Edge-Aware Graph Transformer (EdgeGT-Net): A Meta Optimized, Energy Efficient Architecture for Real-Time Multisensor Stress Monitoring in Wearable Electronics

Continuous, real-time assessment of psychological stress on resource constrained wearables requires models that balance predictive performance with millijoule level energy budgets. This work presents Edge-Aware Graph Transformer (EdgeGT–Net), a unified architecture that integrates 1) a Leaky Integrate and Fire temporal encoder for sparse, event driven representation, 2) a two layer graph convolutional stem that captures inter-sensor physiological dependencies, and 3) a compact multi-head transformer that models long range temporal dynamics. A hybrid genetic-algorithm/particle-swarm optimizer (HGAPSO) jointly tunes bit precision, membrane decay, attention heads, model width, and learning rate under latency and energy constraints. On the WESAD dataset, EdgeGT–Net achieves 92.0% weighted accuracy (F₁ = 0.91, AUROC = 0.95) while requiring only 1.1 × 10⁸ FLOPs, 110 ms latency, and 44 mJ per 16 s window measured on an ARM Cortex–M7. Ablation studies confirm the contribution of each module, and SHAP analysis highlights physiologically plausible channel importance. EdgeGT–Net bridges deep affective computing and ultra-low-power edge deployment, enabling always-on stress monitoring in consumer wearables.