MFSleepNet: An Interactive Multimodal Fusion Framework for Automatic Sleep Staging.

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Sleep Medicine

Background:

• Automatic sleep staging is crucial for diagnosing sleep disorders but faces challenges in integrating multimodal physiological signals like EEG and EOG.
• Existing methods often use simple feature concatenation, failing to capture complex inter-modality relationships essential for accurate sleep analysis.

Purpose of the Study:

• To develop MFSleepNet, a multimodal framework that explicitly models interactions between EEG and EOG signals for improved automatic sleep staging.
• To enhance the integration of heterogeneous physiological signals by enabling bidirectional information exchange and adaptive attention mechanisms.

Main Methods:

• MFSleepNet employs a multimodal feature fusion module for bidirectional EEG-EOG information exchange.
• A gated temporal-channel attention mechanism adaptively emphasizes informative temporal segments and signal channels.
• The framework was evaluated on Sleep-EDF, SHHS, and HSP datasets using epoch-level cross-validation and subject-specific adaptation.

Main Results:

• MFSleepNet significantly outperformed baseline single-modality and multimodal methods in accuracy, Cohen's κ, and Macro-F1 on public datasets.
• Ablation studies confirmed the effectiveness of individual modules, and correlation analysis revealed stage-dependent EEG-EOG relationships.
• External validation showed a performance drop on unseen subjects, addressed by a lightweight subject-specific adaptation strategy.

Conclusions:

• MFSleepNet offers an effective and interpretable solution for multimodal sleep staging by leveraging structured inter-modality interactions.
• Explicit feature interaction and subject-adaptive modeling are critical for improving generalization and practical applicability in sleep analysis.
• The study highlights the need for robust models that account for inter-subject variability in automatic sleep staging.