Research map / Label-efficient and robust learning / Self-supervised and contrastive
Is Limited Participant Diversity Impeding EEG-based Machine Learning?
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.
Studies how EEG ML performance scales with number of participants versus segments per participant, and whether augmentations and LaBraM self-supervised pre-training help under limited participant diversity.
Problem
EEG ML models are trained on many short segments drawn from few participants, so generalisation to unseen participants may be limited by participant diversity rather than total sample size. It is unclear how much participant count versus data per participant matters, and whether common remedies (aug
Approach
Frames EEG data as a two-level (participant, segment) generation process, grounded in statistical learning theory, and disentangles participant count from segments per participant through controlled subsampling grids on large datasets (TUAB, CAUEEG, PhysioNet) with three architectures.
Key results
- Performance rose strongly from 25 to several hundred participants across datasets and models, then diminished; on TUAB, doubling segments per participant gave no substantial gain at fixed participant count.
- On PhysioNet sleep staging, 25 participants with maximum segments matched performance of 400 participants with 5 segments each.
- Augmentations (AmplitudeScaling, FrequencyShift, PhaseRandomisation) gave small, inconsistent effects; PhaseRandomisation hurt TCN and mAtt on PhysioNet while helping LaBraM.
- LaBraM pre-training improved accuracy across most data regimes, with the largest gains when fine-tuning data per participant was small.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase the training participant count from 25 to several hundred | TCN, mAtt, and LaBraM · TUAB, CAUEEG, and PhysioNet | accuracy | Helped |
| Increase segments per participant while keeping participant count fixed | TCN · TUAB and CAUEEG | accuracy | No clear effect |
| Increase segments per participant while keeping participant count fixed | TCN · PhysioNet | accuracy | Helped |
| Add AmplitudeScaling augmentation | TCN, mAtt, and LaBraM · TUAB, CAUEEG, and PhysioNet | accuracy | Mixed |
| Add FrequencyShift augmentation | TCN, mAtt, and LaBraM · TUAB, CAUEEG, and PhysioNet | accuracy | Mixed |
| Add PhaseRandomisation augmentation | TCN and mAtt · PhysioNet | accuracy | Hurt |
| Add PhaseRandomisation augmentation | LaBraM · PhysioNet | accuracy | Helped |
| Increase augmentation benefit by increasing sample size | TCN, mAtt, and LaBraM · TUAB, CAUEEG, and PhysioNet | accuracy | No clear effect |
| Add self-supervised pre-training before fine-tuning | LaBraM Base · TUAB, CAUEEG, and PhysioNet | accuracy | Helped |
| Use LaBraM-compatible preprocessing, including 15-second rather than 30-second epochs (ablation) | TCN · PhysioNet | accuracy | Hurt |
| Subsample consecutive segments instead of uniformly at random (ablation) | TCN · TUAB | accuracy | No clear effect |
Surprising findings
- Increasing segments per participant did little to compensate for a small participant count on TUAB and CAUEEG.
- Standard EEG augmentations did not consistently improve performance and PhaseRandomisation degraded smaller models on PhysioNet.
Limitations
- Focused on generalisation to unseen participants; results may not transfer to participant-specific BCI settings.
- Only three tasks/datasets, and only existing augmentations and a single smallest LaBraM variant were tested.
- Uniform segment subsampling may overstate segment diversity; contiguous control was only run for one setting.
Experimental setup
Datasets: TUAB CAUEEG PhysioNet/CinC Challenge 2018
Benchmarks: TUAB CAUEEG PhysioNet/CinC 2018
Metrics: accuracy (participant-level majority vote)
Baselines: TCN mAtt LaBraM trained from scratch
Builds on: LaBraM TCN mAtt FrequencyShift AmplitudeScaling
Closest papers by meaning
- What do EEG Foundation Models Capture from Human Brain Signals? (NeurIPS 2026 · similarity 0.85)
- NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces (NeurIPS 2026 · similarity 0.85)
- Standing on the Shoulders of Giants: Rethinking EEG Foundation Model Pretraining via Multi-Teacher Distillation (NeurIPS 2026 · similarity 0.84)
- EEG-Language Pretraining for Highly Label-Efficient Clinical Phenotyping (ICML 2025 · similarity 0.84)
- EEG Benchmarking Needs a Task Specification Layer: NeuroDoc for Rulebook-Guided, Executable Benchmark Construction (NeurIPS 2026 · similarity 0.84)
- Large Brain Model for Learning Generic Representations with Tremendous EEG Data in BCI (ICLR 2024 · similarity 0.83)
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.