Abstract

Objective:

This study aimed to systematically evaluate risk prediction models for venous thromboembolism (VTE) in stroke patients and to provide a reference for future model development and clinical research.

Methods:

A systematic search was conducted across multiple databases, including PubMed, Embase, Web of Science, the Cochrane Library, China National Knowledge Infrastructure (CNKI), Wanfang Database, VIP, and SinoMed, to identify studies on VTE risk prediction models in stroke patients. Databases were searched from inception to September 1, 2025. Risk of bias and applicability of the prediction models were assessed using the PROBAST checklist. Meta-analyses were conducted to estimate pooled VTE incidence and the area under the curve (AUC) for model performance using Stata 17.0.

Results:

A total of 2,726 records were retrieved, and seven prediction models were included. Reported VTE incidence ranged from 9.8% to 38.9%, with AUC values between 0.781 and 0.978, indicating moderate to high apparent discriminative performance, however, this should be interpreted with caution in light of the high risk of bias and lack of external validation. None of the seven included models underwent independent external validation, and according to PROBAST, all included models were judged to be at high overall risk of bias. The pooled VTE incidence was 20.8% (95% CI: 14.7%–27.0%), and the pooled AUC across six models was 0.87 (95% CI: 0.81–0.93).

Conclusion:

Among the VTE risk prediction models for stroke patients included in this study, although some models demonstrated favorable predictive performance, all models were judged to be at high risk of bias according to PROBAST. Future research should prioritize the validation and refinement of existing models or the development of new models with more rigorous methodological design, in order to better support clinical decision-making for patients with stroke.

Systematic Review Registration:

https://www.crd.york.ac.uk/prospero/display_record.php?ID=CRD42024603132, PROSPERO CRD42024603132.

1 Introduction

Stroke is a rapidly developing focal or global brain dysfunction lasting more than 24 h or resulting in death (1–4). Stroke is one of the leading causes of death and long-term disability worldwide (5), particularly in low- and middle-income countries (6). There are approximately 10.3 million new cases of stroke worldwide each year (7). In China, stroke has become a major chronic non-communicable disease that poses a serious public health challenge. It is the leading cause of death and disability among adults and is characterized by high incidence, high disability, high mortality, high recurrence, and a substantial economic burden (8).

With continuous advances in stroke treatment technologies and significant improvements in the quality of in-hospital care, stroke-related mortality has declined (9). However, despite these advancements, stroke patients are still prone to secondary complications (10) instance, venous thromboembolism (VTE), including deep vein thrombosis (DVT) and pulmonary embolism (PE), is a frequent yet underdiagnosed complication of stroke characterized by insidious and non-specific symptoms (11). Stroke patients who experience VTE face a significantly increased risk of poor prognosis and death within three months (12). Beyond its impact on patient outcomes, VTE also imposes a considerable burden on healthcare systems, as it is frequently associated with hospitalization, prolonged length of stay, and increased healthcare resource utilization (13), early identification of high-risk patients and timely implementation of preventive interventions are crucial for reducing the incidence of VTE, improving clinical outcomes, and enhancing quality of life in stroke survivors.

Prediction models estimate the probability of the occurrence of a specific outcome by integrating multiple predictors (14). They are widely applied in clinical practice and public health decision-making (15). Prediction models enable clinicians to identify high-risk individuals early and to implement tailored preventive interventions according to model-based risk stratification, thereby improving clinical outcomes (16). In recent years, researchers have developed several VTE prediction models specifically for stroke patients (17–23). However, although several studies have explored VTE risk factors or individual prediction models in stroke patients, systematic comparative evaluations of these models in terms of modeling methods, predictive performance, and sample-related bias remain limited, particularly with respect to recently developed models and those derived from Chinese populations. Therefore, it remains unclear which model is more suitable for stroke patients. Therefore, this study aims to systematically evaluate VTE prediction models for stroke patients both in China and internationally, to provide a reference for future model development and clinical application.

