Utilizing categorical boosting and SHAP to understand key predictors of frequency of use at discharge for completed substance abuse treatments in TEDS-D.

Background: Substance use disorder is a pressing US public health crisis, with 48.7 million people reporting past-year substance use and over 100 000 overdose deaths in 2022. Building on a growing body of machine learning research using the SAMHSA treatment episode Data set-discharges (TEDS-D), this study introduces frequency of use at discharge as an outcome variable and the inclusion of U.S. state as a predictor. Research Design and Methods: First, reason for discharge was evaluated as a dependent variable and compared to frequency of use at discharge. Utilizing 2411 351 completed treatment episodes from TEDS-D 2015-2019, a categorical boosting model was employed to predict frequency of substance use at discharge using 62 predictor variables. Shapley additive explanations (SHAP) calculated global feature importance and decomposed individual predictions. State-level models were compared to the national model using rank-biased overlap (RBO), with a Hawai'i case study demonstrating state-level decision support. Results: Although treatment completion yielded higher rates of no use at discharge than dropout (45% vs 28%), 25% of completed episodes reported active substance use, 9% reported homelessness, and 50% reported unemployment at discharge. The model demonstrated strong predictive performance (AUC = 0.96, sensitivity = 0.85, specificity = 0.95). Treatment and geographic factors outweighed patient attributes, with frequency of use at admission, U.S state of treatment, length of stay, and service type as the four most important features. Among all service types, long-term residential rehabilitation was associated with the lowest SHAP values and predicted probability of substance use at discharge. State-stratified models showed moderate feature importance consistency (mean RBO = 0.59, range: 0.30-0.73). The Hawai'i case study identified elevated probability of use at discharge for younger clients and revealed substantial underreporting of long-term residential and therapeutic community treatment modalities in federal data. Discussion and Implications: State-level variability in treatment outcomes reflects systemic inconsistencies in policy, reporting, and service access that complicate national benchmarking and may obscure the true drivers of patient recovery. Explainable ML provides actionable insights for tailoring treatment and informing data-driven policy to improve outcomes.