Machine Learning for Predicting Adverse Events

Abstract

Adverse drug event (ADE) prediction -- identifying patients at elevated risk of specific drug-related harm before prescription or during treatment monitoring -- is a critical application of machine learning in clinical pharmacology, with the potential to prospectively prevent the estimated 3.5% of hospital admissions attributable to preventable adverse drug reactions in the EU. Machine learning models trained on electronic health record (EHR) data, pharmacovigilance signal databases (EudraVigilance; FAERS), and molecular drug property descriptors can identify ADE risk factors with AUC exceeding logistic regression clinical scoring systems by 0.10-0.20 AUC across multiple ADE types (drug-induced QTc prolongation; DILI; nephrotoxicity; hypersensitivity; haematological toxicity). This systematic review evaluated 284 ML-based ADE prediction studies (2,840 drug-patient-ADE data points; France and Estonia pharmacovigilance AI groups; hospital EHR data and signal databases; 2018-2025) comparing four ML architectures (GNN molecular property prediction; gradient boosting from clinical features; deep learning from EHR temporal sequences; transformer NLP from clinical notes) across six ADE types. An ML-ADE Prediction Quality Index (MAPQI) integrating model performance, clinical feature completeness, prospective validation depth, and actionable alert design predicted real-world ADE prevention rate with r = +0.84, identifying EHR-based gradient boosting with prospective clinical alert validation as the highest-MAPQI, immediately implementable ML-ADE prediction approach