Stroke is a serious condition and a leading cause of death and long-term disability, which places a huge burden worldwide (1). Post-stroke cognitive impairment (PSCI) is a prevalent prognosis and cause of death following a stroke. Stroke patients have a higher incidence of 1-year cognitive impairment than non-stroke populations (2, 3). As society and economy progress, more emphasis is placed on disease and health, especially cognitive impairment. Early identification and diagnosis of PSCI, as well as early prophylaxis and treatment, can help improve stroke patient’s prognosis and reduce social and economic burdens.

Clinical tools for early PSCI diagnosis in stroke patients are currently lacking. Researchers have tried to apply existing cognitive impairment risk prediction models constructed based on the general population to the prediction of PSCI, but their predictive performance was not ideal in stroke patients (4). As a result, researchers have shifted their focus to machine learning (ML) in the hopes of developing more accurate PSCI prediction models. ML is an emerging field in medicine that utilizes computer science and statistics to solve healthcare problems (5). In recent years, ML has been increasingly applied to stroke research, and it was shown that ML-based stroke image prediction can outperform existing prediction tools (6). However, the diversity in mathematical modeling and sensitivity of ML algorithms to factors such as patient sampling, missing data and sample size continue to fuel debates over the accuracy of these models in disease prediction.

The performance of existing stroke prediction models has been inconsistent due to the use of different types of ML (e.g., logistic regression or other alternative) and modeling variables. In these predictive models, there are differences in the types of machine learning utilized, with most researchers using logistic regression while some may consider it lacking and opt for alternative models. Furthermore, we note discrepancies in the selection of modeling variables, which ultimately contributes to the uncertainty of their results. Unfortunately, evidence-based studies investigating the efficiency of ML in the prediction of PSCI are still lacking. As a result, the aim of this study is to examine the predictive accuracy of ML in PSCI and comprehensively summarize the modeling variables included in this prediction model in order to provide a useful reference for the subsequent development of simple clinical prediction tools.

This study was conducted following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines (Supplementary material) (7).

This study has been registered in PROSPERO (CRD42022383476).

Our meta-analysis of 21 original studies demonstrated that ML may be an ideal tool for predicting PSCI. The training set had a c-index of 0.82 (95% CI 0.77–0.87) and sensitivity and specificity of >70%, indicating considerable predictive accuracy in PSCI. Furthermore, the accuracy of the validation set was not significantly lower than that of the training set, indicating that the ML model has good applicability. Currently, LR is the preferred model in clinical practice because it is simple for generating highly accessible nomograms, such as the nomogram on lymph node metastasis developed by the Sloan-Kettering Cancer Center (39–41). In our study, LR was also the preferred model among researchers as it exhibited comparable c-index performance to other ml algorithms while achieving higher sensitivity and specificity. As a result, we conclude that LR demonstrates satisfactory predictive ability for PSCI in this study.

We found that LR is the primary type of model utilized for predicting stroke. LR is a classification algorithm that aims to establish the relationship between features and probability of specific outcomes (42). ML is commonly used to address issues encountered in clinical practice, with supervised learning and unsupervised learning being the most common approaches. Supervised learning primarily focuses on diagnosing and predicting disease prognosis or progression, which involves the process of training, validation, and testing. The training process involves inserting predictive factors into the model and using the model’s inherent parameter calculation rules (e.g., maximum likelihood estimation, iteration) to estimate the optimal model parameters. Selection of modeling variables (feature selection methods) is crucial for the training process and has been a subject of ongoing debate due to their diversification. Furthermore, validation and testing are crucial for a completed model as they reflect the model’s robustness. Unfortunately, in actual research, most studies lacked effective external validation. The original studies included in our analysis predominantly utilized a supervised ML process with single-factor + multi-factor LR model selection method and performed internal validation through random sampling (43–45).

In our study, the c-index of LR did not significantly lag behind other types of ML models, which demonstrates relatively high sensitivity and specificity. Hence, we believe that LR exhibits promising predictive potential for PSCI.

In addition, we found that the major modeling variables for the ML-based PSCI prediction models were age, gender, education level, white matter hyperintensity (WMH), stroke history, stroke severity, lesion volume, lesion site, stroke subtype, and vascular risk factors. These modeling variables were still mainly based on past identified risk factors (race, age, gender, education level, vascular risk factor, stroke severity, and stroke lesion site and volume) (46), and very few or no newly identified risk factors were used for modeling, such as blood proteins [homocysteine (Hcy), C-reactive protein (CRP), low-density lipoprotein cholesterol (LDL-C), total cholesterol (TC)] that have been recognized as effective biomarkers for PSCI (47), cognitive reserve (CR) (48), activity and participation of stroke survivors (49), and intestinal dysbiosis (50). Therefore, the newly identified risk factors should be prioritized for further validation as their efficacy as modeling variables remains uncertain.

