Neural network STST-JEPA learned to determine biological age from EEG

Researchers presented STST-JEPA (Shallow-Target Spatio-Temporal Joint Embedding Prediction Architecture) — a transformer based on self-supervised learning for processing electroencephalography (EEG) both at rest and during cognitive tasks.

How the model works

The model uses latent prediction — prediction of masked token representations against a target updated by exponential moving average. This is augmented with an auxiliary reconstruction signal for 30-second windows of multichannel EEG with spatial-temporal block masking. The architecture is based on transformers, which allows the model to capture long-range dependencies in brain signals.

Training was conducted on two large EEG data corpora: brain.space and Healthy Brain Network (HBN), totaling 47,703 sessions.

Results: biological age prediction

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

Applying light fine-tuning of the final layers allowed the model to rank first on the public NeuralBench leaderboard on tasks:

• Gender classification: balanced accuracy 0.911
• Age prediction: r = 0.749
• Psychopathological composite regression: r = 0.215

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

Why this matters for medicine

Deviation of biological age from chronological age is a biomarker that can indicate neurological and psychiatric disorders, accelerated brain aging, or conversely, exceptional health. EEG is attractive for such screening because it is cheap, portable, and rich in temporal information.

However, EEG models face several challenges: heterogeneity of electrode montage across laboratories (different electrode configurations), small annotated cohorts, dominant non-stationarity at the subject level (temporal signal instability). STST-JEPA solves these problems through training on a large corpus of data in self-supervised mode.

What this means

Robotics and neuroimaging are honest tests for AI. If a model can predict biological age with precision of ±3 years from EEG across the full range from children to elderly people, it demonstrates real value of foundation models in medical applications. The next step is using STST-JEPA for diagnosing specific diseases (depression, autism, dementia) based on deviation of brain age from the norm.