Researchers are exploring new ways to help people recover after a stroke, and one promising approach is embodied neurofeedback training, using motor imagery (NF-MIT).
In this study, participants used motor-imagery to move a robotic hand, while receiving real-time brain feedback via an embodiable, anthropomorphic robotic hand (RH), which was integrated into a closed-looped EEG-based brain-computer interface (BCI).
The idea is that by combining motor imagery with feedback and an embodiable robotic hand, patients may strengthen brain–body connections important for rehabilitation.
Setting up the embodied neurofeedback loop with a robotic hand
Researchers recruited 9 chronic stroke patients for the study, and 9 healthy controls that were matched in age and sex.
To see the cognitive assessment of accepted stroke patients, read the original study linked here.
Participants placed their right hand into a black box, whereas the robotic hand was placed alongside in front of the participant.
In the image, we can see the closed loop system used in the study. The physical setup is pictured at the top, with the robotic hand on the table. See more detail on the loop in the following text.Task description and hand movement classification using EEG
EEG data was collected in real-time using the Smarting mobile EEG amplifier. During the experiment, the participants did three embodied neurofeedback tasks + an acquittance phase. All of these included two hand movements: flexion or extension.
Acquaintance phase: the participant familiarized with the motor-imagery training (MIT). During this segment, the experimenters could also adjust the classifier thresholds. For example, if a participant got stuck in a certain hand state and couldn’t get out.
Rest vs. move task: the participant attempts to move the hand as much as possible during a 30 s window. Then, tries their best not to move it for the next 30 s.
Follow commands task: participants had to carry out the command of the experimenter. They could give 4 different commands: open, close, grasp or remain in the current state. Grasping is described as open and immediate close of the hand.
Announce commands task: the participant would announce their movements, which are the same as in the previous task.
There were 2 phases in the classification algorithm:
- Collecting training EEG data to define the classifiers
- Applying the ensemble classification on real-time EEG data
Training the classifiers
The study derived three probabilistic classifiers. The classifiers are regularized logistic regression models.
The first one had the output probability of flexion as opposed to extension. The second had the probability of flexion compared to remaining open. And the last one’s output is the probability of extension as opposed to remaining closed.
Authors applied BCILAB’s common spatial pattern (CSP) analysis to the EEG data. The output of the CSP-filter is feature matrices, which are used to train the classifiers.
Ensemble classification and real-time data flow
The movement controls depended on the weighted estimates of the three individual classifiers. The usage of the ensemble is determined by the robotic hand’s current state. You can see the real-time flow sketch below:
Schematic representation of the way the classification ensemble works in real-time.How authors evaluated the embodied neurofeedback performance
There was no fixed structure of the command and announce tasks, so authors had to define it post-hoc.
If a movement happened in the intended movement trial – it’s defined as a True Positive (TP). False Negative (FN) would be no movement during the same trial. Let’s look at the intended rest trials. If movement happens during this time, it’s a False Positive (FP). If there’s no movement, as intended, this trial is a True Negative (TN).In the rest vs. move task, they compare the number of movements during rest and move phases.
[results] Offline EEG analysis
In the time-frequency (TF) plots, we see the percentage change in power from baseline. Top two graphs are showing results for stroke patients, and the bottom for control group. Left is the flexion trials, and on the right are extension trials. Solid blue line below the TF plots shows changes in power in the 10-25 Hz range. This range is known as the sensorimotor rhythm range (SMR).As the plots show, we can see that both stroke patients and control showed power reduction in the SMR range. And for both movement types – flexion and extension. This is referred to as event-related desynchronization (ERD) in the SMR range.
But what about the ERD latencies? Do they differ between the groups?
These are defined as the time interval between MI period onset and ERD onset. An ERD onset is the time-point at which the SMR power fell by at least 30%.
Stroke patients showed delayed ERD onset compared to control group. These can possibly indicate slower processing in the damaged neural tissue.
[results] Classifier performance and robotic hand movement
Accuracy levels of the classifier ensemble. Top shows performance during training, and the bottom during the actual embodied neurofeedback task.Training accuracies are above chance level for most participants in both groups. While feedback accuracies were a bit more modest (~58% stroke and ~54% controls). Still, most participants performed at least one task above chance.
There was also no significant difference between two groups.
These results show that both groups could gain some control over the robotic hand. Which marks the first demonstration of this feasibility study.
[results] What do the questionnaire answers say?
Ratings during training and embodied neurofeedback sessions, for both stroke and control groups. The four measures are sense of ownership (SoO), agency (SoA), experiential realness (ER) and MI-action-Binding (MIAB).Participants had to rate multiple statements for all 4 measures. Here’s a few examples of those statements that were shown to the groups:
SoO example: I felt as if the robotic hand was my own hand
SoA: I felt as if I could withhold any robotic hand movements
ER: My imagined movements felt as vivid and real as if they actually happened
MIAB: I experienced my imagined hand flexions and the flexions of the RH as inseparably linked with each other
From the image above, it seems that the participants generally embodied the robotic hand.
Stroke patients met the SoA illusion criterion, but controls didn’t. This may be due to stroke patients’ weaker percepts of their own affected hand. Which makes them more prone to SoA.
Authors also reported no significant correlation between subjective measures and the classification accuracy. However, they do argue embodiment could still provide motivational and cognitive benefits, even if not directly tied to accuracy here.
Final thoughts
The study demonstrates that both healthy participants and stroke patients:
- could embody a robotic hand as part of their body schema,
- and achieved modest but real control of the RH in a self-paced NF-MIT paradigm.
This proof of principle supports the feasibility of self-paced, embodied neurofeedback training. It motivates further research into its potential for stroke motor rehabilitation.
Original paper is linked here.
Curious if this setup is right for your study?
Our team works with researchers across neuroscience, sports science and clinical applications - we'll help you figure out if mobile EEG fits your study design.
