Deep Learning with Convolutional Neural Network for detecting microsleep states from EEG: A comparison between the...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Critical real-time decision-making occupations demand sustained attention.
• Repetitive tasks and extended work periods can lead to microsleeps, brief involuntary lapses in consciousness.
• Microsleeps disrupt cognitive and visual-motor coordination, impacting task performance.

Purpose of the Study:

• To evaluate the efficacy of a convolutional neural network (CNN) in detecting microsleep states.
• To analyze electroencephalogram (EEG) data for microsleep detection during a visuomotor task.
• To compare data imbalance handling techniques for CNN training.

Main Methods:

• Utilized 16-channel EEG data from 8 subjects performing a 1D visuomotor task.
• Analyzed EEG data in 4-second epochs with a 0.25-second step size.
• Employed leave-one-subject-out cross-validation and two imbalance strategies: oversampling and cost-based learning.

Main Results:

• Both oversampling and cost-based learning achieved comparable performance metrics (e.g., AUC_ROC=0.90, AUC_PR=0.41, phi=0.42).
• The dataset exhibited significant imbalance, with 17 responsive states for every microsleep state.
• Cost-based learning demonstrated a substantially reduced training duration compared to oversampling.

Conclusions:

• CNNs show promise for detecting microsleeps from EEG data.
• Cost-based learning is an efficient method for handling imbalanced EEG data in microsleep detection.
• This approach could enhance safety in occupations susceptible to microsleeps.