Artificial intelligence (AI) in pharmacovigilance: A systematic review on predicting adverse drug reactions (ADR) in hospitalized patients

Introduction: Adverse drug reactions (ADRs) significantly impact healthcare systems, leading to increased hospitalization rates and costs. With the growing adoption of artificial intelligence (AI) in healthcare, machine learning (ML) models offer promising solutions for ADR prediction. However, comprehensive evaluations of these models remain limited. Methods: This systematic review synthesized findings from 13 studies that utilized various ML algorithms (regression-based, flexible, and ensemble models) to predict ADRs using data such as patient demographics, laboratory values, and comorbidities. Meta-analysis was conducted to assess the pooled sensitivity and specificity of the models, and a co-authorship and keyword analysis was performed to examine collaborative networks within the field. Results: The included studies primarily focused on model development (77 %), with only 23 % incorporating external validation, raising concerns about generalizability across clinical contexts. Meta-analysis showed pooled sensitivity and specificity of 78.1 % and 70.6 % for development-only studies, while studies with external validation achieved higher sensitivity (81.5 %) and specificity (79.5 %). Co-authorship analysis identified 67 contributors across eight collaboration clusters, indicating a specialized but emerging research field. Discussion: The findings highlight the need for multifactorial models that integrate diverse predictors to improve the performance and reliability of ML-based ADR prediction. Addressing these limitations through rigorous model development and validation processes could enhance the clinical applicability of AI-driven pharmacovigilance, ultimately advancing patient safety and healthcare outcomes.