Integrative multi-modal feature extraction from ECG signals using FusionHeartNet for robust heart disease prediction

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Electrocardiogram (ECG) diagnostics are crucial for early cardiac disorder detection.
• Existing models struggle to integrate temporal, spectral, and spatial ECG dynamics, often relying on unimodal approaches.
• This overlooks cross-domain dependencies and subtle morphological cues in complex arrhythmic patterns.

Purpose of the Study:

• To propose FusionHeartNet, a unified deep learning framework for robust, multi-dimensional ECG analysis.
• To fuse signal- and image-based ECG representations using a novel dual-spectrum feature embedding (DSFE) strategy.
• To enhance the precision of early heart disease detection through advanced ECG analysis.

Main Methods:

• Developed FusionHeartNet, a deep learning framework utilizing a dual-spectrum feature embedding (DSFE) strategy.
• Employed Fourier and wavelet transforms for spectral signatures and GAF/CWT scalograms for spatial morphology.
• Incorporated a multi-focus attention module (MFAM) and heart fusion classifier (HFC), optimized with Bayesian optimization and adaptive learning rate scheduling (BO-ALRS).

Main Results:

• Achieved 98.47% accuracy, 91.67% F1-score, and 0.9311 kappa on the MIT-BIH Arrhythmia Database.
• Demonstrated significant performance improvement over baseline models in ECG arrhythmia classification.
• Validated the effectiveness of the DSFE strategy and MFAM in capturing complex cardiac dynamics.

Conclusions:

• FusionHeartNet offers a clinically viable solution for early heart disease detection.
• The proposed framework sets a new benchmark for robust, multi-dimensional ECG analysis.
• Integrating diverse ECG data representations significantly enhances diagnostic precision.