Bridging the gap between model performance and clinical practice: A scoping review of artificial intelligence models...

Drug-drug interactions (DDIs) are an important source of preventable adverse drug events and remain challenging to address in clinical practice. Conventional rule-based clinical decision support systems are frequently associated with high false-positive rates and alert override exceeding 90%, contributing to alert fatigue. Artificial intelligence (AI) approaches are increasingly applied to DDI prediction; however, their methodological robustness and clinical readiness remain unclear. This scoping review mapped AI-based DDI prediction studies, examining input features, training strategies, validation approaches, performance metrics, and clinical applicability. Of 963 records screened across five databases, 131 studies met the inclusion criteria. Deep learning models predominated (82, 62%), with supervised learning (118, 90%) and internal validation (108, 82%) most common. Most studies relied on chemical structure-based features and curated knowledge bases (notably DrugBank), whereas only 18 (13.7%) studies incorporated patient-level variables. Although 25 (19%) studies reported AUROC, AUPR, and F1-score ≥ 0.90, these findings were largely derived from internally validated datasets, with limited external or clinical validation. AI-based DDI prediction research demonstrates substantial methodological sophistication but remains constrained by limited patient contextualization, interpretability challenges, and insufficient external validation, highlighting a translational gap between algorithmic performance and clinical applicability.