BrainCodes: A Lightweight EEG Foundation Model over Discrete Neural Codes

Kleanthis Avramidis; Ivan Tashev; Dimitra Emmanouilidou

cs.LGOct 21, 2026

BrainCodes: A Lightweight EEG Foundation Model over Discrete Neural Codes

Authors: Kleanthis Avramidis, Ivan Tashev, Dimitra Emmanouilidou

Organizations: Signal Analysis and Interpretation Lab, University of Southern California, Los Angeles, CA 90007, USA · Microsoft Research, Redmond, WA, USA

Abstract

Recent electroencephalography (EEG) foundation models increasingly emphasize model scale, potentially limiting their deployment on resource-constrained devices and in everyday environments. We argue that transferable EEG features can instead be learned with a compact architecture and introduce BrainCodes, a lightweight foundation model pretrained through self-supervised learning on more than 800 hours of recordings from over 1,000 participants. Our architecture combines a quantization-based neural tokenizer with a spatiotemporal encoder that supports arbitrary electrode configurations. The pretraining objective explicitly encourages representations that capture spatial connectivity through channel positional encoding, and sensitivity to narrowband neural oscillations through an auxiliary training objective. Our model matches or surpasses the state of the art across diverse neurophysiological monitoring tasks, including sleep staging, attention assessment, and cognitive load estimation. BrainCodes transfers effectively to low-resource wearable devices while using up to 97% fewer trainable parameters than comparable architectures. Beyond predictive accuracy, our evaluation framework characterizes the geometry and invariances of learned EEG representations across measurable spatial, temporal, and spectral signal properties.

Figures & tables

Figure 1: The pretraining objectives of BrainCodes . A) Contiguous blocks of codes are masked and the model predicts the missing indices, evaluated through mean cross-entropy over the masked positions ( $\mathcal{L}_{\mathrm{tok}}$ ). B) Narrowband bursts are injected into regions of electrodes and the bottleneck is read by linear heads for its frequency, duration and location ( $\mathcal{L}_{\mathrm{osc}}$ ).

Dataset	Category	Channels	Length (s)	Classes	Samples	Participants
Main Bench
TUAB	Clinical	21	10	2	409,455	2,329
TUEV	Events	16	5	6	106,394	370
Siena	Seizures	29	10	2	50,749	14
PhysioNet-MI	Motor	64	4	4	9,838	109
BCI IV-2a	Motor	22	4	4	5,184	9

Table 1: EEG datasets used for downstream inference. ERP stands for event-related potentials.

Methods	Size	TUAB	TUEV	ISRUC	Siena
EEGNet	0.1M	0.764 $\pm$ 0.004	0.388 $\pm$ 0.014	0.715 $\pm$ 0.012	0.749 $\pm$ 0.052
EEGConformer	0.8M	0.776 $\pm$ 0.005	0.407 $\pm$ 0.016	0.740 $\pm$ 0.013	0.756 $\pm$ 0.021
SPaRCNet	0.8M	0.790 $\pm$ 0.002	0.416 $\pm$ 0.026	0.749 $\pm$ 0.008	0.657 $\pm$ 0.038
ContraWR	1.6M	0.775 $\pm$ 0.004	0.438 $\pm$ 0.035	0.740 $\pm$ 0.013	0.655 $\pm$ 0.031
CNN-Transformer	3.2M	0.778 $\pm$ 0.002	0.409 $\pm$ 0.016	0.736 $\pm$ 0.009	0.698 $\pm$ 0.056
ST-Transformer	2.6M	0.797 $\pm$ 0.002	0.398 $\pm$ 0.023	0.738 $\pm$ 0.021	0.753 $\pm$ 0.038

Table 2: Balanced accuracy on the main-bench datasets under full fine-tuning, mean

\pm

s.d. over seeds. Mean nBAC is the chance-normalized mean over all seven datasets, with uncertainty propagated from the per-dataset deviations. Best per column in bold , second best underlined .

