Development and internal validation of an interpretable mammography-based model for malignancy risk stratification in BI-RADS 4 breast lesions

Background BI-RADS category 4 breast lesions are suspicious and generally undergo tissue diagnosis, but malignancy risk remains heterogeneous across BI-RADS 4A, 4B, and 4C. We developed and internally validated an interpretable mammography-based model for risk stratification within this diagnostic category. Methods This single-center retrospective diagnostic prediction study included 441 patients, each contributing one index BI-RADS 4 mammographic lesion. Outcomes were verified histopathologically in 397 lesions and by longitudinal imaging stability in 44. Three primary models and a sensitivity model excluding BI-RADS were evaluated. Internal validation used 1,000 paired patient-level bootstrap resamples in which imputation, encoding, standardization, penalty selection, LASSO selection, and coefficient estimation were repeated. Results Of 441 lesions, 165 were malignant. Malignancy rates were 15.7% (34/216), 46.4% (70/151), and 82.4% (61/74) in BI-RADS 4A, 4B, and 4C, respectively. The final model retained BI-RADS 4B, BI-RADS 4C, architectural distortion, circularity-deficit irregularity score, contour serration score, and normalized GLCM contrast. Optimism-corrected AUCs were 0.770 (95% CI, 0.729-0.815) for BI-RADS only, 0.774 (0.733-0.814) for clinical plus BI-RADS, and 0.841 (0.809-0.877) for the full mammography-based model. Clinical variables did not improve validated performance. The full model improved corrected AUC versus BI-RADS only by 0.072 (0.028-0.108) and versus clinical plus BI-RADS by 0.067 (0.029-0.103). Its corrected calibration intercept was 0.002 (-0.209 to 0.208) and slope was 1.445 (1.140-1.753). Within-category AUCs were 0.785, 0.721, and 0.758 for BI-RADS 4A, 4B, and 4C; the pathology-only corrected AUC was 0.835 (0.799-0.871). Conclusion Structured and quantitative mammographic features improved internally validated discrimination and within-category risk stratification beyond BI-RADS subcategory alone. Because predicted risks were under-dispersed and validation was limited to one center, the model should be used only as a research-stage risk-stratification and prioritization tool, not to replace biopsy. External validation and recalibration are required before clinical implementation.