Background: Machine learning (ML) of the ECG has recently been used to detect non-cardiac phenotypes. We hypothesized that ML of the ECG could be used to detect a variety of co-morbidities with cardiac implications or the impact of common cardiac medications.
Objective: To develop ML architectures to predict clinical comorbidities from the ECG.
Methods: We curated ECGs in 608 well-characterized cardiac patients at 2 academic centers. From the 12-lead ECG, each patients’ corresponding X vector leads in Frank’s XYZ vectorcardiogram model (Fig. A) were transformed to a median beat. We randomly used 80% of data for training and 20% for validation, repeating this for k=5-fold cross validations. Support vector machines (SVM) models were used. Inputs to the SVM were signal characteristics (tsfresh, python) and output was presence or absence of comorbidities.
Results: Patients had age 61.4 ± 14.5 years, and 31.2% were female. This ML model enabled a single ECG beat to provide accuracies of 72% to identify diabetes mellitus and 67% to identify patients with previous coronary revascularization procedures. Furthermore, ML was able to identify use of beta-blocker with accuracy of 70% and the use of antiarrhythmic medications with accuracy of 71% (Fig. B).
Conclusion: In preliminary studies of a modest dataset, ML of the ECG revealed clinical phenotypes with cardiac and arrhythmic implications. Future studies should extend this work to larger datasets, using alternative architectures and explainability analysis to elucidate how single ECG beats identify each pathophysiology.