PaperlayerView exact revision on arXiv

Source-linked AI summary

Lightweight Adaptation of EEG Foundation Models for Stroke Motor Imagery Decoding: Domain Shift and Subject-Level Robustness

Anh T. Nguyen, Zihua Sun, Michelle J. Johnson

arXiv:2609.00282v1cs.CEcs.LG

TL;DR

This paper asks whether pretrained EEG foundation models can reliably support stroke MI decoding despite healthy-to-stroke domain shift. It evaluates head-only models and LoRA adaptation across healthy and stroke datasets, with robustness and subject-level analyses. LoRA improved healthy-dataset performance, but stroke adaptation differed by model and zero-shot transfer remained near chance, supporting target-domain and subject-level evaluation.

  • Problem

    Healthy-dataset performance may not transfer reliably to pathological EEG, while clinically relevant subject-level robustness remains insufficiently tested.

  • Method

    The study compares frozen-backbone and LoRA-adapted EEG foundation models using subject-wise cross-validation, preprocessing ablation, ERD analysis, channel analysis, and subject-level evaluation.

  • Results

    LoRA improved both LaBraM-base and REVE-base on healthy EEG, but on stroke EEG REVE-base reached 0.847±0.194 while LaBraM-base remained near chance; zero-shot transfer remained near chance.

  • Takeaways & Limitations

    Stroke MI decoding requires target-domain adaptation, preprocessing-aware evaluation, subject-level robustness analysis, and physiological interpretation beyond healthy-benchmark performance.

  • Takeaways & Limitations

    The healthy and stroke datasets used different imagined movements and task paradigms, so the performance gap cannot be attributed exclusively to stroke-related neurophysiological domain shift.

Abstract

from arXiv · show

Motor imagery (MI) electroencephalography (EEG) decoding could support post-stroke rehabilitation, but models developed on healthy cohorts may not transfer reliably to pathological EEG. We evaluated whether Low-Rank Adaptation (LoRA) can efficiently adapt three pretrained EEG foundation models (i.e., LaBraM-base, REVE-base, and REVE-large) for binary left- versus right-hand MI decoding. Frozen-backbone head-only baselines and LoRA adaptation were evaluated using subject-wise five-fold cross-validation on the PhysioNet EEG Motor Movement/Imagery Dataset and a binary subset of the UET175 dataset comprising 30 stroke participants. On EEGMMIDB, LoRA increased accuracy to 0.822 for LaBraM-base and 0.957 for REVE-base. On UET175, all head-only models performed near chance. With LoRA, LaBraM-base remained near chance (0.499$\pm$0.009), whereas REVE-base reached 0.847$\pm$0.194 and outperformed REVE-large (0.806$\pm$0.178), indicating that increased model capacity alone did not improve stroke-domain adaptation. The strongest stroke configuration, REVE-base LoRA, was further evaluated using within-cohort leave-one-subject-out cross-validation (LOOCV), showing 0.952 mean accuracy, but subject-wise accuracy ranged from 0.586 to 1.000, revealing a small low-performing tail. Zero-shot transfer from EEGMMIDB to UET175 remained near chance (0.464$\pm$0.072). These findings show that healthy-benchmark performance does not ensure transfer to stroke EEG. Translation of EEG foundation models to pseudo-online or real-time rehabilitation BCIs should therefore include target-domain adaptation and subject-level assessment of temporal informativeness, spatial sensitivity, and physiological discriminability.

I. INTRODUCTION

MI-EEG BCIs are promising for rehabilitation, but clinical transfer remains difficult because stroke EEG is heterogeneous and healthy-dataset performance often degrades. This study examines whether pretrained EEG foundation models and lightweight adaptation improve robustness under healthy-to-stroke domain shift.

  • Stroke EEG decoding is challenging because patient variability, inconsistent signal quality, and pathological differences can reduce transfer performance.
  • Conventional MI pipelines often require task-specific design and remain vulnerable to cross-subject and clinically heterogeneous deployment.
  • EEG foundation models seek reusable representations from large, diverse corpora, while parameter-efficient tuning offers adaptation with limited data and compute.
  • The study benchmarks pretrained models on healthy and stroke MI datasets, then evaluates fine-tuning and robustness analyses under domain shift.
  • Existing evaluations largely emphasize standard or non-clinical tasks, leaving subject-level effectiveness under severe healthy-to-stroke shift insufficiently tested.

