Artificial Intelligence for Drug Safety Across the Lifecycle and Decision Type: A Scoping Review

TL;DR

This review maps how AI is used for drug safety decisions across the drug lifecycle and finds that AI is most commonly applied to patient-level safety prediction and post-marketing safety surveillance.

Contribution

The novel contribution is a lifecycle–decision matrix that clarifies where AI applications are concentrated and identifies gaps in external validation and real-world testing.

Findings

AI applications are concentrated in patient-level safety prediction and post-marketing safety surveillance using EHRs and spontaneous reporting systems.

Common AI methods include gradient boosting, deep neural networks, and natural language processing models.

Most studies use internal validation, with limited external validation and real-world deployment.

Abstract

Background/Objectives: Artificial intelligence (AI) is increasingly applied to drug safety evaluation, yet evidence is dispersed across lifecycle stages and tasks. This scoping review aimed to (1) map how AI supports safety- and treatment-related decision types across the drug lifecycle, and (2) examine evaluation strategies used to assess model reliability for clinical or regulatory use. Methods: Using Arksey and O’Malley’s framework, we searched a major database for studies published in the past decade that applied AI or machine learning to drug safety or medication-related decisions. After screening, we extracted data on lifecycle stage, decision type, AI methods, data sources, and evaluation strategies. A lifecycle–decision matrix was constructed to characterize application patterns. Results: AI applications were concentrated in real-world clinical care × patient-level safety prediction and post-marketing × safety surveillance, using EHRs, spontaneous reporting systems, and clinical text. Common methods included gradient boosting, deep neural networks, graph neural networks, and natural language processing models. This concentration reflects structural incentives favoring safety-oriented applications with readily available data and lower decision liability. Evidence for treatment optimization, regulatory decision modeling, and evidence synthesis was limited. Most studies used internal validation; external validation and real-world deployment were uncommon, indicating early methodological maturity and limited translational readiness. Conclusions: AI demonstrates strong potential to enhance drug safety—particularly in risk prediction and pharmacovigilance—but its use remains uneven across the lifecycle. By situating AI applications within explicit lifecycle stages and decision contexts, this review clarifies where progress has advanced, where translation has stalled, and why these gaps persist. Limited external validation and minimal real-world testing constrain clinical and regulatory adoption. These findings suggest that external validation and real-world testing may contribute to further advances in AI for drug safety.

Figures