This systematic review followed the CHecklist for critical Appraisal and data extraction for systematic Reviews of prediction Modeling Studies (CHARMS) checklist (8) and the PROBAST tool (9) for critical appraisal and risk of bias assessment. The protocol was prospectively registered in PROSPERO (CRD420251009136).

A comprehensive search was conducted across multiple databases, including Cochrane, EMBASE, PubMed and Web of Science. The search strategies were tailored to each database. Core search terms included “stroke,” “intensive care unit,” “mortality,” “prediction model,” and “risk.” The full search strategies for all databases are provided in Supplementary Table S1. This systematic review sought to identify studies published between January 2005 and January 2025 that focus on risk prediction models for short-term mortality in stroke patients admitted to the ICU. The inclusion criteria were based on the Participants, Interventions, Comparisons, Outcomes, Timing, and Setting (PICOTS) framework, as outlined below:

P (Population): Stroke patients admitted to the ICU.

I (Intervention): Risk prediction models for short-term mortality.

C (Comparator): Not applicable.

O (Outcome): Short-term mortality.

T (Timing): During hospitalization.

S (Setting): Stroke patients in the ICU only.

Inclusion criteria: (1) Studies involving hospitalized patients aged 16 years or older; (2) studies that developed or validated prediction models including at least two predictive variables; (3) studies in which the primary outcome was short-term mortality; (4) studies focusing on critically ill stroke patients admitted to the intensive care unit (ICU); (5) studies with observational designs, including prospective and retrospective cohort studies, as well as case–control studies, that developed or validated prediction models.

Exclusion criteria: (1) Studies without full-text availability; (2) conference abstracts, magazine articles, commentaries, opinion pieces, newsletters, and other forms of gray or secondary literature; (3) student dissertations; (4) articles containing erroneous data or methodological flaws; (5) non-English language publications.

Duplicate records were removed using NoteExpress. Two independent reviewers (Zhang Jiali and Li Hui) screened the titles and abstracts of retrieved studies based on the predefined inclusion and exclusion criteria. Any discrepancies were resolved through discussion or by consulting a third reviewer (Fu Yijie). After initial screening, the same two reviewers independently assessed the full texts of potentially eligible articles. Additionally, the reference lists of the included studies were manually searched to identify other relevant publications.

Data extraction was performed by two independent reviewers (Zhang Jiali and Li Hui) using a pre-designed extraction table. The extracted data included basic details such as authorship, publication year, country of origin, study design, and data source. For the prediction models, we collected detailed information on modeling methods, validation types, performance metrics, handling of missing data, and the predictors included in the final model. In studies presenting multiple models, we focused on the model with the highest AUC value during the model development phase. This review does not specifically focus on stroke patients with certain underlying conditions, such as diabetes or hypertension, and the included studies did not stratify based on comorbidities. In case of discrepancies, a third reviewer (Fu Yijie) was consulted to reach consensus.

The risk of bias (ROB) and applicability of the prediction models in the included studies were assessed using the PROBAST tool. This tool evaluates four domains: participants, predictors, outcomes, and analysis, with the first three domains also assessing applicability. Each item is rated as “yes,” “probably yes,” “no,” “probably no,” or “no information.” A domain is considered to have a high risk of bias if at least one item is rated as “no” or “probably no.” If one or more domains are rated as unclear while the others are rated as low risk, the overall ROB is considered unclear. An overall low ROB requires all domains to be rated as low risk. Two independent reviewers (Zhang Jiali and Li Hui) performed the quality assessment using PROBAST. Any discrepancies were resolved by consulting a third reviewer (Fu Yijie).

The meta-analysis was conducted using Stata software, version 18. Standard errors for the AUC values were automatically calculated by Stata during the meta-analysis process using the built-in meta-analysis commands. Given the high degree of heterogeneity observed across the included studies, a random-effects modelwas applied to the meta-analysis. To explore potential sources of heterogeneity, we conducted subgroup analyses based on stroke type, geographic region, and modeling approach. The discriminatory ability of the models was assessed using Area Under the Curve (AUC) from the Receiver Operating Characteristic (ROC) analysis. A higher AUC value (closer to 1.0) indicates better discriminatory power of the model. Additionally, a sensitivity analysis was performed by sequentially excluding individual studies to assess the robustness of the pooled AUC estimates.

