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

Jingying Ma; Feng Wu; Yucheng Xing; Qika Lin; Tianyu Liu; Chenyu Liu; Ziyu Jia; Mengling Feng

cs.LGFeb 19, 2026

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

Authors: Jingying Ma, Feng Wu, Yucheng Xing, Qika Lin, Tianyu Liu, Chenyu Liu, Ziyu Jia, Mengling Feng

Organizations: 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 · Beijing Key Laboratory of Brainnetome and Brain-Computer Interface, Institute of Automation, Chinese Academy of Sciences, Beijing, China ·

Abstract

Electroencephalography (EEG) foundation models (EFMs) have shown strong potential for transferable representation learning, yet their adaptation in realistic settings remains challenging when only a few labeled subjects are available. We show that this challenge stems from a structural mismatch between noisy, limited supervision and the highly plastic parameter space of EFMs, reflected in three key failure modes: overconfident miscalibration, prediction collapse, and representation drift caused by unconstrained parameter updates. To address these challenges, we propose SCOPE, a Structured COnfidence-aware Prototype-guided framework for label-limited EFM adaptation. SCOPE first constructs cohort-level external supervision to provide persistent guidance and further derives confidence-aware pseudo-labels to select reliable unlabeled samples for adaptation. Building on the constructed external supervision, SCOPE introduces ProAdapter, a lightweight prototype-conditioned adapter that modulates frozen EFMs to preserve pretrained representations. Experiments across 50 label-limited adaptation settings, covering 6 EEG tasks, 5 EFM backbones, and 5%-50% training labeled-subject ratios, show that SCOPE consistently achieves strong performance and efficiency.

Figures & tables

Figure 1: Empirical motivation for label-limited EFM adaptation: adapting CodeBrain on ISRUC with 30% labeled subjects reveals (A) large generalization gap; (B) overconfident miscalibration; (C) training collapse across seeds; and (D) layer-wise representation drift. (E) These observations motivate our shift from self-training, sample-level pseudo-labels, and full fine-tuning toward external cohort-level structured supervision, confidence-aware pseudo-labeling, and conditioned adaptation.

Figure 2: Overview of the two-stage SCOPE framework. Left : External structured supervision construction progressively induces class-level prototypes and confidence-aware pseudo-labels for unlabeled data. Right : A frozen EEG foundation model is adapted via lightweight prototype-conditioned adapters in the last layers, using confidence-weighted supervision for controlled adaptation.

Backbone	Method	ISRUC		SEED		Mental		PhysioNet-MI		SHU-MI		CHB-MIT
Backbone	Method	Kappa	WF1	AUPRC	AUC	AUPRC	AUC	Kappa	WF1	AUPRC	AUC	AUPRC	AUC
Non-Pretrained	EEGNet	45.40	52.02	53.39	54.15	40.35	63.75	34.20	50.82	51.60	53.02	16.21	69.26
Non-Pretrained	EEGConformer	47.70	53.93	54.31	53.48	38.92	62.31	34.72	51.38	52.24	53.58	18.38	71.55
Transformer-based backbones
LaBraM	Frozen	48.58	58.35	72.45	73.89	60.83	79.25	22.81	41.80	61.81	63.49	30.53	86.30
	+LoRA	55.61	65.05	67.51	57.28	64.47	54.16	22.79	41.69	61.55	63.12	32.16	86.96

Table 1: Baseline comparisons under the limited-label setting (30% labeled subjects).

Method	CSBrain (Transformer-based)				CodeBrain (SSM-Style)
Method	Kappa	$\Delta$ %	WF1	$\Delta$ %	Kappa	$\Delta$ %	WF1	$\Delta$ %
Frozen	56.27 $\pm$ 1.49	–	64.98 $\pm$ 1.16	–	52.54 $\pm$ 1.40	–	61.75 $\pm$ 1.25	–
	Ablation on Supervision Construction
w/o ETF-guide †	60.04 $\pm$ 2.39	$\downarrow$ 2.38%	68.15 $\pm$ 2.07	$\downarrow$ 2.45%	61.48 $\pm$ 1.72	$\downarrow$ 3.12%	68.87 $\pm$ 1.35	$\downarrow$ 3.15%
w/o Prototype Clustering †	59.16 $\pm$ 1.24	$\downarrow$ 3.82%	67.53 $\pm$ 1.32	$\downarrow$ 3.33%	60.76 $\pm$ 0.94	$\downarrow$ 4.25%	67.91 $\pm$ 0.86	$\downarrow$ 4.50%
w/o Supervision construction	58.30 $\pm$ 1.63	$\downarrow$ 5.22%	66.74 $\pm$ 1.51	$\downarrow$ 4.47%	59.01 $\pm$ 1.45	$\downarrow$ 7.01%	66.78 $\pm$ 1.03	$\downarrow$ 6.09%

Table 2: Ablation on ISRUC with 30% labels. Methods with † additionally use pseudo-labeled data.

Figure 3: Sensitivity analysis on ProAdapter depth and confidence threshold.

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

Parameter-Efficient Self-Supervised Adaptation for EEG-FM under Fixed Computational Budgets
Aug 25, 2026Meghal Dani, Stefanie LiebeEEG Foundation Models Parameter-Efficient Fine-Tuning
University of Tübingen, Germany · Hertie Institute for AI in Brain Health (Hertie AI), Tübingen, Germany · Dept. of Neurology and Epileptology, University Clinic Tübingen, Germany +1
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
ZeroMAG: Zero-Shot Multimodal Adapter Generation for Plug-and-Play EEG Foundation Models
Oct 2, 2026Yubo Wang, Jingying Ma, Xinliang Zhou +8Multimodal Learning EEG Foundation Models
Brown University · Stanford University · National University of Singapore +4

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

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

Meghal Dani, Stefanie Liebe

University of Tübingen, Germany · Hertie Institute for AI in Brain Health (Hertie AI), Tübingen, Germany · Dept. of Neurology and Epileptology, University Clinic Tübingen, Germany+1

EEG Foundation Models Parameter-Efficient Fine-Tuning Data-Efficient Learning Self-Supervised Pre-Training

Open paper AskPDF arXiv

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

Abstract

EEG Foundation Models

Explore similar work

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

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

ZeroMAG: Zero-Shot Multimodal Adapter Generation for Plug-and-Play EEG Foundation Models