Artificial intelligence in hospital fall Prevention: Current Applications, Challenges, and Future Directions

• Machine learning outperforms traditional fall risk tools. • Computer vision enables real-time behavioural monitoring. • Wearable sensors support 24/7 patient movement tracking. • NLP extracts fall risk factors from clinical documentation. • AI-driven systems reduce fall rates by 15–40% in hospitals. Hospital falls represent a critical patient safety challenge, affecting millions of patients globally and resulting in substantial morbidity, mortality, and healthcare costs. Traditional fall prevention strategies, whilst beneficial, often lack the precision and real-time responsiveness needed for optimal patient protection. This narrative review with systematic search examines the current applications, effectiveness, and implementation challenges of artificial intelligence (AI) technologies in hospital fall prevention. A comprehensive search was conducted across PubMed, EMBASE, IEEE Xplore, and Google Scholar databases from January 2015 to October 2024. AI technologies demonstrate promise across four primary domains: machine learning predictive models achieving AUROC of 0.85–0.97 (with calibration reported variably), computer vision systems enabling real-time behavioural monitoring (94–97% detection accuracy in controlled settings), sensor-based technologies providing continuous patient surveillance (89–96% accuracy with multi-sensor fusion), and natural language processing enhancing risk factor extraction from clinical documentation (sensitivity 95% CI in selected studies). These metrics represent primarily single-site, retrospective studies with limited external validation and variable baseline fall rates. Successful implementations report fall reduction rates of 0.9–1.2 falls per 1,000 patient-days (15–40% relative reduction) across various healthcare settings, though baseline rates ranged from 2.8 to 5.1 falls per 1,000 patient-days across different care settings, and secular trends and study design heterogeneity limit causal inference. AI-driven systems offer enhanced prediction accuracy, real-time monitoring capabilities, and personalised risk assessment. However, implementation challenges include alarm fatigue (alert rates and positive predictive value rarely reported), algorithmic bias requiring ongoing fairness audits, liability concerns when AI systems fail to prevent falls, data privacy concerns, integration complexities, clinical workflow adaptation, and substantial cost barriers for smaller institutions. Future developments should prioritize explainable AI systems, multisite external validation with standardised metrics (AUROC, AUPRC, calibration), federated learning approaches, and implementation trials examining both fall rates and care process outcomes.