FHGNet: A Feature-Centric Hierarchical Network with Graph Attention Layer for Supraventricular Tachycardia...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Automated electrocardiogram (ECG) classification is vital for arrhythmia diagnosis.
• Current deep learning models lack physiological rhythm integration and clinical reasoning, limiting reliability and interpretability.
• Ventricular tachycardia (VT) and supraventricular tachycardia (SVT) classification requires enhanced precision and interoperability.

Purpose of the Study:

• To propose FHGNet, a clinically inspired multi-lead oscillatory Transformer framework for improved VT/SVT classification.
• To enhance temporal awareness and capture both intra-beat morphology and inter-beat rhythmic patterns.
• To improve the detection of rare arrhythmia classes and model interpretability.

Main Methods:

• FHGNet integrates R-peak detection, adaptive-length patch extraction with R-wave positional encoding.
• A CNN captures QRS morphology, a Transformer with FANLayer models inter-beat rhythms, and a GAT fuses multi-lead dependencies.
• A two-stage classifier is employed for enhanced rare class detection.

Main Results:

• FHGNet achieved a macro F1-score of 91.35% on the MIT-BIH Supraventricular Arrhythmia dataset, outperforming baselines.
• Ablation studies showed GAT removal decreased F1 by 2.42% and the two-stage design improved minority class recall by 5.82%.
• Attention visualization confirmed focus on clinically relevant features like ST-T segment energy and inter-lead phase differences.

Conclusions:

• FHGNet offers an interpretable, clinically adaptive framework for high-accuracy ECG classification.
• The model's design aligns with clinicians' rhythm analysis logic and diagnostic workflow, enhancing traceability.
• This approach may reduce the need for invasive electrophysiological studies in arrhythmia diagnosis.