Editorials

Artificial intelligence and diagnosis

in general practice

INTRODUCTION
Diagnosis in general practice is difficult.
Within the populations that GPs care for,
“… there is growing interest in using AI to supplement,
serious diseases and disorders are less enhance, or even replace GP diagnostic intelligence.”
prevalent and, generally, at an earlier
stage in the development of symptoms
than in other healthcare settings.1 However, to train AI diagnostic systems.3 Based on variable quality and consistency.
the public, the media, and our specialist an approach known as supervised learning, Several AI-based symptom checkers
colleagues want primary care clinicians to the programmer trains the system by have been developed with the aim of
become much better diagnosticians. GPs defining a set of diagnoses for a range of improving the accuracy and the reliability
are now expected to spot individuals with images labelled by human experts. of any advice, guidance, or signposting.
cancer earlier while also avoiding excessive Although clinical information from In a vignette-based audit of 23 symptom
testing or inappropriate onward referrals. general practice is more ill defined, AI checkers available for use by the public, the
Currently, there is growing interest natural language processing techniques appropriate triage advice was considered to
in using artificial intelligence (AI) to now provide the tools to analyse more have been provided in 57% of evaluations,
supplement, enhance, or even replace GP unstructured data from records or patients. rising to 80% for emergency cases.4
diagnostic intelligence. Advocates of AI have The outputs are then in a form that is more Unfortunately, the major problem with
suggested that such technologies could amenable to further analysis by machine most symptom checkers is that they are
improve diagnostic accuracy (with fewer learning. risk averse, encouraging users to seek
underdiagnoses and overdiagnoses) in AI is a focused technology, and this help from their GP for conditions when
addition to diagnostic efficiency. But others is, on the face of it, at variance with the self-care would be perfectly reasonable
have argued that it will add yet another comprehensive and holistic nature and safe. Striking a sensible balance
information burden during an already busy of primary care medicine. Diagnostic between missing a serious condition and
surgery. It might do very little to improve intelligence in general practice is much overwhelming GP surgeries is tricky. As
patient outcomes, GP stress levels, or more than generating a possible label or many symptoms presenting to GPs are
healthcare finances.2 suggesting a course of action.1 The GP self-limiting, one option to consider is the
AI refers to computing techniques that needs to understand the patient’s context development of AI tools that enable certain
resemble human cognitive functions, in addition to picking up social and symptoms to be monitored over time.5 This
such as learning and problem solving. psychological cues with empathy, care, and symptom-based information might also
Rather than following pre-programmed compassion. Explicit knowledge about the be supplemented by clinical data provided
instructions, AI approaches draw inferences predictive value of symptoms can be taught
directly by individuals using, for example,
from, or discover, patterns in data. Expert to a machine but tacit knowledge such
self-testing or wearable biosensors.
systems emulate human decision making as how to gain an individual’s confidence
Images from an individual’s mobile phone
by using an inference engine to apply rules might not.
of a skin lesion can already be assessed by
to facts held in a knowledge base. Machine In this editorial we will consider some
an AI system using deep learning to rule out
learning — including neural networks or possible roles for AI to augment the current
skin cancers. This presents the opportunity
deep learning — develops models to make diagnostic approach within primary care
to increase the numbers of individuals
decisions, or to classify examples. This settings.
is achieved by using a training algorithm undergoing skin cancer screening and
— supplied with data of suitable quantity THE PATIENT’S PRESENTATION signpost the most appropriate ones to see
and quality — to adjust the weights within their GP for further assessment.6
One of the challenges to improving health
a model to optimise the goodness of fit outcomes from conditions such as cancer,
between the input and the output. THE GP CONSULTATION
depression, and rheumatoid arthritis is to
All AI techniques depend on high-quality get the right individuals to the right place Clinical decision making is often partially
information on which to learn and classify at the right time. Before a person presents dependent on a number of cognitive
clinical data in relation to outcomes. To to their GP, they will often have sought heuristics or ‘rules of thumb’ in order
date, most AI techniques have focused on advice from friends, their family, or the local to aid recall or understand knowledge.
radiology, as medical imaging data have pharmacist. Most will have also checked Unfortunately, as a consequence, GPs may
been systematically stored and collected the internet. But the problem with all these produce a distorted range of differential
for some time and are now readily available interactions is that the advice received is of diagnoses through failing to consider
the relative prevalences of conditions
within a primary care setting. Data that
contradict expectations also tend to be
“… the major problem with most symptom checkers is ignored or dismissed as unimportant,
especially if obtained towards the end of the
that they are risk averse …“ consultation.7
Within UK general practice the earlier

