Memory-augmented graph neural networks for multi-class cardiovascular disease recognition from 12-lead ECG signals

Area of Science:

• Cardiovascular medicine
• Artificial intelligence in healthcare
• Biomedical signal processing

Background:

• Current electrocardiogram (ECG) analysis for cardiovascular disease (CVD) struggles with subtle, pre-symptomatic electrical abnormalities.
• Existing models fail to adequately capture the anatomical coupling between ECG leads, multi-scale temporal dynamics, and the progressive nature of cardiac pathophysiology.

Purpose of the Study:

• To develop an integrated spatiotemporal framework for early cardiovascular risk stratification.
• To address limitations in current predictive models for CVD recognition from ECGs.

Main Methods:

• Implemented a hybrid architecture combining graph convolutions with learnable adjacency matrices for inter-lead dependency modeling.
• Utilized hierarchical dilated temporal convolutions and an external memory bank for rhythm and disease pattern analysis.
• Incorporated cardiac-aware positional encoding and variational inference for timing and uncertainty quantification.

Main Results:

• Achieved high Area Under the Receiver Operating Characteristic Curve (AUROC) for myocardial infarction (0.94), heart failure (0.91), and arrhythmias (0.96).
• Learned graph structures mirrored known cardiac conduction pathways.
• Attention mechanisms highlighted key lead interactions and temporal windows preceding pathological onset.

Conclusions:

• Explicit modeling of cardiac electrical topology and multi-resolution temporal dynamics enhances pre-symptomatic CVD detection.
• The framework advances cardiovascular disease recognition toward proactive risk management with interpretable predictions.