{"entity":{"id":"idea-data-ai-screening-endpoints","kind":"idea","name":"Require stage-shift or interval-cancer endpoints for AI in cancer screening","aka":[],"tldr":"AI for screening should be judged on whether it finds dangerous cancers earlier and misses fewer, not just on whether it agrees with radiologists on old images.","summary":"AI in mammography, lung CT and colonoscopy is evaluated on retrospective detection metrics that reward finding more lesions regardless of clinical significance, which risks overdiagnosis. The proposal requires, for adoption in organised screening programmes, evidence on interval cancer rates, stage distribution of detected cancers and recall rates from prospective studies (randomised or well-designed stepped implementations), with post-implementation monitoring of the same endpoints via registry linkage.","asOf":"2026-09-08","links":[{"label":"Bottleneck evidence (AI that is built but not validated or deployed): Wu et al., How medical AI devices are evaluated: limitations and recommendations from an analysis of FDA approvals (Nature Medicine 2021)","url":"https://doi.org/10.1038/s41591-021-01312-x"}],"tags":[],"related":[],"cancers":[],"sections":["ai-computation","early-detection"],"technologies":["radiology-ai-screening","mammography","low-dose-ct-screening"],"targets":[],"drugs":[],"companies":[],"institutions":[],"pathways":[],"terms":["stage-shift"],"trials":[],"people":[],"bottlenecks":["b-ai-validation","b-overdiagnosis","b-early-detection"],"keyPapers":["paper-wu-nat-med"],"journals":[],"dependsOn":[],"notes":[],"hypothesis":"Judged on interval cancers and stage shift, some AI tools with strong retrospective performance will show no benefit or increased overdiagnosis, while others will reduce interval cancers, and the endpoint requirement will steer development toward the latter.","rationale":"Screening's history (PSA, thyroid ultrasound) shows that detecting more is not the same as helping; MASAI and similar trials show the correct endpoints are measurable within a programme.","test":"Adopt the endpoint requirement in one national screening programme; evaluate two AI tools via stepped implementation with registry-linked interval cancer follow-up over three years.","maturity":"early-clinical","actor":"regulator","cost":"medium","horizonYears":4},"route":"/ideas/idea-data-ai-screening-endpoints/","neighbours":{"section":[{"id":"ai-computation","kind":"section","name":"AI & Computation","route":"/fronts/ai-computation/"},{"id":"early-detection","kind":"section","name":"Early Detection & Screening","route":"/fronts/early-detection/"}],"technology":[{"id":"radiology-ai-screening","kind":"technology","name":"AI in radiology","route":"/technologies/radiology-ai-screening/"},{"id":"low-dose-ct-screening","kind":"technology","name":"Low-dose CT lung screening","route":"/technologies/low-dose-ct-screening/"},{"id":"mammography","kind":"technology","name":"Mammography & tomosynthesis","route":"/technologies/mammography/"}],"term":[{"id":"stage-shift","kind":"term","name":"Stage shift","route":"/terms/stage-shift/"}],"bottleneck":[{"id":"b-ai-validation","kind":"bottleneck","name":"AI that is built but not validated or deployed","route":"/bottlenecks/b-ai-validation/"},{"id":"b-overdiagnosis","kind":"bottleneck","name":"Overdiagnosis and false alarms","route":"/bottlenecks/b-overdiagnosis/"},{"id":"b-early-detection","kind":"bottleneck","name":"The hardest cancers are found late","route":"/bottlenecks/b-early-detection/"}],"paper":[{"id":"paper-wu-nat-med","kind":"paper","name":"How medical AI devices are evaluated: limitations and recommendations from an analysis of FDA approvals","route":"/key-papers/paper-wu-nat-med/"},{"id":"paper-masai-lancet-oncol-2023","kind":"paper","name":"MASAI: AI-supported mammography screening finds more cancers with half the radiologist workload","route":"/key-papers/paper-masai-lancet-oncol-2023/"}]}}