Machine learning (ML) has emerged as a powerful method to improve the diagnosis of cardiovascular diseases through. ML tools can detect abnormalities unavailable to traditional ECG analysis. A clinical problem is rapid and early detection and risk stratification of patients experiencing myocardial ischemia. Clinical diagnosis of myocardial ischemia from 12-lead ECGs lacks precise information about the size and severity of ischemic regions in the heart. Electrocardiographic imaging techniques may be able to address this challenge, however, such approaches add cumbersome and expensive imaging and modeling requirements which make them less suitable for clinical implementation.
We developed an ML tool to predict the size of ischemic regions within the heart using body surface ECG signals. Our dataset consisted of ECGs and ischemic tissue volumes measured from 5 replicates of a large animal model during repeated episodes of ischemic stress. Ischemic volume was measured from intramural needle electrodes in the myocardial volume. From these experiments, we extracted 35,633 ECGs and ischemic volumes, which were split into a 90\% train 10\% test set.
We were able to predict the size of the ischemic regions within a root mean squared error of 2.52~$mm^3$ and detect the presence of ischemia with an area under the receiver operator curve of 0.96.
These results demonstrate a novel application of ECG-ML tools to enhance clinical risk stratification of patients experiencing myocardial ischemia. Combining these ML-based identification of ischemic severity with traditional ECG analysis will allow for rapid, accurate, and robust risk stratification of patients earlier in the disease process. Future studies will look to combine these techniques with physics-based electrocardiographic imaging techniques for localizing myocardial ischemia, and transfer learning techniques to bridge these ML tools from animal to human ECGs.