BaRISTA: Brain-Scale Informed Spatiotemporal Representation of Human Intracranial EEG

Conference: NeurIPS 2025 arXiv: 2512.12135 Code: https://github.com/ShanechiLab/BaRISTA Area: Neuroscience / Foundation Models Keywords: Intracranial EEG, Spatiotemporal Transformer, Spatial Encoding Scale, Masked Reconstruction, Pretraining

TL;DR

BaRISTA systematically investigates spatial encoding scales (electrode/parcel/lobe) for iEEG Transformers, finding that atlas parcel-level encoding combined with spatial masked reconstruction achieves 86.2% AUC on language task decoding (vs. PopT 79.5%). The choice of encoding scale has greater impact than masking strategy, and the model generalizes well across subjects.

Background & Motivation

Background: iEEG provides high spatiotemporal resolution brain activity recordings. Transformer-based pretrained models (PopT, Brant) have been applied to iEEG, but the choice of spatial encoding has not been systematically studied.

Limitations of Prior Work: Electrode positions vary entirely across patients, making channel-level encoding difficult to generalize across subjects. Whether parcel-level or lobe-level encoding is superior, and how spatial encoding interacts with masking strategy, remain open questions.

Key Challenge: Fine-grained spatial resolution (channel) carries the most information but lacks cross-patient consistency; coarse-grained resolution (lobe) is consistent across patients but may sacrifice local information.

Goal: Systematically compare three spatial encoding scales to identify the optimal encoding–masking combination.

Key Insight: Channel, atlas parcel, and lobe are treated as experimental variables and systematically ablated within a masked reconstruction pretraining framework.

Core Idea: By systematically comparing three spatial encoding scales in an iEEG Transformer, the paper identifies atlas parcel-level encoding as the optimal spatial granularity—balancing cross-patient consistency with local information retention.

Method

Overall Architecture

iEEG data (2048 Hz) → Temporal tokenization (Dilated CNN extracting 250 ms patches) → Spatial encoding (learnable embeddings \(E_j\) at channel/atlas parcel/lobe granularity) → token \(S_{ij} = B_{ij} + E_j\) → Spatial masking (randomly selected spatial categories) → Transformer (12 layers / 4 heads / \(d=64\) + RoPE) → EMA target reconstruction (MSE between online encoder and EMA target encoder)

Key Designs

1. Three Spatial Encoding Scales:

   • Channel: MNI coordinates \((x, y, z)\) → learnable embeddings (finest granularity but inconsistent across patients)
   • Atlas parcels: Destrieux atlas parcellation (intermediate granularity, consistent across patients)
   • Lobes: Cerebral lobes + subcortical regions (coarsest but most stable)

2. Spatial Masked Reconstruction Pretraining:

   • Randomly selected spatial categories are masked (e.g., all electrode patches within a given brain region)
   • Online tokenizer \(\mathcal{F}\) and EMA target tokenizer \(\tilde{\mathcal{F}}\) (momentum warmup from 0 to 0.996)
   • \(\mathcal{L} = \frac{1}{|B_{target}|}\sum \|\tilde{B}_{ij} - \hat{B}_{ij}\|_2^2\)

3. Interleaved Spatiotemporal Sequences: Spatial and temporal tokens are interleaved so that attention jointly captures spatiotemporal dependencies.

Loss & Training

• Brain Treebank dataset: 10 epilepsy patients, 26 sessions, 2048 Hz
• Evaluation via pretraining followed by linear probing

Key Experimental Results

Main Results (Downstream Classification AUC %)

Encoding / Masking	Sentence Onset	Speech / Non-Speech
Channel	77.8%	76.4%
Parcel	86.2%	86.9%
Lobe	84.2%	84.1%
PopT baseline	79.5%	77.5%
Brant baseline	76.7%	69.1%

Ablation Study (ANOVA)

Factor	p-value	Effect Size
Encoding scale	p<1e-3	Large
Masking strategy	p=0.01–0.04	Medium
Interaction	Channel encoding + Channel masking yields best match	—

Key Findings

• Parcel encoding significantly outperforms channel encoding (+8.4% sentence onset, +10.5% speech)—anatomical priors are more informative than precise coordinates.
• The effect of encoding scale exceeds that of masking strategy—selecting the right encoding is more critical than selecting the right masking approach.
• Cross-subject generalization: held-out subjects achieve 84.1% AUC (vs. 86.9% with target subjects), confirming that parcel-level encoding facilitates cross-patient generalization.
• Performance scales positively with data volume: continuous improvement is observed as pretraining data increases from 5% to 75%.

Highlights & Insights

• Spatial encoding scale is a critical yet underexplored design choice: prior works default to channel-level encoding, while BaRISTA demonstrates that atlas parcel-level encoding is superior.
• "Intermediate granularity outperforms finest granularity": although channel-level encoding carries the most information, its cross-patient inconsistency leads to poor generalization, whereas atlas parcels balance precision and generalizability.

Limitations & Future Work

• Relies solely on anatomical parcellation; functional brain region encoding is not explored.
• Only spatial masking is evaluated; joint spatiotemporal masking is not tested.
• Experiments are conducted at a single sampling rate (2048 Hz).
• The dilated CNN temporal tokenizer may not be optimal.

Related Work & Insights

• vs. PopT: PopT employs channel-level encoding; BaRISTA demonstrates that parcel-level encoding is superior (+6.7%).
• vs. Brant: Brant uses region-level encoding but provides no systematic ablation; BaRISTA offers a comprehensive analysis.

Rating

• Novelty: ⭐⭐⭐⭐ — Systematic ablation of spatial encoding scales is conducted for the first time.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ — Three scales × multiple masking strategies × ANOVA + cross-subject + data scaling.
• Writing Quality: ⭐⭐⭐⭐ — Rigorous experimental design.
• Value: ⭐⭐⭐⭐ — Provides critical guidance for iEEG foundation model design.
• The hierarchical spatial structure of neural signals calls for multi-scale modeling—coarse and fine granularities are complementary.
• Larger spatial scales (lobe-level) improve decoding performance; pretraining substantially benefits low-data regimes.
• The core contribution lies in the simplicity and effectiveness of the design rationale.
• Experimental results thoroughly validate the central hypothesis.

Related Papers

• [NeurIPS 2025] Representation Consistency for Accurate and Coherent LLM Answer Aggregation
• [AAAI 2026] EEG-DLite: Dataset Distillation for Efficient Large EEG Model Training
• [NeurIPS 2025] S2M-Former: Spiking Symmetric Mixing Branchformer for Brain Auditory Attention Detection
• [NeurIPS 2025] Spiking Brain Compression: Post-Training Second-Order Compression for Spiking Neural Networks
• [NeurIPS 2025] VQToken: Neural Discrete Token Representation Learning for Extreme Token Reduction in Video Large Language Models