	CBraMod	REVE-Base	BrainCodes
Dataset	Best config	Best config	$\mathbf{b}_{c}$	$\bar{\mathbf{b}}_{c}$
EEGMAT (2-class)	0.6050 $\pm$ 0.0200	0.7400 $\pm$ 0.0730	0.7430 $\pm$ 0.0175	0.7097 $\pm$ 0.0225
PhysioNet-MI (4-class)	0.5310 $\pm$ 0.0150	0.5370 $\pm$ 0.0050	0.4137 $\pm$ 0.0076	0.4953 $\pm$ 0.0055
BCIC-IV-2a (4-class)	0.3760 $\pm$ 0.0060	0.5170 $\pm$ 0.0150	0.5023 $\pm$ 0.0096	0.5628 $\pm$ 0.0025
TUEV (6-class)	0.4820 $\pm$ 0.0370	0.5920 $\pm$ 0.0080	0.5406 $\pm$ 0.0052	0.5875 $\pm$ 0.0155
Siena (2-class) ∗	0.7320 $\pm$ 0.0660	0.7400 $\pm$ 0.0070	0.8747 $\pm$ 0.0005	0.7830 $\pm$ 0.0340

Table 3: Linear probe performance on frozen encoders (BAC, mean

\pm

s.d. over 3 seeds).

\mathbf{b}_{c}

is the time-averaged bottleneck feature and

\bar{\mathbf{b}}_{c}

the concatenated one, both as defined in Section 2.3 .

Model	Size	EESM23	Muse-Motor	Muse-Mental	Mean nBAC
LaBraM-Base	5.8M	0.4109 $\pm$ 0.0087	0.4720 $\pm$ 0.0048	0.8245 $\pm$ 0.1446	0.4126 $\pm$ 0.0724
CBraMod	4.0M	0.3993 $\pm$ 0.0163	0.5083 $\pm$ 0.0267	0.8079 $\pm$ 0.1812	0.4178 $\pm$ 0.0918
CSBrain	8.9M	0.3663 $\pm$ 0.0261	0.4452 $\pm$ 0.0349	0.7081 $\pm$ 0.2143	0.3232 $\pm$ 0.1091
REVE-Base	69M	0.4176 $\pm$ 0.0224	0.4485 $\pm$ 0.0240	0.7165 $\pm$ 0.2177	0.3495 $\pm$ 0.1099
BrainCodes	1.9M	0.5254 $\pm$ 0.0251	0.5889 $\pm$ 0.0318	0.8163 $\pm$ 0.0957	0.5128 $\pm$ 0.0514

Table 4: Balanced accuracy on low-resource, consumer wearables. Muse results are means over 4-fold cross-validation; EESM23 over 3 seeds on the test set used in ( Pradeepkumar et al., 2025 ) .

Figure 2: Grounding analysis of BrainCodes. (a–c) Spatial organization: influence of one electrode on another against inter-electrode distance, the fitted decay length $\lambda$ relative to head width, and channel mixing against spatial locality. (d–f) Spectral selectivity under four-way injected rhythm identification, white-versus-pink noise control, and their difference as a preference index.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Methods	Size	Motor	ERP	Memory	Eyes	Mean nBAC
BIOT	3.2M	0.443 $\pm$ 0.079	0.500 $\pm$ 0.000	0.510 $\pm$ 0.018	0.763 $\pm$ 0.049	0.201 $\pm$ 0.037
CBraMod	4.0M	0.614 $\pm$ 0.104	0.777 $\pm$ 0.052	0.574 $\pm$ 0.038	0.839 $\pm$ 0.041	0.466 $\pm$ 0.052
EEGPT	25M	0.313 $\pm$ 0.035	0.668 $\pm$ 0.146	0.520 $\pm$ 0.017	0.797 $\pm$ 0.037	0.264 $\pm$ 0.077
NeuroGPT	79.5M	0.682 $\pm$ 0.083	0.757 $\pm$ 0.048	0.597 $\pm$ 0.029	0.827 $\pm$ 0.036	0.484 $\pm$ 0.043
BrainOmni †	33M	0.585 $\pm$ 0.058	0.723 $\pm$ 0.041	0.518 $\pm$ 0.023	0.852 $\pm$ 0.022	0.408 $\pm$ 0.032
LaBraM †	5.8M	0.630 $\pm$ 0.076	0.822 $\pm$ 0.040	0.526 $\pm$ 0.026	0.799 $\pm$ 0.047	0.450 $\pm$ 0.042

Appendix

Table 5: Model performance on the NeuroRVQ benchmark (balanced accuracy). External numbers are as published in Barmpas et al. (2025) . Sleep-EDF ( Kemp et al., 2000 ) was not included due to computational constraints in both studies. Best per column in bold , second best underlined .

