Machine learning and deep learning models achieve a pooled AUROC of 0.913 for sepsis predictionMachine learning models show high accuracy in predicting sepsis

BACKGROUND: Machine learning (ML) and deep learning (DL) models have been developed for earlier recognition in hospitalized patients, but reported performance varies across datasets, prediction windows, care settings, and validation designs. Interpretation of a single pooled discrimination estimate is therefore uncertain, particularly because public datasets are often reused, and most evidence is retrospective. OBJECTIVE: This study aimed to synthesize the performance of ML- and DL-based sepsis prediction models in hospitalized adults, emphasizing prediction windows and validation maturity, and to separately describe sepsis-related health care burden using Korean national inpatient claims data. METHODS: We conducted a systematic review and meta-analysis of ML and DL models for sepsis prediction in hospitalized adults. The protocol was registered in PROSPERO. Random-effects meta-analysis used the Hartung-Knapp-Sidik-Jonkman approach, with 95% prediction intervals where sufficient studies were available. Interpretation focused on prediction-window subgroups and validation-maturity tiers rather than a single pooled area under the receiver operating characteristic curve (AUROC). Potential nonindependence from repeated use of Medical Information Mart for Intensive Care (MIMIC) and PhysioNet cohorts was examined through dataset-overlap assessment and sensitivity analysis. Separately, Korean Health Insurance Review and Assessment Service National Inpatient Sample data were used to describe length of stay, medical costs, and surgery counts by sepsis-related episode timing; this analysis did not validate an AI model. RESULTS: In total, 34 studies were included, most of which were retrospective model-development or validation studies. Several reused MIMIC- or PhysioNet-derived cohorts, so the 34 reports did not represent 34 fully independent patient populations. The pooled AUROC was 0.913 (95% CI 0.887-0.933), with a 95% prediction interval of 0.660-0.983. In exploratory subgroup analyses, pooled AUROCs were 0.894 (95% CI 0.829-0.936) for models predicting sepsis within 4 hours, 0.926 (95% CI 0.897-0.948) for models predicting more than 4 hours before onset, and 0.858 (95% CI 0.581-0.964) for unclear or unreported prediction windows. Overlapping prediction intervals indicated substantial uncertainty and did not establish superiority of any prediction horizon. Prospective, randomized, and implementation studies were interpreted separately. In the Korean claims analysis, sepsis-related episode groups showed longer observed hospital stays and higher unadjusted medical costs than general inpatient episodes. CONCLUSIONS: Reported ML and DL sepsis prediction models frequently demonstrated good discrimination within individual study settings, but performance in new clinical populations remains uncertain because of extreme heterogeneity, overlapping public data, inconsistent reporting, and limited prospective evaluation. Prediction-window and validation-maturity analyses were more clinically informative than a single pooled AUROC, although exploratory. The Korean claims analysis provided separate contextual evidence of disease burden and should not be interpreted as AI model validation.