Multi-Modality Machine Learning Models to Predict Stroke and Atrial Fibrillation in Patients with Heart Failure

Introduction Atrial fibrillation (AF) and stroke are leading causes of death of heart failure patients. Several ML models have been built using electrocardiography (ECG)-only data, or lab test data or health record data to predict these outcomes. However, a multi-modal approach using wearable ECG data integrated with lab tests and electronic health records (EHRs) data has not been developed. Objective The aim of this study was to apply machine learning techniques to predict stroke and AF amongst heart failure patients from a multi-modal dataset. Methods This study analysed hospitalised patients with heart failure in Hong Kong between 1 January 2010 and 31 December 2016, with the last follow-up of 31 December 2019. The primary outcomes were AF and stroke. The secondary outcomes were all-cause and cardiovascular mortality. ECG-only, non-ECG-only and multimodal models were built to assess feature importance. Four machine learning classifiers and seven performance measures were used to evaluate the performance. Results There are in total 2,868 subjects with heart failure upon admission, among them 1,150 (40.10%) had new onset AF, 668 (23.29%) had new onset stroke/TIA. It was found that accurate and sensitive machine learning models can be created to predict stroke and AF from multimodal data. XGBoost, which was the best algorithm tested, achieved a mean (over 10 iterations) accuracy, AUROC, AUPRC, positive predictive value and negative predictive value of 0.89, 0.80, 0.74, 0.99 and 0.88, respectively, for stroke and 0.78, 0.82, 0.77, 0.77 and 0.79, respectively, for AF. The predictive models, built using multimodal data, were easy to use and had high accuracy. Conclusion Multi-modal machine learning models could be used to predict future stroke and AF occurrences in patients hospitalised for heart failure.