JMIR Publications

ABSTRACT

Background:

Background:

Artificial intelligence (AI) imaging tools are entering national cancer screening programmes, where a tool that performs unevenly across population groups will distribute benefit and harm unevenly at scale. Diagnostic accuracy for these tools is now well described. Whether the evidence base permits any assessment of equity is not.

Objective:

Objective:

To determine how completely studies of AI-based cancer screening imaging report the population composition of their cohorts and the performance of their tools across demographic subgroups, and to identify where in the evidence base equity assessment is currently impossible.

Methods:

Methods:

Cross-sectional study of 51 published reports of AI-based imaging for cancer screening, surveillance or diagnostic workup, drawn from PubMed/MEDLINE and published between 2020 and 2026 across 21 countries. Studies were eligible if they reported at least one diagnostic accuracy metric against a reference standard. Each was classified on a three-level population-composition scale (race or ethnicity not reported; a single national or ancestry group named; a quantified group breakdown), on whether race or ethnicity was evaluated as a performance subgroup, and on socioeconomic and rural/urban reporting. The sample is 1% of an eligible frame of 5,169 PubMed records assembled with a validated high-recall strategy, and was benchmarked against that frame on publication year and cancer site.

Results:

Results:

Of 51 studies, 37 (73%) did not report their cohort’s racial or ethnic composition in any form, 7 (14%) named a single national or ancestry group with no within-cohort breakdown, and 7 (14%) reported a quantified breakdown. Those same 7 studies were the only ones evaluating race or ethnicity as a performance subgroup. All were conducted in the United States (5) or United Kingdom (2); none of the remaining 19 countries contributed one, including every low- and middle-income setting. The 22 studies covering cervical, oral and skin, colorectal and gastro-oesophageal cancers reported none. A socioeconomic proxy appeared in 6 studies (12%), rural/urban status in 20 (39%), any demographic subgroup metric in 14 (27%) and external validation in 25 (49%).

Conclusions:

Conclusions:

Across these 51 studies, equity assessment is possible only for breast, lung and prostate imaging in two high-income countries. This is a reporting failure rather than a finding of fairness: 73% of studies provide no population composition at all, so bias cannot be ruled out or confirmed. Reporting the cohort’s demographic composition, and accuracy stratified by it, should be a precondition for publishing an AI screening tool and for procuring one.