C. Parameter-Efficient Fine-Tuning for EEG

The study adapts pretrained EEG foundation models for binary MI decoding using parameter-efficient methods and compares their preprocessing, patching, spatial encoding, and pooling choices across healthy and stroke datasets.

  • LoRA adapts pretrained EEG backbones by freezing them and inserting trainable low-rank updates, making it suitable for small EEG datasets and limited compute.
  • The experiments use healthy EEGMMIDB and post-stroke UET175 recordings for binary left- versus right-hand MI decoding.
  • LaBraM and REVE both use patch-based Transformer encoders but differ in spatial parameterization and self-supervised pretraining objectives.
  • Both models produce four channel–time tokens per channel from 4.0-second, 200-Hz trials, although LaBraM uses non-overlapping patches and REVE uses overlapping patches.
  • LaBraM uses learned electrode and temporal embeddings, whereas REVE uses electrode coordinates with temporal position and spatial augmentation.
  • LaBraM mean-pools normalized Transformer representations, while REVE uses learned-query pooling to weight channel–time tokens adaptively.

1) Head-Only Baseline:

The head-only baseline freezes each pretrained encoder and trains only a new classification head, measuring the MI information already accessible in the pretrained representations.

  • Frozen-backbone head-only evaluation optimizes only a task-specific classification head while keeping pretrained encoder parameters fixed.

2) LoRA Fine-Tuning:

The LoRA experiments apply low-rank updates to selected attention projections while preserving the pretrained backbones, using subject-wise cross-validation and multiple classification metrics.

  • LoRA freezes the pretrained backbone and inserts trainable low-rank updates into selected linear projections.
  • REVE-large LoRA serves as a capacity-control experiment testing whether a larger backbone improves lightweight adaptation.
  • All experiments use subject-wise five-fold cross-validation and report accuracy, balanced accuracy, Cohen’s κ, and weighted F1-score.

F. Secondary Analyses

Secondary analyses used the strongest primary configuration to test robustness, subject-level generalization, clinical matching, and spatial sensitivity in stroke decoding.

  • Secondary-analysis design: The strongest primary model-and-adaptation setting was reused for follow-up analyses of robustness, preprocessing sensitivity, and stroke-decoding failure cases.Secondary analyses were not generally intended for further model selection.
  • Subject-level robustness: LOOCV held out one of 30 stroke subjects per fold to quantify subject-wise generalization and identify low-performing cases.Each fold used one test subject, two validation subjects, and 27 training subjects.
  • Clinical matching: Low-performing subjects were compared qualitatively with higher-performing subjects matched on lesion location and available severity measures.Matching used NIHSS, mRS, and muscle strength when available, without constituting a formal matched-cohort analysis.
  • Channel ablation: Channel ablations evaluated 22-channel, frontocentral-plus-central, and central-only montages under the same REVE-base LoRA stroke LOOCV protocol.The central-only subset contained C3, CZ, and C4, testing whether motor-region restriction improves robustness.

4) Time-Window Ablation:

The study examined temporal informativeness and physiological failure patterns using REVE-base LoRA, standardized trial inputs, and ERD comparisons across healthy and stroke subjects.

  • Time-window ablation: Time-window ablation compared the full 0–4 s trial with four shorter windows under the stroke LOOCV protocol.Selected windows were zero-padded to the original (C, 800) shape, preserving the model’s patch structure.
  • ERD analysis: ERD analysis used C3, Cz, and C4 in the mu band (8–13 Hz) and beta band (13–30 Hz).ERD was computed from a −1 to 0 s baseline and a 0 to +4 s task window.
  • ERD interpretation: Negative ERD values represented desynchronization, with selected summaries additionally reporting mean ERD over 0.5–3.5 s.Comparisons included one healthy reference, one high-performing stroke subject, and one low-performing stroke subject.
  • Zero-shot transfer: Zero-shot evaluation trained the best-performing model on EEGMMIDB and tested it directly on all 30 stroke subjects without stroke-domain fine-tuning.The 22-channel stroke input was mapped into the 64-channel healthy-dataset space with missing channels zero-padded.

