Parameter-Efficient Self-Supervised Adaptation for EEG-FM under Fixed Computational Budgets

Meghal Dani; Stefanie Liebe

cs.LGAug 25, 2026

Parameter-Efficient Self-Supervised Adaptation for EEG-FM under Fixed Computational Budgets

Authors: Meghal Dani, Stefanie Liebe

Organizations: University of Tübingen, Germany · Hertie Institute for AI in Brain Health (Hertie AI), Tübingen, Germany · Dept. of Neurology and Epileptology, University Clinic Tübingen, Germany · Hertie Institute for Clinical Brain Research, Tübingen, Germany

Abstract

EEG foundation models pretrained via self-supervised learning promise transferable representations, but their generalization remains limited, especially across diverse clinical datasets. Full fine-tuning is impractical for resource-constrained clinical settings due to high computational requirements. In this work, we investigate whether parameter-efficient self-supervised adaptation, updating only 9% of parameters suffices to align representations to target tasks. We evaluate our method on two state-of-the-art models with different pretraining objectives: BIOT (contrastive) and CBraMod (masked reconstruction), and evaluate on three clinical EEG datasets for abnormality detection (TUAB), event classification (TUEV), and seizure detection (CHB-MIT) under both in-distribution and out-of-distribution conditions. SSL adaptation yields consistent gains over linear probing, up to 20x AUCPR. Under a fixed compute budget, peak performance requires only 20--50% of available unlabeled data. Critically, when total window count is fixed, performance remains invariant to patient count, suggesting that performance is dependent on overall temporal window diversity only. Our findings demonstrate that parameter-efficient adaptation enables effective deployment of EEG Foundation models (EEG-FM) with minimal computational overhead and data collection burden. Code available at: https://github.com/c3n-group/efficient-eeg-adapt

Figures & tables

BIOT ( $\|\theta\|{=}3.19\text{M}$ , $\|\theta_{\mathrm{adapt}}\|{=}0.92\text{M}$ )
Dataset	Method	Bal. ACC	AUCPR / Kappa	AUCROC / F1
TUEV	LP	30.52 $\pm$ 1.21	33.86 $\pm$ 2.84	65.87 $\pm$ 1.64
	SA	31.78 $\pm$ 0.46 +1.26	40.45 $\pm$ 1.41 +6.59	69.43 $\pm$ 0.63 +3.56
TUAB	LP	63.93 $\pm$ 0.15	74.97 $\pm$ 0.13	75.08 $\pm$ 0.11
	SA	70.86 $\pm$ 0.23 +6.93	77.39 $\pm$ 0.24 +2.42	78.34 $\pm$ 0.26 +3.26
CHB-MIT	LP	50.74 $\pm$ 0.21	1.53 $\pm$ 0.03	46.93 $\pm$ 0.30

Table 1: Performance comparison between linear probing (LP) versus our parameter-efficient SSL adaptation (SA). Results shown as mean

\pm

SEM across 5 seeds. Blue subscripts indicate absolute improvement; best results in bold . Metrics: Balanced Accuracy / Cohen’s Kappa / Weighted F1 (TUEV); Balanced Accuracy / AUCPR / AUCROC (TUAB, CHB-MIT).

\theta_{\mathrm{adapt}}

: adapted parameters (of total

\theta

in EEG-FM).

Figure 1: SSL adaptation exhibits data efficiency under normalized compute

C{=}N{\times}\mathcal{E}

: performance peaks at

20-50

% of data then plateaus. (a) CHB-MIT seizure detection (AUCPR),

N_{\text{total}}{=}4{,}174

; (b) TUAB abnormality detection (AUCPR),

N_{\text{total}}{=}57{,}614

; (c) TUEV event classification (F1),

N_{\text{total}}{=}86{,}587

. (d) Kernel density estimation confirms saturation at

N/N_{\text{total}}{=}0.2

–

0.5

(repetition factor

r{=}\mathcal{E}{=}12

). Green: BIOT; Blue: CBraMod. Error bars denote SEM (3 data shuffle

{\times}

3 model seeds).

