SpecBraM: What Should an EEG Foundation Model Predict? Masked Band-Power Prediction versus Waveform Reconstruction

Peng Xie; Yequan Bie; Jianda Mao; Kani Chen

cs.LGOct 5, 2026

SpecBraM: What Should an EEG Foundation Model Predict? Masked Band-Power Prediction versus Waveform Reconstruction

Authors: Peng Xie, Yequan Bie, Jianda Mao, Kani Chen

Organizations: The Hong Kong University of Science and Technology

Abstract

Self-supervised EEG models often reconstruct masked waveforms or predict discrete codes. We study a task-aligned alternative: masked band-power prediction (MBP), which predicts fixed narrow-band log spectral energy for masked channel-time patches. This target retains rhythm power relevant to sleep staging while avoiding phase-sensitive waveform reconstruction and a learned codebook. Across three pretraining seeds, we compare band-power and waveform targets with matched backbones, pretraining data (2,388 hours), and training steps, including a 2x2 tokenizer-by-target design. On ISRUC and HMC sleep staging, MBP exceeds raw- and band-waveform reconstruction by 1.6-2.8 balanced-accuracy points with all labels and 4.7-7.3 points with 1% of labels under a strict linear probe; the target effect exceeds the tokenizer effect. Its frozen features reach 0.7916/0.7425 balanced accuracy, versus 0.7636/0.7227 for a matched rich handcrafted spectral baseline, although the gap is about one point with 1% of labels. Full fine-tuning reaches 0.8107/0.7669. The gains do not extend to every task with spectral cues, including motor imagery, depression screening, and vigilance regression. These results support choosing pretraining targets to match the physical quantities and spatial and temporal scales relevant to downstream labels.

Figures & tables

Figure 1: What should an EEG pretext task predict? LaBraM predicts discrete codes from a separately trained vector-quantized tokenizer; CBraMod, CSBrain and REVE reconstruct the raw waveform of masked patches. SpecBraM keeps the masked-patch setting but changes the target : it regresses the log spectral energy of each masked patch in a few fixed frequency bands. The bottom row shows the filterbank-tokenizer variant ; the main model uses a plain linear patch embedding. In our experiments the effect of the target is larger than that of the tokenizer (Sec. 6 ).

	BA (100% labels)		BA (1% labels)		$\Delta$ BA (pts)
Pretext target	ISRUC	HMC	ISRUC	HMC	ISRUC	HMC
Band power (fixed narrow-band filters)	0.7863 $\pm$ 0.0039	0.7394 $\pm$ 0.0025	0.758 $\pm$ 0.002	0.674 $\pm$ 0.009	–	–
band power, leak-free (per-patch filtering)	0.7906 $\pm$ 0.0018	0.7398 $\pm$ 0.0044	0.753 $\pm$ 0.004	0.675 $\pm$ 0.012	+0.4	+0.0
band power, standard bands (per-patch FFT)	0.7827 $\pm$ 0.0021	0.7339 $\pm$ 0.0021	0.725 $\pm$ 0.009	0.638 $\pm$ 0.004	-0.4	-0.6
Raw-waveform reconstruction	0.7677 $\pm$ 0.0028	0.7180 $\pm$ 0.0039	0.711 $\pm$ 0.026	0.623 $\pm$ 0.018	-1.9	-2.1
Band-waveform reconstruction	0.7682 $\pm$ 0.0061	0.7119 $\pm$ 0.0024	0.697 $\pm$ 0.019	0.613 $\pm$ 0.018	-1.8	-2.7

Table 1: Pretext-target study at matched backbone, data (

2{,}388

h), steps (16 epochs) and launcher; only the target (or, last row, the mask axis) changes. Filterbank tokenizer; strict linear probe, mean

{\pm}

std over three pretraining seeds. “1%”: balanced accuracy with

1\%

of the labelled training epochs (mean of five subsets).

\Delta

: difference to the band-power row in balanced-accuracy points.

