#37 Artificial intelligence for early detection of critical events in emergency medicine patients: a systematic review and meta-analysis

Abstract

<h3>Introduction</h3> Emergency Departments face critical challenges in early recognition of patient deterioration within 24–48 hours post-admission with 5–10% experiencing unexpected deterioration within 24 hours. Traditional screening tools such as early warning scores demonstrate limited predictive accuracy (sensitivity 65–80%) and high false-positive rates. This review evaluates artificial intelligence (AI) and machine learning (ML) predictive models for the early detection of critical conditions in emergency department settings. <h3>Methods</h3> This systematic review and meta-analysis were conducted following PRISMA-DTA guidelines. Comprehensive searches of EMBASE, PubMed, ACM Digital Library, IEEE Xplore, Web of Science, and arXiv identified seven studies developing or validating AI/ML predictive models for adult emergency department patients. Studies included 5,945,118 patients with 96,604 critical events (January 2010-March 2025). Primary outcomes included mortality, sepsis, stroke, vasoactive drug requirement, and intubation. Quality assessment utilized PROBAST and TRIPOD tools. Meta-analysis employed random-effects models with subgroup analyses. <h3>Results</h3> ML models exhibited superior performance versus traditional risk scoring systems: pooled area under the curve 0.933 versus 0.844. ML models achieved superior sensitivity 0.824 (95% CI: 0.664–0.917) versus 0.489 (0.292–0.689) for traditional systems, while maintaining comparable specificity (0.938 vs 0.923). Deep learning models achieved the highest performance (AUROC 0.967), followed by ensemble methods (0.928) and gradient boosting (0.914). Sepsis prediction models achieved strong performance (AUROC 0.931–0.970). Substantial heterogeneity (I²&gt;80%) was observed across studies, likely due to clinical heterogeneity, though the effect direction consistently favoured ML models. All studies exhibited a moderate risk of bias due to internal validation approaches. <h3>Conclusion</h3> AI and ML models significantly outperform traditional systems statistically, with potential for clinical utility, with a 68% relative sensitivity improvement demonstrated. Moderate-certainty evidence supports pilot implementation in high-volume emergency departments with robust electronic health record systems. However, the absence of external validation necessitates cautious deployment with rigorous local validation. Future research should focus on prospective multicentre studies with external validation cohorts. <i>*presenting author</i>

Ähnliche Arbeiten