Document Type : Original Article
Authors
1
Eye Research Center, Department of Eye, Amiralmomenin Hospital, School of Medicine, Guilan University of Medical Science, Rasht, Iran
2
Department of nursing, Naghadeh School of Nursing, Urmia University of Medical Sciences, Urmia, Iran
3
M.Sc. in Mathematics, Iran University of Science and Technology, Department of Mathematics and Computer Science, Iran University of Science and Technology
4
PhD in TEFL, Department of Basic sciences, Shoushtar Faculty of Medical Sciences, Shoushtar, Iran
Abstract
Adverse drug reactions (ADRs) and adverse drug events (ADEs) remain major threats to patient safety in hospitals, particularly among older adults, patients with multimorbidity, individuals exposed to polypharmacy, and patients experiencing rapid changes in physiological status. Conventional pharmacovigilance and medication-monitoring approaches depend heavily on retrospective reporting and clinical recognition, which may delay identification of preventable harm. Artificial intelligence (AI), particularly machine learning (ML), natural language processing, and deep learning, offers opportunities to transform pharmacovigilance from a predominantly reactive process into a predictive and continuously updated safety system. This systematic evidence synthesis examined the performance and clinical implications of AI-based models for predicting ADRs and ADEs among hospitalized patients, with particular attention to nursing practice. Evidence published in major biomedical databases synthesized with emphasis on model performance, predictors, validation, explainability, and implications for medication administration and clinical surveillance. Recent systematic reviews indicate that random forest, gradient boosting, support vector machines, regularized regression, and deep-learning architectures frequently used, with pooled predictive performance generally demonstrating moderate discrimination. A 2025 systematic review specifically examining hospitalized patients reported pooled sensitivity and specificity of 78.1% and 70.6%, respectively, for development-only models, 81.5%, and 79.5% for models undergoing external validation. Contemporary evidence also demonstrates increasing use of longitudinal EHR representations, clinical narratives, nursing documentation, and foundation models. However, heterogeneous outcome definitions, missing-data procedures, limited external validation, class imbalance, alert fatigue, algorithmic bias, and insufficient explainability remain substantial barriers to implementation. AI therefore conceptualized as an augmentation technology rather than a replacement for nurses or other clinicians. Effective implementation requires interdisciplinary governance, transparent algorithms, prospective evaluation, workflow integration, education, and continuous monitoring. Properly implemented, AI-supported ADR prediction may strengthen nursing vigilance, prioritize high-risk patients, improve medication safety, and facilitate earlier clinical intervention.
Graphical Abstract
Keywords
Subjects