A comprehensive search across multiple databases (Cochrane, EMBASE, PubMed, and Web of Science) identified a total of 6,874 records. After removing duplicates (n = 20), 6,854 records were screened for title and abstract evaluation. Ultimately, 12 studies meeting the inclusion criteria were included in this review (10–21). Figure 1 presents the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 Flow Diagram detailing the process of literature identification, screening, and selection.

This systematic review included 12 studies that investigated risk prediction models for short-term mortality in ICU stroke patients (10–21). These studies, published between 2022 and 2024, were conducted in the United States, the Netherlands, Brazil, and France. All studies employed a retrospective study design and derived data from existing databases, with most extracting data on the first day of ICU admission or within 24 h (10–13, 15, 19–21). However, four studies did not specify the timing of data extraction (14, 16–18). The total sample size across all studies was 45,939 participants, with individual sample sizes ranging from 236 to 16,592. The highest mortality rate, 41%, was observed in studies focusing on hemorrhagic stroke (HS) (13). Stroke diagnoses were primarily based on the International Classification of Diseases, Ninth and Tenth Revisions (ICD-9 and ICD-10) (10–12, 14–17, 20, 21). Detailed study characteristics are summarized in Table 1.

Table 2 presents the characteristics of the prediction models used in the included studies. The sample sizes used solely for model development varied considerably, ranging from a minimum of 341 to a maximum of 13,274. The modeling approaches primarily consisted of traditional statistical models and machine learning methods. Studies addressing missing data exclusively employed imputation methods (10–12, 14–16, 18–20), while some did not report their approach to handling missing data (13, 17, 21). All studies conducted internal validation; however, only four performed external validation (10, 14, 16, 17), with relatively limited external validation results. Various calibration methods were employed, with calibration curves and Brier scores being the most commonly used. The reported AUC values during model development ranged from 0.761 to 0.977. A total of 7 machine learning models were developed (10, 14–18, 20), with the Multilayer Perceptron (MLP) model achieving an AUC of 0.977 and the Random Forest (RF) model reaching 0.90 (17, 18). These machine learning models demonstrated relatively higher AUC values compared to traditional statistical models, such as logistic regression and LASSO regression, with the latter achieving an AUC of approximately 0.795 (19). The final presentation formats of the models included nomograms and web-based calculators (10–12, 16, 17, 19, 21). These characteristics of the prediction models are detailed in Table 2.

The number of candidate predictors in the models ranged from 20 to 73, with final models containing 5–20 predictors. The predictive factors were categorized into five major groups: demographic and medical history variables, physiological and laboratory indicators, scoring systems and disease severity measures, key clinical interventions, and admission-and nursing-related factors. The four most frequently included predictors were the Glasgow Coma Scale (GCS) (n = 9), age (n = 9), white blood cell count (WBC) (n = 8), and glucose (n = 7), as detailed in Table 3.

