<b>Predicting Dissolution and Gel Viscosity in Hot-Melt Extruded Polymeric Formulations: A Quality-by-Design Framework</b>

On a small, curated formulation dataset (42 records; five formulation/process features), RF provided the best held-out predictive performance (test R² = 0.485) among the baseline models, and feature-importance analysis identified pseudoephedrine drug loading as the dominant predictor. Five-fold grid-search hyperparameter optimization did not materially improve held-out accuracy on this limited dataset, and the negative cross-validated R² values reveal the high variance inherent to small samples.

Combining a second data source into a merged dataset (96 usable records; 17 features) produced a markedly stronger optimized RF model (R² = 0.984, RMSE = 3.56 %, MAE = 2.17 %, MAPE = 7.54 %). This performance is, however, partly attributable to the inclusion of the 60-minute dissolution endpoint as a predictive feature, which introduces target leakage and must therefore be interpreted cautiously. Shapley Additive Explanations (SHAP) showed that drug content uniformity (% RSD) and the 60-minute dissolution readout dominated the combined model, while residual analysis confirmed near-zero bias with no pronounced heteroscedasticity.

Collectively, these results demonstrate a reusable, interpretable, quality-by-design (QbD) workflow for surrogate modeling of HME product quality and highlight the critical roles of data volume, feature-leakage control, and rigorous cross-validation in translating ML from proof of concept to deployment in pharmaceutical process engineering.</p>