Resource-efficient and fair spatial graph attention networks for edge-based multi-label cardiac abnormality detection

Background Cardiovascular mortality claims 20.5 million lives annually, yet deployed ECG artificial intelligence systems exhibit systematic diagnostic disparities across demographic groups — particularly between male and female patients — precluding equitable clinical deployment. Current fairness approaches suppress demographic information, discarding clinically meaningful physiological variation rather than leveraging it. Methods We introduce SC-GAT, a Spatial-Conditioned Graph Attention Network for resource-constrained edge deployment. SC-GAT models the 12 ECG leads as a graph (leads as nodes; edges encoding anatomical adjacency and Pearson signal correlations) and incorporates patient age and sex directly into every attention layer via a low-overhead sigmoid gating mechanism. A fairness-aware multi-task curriculum jointly optimises diagnostic accuracy, demographic representation, and equalized-odds regularisation across three progressive training phases. The model comprises 2.4 M parameters (≤16 MB RAM), compatible with ARM Cortex-M7-class wearable processors. Results Evaluated on PTB-XL (21,837 recordings, 5-fold stratified cross-validation), SC-GAT achieves F1 = 0.8660 and AUROC = 0.9577, outperforming all eight baselines — including transformer and graph-based architectures — with p < 0.05 and Cohen’s d > 2.0. SC-GAT reduces the male–female diagnostic performance gap from 16.67% to 3.93% (76.4% reduction), simultaneously satisfying both ΔDP < 0.05 and ΔEO < 0.10 clinical fairness thresholds within an edge-deployable parameter budget — a combination not demonstrated by any compared baseline. Conclusions SC-GAT empirically falsifies the accuracy–fairness trade-off: treating demographic context as a physiological resource rather than a statistical nuisance simultaneously advances diagnostic accuracy and demographic equity, establishing fairness-through-awareness as a principled design paradigm for equitable cardiac AI.