Development of machine learning models for estimation of disintegration time on fast-disintegrating tablets

Abstract

The disintegration time for solid dosage oral formulations is directly influenced by diverse factors such as molecular properties, physical characteristics, excipient compositions, and formulation-specific attributes. This research addresses the challenge of predicting this parameter by applying advanced machine learning techniques to model disintegration behavior, with improved reliability achieved through Z-score normalization and outlier removal during data preprocessing. The data was used for the fast-disintegrating tablets (FDT) to assess the effects of underlying parameters on the tablet disintegration time. The selected models—Multi-Task Lasso (MTL), Elastic Net (EN), and a stacking ensemble—were chosen to balance feature selection, multicollinearity handling, and predictive performance. The stacking ensemble, which combines the outputs of MTL and EN through a meta-regressor, effectively leverages their complementary strengths, resulting in superior accuracy and robustness. Hyperparameter tuning was performed using the Firefly Optimization Algorithm (FFA), a bio-inspired optimization technique known for its efficiency in navigating high-dimensional search spaces. This ensured optimal model performance and reduced the risk of overfitting, leading to a solution capable of generalizing across various data subsets. Key findings include the identification of the top 10 most influential features, with wetting time emerging as a primary determinant of disintegration behavior. This study reports a new framework that combines machine learning models with advanced optimization techniques for accurate disintegration time prediction of pharmaceutical tablets. Besides increasing the predictive value of the model, the framework also provides valuable understanding of the most influential factors that influence disintegration.