Predicting stroke in Asian patients with atrial fibrillation using machine learning: A report from the KERALA-AF registry, with external validation in the APHRS-AF registry

Atrial fibrillation (AF) is a significant risk factor for stroke. Based on the higher stroke associated with AF in the South Asian population, we constructed a one-year stroke prediction model using machine learning (ML) methods in KERALA-AF South Asian cohort. External validation was performed in the prospective APHRS-AF registry. We studied 2101 patients and 83 were to patients with stroke in KERALA-AF registry. The random forest showed the best predictive performance in the internal validation with receiver operator characteristic curve (AUC) and G-mean of 0.821 and 0.427, respectively. In the external validation, the light gradient boosting machine showed the best predictive performance with AUC and G-mean of 0.670 and 0.083, respectively. We report the first demonstration of ML's applicability in an Indian prospective cohort, although the more modest prediction on external validation in a separate multinational Asian registry suggests the need for ethnic-specific ML models.