STST-JEPA: Shallow-Target Spatio-Temporal Joint Embedding Prediction Architecture For EEG Self-Supervised Learning
arXiv: 2607.06629 | Published: 2026-07-07 (v2: 2026-07-09) | Authors: Roy Segal, Yoni Svechinsky, Tomer Fekete
Core Innovation
Introduces STST-JEPA, a self-supervised transformer for resting-state and task EEG that combines two objectives:
- Latent-prediction objective: Predicting masked-token representations against an EMA-of-tokenizer target
- Auxiliary signal-reconstruction term: Applied to 30-second multi-channel windows under spatiotemporal block masks
Key Results
Pretraining Scale
- 47,703 EEG sessions spanning ages 5-81
- Data from TUH (Temple University Hospital) and Healthy Brain Network (HBN) corpora
- First EEG foundation model to cover the full pediatric-to-older-adult range
Brain-Age Regression
- MAE = 3.06 years (r = 0.924) on 3,367 held-out sessions
- Baseline: ~10 years MAE (predict-the-mean)
- Lightweight attentive probe on frozen pretrained embeddings
NeuralBench x EEGD Leaderboard (Rank-1)
Using native 30-second windows with light task-specific finetuning of final layers:
| Task | Metric | Score |
|---|
| Sex classification | Balanced accuracy | 0.911 |
| Age prediction | Pearson r | 0.749 |
| Psychopathology composite | Pearson r | 0.215 |
Cognitive Efficiency Correlation
- Age-prediction residual (predicted - chronological age) is negatively correlated with cognitive efficiency across several examined tasks
- This validates brain-age deviation as a meaningful biomarker
Architecture Details
STST-JEPA Design
Shallow-Target refers to predicting against an EMA (Exponential Moving Average) of the tokenizer's representations, rather than predicting raw signal or deep features. This approach:
- Provides stable, smooth targets for learning
- Avoids the collapse issues common in self-supervised learning
- Captures meaningful spatio-temporal structure in EEG
Spatiotemporal Block Masking:
- Masks are applied in both spatial (channel) and temporal (time) dimensions simultaneously
- Forces the model to learn cross-channel and cross-temporal dependencies
- More biologically plausible than random masking (models natural data loss patterns)
Dual Objective:
- Primary: JEPA-style latent prediction (predicting representations, not pixels/signal)
- Auxiliary: Signal reconstruction (ensures the learned representations retain signal fidelity)
Why This Works for EEG
- Cross-site heterogeneity: Self-supervised pretraining on massive diverse data learns representations robust to montage differences
- Small labeled cohorts: Frozen embeddings + lightweight probe avoids overfitting
- Subject-level non-stationarity: The EMA target provides stable learning signals despite individual variability
- Full age range: 47K sessions across 5-81 years enables learning developmental trajectories
Comparison to Prior EEG Foundation Models
Previous EEG foundation models struggled with:
- Limited pretraining data (typically <10K sessions)
- Narrow age ranges (often adults only)
- Poor cross-site generalization
- Weak performance on downstream tasks
STST-JEPA addresses all four limitations simultaneously.
Practical Applications
1. Brain-Age as Biomarker
- Use the pretrained model for rapid brain-age estimation from 30-second EEG windows
- Deviation from chronological age tracks neurological and psychiatric burden
- Applicable across full lifespan (pediatric to older adult)
2. EEG Classification/Regression
- Fine-tune the last layers for any downstream EEG task
- Native 30-second window handling (no need for arbitrary epoching)
- Achieves SOTA on sex, age, and psychopathology prediction
3. Cognitive Assessment
- Age-prediction residual correlates with cognitive efficiency
- Can serve as a screening tool for cognitive decline or enhancement
4. Cross-Site EEG Analysis
- Model handles montage heterogeneity natively
- Useful for multi-site studies and clinical deployment
Methodology for Reproduction
- Data: TUH + HBN corpora (47,703 sessions, ages 5-81)
- Preprocessing: Standardize to common montage, handle missing channels
- Pretraining:
- Spatiotemporal block masking on 30-second windows
- Dual objective: JEPA latent prediction + signal reconstruction
- EMA target for stable training
- Downstream:
- Frozen embeddings + lightweight probe (for brain-age)
- Light finetuning of final layers (for classification/regression)
Key Insight
The combination of shallow-target prediction (predicting EMA-of-tokenizer representations) with signal reconstruction creates representations that are both semantically meaningful and signal-faithful — addressing the core tension in self-supervised EEG learning between abstraction and fidelity.
Trigger Words
STST-JEPA, shallow-target JEPA, EEG self-supervised learning, EEG brain-age, EEG foundation model, NeuralBench EEGD, spatiotemporal block masking, EMA tokenizer, cognitive efficiency EEG, latent prediction EEG, EEG representation learning, brain-age biomarker
