Boosting automated sleep staging performance in big datasets using population subgrouping.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Automated sleep staging from electroencephalogram (EEG) data faces challenges due to individual variability and data inconsistencies.
• Current methods often underperform when training and test data exhibit significant differences.

Purpose of the Study:

• To develop a novel method for automated sleep staging that effectively learns relevant individuals based on EEG signal similarities.
• To improve the robustness and accuracy of deep learning models for sleep staging by addressing data diversity and potential artifacts.

Main Methods:

• Patients' EEG data were embedded into a shared feature space, allowing for clustering of individuals with similar statistical relationships.
• A deep learning framework was employed for classification after clustering similar patients.
• The method was validated using 994 patient EEGs from the 2018 Physionet Challenge (≈6,561 hours).

Main Results:

• The proposed clustering approach significantly boosted performance compared to state-of-the-art deep learning methods.
• Average precision score improved from 0.72 to 0.81.
• Average sensitivity score improved from 0.74 to 0.82, and Cohen's Kappa coefficient improved from 0.64 to 0.75 under 10-fold cross-validation.

Conclusions:

• Clustering similar individuals in a shared feature space is an effective strategy for enhancing automated sleep staging.
• This novel method demonstrates superior performance and robustness in classifying sleep stages from EEG data.