Abstract
Background: Postoperative sepsis and related infectious complications remain major causes of morbidity after abdominal surgery. Artificial intelligence (AI) and machine learning (ML) models may support earlier risk stratification, but their generalizability and clinical readiness remain uncertain. Methods: We systematically searched PubMed, Scopus, and CENTRAL from inception to 10 August 2026 for studies developing, validating, or evaluating AI/ML models for postoperative sepsis, postoperative infection/surgical-site infection (SSI), anastomotic leakage, or postoperative pancreatic fistula (POPF) after major abdominal or general surgery. SSI, anastomotic leakage, and POPF were analyzed as distinct postoperative endpoints and were not treated as proxy definitions of clinical sepsis. Study selection and data extraction were performed independently by two reviewers, and risk of bias and applicability were assessed using PROBAST + AI. Owing to substantial clinical and methodological heterogeneity, results were synthesized narratively. The review was registered in PROSPERO (CRD420261492111). Results: Forty-five studies were included: 11 addressed sepsis-related outcomes, including six studies with a primary sepsis-related modeled endpoint whose definitions ranged from consensus Sepsis-3 criteria to coded or SIRS-based definitions and a qSOFA-based proxy; 19 evaluated postoperative infection/SSI; eight anastomotic leakage; five POPF; and two broader postoperative complications. Performance varied according to outcome, surgical population, prediction timing, predictor availability, and validation strategy. Fifteen of 45 studies provided strict temporal, geographic/external-center, or independent-investigator evaluation of the principal model, with variable preservation of performance outside development settings. More complex ML methods did not reliably outperform conventional regression. Explicit calibration analysis was reported in 9/45 studies, decision-curve or net-benefit analysis in 7/45, and no study prospectively evaluated the impact of model-guided care on clinical outcomes. Certainty of evidence was very low for primary postoperative sepsis/septic-shock prediction, low for postoperative infection/SSI, and very low for anastomotic leakage and POPF. Conclusions: Although several AI/ML models showed promising predictive performance, the available evidence remains insufficient to support routine use of AI/ML prediction models across major abdominal surgical populations. Future studies should prioritize clearly defined clinical use cases, multicenter external validation, calibration, prospective workflow evaluation, and demonstration of actionable clinical benefit before routine implementation.