Prediction of LVEF using BiLSTM and Swarm Decomposition-based 24-h HRV Components

Abstract

In this study, we investigated the effectiveness of using hourly Bi-Directional Long Short-Term Memory (BiLSTM) classifiers to predict left ventricle ejection fraction (LVEF) groups of CAD patients using their heart rate variability and Swarm Decomposition components. The 24-hour segmentation of patients' HRV data was performed using Cosinor Analysis. The novel Swarm Decomposition algorithm was then applied on the per-hour HRV data to extract the corresponding oscillatory components (HRV-OCs). The OCs represent the four bands in an HRV data, namely the ultra-low frequency (ULF), very-low frequency (VLF), low frequency (LF), and high frequency (HF). The training and classification process followed a leave-one-out scheme and was done for each per-hour HRV-OC. The highest prediction accuracy of LVEF was observed when using the VLF and HF components of HRV at an early morning hour (03-00-04:00 - average accuracy: 75.6%) and an evening hour (18:00-19:00 - average accuracy: 72.7%), respectively. In addition, the classifier achieved high sensitivity levels in predicting the borderline group (up to 76.7%), which is usually ambiguous and hard to diagnose in clinical practice.