THE ROLE OF AI IN PREDICTING ADVERSE DRUG REACTIONS: ENHANCING PATIENT SAFETY IN PHARMACEUTICAL PRACTICE

Abstract

The introduction of Artificial Intelligence in the pharmacovigilance could be a transformative step to predict Adverse drug reactions and to improve the patients' safety. We evaluate the predictive power of three AI models, Gradient Boosting, Convolutional Neural Networks (CNN), and Long Short Term Memory (LSTM) networks, using both structured Electronic Health Record (EHR) data and unstructured social media data in this study. We evaluate the models using several performance metrics (AUC-ROC, sensitivity, specificity, F1 score) to assess their ability to predict ADRs for various patient demographics. We found CNN to be the best classifier for social media data with an AUC-ROC of 0.91 and 90% sensitivity, and Gradient Boosting to be the best classifier for structured EHR data with an AUC-ROC of 0.89. Feature importance analysis and Shapley Additive explanations (SHAP) provided model interpretability and showed that patient age, drug type, and dosage were significant predictors. The analysis identifies the potential of Natural Language Processing (NLP) in extracting ADR signals from unstructured data sources to supplement traditional pharmacovigilance methods. The study aims to meet regulatory standards in terms of ethical data privacy and model transparency considerations. This work demonstrates that AI models can improve ADR prediction accuracy and contribute to proactive patient safety approaches. The tradeoff between accuracy and interpretability is then applied to clinical applications, and future directions of data standardization and hybrid AI models are explored.