EXPLAINABLE ENSEMBLE MACHINE LEARNING FRAMEWORK FOR PREDICTING DRUG RELEASE BEHAVIOR IN POLYMERIC LONG-ACTING INJECTABLE SYSTEMS

Abstract

Predicting drug release from polymeric long-acting injectable systems remains challenging because release behavior depends on nonlinear interactions among formulation, physicochemical, and temporal variables. This study developed an explainable ensemble machine learning framework for predicting drug release behavior and identifying the most influential release-governing variables. A secondary experimental dataset containing 3,783 observations and 20 variables was analyzed. Data preprocessing included quality inspection, categorical encoding, exclusion of identifier variables, and train-test splitting. Linear Regression, Random Forest, and XGBoost models were developed and evaluated using R², mean absolute error, and root mean square error. Feature importance and residual analyses were additionally performed to assess interpretability and model reliability. XGBoost achieved the best predictive performance with an R² of 0.9849, MAE of 0.0276, and RMSE of 0.0406, outperforming Random Forest and Linear Regression. The strongest predictors were T = 1.0, Drug NHA, SE, Drug Mw, and Time. Residual analysis demonstrated low prediction bias and stable error distribution, supporting the robustness of the optimized model. The findings indicate that explainable ensemble learning can accurately model complex release behavior in polymeric long-acting injectable systems and may support efficient formulation screening, optimization, and data-driven pharmaceutical engineering.