Figure 3: Balanced accuracy as electrodes are removed by farthest-point sampling, with shading marking the standard deviation over seeds. Left: absolute performance, with dotted lines at chance. Right: fraction of each full-montage score, where each full montage sits at its own electrode count. ∗

Figure 4: Low-resource fine-tuning under independently varied participant and within-participant label budgets. (a,b) Mean balanced accuracy of BrainCodes across three seeds on PhysioNet-MI and OpenBMI-ERP. (c) Estimated gain in balanced-accuracy points per doubling of each budget axis. Error bars denote coefficient standard errors.

Ablated component	BCI IV-2a	EEGMAT	EESM23	Mean nBAC
MLP positional encoding	0.631 $\pm$ 0.007	0.698 $\pm$ 0.038	0.510 $\pm$ 0.010	0.439 $\pm$ 0.026
Random token masking	0.640 $\pm$ 0.032	0.753 $\pm$ 0.066	0.477 $\pm$ 0.038	0.466 $\pm$ 0.049
Codebook initialization	0.639 $\pm$ 0.020	0.677 $\pm$ 0.034	0.461 $\pm$ 0.026	0.409 $\pm$ 0.026
Auxiliary objective $\mathcal{L}_{\mathrm{osc}}$	0.607 $\pm$ 0.025	0.670 $\pm$ 0.032	0.444 $\pm$ 0.021	0.383 $\pm$ 0.025
Official BrainCodes	0.639 $\pm$ 0.002	0.772 $\pm$ 0.015	0.525 $\pm$ 0.025	0.498 $\pm$ 0.014

Appendix

Table 6: Ablation study on the pretraining configuration. We report BAC on three sampled datasets from the main benchmark, with the average computed across all datasets in three seeds.

Tokenizer	Backbone	BCI IV-2a	EEGMAT	EESM23	Mean nBAC
BioCodec	BrainCodes	0.639 $\pm$ 0.002	0.772 $\pm$ 0.015	0.525 $\pm$ 0.025	0.498 $\pm$ 0.014
BioCodec	from scratch	0.531 $\pm$ 0.015	0.561 $\pm$ 0.047	0.429 $\pm$ 0.019	0.270 $\pm$ 0.033
Random	from scratch	0.319 $\pm$ 0.010	0.501 $\pm$ 0.002	0.392 $\pm$ 0.030	0.121 $\pm$ 0.013

Appendix

Table 7: Contributions of tokenizer and backbone pretraining. Values are fully fine-tuned mean balanced accuracy

\pm

standard deviation over 3 seeds. The last column reports mean chance-normalized balanced accuracy. Random indicates an identical to BioCodec but randomly initialized RVQ.

Figure 5: Positional bias of the learned channel rotary encoding for a sample attention head from BrainCodes, seeded at each of eight electrodes, where the white marker gives the seed and the color gives the bias assigned to every other site. Top : frontal. Bottom : fronto-temporal and central.

Figure 6: Linear decoding of aperiodic and oscillatory spectral components from frozen representations on PhysioNet-MI, measured by within-channel $R^{2}$ . (a) Decoding of the $1/f$ exponent and mean oscillatory power across the five canonical frequency bands from BioCodec and BrainCodes. (b) Band-specific decoding, with other EEG foundation models shown as reference markers.

Figure 7: Phase assigned by the learned channel rotary encoding to a 10–20 montage, for each of the eight attention heads. Columns: the three rotation pairs of largest frequency magnitude, ordered by decreasing magnitude, and the mean over the remaining five pairs.

