Basic/Translational Science -> Intact Heart Electrophysiology (includes Pharmacology and Optical Mapping) D-SP05 - Young Investigator Awards Competition (ID 34) Special Session

D-SP05-01 - Machine Learning Of Ventricular Monophasic Action Potentials Predicts 3 Year Arrhythmic And All-cause Mortality In Heart Failure (ID 719)


Background: Predictive tools derived from current clinical and electrical markers only modestly predict outcomes in patients with heart failure.
Objective: To test the hypothesis that machine learning of ventricular monophasic action potentials (MAPs), that may reflect remodeling, may improve prediction of sustained ventricular arrhythmias (VT/VF) and mortality.
Methods: We recorded 5706 MAPs in 42 patients with coronary disease and left ventricular ejection fraction (LVEF) ≤ 40%, from right and left ventricles at electrophysiological study (fig. A). We featurized voltage-time series using tsfresh in Python. We then used 70% to train support vector machines (SVM) to VT/VF (appropriate ICD shocks or clinical VT/VF) or all-cause mortality. We tested accuracy in independent 30% cohorts of the data and repeated this for k=10-fold cross validation splits.
Results: Patients had age 64.7 + 13.0 years, 27.0 + 7.6 %. On median follow-up of 2408 days (IQR 920-1489), N=13 patients had VT/VF and N=15 patients died. Fig B shows that the SVM provided an AUC for VT/VF of 0.901, and fig. C shows that the AUC for mortality was 0.912. This exceeded clinical and electrophysiological variables. Explainability analysis suggested that different MAP shapes were linked with arrhythmic or overall mortality.
Conclusion: Machine learning on remodeled in vivo ventricular MAPs predicts long-term outcomes in heart failure. Future studies should extend these data to larger populations and study components of remodeling which contribute to pathology.