Botonis et al.

Journal of NeuroEngineering and Rehabilitation (2022) 19:60

https://doi.org/10.1186/s12984-022-01040-4

RESEARCH Open Access

Wearable airbag technology and machine

learned models to mitigate falls after stroke
Olivia K. Botonis1†, Yaar Harari1,2†, Kyle R. Embry1,2, Chaithanya K. Mummidisetty1, David Riopelle2,3,
Matt Giffhorn1, Mark V. Albert4, Vallery Heike5,6 and Arun Jayaraman1,2*

Abstract
Background: Falls are a common complication experienced after a stroke and can cause serious detriments to
physical health and social mobility, necessitating a dire need for intervention. Among recent advancements, wear-
able airbag technology has been designed to detect and mitigate fall impact. However, these devices have not been
designed nor validated for the stroke population and thus, may inadequately detect falls in individuals with stroke-
related motor impairments. To address this gap, we investigated whether population-specific training data and mod-
eling parameters are required to pre-detect falls in a chronic stroke population.
Methods: We collected data from a wearable airbag’s inertial measurement units (IMUs) from individuals with (n = 20
stroke) and without (n = 15 control) history of stroke while performing a series of falls (842 falls total) and non-falls
(961 non-falls total) in a laboratory setting. A leave-one-subject-out crossvalidation was used to compare the per-
formance of two identical machine learned models (adaptive boosting classifier) trained on cohort-dependent data
(control or stroke) to pre-detect falls in the stroke cohort.
Results: The average performance of the model trained on stroke data (recall = 0.905, precision = 0.900) had statisti-
cally significantly better recall (P = 0.0035) than the model trained on control data (recall = 0.800, precision = 0.944),
while precision was not statistically significantly different. Stratifying models trained on specific fall types revealed
differences in pre-detecting anterior–posterior (AP) falls (stroke-trained model’s F­ 1-score was 35% higher, P = 0.019).
Using activities of daily living as non-falls training data (compared to near-falls) significantly increased the AUC (Area
under the receiver operating characteristic) for classifying AP falls for both models (P < 0.04). Preliminary analysis sug-
gests that users with more severe stroke impairments benefit further from a stroke-trained model. The optimal lead
time (time interval pre-impact to detect falls) differed between control- and stroke-trained models.
Conclusions: These results demonstrate the importance of population sensitivity, non-falls data, and optimal lead
time for machine learned pre-impact fall detection specific to stroke. Existing fall mitigation technologies should be
challenged to include data of neurologically impaired individuals in model development to adequately detect falls in
other high fall risk populations.
Trial registration https://​clini​caltr​ials.​gov/​ct2/​show/​NCT05​076565; Unique Identifier: NCT05076565. Retrospectively
registered on 13 October 2021

†
Olivia K. Botonis and Yaar Harari should be considered co-first authors
*Correspondence: a-jayaraman@northwestern.edu
2
Department of Physical Medicine and Rehabilitation, Northwestern
University, Chicago, IL, USA
Full list of author information is available at the end of the article

© The Author(s) 2022. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which
permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the
original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or
other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line
to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory
regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this
licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by/4.​0/. The Creative Commons Public Domain Dedication waiver (http://​creat​iveco​
mmons.​org/​publi​cdoma​in/​zero/1.​0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.
Botonis et al. Journal of NeuroEngineering and Rehabilitation (2022) 19:60 Page 2 of 14

Keywords: Pre-impact fall detection, Fall mitigation, Machine-learning, Wearable sensors, Stroke, Injury prevention,
Rehabilitation

Background Despite continued development, wearable airbag tech-

Every year, approximately 13 million individuals around nologies are currently only developed with the non-
the world experience a stroke [1, 2]. Falls are one of the neurologically impaired older adult population in mind.
most common medical complications experienced by Furthermore, the internal fall impact detection algo-
individuals after a stroke, reported in up to 65% of the rithms are often developed on data from young par-
stroke population during hospitalization and up to 75% ticipants [21–25]. The algorithms have neither been
in the community [3, 4]. Individuals who have expe- specifically designed nor validated for detecting falls of
rienced a stroke are at an increased vulnerability for individuals presenting with stroke-related motor impair-
falling, related to common correlates of high fall risk ments. A presumption is that the design and computa-
in this population such as impaired mobility, medica- tional models should seamlessly transfer to users in the
tion use, and cognitive impairment [5]. Falling after a stroke population. However, the underlying pathophysi-
stroke can have serious consequences. There are a high ology of a stroke, fundamentally related to the cerebro-
incidence of severe physical injuries, including frac- vascular territory that is compromised in the brain, may
tures, soft tissue and head injuries, and at worst, death manifest with alterations in movement kinematics and a
[6, 7]. Psychologically, individuals often develop a fear loss of ability to control movements. These characteris-
of falling, leading to reduced mobility, increased social tic changes in movement have been observed and quan-
isolation, and significant reduction in quality of life [8, tified in existing literature [26–28] and may translate to
9]. Financially speaking, fall related injuries constitute observed and measurable differences leading up to or
a burden on healthcare systems through prolonged during falls [29–33]. For example, earlier studies have
use of services and incurred high healthcare costs analyzed and compared falls between older able-bodied
[10–12]. Despite evidence that multifactorial reha- and stroke individuals, and found significantly different
bilitation approaches such as improving strength, bal- motor responses including postural stability, trunk con-
ance, and visual impairments can reduce fall incidence trol, fall velocity, and timely step compensation [29, 31–
in the older adult population [13], a recent Cochrane 33]. Furthermore, Dusane et al. found that within a stroke
review concluded that there is little to no evidence of population, kinematic responses differed depending on
interventions that can prevent falls from occurring in the side of the body which a fall was initiated on (paretic
individuals experiencing falls after stroke [14]. There- vs. non-paretic) [34]. Given falls in stroke may have dis-
fore, individuals who suffer from mobility deficits after tinct kinematic profiles, current fall mitigation technol-
a stroke continue to experience falls, frequently and ogy developed on data of generally healthy individuals
repeatedly. Without a way to prevent these falls from might not be sufficiently sensitive or specific to detect
occurring, there is a compelling need to develop meth- falls in individuals who have experienced a stroke. Such
ods and tools to detect these falls before impact with inaccuracy could result in failure to deploy the airbag
the ground and reduce the associated consequences. during a fall or cause unnecessary airbag deployments
One conventional solution to achieve some degree (i.e. false positives) and consequently, may lead to poor
of fall impact mitigation is wearing padded hip protec- user engagement.
tors in or underneath clothing, yet their significance To address these issues, we suggest that systematic fall
in reducing fractures and associated injuries is limited detection models should consider incorporating train-
and their current usage in the community is insignifi- ing data of individuals specific to the intended user
cant (likely due to discomfort and poor compliance) population. Using machine learning to tune movement
[15, 16]. A more novel and recent fall impact mitigation recognition algorithms to unique movements of par-
approach to address these concerns is wearable airbag ticular mobility-impaired populations has been dem-
technology [17–20]. These devices generally include onstrated in various applications for Parkinson’s disease
three design components: (1) at least one sensor, such [35], incomplete spinal cord injury [36, 37], and stroke
as an inertial measurement unit (IMU), to record user [38], yet has not been applied to pre-impact fall detection
motion; (2) a computational model that processes the in stroke populations. Thus, this paper presents consider-
sensor signals to pre-detect fall impact; and (3) an ations for a sensor-based, machine learned wearable air-
inflatable airbag mechanism that deploys upon detec- bag system to demonstrate the importance of pre-impact
tion of a fall, to mitigate contact forces with the ground. fall detection models specific to stroke-related movement
Botonis et al. Journal of NeuroEngineering and Rehabilitation (2022) 19:60 Page 3 of 14

