Sepsis-associated encephalopathy (SAE) refers to widespread brain dysfunction caused by sepsis without central nervous system infection, which requires the identification and exclusion of other encephalopathies as a prerequisite (1). Research indicates that SAE is the predominant form of critical encephalopathy in intensive care units (ICUs) (2, 3). The pathophysiology of SAE is highly complex, clinical diagnosis remains challenging, and the absence of standardized diagnostic criteria leads to significant variability in its occurrence among research (4, 5). A retrospective cohort study from Chinese ICUs indicated an SAE incidence of 48.1% (6), whereas French ICUs reported an incidence of 53.0% (7), and the United States Critical Care Database recorded approximately 70.0% (8). SAE may clinically present as agitation, diminished attention, altered consciousness, delirium, and coma, all of which are closely associated with poor patient outcomes (9, 10). Previous studies have indicated that SAE significantly increases the mortality rate of patients within 30 days or even 1 year (11, 12). Furthermore, SAE may result in extended ICU admissions and enduring cognitive deficits in patients (13). Current studies have shown that the destruction of the blood–brain barrier (BBB) is one of the core factors in the pathogenesis of SAE (14, 15). Individuals with sepsis exhibit a systemic inflammatory response that results in the release of endotoxins and proinflammatory cytokines. These substances stimulate brain endothelial cells and microglia, resulting in the downregulation of tight junction proteins and heightened blood–brain barrier permeability, permitting the ingress of inflammatory agents and toxins into the central nervous system, and inciting neuroinflammation and cerebral dysfunction (16, 17). The diagnosis of SAE remains primarily clinical, often relying on delirium screening tools such as the Confusion Assessment Method for the Intensive Care Unit (CAM-ICU) or neurological assessment scales such as the Glasgow Coma Scale (GCS) (18, 19). However, these methods are highly subjective and lack sensitivity, and early identification remains challenging. Timely diagnosis of SAE can help patients develop personalized treatment plans, avoid risk factors that aggravate SAE as much as possible, and implement neuroprotective measures.

With the development of precision medicine, researchers have paid more attention to the early identification and risk stratification of SAEs. Many studies have developed predictive models to support clinical decision-making (20). These diagnostic models typically combine clinical characteristics, such as vital signs, organ dysfunction scores, and laboratory indicators, and some studies have incorporated specific biomarkers to assess the risk of SAE in patients with sepsis (21, 22). For example, Zhang et al. (23) explored the application of neuron-specific enolase (NSE) and regional cerebral oxygen saturation (rSO₂%) in SAE patients and found that elevated levels of both indicators were significantly associated with decreased 28-day survival rate (p < 0.001), suggesting that they have certain value in SAE risk stratification. Mowafy et al. (24) confirmed in a prospective study that the optic nerve sheath diameter (ONSD) of SAE patients was significantly widened (the critical value was 5.2 mm), supporting ONSD as a non-invasive, bedside early screening method, and proposed a multimodal evaluation strategy integrating clinical examination, imaging and ultrasound, which is expected to optimize the SAE management process. In addition, several studies have employed either traditional logistic regression or advanced machine learning algorithms to develop visualized prediction models for the diagnosis of SAE (25, 26). However, none of these models or predictors have been widely implemented in clinical settings, and considerable variation exists in terms of model performance, predictor selection, and validation approaches. Furthermore, concerns regarding methodological rigor and risk of bias have limited confidence in the clinical utility of these models.

To date, there is still a lack of a comprehensive systematic review of the integration and quality assessment of existing SAE prediction models. This study aims to systematically review the current status of SAE prediction models, summarize their construction process, core variables and performance, use the internationally recognized PROBAST tool (27) to evaluate their methodological quality, and point out some common problems in current research and areas that may need to be optimized in the future. The innovation of this study is that it is the first time to comprehensively evaluate the modeling methods, performance differences and bias risks of SAE prediction models based on systematic review and meta-analysis methods, including the comparison of the interpretability of common algorithms such as logistic regression, random forest, and XGBoost, and explore issues such as model external validation, variable screening, model visualization and clinical transformation. We hope that this study will provide clinical medical staff with a systematic understanding of SAE prediction tools, and also provide certain theoretical support and methodological references for the subsequent development of higher-quality and popularizable SAE early identification models.

This study was conducted under the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020) guidelines (28). The study protocol was registered in PROSPERO (CRD420251062747).

