Prediction models for mortality in patients with sepsis: a systematic review and meta-analysis

Sepsis remains a leading cause of mortality among critically ill patients worldwide. Although an increasing number of prediction models have been published in recent years, their predictive performance, methodological quality, and major predictors have not been comprehensively evaluated in a systematic and quantitative manner. This study aims to evaluate the performance of these models and to identify common predictors associated with sepsis mortality.

We systematically searched PubMed, Embase, Cochrane Library, and Web of Science for studies on sepsis mortality prediction models published up to July 1, 2025. Data were extracted and appraised using the Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modeling Studies (CHARMS), and risk of bias was assessed with the Prediction Model Risk of Bias Assessment Tool for Artificial Intelligence (PROBAST+AI). Meta-analyses were performed to pool area under the curve of the receiver operating characteristic (AUC) metric of externally validated models and the odds ratio (OR) of common predictors. The study was registered in PROSPERO (CRD42024604119).

A total of 84 eligible studies were included, reporting 235 prediction models for sepsis mortality and involving approximately 2.7 million patient records reported across studies, with 461,387 deaths. Only 11(13.10%) studies encompassed model development, internal validation, and external validation. The included studies comprised 78(92.86%) retrospective cohort studies, 57(67.86%) studies developed in intensive care unit (ICU) settings, with MIMIC databases being among the most commonly used data sources. The most prevalent mortality endpoints were in-hospital (n = 38, 45.24%), 28-day (n = 23, 27.38%), and 30-day mortality (n = 16, 19.05%). The included studies employed various modeling approaches, such as logistic regression (LR), random forest (RF), eXtreme Gradient Boosting (XGBoost), and K-nearest neighbors (KNN). Of the included studies, 46(54.76%) evaluated the calibration, 26(30.95%) conducted decision curve analysis, and 25(29.76%) applied SHAP for interpretability, while most did not provide visual model presentation (e.g., nomograms). The reported AUCs ranged from 0.590-0.976 for model development and 0.530-0.992 for internal validation. The pooled AUC of externally validated models, based on one representative model per study, was 0.794 (95% CI: 0.755-0.834), indicating moderate discriminative performance. Overall, the most commonly used predictors were age (n = 19, 22.62%), lactate (n = 13, 15.48%), albumin (n = 8, 9.52%), SOFA score (n = 8, 9.52%), and vasopressor (n = 7, 8.33%). Notably, 64 (76.19%) studies and 50 (65.79%) studies were judged to have a high risk of bias in model development and model evaluation phases, respectively.

Externally validated prediction models generally demonstrate moderate discriminative performance for predicting sepsis mortality, but a substantial proportion of these studies were evaluated as having a high risk of bias. Age, lactate, albumin, SOFA score, and vasopressor use were identified as predictors of mortality. Future studies with larger cohorts, rigorous designs, and multicenter external validation are warranted to improve their generalizability and facilitate clinical implementation.

The unique registration identifier is CRD42024604119, and the publicly accessible website is https://www.crd.york.ac.uk/prospero/.