Automatic sleep staging using fMRI functional connectivity data.

Area of Science:

• Neuroscience
• Brain Imaging
• Machine Learning

Background:

• Sleep stages alter brain activity and functional connectivity, impacting resting-state fMRI (rs-fMRI) data.
• Lack of vigilance monitoring in rs-fMRI experiments can introduce confounds due to participants falling asleep.
• Simultaneous electroencephalography (EEG) and fMRI are technically demanding for sleep scoring.

Purpose of the Study:

• To develop an automatic sleep staging method using only fMRI functional connectivity data.
• To provide vigilance information during rs-fMRI scans without EEG.
• To avoid confounds in rs-fMRI analyses caused by undetected sleep.

Main Methods:

• Feature extraction based on linear correlations between 20 cortical regions and bilateral thalamus regions.
• Utilizing independent component analysis (ICA) for region identification.
• Constructing binary support vector machine (SVM) classifiers for sleep stage discrimination, combined into multiclass classifiers.

Main Results:

• Achieved accuracies over 0.8 in binary sleep stage classification using 60-second fMRI epochs.
• Demonstrated good generalization to two independent datasets with accuracies over 0.8.
• The multiclass classifier effectively applies to any dataset via a sliding window approach.

Conclusions:

• The developed method offers a practical solution for monitoring vigilance levels within an MRI scanner.
• This approach eliminates the need for additional recordings beyond fMRI BOLD signals.
• Accurate vigilance modeling in rs-fMRI analysis prevents confounded inferences and aids the study of vigilance states.