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

Area of Science:

• Neuroscience
• Machine Learning
• Physiological Computing

Background:

• Real-time cognitive load monitoring is crucial but challenged by conventional methods' limitations.
• Existing machine learning models often require extensive data, are computationally intensive, or overlook causal dynamics.

Purpose of the Study:

• To develop an efficient framework for decoding cognitive load using causal spatiotemporal patterns from multimodal peripheral biosignals.
• To enable accurate, real-time cognitive load assessment in resource-constrained settings.

Main Methods:

• Novel feature engineering transformed short biosignal segments into image-like representations (Gramian Angular Difference Fields, Motif Difference Fields).
• Causal interdependencies were assessed using forward-backward copula Granger causality networks.
• A lightweight capsule neural network with self-attention classified fused multimodal features.

Main Results:

• Achieved up to 94% accuracy with 5-second signal segments on benchmark datasets (WESAD, CLAS).
• Demonstrated robust performance (84% accuracy) with 1-second windows, a rarely studied configuration.
• The model features only 323K trainable parameters, balancing complexity and performance.

Conclusions:

• The proposed framework offers a computationally efficient solution for real-time cognitive load assessment.
• It is suitable for resource-constrained environments and biofeedback applications.
• The approach effectively decodes cognitive load using minimal physiological data segments.