To address the challenge of accurately and reliably monitoring individuals' psychological states from multimodal physiological signals, this study aims to solve two core issues of existing deep learning methods: the lack of collaborative modeling and the unreliability of prediction results. This study proposes an innovative framework named Graph Attention-Bayesian Temporal Network (GABT-Net). First, the method uses Graph Attention Network (GAT) to construct each moment's multi-source physiological signals into a "physiological collaboration graph" to explicitly capture the instantaneous collaborative relationships among various indicators. Then, it introduces Bayesian Transformer and performs probabilistic prediction on the time series of collaborative states through Monte Carlo (MC) Dropout technology, thereby realizing the quantification of the model's own uncertainty. Experimental verification conducted on the public Wearable Stress and Affect Detection (WESAD) dataset shows that the GABT-Net model achieves an average accuracy of 96.2% and a macro-average F1-score of 0.9431, both of which outperform the baseline models. In terms of reliability, the Expected Calibration Error (ECE) of this model is 0.021, and the prediction accuracy on samples with the highest confidence reaches 99.10%. Ablation experiment and graph construction strategy analysis further verify the key role of GAT collaborative modeling and Bayesian time series module in performance and reliability.
Cheng et al. (Wed,) studied this question.