Explainable Sleep Stage Classification with Multimodal Electrophysiology Time-series.

Area of Science:

• Computational Neuroscience
• Machine Learning in Healthcare
• Sleep Medicine

Background:

• Deep learning models are increasingly used for automated sleep staging.
• Multimodal data (EEG, EOG, EMG) improve classification performance.
• Model explainability in multimodal sleep staging remains underexplored.

Purpose of the Study:

• To develop and evaluate a novel ablation approach for multimodal sleep stage classification.
• To improve the explainability of deep learning models using EEG, EOG, and EMG data.
• To compare the proposed method with traditional 'zero-out' ablation techniques.

Main Methods:

• Trained a convolutional neural network (CNN) for sleep stage classification using EEG, EOG, and EMG.
• Proposed a new ablation method replacing modalities with line-noise approximations.
• Compared the new method with a 'zero-out' ablation approach.

Main Results:

• Identified relative modality importance consistent with sleep staging guidelines (EEG crucial for most stages, EOG for REM/non-REM).
• EMG showed low importance across all sleep stages.
• The proposed ablation method accentuated modality importance for certain stages (e.g., REM) compared to 'zero-out'.

Conclusions:

• Domain-specific ablation methods offer clearer insights into modality importance for multimodal electrophysiology data.
• The findings provide guidance for applying explainability methods to clinical machine learning.
• Careful consideration of explainability method interaction with clinical data is crucial.