Utility of ECG-Vision for Surveillance of Left Ventricular Dysfunction in Patients With Hypertrophic Cardiomyopathy Initiated on Mavacamten.

Abstract

Background: Frequent transthoracic echocardiograms (TTEs) are required to monitor for left ventricular systolic dysfunction (LVSD) in patients with obstructive hypertrophic cardiomyopathy receiving myosin inhibitors. This requirement may be cumbersome and limit access in underserved and rural areas. The objective of this study was to evaluate the performance of an artificial intelligence (AI)-enabled ECG tool in predicting LVSD in hypertrophic cardiomyopathy patients on mavacamten.

Methods: At Morristown Medical Center/Atlantic Health, 147 patients initiated on mavacamten between June 2022 and June 2025 underwent ECGs and TTEs at baseline and clinically available follow-up visits. A validated AI-enabled ECG algorithm predicted the probability of LVSD. Sensitivity, specificity, negative predictive value, positive predictive value, and area under the curve were calculated for left ventricular ejection fraction <50%, with CIs accounting for repeated paired ECG-TTE observations within patients.

Results: In 147 patients, mean age was 65±14 years; 44% were male. Among 453 paired ECG-TTE observations, 8 LVSD event observations occurred. At an AI probability threshold of 20%, sensitivity was 100% (95% CI, 68%-100%), specificity was 75% (95% CI, 69%-82%), positive predictive value was 7% (95% CI, 3%-11%), and negative predictive value was 100% (95% CI, 99%-100%). The area under the curve was 0.94 (95% CI, 0.90-0.98). Patient-level classifications between AI-enabled ECG and echo-confirmed LVSD were concordant in 127 of 147 patients (86%), inconclusive in 17 of 147 (12%), and discordant in 3 of 147 (2%).

Conclusions: In this single-center exploratory cohort of patients with obstructive hypertrophic cardiomyopathy receiving mavacamten, ECG-Vision left ventricular demonstrated high observed sensitivity and negative predictive value for TTE-defined LVSD, although estimates were imprecise because LVSD events were infrequent. These findings support prospective multicenter validation of AI-enabled ECG as a potential adjunctive triage tool.