Research map / Architectures and efficiency / Attention and transformers
Neural decoding from stereotactic EEG: accounting for electrode variability across subjects
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces seegnificant, a transformer that tokenizes sEEG per electrode with 1D convolutions, applies time and electrode self-attention with MNI-coordinate RBF positional encodings, and trains a shared multi-subject model with subject-specific heads to decode trial-wise response time.
Problem
Cross-subject neural decoding from stereotactic EEG is hard because each subject has a variable number of electrodes at distinct, clinically determined brain locations, so there is no correspondence of recorded sites across individuals.
Approach
Tokenize each electrode independently with a shared temporal 1D convolution, inject 3D MNI location via RBF positional encodings, and alternate self-attention over time (within electrodes) and over electrodes (within timepoints) with subject-specific regression heads over a shared trunk.
Key results
- Multi-subject model achieved test R^2 of 0.54 on pooled data and 0.39 per-subject on average, versus 0.30 for single-subject models.
- Fine-tuning the multi-subject model per subject reached average per-subject R^2 of 0.41.
- Transfer to unseen subjects via few-shot finetuning achieved average R^2 of 0.38, versus about 0.30 for training from scratch.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train jointly on data from all subjects instead of training a separate model per subject | seegnificant · Reaction-time task, per-subject test set | R²: Δ R^{2} = 0.09 | Helped |
| Finetune the multi-subject model for each subject | seegnificant · Reaction-time task, per-subject test set | R²: Δ R^{2} = 0.11 | Helped |
| Pretrain on other subjects and finetune on the left-out subject rather than train from scratch | seegnificant · Reaction-time task, per-subject test set | R²: Δ R^{2} = 0.08 | Helped |
| Transfer the pretrained model to a new subject instead of using the multi-subject model | seegnificant · Reaction-time task, per-subject test set | R²: Δ R^{2} = -0.01 | Hurt |
| Use seegnificant rather than single-subject baseline models | seegnificant, single-subject models · Reaction-time task, per-subject test set | R²: Δ R^{2} ≥ 0.03 | Helped |
| Use the multi-subject model rather than single-subject baseline models | seegnificant, multi-subject model · Reaction-time task, per-subject test set | R²: Δ R^{2} ≥ 0.12 | Helped |
| Finetune the multi-subject model for each subject rather than use baseline models | seegnificant, finetuned multi-subject model · Reaction-time task, per-subject test set | R²: Δ R^{2} ≥ 0.14 | Helped |
| Transfer the pretrained model to a new subject rather than use baseline models | seegnificant, transferred model · Reaction-time task, per-subject test set | R²: Δ R^{2} ≥ 0.11 | Helped |
| Add subject-specific regression heads (ablation) | seegnificant multi-subject model · Reaction-time task, multi-subject test set | R²: Δ R^{2} = 18 | Helped |
| Add attention in the space dimension (ablation) | seegnificant multi-subject model · Reaction-time task, multi-subject test set | R²: Δ R^{2} = 0.10 | Helped |
| Replace separate time and space attention with combined 2D attention and remove positional encoding (ablation) | seegnificant · Reaction-time task, per-subject test set | R²: Δ R^{2} = 0.06 | Hurt |
| Replace separate time and space attention with combined 2D attention and remove positional encoding (ablation) | seegnificant · Reaction-time task, model training | Train time: ∼ 5.5 × | Hurt |
Surprising findings
- Removing spatial positional encoding changed per-subject R^2 by only -0.02 and was not statistically significant (p=0.73).
- Subject-specific regression heads were the most important ablated component, with an R^2 drop of 0.18.
Limitations
- Dataset is private and tied to a single behavioral task (color-change detection), so generalization across tasks is untested.
- Requires a fixed sampling rate across subjects, and electrodes are selected via high-gamma task responsiveness.
- Positional encoding gains are modest and not statistically significant.
Experimental setup
Datasets: in-house sEEG color-change detection task dataset (21 subjects)
Benchmarks: in-house sEEG response time decoding dataset
Metrics: R^2 RMSE
Baselines: Wiener Filter Ridge Regression Lasso Regression XGBoost MLP CNN+MLP 2D attention variant Fourier MNI positional encoding
Builds on: Neural Data Transformer standard transformer block AdamW
Closest papers by meaning
- CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding (NeurIPS 2025 · similarity 0.89)
- MindAlign: Bridging EEG, Vision, and Language for Zero-Shot Visual Decoding (NeurIPS 2026 · similarity 0.88)
- BrainOmni: A Brain Foundation Model for Unified EEG and MEG Signals (NeurIPS 2025 · similarity 0.88)
- What Does the Brain See? Multiview Neural Representation to Demystify the Brain-Visual Alignment (NeurIPS 2026 · similarity 0.88)
- CBraMod: A Criss-Cross Brain Foundation Model for EEG Decoding (ICLR 2025 · similarity 0.88)
- Toward Generalizing Visual Brain Decoding to Unseen Subjects (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.