JMIR Publications

ABSTRACT

Background Stroke has multiple modifiable and unmodifiable risk factors and represents a leading cause of death globally . This study aims to compare risk factors for stroke occurrence in a real-world cohort dataset using explainable machine learning (ML) approaches and conventional statistical models. Methods We assembled a retrospective cohort of high-risk patients who were treated at Ramathibodi Hospital, Thailand from January 2010 to December 2020. Prediction models were developed using logistic regression (LR), and ML models including XGBoost, Explainable Boosting Machine (EBM), Bayesian Network, and Tree-Augmented Naive Bayes. Model performance was compared using C-statistics and F-1 score. Results XGBoost and EBM had high predictive accuracy with C-statistics of 0.89 and 0.87 respectively whereas LR and EBM models yielded similar C-statistics of 0.80. Atrial fibrillation (AF), hypertension (HT), and antihypertensive medication were common significant factors, with AF being the most potent factor in both LR and XGBoost models. In contrast, plasma glucose was the strongest predictor in EBM models. Conclusion This study highlights the benefits of using ML approaches for the prediction of stroke in at risk patients. Further external validation of these ML approaches should be conducted before consideration as a tool adopted for routine clinical practice.