Channel Contribution in Deep Learning Based Automatic Sleep Scoring-How Many Channels Do We Need?

Area of Science:

• Artificial Intelligence in Medicine
• Computational Neuroscience
• Biomedical Signal Processing

Background:

• Machine learning automates sleep stage annotation from polysomnograms.
• Existing multi-channel, multi-modal models show limited improvement over single-channel EEG models.
• The reasons for this performance gap are not fully understood.

Purpose of the Study:

• Investigate if single-channel EEG model features contribute to high performance.
• Determine how multi-channel, multi-modal models utilize information from various channels.
• Evaluate the impact of transferring specific deep learning architectures from single-channel models.

Main Methods:

• Transferred deep learning model features (e.g., CNN filter combinations) from single-channel EEG models to multi-channel, multi-modal models.
• Applied layer-wise relevance propagation (post-hoc) and embedded channel attention networks (intrinsic) for interpretability.
• Quantified the contribution of individual channels to predictive performance.

Main Results:

• Incorporating single-channel model features improved multi-channel, multi-modal model performance.
• Multi-channel, multi-modal models primarily focused on one key channel per modality.
• Remaining channels were utilized to supplement information from the primary focused channel.

Conclusions:

• Specific architectural features from high-performing single-channel EEG models can benefit multi-channel, multi-modal approaches.
• Current multi-channel, multi-modal models exhibit a focused channel usage pattern.
• Developing advanced channel information aggregation techniques is crucial for enhancing sleep scoring performance.