Explainable machine learning models to improve prediction of incident stroke in atrial fibrillation patients using health records, medical imaging and ECG derived metrics

Abstract Background Stroke is a leading cause of death and disability worldwide, affecting 12 million people each year. Atrial fibrillation (AF), the most common cardiac arrhythmia, underlies 20% of all ischaemic strokes and increases risk of thromboembolism 5-fold. Current stratification strategies rely on empirical models, such as CHA2DS2–VASc, to select high-risk AF patients suitable for anticoagulation, but despite their widespread use, these have significant limitations. The rising prevalence of risk factors such as diabetes and hypertension, along with an ageing population, calls for improved stratification strategies. Integrating patient medical imaging and ECG data, which has been proven effective in other cardiology domains, may enhance risk assessment. Aim This study applies explainable machine learning (ML) to predict stroke incidence in AF patients from the UK Biobank study using multimodal data, including patient health records, Cine cardiac magnetic resonance (CMR) imaging and ECG features, and to identify early biomarkers of stroke. Methods Volumetric and functional features were extracted from Cine CMR scans of AF patients using an automated segmentation pipeline. These were then combined with medical records and ECG-derived metrics from a resting 12-lead ECG. The dataset was stratified into paroxysmal (n = 527) and persistent (n = 183) AF patients and used to train a balanced random forest ML model. The model was validated on unseen data using cross validation and compared to CHA2DS2–VASc’s predictions. Shapley additive explanations (SHAP) were used to decompose the model outputs into individual feature contributions, identifying key risk factors. Results For paroxysmal AF patients, the balanced random forest model accurately predicted stroke incidence (C-statistics value, 0.72; 95% confidence interval CI 0.63; 0.79) and outperformed CHA2DS2–VASc on unseen data (C-statistics value, 0.61; 95% CI 0.52; 0.66). The same model also successfully predicted stroke in persistent AF patients (C-statistics value, 0.71; 95% CI 0.61; 0.80) exceeding the performance of CHA2DS2–VASc (C-statistics value, 0.63; 95% CI 0.49; 0.75). SHAP analysis revealed that CMR-derived metrics were the most influential predictors for both paroxysmal and persistent patients, while ECG-derived metrics played a key role in risk prediction mainly for paroxysmal patients (Figure 1). Statistical analysis revealed that patients who suffered from stroke had larger atria, lower atrial ejection fraction and greater P-wave entropy compared to non-stroke patients. These results suggest that stroke patients exhibited more severely impaired function of the atria. Conclusion This study highlights the power of ML in predicting stroke incidence in AF patients from multimodal data. Explainable features in patient medical images and ECG provide valuable insight into risks of stroke during AF progression, and should therefore be considered in patient stratification.Most important risk factors by ML model