Tokenizer	BP	Raw	$\Delta$	wins	BP	Raw	$\Delta$	wins
	ISRUC				HMC
Linear (CBraMod-style patch embedding)	0.7916 $\pm$ 0.0081	0.7755 $\pm$ 0.0034	+1.6 $\pm$ 0.7	3/3	0.7425 $\pm$ 0.0045	0.7141 $\pm$ 0.0062	+2.8 $\pm$ 1.1	3/3
Filterbank	0.7863 $\pm$ 0.0039	0.7677 $\pm$ 0.0028	+1.9 $\pm$ 0.3	3/3	0.7394 $\pm$ 0.0025	0.7180 $\pm$ 0.0039	+2.1 $\pm$ 0.1	3/3
Linear, 1% labels	0.760 $\pm$ 0.002	0.708 $\pm$ 0.010	+5.2 $\pm$ 1.2	3/3	0.685 $\pm$ 0.001	0.613 $\pm$ 0.015	+7.3 $\pm$ 1.6	3/3
Filterbank, 1% labels	0.758 $\pm$ 0.002	0.711 $\pm$ 0.026	+4.7 $\pm$ 2.5	3/3	0.674 $\pm$ 0.009	0.623 $\pm$ 0.018	+5.1 $\pm$ 0.9	3/3

Table 2: Tokenizer

\times

target. Band power (BP) vs. raw-waveform reconstruction (Raw) under both tokenizers; everything else matched. Strict linear probe, mean

{\pm}

std over three pretraining seeds.

\Delta

: paired (same-seed) difference in balanced-accuracy points; “wins”: seeds on which BP

>

Raw.

Figure 2: Target study (Table 1 ): bars are means over three pretraining seeds, dots individual seeds. Blue: band-power targets; orange: waveform reconstruction; grey: band power with frequency-axis masking.

	ISRUC		HMC
Epoch	BP	Raw	BP	Raw
4	0.7719	0.7394	0.7418	0.6714
8	0.7862	0.7630	0.7339	0.7026
16	0.7833	0.7680	0.7398	0.7190

Table 3: Along pretraining (filterbank tokenizer, seed 42, strict linear probe BA). Band power at epoch 4 already exceeds raw-waveform reconstruction at epoch 16.

		ISRUC			HMC
Model	Pretrain (h)	BA	$\kappa$	F1	BA	$\kappa$	F1
Random init (our arch.)	0	0.6884	0.6111	0.6957	0.6256	0.5274	0.6332
BIOT (six-dataset)	mixed	0.7358	0.6658	0.7403	0.6578	0.5623	0.6610
BIOT (SHHS+PREST)	mixed	0.7430	0.6683	0.7411	0.6105	0.5042	0.6175
CBraMod	${\sim}9$ k	0.7278	0.6581	0.7343	0.6190	0.5131	0.6193
LaBraM-Base	${\sim}2.5$ k	0.7523	0.6857	0.7543	0.6323	0.5303	0.6357

Table 4: Strict linear probe on sleep staging. Frozen encoder, mean-pooled features, logistic regression with

C

selected on the validation split, test scored once (deterministic); a 30-s epoch is the mean of its three 10-s window features. Every model we ran sees the identical epochs and splits, each in its own input convention (Sec. 5 ). SpecBraM rows: mean

{\pm}

std over three pretraining seeds. † Quoted from the REVE paper (its own linear-probing protocol). Handcrafted rows: features averaged over the three 10-s windows of each epoch (as for every encoder; better than whole-epoch features), same samples and probe; ∗ 1-Hz log PSD (0.5–45 Hz), relative power in six bands, Hjorth parameters, spectral entropy, 95% spectral edge, line length and kurtosis per channel.

ISRUC
Model	1%	5%	10%	25%	50%	100%
Random init (our arch.)	0.542	0.615	0.644	0.672	0.684	0.689
BIOT	0.655	0.703	0.716	0.728	0.734	0.736
CBraMod	0.660	0.709	0.717	0.726	0.727	0.729
LaBraM-Base	0.680	0.724	0.730	0.741	0.746	0.752
Handcrafted, 5-band	0.707	0.726	0.731	0.739	0.741	0.742

Table 5: Label efficiency (strict linear probe, balanced accuracy vs. fraction of labelled training epochs;

{<}100\%

: mean of five fixed random subsets shared by all models). Our rows: mean over three pretraining seeds.

