Systematic review of AI-based models in pharmacoepidemiology for adverse drug event prediction and detection

TL;DR

This paper reviews how AI models are used to detect and predict adverse drug events in real-world clinical data.

Contribution

The study systematically characterizes AI-based methods for adverse drug event prediction and highlights gaps in validation and model transparency.

Findings

Most studies used structured EHRs or claims data, with limited use of natural language processing.

Tree-based models were most common, while deep learning was less frequently applied.

External validation and explainability methods were rarely used, limiting generalizability.

Abstract

Artificial intelligence (AI) has increasingly been applied in pharmacoepidemiology, yet the methodological landscape of adverse drug event (ADE) prediction remains heterogeneous and insufficiently mapped.

This systematic review aimed to characterize contemporary AI-based approaches used to detect or predict ADEs in real-world clinical data. Following PRISMA 2020 guidelines and a registered protocol (PROSPERO: CRD420251159394), 281 records were screened and 15 studies met the inclusion criteria.

All included studies relied primarily on structured electronic health records (EHRs) or administrative claims, while only a minority incorporated natural language processing (NLP) components, and none used spontaneous reporting systems as the primary analytic datasets. Tree-based ensemble models (e.g., Random Forests, XGBoost) and regularized regression were the most commonly adopted algorithms, whereas deep learning architectures appeared less frequently and typically required temporal or representation-based inputs. Through studies, external or temporal validation was rarely performed and explainability methods were inconsistently applied, limiting generalizability. No standardized benchmarks were identified, and reporting practices varied substantially.

Future work should emphasize rigorous validation, transparent model reporting, and the careful integration of NLP and explainability frameworks to support clinically reliable and scalable pharmacoepidemiological applications.

Figures