SAE is one of the most common neurological complications of sepsis patients in the ICU and is closely related to the poor short-term and long-term prognosis of patients (38, 39). In this study, the incidence of SAE ranged from 15.0 to 62.4%, with a high overall level and significant differences. This difference may be related to the differences in data sources (US Critical Care Database and Chinese Single Center Hospital Data), patient population characteristics, and SAE assessment tools and diagnostic criteria. For example, the average age of the patients in the study by Wang et al. (34) was relatively low (57 years old), the prevalence of hypertension was 43.6%, the proportion of patients using vasopressors was 18.8%, and the incidence of SAEs was only 15.0%. In contrast, in the study by Jin et al. (36) based on the MIMIC database, the average age of the patients was 68 years old, the prevalence of hypertension was 63.6%, the proportion of patients using vasopressors was 64.2%, and the incidence of SAEs was as high as 62.1%. Patients of advanced age, those with comorbid chronic diseases, and those requiring critical intervention may have immune dysfunction and organ fragility, thereby increasing the risk of SAEs. In addition, the patients in the MIMIC database are mainly from the U.S. population, and their basic health conditions, access to medical care, and the intensity of critical care may be different from those of Chinese hospital patients, which may further affect the probability of SAE occurrence. Different diagnostic criteria used in different studies may also lead to systematic differences in incidence. Some studies use only CAM-ICU as a screening tool, while others combine CAM-ICU with GCS score, the latter of which may provide a more comprehensive assessment in a wider range of high-risk groups (40, 41). Although CAM-ICU can identify delirium, it is mainly used to assess delirium symptoms, which is only one of the clinical manifestations of SAE and fails to cover all SAE-related neurological dysfunctions comprehensively.

SAE has a serious adverse effect on the prognosis of patients with sepsis, but early identification and intervention still face great challenges. Commonly used screening tools (GCS and CAM-ICU) mainly rely on the subjective judgment of medical staff, and the evaluation results are susceptible to bias and lack objectivity and consistency (40, 42). With the deepening of SAE research, the construction of accurate prediction models has become a new research focus. However, among the 10 related studies included in this study, the predictive factors used in each model were quite different, and there is no unified combination of predictive variables. We found that some predictive factors such as age, SOFA score, blood sodium level, platelet count, mean arterial pressure (MAP), and neurobiomarker S100β were repeatedly found in multiple studies, suggesting that they may have stable predictive value in SAE prediction. To further explore the relationship between variable consistency and model performance, we compared the AUC performance of studies using high-frequency variables and similar modeling methods. For models built with logistic regression, the AUC of models that included high-frequency predictors (>2) was mostly higher (>0.8); in machine learning models, high-frequency predictors ranked higher in importance and had larger AUC values. Future studies need to further focus on and verify these potential key variables. However, we included relatively few studies, and the relevant results need to be further verified in more models and populations.

This study found that although the included studies differed in terms of modeling methods and variable selection, most models showed good discriminative performance on their respective data sets. Meta-analysis showed that the combined AUC of the model based on logistic regression was 0.85 (95% CI: 0.77–0.93), suggesting that the existing model has certain potential for clinical application. However, there was high heterogeneity among the studies (I² = 91.8%, p < 0.001), and the combined results need to be interpreted with caution. Sensitivity analysis by elimination method showed that heterogeneity may be mainly caused by the study of Jin et al. (36). This may be because the study used the MIMIC database population, which is different from other studies based on Chinese clinical data in terms of characteristic distribution, ethnic composition, modeling sample size, and clinical background of the study population. In addition, the diagnostic criteria for SAE have not been fully unified, and there are certain differences in the assessment tools, judgment time points, and discrimination criteria used in different studies (43). Despite the high heterogeneity, the publication bias assessment results (Egger’s test p > 0.05) did not indicate obvious bias. We evaluated the quality of the included models using the PROBAST tool and found that all models exhibited a high risk of bias, which limits their clinical applicability and generalizability. We also referred to the TRIPOD reporting standards and found that most studies did not report sufficiently on key aspects such as predictor selection, missing value processing, and model validation, which may also be an important reason for the high risk of bias. It is recommended that the construction of SAE prediction models in the future should be carried out as much as possible following the standardized development guidelines for prediction models to improve the reproducibility and horizontal comparison capabilities of the models (44).

