Deep Learning-Based Classification of Reduced Ejection Fraction from Wearable Multimodal Cardiac Signals

Abstract

Background: Cardiovascular diseases are common among elderly people. Some of these diseases cause reduced ejection fraction (EF), which leads to systolic heart failure. The standard medical procedures to identify low EF are echocardiography and right heart catheterization (RHC). However, these procedures are either not suitable for continuous monitoring or invasive. Utilizing advancements in wearable devices and deep learning models, this work investigates the early detection of low EF using multimodal cardiac signals.

Materials and Methods: In this work, a deep learning algorithm is proposed that ensembles a time-frequency two-dimensional CNN (2D CNN) model and a time-domain one-dimensional Convolutional Neural Network (1D CNN) model to detect low EF. The algorithm uses segments of electrocardiogram (ECG) and seismocardiography (SCG) signals captured by a wearable device placed on the subject’s chest. The SCG-RHC dataset was used for training and evaluation of the proposed algorithm. This dataset consists of recorded signals from 73 patients undergoing right heart catheterization.

Results: A standard subject-based five-fold cross-validation approach was used to evaluate the performance of the proposed algorithm. Near-perfect validation performance (accuracy of 99% and F1-score of 99%) was achieved, along with good generalization to unseen subjects.

Conclusion: Despite the limited dataset size, the achieved results demonstrate the potential of using wearable devices combined with the proposed algorithm for the initial screening of low EF. Given its non-invasive nature, remote monitoring of patients with cardiovascular diseases is feasible.