Thanks! The link/thumbnail in the first post gets you to the free version of the article.

Thanks for sharing! For anyone reading this in wake, I posted a quick thread with the 'so what?' + key takeaways:

Add-on "EEG receipt”: spectrogram + Integrated Gradients show what the model fixated on across TP/FP/FN/TN during a full night of sleep. High brain health scorers: clean cycles + early delta/spindles; low scorers: weaker slow waves/spindles + alpha, FP/FN likely the “it’s complicated" cases.

Add-on: We squeezed the 1024-D brain-health embeddings into a 2-D UMAP and surprise: age, brain health score, REM %, fragmentation, delta/slow-oscillation power, and spindle density form clean gradients. The model has built a neurophysiology theme park that hints at brain-health phenotypes.

These gains, and analyses of the learned latent space, suggest the model leverages both known EEG markers and novel features to drive performance. Overall, a multitask end-to-end approach yielded an interpretable, sleep-derived brain health biomarker.

Mortality: In age-adjusted Cox models, +1 SD in Brain Health Score was linked to ~31–35% lower mortality risk (HR≈0.65–0.69, P<0.0001), beating conventional EEG metrics.

Disease: Classification AUROC improved from ~0.50–0.55 (baseline) to ~0.65–0.75 (across conditions like dementia, depression, hypertension).

Results: Our single EEG-derived Brain Health Score tracked all three outcome domains simultaneously — cognition, disease, and mortality — and outperformed the baselines:
Cognition: Correlation with cognitive scores rose from small (demographics-only) to moderate (~r = 0.40).

Benchmarking: We tested whether our learned EEG score adds value beyond standard models. We compared it to a demographics-only baseline, conventional EEG summary features (REM sleep %, spindle density), and classic multivariate ML models built on those features.

We analyzed 36,000 overnight EEG recordings. EEG was fed into an end-to-end multitask deep neural network. It learned a 1024-dimensional latent “brain-health” space by jointly predicting cognitive test scores and disease status. We then distilled this latent space into a single Brain Health Score.

Sleep underpins cognition, disease prevention, and brain health, yet we lack an integrative biomarker. We tested whether a deep-learning model can learn a latent brain-health representation from whole-night EEG and distill it into one Brain Health Score linked to cognition, disease, and mortality.

Brain Health from Sleep EEG: A Multicohort, Deep Learning Biomarker for Cognition, Disease, and Mortality Sleep underpins cognition, disease prevention, and overall brain health, yet objective, integrative biomarkers of brain health remain lacking. We hypothesized that overnight sleep electroencephalog...

I’m working on rebuilding my science bubble here. Sleep/EEG/neuro/psych/biosignals folks, hello!

Starting with good news: Tried alchemy. Didn’t make gold. Made a Brain Health Score from overnight sleep EEG that tracks cognition, disease, and mortality. Philosopher’s Stone is out (NEJM AI)