Abstract 17128: Detection of Hypertrophic Cardiomyopathy From Noisy Single-Lead Electrocardiography
Akshay Khunte, Arya Aminorroaya, Lovedeep S Dhingra, Veer Sangha, Evangelos K Oikonomou, Sounok Sen, Harlan M Krumholz, Christopher M Kramer, Milind Y Desai, Rohan Khera- Physiology (medical)
- Cardiology and Cardiovascular Medicine
Background: Hypertrophic cardiomyopathy (HCM) is a leading cause of sudden cardiac death among young people and be detected from ECGs using artificial intelligence. Wearable devices could allow for broad AI-based screening but obtain noisy ECGs. We report the development of a wearable-adapted AI-ECG model that retains performance for detecting HCM in noisy single-lead ECGs.
Methods: In 11,883 ECGs obtained within 30 days of transthoracic echocardiograms from Yale New Haven Hospital (2015-2021), we defined HCM as interventricular septal wall thickness (end diastole) (IVSd) > 15 mm and moderate to severe left ventricular diastolic dysfunction (LVDD). Standard and noise-adapted models were trained using a 1:10 age-sex matched dataset. For the noise-adapted model, each ECG was included four times in the training set, augmented each time with random gaussian noise within four distinct frequency ranges emulating different real-world noise sources. We evaluated the performance of the noise-adapted and standard models on an independent set of ECGs with four different real-world noisy artifacts, including noise extracted from a portable device ECG, at multiple signal-to-noise ratios (SNRs).
Results: In single-lead ECGs without noise, the noise-adapted model outperformed the standard model, with an AUROC of 0.920 and 0.906, respectively for HCM. With the same ECGs augmented with noise ranging from half (SNR 2) to twice (SNR 0.5) the signal, the noise-adapted model maintained their performance across real-world noise signatures (A). At an SNR of 0.5, the noise adapted model had a significantly greater performance in detecting HCM in ECGs augmented with portable device ECG noise (AUROC 0.917 [0.861-0.974] vs 0.609 [0.398, 0.821]) (B).
Conclusions: We developed a model which accurately screens for HCM in noisy single-lead ECGs. This approach represents a more accessible and scalable method for screening for HCM using wearable device ECGs.