Introduction
Sleep staging—categorizing sleep into phases like REM, N1, N2, and N3—is essential for diagnosing sleep disorders. Traditionally, this requires experts to manually score polysomnography (PSG) data in 30-second epochs, a process that is labor-intensive and prone to variability. While deep learning models have reached human-level performance, they often operate as "black boxes," lacking the interpretability required for clinical trust. Additionally, high-performing models are frequently too computationally heavy for deployment on edge devices or in resource-constrained clinical settings.
Methods
The researchers proposed the Cross-Modal Transformer (CMT), featuring two variants: the Epoch CMT for single 30s segments and the Sequence CMT for modeling temporal transitions across sequences.
The architecture leverages several key innovations:
- Multi-Scale 1D-CNNs: Three parallel convolutional paths with varying kernel sizes capture both local and global features from raw EEG and EOG signals.
[CLS]Token Strategy: Inspired by BERT and Vision Transformers, a learnable classification token aggregates information efficiently, reducing the need for deep attention layers and significantly lowering parameter count.- Cross-Modal Attention: This mechanism learns the relationships between EEG and EOG signals within the same window, improving classification for stages like "Wake" and "N1."
- Inter-Epoch Modeling: The sequence variant uses attention to capture the temporal flow of sleep stages, recognizing that previous epochs influence the current one.
Datasets and Validation
The model was evaluated on two public benchmarks:
- Sleep-EDF Expanded (2018): 153 whole-night recordings from 78 healthy subjects.
- SHHS (Sleep Heart Health Study): A large multicenter study, using a subset of 100 subjects for standardized comparison with existing literature.
Key Results and Efficiency
The Sequence CMT achieved a state-of-the-art accuracy of 84.3% and a Cohen’s Kappa of 0.785 on the Sleep-EDF dataset. Crucially, it maintains this performance while being significantly more efficient than existing models. It is approximately 11.5 times smaller in terms of parameters than SleepTransformer and trains 5 times faster than the current leader, XSleepNet.
Beyond raw metrics, the model offers improved interpretability. Visualization of attention maps reveals that the model focuses on clinically relevant features, such as eye movements for REM sleep and K-complexes for N2 sleep. This alignment with human expert logic is a significant step toward clinical adoption.
Conclusion
The Cross-Modal Transformer bridges the gap between high-performance deep learning and clinical interpretability. By utilizing efficient attention mechanisms and multi-scale feature extraction, it provides a lightweight yet powerful tool for automated sleep staging. Future work will focus on self-supervised pre-training to further enhance performance on smaller, specialized clinical datasets and addressing the quadratic computational cost of self-attention in extremely long recordings.
Summary updated