Abstract
OBJECTIVE: To systematically evaluate the performance and methodological rigor of published prediction models for sepsis-associated encephalopathy (SAE), identify their limitations, and provide guidance for the future development of robust and clinically applicable models. METHODS: We conducted a systematic search across nine English and Chinese databases (from inception to May 2025) for studies developing or validating SAE prediction models in adult sepsis patients. Two researchers independently gathered data, using PROBAST to assess methodological quality, and conducted a meta-analysis of the AUC of logistic regression models. RESULTS: Ten studies were included, encompassing 55,244 patients with sepsis, revealing an incidence of SAE ranging from 15.0 to 62.4%. A total of 29 predictive models were developed, comprising 10 optimal models, primarily utilizing logistic regression or machine learning algorithms. The combined AUC of the five logistic regression models was 0.85 (95% CI 0.77-0.93), exhibiting substantial heterogeneity (I(2) = 91.8%). All models showed a high risk of bias according to the PROBAST evaluation, mainly due to the lack of external validation and methodological shortcomings. CONCLUSION: Current SAE prediction models demonstrate moderate discriminatory ability, but their methodological quality remains poor, and they are not yet suitable for routine clinical application. In the future, standardized SAE definitions and prospective data collection should be strengthened, models should be developed and validated strictly following the TRIPOD guidelines, and model interpretability should be improved to promote clinical application. SYSTEMATIC REVIEW REGISTRATION: https://www.crd.york.ac.uk/PROSPERO/view/CRD420251062747.