Brain-Health Inference from Sleep EEG

Philosopher's Stone is an inference tool that converts a single-channel overnight sleep EEG into a quantitative index of brain health.

It applies a peer-reviewed multi-cohort deep-learning model trained on 36,000 sleep recordings to estimate a Brain Health Score, cognition-related outputs, disease-related outputs, mortality-related physiological patterns, and a 1024-dimensional latent embedding for research workflows.

Scientific Study

If you use or reference this tool, please cite:

Ganglberger, W., Sun, H., Turley, N., ... & Westover, M. B. (2026).
Brain Health from Sleep EEG: A Multicohort, Deep Learning Biomarker for
Cognition, Disease, and Mortality.
NEJM AI, 3(3), AIoa2500487.

Article DOI: https://doi.org/10.1056/AIoa2500487

Repository copies:

• Paper PDF
• Supplementary Appendix

Intended Use

This repository is intended for research use by sleep scientists, neurology and aging researchers, psychiatry researchers, clinical-trial teams, and data scientists working with physiological signals.

This software is not a medical device and is not intended for diagnosis, treatment decisions, or clinical use without independent validation and appropriate regulatory review.

Installation

Python 3.10+ is required. Conda is recommended because the stack includes PyTorch, MNE, scipy, and EEG signal-processing dependencies.

Run the bundled sample end to end:

./run_sample.sh

For application development:

pip install -e .

The package installs a CLI:

philosophers-stone --manifest_csv phi_manifest.csv --outdir phi_out

The legacy source-checkout command remains available:

python philosopher.py --manifest_csv phi_manifest.csv --outdir phi_out

Docker is also supported. See docker/README.md.

Model Checkpoint

The model checkpoint is about 2.3 GB and is intentionally not tracked in Git.

Checkpoint lookup order:

1. PHILOSOPHER_MODEL_FILE, if set.
2. ./model_files/SleepPhilosophersStone.ckpt in a source checkout.
3. ~/.cache/philosophers-stone/model_files/SleepPhilosophersStone.ckpt.

If the checkpoint is missing, inference can auto-download it from:

https://huggingface.co/wolfgang-ganglberger/philosophers-stone

Inputs

The CLI takes a manifest CSV with these columns:

• filepath: path to an EEG file.
• age: age in years.
• sex: 0 for female, 1 for male.

Supported EEG formats:

Format	Requirements
HDF5 (.h5)	Dataset signals/c4-m1, 1-D float array, full night; attributes sampling_rate=200 and unit_voltage="uV".
EDF (.edf)	Must contain a C4-M1 channel or accepted label variant; any sampling rate is resampled to 200 Hz with anti-aliasing.

Sample full-night EEG files are included under sample-data/.

Python API

Preferred import path:

from philosophers_stone import Config, infer_brain_health

result = infer_brain_health(
    eeg_uv,
    fs_hz=200,
    age=65,
    sex=1,
    file_id="study-001",
    cfg=Config(),
)

The legacy import path remains available for compatibility:

from phi_utils.philosopher_utils import Config, infer_brain_health

eeg_uv must be one overnight EEG channel in microvolts. The array-based API does not write files; callers decide where to persist scores, latent features, and optional stage probabilities.

Outputs

Default summary CSV:

phi_out/phi_results.csv

Columns include:

file_id, filepath, age, sex, brain_health_score,
total_cognition_score, fluid_cognition_score,
crystallized_cognition_score, lhl_1 ... lhl_1024

Optional outputs:

• JSON summaries under <outdir>/json/.
• Spectrogram figures under <outdir>/figures/.

Repository Layout

• src/philosophers_stone/: runtime package.
• src/philosophers_stone/_vendor/timm/: minimal vendored timm runtime subset.
• phi_utils/: source-checkout compatibility shims for historic imports.
• sample-data/: bundled sample EEG files.
• paper/: paper and reproducibility assets, not required at runtime.
• docker/: Docker build and usage files.

License

CC BY-NC 4.0 - Attribution-NonCommercial 4.0 International. See LICENSE.