Self-supervised deep representation learning of a foundation transformer model enabling efficient ECG-based assessment of cardiac and coronary function with limited labels

Abstract

Background: Impaired microvascular and vasomotor function is a common consequence of aging, diabetes, and other risk factors, and is associated with adverse cardiac outcomes. Such impairments are not readily identified by standard clinical methods of cardiovascular testing such as coronary angiography and noninvasive single photon emission tomography (SPECT) myocardial perfusion imaging (MPI). We hypothesized that signals embedded within stress electrocardiograms (ECGs) identify individuals with microvascular and vasomotor dysfunction. Methods: We developed and validated a novel convolutional neural network (CNN) using stress and rest ECG data (ECG-Flow) to identify patients with impaired myocardial flow reserve (MFR) on quantitative positron emission tomography (PET) MPI (N=3887). Diagnostic accuracy was validated with an internal holdout set of patients undergoing stress PET MPI (N=963). The prognostic association of ECG-Flow with mortality was then evaluated in a separate cohort of patients undergoing SPECT MPI (N=5102). Results: ECG-Flow achieved good diagnostic accuracy for impaired MFR in the holdout PET cohort (AUC, sensitivity, specificity: 0.737, 71.1%, 65.7%). Abnormal ECG-Flow was found to be significantly associated with mortality in both PET holdout and SPECT MPI cohorts (adjusted HR 2.12 [95% CI 1.45, 2.10], p=0.0001, and 2.07 [1.82, 2.36], p<0.0001, respectively). Conclusion: Signals predictive of microvascular and vasomotor dysfunction are embedded in stress ECG waveforms. These signals can be identified by deep learning methods and are related to prognosis in patients undergoing both stress PET and SPECT MPI.