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

Although deep learning methods have shown great promise for identification of structural and functional cardiac abnormalities using electrocardiographic data, these methods are data hungry, posing a challenge for critically important tasks where ground truth labels are relatively scarce. Impaired coronary microvascular and vasomotor function is difficult to identify with standard clinical methods of cardiovascular testing such as coronary angiography and noninvasive single photon emission tomography (SPECT) myocardial perfusion imaging (MPI). Gold standard data from positron emission tomography (PET) are gaining emphasis in clinical guidelines but are expensive and only available in relatively limited centers. We hypothesized that signals embedded within resting and stress electrocardiograms (ECGs) identify individuals with microvascular and vasomotor dysfunction. We developed and pretrained a self-supervised foundation vision transformer model using a large database of unlabeled ECG waveforms (N=800,035). We then fine-tuned the foundation model for two clinical tasks: the difficult problem of identifying patients with impaired myocardial flow reserve (AI-MFR), and the relatively easier problem of detecting impaired LVEF (AI-LVEF). A second ECG database was labeled with task-specific annotations derived from quantitative PET MPI (N=4167). Diagnostic accuracy of AI predictions was tested in a holdout set of patients undergoing PET MPI (N=1031). Prognostic evaluation was performed in the PET holdout cohort, as well as independent cohorts of patients undergoing pharmacologic or exercise stress SPECT MPI (N=6635). The diagnostic accuracy of AI-MFR with SSL pretraining increased significantly compared to supervised training (AUROC, sensitivity, specificity: 0.758, 70.1%, 69.4% vs. 0.632, 66.1%, 57.3%, < 0.0001). SSL pretraining also produced a smaller increase in AI-LVEF accuracy (AUROC, sensitivity, specificity: 0.946, 89.4%, 85.9% vs. 0.918, 87.6%, 82.5%, < 0.02). Abnormal AI-MFR was found to be significantly associated with mortality risk in all three test cohorts (Hazard Ratio (HR) 2.61 [95% CI 1.83, 3.71], < 0.0001, PET cohort; HR 2.30 [2.03, 2.61], < 0.0001, pharmacologic stress SPECT cohort; HR 3.76 [2.36, 5.99], < 0.0001, exercise stress SPECT cohort). SSL pretraining of a vision transformer foundation model enabled identification of signals predictive of impaired MFR, a hallmark of microvascular and vasomotor dysfunction, and impaired LV function in resting and stress ECG waveforms. These signals are powerful predictors of prognosis in patients undergoing routine noninvasive stress testing and could enable more efficient diagnosis and management of these common conditions.