Machine Learning and Large Language Model Decision Support for Medication-Error Prevention Across the Medication-Use Process: A Systematic Review with Implications for Saudi Arabia

Objectives: Medication errors are a major source of preventable harm. We systematically reviewed workflow integrated artificial intelligence (AI) decision support (machine learning [ML], natural language processing, computer vision, and large language models [LLMs]) aimed at detecting, intercepting, or preventing medication errors, and considered implications for Saudi Arabia.Methods: We searched MEDLINE (PubMed), Embase, Scopus, Web of Science Core Collection, IEEE Xplore, and the Cochrane Library for English language studies published from January 2010 to February 2026. Eligible studies used real clinical/pharmacy data and evaluated an AI component designed to prevent or detect medication errors across prescribing/order verification, dispensing, administration, reconciliation, or medication directions. Two reviewers screened and extracted data independently; risk of bias was assessed with PROBAST, QUADAS2, or ROBINS-I.Results: A total of 16 studies met inclusion criteria. Nine evaluated prescribing/order verification tools (including look alike/sound alike and outlier detection), 3 evaluated administration support (including computer vision verification), 2 evaluated reconciliation/triage, 1 evaluated dispensing verification, and 1 evaluated an LLM for medication directions. Performance was often high for constrained tasks, but external validation, calibration, and explicit alert threshold governance were uncommon. Deployed evaluations suggested potential reductions in errors or near misses, but most impact evidence was observational, and alert burden/overrides were inconsistently reported. Heterogeneity precluded meta-analysis; using GRADE, certainty of evidence for error reduction was generally low.Conclusion: Workflow integrated AI decision support can augment medication safety, particularly via high plus pressure variation surveillance and task specific verification, but the current evidence base provides limited causal certainty for harm reduction. For Saudi Arabia, scale up should be phased and governed through interoperability, clinical decision support governance, human factors evaluation, Saudi Food and Drug Authority aligned software as a medical device pathway, and post deployment monitoring.