Abstract
ObjectivesIn hypertrophic cardiomyopathy (HCM), specific ECG abnormalities are observed. Therefore, ECG is a valuable screening tool. Although several studies have reported on estimating the risk of developing fatal arrhythmias from ECG findings, the use of ECG to identify the severity of heart failure (HF) by applying deep learning (DL) methods has not been established.MethodsWe assessed whether data-driven machine-learning methods could effectively identify the severity of HF in patients with HCM. A residual neural network-based model was developed using 12-lead ECG data from 218 patients with HCM and 245 patients with non-HCM, categorised them into two (mild-to-moderate and severe) or three (mild, moderate and severe) severities of HF. These severities were defined according to the New York Heart Association functional class and levels of the N-terminal prohormone of brain natriuretic peptide. In addition, the patients were divided into groups according to Kansas City Cardiomyopathy Questionnaire (KCCQ)-12. A transfer learning method was applied to resolve the issue of the low number of target samples. The model was trained in advance using PTB-XL, which is an open ECG dataset.ResultsThe model trained with our dataset achieved a weighted average F1 score of 0.745 and precision of 0.750 for the mild-to-moderate class samples. Similar results were obtained for grouping based on KCCQ-12. Through data analyses using the Guided Gradient Weighted-Class Activation Map and Integrated Gradients, QRS waves were intensively highlighted among true-positive mild-to-moderate class cases, while the highlighted part was highly variable among true-positive severe class cases.ConclusionsWe developed a model for classifying HF severity in patients with HCM using a deep neural network algorithm with 12-lead ECG data. Our findings suggest that applications of this DL algorithm for using 12-lead ECG data may be useful to classify the HF status in patients with HCM.
Subject
Cardiology and Cardiovascular Medicine