Abstract: OBJECTIVE: Nowadays, machine learning algorithms are gradually being applied to predict stroke and cardiovascular disease. Compared with traditional regression models, machine learning can learn from data to achieve high prediction accuracy by exploring the flexible relationship between a large number of predictive features and outcome variables, providing a new method for the formulation of individualized treatment and rehabilitation programs. This study aims to systematically evaluate stroke functional recovery and prognosis prediction models based on machine learning, comprehensively assessing their predictive performance and clinical application potential to provide references for the development, application, and promotion of related predictive models. 
METHODS: This review was conducted following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines. Relevant literature on stroke prognosis prediction using machine learning methods was selected by searching PubMed, EMbase, Web of Science Core Collection, CNKI, WanFang, and the China Biomedical Literature Database, with the search period from January 1, 2014, to July 1, 2024. Two researchers independently screened the literature and extracted data based on inclusion and exclusion criteria, using the Prediction model Risk Of Bias ASsessment Tool (PROBAST) to assess model quality. 
RESULTS: (1) A total of 3 126 articles were obtained in the preliminary search. After screening and exclusion, 18 articles were finally included. 150 prediction models were constructed using 13 machine learning methods. The three most frequently used methods are Logistic Regression, Random Forest, and Extreme Gradient Boosting (XGBoost). Only one study was externally validated. Eight studies reported how the missing data were handled. (2) In terms of outcome indicators, 8 studies used the combination of clinical data and imaging data to build models, 9 studies only used clinical data to build models, and 1 study only used imaging data to build models. (3) Each of the 18 studies gave the most important characteristics of the study, with the most mentioned being the National Institute of Health Stroke Scale and age. All studies reported area under curve values ranging from 0.74 to 0.96, with the highest area under curve being 0.96. The overall risk of bias in all models was high. The high risk of bias in the field of model analysis was the main reason for the high risk of overall bias in all models. (4) The results of meta-analysis showed that age and National Institute of Health Stroke Scale score had significant influence on stroke prognosis, with age [MD=8.49, 95%CI(6.24, 10.75), P < 0.01] and National Institute of Health Stroke Scale score [MD=4.78, 95%CI(2.56, 7.00), P < 0.01].  
CONCLUSION: This study systematically evaluated the predictive model of functional recovery and prognosis of stroke based on machine learning, and all the models have good predictive potential. However, future studies should increase the sample size of the included model, adopt prospective studies, and add external validation of the model to improve the stability and prediction accuracy of the model, control the risk of bias, and contribute to the validation and promotion of the model in practical clinical applications. At the same time, the interpolation of missing values is more transparent and accurate. Although existing machine learning models show good predictive performance, it is also important to focus on the functionality and usability of the model, and the inclusion of features will reduce ease of use. We should develop easy to use model interfaces and user-friendly clinical tools to enable medical staff to better apply the model for clinical decision. 

Key words: machine learning, stroke, prognosis prediction, functional recovery, systematic review