Background: The value of the electrocardiogram (ECG) for predicting long-term cardiovascular outcomes is not well defined. Machine learning methods are well suited for analysis of highly correlated data such as that from the ECG.
Methods: Using demographic, clinical, and 12-lead ECG data from the Third National Health and Nutrition Examination Survey (NHANES III), machine learning models were trained to predict 10-year cardiovascular mortality in ambulatory U.S. adults. Predictive performance of each model was assessed using area under receiver operating characteristic curve (AUROC), area under precision-recall curve (AUPRC), sensitivity, and specificity. These were compared to the 2013 American College of Cardiology/American Heart Association Pooled Cohort Equations (PCE).
Results: 7,067 study participants (mean age: 59.2 +/- 13.4 years, female: 52.5%, white: 73.9%, black: 23.3%) were included. At 10 years of follow up, 338 (4.8%) had died from cardiac causes. Compared to the PCE (AUROC: 0.668, AUPRC: 0.125, sensitivity: 0.492, specificity: 0.859), machine learning models only required demographic and ECG data to achieve comparable performance: logistic regression (AUROC: 0.754, AUPRC: 0.141, sensitivity: 0.747, specificity: 0.759), neural network (AUROC: 0.764, AUPRC: 0.149, sensitivity: 0.722, specificity: 0.787), and ensemble model (AUROC: 0.695, AUPRC: 0.166, sensitivity: 0.468, specificity: 0.912). Additional clinical data did not improve the predictive performance of machine learning models. In variable importance analysis, important ECG features clustered in inferior and lateral leads.
Conclusions: Machine learning can be applied to demographic and ECG data to predict 10-year cardiovascular mortality in ambulatory adults, with potentially important implications for primary prevention.