2 Methods2.1 Design

We reported this review according to the guidelines of the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) (24). This study was registered prospectively in PROSPERO (CRD42024578643).

2.2 Retrieval strategy

The databases searched included PubMed, Embase, Web of Science, Cochrane Library, CNKI, Wanfang Database, VIP, and SinoMed. The search period spanned from the inception of each database to September 1, 2025, using the following keywords: “stroke”, “venous thromboembolism”, “risk prediction model”, “risk score”, “risk factors”, “predictors”, and “nomogram”, and “model”.

2.3 Eligibility criteria

The inclusion criteria for this study were developed according to the PICOTS (25) framework and detailed in Table 1. Studies that were limited to VTE subgroups, as well as reviews, case reports, and other informally published or duplicated studies, were excluded. Additionally, studies for which the full text was unavailable or for which the model details were incomplete were also excluded.

PICOTS frameworkDefinitionPopulationPatients aged ≥18 who meet the diagnostic criteria for stroke.InterventionPublished risk prediction models for VTE in stroke patients incorporating at least two predictors.ComparatorNot applicable.OutcomeOutcomes focused specifically on VTE rather than its subgroups.TimingTiming involves predicting results based on the patient's baseline information, clinical scoring scales, and laboratory test results.SettingSettings focus on personalized prediction for stroke patients, facilitating early identification of high-risk VTE groups and enabling the implementation of targeted preventive measures.

Criteria for study inclusion in the systematic review.

2.4 Literature screening and data extraction

Two researchers trained in evidence-based medicine independently screened the literature and extracted data based on the inclusion and exclusion criteria. Disagreements were resolved through discussion with a third reviewer (26). The data collection table was developed based on the CHARMS checklist for critical appraisal and data extraction in systematic reviews of prediction modeling studies (27), and includes the following: (1) basic characteristics of the included studies (author, publication date, country, data source, and sample size), (2) establishment of the prediction model (modeling method, number of models, handling of missing data, treatment of continuous variables, number of predictors, and factors included), (3) performance and presentation of the prediction model (model performance, model validation, and model presentation), and (4) quality evaluation of the prediction model (research subjects, predictors, results, and statistical analysis).

2.5 Quality evaluation

Two researchers independently used the prediction model risk of bias assessment tool (PROBAST) (28) to assess the risk of bias and applicability of the included studies. Discrepancies in the evaluation results were resolved through discussion with the third party.

When multiple predictive models are developed using different methodologies within a single study, the model recommended by the authors as the best is prioritized for evaluation. If no model is explicitly defined, the one with the highest predictive performance is selected for assessment. The risk of bias assessment evaluates four key areas: research subjects, predictors, results, and statistical analysis, consisting of 20 questions in total. Each answer was as “yes/maybe”, “no/maybe not”, or “no information”. The evaluation results for each area were classified into three categories: low, high, or unclear risk of bias. If all four areas indicated low risk, the study is considered to have a low risk of bias. A high risk in any area, or a lack of external validation despite low risk in all areas, results in a high overall risk of bias. If any area has an unclear result, while the others showed a low risk of bias, the study is classified as having unclear overall risk of bias. The applicability evaluation focuses on three areas: research subjects, predictors, and results, and follows a process similar to the risk of bias evaluation.

2.6 Statistical analysis

Meta-analysis was performed using Stata 17.0 to evaluate the incidence of VTE in stroke patients and the area under the curve (AUC) of the prediction models. Pooling of AUC values was conducted to provide an overall summary of the discriminative performance reported in the included studies. The pooled AUC was interpreted as a descriptive rather than inferential estimate and was not intended for direct comparison between models. Heterogeneity was assessed using the I2 statistic and the Cochrane Q test. The I2 statistic was used to quantify heterogeneity, with values of 25%, 50%, and 75% representing low, moderate, and high heterogeneity, respectively. A fixed-effects or random-effects model was selected according to the degree of heterogeneity (29).

