Contributor

Name: Yihan Zhang
NetID/Email: yihan20 yihan20@illinois.edu

Type of contribution & Original Paper

model contribution
original paper: https://arxiv.org/abs/2411.04644

High-level description

Wav2Sleep is a multi-modal architecture designed for automated sleep staging using synchronized raw signals (e.g., ECG and Respiratory).

Key Implementation Details:
• Temporal Encoders: Utilizes specialized CNN-based feature extractors for different sampling frequencies.
• Transformer Backbone: Implements a global Transformer block to capture long-range dependencies between sleep epochs.
• Stochastic Masking: Includes a robust fusion mechanism that handles missing modalities during training, as described in the original paper.
• Custom Output Handling: Implements a sequence-aware output head to bypass standard label processing limitations in pyhealth, ensuring compatibility with sequence prediction tasks.

File guide listing which files to review

• pyhealth/models/wav2sleep.py: Core model architecture and logic.
• pyhealth/models/init.py: Registration of Wav2Sleep.
• tests/core/test_wav2sleep.py: Unit tests covering instantiation, forward pass, output shapes, and gradient computation.
• examples/sleep_staging_wav2sleep.py: Comprehensive ablation study on synthetic data.
• docs/api/models/pyhealth.models.wav2sleep.rst: API documentation.
• docs/api/models.rst: Added Wav2Sleep entry to models table

Ablation Study Summary

I evaluated the model's sensitivity to two key hyperparameters using synthetic sleep staging data (5-stage classification). The experimental setup follows the data structure of the SHHS dataset supported by PyHealth, but uses synthetic tensors for fast reproducibility.

1. Embedding Dimension Ablation: testing the impact of latent space capacity.

2. Transformer Layers Ablation: testing the impact of architectural depth.

Findings & Conclusion
Optimal Capacity: The model performs best with a 128-dimensional embedding. Increasing to 256 leads to overfitting on small-scale synthetic data.
Depth Efficiency: A 2-layer Transformer backbone is sufficient for capturing temporal dependencies in these signals; additional layers do not yield significant gains.