Are artificial intelligence/machine learning (AI/ML) algorithms ready for implementation in community and primary care settings to facilitate the early detection of skin cancer? A systematic review.

Abstract

Background Skin cancers occur very commonly worldwide. Prognosis and disease burden are highly dependent on cancer type and disease stage at diagnosis. We systematically reviewed AI/ML algorithms aiming to facilitate early diagnosis of skin cancers, focusing on their application in primary care.

Methods We searched Medline, Embase, SCOPUS, and Web of Science (01/01/2000-9/08/2021), including all studies providing evidence on applying AI/ML algorithms to the early diagnosis of skin cancer, all study designs and languages. Primary outcome was diagnostic accuracy for skin cancers. Secondary outcomes included: AI/ML methods, evaluation approach, cost-effectiveness, and acceptability.

Findings We identified 14,224 studies. Only 2 studies used data from low prevalence settings, so we report data from all 272 studies that could have relevance in primary care. Primary outcomes showed reasonable mean diagnostic accuracy: melanoma (89.5% (range 59.7-100%)), keratinocyte carcinomas (86.7% (70.0-99.7%)). Secondary outcomes demonstrated heterogeneity of AI/ML modalities and study designs, with high levels of incomplete reporting.

Interpretation Few studies used low prevalence population data to train and test their algorithms, therefore widespread adoption into community and primary care practice cannot currently be recommended. We propose a methodological checklist for use in development of new AI/ML algorithms to detect skin cancer, to facilitate their design, evaluation, and implementation.