Methods	Size	Balanced Accuracy	AUC-PR	AUROC
EEGNet	0.1M	0.7642 $\pm$ 0.0036	0.8299 $\pm$ 0.0043	0.8412 $\pm$ 0.0031
EEGConformer	0.8M	0.7758 $\pm$ 0.0049	0.8427 $\pm$ 0.0054	0.8445 $\pm$ 0.0038
SPaRCNet	0.8M	0.7896 $\pm$ 0.0018	0.8414 $\pm$ 0.0018	0.8676 $\pm$ 0.0012
ContraWR	1.6M	0.7746 $\pm$ 0.0041	0.8421 $\pm$ 0.0104	0.8456 $\pm$ 0.0074
CNN-Transformer	3.2M	0.7777 $\pm$ 0.0022	0.8433 $\pm$ 0.0039	0.8461 $\pm$ 0.0013
ST-Transformer	2.6M	0.7966 $\pm$ 0.0023	0.8521 $\pm$ 0.0026	0.8707 $\pm$ 0.0019

Appendix

Table 8: Model performance on abnormality detection (TUAB, 2-class).

Methods	Size	Balanced Accuracy	Cohen’s $\kappa$	Weighted F1
EEGNet	0.1M	0.3876 $\pm$ 0.0143	0.3577 $\pm$ 0.0155	0.6539 $\pm$ 0.0120
EEGConformer	0.8M	0.4074 $\pm$ 0.0164	0.3967 $\pm$ 0.0195	0.6983 $\pm$ 0.0152
SPaRCNet	0.8M	0.4161 $\pm$ 0.0262	0.4233 $\pm$ 0.0181	0.7024 $\pm$ 0.0104
ContraWR	1.6M	0.4384 $\pm$ 0.0349	0.3912 $\pm$ 0.0237	0.6893 $\pm$ 0.0136
CNN-Transformer	3.2M	0.4087 $\pm$ 0.0161	0.3815 $\pm$ 0.0134	0.6854 $\pm$ 0.0293
ST-Transformer	2.6M	0.3984 $\pm$ 0.0228	0.3765 $\pm$ 0.0306	0.6823 $\pm$ 0.0190

Appendix

Table 9: Model performance on event type classification (TUEV, 6-class).

Methods	Size	Balanced Accuracy	Cohen’s Kappa	Weighted F1
EEGNet	0.1M	0.7154 $\pm$ 0.0121	0.7040 $\pm$ 0.0173	0.7513 $\pm$ 0.0124
EEGConformer	0.8M	0.7400 $\pm$ 0.0133	0.7143 $\pm$ 0.0162	0.7634 $\pm$ 0.0151
SPaRCNet	0.8M	0.7487 $\pm$ 0.0075	0.7097 $\pm$ 0.0132	0.7624 $\pm$ 0.0092
ContraWR	1.6M	0.7402 $\pm$ 0.0126	0.7178 $\pm$ 0.0156	0.7610 $\pm$ 0.0137
CNN-Transformer	3.2M	0.7363 $\pm$ 0.0087	0.7129 $\pm$ 0.0121	0.7719 $\pm$ 0.0105
ST-Transformer	2.6M	0.7381 $\pm$ 0.0205	0.7013 $\pm$ 0.0352	0.7681 $\pm$ 0.0175

Appendix

Table 10: Model performance on sleep staging (ISRUC, 5-class).

Methods	Size	Balanced Accuracy	AUC-PR	AUROC
EEGNet	0.1M	0.7487 $\pm$ 0.0521	0.3753 $\pm$ 0.0867	0.8687 $\pm$ 0.0527
EEGConformer	0.8M	0.7556 $\pm$ 0.0210	0.2091 $\pm$ 0.0786	0.8159 $\pm$ 0.0261
SPaRCNet	0.8M	0.6572 $\pm$ 0.0381	0.3164 $\pm$ 0.0659	0.7334 $\pm$ 0.0857
ContraWR	1.6M	0.6546 $\pm$ 0.0311	0.3711 $\pm$ 0.0405	0.7819 $\pm$ 0.0596
CNN-Transformer	3.2M	0.6982 $\pm$ 0.0560	0.3835 $\pm$ 0.0709	0.8719 $\pm$ 0.0403
ST-Transformer	2.6M	0.7527 $\pm$ 0.0381	0.3636 $\pm$ 0.0252	0.8884 $\pm$ 0.0091

Appendix

Table 11: Model performance on seizure detection (Siena, 2-class).

