Machine learning based prediction models for first stroke in community primary care: a systematic review and meta-analysis

Background Stroke is the third leading cause of death and the fourth leading cause of disability globally, particularly in low- and middle-income countries, where the disease burden is increasing significantly. Risk prediction of first-stroke is crucial for stroke prevention. Although artificial intelligence (AI)-based prediction models have rapidly developed in recent years, their clinical utility in primary community healthcare settings remains unclear. This study aimed to systematically review and analyze the application of algorithms, especially machine learning (ML)-based ones, in first-stroke risk prediction models within community settings.Methods We searched PubMed, EMBASE, and the Cochrane Library from inception to 30 June 2024 for studies developing or validating multivariable stroke risk models for primary care. We analyzed AUC/C-statistic values with 95% confidence intervals (CI) and conducted a meta-analysis.Results A total of 43 studies encompassing 93 models were included. The Cox regression model was the most common traditional method (n = 32, 74.42% of traditional models). Among ML models, 25 algorithms were identified, with logistic regression being the most frequent (n = 8, 18.60%). The meta-analysis showed a pooled AUC of 0.79 (95% CI: 0.77-0.81) for the Cox regression model, while logistic regression, random forest, and eXtreme Gradient Boosting models yielded 0.76 (95% CI: 0.64-0.89), 0.77 (95% CI: 0.61-0.93), and 0.77 (95% CI: 0.59-0.94), respectively. Notably, 76.7% of the included studies had a high risk of bias, and 20.9% raised high concerns regarding their applicability to community primary healthcare settings. ML models did not outperform traditional regression models, with significant performance variability observed.Conclusion Despite the increasing use of ML models for first-stroke risk prediction, their clinical utility in community primary healthcare remains unverified. Current ML models show no significant advantage over traditional regression methods and face challenges in algorithm innovation, data standardization, and external validation. Furthermore, the prevalence of methodological flaws and applicability concerns among existing studies underscores the need for caution. Future research should focus on standardized validation and real-world clinical analysis to develop effective stroke prevention tools.