Estimating distribution shifts for predicting cross-subject generalization in electroencephalography-based mental...

Area of Science:

• Neuroscience
• Machine Learning
• Human-Computer Interaction

Background:

• Assessing mental workload in real-world settings is crucial for worker performance, especially in tasks requiring sustained attention.
• Electroencephalography (EEG) is used for mental workload assessment, but its correlates vary significantly across subjects and physical strain, hindering reliable cross-user models.
• Existing domain adaptation methods for machine learning struggle with EEG data due to unmet assumptions about data distributions.

Purpose of the Study:

• To propose a strategy for estimating marginal and conditional shifts between EEG data distributions from different subjects.
• To investigate the impact of various normalization strategies on statistical shifts and their relationship with mental workload prediction accuracy.
• To provide a method for quantitatively assessing domain adaptation strategies in the context of EEG data.

Main Methods:

• Collected EEG data from individuals performing mental tasks while running on a treadmill and cycling on a stationary bike.
• Developed a strategy to estimate two types of discrepancies: marginal and conditional shifts between data distributions.
• Explored the effects of common normalization strategies used to reduce cross-subject variability in EEG data.

Main Results:

• The proposed approach effectively estimates statistical shifts between EEG data from different subjects.
• Different normalization strategies significantly impact these statistical shifts.
• The identified statistical shifts correlate with the accuracy of mental workload prediction on unseen participants.

Conclusions:

• The proposed method offers insights into data distribution assumptions for EEG-based mental workload analysis.
• Understanding and quantifying statistical shifts is vital for developing robust, cross-subject mental workload prediction models.
• This work highlights the importance of normalization strategies in mitigating cross-subject variability and improving model generalization.