Taguchi–machine learning hybrid framework for optimization of particulate drug delivery systems

Abstract

Optimizing particulate drug carriers requires balancing multiple formulation parameters to achieve target physicochemical properties while minimizing experimental burden. In this study, a hybrid optimization framework integrating a Taguchi orthogonal array (OA) design with statistical modeling and machine-learning–based interpretability is implemented and demonstrated using doxorubicin-loaded chitosan microspheres (DOX–CS MSs) as a model system. Microspheres were produced via a water-in-oil emulsion crosslinking method using an acidified chitosan solution and mineral oil, hardened by glutaraldehyde-mediated crosslinking and characterized by FTIR, XRD, and FESEM to confirm chemical structure, crystallinity, and spherical morphology. The optimization targeted a particle size of 5–7 μm and encapsulation efficiency (EE) exceeding 90 %. An initial L9 Taguchi OA design efficiently narrowed the formulation space by varying chitosan concentration (1–3 % w/v), glutaraldehyde concentration (1.5–5 % v/v), and crosslinking time (3–5 h), yielding nine core formulations. Pearson/Spearman correlation, second-order polynomial regression (Poly2), and Gradient Boosting Machine (GBM) models applied on experimental data quantified parameter influences and predicted performance. SHapley Additive exPlanations (SHAP) analysis identified chitosan concentration as the primary determinant of both size and EE, with glutaraldehyde content exerting secondary, synergistic effects. Poly2 response-surface modeling achieved high predictive accuracy (R2 = 0.983 for size; R2 = 0.986 for EE) and yielded explicit regression equations for real-time formulation targeting. This hybrid Taguchi–ML approach enables rapid factor prioritization, reveals nonlinear interactions overlooked by conventional Taguchi analysis, and offers transparent ML interpretability. Beyond chitosan-based carriers, proposed framework offers a generalizable and scalable route to rational formulation design in complex particulate systems for targeted biomedical applications. © 2025 The Authors.