The prediction models included in this study also have certain positive clinical significance. First, there is the significant role of specific biomarkers. The best-performing logistic model in Zhang et al. (37) combined with neurospecific indicators (cerebrospinal fluid BNIP3L and ONSD), made the model have high accuracy (AUC greater than 0.9) and good clinical benefits. Similarly, Mei et al. (25) used indicators such as cerebral circulation time, S100β, and middle cerebral artery PI to construct a logistic regression model, and the results showed that the model had clinical benefits in the range of 0–100%. This also suggests that compared with simple systemic indicators, combining direct measurement indicators of brain injury or cerebral perfusion can significantly improve the diagnostic accuracy of SAE. However, these studies also have certain problems: both studies were single-center with small sample sizes, and routine cerebrospinal fluid analysis and cranial ultrasound measurements in clinical ICUs are difficult. We also summarized some other routine clinical variables of SAE patients (such as age, SOFA score, MAP, and blood sodium level). These variables not only appear repeatedly in multiple studies but also have good accessibility and ease of application in clinical practice. Future optimization of prediction models should focus on combining special indicators related to brain injury with these routine clinical variables and further improve the accuracy and clinical practicality of early diagnosis of SAE by constructing a multi-dimensional and multi-level prediction framework. In addition, an important consideration is the choice between developing a machine learning model or a traditional logistic regression model. This study found that complex machine learning methods (such as XGBoost and neural networks) can achieve slightly better predictive performance than traditional logistic regression models in the MIMIC dataset. However, this advantage is not absolute. Zhang et al. (37) showed that in the clinical dataset they collected, the neural network was not better than the logistic regression model, and the latter still had good discrimination ability. In general, the selection of predictors and data quality may have a greater impact on model performance than the modeling algorithm itself (45). The logistic regression model has good interpretability and is easy to convert into a clinical evaluation tool, which is easy for clinical medical staff to understand and apply. It still has important clinical value at the current stage. Machine learning methods have unique advantages in dealing with complex feature interactions and nonlinear relationships and are particularly suitable for large-scale and diverse datasets (46). However, its interpretability is poor, and targeted algorithm adjustments are required. It lacks the trust of clinical medical staff, which affects its widespread application in actual clinical settings (47, 48). In the development of SAE prediction models, these two methods can be combined to improve model performance and ensure the interpretability and clinical scalability of the model (49).

All ten studies included in this review were published within the past 5 years, reflecting growing interest in the early prediction of SAE. However, the lack of high-quality external validation remains a key limitation. Future research should focus on developing large-scale, multicenter, prospective models with external validation to assess their stability and clinical utility (50, 51). The high-frequency risk factors summarized in this study can provide a reference for clinical work, and medical staff can strengthen the monitoring and management of high-risk patients in practice. It is important to note that some of the inherent risks identified by the model cannot be directly changed (such as patient age, chronic comorbidities, etc.), so the role of the prediction model focuses more on risk stratification and reminding clinical attention to high-risk groups. Although some of the included models reported higher AUC values (such as above 0.9), it should be pointed out that the improvement in discriminative ability does not mean that the model is reliable in actual clinical applications. All models have the risk of bias in the PROBAST assessment, especially in the analysis domain-related items (such as missing data processing methods and predictor screening strategies). In the future, the construction of SAE prediction models should focus on methodological rigor. We suggest that improvements can be made in the following aspects: (1) There are differences in the definition of SAE. The use of recognized standard definitions will help improve the verifiability and comparability of the model in different data sets; (2) Prospective data collection: More prospective studies are needed to ensure the uniformity of prediction data collection and standardize the outcome evaluation process; (3) Onset time processing: Current studies have not yet clarified which stage of the sepsis course the model is applicable to. In the future, dynamic variables (such as time series or continuous organ dysfunction trajectories) can be considered to improve the prediction ability of SAE, although this will increase the complexity of the model; (4) Variable screening methods urgently need to be standardized and transparent: At present, some studies still only use univariate analysis to screen predictive variables, which may miss important interaction terms or confounding factors, reducing the predictive ability of the model. It is recommended that future studies use more systematic and robust feature selection methods such as LASSO regression, recursive feature elimination, and Boruta algorithm; and (5) Transparency and reporting: Some studies did not clearly report the proportion of missing data, the processing method, and the final form of the model. In addition, the model evaluation indicators are still not comprehensive. Although AUC is a commonly used performance indicator, it is recommended to report indicators such as Brier score, calibration slope, and decision curve analysis (DCA) at the same time to more comprehensively reflect the predictive ability and clinical practicality of the model at different risk thresholds. Future studies should try to follow reporting standards such as TRIPOD to improve research quality and transparency.

This study has some limitations. The number of included studies is relatively limited, and there is high heterogeneity among the studies. The inclusion of only five studies in the meta-analysis may affect the robustness of the results. Although the data summarized in this study came from different regions, the authors of the included studies were all from China, which may have a certain regional bias.

This study evaluated a variety of SAE prediction models. The results showed that the existing SAE prediction models performed well in terms of discrimination ability, but they all had a high risk of bias and were not recommended for use in current clinical practice. In the future, in the process of model development, it is necessary to pay attention to the quality of the model, reduce the risk of bias, conduct high-quality external validation, and ensure the ease of interpretability of the model.