3 Results3.1 Literature screening process and results

A search of the relevant databases initially identified 2,726 articles, of which 2,516 remained after removing 210 duplicates. These articles were then screened according to predefined inclusion and exclusion criteria. Initially, 2,470 articles were excluded following a review of the title and abstract, and 46 articles remained and were evaluated in full text. Of these, 23 articles focused exclusively on outcomes related to the VTE subgroup, 9 articles addressed risk factors without developing prediction models, 3 articles included fewer than two predictors, and 4 articles were inaccessible. Consequently, 39 articles were excluded. Finally, 7 articles were included in this study, a total of 7 studies were included (17–23). The literature screening process and results are depicted in Figure 1.

PRISMA flow diagram of the literature search and study selection process.

3.2 Basic characteristics of the included literature

The seven studies included in this analysis were published between 2022 and 2025, all using a retrospective design. Five of them are Chinese articles (17, 19, 20, 22, 23), and the other two are English articles (18, 21), respectively. One study was sourced from the MIMIC-IV database (18), while another utilized data from the Henan Stroke Cohort (19). We also found that five studies focused on stroke patients in general (17–20, 22), the other two studies specifically addressed ischemic stroke (23) and hemorrhagic stroke (21). The total sample size ranged from 185 to 675, with the number of VTE cases among stroke patients varying from 26 to 164, with an incidence rate between 9.8% and 38.9%. The basic characteristics of the included studies were summarized in Table 2.

Author (year)ParticipantsData sourceCases/sample size (%)Su Wei (2025)aStroke PatientsThe neurology department of a general hospital (2022–2024)72/185 (38.9%)Folin Lan (2024)Stroke PatientsMIMIC-IV Database (2008–2019)26/266 (9.8%)Wang Rongronga (2023)Stroke PatientsHenan Stroke Cohort (2019–2021)164/675 (24.3%)Hu Yonghuana (2023)Stroke patients aged ≥ 18 years and hospitalized for ≥ 48 hHIS system of a tertiary hospital (2021–2022)49/409 (11.98%)Liu Jiea (2022)Acute stroke patients aged ≥ 18 and hospitalized within 24 h after onset.EMR of one hospital (2018–2020)38/215 (17.67%)Cao Huaa (2022)Acute ischemic stroke patients aged ≥ 18 years and diagnosed within 72 h after onsetThe neurology department of a general hospital (2019–2020)55/238 (23.1%)Liu Shucheng (2022)Primary ICH patients aged ≥ 18 years who undergo screening within 24 h of onsetNeurosurgery intensive care unit of a certain hospital (2019–2021)83/369 (22.5%)

Overview and characteristics of the included studies.

a

This study was published in Chinese.

3.3 Construction of prediction model

A total of nine VTE risk prediction models for stroke patients were developed across seven studies. One study developed two prediction models using logistic regression and random forest (19), while the remaining six studies used logistic regression to develop their models (17, 18, 20–23). In addressing missing data, the studies employed varying approaches. One study utilized the random forest method for datasets with a missing rate below 20% (18), while another applied multiple imputation methods (19). Four studies employed complete case analysis (17, 20, 22, 23), and one study did not specify whether missing data were present or how they were handled (21). Regarding the treatment of continuous variables, methodological differences were also apparent. Four studies preserved the continuous nature of these variables (17, 18, 21, 23), whereas the remaining three studies converted all continuous variables into categorical variables (19, 20, 22). A comprehensive summary of the prediction models is presented in Table 3.

Author (year)Model development methodModel numbersMissing data handlingContinuous variables processing methodFinal predictorsSu Wei (2025)LR1Complete-case analysisMaintain continuityIntracranial hemorrhage volume, GCS score, FIB, CRP, D-dimerFolin Lan (2024)LR1Exclude if missing >20%, otherwise RF imputationMaintain continuityLog-formed D-dimer, PTT, Lung infection, MCHWang Rongrong (2023)LR, RF2Multiple imputationCategorical variablesAge, ADL, Muscle strength, UA, Length of stay, Fib, TC, D-dimer,Hu Yonghuan (2023)LR1Complete-case analysisCategorical variablesAge, History of VTE, ADL, D-dimerLiu Jie (2022)LR1Complete-case analysisCategorical variablesHistory of VTE, D-dimer, Recent surgical history, Bed rest time, Central venous catheterization, Prophylactic anticoagulationCao Hua (2022)LR1Complete-case analysisMaintain continuityCRP, D-dimerLiu Shucheng (2022)LR1NAMaintain continuityGCS score, NIHSS score, D-dimer

Basic information of the prediction models.