Methods	Size	Balanced Accuracy	Cohen’s $\kappa$	Weighted F1
EEGNet	0.1M	0.5814 $\pm$ 0.0125	0.4468 $\pm$ 0.0199	0.5796 $\pm$ 0.0115
EEGConformer	0.8M	0.6049 $\pm$ 0.0104	0.4736 $\pm$ 0.0171	0.6062 $\pm$ 0.0095
SPaRCNet	0.8M	0.5932 $\pm$ 0.0152	0.4564 $\pm$ 0.0234	0.5937 $\pm$ 0.0147
ContraWR	1.6M	0.5892 $\pm$ 0.0133	0.4527 $\pm$ 0.0248	0.5918 $\pm$ 0.0116
CNN-Transformer	3.2M	0.6053 $\pm$ 0.0118	0.4725 $\pm$ 0.0223	0.6041 $\pm$ 0.0105
ST-Transformer	2.6M	0.6035 $\pm$ 0.0081	0.4712 $\pm$ 0.0199	0.6053 $\pm$ 0.0075

Appendix

Table 12: Model performance on motor imagery (PhysioNet-MI, 4-class).

Methods	Size	Balanced Accuracy	Cohen’s $\kappa$	Weighted F1
EEGNet	0.1M	0.4482 $\pm$ 0.0094	0.2693 $\pm$ 0.0121	0.4226 $\pm$ 0.0108
EEGConformer	0.8M	0.4696 $\pm$ 0.0106	0.2924 $\pm$ 0.0141	0.4533 $\pm$ 0.0128
SPaRCNet	0.8M	0.4635 $\pm$ 0.0117	0.2847 $\pm$ 0.0147	0.4432 $\pm$ 0.0126
ContraWR	1.6M	0.4678 $\pm$ 0.0125	0.2905 $\pm$ 0.0160	0.4413 $\pm$ 0.0142
CNN-Transformer	3.2M	0.4600 $\pm$ 0.0108	0.2800 $\pm$ 0.0148	0.4460 $\pm$ 0.0114
ST-Transformer	2.6M	0.4575 $\pm$ 0.0145	0.2733 $\pm$ 0.0198	0.4471 $\pm$ 0.0142

Appendix

Table 13: Model performance on motor imagery (BCIC-IV-2a, 4-class).

Methods	Size	Balanced Accuracy	AUC-PR	AUROC
EEGNet	0.1M	0.6770 $\pm$ 0.0116	0.5763 $\pm$ 0.0102	0.7321 $\pm$ 0.0108
EEGConformer	0.8M	0.6805 $\pm$ 0.0123	0.5829 $\pm$ 0.0134	0.7424 $\pm$ 0.0128
SPaRCNet	0.8M	0.6879 $\pm$ 0.0107	0.5825 $\pm$ 0.0193	0.7418 $\pm$ 0.0132
ContraWR	1.6M	0.6631 $\pm$ 0.0097	0.5787 $\pm$ 0.0164	0.7332 $\pm$ 0.0082
CNN-Transformer	3.2M	0.6779 $\pm$ 0.0268	0.5777 $\pm$ 0.0285	0.7258 $\pm$ 0.0336
ST-Transformer	2.6M	0.6631 $\pm$ 0.0173	0.5672 $\pm$ 0.0259	0.7132 $\pm$ 0.0174

Appendix

Table 14: Model performance on mental stress detection (EEGMAT, 2-class).

Models	Size	Balanced Accuracy	Cohen’s $\kappa$	Weighted F1
BIOT	3.2M	0.3858 $\pm$ 0.0085	0.3406 $\pm$ 0.0096	0.4888 $\pm$ 0.0124
BIOT-TFM	3.2M	0.3952 $\pm$ 0.0170	0.3603 $\pm$ 0.0252	0.5033 $\pm$ 0.0165
LaBraM-Base	5.8M	0.3890 $\pm$ 0.0182	0.3322 $\pm$ 0.0232	0.4827 $\pm$ 0.0157
LaBraM-TFM	5.8M	0.4004 $\pm$ 0.0086	0.3475 $\pm$ 0.0128	0.4864 $\pm$ 0.0118
TFM-Tokenizer	1.9M	0.4148 $\pm$ 0.0209	0.3883 $\pm$ 0.0233	0.5174 $\pm$ 0.0141
Reproduced:

Appendix

Table 15: Model performance on low-resource sleep staging (EESM23, 6-class).

