Artificial Intelligence Versus Logistic Regression for Postoperative Mortality Prediction: A Systematic Review and Meta-Analysis

Postoperative mortality remains a major global health burden, and accurate risk prediction is essential for perioperative decision-making and quality improvement.We conducted a systematic review and metaanalysis comparing artificial intelligence (AI)-based models with logistic regression (LR) for postoperative mortality prediction in adult surgical populations.PubMed, Excerpta Medica database (Embase), and the Cochrane Library were searched from inception to February 2, 2026.The review was registered in the International Prospective Register of Systematic Reviews (PROSPERO) on May 5, 2026 (registration number: CRD420261365093).Eligible studies directly compared AI-based and LR models within the same adult surgical population and reported the area under the receiver operating characteristic curve (AUC) for 30-day and/or 90-day postoperative mortality.Risk of bias was assessed using the Prediction model Risk Of Bias Assessment Tool (PROBAST).A random-effects meta-analysis with restricted maximum likelihood and Hartung-Knapp adjustment was performed to pool the between-model difference in AUC (ΔAUC).The search identified 977 records; after removal of 70 duplicates, 907 records were screened, 68 reports were assessed for eligibility, and 11 retrospective cohort studies including 372,343 participants were included.All studies showed high or unclear risk of bias.Seven studies contributed to the 30-day mortality synthesis, in which AI-based models showed better discrimination than LR (ΔAUC: 0.0575, 95% confidence interval (CI): 0.0128-0.1022;I² = 89.0%).Six studies contributed to the 90-day mortality synthesis, also favoring AI-based models (ΔAUC: 0.0509, 95% CI: 0.0078-0.0940;I² = 39.9%).AI-based models showed modest pooled improvements in discrimination compared with LR for postoperative mortality prediction.However, confidence in these findings is limited by high or unclear risk of bias, the small number of eligible studies, and heterogeneity, particularly in the 30-day analysis.