Figure 3: Label efficiency under the strict linear probe. Our curves: mean over three pretraining seeds with

{\pm}1

std band.

Task-specific models
		ISRUC			HMC
Method	Pretrain (h)	BA	$\kappa$	Weighted F1	BA	$\kappa$	Weighted F1
EEGNet ( Lawhern et al. 2018 )	–	0.7154 $\pm$ .0121	0.7040 $\pm$ .0173	0.7513 $\pm$ .0124	0.6534 $\pm$ .0122	0.5886 $\pm$ .0201	0.6536 $\pm$ .0168
EEGConformer ( Song et al. 2023 )	–	0.7400 $\pm$ .0133	0.7143 $\pm$ .0162	0.7634 $\pm$ .0151	0.7149 $\pm$ .0086	0.6432 $\pm$ .0055	0.7080 $\pm$ .0039
SPaRCNet ( Jing and others 2023 )	–	0.7487 $\pm$ .0075	0.7097 $\pm$ .0132	0.7624 $\pm$ .0092	0.4756 $\pm$ .1109	0.3147 $\pm$ .1315	0.4108 $\pm$ .1310
ContraWR ( Yang et al. 2023b )	–	0.7402 $\pm$ .0126	0.7178 $\pm$ .0156	0.7610 $\pm$ .0137	0.4242 $\pm$ .0541	0.2340 $\pm$ .0554	0.2987 $\pm$ .0288

Table 6: Sleep staging, fine-tuned (5-class, 30-s epochs in sequences of 20, one shared 1-layer sequence transformer; identical subject-disjoint splits as Zhou et al. (2025) ). Baselines as reported by Zhou et al. (2025) (REVE: its paper, ISRUC only). Our rows: mean

{\pm}

std over three pretraining seeds (one fine-tuning run each), so the spread includes pretraining variance, unlike the baselines’ fine-tuning-seed spread.

	MA (stress, 2-class)			Mumtaz (MDD, 2-class)
Model	BA	AUC-PR	AUROC	BA	AUC-PR	AUROC
SpecBraM	0.752 $\pm$ 0.014	0.665 $\pm$ 0.034	0.840 $\pm$ 0.020	0.935 $\pm$ 0.003	0.991 $\pm$ 0.005	0.990 $\pm$ 0.005
Raw-waveform recon. (same arch.)	0.666 $\pm$ 0.022	0.606 $\pm$ 0.026	0.773 $\pm$ 0.024	0.935 $\pm$ 0.020	0.983 $\pm$ 0.009	0.983 $\pm$ 0.008
SpecBraM, filterbank tokenizer	0.693 $\pm$ 0.047	0.655 $\pm$ 0.064	0.816 $\pm$ 0.045	0.941 $\pm$ 0.023	0.990 $\pm$ 0.008	0.990 $\pm$ 0.008
Raw recon., filterbank tokenizer	0.730 $\pm$ 0.020	0.699 $\pm$ 0.041	0.810 $\pm$ 0.026	0.922 $\pm$ 0.012	0.978 $\pm$ 0.007	0.979 $\pm$ 0.007
CBraMod (official)	0.653	0.515	0.708	0.893	0.973	0.973

Table 7: Strict linear probe beyond sleep: binary tasks. Samples, channels and subject-disjoint splits are produced by CBraMod’s own preprocessing scripts, so every model is probed on identical samples. Mental Arithmetic (MA) has only 4 test subjects. † CBraMod’s fine-tuned result from its paper, for reference.

