Accurate prediction of stroke outcome is crucial for patient management and treatment planning. However, clinical prediction can be challenging, especially for inexperienced physicians. Artificial intelligence (AI), particularly machine learning (ML) and deep learning (DL), offers the potential to improve prognostic accuracy. AI algorithms can analyze complex medical data, including imaging and clinical variables, to predict stroke outcomes more objectively than traditional methods. While several studies have explored the use of AI in stroke prognosis, the overall accuracy and the performance of different algorithms remain unclear. This systematic review and meta-analysis aimed to comprehensively assess the predictive performance of AI models in predicting stroke outcomes and compare the accuracy of various algorithms. The study sought to determine if AI models could provide a reliable and clinically useful tool to assist clinicians in predicting stroke outcomes.

Previous studies exploring AI for stroke outcome prediction have used various algorithms, including logistic regression (LR), random forests (RF), support vector machines (SVM), and deep neural networks (DNN). While some studies reported high accuracy (AUCs up to 0.936), others showed inconsistencies. Tree-based algorithms like RF have shown promise due to their interpretability, achieving high accuracy in some studies, though sample sizes were often limited. SVM models also demonstrated moderate to high accuracy, but their use of neuroimaging data often limited interpretability. DNNs, the most complex algorithms, showed good accuracy in larger sample analyses, but often lacked interpretability. The overall accuracy of AI models varied across studies, highlighting the need for a comprehensive meta-analysis to consolidate existing evidence.

This study followed PRISMA guidelines. PubMed, Embase, and Web of Science databases were searched from inception to February 2023 using keywords related to acute stroke, AI, prognosis, and outcome. Inclusion criteria included studies using AI to predict acute stroke prognosis in cohorts with complete data. The Quality Assessment of Diagnostic Accuracy Studies (QUADAS) tool was used to assess the risk of bias in included studies. The area under the curve (AUC) was used as the primary outcome measure to assess the predictive accuracy of AI models. A random-effects model was used for meta-analysis, and heterogeneity was evaluated using the Q statistic and I-squared test. Subgroup analyses were conducted to compare the AUCs across different AI algorithms.

The search yielded 1,241 publications, with seven studies (including 4,379 ischemic stroke patients) ultimately included in the meta-analysis. The included studies had low to moderate risks of bias. No significant heterogeneity was detected (I² = 27.67%). The pooled AUC under the fixed-effects model was 0.872 (95% CI 0.862–0.881), demonstrating good overall predictive accuracy of AI models in predicting stroke outcomes. Subgroup analysis revealed varying AUCs among different algorithms. The DL subgroup had an AUC of 0.888 (95% CI 0.872–0.904), LR subgroup had an AUC of 0.852 (95% CI 0.835–0.869), RF subgroup had an AUC of 0.863 (95% CI 0.845–0.882), SVM subgroup had an AUC of 0.905 (95% CI 0.857–0.952), and Xgboost subgroup had an AUC of 0.905 (95% CI 0.805–1.000). SVM and Xgboost models demonstrated the highest accuracy.

This meta-analysis provides the first comprehensive assessment of AI models' performance in predicting stroke outcomes. The high pooled AUC indicates that AI models offer a valuable tool to aid clinicians in predicting stroke outcomes, particularly ischemic stroke. The superior performance of SVM and Xgboost models may be attributed to their ability to handle non-linear relationships in data. The results are promising compared to traditional prognostic scores like the Virtual International Stroke Trials Archive (AUC 0.808 for functional outcome, 0.706 for survival) and the CoRisk score (AUC 0.819), although these traditional methods possess better interpretability. The use of data readily available in the emergency department enhances the clinical applicability of AI models. Despite the good predictive accuracy of DL models, their complexity and lack of interpretability might limit their clinical use. Future research should focus on developing more interpretable models and using larger datasets to improve the generalizability of AI models.

AI models demonstrate high accuracy in predicting stroke outcomes, providing valuable support for clinical decision-making. While SVM and Xgboost algorithms showed superior performance, further research is needed to optimize AI models for improved interpretability and wider clinical applicability. Larger datasets are necessary to validate these findings and explore the potential of AI in personalizing stroke care.

The meta-analysis focused primarily on ischemic stroke due to limited data on hemorrhagic stroke. The assessment of predictive accuracy was limited to AUC, without detailed analysis of sensitivity, specificity, and accuracy due to incomplete data. The sample sizes in the included studies were relatively small, hindering the generalizability of the findings. Potential publication bias cannot be fully excluded. The reasons behind the choice of specific algorithms by researchers in the included studies remain unknown.