Predictors of guideline-concordant treatment intensification in metastatic castration-sensitive prostate cancer: an explainable machine learning study

Background Treatment intensification with androgen deprivation therapy (ADT) combined with docetaxel, an androgen receptor pathway inhibitor, or selected triplet regimens is recommended for eligible patients with metastatic castration-sensitive prostate cancer (mCSPC). Despite increasing adoption, implementation remains variable in routine clinical practice. Objective To characterize temporal trends in first-line systemic treatment for mCSPC and identify factors associated with receipt of guideline-concordant treatment intensification using exploratory explainable machine learning. Methods National treatment patterns were evaluated using the Pfizer Oncology Quality Care Insight Tool derived from IQVIA Anonymized Patient Longitudinal Data, and institutional patterns were evaluated within Avera Health. The exploratory cohort included 42 patients with documented mCSPC treated between 2018 and 2023. Guideline-concordant treatment intensification was defined as ADT combined with docetaxel or a novel hormonal therapy within 90 days of ADT initiation. Potential outcome-derived variables were excluded before modeling. Six predictive approaches were evaluated using repeated nested stratified cross-validation, with 5-fold outer cross-validation repeated 20 times and 5-fold inner cross-validation repeated three times. Recursive feature elimination was performed within outer training folds, and class imbalance was addressed using SMOTE within resampling. Results Thirty-two (32) of forty-two (42) patients (76.2%) received guideline-concordant treatment intensification. Treatment-pattern analysis showed increasing use of first-line novel hormonal therapy and declining ADT monotherapy over time. Under repeated nested validation, discrimination was modest and variable across models. The Decision Tree achieved the highest patient-level AUC (0.675; 95% CI, 0.421–0.929), whereas Random Forest achieved an AUC of 0.650 (95% CI, 0.383–0.917), accuracy of 85.7%, precision of 88.2%, recall of 93.8%, and F1 score of 0.909. Feature-selection stability identified insurance coverage for novel hormonal therapy as the most consistently selected predictor, followed by positive core proportion and smoking history. These three variables were retained in the final explanatory Random Forest model, and global SHAP and Random Forest importance analysis showed broadly concordant rankings. Conclusions Treatment intensification for mCSPC increased substantially in routine practice. Repeated nested validation produced more conservative and variable performance estimates than the original single-split analysis, underscoring the exploratory nature of this small cohort. Healthcare access and selected clinical factors warrant further evaluation in larger, multicenter datasets.