B. Primary Analysis: Head-Only Baseline and LoRA Adaptation

Frozen-backbone models transferred poorly to stroke EEG, while LoRA adaptation produced architecture-dependent gains; REVE-base offered the strongest practical stroke configuration without benefit from stronger ASR preprocessing.

  • Head-only baseline: 0.484±0.016, 0.495±0.022, and 0.492±0.014 accuracy kept LaBraM-base, REVE-base, and REVE-large near chance on stroke head-only decoding.The same models reached 0.546±0.010, 0.619±0.018, and 0.697±0.023 on EEGMMIDB.
  • LoRA on healthy EEG: 0.822 and 0.957 accuracy were achieved by LaBraM-base and REVE-base LoRA on EEGMMIDB, rising from 0.546 and 0.619.These results show strong healthy-domain gains from lightweight adaptation.
  • LoRA on stroke EEG: 0.499±0.009 versus 0.847±0.194 accuracy showed sharply divergent stroke performance for LaBraM-base and REVE-base after LoRA adaptation.LaBraM-base remained near chance, whereas REVE-base motivated selection as the main stroke configuration.
  • Model capacity: 0.806±0.178 versus 0.847±0.194 accuracy showed that REVE-large LoRA did not exceed REVE-base LoRA on UET175.The comparison indicates that greater model capacity alone did not improve stroke-domain adaptation.
  • ASR ablation: 0.719±0.243 versus 0.847±0.194 accuracy showed that ASR reduced REVE-base LoRA performance on UET175.ASR also reduced the evaluated EEGMMIDB REVE-base LoRA result from 0.957±0.056 to 0.896±0.060.

2) LOOCV Reveals a Small Low-Performing Subject Tail:

REVE-base LoRA achieved high mean stroke-decoding accuracy, but subject-level LOOCV exposed a small low-performing tail. Ablations further showed that spatial and temporal preprocessing effects were heterogeneous across subjects.

  • Subject-level LOOCV: 0.9524 mean accuracy was achieved across 30 stroke subjects, but subject-wise accuracy ranged from 0.586 to 1.000.Twenty-seven subjects exceeded 0.90 accuracy, while ID01 and ID02 achieved 0.586 and 0.614, respectively.
  • Channel ablation: C-only input reached 0.9661 mean accuracy versus 0.9524 for S22, but the difference was not statistically significant.The paired subject-wise comparison yielded p = 0.37.
  • Channel ablation: ID02 improved from 0.6136 to 0.9773 with C-only input, whereas ID01 decreased from 0.5862 to 0.5517.Only 7 subjects improved, 10 decreased, and 13 remained unchanged under C-only input.
  • Time-window ablation: 0.5–2.5 s achieved the highest tested mean accuracy, reaching 0.9605 with Cohen’s κ = 0.9205.The result suggests that discriminative information was concentrated in an early post-cue interval rather than uniformly distributed across the full trial.
  • Time-window ablation: ID02 improved from 0.6136 to 0.7273 under the 0.5–2.5 s window, whereas ID01 decreased from 0.5862 to 0.4483.Temporal cropping therefore benefited some low-performing subjects but not all.
  • Physiological analysis: ERD analyses showed that low-performing subjects retained MI-related activity but differed in task-dependent spatial organization and frequency-band structure.The comparison used ID01 and ID02 alongside clinically matched higher-performing participants.

5) Clinical Matching Shows That Failure Is Not Explained Solely by Gross Lesion Severity:

Clinically matched stroke participants could differ substantially in decoding performance despite similar lesion locations and severity measures. Their ERD patterns indicate that decoding difficulty reflects non-discriminability of task-related structure rather than absent MI activity.

  • Clinical matching: ID01 achieved 0.586 accuracy versus 0.983 for clinically matched ID09 despite right temporal infarcts and identical available severity measures.Both participants had NIHSS=2, mRS=2, and muscle strength 4/5.
  • Clinical matching: ID02 achieved 0.614 accuracy versus 0.766 for ID05 despite shared left lentiform-region involvement and comparable clinical severity measures.The comparison indicates that gross clinical severity and lesion location alone did not fully explain decoding difficulty.
  • ERD comparison: Both low-performing participants exhibited substantial mu-band ERD, while their spatial distribution, task specificity, and frequency-band organization differed from higher-performing comparisons.Thus, MI-related neural activity was not completely absent in the low-performing participants.
  • ERD comparison: ID01 showed abnormal spatial organization with weak beta modulation, whereas ID02 showed nearly symmetric mu ERD and minimal beta modulation.The higher-performing comparison participants exhibited more distinct task-dependent spatial and frequency-band patterns.
  • Interpretation: Successful decoding appeared to depend less on ERD magnitude than on stable, task-dependent spatial and frequency-band structure.The authors characterize the low-performing patterns as physiological non-discriminability rather than complete absence of MI-related activity.

