JMIR Publications

ABSTRACT

Background:

Adverse drug events (ADEs) are a major cause of morbidity and mortality. Recent advances in artificial intelligence (AI), particularly deep learning, have enabled the development of models specifically designed for the prediction and detection of ADEs across all stages of drug development.

Objective:

This scoping review aims to provide a comprehensive overview of how AI methods are applied to predict and detect ADEs throughout the drug development pipeline, from preclinical research to post-market surveillance.

Methods:

We conducted a scoping review in accordance with PRISMA-ScR guidelines. A systematic search of PubMed, Web of Science, and Google Scholar identified 1,802 records published between January 2015 and December 2022. After screening and eligibility assessment, 81 studies were included in the final analysis. Inclusion criteria focused on articles using AI to analyze ADEs. Data extraction covered, among other elements, algorithm type, method type, features, data sources, prediction tasks, evaluation metrics, and application stage.

Results:

Among the 81 included studies, 37 addressed pre-market ADE prediction and 44 focused on post-market detection. Most studies originated from the United States and China. The liver and heart were the most commonly studied organs due to their critical roles in systemic drug response. A shift from traditional methods to deep learning approaches is emerging, with transformers in particular becoming increasingly dominant. Commonly used datasets included SIDER, EHR from clinical notes, n2c2 clinical notes, and DrugBank. Text embeddings were the dominant feature representation for detection tasks. Evaluation metrics differed by phase: AUROC was prevalent in pre-market prediction (n=32), while F1-score dominated post-market detection (n=39). Major challenges included the lack of detailed dosage data, limited integration of molecular target and pathway information, and class imbalance in datasets, all of which affect model interpretability and performance assessment.

Conclusions:

Although still emerging, the application of AI in ADE analysis shows significant promise. Our review highlights that various deep learning approaches have already been successfully implemented. As these technologies continue to evolve, they are expected to enhance drug safety, reduce healthcare costs, and support timely pharmacovigilance. Improvements in data quality, model interpretability, and methodological robustness will be essential to facilitate broader clinical adoption. Clinical Trial: Not applicable