impairments. We hypothesize that for fall detection in Northwestern University Institutional Review Board
the stroke population, a pre-impact fall detection model (NUIRB, IL, USA). All study procedures were carried out
trained on data from a stroke population would perform in accordance with the standards listed in the Declaration
better than models trained on data from a control pop- of Helsinki 1964. This study is a registered clinical trial on
ulation. In other words, failure to train a model on fall ClinicalTrials.gov (registry number NCT05076565) with
movements specific to individuals with a history of stroke the National Library of Medicine (NLM) at the National
will result in decreased pre-impact detection perfor- Institutes of Health (NIH).
mance for users of the stroke population. Furthermore, Participants wore the airbag device and protective gear
we explore secondary considerations for model develop- (e.g., helmet, neck/knee/elbow guards, padded shorts)
ment, including dataset activity composition, severity of while performing falls and non-falls onto a padded mat
gait impairments across users, and lead time parameter under supervision. For the purposes of this study, a fall
tuning. was defined as an event in which the participant loses
balance and results in a terminal position on the ground
Methods (e.g., slips, trips, and falls from chair) [39]. Within falls,
Device concept and prototype we defined subcategories for the directionality of the fall,
Data were recorded via the Wolk Hip Airbag (Wolk De namely in the lateral direction (i.e., lateral fall) and ante-
Heupairbag; Wolk Company, Netherlands), a commer- rior–posterior direction (i.e., AP fall). Participants were
cial airbag system (Fig. 1). Wolk Airbag is a Conformitè instructed to respond to loss of balance using any natu-
Europëenne (CE) marked and Federal Communications ral technique, such as using an arm to catch themselves
Commission (FCC) approved smart system designed to or taking multiple steps, to encourage realistic behavior.
mitigate falls in the healthy older adult population. The Non-falls were performed and stratified for low or high
device is a lightweight, battery-powered belt available movement complexity, in this case related to loss of bal-
in four sizes designed to be worn underneath clothing. ance. Activities of daily living (ADLs) are lower com-
The device includes three fixed IMU sensors (one on the plexity non-fall movements which do not include loss
side of each hip and one on the lower back, in line with of balance (e.g., walking, sitting, lying, sit-to-stands,
the L3 vertebrae) and an onboard computing unit. Each jumping), and are often used to discriminate against fall
IMU sensor contains an accelerometer (range ± 16 g) and events for fall detection technology. On the higher com-
gyroscope sensor (range ± 2000 deg/s) collecting data in plexity end, near-falls are non-fall movements in which
all three axes (x, y, and z) at a sampling rate of 500 Hz. the participant loses balance but recovers without hitting
The device has built-in models for pre- and post-impact the ground, such as recovery from slips, trips, and lateral
fall detection, which send a command signal for deploy- perturbations. The order of the event types (i.e., falls,
ment to the two embedded CO2 cartridges in the event non-falls) and their subcategories (e.g., trips, slips, near-
a fall impact is pre-detected. For this study, a modified falls, ADLs) were randomized between participants. Each
version of the Wolk airbag was utilized. The internal event was observed and labeled by the research team,
circuitry had been altered to incorporate data logging encoding the type of fall activity and its context. Video
capabilities to an external SD card to store the raw IMU recordings of all activities were collected for fall activity
data, and the deployment signal was inhibited to pre- confirmation, post-session processing, and data analy-
vent cartridge activation during the supervised falls data sis. All events were performed at least twice to capture
collection. intra-subject variability. Participants did not use assistive
mobility devices during the session with the exception of
Study design and data collection orthoses if necessary.
Stroke cohort eligibility criteria included individu-
als within the age range of 18–85, diagnosis of stroke at Development of a fall detection model for individuals
least 6 months ago, and the ability to sit unsupported and with stroke‑related impairments
walk independently (at least with an assistive device). The The collected raw IMU data and event times were pro-
control cohort were recruited from the same age range vided as input to an automated custom Python program
with no serious conditions, injuries, or history of back (Fig. 1). Accelerometer data were filtered by a fourth-
pain. Exclusion criteria for both cohorts included per- order band pass Butterworth filter (0.1–50 Hz) to remove
sons on anti-coagulants, severe osteoporosis, pregnant high frequency noise. The pre-impact data window
women, cognitive deficits, and a comorbidity that inter- was extracted based on the fall impact and the accept-
feres with the efficacy of the study or increases minimal able intervention time, which is critically constrained
risk of injury. All participants signed informed consent by the device lead time (i.e., the time prior to impact by
before study participation, which was approved by the which the algorithm must decide to deploy the airbag for
Botonis et al. Journal of NeuroEngineering and Rehabilitation (2022) 19:60 Page 4 of 14

Fig. 1 Airbag Device and Model Pipeline. Sequence of steps in processing and developing the fall prediction models. Kinematic data were
collected from individuals with a history of stroke (n = 20) and individuals who had not experienced a stroke (i.e. control, n = 15) while wearing an
IMU airbag device. Raw IMU accelerometer and gyroscope signals were filtered and segmented for the pre-impact fall data window, ending at the
detected impact time minus the selected lead time duration. Statistical features were extracted from the pre-impact fall data window and labeled
according to cohort membership (control or stroke) and activity type (fall or non-fall, fall type). All features were used as input to two models, each
model trained on either control or stroke data. The control-trained model was trained on all control features, and each subject of the stroke cohort
was tested iteratively. In the stroke-trained model, a single subject was iteratively held out for testing while the remaining subjects were used for
model training (leave-one-subject-out cross-validation scheme). The dashed grey rectangle signifies an iterative selection process of each held out
test subject. Performance metrics to pre-impact detect falls in stroke were compared between the two exclusive cohort trained models

successful intervention) [23]. The minimal lead time for impact, i.e., the last possible moment to classify the fall
the device in use is 75 ms. Thus, for each fall event, the (Fig. 1). Non-fall data were randomly evaluated within
program automatically selected an evaluation data win- the latter half of the event time region to capture near-fall
dow (250 frames) ending 75 ms before the identified fall event features in close proximity to the user’s maximal
Botonis et al. Journal of NeuroEngineering and Rehabilitation (2022) 19:60 Page 5 of 14

