Background Brain-computer interfaces (BCIs) hold significant potential for post-stroke motor recovery, yet active movement-based BCIs face limitations in generalization due to inter-subject variability. This study investigates passive movement-based BCIs, driven by exoskeleton-guided rehabilitation, to address these challenges by evaluating electroencephalogram (EEG) responses and algorithmic generalization in both healthy subjects and stroke patients. Methods EEG signals were recorded from 20 healthy subjects and 10 stroke patients during voluntary and passive hand movements. Time and time-frequency domain analyses were performed to examine the event-related potential (ERP), event-related desynchronization (ERD), and synchronization (ERS) patterns. The performance of two BCI algorithms, Common Spatial Patterns (CSP) and EEGNet, was evaluated in both within-subject and cross-subject decoding tasks. Results Time-domain and time-frequency analyses revealed that passive movements elicited stronger, more consistent ERPs in healthy subjects, particularly in bilateral motor cortices (contralateral: <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="12984_2025_1627_Article_IEq1.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="94" /> </InlineMediaObject> <EquationSource Format="TEX">$-7.29\pm 4.51$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>-</mo> <mn>7.29</mn> <mo>±</mo> <mn>4.51</mn> </mrow> </math></EquationSource> </InlineEquation> μV; ipsilateral: <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="12984_2025_1627_Article_IEq2.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="94" /> </InlineMediaObject> <EquationSource Format="TEX">$-4.33\pm 3.69$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>-</mo> <mn>4.33</mn> <mo>±</mo> <mn>3.69</mn> </mrow> </math></EquationSource> </InlineEquation> μV). Stroke patients exhibited impaired mu/beta ERD/ERS in the affected hemisphere during voluntary movements but demonstrated EEG patterns during passive movements resembling those of healthy subjects. Machine learning evaluation highlighted EEGNet’s superior performance, achieving 84.19% accuracy in classifying affected vs. unaffected movements in patients, surpassing healthy subject left-right discrimination (58.38%). Cross-subject decoding further validated passive movement efficacy, with EEGNet attaining 86.00% (healthy) and 72.63% (stroke) accuracy, outperforming traditional CSP methods. Conclusions These findings underscore that passive movement elicits consistent neural responses, thereby enhancing the generalizability of decoding algorithms for stroke patients. By integrating exoskeleton-evoked proprioceptive feedback, this paradigm reduces inter-subject variability and improves clinical feasibility. Future work should explore the application of exoskeletons in the combination of active and passive movement for stroke rehabilitation.

Exoskeleton-guided passive movement elicits standardized EEG patterns for generalizable BCIs in stroke rehabilitation

期刊论文
出版模式：

开放获取
发表日期：

2025年4月26日

查看全文

Xinyi Zhang,
Lanfang Xie,
Wanting Liu,
Shaoying Liang,
Liyao Huang,
Mingjun Wang,
Lingling Tian,
Li Zhang,
Zhen Liang,
Hai Li,
Gan Huang

摘要

Background

Brain-computer interfaces (BCIs) hold significant potential for post-stroke motor recovery, yet active movement-based BCIs face limitations in generalization due to inter-subject variability. This study investigates passive movement-based BCIs, driven by exoskeleton-guided rehabilitation, to address these challenges by evaluating electroencephalogram (EEG) responses and algorithmic generalization in both healthy subjects and stroke patients.

Methods

EEG signals were recorded from 20 healthy subjects and 10 stroke patients during voluntary and passive hand movements. Time and time-frequency domain analyses were performed to examine the event-related potential (ERP), event-related desynchronization (ERD), and synchronization (ERS) patterns. The performance of two BCI algorithms, Common Spatial Patterns (CSP) and EEGNet, was evaluated in both within-subject and cross-subject decoding tasks.

Results

Time-domain and time-frequency analyses revealed that passive movements elicited stronger, more consistent ERPs in healthy subjects, particularly in bilateral motor cortices (contralateral: $-7.29\pm 4.51$ $- 7.29 \pm 4.51$ μV; ipsilateral: $-4.33\pm 3.69$ $- 4.33 \pm 3.69$ μV). Stroke patients exhibited impaired mu/beta ERD/ERS in the affected hemisphere during voluntary movements but demonstrated EEG patterns during passive movements resembling those of healthy subjects. Machine learning evaluation highlighted EEGNet’s superior performance, achieving 84.19% accuracy in classifying affected vs. unaffected movements in patients, surpassing healthy subject left-right discrimination (58.38%). Cross-subject decoding further validated passive movement efficacy, with EEGNet attaining 86.00% (healthy) and 72.63% (stroke) accuracy, outperforming traditional CSP methods.

Conclusions

These findings underscore that passive movement elicits consistent neural responses, thereby enhancing the generalizability of decoding algorithms for stroke patients. By integrating exoskeleton-evoked proprioceptive feedback, this paradigm reduces inter-subject variability and improves clinical feasibility. Future work should explore the application of exoskeletons in the combination of active and passive movement for stroke rehabilitation.