Exploring osteoporosis risk in breast cancer patients after comprehensive treatment via explainable artificial intelligence algorithms.

BACKGROUND: Osteoporosis is a common complication among long-term breast cancer (BC) survivors. Assessment of osteoporosis risk of patients before therapy is essential for timely intervention and long-term health management. METHODS: We included female BC patients treated between 2018 and 2022 at a tertiary hospital in China. Baseline characteristics, hematologic parameters, and pathological subtypes were collected and stratified into four molecular groups based on hormone receptor and HER2 status. Twelve machine learning (ML) algorithms were developed, and the best-performing models were interpreted using SHAP analysis. RESULTS: Among 1314 patients, 546 (41.6%) were diagnosed with osteoporosis, predominantly in HR-positive subtypes. Random Forest achieved the best performance in HR-/HER2-patients (validation AUC = 0.773), with platelet count, LDH, age, γ-GTP, uric acid, and globulin as key predictors. In HR-/HER2+ patients, Gradient Boosting performed optimally (AUC = 0.942), with hematologic and metabolic markers as major contributors. For HR+/HER2-and HR+/HER2+ groups, Boosting models reached AUCs of 0.832 and 0.771, respectively, with nodal stage, age, platelet count, uric acid, and liver/renal indices as leading predictors. SHAP dependence plots revealed critical interactions, such as age with platelet count and nodal stage with bilirubin. CONCLUSIONS: Osteoporosis risk varies substantially across molecular subtypes of BC and is shaped by both clinical and biochemical factors. ML combined with explainable AI provides accurate prediction and highlights key risk determinants, offering a potential evaluated tool for personalized bone health management in BC survivors.