instability for the task. Any activities with missing or impaired stroke users. To preliminarily assess the sensi-
defective data (due to occasional hardware damage of a tivity of the models to severity representation, subjects
cable or connection during recording) were flagged and from the available stroke cohort who displayed indica-
removed from analyses. The following features were tors of severe gait impairments during a standardized
engineered for each axis of the acceleration and angular gait assessment were identified and collectively labeled as
velocity from the IMU signals: min, median, max, inter- “unstable ambulators” (see Additional file 2: Table S2). A
quartile range, standard deviation, skew, and kurtosis compilation of test sets was constructed using a “leave-
(see Additional file 1: Table S1) [40]. one-group-out” (LOGO) modeling scheme, each group
The calculated features were used as input to develop a subset of five stroke cohort subjects labeled by quan-
a machine learning classifier for pre-impact fall detec- tity of unstable ambulators being tested on (see Addi-
tion based on an adaptive boosting algorithm [41]. Gra- tional file 3: Figure S1). Iterative model evaluation was
dient boosting algorithms, such as AdaBoost, have been performed for each test group in both the control- and
noted for superior fall detection performance as com- stroke-trained models. Please see the Additional file 5 for
pared to other methods including k-Nearest Neighbor detailed information on this analysis.
(KNN), Support Vector Machine (SVM), and Random Lastly, model performance was investigated for var-
Forest (RF) [42]. The AdaBoost classifier is a multiple ied lead times to determine the optimal lead time which
decision tree classifier that, in sequence, chooses a selec- maximizes model performance for each population-spe-
tive weight by which the average contributes to the final cific model. Lead time, tl , is defined as the time before the
prediction of the test data set. Two identical AdaBoost impact by which the fall must be detected [23]. Instead
models (50 estimators, learning rate = 1) were developed of viewing this lead time parameter as a constraint, we
to detect falls of individuals with stroke-related impair- rather investigate and define an optimal lead time,tl ∗, as
ments. Each model was exclusively trained on data from the duration of time before a fall impact, tp, to end the
a homogeneous population, namely, (i) a model trained data evaluation window, te , resulting in a maximized fall-
only on data of non-stroke participants (the “control- classification performance. The start, ts, and end, te , time
trained” model), and (ii) a model trained only on data of of the data evaluation window for different lead times, tl ,
post-stroke participants (the “stroke-trained” model). A are defined as the following, where w is the duration of
“leave-one-subject-out” (LOSO) modeling scheme was the data window (250 frames or 500 ms):
used to structure train and test sets in the stroke-trained
ts = tp − tl − w
model. The control-trained model was trained on data of
all control subjects and iteratively tested on each stroke
cohort participant (Fig. 1). Both models utilized identi- te = tp − tl
cal processing techniques and features for iterative model
For all analyses, the evaluated metrics on model perfor-
development and testing. Events were predicted and
mance were recall, precision, ­F1-score, receiver operat-
labeled as falls or non-falls.
ing characteristic curve (ROC), and area under the curve
One of the key factors in developing a generalized and
(AUC). In this study, recall is defined as the percentage of
reliable fall detection model is the type of data used for
true falls detected out of all true falls performed, and pre-
the non-fall events. In much of the existing research, falls
cision is defined as the percentage of true falls detected
are improperly classified against static ADLs only, which
out of all the events predicted as falls. The ­F1-score con-
often yields inflated and overly optimistic fall detection
siders both the recall and precision scores to measure
performance. Real life applications require classification
overall performance, given by the following equation:
of more complex motions, such as near-fall events. In
order to demonstrate the importance of non-falls diver- recall · precision
sity, fall detection models that were trained and tested F1 = 2 ·
recall + precision
using exclusively ADLs or near-falls as the non-fall events
were compared. Additionally, as an exploratory analy- While performance metrics are calculated for every
sis, we investigated whether models developed to detect model iteration (i.e. each train and test split), the aver-
falls in post-stroke individuals should include train- age and standard deviation of model performance are
ing data not only of falls from the stroke population but reported to convey overall model performance and
furthermore, falls of individuals with varying levels of reduce complexity for between-model comparisons.
stroke-related impairment severity. Models which are not Two-tailed paired t tests were performed to determine
adequately trained on movements of stroke individuals the relative performance of two models using different
with severe motor impairments may result in a decreased training data. Both models were used to make predic-
pre-impact fall detection performance for severely tions for an identical test set, from which average recall,
Botonis et al. Journal of NeuroEngineering and Rehabilitation (2022) 19:60 Page 6 of 14