	FACED (emotion, 9-class)			PhysioNet-MI (4-class)
Model	BA	$\kappa$	F1	BA	$\kappa$	F1
SpecBraM	0.241 $\pm$ 0.011	0.146 $\pm$ 0.012	0.240 $\pm$ 0.011	0.388 $\pm$ 0.013	0.184 $\pm$ 0.017	0.386 $\pm$ 0.013
Raw-waveform recon. (same arch.)	0.216 $\pm$ 0.005	0.118 $\pm$ 0.007	0.213 $\pm$ 0.007	0.397 $\pm$ 0.008	0.196 $\pm$ 0.011	0.397 $\pm$ 0.009
SpecBraM, filterbank tokenizer	0.226 $\pm$ 0.002	0.130 $\pm$ 0.002	0.226 $\pm$ 0.002	0.383 $\pm$ 0.015	0.178 $\pm$ 0.021	0.382 $\pm$ 0.016
Raw recon., filterbank tokenizer	0.204 $\pm$ 0.016	0.105 $\pm$ 0.017	0.202 $\pm$ 0.014	0.380 $\pm$ 0.016	0.173 $\pm$ 0.021	0.380 $\pm$ 0.016
CBraMod (official)	0.171	0.069	0.167	0.312	0.083	0.310

Table 8: Strict linear probe beyond sleep: multi-class tasks (protocol as Table 7 ). FACED: 9-class emotion (chance

0.111

); PhysioNet-MI: 4-class motor imagery (chance

0.25

). ‡ CBraMod with a frozen backbone and a trained classifier, from its paper (Table 18). † CBraMod fine-tuned, from its paper.

	LP, 0.5–45 Hz		LP, 4–45 Hz		FT, 0.5–45 Hz
Model	Pearson	$R^{2}$	Pearson	$R^{2}$	Pearson	$R^{2}$
SpecBraM	0.565 $\pm$ 0.012	0.166 $\pm$ 0.035	0.619 $\pm$ 0.006	0.250 $\pm$ 0.012	0.587 $\pm$ 0.033	0.260 $\pm$ 0.083
Raw-waveform recon. (same arch.)	0.548 $\pm$ 0.031	0.209 $\pm$ 0.038	0.572 $\pm$ 0.022	0.261 $\pm$ 0.051	0.626 $\pm$ 0.039	0.357 $\pm$ 0.049
SpecBraM, filterbank tokenizer	0.540 $\pm$ 0.021	0.176 $\pm$ 0.035	0.611 $\pm$ 0.007	0.257 $\pm$ 0.021	–	–
Raw recon., filterbank tokenizer	0.564 $\pm$ 0.005	0.221 $\pm$ 0.029	0.549 $\pm$ 0.011	0.213 $\pm$ 0.017	–	–
LaBraM-Base (official)	0.532	0.162	–	–	–	–

Table 9: SEED-VIG vigilance regression (PERCLOS) on the split stated in the CBraMod paper (experiments 1–15/16–19/20–23 in subject order). Linear probe = frozen encoder + ridge regression (regularization on validation

R^{2}

); FT = full fine-tuning (MSE, epoch selected on validation

R^{2}

). Input band 0.5–45 Hz (standard) or 4–45 Hz (high-passed; the cutoff the validation split selects for both linear-tokenizer targets). Our rows: mean

{\pm}

std over three pretraining seeds. † From the CBraMod paper.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Dim.	Nominal	Peak	$-3$ dB band	In nominal band
“ $\delta$ ”	2 Hz	4.6 Hz	2.9–6.3 Hz	31% (63% in $\theta$ )
“ $\theta$ ”	6	6.0	3.7–8.5	68%
“ $\alpha$ ”	10	10.0	7.2–12.8	71%
“ $\beta$ ”	21	21.0	18.1–23.9	100% (mid-band only)
“ $\gamma$ ”	38	38.0	35.1–40.9	100% (mid-band only)

Appendix

Table 10: Measured response of the five fixed target filters (65 taps at 256 Hz). Last column: share of each filter’s energy within 0–45 Hz falling into its nominal band.

