Artificial intelligence enabled electrocardiogram for mortality and ventricular arrhythmia prediction

Abstract

Abstract Background Artificial intelligence-enhanced electrocardiograms (AI-ECG) can be used to predict risk of events but existing models suffer from lack of actionability, explainability and biological plausibility, and therefore are not used clinically. Purpose We sought to address these limitations of previous AI-ECG approaches by developing the AI-ECG risk estimator (AIRE). Methods 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 create a subject-specific survival curve using a single ECG. AIRE was trained to predict time-to-mortality and fine-tuned for the additional endpoints described below. Results AIRE outputs a subject-specific survival curve that predicts not only mortality but time-to-mortality, using only a single ECG (Figure 1A). AIRE accurately predicts risk of all-cause mortality (C-index 0.775 (0.773-0.776), Figure 2) and can differentiate prognosis in high and low risk subjects even when only cardiologist reported normal ECGs are considered (Figure 2). AIRE also accurately predicted cardiovascular (CV) death 0.832 (0.831-0.834), future ventricular arrhythmia (0.760 (0.756-0.763)), future atherosclerotic cardiovascular disease (0.696 (0.694-0.698)) and future heart failure (0.787 (0.785-0.889))). In particular 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 ) and in dilated cardiomyopathy (C-index 0.690 (0.669-0.711) vs 0.608 (0.584 – 0.631), p < 0.0001). We used a variational autoencoder (VAE) to visualise the features most correlated with the mortality predictions. Features of QRS morphology, particularly broader and more left bundle branch block morphologies, inverted and biphasic T waves as well as ST segment changes were associated with high predicted mortality (Figure 1B). We externally validated our findings in four diverse, transnational cohorts from Brazil and the UK including volunteers and primary care subjects (Figure 2), and found AIRE accurately predicted all-cause mortality in all four external cohorts. Through phenome- and genome-wide association studies, we identified candidate biological pathways for the prediction of increased risk, including changes in cardiac structure and function (LV mass, LVEF, LV trabeculation), and genes associated with cardiac structure, QT interval, biological aging and metabolic syndrome (VGLL2, KCNQ1, CCDC91, TBX3). Conclusion AIRE is an explainable, biologically plausible and actionable AI-ECG risk estimation platform that has the potential for use worldwide across a wide range of clinical contexts for short- and long-term risk estimation.Figure 1Figure 2