Unsupervised domain transfer: Overcoming signal degradation in sleep monitoring by increasing scoring realism

1. Problem Statement

The paper addresses the challenge of signal degradation in mobile and "in-the-wild" sleep monitoring. While deep learning models (like U-Sleep) perform well on clean, clinical polysomnography (PSG) data, their performance often collapses when applied to real-world data affected by noise, device variability, or recording artifacts.

• The Core Issue: Standard supervised transfer learning requires labeled ground-truth data from the target domain (the noisy data), which is rarely available in practical deployment scenarios.
• The Goal: To develop an unsupervised domain adaptation method that can adapt a pre-trained sleep scoring model to distorted target data without access to manual labels, thereby maintaining scoring accuracy despite signal degradation.

2. Methodology

The authors propose a "Discriminator-guided fine-tuning" framework based on adversarial domain adaptation. The core hypothesis is that "realistic" hypnograms (sleep stage sequences) occupy a sparse subspace of all possible sequences. By forcing the model to generate hypnograms that look "realistic" (i.e., indistinguishable from those generated on clean data), the model can effectively disentangle signal from noise.

2.1 Architecture

The system consists of three main components:

1. Source Encoder ($E_s$): A frozen, pre-trained U-Sleep encoder that processes clean data. It provides a fixed feature representation of "clean" sleep.
2. Target Encoder ($E_t$): Initialized with the same pre-trained weights as $E_s$ but fine-tuned during training. It processes distorted target data.
3. Shared Decoder ($Dec$): A frozen U-Sleep decoder that converts features into sleep stage predictions.
4. Discriminator ($D$): A lightweight Transformer-based binary classifier. It takes the feature outputs from both encoders and attempts to distinguish whether they originated from the source (clean) or target (distorted) domain.

2.2 Training Strategy

The training involves a min-max game between the Target Encoder and the Discriminator, guided by three loss functions:

• Adversarial Loss ($L_{adv}$): The Discriminator is trained to correctly classify source vs. target features. Conversely, the Target Encoder is trained to fool the Discriminator (making target features look like source features).
• Anchor Loss ($L_{anchor}$): To prevent the model from collapsing or drifting too far from the pre-trained knowledge, an anchor loss penalizes the difference between the features extracted by $E_t$ and $E_s$ when both are fed the same clean source input. This enforces consistency and assumes a "covariate shift."
• Optimization: The Target Encoder is updated to minimize $\alpha L_{adv} + \beta L_{anchor}$ (with $\alpha=1, \beta=0.1$), while the Discriminator minimizes the binary cross-entropy loss.

2.3 Data and Distortions

The study uses a diverse collection of open-source datasets (e.g., MASS, ISRUC, DOD) as the Source Domain. To create the Target Domain, the authors artificially introduced three types of realistic signal distortions:

1. White Noise: Simulating broken channels (replacing EEG or EOG channels with white noise).
2. Amplifier Overload: Simulating electrode connection issues via bi-exponential transient spikes.
3. Spectral Deformation: Simulating hardware/physiological frequency shifts using band-pass filters (removing low frequencies).

3. Key Contributions

1. Realism as a Constraint: The paper introduces the concept of using "hypnogram realism" as a guiding constraint for unsupervised adaptation. It posits that correct sleep staging follows specific temporal patterns (sparsity in the hypnogram space) that can be learned via adversarial training.
2. Unsupervised Adaptation for Signal Degradation: Unlike previous works focusing on domain shifts between datasets, this approach specifically targets signal artifacts (noise, spikes, filtering) without requiring any ground-truth labels from the noisy domain.
3. Comprehensive Evaluation: The study provides a rigorous comparison against:
   • The pre-trained baseline.
   • A "Benchmark" model trained in a fully supervised manner on the distorted data (representing the theoretical upper bound).
   • A "Chance Kappa" analysis to ensure the model isn't just outputting generic, repetitive sleep patterns.

4. Results

The method was evaluated on thousands of recordings across different distortion types.

• Performance Gains:
  • White Noise: Fine-tuning improved Cohen's Kappa by 0.16 to 0.29 (depending on whether EEG or EOG was distorted) compared to the pre-trained model.
  • Amplifier Overload: Improvements of 0.08 to 0.11 in Kappa were observed across different distortion frequencies.
  • Spectral Deformation: Improvements ranged from 0.03 to 0.16, with the most significant gains seen in aggressive high-pass filtering scenarios.
• Comparison to Benchmarks: While the unsupervised method significantly outperformed the pre-trained baseline, it did not reach the performance of the supervised benchmark models (which had access to ground truth labels).
• Real-World Test: When tested on a real-world domain mismatch (transferring from the source datasets to the Tabar et al. dataset), the improvement was insignificant. The authors attribute this to the small size of the target dataset (120 recordings) compared to the thousands required for the method to converge effectively.
• Statistical Significance: All improvements over the baseline were statistically significant ($p < 10^{-4}$) and well outside the "chance" kappa distribution.

5. Significance and Limitations

• Significance:
  • The approach demonstrates that adversarial learning can effectively handle signal degradation in sleep monitoring without manual labels.
  • It validates the hypothesis that enforcing "realism" in sleep stage sequences helps models ignore noise and focus on physiological patterns.
  • It offers a potential pathway for deploying sleep monitoring in "in-the-wild" settings where signal quality is unpredictable.
• Limitations & Future Work:
  • Data Efficiency: The method currently requires thousands of training samples to be effective. The authors note a need to reduce this requirement by 1–2 orders of magnitude for clinical viability.
  • Gap to Supervised Performance: The unsupervised method still lags behind supervised fine-tuning, suggesting that while it recovers much of the lost performance, it cannot fully reconstruct the signal quality of clean data.
  • Complex Mismatches: The method struggled with complex, real-world domain shifts (different datasets) compared to synthetic distortions, indicating that the "covariate shift" assumption may not hold for all types of domain mismatches.

In summary, the paper presents a promising unsupervised technique for robust sleep scoring that leverages the structural constraints of sleep data to overcome signal degradation, though further development is needed to reduce data requirements and close the gap with supervised methods.