Band	Range (Hz)	$m_{b}$	$\mathrm{sd}[\log P_{b}]=\sqrt{\psi^{\prime}(m_{b})}$
$\delta$	0.5–4	3	0.63
$\theta$	4–8	3	0.63
$\alpha$	8–13	4	0.53
$\beta$	13–30	13	0.28
$\gamma$	30–45	12	0.29

Appendix

Table 11: Ideal rectangular bands at our patch length (

N{=}200

f_{s}{=}256

Hz,

\Delta f{=}1.28

Hz): number of Fourier bins and the noise of

\log P_{b}

(Prop. 3 ). Applies to the standard-band FFT variant, not to the filterbank target.

Explore similar work

EEG-JEPA: Structured Latent Prediction for EEG Foundation Models
Jul 31, 2026Jinhao Li, Zhiyuan Ma, Xueqiao Han +8Electroencephalography Foundation Models Electroencephalography
Tsinghua Laboratory of Brain and Intelligence, Tsinghua University · School of Basic Medical Sciences, Tsinghua Medicine, Tsinghua University · School of Biomedical Engineering, Tsinghua Medicine, Tsinghua University +2
MANAS-2: Constrained Reconstruction for EEG Foundation Models
Sep 12, 2026Arvasu Kulkarni, Aditya Ray Mishra, Jeet Bandhu Lahiri +5Electroencephalography Foundation Models 3D Masked Autoencoders
Mannas AI · Indian Institute of Technology, Mandi
BandVQ: Band-Wise Vector-Quantized EEG Foundation Model
May 24, 2026Jamiyan Sukhbaatar, Satoshi Imamura, Toshihisa TanakaElectroencephalography Foundation Models Electroencephalography
National University of Mongolia, Ulaanbaatar, Mongolia · Tokyo University of Agriculture and Technology, Tokyo, Japan

Jul 31, 2026, eess.SPJ/K move · Enter open · S save

EEG-JEPA: Structured Latent Prediction for EEG Foundation Models

Jinhao Li, Zhiyuan Ma, Xueqiao Han, Zhongye Xia+7

Tsinghua Laboratory of Brain and Intelligence, Tsinghua University · School of Basic Medical Sciences, Tsinghua Medicine, Tsinghua University · School of Biomedical Engineering, Tsinghua Medicine, Tsinghua University+2

Electroencephalography Foundation Models Electroencephalography Latent Prediction Electrocardiogram Foundation Models Pretraining

Open paper AskPDF arXiv

Electroencephalography (EEG) foundation models aim to learn reusable representations from large-scale unlabeled recordings. A common pretraining strategy is masked waveform reconstruction, but applying supervision directly to noisy EEG may encourage models to recover predictable background activity, acquisition effects, and artifacts rather than neural structure that transfers across tasks. This raises a central question: what should an EEG foundation model predict to learn transferable representations? We introduce EEG-JEPA a structured latent-prediction framework for EEG foundation modeling. Rather than reconstructing masked voltage samples, a masked context encoder and predictor infer contextual latent states produced by an exponential-moving-average target encoder that observes the complete input. EEG-JEPA organizes target design along three complementary dimensions: target content specifies what representation is predicted, target support specifies where prediction occurs over structured electrode--time regions through Neurotopology-Aware Multi-scale Electrode-Temporal Masking (N-MET), and target depth specifies at which encoder layers supervision is applied. Together, these designs shift EEG pretraining from recovering missing measurements to inferring latent states from structured electrode--time context. We evaluate EEG-JEPA through controlled objective comparisons, frozen multitask transfer, and full fine-tuning. Under the same backbone, pretraining corpus, and training duration, EEG-JEPA improves the 14-task frozen macro balanced accuracy from 40.49% to 50.42% over CBraMod-style masked waveform reconstruction. Multi-source continuation further raises this result to 52.94%, the highest average among the EEG foundation models evaluated on EEG-FM-Bench. Under protocol-matched full fine-tuning, EEG-JEPA also improves the nine-task average balanced accuracy from 68.98% to 70.65%.

Abstract

Explore similar work

EEG-JEPA: Structured Latent Prediction for EEG Foundation Models

MANAS-2: Constrained Reconstruction for EEG Foundation Models

BandVQ: Band-Wise Vector-Quantized EEG Foundation Model