To cast a wide net in our search, we scoured both Chinese and English databases due to the vast population and widespread use of these languages. The databases under the microscope were the China National Knowledge Infrastructure (CNKI), Wanfang, the China Science and Technology Journal Database (VIP), SinoMed, PubMed, Web of Science, the Cochrane Library, CINAHL, and Embase. Our search was exhaustive, covering the birth of these databases up to September 30, 2024. We combed through with the following search terms: “Stroke,” “Cerebral stroke,” “patient discharge,” “disposition,” “outcome,” “destination,” “prediction model,” “Risk factors,” “Predictors” and “Risk score,” along with their respective variations. The nitty-gritty of our search tactics is detailed in Supplemental material. Moreover, we delved into reference lists and pertinent systematic reviews manually to uncover any pertinent studies that might have been overlooked for inclusion in our analysis.

In conducting our systematic review, we relied on the PICOTS system, which is endorsed by the Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modeling Studies (CHARMS) checklist (15). This system is instrumental in shaping the review’s objectives, the search methodology, as well as the criteria for including and excluding studies (16). The core components of our systematic review are outlined here:

P (Population): Stroke patients.

The inclusion criteria were: (1) Study subjects: Conforms to the diagnostic standards for stroke guidelines for therapy established by the Neurology Branch of the Chinese Medical Association (17), MRI or CT scans were used to diagnose stroke; aged ≥18 years; (2) Study types: include cohort studies, cross-sectional studies, and case–control studies; (3) Study content: focused on the construction and validation of risk prediction models for discharge disposition in stroke patients; (4) Predict the outcome: discharge to home or non-home setting; (5) Language type: Chinese or English textual form.

The exclusion criteria were: (1) Articles that identified risk factors but did not formulate a predictive model; (2) Study included fewer than two predictors in the model; (3) Studies with incomplete data or full text; (4) reviews, systematic reviews, and meta-analyses; (5) Conference papers.

Two writers (XCR and HL) individually imported all retrieved records into Zotero,¹ an open-source, free reference management software. The duplicate studies were then first removed. Second, the rest of papers were assessed at random based on their abstract and title to see if they qualified for inclusion. The papers were then further examined in full text in accordance with the inclusion and exclusion criteria, and more potentially suitable research were located by looking through the reference lists of all eligible studies. When the authors could not agree on which research were eligible, authors (XCR, HL, and XLJ) talked it out and came to an agreement.

Two authors (XCR and HL) individually extracted search results and checked with each other to minimize bias and ensure the consistency of collected data. According to the CHARMS checklist (Checklists for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modeling Studies) (15) used for systematic review data extraction of risk prediction models, the following data were extracted from the included studies: basic information included details such as author, publication year, research design, participants, data source, and sample size. Variable selection method, model building methodologies, validation kinds, performance measures, continuous variable handling techniques, final predicted variables, and presentation styles are all examples of model information. A third reviewer (XLJ) confirmed the information gathered during data extraction. To make sure they all agreed, the three researchers discussed and worked out any discrepancies. The two reviewers had extremely high consistency during the whole-text screening phase, with a Cohen’s kappa coefficient of 0.86 [Cohen’s kappa≥0.75 indicates extremely good consistency in the screening results (18)].

The quality of the included studies was assessed based on the TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis) guidelines (19), which offer a template for the transparent and comprehensive reporting of a multivariate predictive model for individual prognosis or diagnosis. Furthermore, the PROBAST (Predictive Model Risk of Bias Assessment Tool) was used to assess the risk of bias and applicability of the included studies in the present research on predictive modeling (20). The risk of bias was assessed in four domains: participants, predictors, outcomes, and analysis methods. Each domain comprised 2–9 questions, and the answers were scored as “Yes/Likely,” “No/Unlikely,” or “No Information,” representing low risk of bias, high risk of bias, or unclear risk of bias, respectively. The domain of risk of bias was judged as low risk if all questions were scored as low risk; if one or more questions were scored as unclear, the domain was judged as unclear; and if one or more questions were scored as high risk, the domain was judged as high risk. Applicability was assessed in three fields—“participants,” “predictors” and “outcome measures” (excluding the critical issues)—with criteria similar to those used for the bias assessment (21, 22). Two researchers (XCR and HL) used the PROBAST tool to independently assess the risk of bias and the applicability of the prediction models included in the present study. A third researcher (XLJ) was consulted for a conclusion in cases of disagreement.

