Philosopher's Stone

Abstract

This repository contains data and code for the paper Brain health from sleep EEG: A multi-cohort, deep learning biomarker for cognition, disease and mortality. We developed an end-to-end deep learning model, the Philosopher’s Stone, that learns directly from overnight sleep EEG to generate a latent brain health space and a single brain health score. This biomarker predicts cognition, disease, and mortality more accurately than demographic or classical EEG models and integrates both known physiological markers and novel EEG patterns.

Full code and open-source, pre-trained model are available at: https://github.com/bdsp-core/philosophers-stone

Model outputs (Brain Health Score, 1024 latent space) for all data used in the paper, training code and analysis workflows are provided via the linked AWS storage.

Background

Sleep EEG contains rich information about brain and systemic health, but traditional analyses rely on predefined features that may miss subtle patterns. We aimed to develop a scalable, data-driven biomarker that links sleep EEG to cognition, disease, and mortality across large, diverse populations.

We analyzed 36,000 polysomnography recordings from 27,000 subjects across six cohorts (FHS, MESA, MrOS, SOF, KoGES, MGH). A multi-task deep neural network processed raw EEG (time-series and spectrograms) to learn a 1024-dimensional latent representation. From this, we derived a single brain health score optimized to reflect better cognition, lower disease risk, and longer survival. Performance was benchmarked against demographic and expert-defined EEG features using cross-validation.

Software Description

The repository provides deidentified model outputs, latent embeddings, and the trained Philosopher’s Stone model. Inputs include overnight EEG recordings and corresponding cognitive, disease, and survival labels across cohorts. Outputs include the brain health score, latent space embeddings, and task-specific predictions (cognition, disease, sleep metrics).

Installation and Requirements

To use the trained model and apply it on your data, please find everything in the GitHub repository.

An API allows users to compute latent representations and brain health scores from new EEG data. Example notebooks reproduce key results, including prediction analyses, UMAP visualizations, and salience maps.

Usage Notes

See GitHub

Release Notes

Version 1.0

Ethics

All data were collected under IRB-approved protocols at participating institutions, including BIDMC (#2024P000804) and MGH (#2013P001024), with waivers of consent where applicable. External cohort data (FHS, MESA, MrOS, SOF, KoGES) were accessed under institutional agreements. All shared data are deidentified in accordance with HIPAA and cohort-specific guidelines.

Acknowledgements

This work was supported by NIH grants R01AG073410, R01HL161253, RF1AG064312, RF1NS120947, R01NS126282, R01AG073598, R01NS131347, R01NS130119, and cohort-specific funding (FHS, MESA, MrOS, SOF, KoGES). We thank all study participants and staff for their contributions.

Conflicts of Interest

M.B. Westover is co-founder and equity holder of Beacon Biosignals. R.J. Thomas reports patents/licensing with MyCardio LLC and DeVilbiss-Drive, and consulting for Jazz Pharmaceuticals, Guidepoint, and GLG Councils. R. Au is a scientific advisor to Signant Health and NovoNordisk. The remaining authors declare no competing interests.

References

1. Brain health from sleep EEG: A multi-cohort, deep learning biomarker for cognition, disease and mortality; Ganglberger et al., 2025