Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Structured Prototype-Guided Adaptation for EEG Foundation Models

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

TL;DR

SCOPE, a Structured COnfidence-aware Prototype-guided adaptation framework for EFM fine-tuning, is proposed and demonstrated that SCOPE consistently achieves strong performance and efficiency under label-limited cross-subject settings.

Abstract

Electroencephalography (EEG) foundation models (EFMs) have achieved strong performance under full fine-tuning but exhibit poor generalization when subject-level supervision is limited, a common constraint in real-world clinical settings. We show that this failure stems not merely from limited supervision, but from a structural mismatch between noisy, limited supervision and the highly plastic parameter space of EFMs. To address this challenge, we propose SCOPE, a Structured COnfidence-aware Prototype-guided adaptation framework for EFM fine-tuning. SCOPE follows a two-stage pipeline. In the first stage, we construct reliable external supervision by learning geometry-regularized task priors, constructing balanced class-level prototypes over the resulting embeddings, and producing confidence-aware pseudo-labels from their agreement to filter unreliable signals on unlabeled data. In the second stage, we introduce ProAdapter, which adapts frozen EEG foundation models via a lightweight adapter conditioned on the structured prototypes. Experiments across three EEG tasks and five foundation model backbones demonstrate that SCOPE consistently achieves strong performance and efficiency under label-limited cross-subject settings.

Structured Prototype-Guided Adaptation for EEG Foundation Models

TL;DR

SCOPE, a Structured COnfidence-aware Prototype-guided adaptation framework for EFM fine-tuning, is proposed and demonstrated that SCOPE consistently achieves strong performance and efficiency under label-limited cross-subject settings.

Abstract

Electroencephalography (EEG) foundation models (EFMs) have achieved strong performance under full fine-tuning but exhibit poor generalization when subject-level supervision is limited, a common constraint in real-world clinical settings. We show that this failure stems not merely from limited supervision, but from a structural mismatch between noisy, limited supervision and the highly plastic parameter space of EFMs. To address this challenge, we propose SCOPE, a Structured COnfidence-aware Prototype-guided adaptation framework for EFM fine-tuning. SCOPE follows a two-stage pipeline. In the first stage, we construct reliable external supervision by learning geometry-regularized task priors, constructing balanced class-level prototypes over the resulting embeddings, and producing confidence-aware pseudo-labels from their agreement to filter unreliable signals on unlabeled data. In the second stage, we introduce ProAdapter, which adapts frozen EEG foundation models via a lightweight adapter conditioned on the structured prototypes. Experiments across three EEG tasks and five foundation model backbones demonstrate that SCOPE consistently achieves strong performance and efficiency under label-limited cross-subject settings.
Paper Structure (47 sections, 2 theorems, 26 equations, 14 figures, 17 tables, 1 algorithm)

This paper contains 47 sections, 2 theorems, 26 equations, 14 figures, 17 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Scalp EEG Foundation Models.
  4. Parameter-efficient adaptation for foundation models.
  5. Methodology
  6. Problem Setting and Framework Overview
  7. External structured supervision construction
  8. Prototype-Conditioned Adaptation
  9. Training Objective
  10. Experiments
  11. Datasets
  12. Experiment Settings
  13. Performance Comparison.
  14. Ablation Study
  15. Sensitivity Analysis
  16. ...and 32 more sections

Key Result

Proposition 3.1

Let $\{\tilde{\mathbf{w}}_{k}\}_{k=1}^{K}\subset\mathbb{S}^{d-1}$ be the normalized classifier weights. If they satisfy the simplex equiangular tight frame condition $\tilde{\mathbf{w}}_{k}^{\top}\tilde{\mathbf{w}}_{k'} = -\frac{1}{K-1}, \qquad \forall\, k\neq k',$, the minimum pairwise angular sepa

Figures (14)

  • Figure 1: CodeBrain under full fine-tuning on the ISRUC dataset. (A) Training and validation Kappa trajectories, illustrating the generalization gap under 30% versus 100% subjects supervision. (B) Validation Kappa across five random seeds when training with 30% labeled subjects, showing high sensitivity to initialization. (C) Calibration curve on the validation set under 30% subjects supervision, indicating overconfident predictions.
  • Figure 2: CodeBrain (8-layers EFM backbone) under full fine-tuning on the ISRUC dataset with 30% labeled subjects. (A) Early-stage training dynamics across different random seeds, showing instability and occasional collapse. (B) Layer-wise evolution of parameter space similarity during training, measured by centered kernel alignment (CKA), with low, mid, and deep layers.
  • Figure 3: 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.
  • Figure 4: Sensitivity analysis of ProAdapter depth on ISRUC.
  • Figure 5: Sensitivity analysis of confidence threshold on ISRUC.
  • ...and 9 more figures

Theorems & Definitions (2)

  • Proposition 3.1
  • Lemma 3.2