Error-Related Negativity-Based Robot-Assisted Stroke Rehabilitation System: Design and Proof-of-Concept

Abstract

Conventional rehabilitation systems typically execute a fixed set of programs that most motor-impaired stroke patients undergo. In these systems, the brain, which is embodied in the body, is often left out. Including the brains of stroke patients in the control loop of a rehabilitation system can be worthwhile as the system can be tailored to each participant and, thus, be more effective. Here, we propose a novel brain-computer interface (BCI)-based robot-assisted stroke rehabilitation system (RASRS), which takes inputs from the patient's intrinsic feedback mechanism to adapt the assistance level of the RASRS. The proposed system will utilize the patients' consciousness about their performance decoded through their error-related negativity signals. As a proof-of-concept, we experimented on 12 healthy people in which we recorded their electroencephalogram (EEG) signals while performing a standard rehabilitation exercise. We set the performance requirements beforehand and observed participants' neural responses when they failed/met the set requirements and found a statistically significant (p < 0.05) difference in their neural responses in the two conditions. The feasibility of the proposed BCI-based RASRS was demonstrated through a use-case description with a timing diagram and meeting the crucial requirements for developing the proposed rehabilitation system. The use of a patient's intrinsic feedback mechanism will have significant implications for the development of human-in-the-loop stroke rehabilitation systems.

Keywords: Training-ErrPs; assist-as-needed (AAN); brain-computer interface (BCI); error-related potentials (ErrP); robot-therapy; single-trial classification; stroke rehabilitation.

Figure 1

Setup of the event-related negativity (ERN)-based robot-assisted stroke rehabilitation system (RASRS). The rehabilitation robot would assist the patient in performing the rehabilitation exercise as required. The instruction monitor would guide the patient in performing the rehabilitation exercise by showing relevant cues on the screen and would also show the real-time feedback of the movement of the impaired limb with the help of the Azure Kinect sensor, so that the patient can know how much exercise has been performed and how much it is left. The electroencephalogram (EEG) sensors record the EEG activity from the patient's brain. The control unit would process the EEG data, the Kinect sensor data, give inputs to the instruction monitor, and signals the rehabilitation robot to increase/decrease the robotic assistance.

Figure 2

Four typical trials of the ERN-RASRS illustrated using a timing diagram. The durations in the figure are not scaled. Event e marks the time when the ERN associated with the Training-ErrP signal elicits. Event d marks the time when the ERN signal is detected. x shows the minimum level of assistance provided by the rehabilitation robot throughout the exercise trial. The x+1 shows the level of assistance provided by the rehabilitation robot on the detection of the ERN signal. Event i marks the time when robotic assistance is increased to level x+1. Event f marks the time when the current trial finishes and the robotic assistance is decreased to level x. Event s marks an exercise trial in which almost the whole exercise is performed only with the minimum level of robotic assistance (i.e., x); thus, the difficulty level of the exercise is increased, and the minimum level of robotic assistance (i.e., x) is decreased.

Figure 3

(A) A fixation cross marked the start of a trial. (B) A pre-recorded video showed the rehabilitation exercise to be performed. (C) A 3-2-1 timer. (D) Instruction to the participants to start performing the rehabilitation exercise depicted in the video. (E) Participants were asked to complete the rehabilitation exercise before this “Time's up!” screen appeared. This marked the completion of one exercise trial, and subsequently, the next trial started with the fixation cross. The epoch of interest for EEG analyses is marked in red on the timeline, immediately after the “Time's up!” screen appeared.

Figure 4

Graphical representation of an error trial structure and the evaluation period. All occurrences that are neither labeled nor detailed inherit their description from the corresponding preceding node. Notably, in no detection, no error event was detected in the evaluation period despite the presence of an error event.

Figure 5

The grand-average waveform of error-related negativity, correct-related activity, and Training-ErrP signal at the Fz electrode location. The green-bars mark the time-points when the Training-ErrP signal is statistically significantly different against zero after cluster-based multiple comparisons correction. For illustration purposes, the waveforms have been smoothed out with a 10 Hz low-pass finite impulse response (FIR) filter.

Figure 6

Topographical scalp maps of the grand-average error-related negativity and correct-related activity signals at two time-points, where the Training-ErrP signal was observed to be prominent.

Figure 7

Row-normalized confusion matrix for single-trial binary classification of error-related negativity (ERN) epochs against the correct-related activity (CRA) epochs, using 5-fold cross-validation.

Figure 8

Percentage of error trials that were correctly detected, incorrectly detected, or no detection was observed, as a function of the sensitivity. The error bars represent the 99% confidence interval for the averages. The incidence-detection rate (IDR) metric, represented using pink line, is highest at sensitivity level 1. The chance-level IDR, calculated using a permutation test, is also shown in orange.