Abstract
OBJECTIVE: To systematically review the outcome constructs, modeling characteristics, and methodological quality of existing sepsis-associated brain dysfunction (SABD) risk prediction models, with the aim of explaining why current models are difficult to reproduce or translate into practice, and of proposing standardized directions for future research. METHODS: A systematic review was conducted by searching CNKI, Wanfang, VIP, SinoMed, PubMed, CINAHL, Cochrane Library, Embase, and Web of Science from database inception to April 2025. Studies developing or validating SABD risk prediction models were included, with outcomes defined as sepsis-associated encephalopathy (SAE) or sepsis-associated delirium (SAD). Model characteristics were extracted according to the CHARMS checklist, and methodological quality was assessed using the Prediction model Risk of Bias Assessment Tool (PROBAST). RESULTS: Twelve studies involving 24 risk prediction models were included, of which four studies evaluated SAD as the outcome and eight evaluated SAE. Substantial heterogeneity was observed in outcome definitions, modeling strategies, and variable selection approaches. Calibration was reported in 10 studies, internal validation in nine studies, and both internal and external validation in one study. According to PROBAST, three studies had high applicability concerns and nine had low applicability concerns. All included studies were assessed as having a high risk of bias, predominantly in the analysis domain. CONCLUSION: Current risk prediction modeling studies for SAD and SAE remain exploratory, and high risk of bias together with insufficient validation limits their reliable clinical translation. Future research should adhere to the PROBAST and TRIPOD guidelines, conduct multicenter prospective studies, and standardize modeling and validation procedures. SYSTEMATIC REVIEW REGISTRATION: https://www.crd.york.ac.uk/, identifier CRD420251014680.