Participants
Patients with hemiplegia after early subacute (from 1 week to 5 weeks) ischemic stroke were prospectively enrolled. Patients with cognitive impairment (Mini-Mental State Examination score < 16), delirium, depression, anxiety, or other uncontrolled medical diseases were excluded. Initially, 20 patients consented to participate and were enrolled in the study. However, three patients dropped out due to discomfort in maintaining posture and wearing the EEG, one patient was excluded due to an inability to follow instructions caused by decreased attention, and one patient was excluded due to poor EEG quality. Consequently, a total of 15 patients were included in the final analysis. This study was approved by the Institutional Review Board of Keimyung University Dongsan Hospital (IRB number 2021-04-112-003). All participants signed the written consent form.
Study design
This randomized cross-over study consisted of two experiments: neurofeedback and sham (Fig. 1). Each experiment consisted of four blocks (three blocks of resting, grasp, resting, and intervention and one block of resting and grasp). In the resting session, a white fixation cross with a black background appeared and was fixed for 2 minutes without upper extremity movement. In the grasp session, the same white fixation cross appeared, and the participants were instructed to grasp and release their weak hand at a frequency of about 1 Hz for 3 minutes. Subsequently, another 2-minute resting session followed. Then, the participants spent 3 minutes imagining movements in the neurofeedback session, during which a boxing image was generated based on mu suppression detected from the EEG. However, during the sham condition, an EEG recorded from another participant's neurofeedback session was used to trigger the appearance of the boxing image on the screen instead of the participant's real EEG. The first patient underwent neurofeedback first, with the sham condition using their own EEG. For subsequent patients, the sham EEG was randomly selected from the neurofeedback sessions of the previous patients. The two experiments were conducted in a randomized crossover design with a washout period of at least 3 days.
Neurofeedback mechanism
The neurofeedback system is based on the characteristics of the asymmetry of mu suppression in the motor cortex. The EEG data from the C3 and C4 channels were filtered using a bandpass filter (a 5th-order Butterworth) that passed signals between 8 and 12 Hz. The EEG signal was then segmented with an epoch size of 1 second, overlapping every 1/16 second to calculate the mu band power.
First, we used the mean and standard deviation of the mu power of the C3 and C4 brainwaves acquired for a 1-minute resting session to normalize the data such that the incoming mean mu power is 0 when it is at the mean value and + 1 or -1 at a point that is 2 standard deviations away from the mean. The output index can, therefore, move along a line that outputs + 1 when the incoming mu power is at the mean minus 2 standard deviations, 0 when it is at the mean, and − 1 when it is at the mean plus 2 standard deviations; thus, the index reflects higher values when the mu power is lower and lower values when it is higher.
After obtaining normalized levels for both the ipsilesional and contralesional sides, contrast scores were calculated for each motor imagery training session. For paretic upper limb motor imagery training, the contrast score was obtained by subtracting the contralesional index from the ipsilesional index. This score represented the relative strength of ipsilesional mu suppression compared with contralesional mu suppression. When the contrast score exceeded 0, it accumulated, and once the accumulated score reached 2, an arbitrarily determined threshold, a punch to the target in the boxing game was triggered.
EEG acquisition
EEG data were collected from 19 dry electrodes using DSI-24 (Wearable Sensing, San Diego, CA, USA). The electrodes were arranged based on the International 10–20 system, and the data were sampled at 300Hz frequency.
The recorded EEG was band pass filtered with a 4–30 Hz and rejected artifacts. Then, it was segmented with a 2-second epoch. The epochs with an average amplitude exceeding 150 uV were excluded. Baseline correction and the linear detrend were performed for each epoch. The mu band power was extracted using power spectrum density analysis, and the values of the mu band power were averaged. The mu band power for each grasp session was compared with that of resting session. The degree of mu suppression was calculated using the log ratio. In the EEG topography analysis, the lesion side was standardized to the left hemisphere. Consequently, the EEG data from patients with left-sided lesions were used without modification, whereas the EEG data from patients with right-sided lesions were horizontally flipped.
Statistical analysis
Mu suppression in the motor and parietal cortices was used for analysis. Repeated-measures analysis of variance (ANOVA) for intervention (neurofeedback vs sham) x time (from task 1 to task 3 or from grasp 1 to grasp 4) was used to identify significant factors associated with mu suppression. If there was a statistical significance for intervention or time, post-hoc analysis to compare the effect between neurofeedback and sham or between time points was conducted using Bonferroni-corrected p-values. Statistical significance was set at p < 0.05. All analyses were performed using R 4.4.0 (R Foundation for Statistical Computing, Vienna, Austria).