Biosignals Based Automated Driver Cognitive Load Assessment Using a Pre-Trained Transformer

Assessing driver cognitive load (DCL) is essential for enhancing driving safety and performance. This study introduces a novel method that leverages a pre-trained biosignal transformer (BIOT) to classify DCL using unimodal and multimodal multivariate biosignals. The proposed transformer-based model features an encoder-only architecture with efficient multi-head self-attention to process input tokens derived from various biosignals. Our approach includes Fourier-based biosignal tokenization techniques and a perturbation module as a data augmentation layer. Additionally, we employed a hierarchical spatiotemporal embedding framework, using additive absolute position embeddings for temporal encoding and additive learnable tokens for each channel for spatial encoding. Experimental results demonstrate that our model, trained on raw biosignals, outperforms previous methods, highlighting its potential for monitoring DCL using wearable devices. It achieved error rates that were at least 2.4% and up to 20% lower in ternary classification tests when compared to the baseline models. The findings suggest that integrating a pre-trained transformer-based architecture with perturbation-based augmentation can significantly enhance the model's accuracy and robustness. Our results offer a promising direction for future research and development in intelligent transportation systems.