Predicting Adverse Drug Reactions Using Machine Learning: A Hybrid Model Stacking Approach

This study leverages the SIDER dataset to predict the probability of Adverse Drug Reactions (ADRs) through a hybrid machine learning pipeline. The primary scientific objective was to engineer a stacked generalization framework capable of capturing the intricate, non-linear dependencies between drug indications and side effects—relationships often missed by single-model architectures due to high data cardinality and class imbalance. By aggregating the predictive outputs of three gradient boosting algorithms (LightGBM, XGBoost, and CatBoost) into a Random Forest meta-learner, the study aimed to systematically mitigate the variance and bias inherent in individual models, thereby maximizing predictive robustness and accuracy for pharmacovigilance.