Artificial Intelligence and Machine Learning in Pharmacovigilance: A Systematic Review of Models for Drug Safety in Polypharmacy

Background: Polypharmacy is increasingly prevalent worldwide and is strongly associated with adverse drug reactions (ADRs) and drug–drug interactions (DDIs). Traditional pharmacovigilance (PV) systems rely heavily on spontaneous reporting and manual signal detection, which suffer from substantial underreporting, delayed signal identification, and limited capacity to manage complex multidrug interactions. Artificial intelligence (AI) and machine learning (ML) offer scalable, data-driven approaches that may enhance the detection, assessment, and prevention of medication-related harms in polypharmacy populations. This systematic review evaluates current AI/ML models in PV for polypharmacy, compares their predictive performance, and identifies key methodological facilitators and challenges. Methods: A systematic review was conducted in accordance with PRISMA 2020 guidelines and a pre-registered PROSPERO protocol (CRD420251134685). Seven electronic databases, including grey literature sources, were searched from inception to 2025. Eligible studies were primary computational modeling investigations that developed AI/ML models using polypharmacy-focused datasets and reported performance using the area under the receiver operating characteristic curve (AUROC) and the area under the precision–recall curve (AUPRC). Study quality was appraised using the Mixed Methods Appraisal Tool (MMAT). Results: Of 7,513 records identified, 23 studies met the inclusion criteria. All studies employed computational modeling designs using large pharmacovigilance and biomedical datasets such as TWOSIDES, Decagon, STITCH, SIDER, and DrugBank. Reported AUROC values ranged from 0.553 to 0.9993, while AUPRC values ranged from 0.112 to 0.999. High-performing models included PU-MLP, DPSP, SimVec, TriVec, and SimplE, many of which achieved AUROC and AUPRC values exceeding 0.95. Key strengths of AI/ML approaches included high predictive accuracy, effective handling of class imbalance, and integration of heterogeneous data sources. Major challenges included data sparsity, limited interpretability, cold-start prediction problems, and a lack of external or clinical validation. Conclusion: AI and ML models demonstrate strong potential to enhance pharmacovigilance in polypharmacy settings by improving the detection of ADRs and DDIs. Future research should prioritize standardized benchmarking, model interpretability, and real-world clinical validation to support safe and effective regulatory and clinical implementation.