COGNITIVE LOAD CLASSIFICATION FROM FUNCTIONAL NEAR-INFRARED SPECTROSCOPY DATA

Predicting cognitive workload using physiological sensors, such as functional near-infrared spectroscopy (fNIRS), has gained significant attention in recent years. However, many existing methods are limited by small datasets and narrow brain channel coverage, restricting their generalizability across participants, tasks, and sessions. In this thesis, we introduce two innovative approaches aimed at overcoming these limitations: a CNN-BiGRU model with self-supervised label augmentation (SLA) and a block-wise domain adaptation method (BWise-DA). Our CNN-BiGRU-SLA model, grounded in cognitive load theory, leverages attention mechanisms and multi-label classification to predict working memory load (WML) and visual processing load (VPL) across multiple participants and sessions. Using leave-one-participant-out (LOOCV) validation, we achieved F1-scores of 91.79% and 89.07% for binary classification of WML and VPL, respectively, and 79.72% and 79.68% for multi-level classification. The largest barrier to the application of fNIRS data is the high variance, which can arise between subjects and across sessions. In this thesis, we highlight that such variance can also occur within different blocks of the same session for the same subject. To address this, we introduce the BWise-DA approach, which effectively handles intra-subject and inter-subject variability by treating different blocks within the same session as distinct domains. This method minimizes intra-class domain discrepancy while maximizing inter-class domain distinction, enhancing the model’s ability to generalize. Additionally, we propose an MLPMixer-based model to further refine workload prediction. Experimental evaluations on three publicly available datasets—two involving n-back tasks and one involving finger tapping—demonstrate that our methods outperform baseline models. Furthermore, a visualization study shows that the models accurately focus on brain regions associated with the relevant tasks, confirming the model’s interpretability. These contributions establish a strong foundation for fNIRS-based workload prediction, paving the way for more accurate and generalizable models suitable for real-world applications.