STST-JEPA neural network learns to estimate biological age from EEG

Researchers have introduced STST-JEPA (Shallow-Target Spatio-Temporal Joint Embedding Prediction Architecture), a transformer based on self-supervised learning that processes electroencephalography (EEG) both at rest and during cognitive tasks.

How the model works

The model uses latent prediction: it predicts masked token representations against a target updated by an exponential moving average. An auxiliary reconstruction signal is added for 30-second windows of multichannel EEG with spatio-temporal block masking. The transformer-based architecture lets the model capture long-range dependencies in brain signals.

Training used two large EEG corpora, brain.space and the Healthy Brain Network (HBN), totaling 47,703 sessions.

Results: biological age prediction

Brain age (biological age) prediction is one of the key applications. On 3,367 test sessions, the model reached a mean absolute error (MAE) of 3.06 years with a correlation of r = 0.924. For comparison, the baseline approach (guessing the mean value) gives an error of about 10 years.

Lightweight fine-tuning of the final layers put the model in first place on the public NeuralBench leaderboard for these tasks:

• Sex classification: balanced accuracy 0.911
• Age prediction: r = 0.749
• Psychopathology composite regression: r = 0.215

Interestingly, the residual error in the model's age prediction correlates negatively with cognitive performance on several of the tasks tested.

Why it matters for medicine

The gap between biological and chronological age is a biomarker that can point to neurological and psychiatric disorders, accelerated brain aging, or, conversely, exceptional health. EEG is attractive for this kind of screening because it is cheap, portable and rich in temporal information.

However, EEG models face several problems: montage heterogeneity between labs (different electrode configurations), small labeled cohorts, and dominant subject-level nonstationarity (signal instability over time). STST-JEPA addresses these problems by training on a large corpus in a self-supervised regime.

What this means

Robotics and neuroimaging are honest tests for AI. If a model can predict biological age from EEG to within ±3 years across the full range from children to older adults, it shows the real value of foundation models in medical use. The next step is to use STST-JEPA to diagnose specific conditions (depression, autism, dementia) based on how far brain age deviates from the norm.