---
license: mit
tags:
  - sleep-staging
  - wav2sleep
  - polysomnography
  - time-series
  - pytorch
library_name: wav2sleep
pipeline_tag: other
---

# wav2sleep-eog

EOG-based sleep staging (5-class: Wake, N1, N2, N3, REM)

## Model Description

This is a **wav2sleep** model for automatic sleep stage classification from electrooculography (EOG).
wav2sleep is a unified multi-modal deep learning approach that can process various combinations
of physiological signals for sleep staging.

- **Paper**: [wav2sleep: A Unified Multi-Modal Approach to Sleep Stage Classification](https://arxiv.org/abs/2411.04644)
- **Repository**: [GitHub](https://github.com/joncarter1/wav2sleep)
- **Conference**: ML4H 2024

## Model Details

| Property | Value |
|----------|-------|
| **Input Signals** | EOG-L, EOG-R |
| **Output Classes** | 5 |
| **Architecture** | Non-causal (bidirectional) |

### Signal Specifications

| Signal | Samples per 30s epoch |
|--------|----------------------|
| ECG, PPG | 1,024 |
| ABD, THX | 256 |
| EOG-L, EOG-R | 4,096 |

## Usage

```python
from wav2sleep import load_model

# Load model from Hugging Face Hub
model = load_model("hf://joncarter/wav2sleep-eog")

# Or load from local checkpoint
model = load_model("/path/to/checkpoint")
```

For inference on new data:

```python
from wav2sleep import load_model, predict_on_folder

model = load_model("hf://joncarter/wav2sleep-eog")
predict_on_folder(
    input_folder="/path/to/edf_files",
    output_folder="/path/to/predictions",
    model=model,
)
```

## Training Data

The model was trained on polysomnography data from multiple publicly available datasets
managed by the National Sleep Research Resource (NSRR), including SHHS and MESA.

## Citation

```bibtex
@misc{carter2024wav2sleep,
    title={wav2sleep: A Unified Multi-Modal Approach to Sleep Stage Classification from Physiological Signals},
    author={Jonathan F. Carter and Lionel Tarassenko},
    year={2024},
    eprint={2411.04644},
    archivePrefix={arXiv},
    primaryClass={cs.LG},
}
```

## License

MIT