Abstract 17073: Evaluation of Impact of Septal Reduction Therapy on Deep Learning-Based Electrocardiographic Markers Hypertrophic Cardiomyopathy

Background: Deep learning-based models could identify hypertrophic cardiomyopathy (HCM) signatures on ECG, with emerging evidence of being able to track disease modification with mavacamten. To explore the mechanism of HCM detection on ECG, we developed an AI-ECG model for echocardiographically confirmed HCM and evaluated HCM detection after septal reduction therapy (SRT, alcohol septal ablation [ASA] or ventricular myectomy [VM]).

Methods: The AI-ECG model for HCM detection was developed in a cohort of 25,652 ECGs (1:10, age-sex matched HCM to control) obtained within 30 days of a transthoracic echocardiogram (TTE) at Yale (2015-21). We defined HCM by end-diastolic interventricular septal wall thickness > 15 mm with moderate to severe diastolic dysfunction on TTE. The model was developed using a convolutional neural network and transfer learning. HCM phenotype was defined by high predicted probabilities for HCM and was assessed on ECGs spanning SRT.

Results: Our AI-ECG model for HCM had an AUROC of 0.93 (0.88-0.97), sensitivity of 0.90, and specificity of 0.89 in the held-out test set. The mean HCM probability among HCM patients vs control was 0.51±0.33 vs 0.07±0.20 (P<.001). The mechanism of detection was evaluated in 1400 ECGs from 78 patients (mean age 66 y; 47% women) spanning SRT (20 ASA, 58 VM). In this group, the mean HCM probability for all ECGs performed before & after SRT was 0.49±0.34 & 0.72±0.30, respectively. The mean HCM probability in the closest ECGs per patient before SRT (median 35 days before SRT) was 0.53±0.35, and in farthest ECGs after SRT (median 536 days) was 0.70±0.32 (P diff = 0.002).

Conclusion: We developed and validated an AI-ECG model that detects HCM from 12-lead ECG. The model does not merely detect thickened septum, and the continued identification of HCM signature after SRT suggests HCM prediction based on disordered myocardial kinetics. This supports the role of AI-ECG in both detection and monitoring of phenotypic effects of therapies in HCM.