Machine Learning for Predictive Risk Evaluation in Drug Substitution: A Data-Driven Approach

This study investigates the potential risks associated with substitute drugs compared to their original counterparts by leveraging advanced machine learning (ML) and deep learning (DL) techniques. Utilizing a comprehensive dataset comprising over 250,000 medication records enriched with a specially constructed side effect risk scoring system, we performed extensive data preprocessing and feature engineering. Key engineered features include binary encodings of the top side effects, statistical risk summaries of alternative medications, and one-hot encodings of chemical and action classes. A diverse set of classical ML models-including Logistic Regression, Random Forest, XG-Boost, LightGBM, and Support Vector Classifier-alongside deep learning architectures such as Deep Neural Networks (DNN), were trained and rigorously evaluated. The results indicate that gradient boosting models, particularly XGBoost and Light-GBM, consistently outperform others in accurately identifying risky drug substitutes, effectively handling class imbalance and complex feature interactions inherent in medical data. The incorporation of domain-specific risk scoring and comprehensive feature engineering significantly contributed to improved predictive performance and clinical interpretability. This work presents a robust predictive framework that can aid healthcare professionals in making informed, risk-aware decisions regarding drug substitution, ultimately aiming to enhance patient safety and treatment outcomes. Future directions include external validation on diverse patient cohorts and integration of explainable AI methods to facilitate clinical adoption.