Machine learning and cox model-based prediction of CDK4/6 inhibitor outcomes in HR+/HER2 - metastatic breast cancer: a multicenter real-world study.

BACKGROUND: Cyclin-dependent kinase 4/6 inhibitors (CDK4/6is) are standard therapy for HR+/HER2 - metastatic breast cancer (MBC), yet treatment outcomes vary substantially and individualized prognostic tools based on real-world data remain limited. This study evaluated real-world effectiveness and developed prognostic models integrating Cox regression and machine learning. METHODS: This multicenter retrospective study included 1,008 HR+/HER2 - MBC patients treated with CDK4/6is across 20 cancer centers in central China. Treatment patterns were analyzed in the overall cohort. PFS and prognostic factors were evaluated in patients receiving first- or second-line CDK4/6is using Kaplan-Meier and Cox regression analyses. These patients were randomly divided into training and validation cohorts (7:3). A Cox model and seven machine learning algorithms (GBM, RSF, Lasso-Cox, CoxBoost, XGBoost, SuperPC, and plsRcox) were developed and compared using time-dependent AUC, calibration, and decision curve analysis. RESULTS: CDK4/6is were used as first- and second-line therapy in 65.68% and 24.60% of patients. Median PFS was significantly longer in first-line versus second-line treatment (38.0 vs. 18.8 months, P < 0.001). Independent prognostic factors included Luminal B subtype, liver metastasis, and second-line treatment, while HER2 IHC 1 + and longer disease-free interval were favorable predictors. The Cox model showed good discrimination (AUCs: 0.731, 0.719, and 0.704). Among machine learning models, GBM and RSF showed higher discrimination in the training cohort but only moderate performance in the validation cohort. CONCLUSION: CDK4/6is demonstrated substantial real-world effectiveness. Both Cox and machine learning models enabled individualized prognostic prediction. GBM and RSF showed relatively better predictive performance, but external validation is still required before clinical application.