We assessed the risk of bias (ROB) and applicability of all included prediction models using the PROBAST tool. Our analysis revealed that all models exhibited a high risk of bias and low applicability. Detailed evaluation results are provided in Table 4, and the risk of bias is illustrated in Figure 2. Since all 12 studies employed retrospective designs and data were sourced from databases, the participant domain for all studies was rated as “high risk” (10–21). In the predictor domain, 11 studies included predictors that required subjective interpretation or were judged by different individuals with varying levels of experience (10–13, 15–21), such as the Glasgow Coma Scale (GCS) and the Charlson Comorbidity Index. Additionally, retrospective studies may involve inconsistent collection of data over time, which can impact model accuracy. Furthermore, these studies are more prone to recall bias, which could lead to the assessment of predictors with knowledge of the outcome. Therefore, all studies were rated as having a high ROB in the predictor domain. In the outcome domain, all outcome data were sourced from databases, and the use of blinding was not explicitly stated. Ten studies could not confirm whether outcomes were determined without knowledge of the predictor variables (10–12, 14–18, 20, 21). Two studies were rated as “high risk” because they included predictors directly related to the outcome definition (13, 19), the collection time of the predictors was not clearly stated in four studies (14, 16–18). Therefore, two studies were rated as high risk in the outcome domain (13, 19), while 10 studies were rated as “unclear” regarding the outcome assessment (10–12, 14–18, 20, 21). In the analysis domain, seven studies were rated as “high risk” due to sample size, the number of predictors, or improper handling of variables (10, 14, 15, 17–19, 21), while five studies were rated as “unclear” due to insufficient data handling (11–13, 16, 20). Specifically, two studies did not report the number of event outcomes used for model validation (10, 13), and six studies did not clearly specify the number of candidate predictors (11, 12, 16, 17, 19, 21). Two studies had fewer than 20 outcome events per predictor (14, 15), and one study categorized continuous variables during analysis (10). Furthermore, nine studies did not specify whether continuous and categorical predictors were appropriately handled (11–13, 15–18, 20, 21). Additionally, three studies used univariate analysis to select predictors (17, 19, 21), and one study discretized all continuous variables (10). Regarding model performance evaluation, one study employed an insufficient evaluation method and did not mention any calibration methods (15). All 12 studies were rated as having low applicability. In the participant domain, all studies were rated as having low applicability. Two studies focused on special populations: one targeting the elderly (12), and another focusing on mechanically ventilated patients with aneurysmal subarachnoid hemorrhage (aSAH) (19). The remaining 10 studies (10, 11, 13–18, 20, 21), being retrospective, were limited by data completeness, with missing key variables, and potential biases in the definition and selection of participants, resulting in low applicability. In the predictor domain, some predictors, such as the presence of metastatic solid tumors, may not be universally available in all hospitals, raising concerns about applicability (11, 12, 21). Additionally, in the remaining nine studies, retrospective data limitations led to incomplete recording of key variables at all-time points, further impacting the applicability of the predictors. In the outcome domain, since all studies rely on retrospective database data, the data sources may have certain limitations, which could restrict their external applicability.

Due to the lack of detailed AUC confidence interval values in some studies, only eight models from six articles were included in the meta-analysis (11–13, 16, 19, 20). The study by Mariëlle Kvan Valburg includes models for two distinct study populations (13), while Yuxin Wang’s study includes models for two different prediction time points (20), resulting in eight models from six articles. The pooled analysis showed an AUC of 0.82 (95% CI: 0.80–0.85), indicating that these models demonstrate good discriminative ability in predicting short-term mortality risk in ICU stroke patients (Figure 3). However, heterogeneity analysis revealed significant variability between the studies (I² = 80.1%, p = 0.000). To further explore the potential sources of heterogeneity, we conducted subgroup analyses based on stroke type, modeling methods, and geographical region (Figure 4). Specifically, studies were categorized into ischemic stroke (3 models) and hemorrhagic stroke (4 models); traditional modeling methods (5 models) and machine learning methods (3 models); as well as studies conducted in America (5 models) and Europe (3 models). The subgroup analyses revealed varying degrees of heterogeneity within each subgroup. For instance, lower heterogeneity was observed in studies using traditional modeling methods (I² = 54.9%) and those conducted in Europe (I² = 28.7%), compared to higher heterogeneity in machine learning-based models (I² = 90.9%) and studies from America (I² = 83.1%). However, the between-group differences were not statistically significant (p > 0.05), indicating that stroke type, modeling method, and geographical region did not fully account for the observed heterogeneity. To assess the robustness of our findings, we performed a sensitivity analysis by sequentially excluding each study (Figure 5). The pooled AUC values remained stable, ranging from 0.82 to 0.85, indicating that the overall estimate is robust and reliable.

(1) Due to the lack of AUC values for internal and external validation, we used the AUC values from model construction for the meta-analysis. (2) The lack of AUC confidence interval values for some models resulted in the inclusion of only 8 models from 6 studies, potentially affecting the meta-analysis results. (3) The absence of key performance metrics such as sensitivity, specificity, and accuracy limited the comprehensiveness of the meta-analysis. (4) Only English-language studies were included, which may have excluded relevant research published in other languages, potentially affecting the comprehensiveness of the results.