Use Stata 17.0 software to perform analysis using the area under the curve (AUC) and 95% confidence interval (CI) as effect statistics. Employ the Q test and I² statistic to assess heterogeneity among multiple studies. If p > 0.1 or I² < 50%, it indicates low heterogeneity between studies, and a fixed-effects model is used for combined analysis (23); if p ≤ 0.1 or I² ≥ 50%, it suggests significant heterogeneity among studies, and a random-effects model is adopted, along with sensitivity analysis (24). To assess publication bias, Egger’s test and a funnel plot were employed. At α = 0.05, the significance level is established. There is less chance of publication bias when p > 0.05.

Ethical approval was unnecessary in this study, because it was a systematic review of existing articles, and no individual patient data were handled.

A total of 4,059 indexed documents were found in the first search. The titles and abstracts of 3,239 documents were evaluated after 820 duplicate records discovered in all databases were eliminated. Fifty six papers were included for additional review after this screening procedure. In the course of the following assessment, 15 studies were ruled out because they either did not create prediction models or only addressed risk factors. Furthermore, seven studies were deemed incompatible with the review’s target population, 10 studies included fewer than two predictors, and 11 studies were unable to provide necessary data. The PRISMA flow diagram, which depicts the thorough literature screening procedure and outcomes, is displayed in Figure 1.

The review’s studies were released between 2008 and 2024. Four of these were carried out in Japan, three in the United States, two in Australia, and one each in China, the United Kingdom, Switzerland, South Korea, and France. Regarding study design, six were prospective studies (including two multicenter studies), and eight were retrospective studies. The subjects consisted of two categories, with nine studies focused on stroke patients in the acute phase and five studies focusing on stroke patients undergoing rehabilitation in the sub-acute phase. Predicting discharge outcomes mainly involves two scenarios: home and non-home discharge. The former includes patients being discharged back to their own homes or the homes of relatives or friends, while the latter includes patients dying during hospitalization or being discharged to inpatient rehabilitation facilities, professional nursing facilities, nursing homes, secondary hospitals, rehabilitation centers, etc. The incidence of home discharge ranges from 15.1 to 85%, and the incidence of non-home discharge ranges from 15 to 84.9%. The number of people included in these studies varied, with sample sizes ranging from 296 to 74,425. Table 1 provides specifics.

A total of 22 prediction models were constructed across the 14 studies. Cui et al. (25) developed 6 models and before deciding on the best one for creating nomograms, Lensky et al. (26) developed 3 models, Brauer et al. (7) developed 2 models; 1 model was built in each of the other studies. In terms of variable selection, Cui et al. (25) (8.3%) optimal model used the Feature Importance analysis, Lensky et al. (26) (8.3%) applied principal component analysis, Itaya et al. (27) and Stineman et al. (28) (16.7%) applied the backward variable selection method, Brauer et al. (7) (8.3%) used a multinomial forward stepwise logistic regression, and the other 7 studies (58.3%) utilized univariate analysis to select variables. Regarding modeling methods, Cui et al. (25), Lensky et al. (26), Veerbeek et al. (29), and Berker et al. (30) employed machine learning method, such as the support vector machine (SVM), random forests (RF), naive bayes (NB), gradient boosting (GB), KNearest Neighbors (KNN) (25), AdaBoost, Bootstrap (26), and CART decision tree (29), while other 10 models used logistic regression (LR). The performance indicators mentioned in this research report include model Calibration, sensitivity, specificity, accuracy rate and AUC. For all these indicators, the closer the value is to 1.0, the better the performance. Thirteen models provided the area under the curve (AUC), which varied between 0.75 and 0.95; one model (31) reported the C statistic. Nine studies addressed calibration, and the most commonly used approach was the Hosmer-Lemeshow test. Only one study (25) compared machine learning algorithms with other classical statistical methods (including logistic regression). Cui et al. (25) compared the LR, SVM, RF, NB, GB, and KNN models using data from 523 stroke patients and found that the RF model was the best. Additionally, the AUC values of four of the machine learning algorithms were higher than that of the logistic regression model, while the KNN model was slightly lower than the LR model (LR: 0.85, SVM: 0.92; RF: 0.95; NB: 0.93; GB: 0.93; KNN: 0.84). Details are provided in Tables 2, 3.

The prediction models in the majority of the 14 investigations were validated either internally or externally, proving their generalizability and robustness. In particular, nine research were validated internally, while four studies were validated externally. Two articles were combined with internal and external validation, and two articles did not perform any validation after model construction. Regarding model presentation, 7 used risk scoring system, 5 used formula of risk score obtained by regression coefficient of each factor, and 1 studies used nomograms. The total amount of predictors that are included ranges from two to ten. The most often utilized factors in these models are age, the Functional Independence Measure (FIM) cognitive function score, the National Institutes of Health Stroke Scale (NIHSS) score at admission, and the FIM motor function score, which appear in six, five, five, and four studies, respectively. Other commonly used predictive indicators include the modified Rankin Scale (mRS) score at admission and stroke types, which appear in three studies. Barthel index (BI), Enteral feeding, pre-admission residence, marital status, paralysis degree, living alone and comorbidities, which appear in two studies. Additional details are available in Table 4.

