Heart failure diagnosis and ejection fraction classification via feature fusion model using non-contact vital sign signals

Abstract

Background and objectives:

Ballistocardiography (BCG) has emerged as a promising modality for home-based heart failure (HF) monitoring, yet existing single-dimensional manual feature analyses fail to adequately characterize left ventricular ejection fraction (LVEF $<$ 40%) dynamics. We address this limitation by developing a hybrid feature fusion framework that synergizes manual feature engineering with deep learning for improved HF diagnosis and LVEF classification.

Methods:

83 participants were recruited from a hospital, with their samples categorized into two (healthy and HF) and three classes (healthy, LVEF $\ge$ 40% HF, and LVEF $<$ 40% HF) based on clinical diagnosis. Non-contact vital signs were collected from supine participants using a piezoelectric sensor, and the BCG and respiratory signals were isolated using filters. We developed a model that integrates manual with deep features extracted from BCG and respiratory signals, to enhance the accuracy of HF diagnosis and LVEF classification. Additionally, we designed a multi-scale ResNet-BiLSTM network model to extract deep features from the signals, effectively capturing dynamic changes and intrinsic patterns across various time scales.

Results:

Ablation experiments show that the proposed method outperforms traditional manual methods, achieving classification accuracies of 98.20% and 98.76% for two and three-class HF classification under five-fold cross-validation, respectively.

Conclusions:

This study establishes a healthcare-oriented framework for at-home diagnosis of HF and LVEF classification, facilitating rapid preliminary screening and auxiliary diagnosis in non-clinical settings.