Explainable machine learning for breast mass characterization and malignancy risk stratification: multimodal integration of AI-derived structured digital breast tomosynthesis features and peripheral blood immune-inflammatory biomarkers

Abstract

Background Accurate discrimination between benign and malignant breast masses is essential for biopsy decisions and individualized management. This study aimed to develop an interpretable machine-learning model integrating artificial intelligence (AI)-structured digital breast tomosynthesis (DBT) mass features with peripheral blood immune-inflammatory indices. Methods This retrospective study included 382 patients with 401 pathologically confirmed breast mass lesions (299 benign and 102 malignant) who underwent DBT before pathological examination. AI-structured DBT descriptors, clinical variables, and hematologic inflammatory indices were collected. Lesions were divided into training and testing cohorts using stratified random sampling at a 7:3 ratio. Least absolute shrinkage and selection operator (LASSO) logistic regression was used for feature selection, and five machine-learning models were trained. Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC), calibration, decision curve analysis, Brier score, and confusion matrix. SHapley Additive exPlanations (SHAP) analysis was used to interpret the final model, and predicted probabilities were used for descriptive risk stratification. Results Logistic LASSO regression selected nine predictors: suspicious calcifications, irregular mass shape, age, long-axis diameter, neutrophil-to-lymphocyte ratio (NLR), short-axis diameter, oval mass shape, absence of calcifications, and obscured margin. In the testing cohort, the logistic regression model achieved the highest AUROC (0.884, 95% CI: 0.807–0.950), with an accuracy of 0.860, sensitivity of 0.806, specificity of 0.878, and negative predictive value of 0.929. SHAP analysis results indicate that suspicious calcifications, advanced age, irregular morphology, elevated NLR, and larger lesion size may be the primary factors contributing to the prediction of malignancy. The malignancy rate observed in this study ranged from 2.5% in the low-risk group to 67.6% in the high-risk group. Conclusion The logistic regression (LR) model integrating AI-derived structured DBT features, clinical variables, and peripheral blood immune-inflammatory biomarkers demonstrated favorable performance for breast mass classification and malignancy risk stratification, with a test-set AUROC of 0.884. Prospective multicentre external validation is required.