Age-stratified evidence gates for a frozen mammography ensemble in the CMMD diagnostic cohort
Objective. To measure how independent age-group evidence gates change the below-threshold fraction of a frozen mammography classifier in a selected diagnostic cohort. Methods. We selected one pathology-annotated breast per CMMD subject aged 40–74, froze participant-disjoint 20% tuning, 30% certification and 50% evaluation partitions, and applied the five released GMIC members without training. Thresholds were chosen on tuning only. The primary 5% malignant-miss cap used one-sided Clopper–Pearson bounds with family error 0.05 over two age groups. Results. Of 1,359 intended participants, 1,355 had valid two-view scores; 4 retained review because of imaging or linkage failures. The evaluation partition contained 683 participants, including 529 malignant observations. Valid-score AUROC was 0.783 (stratified bootstrap 95% interval 0.747–0.818). The ungated global policy assigned 57/683 (8.3%) below threshold and assigned 32/529 malignant observations below threshold (6.0%). Its false-omission fraction was 56.1%. The older certification group contained 62 malignancies and 0 errors; its upper bound was 5.78%. The primary subgroup gate failed and assigned 0 observations below threshold. Conclusion. The locked global candidate and the subgroup gate therefore behave differently in this cohort. This single-cohort diagnostic audit supports neither a screening-workload estimate nor a nontrivial age-group safety certificate.
Authors
- S. M. Abdulla Al Mamun
Institutions
- University of Manchester (GB)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24369294
- Primary Topic
- AI in cancer detection
- Type
- preprint