Advancing Predictive Modeling in Behavioral Health with Gradient Boosted Bootstrap Model

This study advanced predictive modeling in behavioral health by developing and evaluating a Variance Adaptive Gradient Boosted Bootstrap Model (VAGBBM) designed to improve stability, calibration, and interpretability in heterogeneous clinical data. Administrative records from 1,983 adults enrolled in a community-based substance use disorder treatment program were analyzed. The proposed model integrated bootstrap resampling with gradient boosting to address bias variance tradeoffs common in high dimensional observational settings. Predictive performance was evaluated on a held-out test set using standard discrimination and calibration metrics. Model interpretability was assessed using SHapley Additive exPlanations (SHAP) to generate both global and individual level explanations. The model demonstrated strong predictive performance, with excellent discrimination (area under the receiver operating characteristic curve = 0.92) and good calibration (Brier score = 0.09). Explainability analyses suggested that behavioral engagement indicators, including attendance patterns and client-initiated disengagement, were substantially more predictive of treatment completion than demographic, socioeconomic, or psychiatric characteristics. Individual level explanations suggested how dynamic engagement processes shaped predicted risk, supporting clinically actionable interpretation. These findings emphasize the value of modeling treatment retention as a dynamic behavioral process rather than relying solely on static baseline characteristics. Methodologically, the VAGBBM provides a transparent and robust framework for predictive modeling in behavioral health research. Substantively, results emphasize opportunities for precision intervention strategies that target early disengagement to improve treatment retention and equity in care delivery.