An iterative cross-subject negative-unlabeled learning algorithm for quantifying passive fatigue.

Area of Science:

• Neuroscience
• Machine Learning

Background:

• Detecting passive fatigue from EEG data is challenging due to the lack of labeled fatigue states.
• Existing methods often rely on manual fatigue state labeling, which is subjective and time-consuming.

Purpose of the Study:

• To propose an iterative negative-unlabeled (NU) learning algorithm for cross-subject detection of passive fatigue.
• To develop a method for quantifying driving fatigue scores from unlabeled EEG data.

Main Methods:

• An iterative NU learning algorithm was developed, utilizing labeled alert and unlabeled driving EEG data from 29 subjects.
• The algorithm iteratively identified the most fatigued EEG data blocks and computed fatigue scores.
• Repeated measures correlations were performed between fatigue scores and EEG band powers (theta, alpha, beta).

Main Results:

• The proposed algorithm achieved an average accuracy of 93.77% ± 8.15% in fatigue detection across subjects.
• Fatigue scores showed significant positive correlation with theta band power and negative correlation with beta and alpha band powers.
• The results demonstrated superior performance compared to various baseline methods.

Conclusions:

• The iterative NU learning algorithm effectively labels and quantifies passive fatigue from unlabeled EEG data in a cross-subject manner.
• The algorithm shows significant promise for real-world fatigue detection applications using EEG.
• Correlations with specific EEG band powers validate the algorithm's ability to capture fatigue-related neural changes.