precision, ­F1-score, and AUC metrics were calculated. training stage of each model (see Additional file 4: Fig. S2
Each of these performance metrics were recorded per for an illustration of model feature importance).
test iteration in both models, resulting in paired distri- Cohort-dependent models were compared to detect
butions of average performance metrics. For most tests, falls of individuals with stroke-related impairments. For
basic statistical significance is reported if P < 0.05. For training on all categories of falls and non-falls, the stroke-
some instances where repeated t tests were used, signifi- trained model resulted in higher average recall while the
cance after accommodating for the family-wise error rate control-trained model resulted in higher average preci-
is also reported using the Holm-Bonferroni method [43]. sion (Fig. 2). However, the difference in recall was sta-
The Holm-Bonferroni method proceeds as follows: for tistically significant (P = 0.0035), while the difference
each of the m repeated t tests, sort their corresponding P in precision was not. When broken down into specific
values from lowest to highest (P1, P2,…, Pm). Starting with fall types (i.e. lateral falls and AP falls) classified against
the lowest P value, P1, check if: non-fall activities, the stroke-trained model had statisti-
α cally higher recall for both lateral (P = 0.028) and AP falls
Pk < , (P = 0.0036). The control model had 5.7% higher preci-
m+1−k
sion for lateral falls, while the stroke-trained model had
where α is the desired significance level (α = 0.05 10.7% higher precision for AP falls, though neither differ-
throughout this paper). While checking ascending val- ence was statistically significant.
ues of Pk = P1, P2,…, Pm, determine the first Pk such that For lateral falls, the stroke-trained model and con-
the above inequality is satisfied. All subsequent values of trol-trained model had approximately equal AUC and
P (Pk , Pk+1 , . . . , Pm) are considered significant. This pro- ­F1-scores, with less than 3.5% change and no statistical
cedure ensures that the family-wise error rate does not difference. For AP falls, the stroke-trained model had
exceed α, and is less overly conservative than the popular 3.5% higher AUC and 35% higher ­F1-score (P = 0.019).
Dunn-Bonferroni method.
Comparison of data types used for fall and non‑fall events
In addition to evaluating performance for directionally
Results distinct fall types, non-falls of different movement com-
Fall mitigation model performance: control‑ vs plexities (i.e., ADLs, near-falls) were assessed (Table 2).
stroke‑trained models Table 2 shows the results of paired t tests that were used
A summary of participant demographics and session- to compare the distribution of each performance metric
related characteristics are given in Table 1. In total, 842 (recall, precision, ­F1-score, and AUC) between the vali-
falls (610 lateral falls and 232 AP falls) and 961 non-falls dation set of stroke-trained and control-trained models.
(562 ADLs and 399 near-falls) were performed in the Tests that resulted in a statistically significant value (as
experiment. The control cohort constitutes a greater defined by P < 0.05) are indicated by a single *. Tests that
number of recorded movement samples within all sub- are significant, even after compensating for the family-
categories of falls and non-falls, with the exception of wise error rate of using repeated t tests for all three fall
lateral falls. In total, 126 features per fall sample were types (all falls, lateral falls, AP falls), are marked with **.
included in the feature matrix, provided as input to the Using ADLs as non-fall data resulted in generally higher
model performance compared to using near-falls as non-
fall data. This effect was most pronounced in the con-
trol model for AP falls, where using ADLs as non-fall
Table 1 Participant demographic and session information data resulted in statistically significantly higher recall
(P = 0.0065), AUC (P = 0.042), and F ­ 1-score (P = 0.037).
Characteristics Unit Stroke (n = 20) Controls (n = 15)
The difference was less pronounced for the stroke-trained
Mean Age (std) years 53 (13) 40 (16) model, which only has statistically significantly higher
Sex – 10 M / 10 F 8M/7F AUC when using ADL data, both when analyzing AP falls
Mean Height (std) cm 168.3 (12.0) 170.7 (12.8) (P = 0.033) or the group of all falls (P = 0.021), but not for
Mean Weight (std) kg 80.6 (21.5) 80.3 (12.3) the group of lateral falls alone.
Mean BMI (std) kg/m2 28.1 (5.15) 27.9 (5.97)
Time Since Stroke (std) years 7.47 (4.58) – Model performance for a spectrum of stroke‑related
Mean Gait Speed (std) m/s 0.88 (0.29) 1.22 (0.19) impairment
Falls Count (Lateral/AP) – 439 (327/112) 403 (283/120) The above models evaluated training data based on
Non-Falls Count (ADL/ – 415 (259/156) 546 (303/243) binary presence or absence of stroke, yet within the
Near)
stroke population there is a spectrum of severity levels
Botonis et al. Journal of NeuroEngineering and Rehabilitation (2022) 19:60 Page 7 of 14

Fig. 2 Pre-impact fall detection performance of post-stroke individuals’ falls tested on control- and stroke-trained models. Average
performance metrics of recall, precision, ­F1-score, and AUC for pre-impact fall detection of post-stroke individuals’ falls using an AdaBoost Classifier
trained on control (blue) or stroke (red) cohort data. Each model is stratified and compared across different fall types: (A) all falls, (B) lateral falls, (C)
AP falls using paired t tests. Tests that resulted in a statistically significant difference (as defined by P < 0.05) are indicated by a single *. Tests that
are significant after compensating for the family-wise error rate of using repeated t tests for all three fall types, as defined by the Holm-Bonferroni
method, are marked with **
Botonis et al. Journal of NeuroEngineering and Rehabilitation (2022) 19:60 Page 8 of 14

Table 2 Effect of non-fall type (ADLs or near-falls) on pre-impact fall detection for stroke- and control-trained models
Recall Precision F1-Score AUC​
M SD M SD M SD M SD

Stroke-trained model
All falls
ADL 0.95 0.07 0.94 0.06 0.94 0.05 0.97 0.04
Near-fall 0.93 0.07 0.92 0.08 0.92 0.06 0.92 0.08
P 0.53 0.3 0.27 0.02*
Lateral falls
ADL 0.93 0.08 0.96 0.06 0.94 0.06 0.97 0.05
Near-Fall 0.92 0.08 0.93 0.10 0.92 0.07 0.95 0.05
P 0.73 0.28 0.33 0.19
AP Falls
ADL 0.79 0.23 0.84 0.20 0.79 0.19 0.97 0.05
Near-Fall 0.80 0.23 0.89 0.19 0.82 0.19 0.88 0.16
P 0.81 0.44 0.65 0.03*
Control-trained model
All falls
ADL 0.88 0.12 0.92 0.07 0.89 0.08 0.94 0.08
Near-fall 0.85 0.10 0.96 0.06 0.90 0.08 0.91 0.13
P 0.42 0.07 0.85 0.37
Lateral falls
ADL 0.91 0.13 0.97 0.04 0.94 0.08 0.98 0.04
Near-fall 0.89 0.07 0.95 0.08 0.92 0.06 0.96 0.05
P 0.56 0.25 0.45 0.15
AP falls
ADL 0.75 0.22 0.82 0.25 0.76 0.21 0.94 0.07
Near-Fall 0.49 0.34 0.76 0.40 0.57 0.34 0.86 0.15
P 0.01** 0.58 0.04* 0.04*
Average recall, precision, F1-score, and AUC reported for stroke-trained (top) and control-trained (bottom) models to compare the effect of non-fall type (ADL, near-
fall) on model performance for each category of fall types (all, lateral, AP). For each fall type, P values are reported to compare the significance when using ADLs versus
near-falls for the non-falls type in the model. Tests that resulted in a statistically significant value (as defined by P < 0.05) are indicated by a single *. Tests that are
significant after compensating for the family-wise error rate of using repeated t tests for all three fall types are marked with **

related to mobility and fall risk. Five stroke cohort sub- unstable ambulators in the test set. Consistent with pre-
jects were identified and labeled as unstable ambulators vious findings, fall detection recall and precision face a
(i.e., displaying some form of gait instability) from the general tradeoff in performance between the control- and
available dataset (see Additional file 5). The distribu- stroke-trained models. However, these results suggest fall
tion of model performance is demonstrated below in performance may be correlated to the quantity of unsta-
Fig. 3, separated and ordered by an increasing quantity of ble ambulators within the test set. In general, an increase

(See figure on next page.)

Fig. 3 The effect of an increasing quantity of test subjects with severe stroke-related impairments on pre fall detection. Average recall,
precision, ­F1-score, and AUC are displayed for unique, randomly selected subgroups of stroke individuals (n = 5 per subgroup) tested on both
control- and stroke-trained models. Performance is stratified by the quantity of stroke individuals with severe stroke-related impairments (i.e.
unstable ambulators) included in the test set. For zero to three unstable ambulators in a test set, 100 unique subgroups were input into both
models. For four or five unstable ambulators in a test set, the maximum number of unique possible subgroups were used, i.e. 75 and 1 subgroup(s),
respectively. Visual trends are displayed to demonstrate how an increase in the severity of the test group impacts model performance for each
subcategory of fall types: (A) all falls, (B) lateral falls, (C) AP falls. Notably, control-trained model performance generally declines with an increase in
unstable ambulators tested upon, while stroke-trained models are unaffected or even result in an improved performance with an increased number
of unstable ambulators tested upon
Botonis et al. Journal of NeuroEngineering and Rehabilitation (2022) 19:60 Page 9 of 14

Fig. 3 (See legend on previous page.)

