Philosophers-Stone — Brain-Health Inference from Single-Channel Sleep EEG
Philosophers-Stone is a lightweight inference tool that converts a single-channel overnight sleep EEG into a quantitative index of brain health.
It applies a validated multi-cohort deep-learning model trained on 36,000 sleep recordings to estimate cognitive performance, disease likelihoods, and mortality-related physiological patterns.
The tool runs in seconds and outputs both a single Brain Health Score and a 1024-dimensional latent embedding suitable for research and biomarker discovery.
Code repository: https://github.com/wgbrain/Philosophers-Stone
Scientific study
If you use or reference this tool, please cite the peer-reviewed study:
Ganglberger, W., Sun, H., Turley, N., et al. and Westover, M.B. (2026) "Brain Health from Sleep EEG: A Multicohort, Deep Learning Biomarker for Cognition, Disease, and Mortality", NEJM AI, 3(3), DOI: 10.1056/AIoa2500487.
Available here.
Intended use
This model is intended for research use in sleep science, neurology, aging, and biomarker discovery.
It is not intended for clinical diagnosis, treatment decisions, or emergency use.
Who is this for?
- Sleep scientists
- Neurologists and dementia researchers
- Aging and cognitive-decline investigators
- Psychiatry researchers
- Data scientists working with physiological signals
- Clinical-trial teams exploring EEG-based biomarkers
What you get
- Brain Health Score (single interpretable metric)
- 1×1024 latent brain-health embedding (AI-derived sleep features)
- Predictions for cognition, disease risk, and mortality-related physiology
- Optional outputs: spectrograms and per-recording JSON summaries
Model provenance
This checkpoint implements the multi-task deep-learning framework described in:
Ganglberger, W., Sun, H., Turley, N., et al. and Westover, M.B. (2026) "Brain Health from Sleep EEG: A Multicohort, Deep Learning Biomarker for Cognition, Disease, and Mortality", NEJM AI, 3(3), DOI: 10.1056/AIoa2500487.
Requirements
- Python >= 3.10
- PyTorch 2.x (CUDA recommended)
- pandas, numpy, mne (for EDF), h5py, matplotlib, tqdm, psutil
Install dependencies:
pip install torch pandas numpy mne h5py matplotlib tqdm psutil
Model file
This Hugging Face repository hosts the checkpoint file used by the Philosophers-Stone codebase.
The current GitHub project auto-downloads it from this model repo when the local file is missing.
Inputs
Manifest CSV
A CSV with columns:
filepathage(years)sex(0=female, 1=male)
EEG file requirements
Philosophers-Stone accepts single-channel overnight EEG in HDF5 (.h5) or EDF (.edf) format.
Preferred channel: C4-M1.
| Format | Requirements |
|---|---|
HDF5 (.h5) |
Dataset: signals/c4-m1 (1-D float array, full night) Attributes: sampling_rate=200, unit_voltage="uV" Extra channels and annotations are ignored Manifest uses absolute paths |
EDF (.edf) |
Must contain a C4-M1 channel (label variants allowed) Any sampling rate accepted; auto-resampled to 200 Hz with anti-aliasing |
Sample full-night EEG data is included in the GitHub repository under ./sample-data/.
Quick start
python philosopher.py \
--manifest_csv phi_manifest.csv
Outputs
- Summary CSV (
phi_out/phi_results.csv) - Brain Health Score
- Disease and cognition-related outputs
- Latent embedding (
lhl_1tolhl_1024) - Optional JSON files under
phi_out/json/ - Optional spectrograms under
phi_out/figures/
Limitations
- The model is designed for single-channel overnight sleep EEG, primarily C4-M1.
- Performance may degrade when acquisition hardware, montages, preprocessing, or cohort characteristics differ from the development setting.
- Outputs should be interpreted as research biomarkers, not as standalone medical conclusions.
Performance tips
- Use a GPU if available
- Keep
batch_size=1
License
CC BY-NC 4.0 — Attribution-NonCommercial 4.0 International.
See the GitHub repository LICENSE file for details.