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

Area of Science:

• Biomedical Engineering
• Cardiology
• Artificial Intelligence in Healthcare

Background:

• Ballistocardiography (BCG) shows potential for remote heart failure (HF) monitoring.
• Current manual feature analysis of BCG is insufficient for characterizing left ventricular ejection fraction (LVEF) dynamics, especially LVEF < 40%.
• A need exists for advanced methods to improve HF diagnosis and LVEF classification using BCG.

Purpose of the Study:

• To develop and validate a hybrid feature fusion framework integrating manual and deep learning features from BCG and respiratory signals.
• To enhance the accuracy of heart failure diagnosis and left ventricular ejection fraction classification.
• To establish a non-clinical framework for at-home HF and LVEF assessment.

Main Methods:

• Recruited 83 participants, classifying them into healthy, HF (LVEF ≥ 40%), and HF (LVEF < 40%) groups.
• Collected non-contact vital signs using piezoelectric sensors, isolating BCG and respiratory signals.
• Developed a hybrid model combining manual features with deep features from a multi-scale ResNet-BiLSTM network for signal analysis.

Main Results:

• The proposed hybrid method achieved superior performance compared to traditional manual methods.
• Achieved high classification accuracies: 98.20% for two-class (healthy vs. HF) and 98.76% for three-class (healthy, HF LVEF ≥ 40%, HF LVEF < 40%) HF classification.
• Results were validated using five-fold cross-validation.

Conclusions:

• The developed healthcare-oriented framework enables effective at-home diagnosis of HF and LVEF classification.
• Facilitates rapid preliminary screening and auxiliary diagnosis in non-clinical settings.
• Demonstrates the potential of hybrid deep learning approaches for remote cardiovascular health monitoring.