OnCo
ideasIdea

Judge skin cancer AI by the thick melanomas it prevents, not the thin ones it finds

Melanoma diagnoses have soared while deaths barely changed, a sign of overdiagnosis. AI skin apps should be judged on whether dangerous thick melanomas fall, not how many spots they flag.

Melanoma incidence has risen roughly six-fold in the US since 1975 with much smaller mortality change. AI dermatology tools risk amplifying detection of indolent in-situ lesions. Propose that regulatory and reimbursement evaluation of AI skin triage require thick (over 1 mm) melanoma incidence and biopsy-to-melanoma ratio as endpoints.

Hypothesis
AI triage tools judged by detection volume increase in-situ diagnoses and biopsies without reducing thick melanoma incidence; tools optimised against thick-melanoma endpoints adopt more conservative operating points.
Rationale
The metric determines the algorithm's threshold.
What would test it
Cluster RCT of AI triage in primary care with three-year thick melanoma incidence and biopsy counts.
Maturity
speculative
Who has to act
regulator
Cost to try
Medium ($1M to $50M)
Years to first evidence
4
Bottlenecks it attacks

Connected

4top