🤖 machine learning

Structured Prototype-Guided Adaptation for EEG Foundation Models

This paper proposes SCOPE, a structured confidence-aware prototype-guided framework that addresses the challenges of adapting EEG foundation models to label-limited settings by mitigating overconfident miscalibration and representation drift through cohort-level supervision and a lightweight prototype-conditioned adapter.

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

Published 2026-05-11

📖 4 min read☕ Coffee break read

CC BY 4.0

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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Over-Eager Student"

Imagine you have a brilliant student (an EEG Foundation Model) who has read every book in the library. They know everything about brain waves because they were trained on massive amounts of data. This is great!

However, in a real hospital, a doctor only has a few patients with confirmed diagnoses (labeled data) and thousands of patients with no notes (unlabeled data).

When you try to teach this brilliant student using just those few confirmed cases, something goes wrong. Instead of learning the new specific rules, the student gets overconfident. They start guessing wildly based on the few examples they have, ignoring their vast general knowledge. They "overfit," meaning they memorize the few examples perfectly but fail completely when they see a new patient. It's like a student who memorizes the answers to three practice questions but fails the real exam because they didn't understand the concept.

The Solution: SCOPE

The authors created a new method called SCOPE (Structured COnfidence-aware Prototype-guided adaptation). Think of SCOPE as a smart tutor that helps the student learn from the few confirmed cases without making them forget their general knowledge or guess wildly on the unknown cases.

SCOPE works in two main stages:

Stage 1: Building a "Map" and a "Trust Filter"

Before the student starts studying the new cases, the tutor builds a guide.

Creating a "Map" (Structured Prototypes):
Imagine the different types of brain states (like "sleep," "awake," "anxious") are different cities on a map. Usually, when you have very few examples, the map gets messy, and the cities blur together.
SCOPE draws a clear, structured map first. It uses math to ensure the "cities" (classes) are spaced out evenly and that there are distinct neighborhoods within each city to handle different variations. This map acts as a persistent guide so the student knows where they are supposed to be, even if the data is noisy.
The "Trust Filter" (Confidence-Aware Pseudo-Labeling):
The tutor looks at the thousands of patients with no notes. Instead of blindly guessing what their condition is, the tutor asks two different questions:
- "Based on the general map, what is this?"
- "Based on the specific neighborhoods, what is this?"
  If both answers agree, the tutor says, "Okay, this looks reliable." If they disagree, the tutor says, "I'm not sure, let's ignore this one for now." This prevents the student from learning from bad guesses.

Stage 2: The "Lightweight Glasses" (ProAdapter)

Now, it's time for the student to study.

The Problem with Old Methods: Usually, to teach a new task, you tell the student to rewrite their entire brain (Full Fine-Tuning). This is dangerous because they might erase all the knowledge they gained from reading the library.
SCOPE's Approach: Instead of rewriting the brain, SCOPE puts a pair of specialized glasses on the student. These are called ProAdapters.
- The student's main brain (the Foundation Model) stays frozen and untouched.
- The glasses are lightweight and only adjust the very top layers of the student's thinking.
- Crucially, these glasses are conditioned on the Map. They force the student to look at the new data through the lens of the structured guide built in Stage 1.

Why This Works Better

The paper tested this on six different brain tasks (like sleep staging and emotion detection) and five different types of AI models.

The Result: SCOPE consistently beat the old methods.
The Analogy: While other methods tried to force the student to memorize the few examples (leading to confusion), SCOPE gave the student a clear map and a filter to ignore bad guesses, all while keeping their original library knowledge intact.
Efficiency: It did this by changing only a tiny fraction of the model's parameters (like adjusting the glasses rather than rebuilding the brain), making it fast and cheap to run.

Summary

In short, SCOPE solves the problem of teaching powerful AI brain models when you don't have enough labeled data by:

Building a structured map of what the data should look like.
Using a trust filter to only learn from high-quality guesses.
Putting lightweight glasses on the AI to guide its thinking without breaking its existing knowledge.

This allows the AI to adapt to new, real-world clinical situations reliably, even when a doctor only has a handful of confirmed cases to work with.

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

1. Problem Statement

Electroencephalography (EEG) foundation models (EFMs) have demonstrated strong performance when fully fine-tuned on abundant labeled data. However, in real-world clinical settings, reliable subject-level annotations are scarce, often leaving only a small fraction of labeled subjects (e.g., 30% or less) alongside abundant unlabeled recordings.

The authors identify that standard full fine-tuning fails under these label-limited conditions not merely due to insufficient data, but due to a structural mismatch between noisy, limited supervision and the highly plastic parameter space of EFMs. Empirical analysis reveals three critical failure modes:

Unreliable Self-Guidance: EFMs exhibit overconfidence but inaccuracy under limited supervision, making self-training (relying on model confidence) prone to error amplification.
Prediction Collapse: Pseudo-labeling strategies often fail due to ambiguous class boundaries in continuous brain states and class imbalance, leading to biased learning toward dominant modes.
Parameter Space Vulnerability: Full fine-tuning causes significant representation drift across layers, where noisy gradients overwrite transferable pretrained representations, leading to irreversible optimization drift.

