Data driven analysis of tablet design via machine learning for evaluation of impact of formulations properties on the disintegration time

Abstract

This study investigated the use of advanced machine learning techniques to model disintegration time for solid dosage oral formulations. The input features encompass molecular properties, physical attributes, excipient compositions, and formulation characteristics. An Isolation Forest algorithm is employed for outlier detection, while a Standard Scaler normalization technique ensures consistent feature scaling. Feature selection is conducted using Conditional Mutual Information (CMI) to identify the most informative predictors. The study compares three regression models: Local Polynomial Regression (LPR), Gaussian Process Regression (GPR), and Deep Gaussian Process Regression (DGPR). Results indicate that DGPR outperforms others, obtaining the highest R² scores and the lowest error rates across training, validation, and testing phases. Interpretability of the DGPR model is enhanced using SHAP (SHapley Additive exPlanations), providing insights into feature importance and their effects on predictions. Additionally, the Hunter-Prey Optimization algorithm is utilized to optimize hyperparameters, demonstrating its efficacy in balancing exploration and exploitation. This research shows that DGPR can accurately model complex relationships in pharmaceutical datasets and provides both predictive accuracy and interpretability.