LR, logistic regression, RF, random forest, NA, not reported, PTT, partial thromboplastin time, MCH, mean corpuscular hemoglobin, ADL, activity of daily living, UA, uric acid, FIb, fibrinogen, TC, total cholesterol, CRP, C-reactive protein, GCS, Glasgow Coma Scale, NIHSS, National Institute of Health stroke scale.

3.4 Performance of prediction models

This study analyzed AUC values for the included models, ranging from 0.78 to 0.98, with sensitivity ranging from 70% to 94.4%, and specificity ranging from 67.35% to 91.6%. All seven studies reported both discrimination and calibration metrics. Among them, five studies assessed calibration using the Hosmer–Lemeshow test, and five studies reported calibration curves. However, more informative measures of calibration, such as calibration slope, intercept, or Brier score, were not reported in any of the included studies. In terms of model validation, one study did not conduct internal validation (22). The other six studies used internal validation using random split validation, k-fold cross-validation, bootstrap repeated sampling (17–21, 23). None of the seven studies performed external validation. For model presentation, six studies primarily utilized nomogram score analysis to illustrate their predictions (17–21, 23), while one study proposed a risk score formula (22). A detailed summary of the performance and presentation formats of the prediction models is provided in Table 4.

Author (year)Model performanceCalibration methodValidation methodologyModel presentationSu Wei (2025)A: 0.98 (0.94–0.99)CCRandom splitNomogramB: 0.96 (0.89–0.99)Folin Lan (2024)A: 0.88H-L test, CCRandom splitNomogramB: 0.88Wang Rongrong (2023)A: 0.92 (0.90–0.95)H-L testRandom split 5-fold crossNomogramB: 0.90 (0.85–0.95)Hu Yonghuan (2023)A: 0.78 (0.71–0.85)H-L test, CCBootstrapNomogramB: 0.74 (0.67–0.82)Liu Jie (2022)A: 0.81 (0.73–0.89)H-L test, CCNoneRisk score formulaCao Hua (2022)A: 0.90 (0.85–0.96)CCBootstrapNomogramLiu Shucheng (2022)A: 0.79 (0.74–0.85)H-L test, CCBootstrapNomogram

Performance and presentation formats of the prediction models.

AUC, Area under the curve (0.5–0.7 as poor discrimination, 0.7–0.8 as moderate discrimination, 0.8–0.9 as good discrimination, and 0.9–1.0 as excellent discrimination), A, development cohort, B, validation cohort, H-L test, Hosmer-Lemeshow Goodness of fit test, CC, calibration curve. The study by Folin Lan (2024) was excluded from the AUC meta-analysis because confidence intervals were not reported.

3.5 Literature quality evaluation3.5.1 Risk assessment of bias

All seven studies were assessed with a high overall risk of bias. Primarily, due to the retrospective nature of the studies, it may introduce recall bias. Additionally, the data in these studies were not originally collected with the intent of developing or validating prediction models. Consequently, key predictors related to VTE in stroke patients may not have been fully documented in the cases. In the area of predictors, five studies were assessed to have a low risk of bias (17, 20–23), whereas two were rated as an unclear risk of bias (18, 19). This uncertainty arose primarily from the lack of reported quality control during the collection and evaluation of predictors, leading to inconsistencies in both timing and method of data acquisition. Furthermore, the absence of standardized training or guidelines for raters may have further contributed to variability and potential bias. In the assessment of results, six studies were rated as having a low risk of bias. However, one study (18) failed to specify the criteria used to define outcomes, raising concerns about the reliability and reproducibility of its findings.

