Machine learning methods for predicting adverse drug events: A systematic review.

Area of Science:

• Pharmacovigilance
• Health Informatics
• Machine Learning in Healthcare

Background:

• Predicting adverse drug events (ADEs) is vital for patient safety and cost reduction in outpatient care.
• Traditional methods face limitations with complex healthcare data, prompting exploration of machine learning (ML).
• The effectiveness of ML models for ADE prediction in real-world outpatient settings requires systematic evaluation.

Purpose of the Study:

• To systematically review machine learning algorithms applied to adverse drug event prediction in outpatient settings.
• To analyze study characteristics, ML methods, performance metrics, and risk of bias in existing research.
• To identify current limitations and future directions for ML-based ADE prediction.

Main Methods:

• Systematic literature search of MEDLINE and Embase up to December 2024.
• Inclusion of studies developing or validating ML models for ADE prediction in outpatient or similar large-scale data.
• Assessment of study characteristics, ML algorithms, performance (AUC), and risk of bias using the PROBAST tool.

Main Results:

• 191 ML implementations across 59 studies were analyzed; Logistic Regression, Random Forest, and XGBoost were common.
• Most studies (85%) reported moderate to high internal validation performance (AUC > 0.70).
• Significant methodological gaps exist, including poor handling of class imbalance (33.9%) and limited external validation (18.6%).

Conclusions:

• Machine learning models, particularly ensemble methods, demonstrate potential for predicting outpatient adverse drug events.
• Current limitations in addressing class imbalance and conducting external validation hinder widespread clinical adoption.
• Future research must prioritize rigorous methodologies, external validation, and integration with pharmacovigilance practices for reliable ADE prediction.