Botonis et al. Journal of NeuroEngineering and Rehabilitation (2022) 19:60 Page 10 of 14

in the tested quantity of unstable ambulators decreased lead times of 150 ms for the control-trained model and
the recall and precision of the control model, while the 300 ms for the stroke-trained model. As shown for both
stroke-trained model was more or less unaffected with an models, the lowest lead time (50 ms) did not result in
increase in the quantity of unstable ambulators tested on. the highest-performing lead time and in addition, the
Accordingly, this led to a larger gap between the recall decrease or increase in AUC was not linear.
and a narrowed gap in precision between the two mod-
els, suggesting differences in the trained models’ ability to
predict falls of more severely impaired stroke individuals. Discussion
Here we present for the first time research on a wearable
Model performance for different lead times airbag technology specific to detecting falls in individuals
To test the effect of the lead time on the model’s per- that have experienced a stroke, a population at high fall
formance, we compared the AUC-ROC curves for both risk with minimal evidence for reliable fall intervention
models to classify AP falls against each non-fall subtype strategies [14]. In this work, we investigate the impor-
separately (ADLs and near-falls) with varied lead times tance of population-specific models to detect falls in
(50, 100, 150, 200, 300, 400, 500 ms). The results show stroke by purposefully altering model-dependent train-
that for the events recorded in this study, the lead times ing data sets and analyzing the consequent effects on
resulting in the best performance (i.e., highest AUC val- performance. The resulting model comparisons support
ues) differed between control- and stroke-trained mod- our primary hypothesis, i.e. that stroke-specific train-
els (Fig. 4). For classifying the unstable ambulators’ AP ing data improves pre-impact fall detection in stroke,
falls against ADLs, the control-trained model performed and demonstrate the influence of activity composition,
best at 100 ms lead time (95.9% AUC) while the stroke- impairment severity, and lead time on stroke-specific
trained model performed best within 150 ms–300 ms model performance. Incorporating a population-specific
lead time (94.5% AUC). This lead time difference main- model into fall mitigation technology tailored to the
tained a similar ratio when classifying against near- motion of ambulatory individuals after stroke could help
falls, with peak AUC scores of 88.7% in both models at

Fig. 4 Pre-impact fall detection optimal lead time dependent on model and non-fall type. An AdaBoost Classifier trained on control or
stroke data were used to classify AP falls against ADLs or near-falls for a test subgroup of unstable ambulatory subjects (n = 5). AUC-ROC curves
are displayed for varied lead times (50, 100, 150, 200, 300, 400, 500 ms) for classifying AP falls from ADLs in the control-trained model (A) and
stroke-trained model (B), and from near-falls in the control-trained model (C) and stroke-trained model (D). The control-trained model performs
similarly to the stroke-trained model’s maximal performance with a ~ 50% lead time reduction
Botonis et al. Journal of NeuroEngineering and Rehabilitation (2022) 19:60 Page 11 of 14

to reliably mitigate the impact of falls and their related detected further upstream from the fall impact, perhaps
consequences in the stroke population. during the ambulation period leading up to a fall. How-
This work supports recent initiatives for “data-centric ever, this trend was not linear in either model. In some
artificial intelligence”, encouraging the formulation of cases, for the stroke-trained model, increasing the lead
better datasets (specific to an end model’s intended goal) time over 300 ms decreased the model performance. This
rather than overdeveloping algorithms around poor data finding suggests that investigating a variety of lead times
[44]. While studies have presented wearable airbag tech- may improve model performance evaluation, and could
nologies for fall mitigation, their target population was furthermore be beneficial to optimization of fall mitiga-
non-neurologically impaired and otherwise healthy older tion technologies (i.e. decreased allocation of expenses to
adults and thus, detection models were developed and maximally reducing the lead time of the device when not
validated based on data of healthy individuals. To the best optimal). However, lead times should be optimized with
of our knowledge, the fall mitigation model presented in the end user environmental domain in mind, especially
the current study is the first which intentionally encom- during device validation. It may be of interest in future
passes movement data of stroke survivors in the training work to implement a time series classification structure
set of the detection model. Results demonstrate that the in which previous movement states are used to track falls
model performance for detecting falls in individuals after progressing over time, such as in a long short-term mem-
stroke is improved (increased recall and ­F1-score) when ory network. Such a model may benefit users requiring
the model is trained using movement data of individu- different windows of time and selection criteria to con-
als after stroke, as opposed to data of healthy individu- fidently pre-impact detect a fall at the earliest possible
als. In this, a trend in the recall-precision tradeoff was stage.
observed, namely, stroke-trained models performed with For this dataset, the direction of a fall seems to be of
an averaged higher recall while control-trained models particular importance for model performance. Specifi-
performed with an averaged higher precision (though cally, while the stroke-trained model is much better at
the precision difference was not statistically significant). detecting AP falls in post-stroke participants than the
The resulting difference in recall between the control- control-trained model, this improvement is less sub-
and stroke-trained models suggests that current fall stantial when distinguishing falls to the side. This may
mitigation technology developed on data of able-bodied be because stroke-related impairments do not affect the
individuals could be further improved to better protect kinematics of all movements and functional activities
individuals with neuromotor impairments such as stroke, equally. For example, Hollands et al. studied full-body
supporting our hypothesis to consider stroke-specific kinematics of turning 180 degrees during the Timed Up
data in the model development. and Go test [49] for patients following stroke compared
Development and validation of generalized fall detec- to healthy age-matched controls. They found that while
tion models requires not only the aforementioned inclu- stroke survivors did take longer to turn, there were no
sion of data from the intended user population, but also significant kinematic differences demonstrated in turn
a diverse and complex set of non-fall data [45, 46]. The performance or in axial segment coordination dur-
results demonstrate that using only ADLs as non-fall ing turning between the two groups [50]. Extrapolating
data may result in inflated model performance compared these results to our study, perhaps lateral falls for post-
to the models’ true ability to detect falls in everyday life stroke individuals and healthy controls appear kinemati-
[47, 48]. This was especially apparent for the condition cally similar enough in IMU output that no significant
in classifying AP falls from near-falls rather than ADLs, improvements occur when training the model with
which significantly affected the performance of both con- stroke-specific data compared to control data. Nonethe-
trol- and stroke-trained models. less, the need for stroke-specific fall detection training
For the model presented in this work, selection of the relies on the hypothesis that there are significant and
lead time has a meaningful impact on model training and measurable differences in movement of individuals with
evaluation. Using the shortest possible lead time may not stroke-related impairments, which we believe is demon-
result in the best window of data to assess the model’s strated by the improved results for classification of AP
true ability to detect falls. The results of this study convey falls. These nuances indicate a potential need for more
that the control-trained models utilized data closer to the research to distinguish which movements and falls are
time of impact (shorter lead time) than the stroke-trained most effected kinematically by stroke-related impair-
models to maximize performance. This could indicate ments, so that stroke-specific model training can be min-
that the pre-impact motion features of control versus imized when control movement data is sufficient.
stroke individuals differ relative to the time of the fall The current study design includes limitations. First, the
impact itself. Stroke features related to falling were better fall detection model was developed based on a limited
Botonis et al. Journal of NeuroEngineering and Rehabilitation (2022) 19:60 Page 12 of 14