The overall quality of the studies included in this analysis was rather excellent, covering more than 70% of the reporting items, according to our quality evaluation, which was based on the TRIPOD guidelines (19). Nevertheless, we also observed certain restrictions or ambiguities in specific reporting elements. For example, despite Lensky’s study (26) being comprehensive, it is deficient in providing sufficient details about its predictive model parameters and their applications. In their comprehensive report, Cho’s study (11) failed to define the parameters of their prediction model or the methodology for determining the sample size. Despite the fact that the TRIPOD rules offer us a reliable instrument for evaluating the quality of predictive models, this evaluation might not adequately capture all the information essential to predictive modeling because of the models’ complexity and diversity. For instance, sample size has a big impact on how accurate and robust models are, and choosing and modifying model parameters is essential for predictive power and model optimization. Detailed quality assessment in Supplementary Table S6.

Only five of the listed papers were suitable for synthesis because of inadequate reporting on the model development specifics. The validation model used a random effects model to calculate the combined AUC, which was 0.80 (95% CI [0.75–0.86]), indicating good overall predictive performance (Figure 3). The heterogeneity test revealed significant heterogeneity (p < 0.001, I² = 95.2%), and a random effects model was used. Changes in study design, sample size, predictor selection, and outcome definition may be the cause of the significant variability among studies. The findings’ robustness was demonstrated by the results’ stability even after any one study was eliminated. In addition, the sensitivity analysis confirmed the robustness of the results, with no single study altering the magnitude of the combined effect (Figure 4). The Egger’s test indicated no significant asymmetry (p = 0.757) (Figure 5). The funnel plot showed a roughly symmetrical distribution of the scatter points around the pooled effect size, with all points falling within the pseudo 95% confidence limits (Supplementary Figure 1), suggesting a low likelihood of publication bias. The meta-regression analysis identified the country of the participants (p = 0.5235), to be not a significant source of the heterogeneity.

Following a thorough screening and search of the literature, we found 14 original research. The risk prediction algorithms now in use for stroke patients’ discharge disposition are still in the early stages of development. All of these investigations included model construction and showed strong performance, with AUC values above the 0.7 threshold and ranging from 0.75 to 0.95. It is noteworthy that 2 studies (27, 29, 36) performed internal and external validation. This step is essential for assessing the generalization ability of the model, detecting overfitting, boosting prediction accuracy, and guaranteeing the dependability and significance of the findings (40). AUC values of 0.80 (95% CI [0.75–0.86]) were also noted for validation models, coupled with notable heterogeneity that was probably brought on by different demographic features, predictive factors, and methodology. Furthermore, most of the included studies presented their model results in the form of a risk-scoring system. The simplified scoring system is relatively simple and easy to understand clinically, and can achieve intuitive, convenient and effective individualized risk prediction, thus promoting the clinical application of healthcare professionals (41). Overall, these models have strong predictive power for discharge non-home disposition of stroke patients, showing favorable performance. However, the model’s clinical application was limited because all analyzed papers were deemed to have a high bias risk based on the PROBAST checklist.

First, the data sources are biased. Eight of the studies were retrospective analyses, which perhaps increased recollection bias and impacted the model’s quality. The majority of the data came from single-center research with small sample sizes. Multicenter, large sample, high-quality prospective studies should be given priority in future research in order to reduce recollection bias and enable the adaptation of prediction models to a larger patient population.

Second, there is also a bias in the model construction. (1) In terms of variable selection, univariate selection was one of the most commonly employed methods (58.3% of studies), and it ignored the interaction between variables and potential problems of collinearity, which may lead to inaccurate prediction results (42). (2) Eight research have explicitly eliminated missing data, and four studies have not reported missing values. The link between predictor factors and outcomes may become biased as a result of this approach; even in the absence of bias, missing data reduces precision and increases the confidence interval (43). (3) Ten research built their models using logistic regression, which limited their capacity to capture intricate interactions between variables and impacted the model’s accuracy and stability. We propose that more sophisticated variable screening techniques, including LASSO regression, which can address the issue of multicollinearity among variables and assist in identifying the most predictive variables (44), be used to increase the model’s stability and predictive power. In order to lessen the negative effects of missing data on statistical analysis and model stability, techniques like multiple imputation and single imputation should be applied when addressing missing data. Cui et al. (25), Lensky et al. (26), Veerbeek et al. (29), and Berker et al. (30) developed models using machine learning methods in this systematic review. Despite their potential to enhance prediction accuracy (45), machine learning algorithms showed limited advantages in this review. We assume that variables like random data set division, univariate analysis-based variable screening methods, and inadequate sample size may be connected to this phenomenon.