V. DISCUSSION

LoRA adaptation separated healthy-domain representation quality from stroke-domain transferability: REVE-base adapted strongly on UET175, whereas LaBraM-base remained near chance and zero-shot transfer failed. Subject-level analyses further showed heterogeneous failure patterns and sensitivity to preprocessing and input selection.

  • Foundation Model Adaptation Under Stroke Domain Shift: LoRA improved LaBraM-base and REVE-base on EEGMMIDB, but only REVE-base remained strongly adaptable on the UET175 stroke subset.The comparison distinguishes useful pretrained MI representations from representations that remain compatible with stroke EEG under domain shift.
  • Foundation Model Adaptation Under Stroke Domain Shift: REVE-base outperformed REVE-large under LoRA on UET175, indicating that increased backbone capacity alone did not improve stroke-domain adaptation.The authors identify REVE-base as the best performance-efficiency tradeoff among evaluated configurations.
  • Foundation Model Adaptation Under Stroke Domain Shift: Direct healthy-to-stroke transfer failed, supporting target-domain adaptation as necessary for applying healthy-trained MI decoders to stroke EEG.The REVE-base LoRA model trained on EEGMMIDB performed below chance with negative Cohen’s κ on UET175.
  • Preprocessing Sensitivity: ASR-based preprocessing reduced performance across evaluated settings, suggesting that preprocessing can alter input statistics unfavorably for pretrained EEG pipelines.The interpretation is limited to the evaluated configurations and does not establish that ASR is universally harmful.
  • Subject-Level Robustness: LOOCV revealed strong average REVE-base LoRA performance but a small low-performing subject tail that cohort means did not represent.Within-cohort LOOCV should be interpreted as a robustness analysis rather than external validation.
  • Subject-Level Robustness: Low-performing subjects differed in response to channel and time-window restrictions, with ID02 showing input-selection sensitivity and ID01 remaining difficult across configurations.The C-only improvement for ID02 was not consistent across the cohort and did not reach group-level statistical significance.

D. Physiological Interpretation From ERD Analysis

ERD analyses showed that low decoding performance can coexist with substantial motor-related activity when its spatial and frequency organization is weak. The findings therefore support subject-level physiological assessment alongside model evaluation in rehabilitation BCI studies.

  • Physiological Interpretation From ERD Analysis: Both ID01 and ID02 showed mu-band ERD, but their patterns were weakly task-discriminative, with abnormal lateralization or minimal beta modulation.The presence of MI-related activity did not imply clear left-versus-right class structure.
  • Physiological Interpretation From ERD Analysis: The healthy reference and high-performing stroke subject showed more organized task-related ERD, whereas the low-performing stroke subject showed large but non-specific ERD.This comparison indicates that ERD magnitude alone may not be sufficient for decoding.
  • Physiological Interpretation From ERD Analysis: Strong neural activation may still yield poor classification when left- and right-hand imagery produce similar or abnormally lateralized responses.Binary MI decoding requires consistent class-specific structure, not merely large overall ERD.
  • Implications for Rehabilitation BCI: Healthy benchmark performance should not be treated as sufficient evidence for stroke rehabilitation deployment because zero-shot healthy-to-stroke transfer failed.The paper recommends target-domain adaptation and subject-level robustness analysis for clinical evaluation.
  • Implications for Rehabilitation BCI: Future calibration should assess temporal informativeness, spatial sensitivity, and physiological discriminability on a subject-specific basis.Channel and time-window ablations are presented as sensitivity analyses rather than universal preprocessing prescriptions.
  • Limitations: The healthy-to-stroke performance gap cannot be attributed exclusively to stroke neurophysiology because the datasets used different imagined movement paradigms.The stroke cohort also comprised 30 subjects, and LOOCV did not replace external validation.
Loading 2609.00282v1…