Figure 2: SSL adaptation performance is patient-agnostic under fixed computational budget

C

. At constant

N{=}P{\times}W

, varying patient count

P

yields flat performance curve: (a) CHB-MIT seizure detection (AUCPR),

P{\in}[3,8,11,16,22]

; (b) TUEV event classification (F1),

P{\in}[40,80,160,230]

; (c) TUAB abnormality detection (AUCPR),

P{\in}[900,1200,1500,1650,1800]

. Green: BIOT; Blue: CBraMod. Error bars denote SEM over 3 data shuffle

{\times}

3 model initialization seeds.

Model	Layers	Param.	AUCROC	AUCPR	Bal. Acc
BIOT	$L$	0.92M	88.67	35.62	60.62
	$L,L-1$	1.71M	88.44	35.05	60.82
	All	3.19M	88.60	36.79	62.18
CBRAMOD	$L$	0.44M	87.16	31.25	67.60
	$L,L-1$	0.84M	86.62	21.90	58.69
	All	4.4M	88.20	29.41	50.98

Table 2: Layer unfreezing ablation on CHBMIT dataset with one seed. Unfreezing only the final layer (

L

) provides optimal parameter efficiency.

EEG Foundation Models

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

See all 100 papers

Channel Adaptation for EEG Foundation Models: A Systematic Benchmark Across Architectures, Tasks, and Training Regimes
Apr 25, 2026Kuntal Kokate, Bruno Aristimunha, Dung Truong +1Fine-Tuning EEG Decoding
Swartz Center for Computational Neuroscience, Institute for Neural Computation, University of California San Diego, La Jolla, CA 92093, USA · Yneuro, Paris, France · Centre de Recherche Cerveau et Cognition, CNRS, Université Paul Sabatier, Toulouse, France
SCOPE: Structured Prototype-Guided Adaptation for EEG Foundation Models with Limited Labels
Feb 19, 2026Jingying Ma, Feng Wu, Yucheng Xing +5EEG Foundation Models Few-Shot Domain Adaptation
Saw Swee Hock School of Public Health, National University of Singapore, Singapore · Guangzhou Research Translation and Innovation Institute, National University of Singapore, Guangzhou, China · College of Computing and Data Science, Nanyang Technological University, Singapore +3
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

Apr 25, 2026J/K move · Enter open · S save

Channel Adaptation for EEG Foundation Models: A Systematic Benchmark Across Architectures, Tasks, and Training Regimes

Kuntal Kokate, Bruno Aristimunha, Dung Truong, Arnaud Delorme

Swartz Center for Computational Neuroscience, Institute for Neural Computation, University of California San Diego, La Jolla, CA 92093, USA · Yneuro, Paris, France · Centre de Recherche Cerveau et Cognition, CNRS, Université Paul Sabatier, Toulouse, France

Fine-Tuning EEG Decoding Negative Transfer EEG Foundation Models

Open paper AskPDF arXiv

Scaling EEG foundation models requires pooling data across heterogeneous electrode montages, a prerequisite both for larger pretraining corpora and for downstream deployment. We present the first systematic comparison of four channel adaptation methods (Conv1d projection, spherical spline interpolation (SSI), source-space decomposition, and Riemannian re-centering) across five pretrained EEG foundation models (5M--157M parameters), five downstream tasks, and two training regimes with 10--15 random seeds each. We find that rigid-montage models (BENDR, Neuro-GPT) require external adaptation, while flexible models (EEGPT, CBraMod) match or exceed it natively when fine-tuned but benefit from external methods under frozen-encoder deployment. A probe-SFT asymmetry exists: external adaptation can cause severe negative transfer during fine-tuning of flexible models. The optimal method is architecture-dependent (Conv1d for BENDR, SSI/Riemannian for Neuro-GPT, source-space decomposition for depression detection), and 5M-parameter CBraMod outperforms models up to 31

\times

larger on 4/5 datasets, consistent with independent findings that compact EEG-specific architectures can match larger models.

Abstract

EEG Foundation Models

Explore similar work

Channel Adaptation for EEG Foundation Models: A Systematic Benchmark Across Architectures, Tasks, and Training Regimes

SCOPE: Structured Prototype-Guided Adaptation for EEG Foundation Models with Limited Labels

Lightweight Semantic EEG Foundation Model for Frozen Cross-Disorder Transfer