Can artificial intelligence/machine learning (AI/ML) techniques aid the early detection of skin cancer in primary care settings?

Conference:

SAPC ASM 2021 - virtual

Talk Code:

P1.10.5

Presenter:

Owain Jones

Co-authors:

K Prathivadi(2), CKI Ranmuthu(2), MS Islam(2), N Calanzani(1), M van der Schaar(3,4), J Emery(5,1), RN Matin(6), H Williams(7), FM Walter(1,8)

Author institutions:

Department of Public Health & Primary Care University of Cambridge, Department of Applied Mathematics and Theoretical Physics University of Cambridge, Centre for Cancer Research University of Melbourne, Department of Dermatology Churchill Hospital Oxford, Centre of Evidence Based Dermatology University of Nottingham, Wolfson Institute for Preventive Medicine Queen Mary University of London

Problem

Most people who are concerned about a skin lesion first present in primary care, where primary care clinicians need to distinguish rare melanomas and other skin cancers from common benign lesions. There has been a recent boom in the application of AI/ML techniques in medicine, including in diagnosis of skin cancer. There is some evidence that these technologies can match the diagnostic performance of experienced dermatologists. If AI/ML techniques could reproduce this level of performance in primary care settings they could aid GPs in the triage of suspicious skin lesions, potentially leading to earlier diagnosis, improved outcomes for patients, and reduced burden on overstretched secondary care services.

Approach

Our aim was to identify AI/ML technologies that could be used in primary care settings to aid detection of skin cancer. We performed a systematic review, extended through a scoping review of commercially available AI technologies. We searched Medline, Embase, SCOPUS, and Web of Science bibliographic databases from 01/01/2000 to 25/08/2020, looking for primary research of any study type, in any language, that provided evidence on the accuracy of AI/ML technologies in the assessment of skin cancer, their potential for implementation in primary care settings, and barriers to implementation.

Findings

10,456 studies were identified; 198 met inclusion criteria. Two thirds of studies used neural network methodology. The study populations varied: while none used primary care patients or images alone, 196 studies used secondary care data, and 2 used a mixture of primary and secondary care data. Half the studies included images from the International Skin Imaging Collaboration image database, in publications from 2016 onwards. Database size ranged from 100 to 131,873 images. Only 6 studies performed validation of their AI/ML technique in an independent dataset, and only 2 studies were prospective. We found no data on implementation barriers or cost-effectiveness. Risk-of-bias assessment highlighted a wide range in study quality. Marked heterogeneity between study design and outcomes measures made meta-analysis unfeasible.

Consequences

AI/ML techniques applied to the triage of skin lesions have great potential to support the early detection of skin cancer in primary care, potentially leading to improved outcomes for patients, improved survival, and reduced burden on secondary care services. However, there is a notable absence of primary care data in the development and validation of these technologies, meaning we are unable to comment on their accuracy and suitability for use in primary care settings at the current time. There is also a lack of evidence on cost-effectiveness, acceptability to patients and clinicians, and implementation barriers. Further research is required to build the evidence base and ensure these technologies are safe and effective enough for implementation into clinical practice.

Submitted by:

Owain Jones

Funding acknowledgement:

This research was commissioned and funded by the National Institute for Health Research (NIHR) Policy Research Programme, conducted through the Policy Research Unit in Cancer Awareness, Screening and Early Diagnosis, PR-PRU-1217-21601. The views expressed are those of the author(s) and not necessarily those of the NIHR, the Department of Health and Social Care. This work was also supported by the CanTest Collaborative (funded by Cancer Research UK C8640/A23385) of which FMW is a Director, and JE is an Associate Director. OTJ is supported by a Cancer Research UK Cambridge Centre clinical research fellowship. The views expressed in this publication are those of the authors and not necessarily those of the National Health Service, the NIHR or the Department of Health. The funding sources had no role in the study design, data collection, data analysis, data interpretation, writing of the report or in the decision to submit for publication.