Efficient cognitive load decoding using causal spatiotemporal patterns from multimodal physiological signals

Monitoring cognitive load is critical in diverse, demanding environments, yet conventional assessment methods face limitations in real-time applicability. While machine learning approaches using physiological signals show promise, they often require long data segments, exhibit high computational complexity, or neglect underlying causal dynamics. This paper proposes an efficient framework for cognitive load decoding using causal spatiotemporal patterns derived from multimodal peripheral biosignals. We introduce a novel feature engineering pipeline that transforms short signal segments into image-like representations capturing temporal dynamics via Gramian Angular Difference Fields and Motif Difference Fields, alongside causal interdependencies assessed using forward-backward copula Granger causality networks. These fused multimodal features are classified using a lightweight capsule neural network employing a self-attention routing mechanism. To evaluate the proposed solution, we conducted experiments on two widely used benchmark datasets, WESAD and CLAS. Our model achieved up to 94% accuracy using only 5-second signal segments, and maintained robust performance (84% accuracy) even with 1-second windows, a configuration rarely addressed in prior research. The proposed architecture includes only 323K trainable parameters, offering a favorable balance between model complexity and classification performance. The results confirm the framework's potential for computationally efficient, real-time cognitive load assessment suitable for resource-constrained environments and biofeedback applications.