sample size. While this dataset is novel in its content of unstable ambulators in the test set changed performance
falls initiated from individuals with a history of stroke, as hypothesized, it remains uncertain whether perfor-
future studies could expand upon the variety of data col- mance change is attributed to true movement severity
lected by using more participants with a wider range of in this sub group, or rather to a specific movement pat-
demographics and impairment levels. A related con- tern which our criterion selected for. In both cases, the
founding factor was that participants within the stroke results demonstrate that individuals with distinct move-
cohort performed fewer falls and non-falls on aver- ment kinematics will experience decreased user pro-
age than control participants, due to considerations of tection when not represented in the training set. It is
safety and experiment duration. To reduce the effect of recommended that future studies incorporate clinically
data imbalance between cohorts, the stroke model was validated assessment tools in addition to gait kinemat-
trained on data from 19 post-stroke individuals, and the ics to label stroke impairment severity and thus, ensure
control model on data from 15 control individuals. For that the final model accounts for individuals across the
each tested stroke subject, this resulted in a stroke train- impairment spectrum. It may also be useful to consider
ing data set with approximately 27.8% fewer non-fall the symptomatic cause of increased fall risk (e.g. hemipa-
events and 3.5% more fall events (only exceeding the con- resis vs loss of coordination) to further ensure inclusion
trol cohort in lateral fall count, in which there was no sta- of specific participant profiles. This stratification may
tistically significant difference in classification) as shown allow for a more generalized, universal machine learning
in Table 1. This training data arrangement was used while model for stroke impairment, or allow user input of an
comparing the efficacy of the cohort-dependent trained identified stroke presentation to fine tune parameters of a
models (Fig. 2) and evaluating the effect of non-fall fall detection model accordingly.
type on fall detection (Table 2). While the training data While an aim of this study was to demonstrate the
sets as described provide a compromise between sub- effect of non-falls complexity on model performance,
ject count and the quantity of trained samples between more complex activities that occur in day-to-day living
the two models, a leave-one-group-out schematic was or which contribute to falling could have been included.
additionally attempted to balance subject count in the Additionally, the models in this study were developed
analysis of stroke-related impairments (Fig. 3). In this using the AdaBoost classification method, yet it is pos-
case, the stroke training set had the same number of sub- sible that other machine learning approaches which were
jects, fewer falls, and fewer non-falls, but still has better not investigated for the purposes of this study might
or equal performance metrics. We believe this elevated result in better performance. Rather than present the
performance, despite having fewer data to train with, is best classification model at this stage, this study aimed
strong support for our central message: data from stroke to demonstrate how training data composition can affect
subjects, especially of representative severity, is greatly classification performance related to falls in stroke popu-
beneficial for training pre-impact fall detection algo- lations. While the statistical features used in these mod-
rithms for the stroke population. els did reveal statistically significant differences in the
The method used to select participants with severe cohort-based model predictions, we aim to extract more
impairments in the stroke cohort is approximate and clinically intuitive features with gold standard validation
includes several assumptions. While many studies techniques in future work, as to better understand the
observe gait characteristics as indicators of post-stroke kinematic differences between control and stroke popu-
mobility deficits, such as gait speed and changes in trunk lations. Finally, the airbag system was not deployed into
asymmetry and instability [48, 51–53], these measures a real-world setting, which is one of the next steps in our
may not be the most indicative for risk of falling or move- research. Conclusions made in regard to control- and
ment severity. For the available data, it was not possible stroke-trained models should be validated by the deploy-
to assess whether greater range of motion about the lower ment of the model and prototype into the community
trunk during gait was truly correlated to poor instabil- for those individuals at risk of falling. Capturing real-fall
ity or was rather a by-product of compensatory mecha- data and testing model reliability and feasibility in both
nisms [53]. In addition, five of the stroke subjects could the community and clinical settings could strengthen the
not be assessed for severity due to missing data. For true arguments made by this study.
fall risk diagnosis in individuals with a history of stroke,
numerous factors are taken into consideration and cul- Conclusions
minated to assess safe ambulation, including sensorimo- In summary, we present a sensor-driven wearable airbag
tor deficits, poor balance in a variety of contexts, visual technology for pre-detecting falls in individuals’ post-
and cognitive impairment, medication use, and impulsiv- stroke. Results demonstrate the importance of devel-
ity [5, 54]. While an increased quantity of the identified oping fall mitigation technology which utilizes motion
Botonis et al. Journal of NeuroEngineering and Rehabilitation (2022) 19:60 Page 13 of 14

data specific to the stroke population. Even further, Declarations

these results demonstrate that a “one-size-fits-all” pre-
Ethics approval and consent to participate
impact fall detection model based on healthy data may All participants signed informed consent before study participation, which
not extend to protecting other neurological, orthopedic, was approved by the Northwestern University Institutional Review Board
and neuromuscular conditions, including but not lim- (NUIRB, IL, USA). All the study procedures were carried out in accordance with
the standards listed in the Declaration of Helsinki 1964.
ited to Parkinson’s Disease, lower-limb amputation, and
multiple sclerosis. Our insights could help researchers, Consent for publication
clinicians, and companies to develop better fall detection Not applicable.

