Classification of error-related potentials evoked during stroke rehabilitation training

Area of Science:

• Neuroscience
• Rehabilitation Engineering
• Machine Learning

Background:

• Error-related potentials (ErrPs) are brain signals indicating error perception.
• ErrPs have potential applications in developing adaptive robotic stroke rehabilitation systems.
• Feasibility of using ErrPs for real-time assist-as-needed (AAN) systems is unexplored.

Purpose of the Study:

• To evaluate and compare machine learning and deep learning methods for classifying single-trial ErrPs.
• To assess the feasibility of using ErrPs for real-time AAN robotic stroke rehabilitation.

Main Methods:

• Utilized electroencephalogram (EEG) data from 13 stroke patients during upper-limb rehabilitation.
• Employed two classification approaches: xDAWN spatial filtering with support vector machines (SVM) and convolutional neural network (CNN)-based double transfer learning.
• Analyzed latency, cross-subject, and asynchronous classification for real-time feasibility.

Main Results:

• Achieved a mean area under the ROC curve of 0.838 and mean accuracy of 0.842 for within-subject ErrP classification.
• Demonstrated classification performance significantly above chance level (p<0.05).
• Latency and cross-subject analyses supported the feasibility of real-time application.

Conclusions:

• Single-trial ErrP classification is feasible for real-time AAN robotic stroke rehabilitation.
• Machine learning and deep learning approaches show promise for developing adaptive rehabilitation systems.
• ErrPs can serve as a reliable measure to trigger or modulate robotic assistance in human-in-the-loop systems.