EA-DualGAT: An error-aware dual-graph attention framework for Parkinson's disease diagnosis from wearable plantar pressure gait signals.

Timely diagnosis of Parkinson's disease (PD) is essential for timely intervention and effective disease management. Wearable sensing technologies provide objective gait measurements and have become a promising, low-cost solution for automated PD diagnosis. However, existing deep learning and graph-based methods primarily focus on individual gait representations or fixed graph structures, often overlooking relationships among gait samples, difficult samples, and adaptive graph refinement during training. To address these limitations, this study proposes EA-DualGAT, an Error-Aware Dual Graph Attention Network (GAT) for PD diagnosis using wearable gait signals. A time-series autoencoder is first employed to extract latent representations from gait cycles acquired by 16 plantar pressure sensors measuring vertical ground reaction force (vGRF). These latent representations are subsequently used as graph node features to construct a gait-cycle similarity graph that captures structural dependencies among gait cycles. The proposed architecture combines several components: a residual multi-head GAT with Jumping Knowledge (JK) aggregation, a dynamically constructed Error Graph processed by a dedicated Error-GAT encoder, an Error Memory Bank (EMB), attention-based feature fusion, and self-distillation using an exponential moving average (EMA) teacher. These components work together to progressively refine difficult samples and improve representation learning. Experimental results demonstrate the effectiveness of the proposed framework, achieving 99.87% accuracy, 99.92% precision, 99.82% recall, and 99.87% F1-score on the test set. These findings indicate that EA-DualGAT provides an accurate and computationally efficient framework for wearable sensor-based PD diagnosis on the evaluated dataset, highlighting its potential for deployment in intelligent clinical decision-support systems and remote healthcare applications.