models and advance fall mitigation technologies for pop- Competing interests
ulation-specific individuals at high risk of falling. The use The authors have no competing interests to disclose.
of such devices could help individuals after stroke and
Author details
other conditions to reduce the risk of fractures and inju- 1
Max Nader Rehabilitation Technologies and Outcomes Lab, Shirley Ryan Abili-
ries and reduce their fear of falling, thus improving their tyLab, Chicago, IL, USA. 2 Department of Physical Medicine and Rehabilitation,
overall health and their quality of life. Northwestern University, Chicago, IL, USA. 3 Northwestern University Feinberg
School of Medicine, Chicago, IL, USA. 4 Department of Computer Science
and Engineering, Department of Biomedical Engineering, University of North
Texas, Denton, TX, USA. 5 Department of BioMechanical Engineering, Delft
Abbreviations University of Technology, Delft, The Netherlands. 6 Department of Rehabilita-
AP: Anterior–posterior; ADL: Activity of daily living; AUC​: Area under the tion Medicine, Erasmus MC, Rotterdam, The Netherlands.
receiver operating characteristic; IMU: Inertial measurement unit.
Received: 8 February 2022 Accepted: 26 May 2022
Supplementary Information
The online version contains supplementary material available at https://​doi.​
org/​10.​1186/​s12984-​022-​01040-4.
References
Additional file 1: Table S1. Statistical features for pre-impact fall detec- 1. Feigin VL, Krishnamurthi RV, Theadom AM, Abajobir AA, Mishra SR,
tion model. Ahmed MB, et al. Global, regional, and national burden of neurological
Additional file 2: Table S2. Selection of unstable ambulators in the stroke disorders during 1990–2015: a systematic analysis for the Global Burden
cohort. of Disease Study 2015. Lancet Neurol. 2017;16:877–97.
2. Lindsay MP, Norrving B, Sacco RL, Brainin M, Hacke W, Martins S, et al.
Additional file 3: Fig. S1. Modeling scheme to stratify stroke-related World Stroke Organization (WSO): Global Stroke Fact Sheet 2019. Int J
impairments. Stroke. 2019;14:806–17.
Additional file 4: Fig. S2. Feature importance for a fully-trained stroke 3. Ashburn A, Hyndman D, Pickering R, Yardley L, Harris S. Predicting people
pre-impact fall detection model. with stroke at risk of falls. Age Ageing. 2008;37:270–6.
4. Langhorne P, Stott DJ, Robertson L, MacDonald J, Jones L, McAlpine C,
Additional file 5. Supplementary methods addressing detailed work and et al. Medical complications after stroke: a multicenter study. Stroke.
additional files Table S1, Table S2, Fig. S1, and Fig. S2. 2000;31:1223–9.
5. Xu T, Clemson L, O’Loughlin K, Lannin NA, Dean C, Koh G. Risk factors for
falls in community stroke survivors: a systematic review and meta-analy-
Acknowledgements
sis. Arch Phys Med Rehabil. 2018;99(3):563–73.
The authors are thankful to the participants for their time and effort in being a
6. Schmid AA, Klar Yaggi H, Burrus N, McClain V, Austin C, Ferguson J, et al.
part of this research. We would also like to thank Filippo van Hellenberg Hubar,
Circumstances and consequences of falls among people with chronic
Michiel Van Lohuijzen, and Randy Stakelbeek from Wolk Company for their
stroke. J Rehabil Res Dev. 2013;50:1277–86.
hardware and firmware support. We also like to extend our thanks to Emily
7. World Health Organization. Global health estimates 2016: estimated
Blankenberger, SPT for the help with data collection.
deaths by age, sex and cause. 2018.
8. Schmid AA, Rittman M. Fear of falling: an emerging issue after stroke. Top
Author contributions
Stroke Rehabil. 2007;14:46–55.
OKB and YH are co-first authors. Conception, design, and study direction: YH,
9. Andersson ÅG, Kamwendo K, Appelros P. Fear of falling in stroke patients:
CKM, MVA, VH, and AJ. Data acquisition: OKB, YH, CKM, DR. Data analysis: OKB,
relationship with previous falls and functional characteristics. Int J Rehabil
YH, and KE. Manuscript writing: OKB, YH, KE, CKM, DR, MG, MVA, VH, and AJ. All
Res. 2008;31:261–4.
authors read and approved the final manuscript.
10. Stevens JA, Corso PS, Finkelstein EA, Miller TR. The costs of fatal and non-
fatal falls among older adults. Inj Prev. 2006;12:290–5.
Funding
11. Carroll NV, Slattum PW, Cox FM. The cost of falls among the community-
This study was funded by the Department of Health and Human Services,
dwelling elderly. J Manag Care Pharm. 2005;11:307–16.
National Institute on Disability, Independent Living, and Rehabilitation
12. Bh A, Fp R, Me W. The cost and frequency of hospitalization for fall-related
Research (NIDILRR), RERC Grant # 90REGE0003-01-00.
injuries in older adults. Am J Public Health. 1992;82:1020–3.
13. Stevens JA, Burns E. A CDC Compendium of effective fall interventions.
Availability of data and materials
2015.
De-identified data are available from the corresponding author upon reason-
14. Denissen S, Staring W, Kunkel D, Pickering RM, Lennon S, Geurts AC,
able request.
et al. Interventions for preventing falls in people after stroke. Cochrane
Database Syst Rev. 2019;10(10):CD008728.
15. Van Schoor NM, Smit JH, Twisk JWR, Bouter LM, Lips P. Prevention of hip
fractures by external hip protectors: a randomized controlled trial. J Am
Med Assoc. 2003;289:1957–62.
Botonis et al. Journal of NeuroEngineering and Rehabilitation (2022) 19:60 Page 14 of 14

