The review team conducted systematic searches of PubMed, Embase, Web of Science, and the Cochrane Library through April 8, 2026, and screened records against prespecified criteria. Studies had to include adults (≥18 years) with sepsis defined by Sepsis-3.0, develop or validate machine learning mortality models, and report performance metrics or sufficient data to derive them. Cohort designs only were eligible. After screening 9,550 records and reviewing 135 full texts, 72 studies met inclusion and were analyzed.
The pooled dataset comprised 484,826 patients across the 72 included cohorts. The vast majority of studies (67 of 72) were performed in China; four studies came from Italy, Poland, South Korea, and Turkey. Most studies reported short-term mortality endpoints (71 studies); only one reported long-term mortality. Frequent predictors across models included age, lactate, heart rate, SOFA score, respiratory rate, blood urea nitrogen, white blood cell count, platelet count, systolic blood pressure, and body temperature. Multiple imputation was the most common missing-data strategy; LASSO and univariate-plus-multivariable selection were widely used for predictor selection.
Investigators evaluated 28 modeling approaches in total: 22 traditional machine learning algorithms and 6 deep learning algorithms. The most commonly used model was logistic regression (reported in 63 studies), followed by random forest (22 studies) and support vector machine (19 studies). Tree-based ensemble methods such as random forest and XGBoost were also frequently tested. Deep learning variants (CNN, DNN, LSTM, MLP, and a combined DL-CMT) were included in some studies. The review classified reported model performance into training and validation sets; internal and external validations were grouped together under the validation category for pooled analyses.
For short-term mortality among patients with sepsis, pooled performance on training data yielded an overall AUC of 0.82 across 46 studies and 15 model types. Validation-set performance pooled across 50 studies (25 models) gave an AUC of 0.79. Logistic regression demonstrated stable performance, with a reported validation AUC of 0.81. Integrated tree-based methods (random forest and XGBoost) showed validation AUCs near 0.81–0.83. LightGBM and DL-CMT achieved the highest AUCs in training and validation sets respectively in the source, but both of those estimates were based on single-study data, and the authors cautioned against overinterpreting single-study highs. Heterogeneity across pooled analyses was substantial in many comparisons.
The meta-analysis separated performance by complication. Sepsis-associated encephalopathy (SAE) showed combined evidence across three studies with an AUC of 0.81. Sepsis-induced coagulopathy (SIC) had an AUC of 0.80 across four studies. Sepsis-associated acute kidney injury (SAKI) demonstrated moderate and stable performance with a validation AUC of 0.79 based on four studies. Septic shock reported a very high AUC (0.95) in the source, but this estimate came from a single study. Sepsis-associated delirium had limited predictive utility (AUC = 0.67, single study). Sepsis-induced myocardial injury and sepsis-induced cardiomyopathy had only single-study or sparse evidence; cardiomyopathy showed attenuation from training to validation (AUC 0.84 vs 0.78 in the single-study report).
The authors used PROBAST to assess methodological quality. They judged 17 studies as high overall risk of bias and 55 as unclear overall risk. Bias concentrated in the Analysis domain, whereas Participants, Predictors, and Outcome domains were generally rated low risk. Applicability was rated low concern in 69 studies and high concern in three. The pooled estimates frequently exhibited high statistical heterogeneity, and for outcomes with ten or more studies the authors assessed publication bias using funnel plots and Egger’s test.
Long-term mortality predictions were effectively absent. The source reported only single-study long-term estimates for a few complications (for example, SAE and SCRS), so no robust conclusions can be drawn. The study team emphasized that external validation was infrequent and that many high-performing single-study models lacked replication. Limitations documented in the source include concentration of studies in a single country, dominance of retrospective designs, frequent unclear risk of bias in analysis methods, substantial heterogeneity across studies, and insufficient external validation. The authors called for higher-quality, externally validated studies and more research on under-investigated complications before clinical deployment of these models can be advised.