Multichannel Sleep Stage Classification and Transfer Learning using Convolutional Neural Networks.

Area of Science:

• Computational neuroscience
• Sleep medicine
• Machine learning applications in healthcare

Background:

• Current sleep medicine heavily relies on manual analysis of polysomnography (PSG) signals, including electroencephalogram (EEG), electromyogram (EMG), and electrooculogram (EOG).
• Automated sleep stage classification using machine learning, particularly deep learning models like Convolutional Neural Networks (CNNs), offers a potential solution to improve efficiency and consistency.

Purpose of the Study:

• To compare the performance of existing CNN approaches for automated sleep stage classification across diverse patient populations.
• To evaluate the impact of incorporating additional sensor data (EOG, EMG) alongside EEG.
• To develop and validate a transfer learning strategy to address data scarcity for less prevalent sleep disorders.

Main Methods:

• Comparative analysis of established CNN models using four distinct sleep databases (pathological and physiological subjects).
• Evaluation of CNN performance with and without supplementary EOG and EMG sensor data.
• Implementation of a transfer learning approach: pre-training a model on large public datasets and fine-tuning on smaller, specific datasets (e.g., REM Behaviour Disorder).

Main Results:

• The best-performing CNN model achieved Cohen's Kappa scores of 0.75 for healthy subjects and 0.64 for patients with sleep disorders.
• Inclusion of EOG and EMG data demonstrated advantages for sleep classification accuracy.
• The proposed transfer learning method significantly improved sleep classification by 24.4% on a private REM Behaviour Disorder database.

Conclusions:

• CNNs are effective tools for automated sleep stage classification from PSG data, with performance varying across different subject groups.
• Multimodal sensor data (EEG, EMG, EOG) enhances classification accuracy.
• Transfer learning is a viable and effective strategy to overcome data limitations in deep learning for rare sleep disorders, improving diagnostic capabilities.