Accuracy of Artificial Intelligence-Based Models versus Traditional Scoring Systems (APACHE, SOFA, SAPS) for Predicting Mortality in ICU Patients: A Systematic Review and Meta-Analysis

IntroductionReliable estimation of mortality among critically ill patients is crucial for guiding clinical decisions and optimizing ICU performance. Traditional scoring systems such as APACHE, SOFA, and SAPS are commonly applied, though their predictive capacity is constrained by their reliance on static structures and linear modeling assumptions. Artificial intelligence-based models provide flexible, data-oriented prediction strategies, yet their comparative accuracy remains unclear. This study systematically reviewed and meta-analyzed the performance of Artificial intelligence-based models versus conventional ICU scores for predicting in-hospital mortality in adults admitted to ICU. Materials and MethodsLiterature searches were performed in PubMed, Embase, Web of Science, Scopus and the Cochrane Library from January 2015 to August 2025 for studies comparing AI models with traditional scoring systems. Studies were included if they provided diagnostic performance indicators including AUC, sensitivity, or specificity. Risk of bias was assessed using PROBAST, and pooled statistical estimates were derived through bivariate random-effects modeling with Fisher s Z- transformation. Subgroup analyses examined AI modality, ICU type, and geographic region. ResultsEleven studies involving over one million ICU admissions met inclusion criteria. Two studies (Huang 2023; Lim 2024) provided complete 2 by 2 data for meta-analysis. Pooled sensitivity and specificity for AI models were 0.875 (95% CI: 0.840 - 0.904) and 0.857 (95% CI: 0.845 - 0.868), respectively. AI models achieved higher AUCs (0.82 - 0.90) than APACHE II (0.70 - 0.78), SOFA (0.68 - 0.75), and SAPS II (0.70 - 0.79). Deep learning and ensemble methods performed best across ICU settings and regions. ConclusionAI-based models outperform conventional scoring systems in predicting ICU mortality. Their integration into critical care could enhance early risk stratification and precision prognostication. HighlightsThis meta-analysis highlights that artificial intelligence-based predictive models demonstrated superior predictive performance than conventional ICU scoring systems (APACHE, SOFA, and SAPS) in predicting in-hospital mortality, with higher pooled sensitivity, specificity, and overall discriminative accuracy, particularly for deep learning and ensemble approaches across diverse ICU settings.