It was reported that common cognitive screening tools have similar predictive accuracy in PSCI. Although the MoCA has significantly better sensitivity in PSCI prediction than other cognitive screening tools, its specificity is less than desirable (51, 52). This demonstrates that there is a lack of effective prediction models for the early screening of PSCI. However, our findings showed that ML has considerably high predictive accuracy (c-index, sensitivity, and specificity) in PSCI and is a promising tool for predicting PSCI.

A recent systematic review indicated that although PSCI has unique risk factors (e.g., Vascular risk factors, lifestyle, overweight, and obesity), it is currently unclear whether the intervention of these risk factors can effectively reduce the incidence of PSCI. Most approaches for lowering PSCI incidence are still dependent on effective prophylaxis for stroke (46). Therefore, effective prediction tools for the early identification and diagnosis of PSCI are urgently needed. Despite the uncertainty in the intervention measures for post-stroke cognitive functions, some researchers found that physical activity intervention and noninvasive brain stimulation can improve post-stroke cognitive functions compared with conventional care (53). Though, the ≥2-year improvement in PSCI after intervention was small (54). Moreover, patients with cognitive impairment have significantly increased risks of subsequent ischemic and fatal stroke (55, 56). Hence, early identification of appropriate treatment and rehabilitation measures are critical for improving the health and life expectancy of PSCI patients. Our study demonstrated the feasibility of ML in the development of PSCI prediction tools and that ML is also an important means for PSCI prediction.

Given the low number of PSCI prediction models for hemorrhagic stroke included in this study (n = 3) (19), the predictive accuracy of ML vs. common cognitive screening tools in PSCI in hemorrhagic stroke patients remains unclear and warrants further investigation.

ML plays an important role in the clinical management of stroke and improvement of the accuracy and efficiency of stroke prediction, diagnosis, personalized treatment, and prognosis assessment (57). For prediction of stroke risk, ML algorithms can be trained using patient data to establish predictive models and estimate the risk of stroke based on individual patient information, clinical indicators, and biomarkers. As for stroke diagnosis, ML can learn and identify radiological features of stroke and assist physicians with early and accurate diagnosis. In addition, ML can predict the efficacy and safety of different treatment options based on the patient’s personal information, medical history, and clinical manifestations, enabling physicians to develop personalized treatment strategies. Furthermore, ML algorithms can predict post-stroke recovery and long-term prognosis based on patients’ clinical and biomarker data. ML has been extensively used in stroke diagnosis, particularly in brain imaging, with SVM being the optimal model for stroke imaging (6, 44, 57). However, in our study, SVM exhibited inferior sensitivity to LR despite higher c-index, and the model size was limited (n = 1). Therefore, further exploration and development of SVM in predicting PSCI are warranted. We can attempt to optimize the accuracy of PSCI prediction by using different SVM models and parameter settings. SVM has various variants, such as non-linear SVM, multi-kernel SVM, and support vector regression, which are selected based on specific circumstances. Additionally, the combination of SVM with other ML methods can be explored for PSCI prediction. For instance, integrating SVM with deep learning techniques can improve the accuracy and robustness of predictions when analyzing images or text data. Moreover, extensive clinical validation studies are required to assess the actual effectiveness of SVM in PSCI prediction. The application value of SVM in PSCI prediction can be comprehensively assessed by collecting more data from PSCI patients and evaluating the models on independent validation sets. In conclusion, SVM, as a widely used ML method, has untapped potential in PSCI prediction. Continuous learning and research efforts can further refine and optimize the application of SVM in PSCI prediction, providing more accurate diagnostic and treatment decision support for clinical practitioners.

For this systematic review, the literature search was performed up until December 2022 and additional studies on this topic may become available subsequently. Hence, a regular review of the literature is recommended to obtain the most updated progress on this research topic.

ML has considerable predictive accuracy and is a promising prediction tool for PSCI. Therefore, future studies should concentrate on constructing ML models based on multi-racial, multi-center, and large-cohort samples and transforming them into simple clinical scoring tools with wide application. This will undoubtedly help with the development of follow-up strategies or rehabilitation measures for stroke patients to reduce their risk of developing cognitive impairment.

The original contributions presented in the study are included in the article/Supplementary material, further inquiries can be directed to the corresponding authors.

All authors listed have made a substantial, direct, and intellectual contribution to the work and approved it for publication.

This study was supported by the National Natural Science Foundation of China (31960178 and 82160923), Applied Basic Research Programs of Science and Technology Commission Foundation of Yunnan Province (2019FA007), Key Laboratory of Traditional Chinese Medicine for Prevention and Treatment of Neuropsychiatric Diseases, Yunnan Provincial Department of Education, Scientific Research Projects for High-level Talents of Yunnan University of Chinese Medicine (2019YZG01), Young Top-Notch Talent in 10,000 Talent Program of Yunnan Province (YNWR-QNBJ-2019-235), National Science and Technology Innovation 2030 Major Program (2021ZD0200900), Yunnan Key Research and Development Program (202103AC100005), and Yunnan Province Fabao Gao Expert Workstation Construction Project (202105AF150037).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.