(A-180) Machine Learning Methods for Predicting Adverse Drug Events: A Systematic Review

Background: Predicting adverse drug events (ADEs) in outpatient settings is crucial for improving medication safety, identifying high-risk patients, and reducing healthcare costs. Traditional methods struggle with the complexity of healthcare data, while machine learning (ML) models offer improved prediction capabilities. However, the effectiveness of ML in ADE prediction remains unclear.

Objectives: To evaluate the types and performance of ML algorithms used for ADE prediction in outpatient settings and assess the methodological rigor and generalizability of these models.

Methods: A systematic review of MEDLINE and Embase up to December 2024, identified studies developing or validating ML models for ADE prediction. Study characteristics, ML methods, ADE types, model performance, and risk of bias were assessed using the PROBAST tool. Inclusion criteria included studies with ML models developed or validated using outpatient data. Exclusion criteria involved studies that did not employ ML methods for ADE prediction.

Results: Fifty-nine studies (191 ML implementations) were included. Logistic Regression, Random Forest, and XGBoost were most commonly used. The majority of studies (40/59, 67.8%) reported AUC, with 85% of them (34/40) having moderate to high performance (AUC-ROC > 0.70) for internal validation. Only 33.9% of all studies (20/59) addressed class imbalance, and 18.6% (11/59) conducted external validation. Methodological concerns included missing data handling and lack of external validation. Despite these limitations, evidence from the reviewed studies suggests that ML models, when properly implemented, have the potential to enhance patient safety by identifying at-risk patients and informing preventive strategies.

Conclusions: ML models, especially ensemble methods such as Random Forest and XGBoost, can predict ADEs in outpatient settings, but challenges such as class imbalance and limited external validation hinder clinical applicability. To overcome these challenges, future research should focus on adopting more rigorous methodologies and developing specialized frameworks for ML-based ADE prediction that build upon existing pharmacovigilance standards. This would facilitate the development of more valid and generalizable models that can be effectively integrated into clinical workflows, enabling regular monitoring and improvement of medication safety.