In the area of statistical analysis, all seven studies exhibited a high risk of bias. Importantly, the high overall risk of bias observed in the included studies was primarily driven by deficiencies in the analysis domain, rather than the absence of external validation alone. Several methodological shortcomings were identified, and the main reasons included that six studies were unable to satisfy the recommended number of events per variable (EPV) ≥20 for each independent variable (17, 18, 20–23), three studies transformed all continuous variables into categorical variables (19, 20, 22), one study did not disclose the method used to handle missing data (21), and six studies relied on univariate analysis to screen predictors. Additionally, two studies did not address the key issues such as overfitting, underfitting, or optimal fitting during their modeling prediction (18, 22), and two studies failed to report the coefficients of predictors within their models (18, 21). The PROBAST assessment results were summarized in Table 5.

Author (year)Study typeROBApplicabilityOverallParticipantsPredictorsOutcomeAnalysisParticipantsPredictorsOutcomeRisk of BiasApplicabilitySu Wei (2025)B−++−+++−+Folin Lan (2024)B−??−+++−+Wang Rongrong (2023)B−?+−+++−+Hu Yonghuan (2023)B−++−+++−+Liu Jie (2022)A−++−+++−+Cao Hua (2022)B−++−−++−−Liu Shucheng (2022)B−++−−++−−

PROBAST results for the included studies.

PROBAST, Prediction model Risk Of Bias Assessment Tool, ROB, risk of bias.

A, indicates “development only”; B, indicates “development and validation in the same publication”; +, indicates low ROB/low concern regarding applicability; −, indicates high ROB/high concern regarding application;?, indicates unclear ROB/unclear concern regarding applicability.

3.5.2 Applicability risk assessment

Five studies demonstrated a low risk of applicability concerns (17–20, 22), whereas two studies (21, 23) were classified as having a high risk of applicability. The high risk in these studies stemmed from their limited focus. The two studies focused on hemorrhagic stroke (21) and acute ischemic stroke (23), respectively. Consequently, the present predictive models developed in these studies are lacking in generalizability to the broader population of stroke patients.

3.6 Meta-analysis results

A meta-analysis was conducted to estimate the incidence of VTE in stroke patients across the seven included studies. Substantial heterogeneity was observed among the studies (I2 = 93.4%, Cochran's Q = 84.91, df = 6, P < 0.001).

Sensitivity analysis using a leave-one-out approach demonstrated that exclusion of any single study did not materially change the pooled estimate, indicating the robustness of the overall result. However, heterogeneity remained consistently high across all iterations. Due to the limited number of included studies (n = 7), formal subgroup analyses were not performed, as this would result in an insufficient number of studies per subgroup and reduce statistical reliability. Given the substantial between-study variability, a random-effects model was applied to account for heterogeneity. The pooled incidence of VTE in stroke patients was 20.8% (95% CI: 14.7%–27.0%), as shown in Figure 2. Forest plot of pooled VTE incidence in stroke patients. The horizontal axis represents the proportion of VTE incidence. Weights were assigned using the DerSimonian–Laird random-effects model.

Forest plot of pooled VTE incidence in stroke patients.

Subsequently, a meta-analysis was conducted to summarize the AUC values reported for the prediction models. One model was excluded due to the absence of a reported 95% confidence interval (18), and six models were included in the analysis (17, 19–23). Among these, one study developed two models based on the same dataset, and one model was selected to avoid duplication (19). Substantial heterogeneity was observed (I2 = 92.1%, Cochran's Q = 58.29, df = 5, P < 0.001). Given the substantial between-study variability, a random-effects model was used, yielding a pooled AUC of 0.87 (95% CI: 0.81–0.93), as shown in Figure 3. Forest plot of pooled AUC values for VTE risk prediction models in stroke patients. The horizontal axis represents AUC values. Weights were assigned using the DerSimonian–Laird random-effects model.