16. Santesso N, Carrasco-Labra A, Brignardello-Petersen R. Hip protectors for 40. Shawen N, Lonini L, Mummidisetty CK, Shparii I, Albert MV, Kording K,
preventing hip fractures in older people. Cochrane Database Syst Rev. et al. Fall detection in individuals with lower limb amputations using
2014. https://​doi.​org/​10.​1002/​14651​858.​CD001​255.​pub5. mobile phones: machine learning enhances robustness for real-world
17. Tamura T, Yoshimura T, Sekine M, Uchida M, Tanaka O. A wearable airbag applications. JMIR mHealth uHealth. 2017;5:e151.
to prevent fall injuries. IEEE Trans Inf Technol Biomed. 2009;13:910–4. 41. Freund Y, Schapire RE. A decision-theoretic generalization of on-line
18. Fukaya K, Uchida M. Protection against impact with the ground using learning and an application to boosting. Lect Notes Comput Sci (includ-
wearable airbags. Ind Health. 2008;46:59–65. ing Subser Lect Notes Artif Intell Lect Notes Bioinformatics). Springer
19. Shi G, Chan CS, Li WJ, Leung KS, Zou Y, Jin Y. Mobile human airbag system Verlag; 1995. p. 23–37.
for fall protection using mems sensors and embedded SVM classifier. IEEE 42. Zurbuchen N, Bruegger P, Wilde A. A Comparison of Machine Learning
Sens J. 2009;9:495–503. Algorithms for Fall Detection using Wearable Sensors. 2020 Int Conf
20. Ahn S, Choi D, Kim J, Kim S, Jeong Y, Jo M, et al. Optimization of a pre- Artif Intell Inf Commun ICAIIC 2020. Institute of Electrical and Electronics
impact fall detection algorithm and development of hip protection Engineers Inc.; 2020;427–31.
airbag system. Sensors Mater. 2018;30:1743–52. 43. Aickin M, Gensler H. Adjusting for multiple testing when reporting
21. Rescio G, Leone A, Siciliano P. Supervised machine learning scheme for research results: the Bonferroni vs Holm methods. Am J Public Health.
electromyography-based pre-fall detection system. Expert Syst Appl. 1996;86:726–8.
2018;100:95–105. 44. ’Ng A. A Chat with Andrew on MLOps: From Model-centric to Data-
22. Zhen T, Mao L, Wang J, Gao Q. Wearable preimpact fall detector using centric AI. DeepLearningAI. 2021. https://​www.​youtu​be.​com/​watch?v=​
SVM. Proc Int Conf Sens Technol ICST. 2016;1–6. 06-​AZXmw​Hjo. Accessed 23 Aug 2021.
23. Aziz O, Russell CM, Park EJ, Robinovitch SN. The effect of window size and 45. Broadley RW, Klenk J, Thies SB, Kenney LP, Granat MH. Methods for the
lead time on pre-impact fall detection accuracy using support vector real-world evaluation of fall detection technology: a scoping review. Sen-
machine analysis of waist mounted inertial sensor data. 2014 36th Annu sors. 2018;18:2060.
Int Conf IEEE Eng Med Biol Soc EMBC 2014. Institute of Electrical and 46. Pang I, Okubo Y, Sturnieks D, Lord SR, Brodie MA. Detection of near falls
Electronics Engineers Inc.; 2014. p. 30–3. using wearable devices. J Geriatr Phys Ther. 2019;42:48–56.
24. Zurbuchen N, Wilde A, Bruegger P. A machine learning multi-class 47. Özdemir A, Barshan B. Detecting falls with wearable sensors using
approach for fall detection systems based on wearable sensors with a machine learning techniques. Sensors MDPI AG. 2014;14:10691–708.
study on sampling rates selection. Sensors. 2021;21:938. 48. Punt M, Bruijn SM, Wittink H, Van De Port IG, Van Dieën JH. Do clinical
25. Leone A, Rescio G, Siciliano P. Fall risk evaluation by surface electromyo- assessments, steady-state or daily-life gait characteristics predict falls in
graphy technology. 2017 Int Conf Eng Technol Innov Eng Technol Innov ambulatory chronic stroke survivors? J Rehabil Med. 2017;49:402–9.
Manag Beyond 2020 New Challenges, New Approaches, ICE/ITMC 2017 49. Podsiadlo D, Richardson S. The timed “Up & Go”: a test of basic functional
- Proc. 2018. mobility for frail elderly persons. J Am Geriatr Soc. 1991;39:142–8.
26. do Garcia FV, da Cunha MJ, Schuch CP, Schifino GP, Balbinot G, Pagnussat 50. Hollands KL, Hollands MA, Zietz D, Wing AM, Wright C, van Vliet P. Kin-
AS. Movement smoothness in chronic post-stroke individuals walking ematics of turning 180° during the timed up and go in stroke survivors
in an outdoor environment—a cross-sectional study using IMU sensors. with and without falls history. Neurorehabil Neural Repair. 2009;24:358–
PLoS ONE. 2021;16: e0250100 67. https://​doi.​org/​10.​1177/​15459​68309​348508.
27. Kao PC, Dingwell JB, Higginson JS, Binder-Macleod S. Dynamic instability 51. Hacmon RR, Krasovsky T, Lamontagne A, Levin MF. Deficits in interseg-
during post-stroke hemiparetic walking. Gait Posture. 2014;40:457–63. mental trunk coordination during walking are related to clinical balance
28. Ds M, Jj E. Altered timing of postural reflexes contributes to falling in and gait function in chronic stroke. J Neurol Phys Ther. 2012;36:173–81.
persons with chronic stroke. Exp brain Res. 2006;171:459–68. 52. Isho T, Usuda S. Association of trunk control with mobility performance
29. Honeycutt CF, Nevisipour M, Grabiner MD. Characteristics and adaptive and accelerometry-based gait characteristics in hemiparetic patients with
strategies linked with falls in stroke survivors from analysis of laboratory- subacute stroke. Gait Posture. 2016;44:89–93.
induced falls. J Biomech. 2016;49:3313–9. 53. Van Criekinge T, Saeys W, Hallemans A, Velghe S, Viskens PJ, Vereeck L,
30. Punt M, Bruijn SM, van Schooten KS, Pijnappels M, van de Port IG, Wittink et al. Trunk biomechanics during hemiplegic gait after stroke: a system-
H et al. Characteristics of daily life gait in fall and non fall-prone stroke atic review. Gait Posture. 2017;54:133–43.
survivors and controls. J Neuroeng Rehabil. 2016;13:67 54. O’Sullivan S, Schmitz T, Fulk G. Physical Rehabilitation, 6th edition. Fac
31. Salot P, Patel P, Bhatt T. Reactive Balance in Individuals With Chronic Bookshelf. 2014. https://​hsrc.​himme​lfarb.​gwu.​edu/​books/​85. Accessed 23
Stroke: Biomechanical Factors Related to Perturbation-Induced Backward Aug 2021.
Falling [Internet]. 2016. https://​acade​mic.​oup.​com/​ptj/​artic​le/​96/3/​338/​
26865​36
32. Patel PJ, Bhatt T. Does aging with a cortical lesion increase fall-risk: Publisher’s Note
examining effect of age versus stroke on intensity modulation of Springer Nature remains neutral with regard to jurisdictional claims in pub-
reactive balance responses from slip-like perturbations. Neuroscience. lished maps and institutional affiliations.
2016;333:252–63.
33. Joshi M, Patel P, Bhatt T. Reactive balance to unanticipated trip-like per-
turbations: a treadmill-based study examining effect of aging and stroke
on fall risk. Int Biomech. 2018;5:75–87.
34. Dusane S, Gangwani R, Patel P, Bhatt T. Does stroke-induced sensorimo-
tor impairment and perturbation intensity affect gait-slip outcomes? J
Ready to submit your research ? Choose BMC and benefit from:
Biomech. 2021;118:110255.
35. Albert MV, Toledo S, Shapiro M, Koerding K. Using mobile phones for
• fast, convenient online submission
activity recognition in Parkinson’s Patients. Front Neurol. 2012;3:158.
36. Albert MV, Azeze Y, Courtois M, Jayaraman A. In-lab versus at-home activ- • thorough peer review by experienced researchers in your field
ity recognition in ambulatory subjects with incomplete spinal cord injury. • rapid publication on acceptance
J Neuroeng Rehabil. 2017;14:1–6.
• support for research data, including large and complex data types
37. Sok P, Xiao T, Azeze Y, Jayaraman A, Albert MV. Activity recognition for
incomplete spinal cord injury subjects using hidden markov models. IEEE • gold Open Access which fosters wider collaboration and increased citations
Sens J. 2018;18:6369–74. • maximum visibility for your research: over 100M website views per year
38. O’Brien MK, Shawen N, Mummidisetty CK, Kaur S, Bo X, Poellabauer C,
et al. Activity recognition for persons with stroke using mobile phone At BMC, research is always in progress.
technology: toward improved performance in a home setting. J Med
Internet Res. 2017;19:e184. Learn more biomedcentral.com/submissions
39. World Health Organization. Falls. 2021;