AE-HGNN: attention-enhanced hypergraph neural networks for interpretable stress prediction through higher-order dependency modeling.

Stress is a significant contributor to the deterioration of mental and physical health, including conditions such as anxiety, depression, and cardiovascular issues. Traditional machine learning techniques like SVM, Decision Trees, and DNNs have been used for stress prediction using physiological and behavioral data, but they struggle to capture complex, high-order interactions among stress indicators and require extensive preprocessing of noisy signals. To address these limitations, a novel Stress Prediction System based on Attention Enhanced Hypergraph Neural Networks (AE-HGNN) is proposed that models higher-order relationships among environmental and behavioral factors such as humidity, temperature, and step count. The proposed attention mechanism adds attention weights to each factor; unlike conventional graph-based models limited to pairwise relationships, the attention enhanced hypergraph neural network leverages hyperedges to represent multi-node dependencies, significantly improving classification performance and robustness. The primary contributions of this work involve the design of an Attention Enhanced-HGNN architecture for stress classification that models higher-order dependencies by assigning attention weights to each factor. The proposed model achieves a test accuracy of 99.75% (5-fold cross-validated mean: 98.44% ± 0.56%) with optimized hyperparameters including a learning rate of 0.01, a hidden dimension of 128, and an epoch size of 150. This performance significantly outperforms existing methods such as Random Forest (87.3%), SVM (82.8%), and standard DNNs (90.7%), confirming the model's potential for non-invasive, real-time stress monitoring via wearable technologies and mobile health platforms.