Integrative ensemble learning framework for forecasting controlled drug release based on Raman spectral signatures.

Modeling drug-release kinetics from polysaccharide-coated oral controlled-release formulations remains challenging due to nonlinear diffusion–dissolution behavior, complex polymer–drug interactions, and the limited interpretability of conventional machine-learning approaches. In this study, we develop and validate a predictive framework for targeted colonic delivery of 5-aminosalicylic acid (5-ASA) from polysaccharide-coated solid oral dosage forms using Raman spectroscopy–derived molecular fingerprints and time-resolved dissolution data. The dataset comprises 155 formulation samples, each characterized by more than 1,500 Raman spectral features, categorical formulation variables (polysaccharide type and release medium), and drug-release measurements at 2, 8, and 24 h collected under simulated physiological conditions. A dual-optimizer, dual-ensemble learning strategy is introduced, integrating the Puma Optimizer Algorithm (POA) and Black-Winged Kite Algorithm (BWKA) within a Damsphere Weighted Ensemble (DWE) of XGBoost regression and AdaBoost models. The complementary exploration–exploitation dynamics of the two optimizers enhance convergence stability and generalization, yielding strong predictive performance under five-fold cross-validation (RMSE = 0.038; R2 = 0.991). Feature-level analysis based on F-statistics highlights release time, dissolution medium, and chemically meaningful Raman bands as dominant predictors, consistent with diffusion- and erosion-controlled release mechanisms in polysaccharide-coated systems. From a pharmaceutical perspective, the proposed framework reduces experimental burden while maintaining mechanistic interpretability, supporting Quality by Design (QbD) and green pharmaceutics principles. Owing to its modular architecture, the approach is readily extensible to other polymer-based oral controlled-release formulations and spectroscopic modalities.