Multi-Institutional Validation of a Mammography-Based Breast Cancer Risk Model.
Level 3 - non-randomized controlled study
Retrospective multi-institutional prognostic validation cohort study
PubMed 34767469 · doi:10.1200/JCO.21.01337
What was done
Researchers validated Mirai, an artificial intelligence-based breast cancer risk prediction model, across seven screening populations across five countries (MGH, Novant, and Emory in the USA; Maccabi-Assuta in Israel; Karolinska in Sweden; Chang Gung Memorial Hospital in Taiwan; and Barretos in Brazil). They collected screening mammograms and pathology-confirmed breast cancer diagnoses and evaluated performance using Uno's concordance index (C-index) for predicting cancer risk 1 to 5 years post-mammogram.
What was found
The dataset included 128,793 mammograms from 62,185 patients, with 3,815 incident cancer diagnoses within 5 years. Mirai achieved the following C-indices across sites: MGH, 0.75 (95% CI, 0.72 to 0.78); Novant, 0.75 (95% CI, 0.70 to 0.80); Emory, 0.77 (95% CI, 0.75 to 0.79); Maccabi-Assuta, 0.77 (95% CI, 0.73 to 0.81); Karolinska, 0.81 (95% CI, 0.79 to 0.82); Chang Gung, 0.79 (95% CI, 0.76 to 0.83); and Barretos, 0.84 (95% CI, 0.81 to 0.88).
Why it matters
This multi-national validation demonstrates that a single deep-learning mammography risk model generalizes effectively across distinct geographic, demographic, and institutional imaging datasets without substantial performance degradation.
Limits
The study relies on retrospective screening cohorts. The abstract does not provide head-to-head performance comparisons against standard clinical risk tools (e.g., Tyrer-Cuzick or BCRAT), calibration metrics, subgroup breakdowns by breast density or race, or clinical impact on screening outcomes.
Cited by
- context AI models can detect certain tumors like breast cancer years before a radiologist can see them.