Block-Wise Domain Adaptation for Workload Prediction from fNIRS Data

Area of Science:

• Neuroscience
• Cognitive Science
• Biomedical Engineering

Background:

• Functional near-infrared spectroscopy (fNIRS) measures cortical hemodynamic activity non-intrusively.
• Predicting cognitive workload from fNIRS data faces challenges in generalizing across subjects and sessions.
• Existing methods often fail to perform well on unseen subjects due to high inter-subject and intra-subject data variability.

Purpose of the Study:

• To develop a robust method for predicting cognitive workload from fNIRS data that generalizes across subjects and sessions.
• To address the challenges of inter-subject and intra-subject variability in fNIRS data.
• To improve the applicability of cognitive workload prediction models in real-world settings.

Main Methods:

• Proposed a block-wise domain adaptation (BWise-DA) method to minimize intra-session variance by treating blocks from the same subject/session as different domains.
• Minimized intra-class domain discrepancy and maximized inter-class domain discrepancy.
• Introduced an MLPMixer-based model for workload prediction and a contrastive learning approach.

Main Results:

• The proposed BWise-DA method and MLPMixer model outperformed three baseline models on three public workload datasets (n-back and finger-tapping tasks).
• The contrastive learning method improved the performance of baseline models.
• Visualization confirmed that the models focused on relevant brain regions.

Conclusions:

• The BWise-DA method effectively enhances the generalization of cognitive workload prediction models using fNIRS data.
• The MLPMixer-based approach combined with contrastive learning offers a promising direction for accurate and reliable workload assessment.
• The findings support the use of fNIRS for real-world cognitive workload monitoring.