Predictive modeling of controlled drug release from polysaccharide-based systems using gradient boosting and metaheuristic optimization

Abstract Accurate prediction of drug release kinetics from polysaccharide-based delivery systems is essential for rational formulation design. In this study, a hybrid machine learning framework integrating Raman spectroscopy with formulation descriptors is developed to model drug release profiles across different polysaccharide matrices. A dataset comprising 155 experimental instances from 13 formulation groups is used, including 1,675 Raman spectral variables, categorical medium descriptors, and temporal information. Feature selection using F-statistics reduces the spectral space to 17 informative Raman peaks, which are combined with medium and time as model inputs. Extreme Gradient Boosting (XGB) and Light Gradient Boosting (LGB) models are optimized using Swarm-Assisted Bayesian Optimization (SABO) and Quantum-Inspired Optimization (QIO), forming four hybrid predictors (XGSO, XGQO, LGSO, and LGQO). The optimized hybrid models achieve superior predictive performance compared to single learners, with XGSO and LGSO yielding the lowest prediction errors (test RMSE = 0.065 and 0.077, respectively, and R 2 = 0.961 and 0.939). SHapley Additive exPlanations (SHAP) reveal that Raman bands in the 940–990 cm –1 and 470–510 cm –1 regions, associated with glycosidic backbone vibrations of polysaccharides, exert the strongest influence on release prediction, together with time and medium effects. These results demonstrate that the proposed framework not only improves predictive accuracy but also captures chemically meaningful relationships between polymer structure and macroscopic drug release kinetics, supporting its potential for data-driven formulation optimization.