Predictive performance of mortality risk models in sepsis-associated encephalopathy: a systematic review and meta-analysis

Background Sepsis-associated encephalopathy (SAE) is a common central nervous system complication in patients with sepsis and is associated with a significant increase in in-hospital and long-term mortality. In recent years, a variety of mortality risk prediction models based on traditional statistics and machine learning have been proposed to identify high-risk patients. However, the performance differences and clinical applicability of these models remain poorly evaluated. Therefore, this study systematically reviews and quantitatively integrates the performance of existing SAE mortality risk prediction models, compares the predictive efficacy of different modeling approaches, and provides a basis for clinical risk assessment and model optimization. Methods This study adhered to the PRISMA statement, and the study protocol was registered with PROSPERO (CRD420251062130). A systematic search of PubMed, Embase, Web of Science, the Cochrane Library, and CINAHL databases was conducted from inception to May 2025. Studies investigating mortality risk prediction models for adult ICU patients with SAE were included. Model characteristics, predictors, and performance indicators were extracted. A random-effects model was used to perform a meta-analysis of the area under the curve (C statistic). Subgroup comparisons were also performed according to the modeling approach. Results Seven studies were included, and 23 SAE mortality risk prediction models were constructed and evaluated (three using machine learning, three using logistic regression, and one using Cox regression). The overall pooled AUC was 0.82 (95% CI: 0.78–0.86), demonstrating good discrimination. Subgroup analysis showed that the pooled AUC for the machine learning model was 0.85 (95% CI: 0.80–0.90), slightly higher than the 0.80 (95% CI: 0.74–0.87) for the traditional regression model, with no statistically significant difference (p = 0.283). PROBAST assessment indicated that most studies had a certain risk of bias, primarily in model development and validation, such as limited sample size, inconsistent variable selection, and insufficient internal validation. Conclusion Existing SAE mortality risk prediction models generally have good discriminative performance. Machine learning models have a slight advantage in terms of AUC, but insufficient interpretability and external validation still limit their widespread application. Future research should standardize SAE definitions, expand sample sizes, strengthen external validation, and integrate interpretable artificial intelligence with multimodal data fusion techniques to improve model robustness and clinical applicability.