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

BACKGROUND AND OBJECTIVE: Cardiovascular disease recognition from electrocardiogram (ECG) requires capturing subtle pre-symptomatic electrical abnormalities that manifest across spatially distributed leads and evolve temporally. Current predictive models inadequately address three critical challenges: the anatomical coupling between ECG leads reflecting cardiac electrical propagation, the multi-scale temporal dynamics spanning millisecond depolarization to circadian rhythms, and the progressive nature of cardiac pathophysiology. We developed an integrated spatiotemporal framework to model these interdependent mechanisms for early cardiovascular risk stratification. METHODS: We implemented a hybrid architecture combining graph convolutions with learnable adjacency matrices to model dynamic inter-lead dependencies based on cardiac electrical topology, hierarchical dilated temporal convolutions capturing rhythms from QRS complexes to heart rate variability patterns, and an external memory bank encoding disease-specific pattern prototypes. Cardiac-aware positional encoding preserved electrophysiological timing relationships, while variational inference quantified recognition uncertainty. We analyzed 10-second 12-lead ECG recordings evaluating myocardial infarction, heart failure, and arrhythmia recognition using area under the receiver operating characteristic curve (AUROC), sensitivity, and specificity with cross-validation. RESULTS: The framework achieved AUROC of 0.94 (sensitivity: 89.2%, specificity: 91.3%) for myocardial infarction, 0.91 (sensitivity: 86.7%, specificity: 88.1%) for heart failure, and 0.96 (sensitivity: 92.8%, specificity: 93.6%) for arrhythmias. Learned graph structures corresponded to known cardiac conduction pathways. Attention mechanisms identified precordial lead interactions during ischemic episodes and temporal windows associated with pathological onset, aligning with clinical prodromal phases. CONCLUSIONS: Explicit modeling of cardiac electrical topology and multi-resolution temporal dynamics enables detection of pre-symptomatic pathophysiological patterns, advancing cardiovascular disease recognition toward proactive risk management with mechanistically interpretable predictions.