Dataset	Balanced Accuracy	Cohen’s $\kappa$	Weighted F1
Muse-Motor (3-class)	0.5885 $\pm$ 0.0318	0.3650 $\pm$ 0.0516	0.5800 $\pm$ 0.0298
Muse-Mental (3-class)	0.8160 $\pm$ 0.0957	0.7097 $\pm$ 0.1287	0.8125 $\pm$ 0.0793

Appendix

Table 16: Detailed results (over 4 cross-validation folds) of BrainCodes on the Muse tasks.

Dataset	Head	Peak LR	Epochs (warmup)	Batch
TUAB	$\bar{b}_{c}$	5×10 −4	10 (2)	128
TUEV	$\bar{b}_{c}$	5×10 −4	10 (2)	32
Siena	$b_{c}$	5×10 −4	20 (2)	32
ISRUC	$b_{c}$	5×10 −4	20 (2)	64
PhysioNet-MI	$b_{c}$	5×10 −4	20 (2)	64
BCI IV-2a	$\bar{b}_{c}$	2×10 −4	20 (2)	32

Appendix

Table 17: Fine-tuning settings of BrainCodes for the results reported in the main benchmarks.

EEG Foundation Models

Electroencephalography (EEG) foundation models learn reusable representations of brain signals for transfer across people, tasks, and recording setups.

See all 101 papers

A Multi-dimensional Framework for Evaluating Generalization in EEG Foundation Models
May 27, 2026Aditya Kommineni, Emily Zhou, Kleanthis Avramidis +2Neural Network Generalization EEG Decoding
Signal Analysis and Interpretation Laboratory, University of Southern California, USA
BandVQ: Band-Wise Vector-Quantized EEG Foundation Model
May 24, 2026Jamiyan Sukhbaatar, Satoshi Imamura, Toshihisa TanakaCross-Subject EEG Decoding Masked Autoencoders
National University of Mongolia, Ulaanbaatar, Mongolia · Tokyo University of Agriculture and Technology, Tokyo, Japan
Lightweight Semantic EEG Foundation Model for Frozen Cross-Disorder Transfer
Oct 4, 2026Rita Huan-Ting Peng, Nhat BuiEEG Seizure Detection Self-Supervised Pre-Training
University of Illinois Urbana–Champaign, Urbana, IL, USA · Carle Foundation Hospital, Urbana, IL, USA

May 27, 2026J/K move · Enter open · S save

A Multi-dimensional Framework for Evaluating Generalization in EEG Foundation Models

Aditya Kommineni, Emily Zhou, Kleanthis Avramidis, Tiantian Feng+1

Signal Analysis and Interpretation Laboratory, University of Southern California, USA

Neural Network Generalization EEG Decoding EEG Foundation Models

Open paper AskPDF arXiv

Evaluating foundation models under appropriate adaptation settings is essential for understanding the quality and transferability of the learned representations. Recent EEG foundation models have demonstrated promising transfer capabilities across tasks and datasets, motivating their growing use in neurotechnology and clinical applications. However, these models are typically evaluated under full fine-tuning on well-curated downstream datasets, a setting that does not reflect biomedical domain constraints such as limited labeled data, reduced sensor coverage, or parameter-efficient adaptation. In this work, we propose a multi-dimensional evaluation framework for assessing EEG models under realistic low-resource conditions. Empirical analysis of both supervised EEG models and recent EEG foundation models, including LaBraM, CSBrain, and CBraMod, across 6 different datasets is performed under the proposed multi-dimensional evaluation framework. We find that EEG foundation models consistently provide performance gains on long-context tasks such as sleep stage prediction and mental health state classification. In contrast, for short-window Brain Computer Interface style tasks, supervised models achieve comparable despite having substantially fewer parameters. Additional analyses demonstrate that current foundation models provide limited robustness to short-window tasks and channel constrained settings. Together, these findings motivate the use of multi-dimensional evaluation protocols that characterize model behavior under realistic use constraints.

Abstract

EEG Foundation Models

Explore similar work

A Multi-dimensional Framework for Evaluating Generalization in EEG Foundation Models

BandVQ: Band-Wise Vector-Quantized EEG Foundation Model

Lightweight Semantic EEG Foundation Model for Frozen Cross-Disorder Transfer