Early warning of operators’ thermal-discomfort-induced cognitive impairment via ECG-fNIRS monitoring and dynamic spatio-temporal graph learning model

Thermal discomfort significantly impairs cognitive functions, leading to human errors and even severe system accidents. Existing deep learning methods for cognitive assessment via physiological monitoring often neglect the spatiotemporal dynamics of neurophysiological signals, focusing primarily on real-time classification. To address this limitation, this paper proposes DCATT (Diffusion Convolutional Attention Model), a novel spatiotemporal framework for multistep prediction of operators’ abnormal cognitive states under heat stress, comprising three key modules: (1) diffusion graph convolution over functional brain network to capture dynamic brain activity and enhance short-term forecasting; (2) multi-head self-attention to aggregate long-range temporal dependencies for improved long-term prediction; and (3) a position-encoding-as-query cross-attention mechanism that adaptively projects prior physiological embeddings into stable, latent representations of future cognitive states. Evaluated on a multimodal dataset of operators under thermal stress, DCATT predicts cognitive impairment 30 s in advance with 72% accuracy, achieving an AUC of 72.4% and an F1-score of 67.4%, outperforming current benchmarks by 4.2 and 5.6 percentage points, respectively. The model also demonstrates strong interpretability, highlights task-relevant brain regions, and exhibits robust generalization across subjects, sensor noise, and missing-channel conditions. DCATT offers a promising framework for early warning of cognitive impairment and dynamic risk management in emergency scenarios, particularly enhancing process safety and productivity in human-cybernetic systems in the Industry 5.0 era.