Prediction of tablet disintegration time based on formulations properties via artificial intelligence by comparing machine learning models and validation

This research assesses multiple predictive models aimed at estimating disintegration time for pharmaceutical oral formulations, based on a dataset comprising nearly 2,000 data points that include molecular, physical, compositional, and formulation attributes. Drug and formulation properties were considered as the inputs to estimate the output which is tablet disintegration time. Advanced machine learning methods, including Bayesian Ridge Regression (BRR), Relevance Vector Machine (RVM), and Sparse Bayesian Learning (SBL) were utilized after comprehensive preprocessing involving outlier detection, normalization, and feature selection. Grey Wolf Optimization (GWO) was utilized for model optimization to obtain optimal combinations of hyper-parameters. Among the models, SBL stood out for its superior performance, achieving the highest R² scores and the lowest Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE) error rates in both the training and testing phases. It also demonstrated robustness and effectively avoided overfitting. SHapley Additive exPlanations (SHAP) analysis provided valuable insights into feature contributions, highlighting wetting time and sodium saccharin as key factors influencing disintegration time.