UCLA Review: AI Flags 5-78% of Interval Breast Cancers Retrospectively but Lacks Prospective Outcome Data

The UCLA-led analysis examined two AI applications for screening mammography: detection of retrospectively visible interval cancers and identification of pre-visible risk patterns. Interval cancers, diagnosed after negative mammograms, often prove more aggressive. Retrospective datasets showed AI assigning elevated scores to 23.1% of future interval cases three rounds prior and 39.4% immediately before diagnosis, yet these remain post-hoc observations on archived images rather than real-time workflow changes.

Original coverage understated the gap between retrospective detection and clinical utility. Most included studies lacked control arms or blinding, and none reported reductions in advanced-stage diagnoses or mortality. A 2023 Lancet Digital Health trial of AI-supported screening similarly found higher detection rates but no difference in interval cancers at two-year follow-up, highlighting the need for randomized designs that track downstream outcomes rather than surrogate accuracy metrics.

Next steps require prospective cluster-randomized trials embedding AI into routine double-reading workflows with prespecified endpoints of interval cancer rate, stage shift, and biopsy rates. Ongoing European studies with 100,000-plus participants will report in 2027-2028; until then, claims of immediate life-saving impact remain unsupported by level-1 evidence.