Prediction of mortality, future arrhythmia and cardiovascular disease: an artificial intelligence-enhanced electrocardiography platform

Abstract

Abstract Background Artificial intelligence-enhanced electrocardiography (AI-ECG) can be used to identify existing disease, but could additionally be used to predict occurrence of future disease and death. Purpose We developed an AI-ECG platform that predicts mortality and a wide spectrum of future arrhythmia and cardiovascular disease. Methods The AI-ECG risk estimation platform (AIRE) was developed in a dataset of 1,163,401 ECGs from 189,539 patients from a secondary care setting in the USA. Uniquely, AIRE uses deep learning with a residual convolutional neural network with a discrete-time survival loss function to output subject-specific survival curves (Fig 1A). AIRE was fine-tuned to additionally predict the endpoints described below. Results AIRE accurately predicts risk of all-cause mortality (C-index 0.775 (0.773-0.776) and can differentiate prognosis in high- and low-risk subjects, even when only cardiologist-defined normal ECGs are considered (Fig 2B). Additionally, in Cox regression analyses, AIRE was superior to conventional risk factors (Fig 1B). AIRE accurately predicted future ventricular arrhythmia (c-index 0.760 (0.756-0.763)), future complete heart block (CHB) (0.809 (0.805-0.814), future atrial fibrillation (0.753 (0.751-0.756), future atherosclerotic cardiovascular disease (0.696 (0.694-0.698)) and future heart failure (0.787 (0.785-0.789)) . AIRE was superior to left ventricular ejection fraction (LVEF) in predicting risk of future ventricular arrhythmias in subjects with LVEF <50% (C-index 0.649 (0.638 – 0.660) vs 0.615 (0.603 – 0.626) p < 0.0001 ). For CHB prediction, AIRE was able to further risk stratify subjects with bifascicular block and syncope into low (1.9% 5 year CHB), intermediate (8.8%), and high risk (20.7%) groups, and could guide treatment decisions. We externally validated our findings in four diverse, transnational cohorts from Brazil and the UK, including volunteers, primary care subjects and patients with Chagas disease, and found AIRE accurately predicted all-cause mortality in all four external cohorts (Fig 2). Non-mortality endpoints were successful externally validated in the UKB. In explainability analyses, we found features of QRS morphology, low QRS voltage, poor precordial R wave progression, inverted and biphasic T waves and ST segment changes were the most significant morphological features associated with high predicted mortality (Fig 1D). Through phenome- and genome-wide association studies (GWAS), we identified candidate biological pathways including changes in cardiac structure (LV mass, LV trabeculation and left atrial size). GWAS identified variants adjacent to VGLL2, KCNQ1, CCDC91 and TBX3 which have been described in association with QT interval, biological aging and metabolic syndrome (Fig 1E). Conclusion AIRE could be used worldwide across a range of clinical contexts for risk estimation of not only mortality but a wide spectrum of future arrhythmia and cardiovascular diseases.Figure 1Figure 2