Domain adaptation for robust workload level alignment between sessions and subjects using fNIRS.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Functional near-infrared spectroscopy (fNIRS) measures brain activity, but data variability across sessions and subjects (domain shift) hinders accurate workload classification.
• Aligning fNIRS data across different experimental conditions and participants is crucial for reliable brain-computer interfaces and cognitive state monitoring.

Purpose of the Study:

• To investigate the efficacy of domain adaptation techniques, specifically Gromov-Wasserstein (G-W) and fused Gromov-Wasserstein (FG-W), for aligning fNIRS data.
• To improve the classification of working memory workload levels in n-back tasks by addressing domain shift challenges.

Main Methods:

• Applied G-W for session-by-session fNIRS data alignment and FG-W for subject-by-subject alignment.
• Utilized labeled data from one session/subject to classify trials in another session/subject during n-back tasks.
• Compared G-W and FG-W performance against supervised methods like SVM, CNN, and RNN, and assessed the impact of motion artifact removal.

Main Results:

• G-W achieved 68% ± 4% accuracy for session-by-session alignment, while FG-W reached 55% ± 2% for subject-by-subject alignment.
• Both G-W and FG-W significantly outperformed SVM, CNN, and RNN classifiers.
• Effective removal of motion artifacts was shown to be critical for enhancing alignment performance.

Conclusions:

• Domain adaptation methods, particularly G-W and FG-W, demonstrate significant potential for aligning fNIRS data.
• These techniques enable more robust classification of mental workload across varying experimental sessions and subjects.
• fNIRS data alignment using domain adaptation offers a promising avenue for advancing brain-computer interfaces and cognitive workload assessment.