Existing parameter-efficient fine-tuning (PEFT) methods and generic semi-supervised learning approaches (e.g., FixMatch, FineSSL) have been shown to be unstable or suboptimal when applied directly to EFMs in cross-subject, label-limited scenarios.

2. Methodology: SCOPE Framework

The authors propose SCOPE (Structured COnfidence-aware Prototype-guided adaptation), a two-stage framework designed to adapt EFMs without overwriting pretrained knowledge.

Stage I: External Structured Supervision Construction

Instead of relying on the EFM's own predictions for self-training, SCOPE constructs external supervision using a lightweight Task-Prior Network (TPN) trained only on the small labeled set.

Cohort-Level Structured Prototypes:
- Inter-Class Geometry: The TPN employs a Task-Prior Network with an Equiangular Tight Frame (ETF) regularizer. This enforces maximal angular separation between class weights, ensuring class-separable embeddings.
- Intra-Class Variability: To handle diverse patterns within a class, the framework learns multiple balanced prototypes per class using Sinkhorn-Knopp optimization. This ensures a balanced assignment of samples to prototypes, preventing collapse into dominant modes.
Confidence-Aware Sample-Level Pseudo-Labeling:
- Pseudo-labels are generated only when the task-prior prediction (emphasizing inter-class separation) agrees with the prototype-based prediction (capturing intra-class variation).
- A Dempster-Shafer fusion mechanism combines these two evidence sources to quantify reliability.
- Samples are selected for adaptation only if their fused confidence score exceeds a predefined threshold, filtering out unreliable signals.

Stage II: Prototype-Conditioned Adaptation (ProAdapter)

To address the vulnerability of the EFM's parameter space, SCOPE freezes the backbone and introduces a lightweight ProAdapter inserted into the last $L$ layers of the network.

Mechanism: The adapter uses feature-wise modulation (scaling $\alpha$ and shifting $\beta$ ) conditioned on the sample's similarity vector to the learned prototypes.
Architecture:
- It generates base modulation parameters from the layer's own statistics (self-conditioning).
- It generates an additional modulation term from the prototype similarity vector, explicitly gated and bounded to ensure stable adaptation.
- This allows the structured geometry of the external supervision to guide intermediate feature adaptation without modifying the pretrained weights.

Training Strategy

SCOPE employs a warm-up adaptation strategy. Training begins with supervised updates on the labeled set, then gradually incorporates unlabeled data using the confidence-weighted pseudo-labels. This prevents noisy supervision from dominating early training stages.

3. Key Contributions

External Cohort-Level Structured Prototypes: The first framework to construct persistent, geometry-regularized prototypes (via ETF and Sinkhorn-Knopp) to provide stable external guidance for EFMs, addressing the issue of class mixing and mode collapse.
Confidence-Aware Sample-Level Pseudo-Labeling: A novel labeling strategy that fuses task-prior and prototype predictions using Dempster-Shafer theory to filter uncertain samples, mitigating the overconfidence of EFMs.
Prototype-Conditioned Adaptation (ProAdapter): A lightweight adapter that injects prototype geometry into frozen EFMs via feature modulation, preserving pretrained representations while adapting to specific tasks.
Comprehensive Evaluation: A systematic study across 6 EEG tasks, 5 foundation model backbones (including Transformer and State-Space Model architectures), and varying label ratios (5%–50%).

4. Experimental Results

The authors evaluated SCOPE on six public datasets (ISRUC, SEED, Mental Arithmetic, PhysioNet-MI, SHU-MI, CHB-MIT) using five different EFM backbones (LaBraM, CBraMod, CSBrain, EEGMamba, CodeBrain).

Performance: SCOPE consistently outperformed baselines, including full fine-tuning, frozen backbones, LoRA, and existing semi-supervised methods (FixMatch, FineSSL).
- On the ISRUC dataset with 30% labeled subjects, SCOPE improved Kappa by +7.61 (LaBraM) and +10.92 (CodeBrain) compared to frozen backbones.
- It demonstrated robustness across different architectural families (Transformer vs. SSM) and datasets where other methods failed or collapsed.
Efficiency: The method achieves these gains with only 2–5% trainable parameters and significantly reduced training time and GPU memory compared to full fine-tuning.
Ablation Studies:
- Removing the supervision construction module caused the largest performance drops (5–7%), confirming that self-training is insufficient for EFMs.
- Removing ProAdapter led to significant degradation (up to 10% on SSM backbones), highlighting the necessity of prototype-conditioned modulation.
- Confidence weighting and warm-up strategies were shown to be critical for stability.

5. Significance and Claims

The paper claims to be the first systematic study on adapting EEG foundation models under label-limited practical deployment conditions.

The authors argue that the unreliability of full fine-tuning in low-label regimes stems from a structural mismatch, not just data scarcity. By shifting from self-training and sample-level pseudo-labels to external cohort-level structured supervision and conditioned adaptation, SCOPE provides a more reliable and data-efficient pathway for deploying EFMs in real-world clinical scenarios where expert annotations are limited. The work suggests that preserving pretrained representations through lightweight, geometry-guided modulation is superior to direct parameter updates in these settings.