Finally, there are some limitations to the model validation. The predictive power of the model may be influenced by population and regional differences, thus highlighting the need for comprehensive validation during model development. For example, Kim et al. (32) study used a high sample size and a prospective cohort design, but it omitted internal validation. While random split validation is typically thought of as an internal validation method, it does not address the matter of model overfitting (46). Furthermore, only four studies performed external validation, and the remaining articles lacked such validation, limiting the generalization and applicability of the model (47). To increase the model’s generalizability, future research should place a strong emphasis on external validation, particularly across various geographies, racial groups, cultural contexts, and lifestyle characteristics. Different phases of the disease and variations in stroke types (such as hemorrhagic, mixed, and ischemic strokes) should also be taken into account. The degree of social support, frailty, and treatment strategy may also affect predictive performance. Taking all of these things into account will help to increase the model’s applicability and dependability.

Between two and ten predictors were included in the 22 models in this study, which were primarily divided into four groups: Sociodemographic factors: (e.g., age, marital status, residence before admission, lived alone), Clinical factors: (e.g., type of stroke, degree of paralysis, mRS, NIHSS, motor cognitive function score), Psychological factors: (e.g., depression), Comorbidities: (e.g., dyslipidemia, high blood pressure, acute heart attack, cerebral hemorrhage), Treatment: (e.g., mechanical ventilation, tube feeding, ICU admission). Some similarities were found despite differences in predictor selection brought about by research types and included variables. Age, NIHSS as well as FIM cognitive function and motor function scores at admission, are high-frequency predictive factors. Age was important predictor of institutional long-term care admission directly from the hospital after an acute stroke. Multiple studies (11, 25, 31, 33) have found a positive association between age and discharge non-home placement, where older patients may have more comorbidities with less support at home and may require further medical care and monitoring at institutions. Studies reported a strong correlation between non-home firing and initial neurological function as assessed by the NIHSS score (25, 28, 30, 31). High NIHSS scores are a common risk factor for neurological deficit after severe stroke and poor recovery after stroke. FIM motor and cognitive scores at admission showed significantly high score weights, with the former showing the highest predictive power in the model (30), a finding supported by Stineman et al. (28) Instead, Kim et al. (32) and Choi-Kwon et al. (11) state that cognitive FIM is the most important factor. This may be because patients with low cognitive ability often struggle to self-discharge or reunite with family, as cognitive and emotional disorders impose a greater burden on caregivers than physical disability. Further study is required to evaluate the effectiveness of the place where a patient gets referred at discharge, and discharge decisions should take into account patient-specific biopsychosocial characteristics that may take precedence over isolated outcomes of the outcome measures.

Therefore, in order to promptly identify high-risk patients, early screening should concentrate on these common variables. Nursing intervention is limited, though, as this study also revealed that many of the present factors are hard to directly alter. Future studies should look into incorporating controllable characteristics such stroke knowledge, psychological state and medication adherence in an effort to achieve individualized care interventions and prevent needless hospitalizations and resource misallocation.

Applying credible models helps to identify patients who require non-home discharge to higher levels of care, enabling early intervention plans. To guarantee the precision and legitimacy of the best outcomes, future research efforts should follow the TRIPOD guidelines. This means that investigators should prioritize the implementation of studies with a larger sample size and adopt a prospective study design. Numerous included research had difficulties with sample size, selecting predictors, and handling continuous variables. The application of sophisticated machine learning techniques in model construction can help with some of these issues. However, the current dearth of appropriate demonstration tools is a disadvantage of the machine learning models. Consequently, researchers need to select the right model creation techniques based on the particular circumstance. Although all of the models performed well overall, there was still a significant chance of bias. The number of events, processing of continuous variables, methods for selecting predictor variables, data complexity, model calibration and fitting, time interval between assessment of predictor variables and outcome determination, and study design (cohort or case–control study) all require improvement. Future research should concentrate on developing new models through larger, rigorously designed studies including multicenter external validations and improved reporting transparency, in order to increase the prediction models’ usefulness for evaluating discharge disposition risks in stroke patients. With the continuous progress in research methods and data processing technologies, the model will be continuously optimized and adjusted to better adapt to clinical needs, and thus provide a more accurate risk assessment tool in clinical practice.