DA-GAT-v2, a Demographic-Aware Graph Attention Network, achieves macro F1 of 0.8952 and AUROC of 0.9762 on 12-lead ECG interpretation while reducing the male-female diagnostic performance gap from 15.42% to 1.75%, demonstrating that 'diagnostic accuracy and demographic fairness in ECG AI are complementary objectives achievable through principled architectural design.'
DA-GAT-v2 achieved a macro F1 of 0.8952 and AUROC of 0.9762 on the PTB-XL dataset, surpassing all compared baselines.
DA-GAT-v2 reduced the male-female diagnostic performance gap from 15.42% to 1.75%.
Cross-dataset validation on Chapman-Shaoxing confirmed generalization with a minimal F1 degradation of 0.0150.
A lead-wise Temporal Convolutional Encoder (TCE) was developed to replace coarse statistical node features with 128-dimensional morphologically rich embeddings.
A dynamic α-Net was introduced to predict patient-specific inter-lead graph topologies by adaptively balancing anatomical adjacency and signal correlation.
Feature-wise Linear Modulation (FiLM) was integrated into every graph attention layer to enable demographic conditioning.
A three-stage curriculum training strategy with composite fairness regularization loss was employed, combining equalized odds and demographic parity constraints.
Ablation studies quantified each architectural component's independent contribution to performance and fairness.
Fairness evaluation was limited to sex and age subgroups due to the absence of race and ethnicity metadata in both datasets.
This research introduces DA-GAT-v2, an artificial intelligence system designed to interpret 12-lead electrocardiograms (ECGs) — the standard heart tracing test used in clinical settings — both accurately and fairly across male and female patients. Many existing AI diagnostic tools perform significantly better for one group than another, a problem the authors address not by patching the bias after the fact, but by building fairness directly into the system's design. The model uses three new technical components: a specialized encoder that extracts detailed heart signal patterns specific to sex and age, a flexible graph structure that adapts to individual patients' heart anatomy, and a modulation mechanism that adjusts how the system processes features based on demographic information. Tested on over 21,500 ECG recordings from the PTB-XL dataset, DA-GAT-v2 outperformed all other compared methods in detecting cardiac abnormalities, achieving high accuracy scores (macro F1: 0.8952, AUROC: 0.9762). More notably, the performance gap between male and female patients — which was 15.42% before this approach — was reduced to just 1.75%, falling well within a defined clinical acceptability threshold. When the model was tested on a separate dataset of over 10,600 recordings from Chapman-Shaoxing, it maintained strong performance with only a small drop in accuracy, suggesting the approach generalizes beyond the data it was trained on. This research suggests that building demographic awareness into the core architecture of an AI diagnostic tool — rather than applying corrections afterward — can simultaneously improve both accuracy and fairness. The findings are relevant to the ongoing challenge of deploying AI tools in healthcare settings where performance disparities across patient groups can have real consequences for diagnosis and treatment. The authors note that the fairness analysis was limited to sex and age because the datasets did not include race or ethnicity information, and further evaluation would be needed before clinical deployment.
Naeem M. (2026). Demographic-aware temporal graph attention for fair and accurate cardiac abnormality detection in 12-lead ECG.. Scientific reports. https://doi.org/10.1038/s41598-026-54206-8