or classroom
use is granted without fee provided that copies are not made or distributed for profit or
commercial advantage and that copies bear this notice and the full citation on the first page.
Copyrights for third-party components of this work must be honored. For all other uses,
contact the owner/author(s).

©2020 Copyright held by the owner/author(s).

1073-0516/2020/08-ART34
DOI: https://doi.org/10.1145/3398069

Publication History: Received September 2019; revised April 2020; accepted May 2020