324 British Journal of General Practice, July 2019

 ADDRESS FOR CORRESPONDENCE
“AI could assist GPs in recognising and overcoming Nick Summerton
Yorkshire Medical Chambers, 7 Hall Walk, Welton,
cognitive biases …” Brough, East Yorkshire HU15 1PN, UK.
Email: nsummerton@doctors.org.uk

diagnosis of cancer continues to remain In addition, AI might also point GPs

a challenge, and this might reflect the towards new diagnostic approaches, for
inappropriate use of certain heuristics. example, the analysis of an individual’s
According to Sheringham et al, GPs are not social media data history to assist in the
more likely to initiate cancer investigations earlier recognition of depression.11
for individuals with higher-risk symptoms The clinical information required to
and, also, do not investigate everyone with distinguish between, for example, two
the same symptoms equally.8 courses of action in general practice is
REFERENCES
AI could assist GPs in recognising and distinct from that required to make a 1. Summerton N. Primary care diagnostics.
overcoming cognitive biases, but only if the precise clinical diagnosis in a specialist Abingdon: Radcliffe Medical Press, 2011.
system has been trained on the correct setting. In keeping with this, it is of some 2. Academy of Medical Royal Colleges. Artificial
data, and from the appropriate population. concern that the number of tests being intelligence in healthcare. London: AoMRC,
An AI diagnostic tool developed using ordered by GPs has risen over three-fold 2019.
clinical information from the UK might not between 2000 and 2015.12 Such an increase 3. Tang A, Tam R, Cadrin-Chênevert A, et al.
provide the correct clinical guidance to a will have adverse effects on patients in Canadian Association of Radiologists White
Paper on artificial intelligence in radiology.
primary care clinician in India. Moreover, addition to healthcare finances. Therefore, Can Assoc Radiol J 2018; 69(2): 120–135.
the AI outputs need to support GP decision when faced with some common symptoms 4. Semigran HL, Linder JA, Gidengil C, Mehrotra
making where, for example, precise — such as tiredness or fatigue — it might be A. Evaluation of symptom checkers for self
diagnostic labels are less important than that AI can be harnessed in order to support diagnosis and triage: audit study. BMJ 2015;
deciding on an appropriate course of action.1 more rational approaches to testing within 351: h3480.
In enhancing a GP’s diagnostic abilities, general practice.13 5. Kroenke K, Jackson JL. Outcome in general
To conclude, AI cannot replace GP medical patients presenting with common
it is also important to appreciate that
symptoms: a prospective study with a 2-week
AI systems do not get tired or irritable! diagnostic intelligence, but could augment and a 3-month follow-up. Fam Pract 1998;
Moreover, AI might have a key role in it. However, before any large-scale roll- 15(5): 398–403.
assessing and improving GP communication out is considered, there is a requirement 6. Zakhem GA, Motosko CC, Ho RS. How should
skills.9 for AI systems to be subjected to careful artificial intelligence screen for skin cancer
More broadly, if we are really committed evaluation.14 We need to ensure that the and deliver diagnostic predictions to patients?
JAMA Dermatol 2018; 154(12): 1383–1384.
to improving the quality of primary care AI–GP combination enhances diagnostic
7. Tversky A, Kahneman D. Judgement under
cancer diagnoses, then educating new GPs decision making and health outcomes
uncertainty: heuristics and biases. Science
about cancer symptoms using retrospective without simply producing anxious patients 1974; 185(4157): 1124–1131.
studies of historical clinical data seems or inflicting a further burden on healthcare 8. Sheringham J, Sequeira R, Myles J, et al.
unlikely to have much impact on cancer budgets. Variations in GPs’ decisions to investigate
prognoses. In order to significantly improve suspected lung cancer: a factorial experiment
patient outcomes, we urgently need more Nick Summerton, using multimedia vignettes. BMJ Qual Saf
GP, Yorkshire Medical Chambers, Welton, Brough, 2017; 26(6): 449–459.
and better information on the features that East Yorkshire. 9. Ryan P, Luz S, Albert P, et al. Using
might indicate early-stage cancers within a artificial intelligence to assess clinicians’
primary care setting. communication skills. BMJ 2019; 364: l161.
Martin Cansdale,
Principal Data Scientist, Your.MD Limited, 10. Gray DJP. The role of the general practitioner
CONCLUSION Bloomsbury, London. in the early detection of malignant disease.
By its ability to analyse masses of data and to Trans Hunterian Soc 1966; 25: 135–179.
come up with new insights and associations, 11. Eichstaedt JC, Smith RJ, Merchant RM, et
Provenance al. Facebook language predicts depression
AI has the potential to furnish GPs with Freely submitted; externally peer reviewed. in medical records. Proc Natl Acad Sci U S A
some fresh ideas and understandings. 2018; 115(44): 11203–11208.
For example, it may highlight the potential 12. [No authors listed]. Temporal trends in
Competing interests
importance of other non-traditional features Nick Summerton sits on the Clinical Advisory Group use of tests in UK primary care, 2000–15:
in achieving an earlier cancer diagnosis, for Your.MD Ltd. Martin Cansdale is employed retrospective analysis of 250 million tests.
such as, measurable changes in a patient’s by Your.MD Ltd. Your.MD Ltd has developed an BMJ 2019; 364: l444.
lifestyle, or health-seeking behaviour when AI-based symptom checker. 13. Koch H, van Bokhoven MA, ter Riet G, et
al. Ordering blood tests for patients with
they consult with a non-specific symptom
unexplained fatigue in general practice:
such as fatigue.10 DOI: https://doi.org/10.3399/bjgp19X704165 what does it yield? Results of the VAMPIRE
trial. Br J Gen Pract 2009; DOI: https://doi.
org/10.3399/bjgp09X420310.
14. Fraser H, Coiera E, Wong D. Safety of patient-
facing digital symptom checkers. Lancet
2018; 392(10161): 2263–2264.

British Journal of General Practice, July 2019 325