Real time classification between uni- and bimanual motor imagery task for BCI controlled functional electrical stimulation

Vuckovic, A. and Pangaro, S. (2018) Real time classification between uni- and bimanual motor imagery task for BCI controlled functional electrical stimulation. 7th International BCI Meeting, Pacific Grove, CA, 21-25 May 2018.

Full text not currently available from Enlighten.


Introduction Most Brain Computer Interface (BCI) systems based on EEG, used for rehabilitation, are designed to classify motor imagery (MI) of the left and right hand or between one hand and the resting state. However, lots of activities or daily living require combination of uni and bimanual tasks. Therapies based on bimanual practice after stroke have positive therapeutic effect [1]. Bimanual control of BCI operated orthoses would increase their usability. A challenging aspect in creating a uni vs bimanual classifier is that bimanual movements do not necessarily result in a distinctive pattern of neural activity [2]. Materials, Methods and Results Ten adult, right handed able bodied volunteers took part in the study. On the first day they participated in a cue-based off-line imagery task. Those whose offclassification accuracy exceeded 70% took part in a subsequent on-line BCI task about a week later. Offline session: Participants sit comfortably 1 m from a computer screen. At t=0s a warning sign appeared on the computer screen and stayed there until t=4s. At t=1s a warning sign was overlaid by an execution cue ( for MI of the right hand, for the left hand and for both hands) which remained on the screen for 1.25s. Depending on a cue, participants imagined waving with one or both hands from t=1s till t=4s. Cues appeared in a semi random order. One hundred trials of each of three types of MI were performed, divided into 10 shorter sub-sessions. EEG was recorded with linked ear reference using usbamp (Guger technologies, Austria) with 31 electrodes placed over the sensory-motor cortex and one for EOG. Noise, primarily EOG, was removed by calculating independent components and removing noisy components before returing back to EEG domain. Following this, event-related spectral perturbation was calculated in EEGlab [3] for each task and each participant separately to determine one or two frequency bands with strongest event related desynchronization , in pre-defined frequency bands: 8-12, 16-24, 16-30 and 30-35 Hz. Common spatial patterns (CSP), ranging from 2 to 32, were calculated for either full band 8-40 Hz or for CSP in selected bands (CSPb) for the left and right, left and both hands and right and both hands. Linear discriminant analysis classifier (leave-one out procedure) was implemented in Biosig [4]. The average classification accuracy for CSP and CSPb respectively was 74%±9%, and 74±10% for left vs right hand; 69±8% and 73±7% for right vs both hands; 70 ±8% and 71±7% for left vs both hands. For CSP only 3 participants achieved a classification accuracy higher than 70% as opposed to 6 participants for CSPb, which was used for on-line classification. On-line sessions: The experimental procedure was similar to the off-line but a bar proportional to the on-line accuracy overplayed the execution cues on the computer screen. Unmodified classifier from the off-line session was used. Functional electrical stimulation (frequency 33 Hz, pulse duration 250 s, amplitude 6-11 mA, duration 2s) was delivered through bipolar pair of electrodes to right or to the left and right hand extensor muscles with intensity sufficient to produce visible muscle contraction. In one sub-session participants were asked to perform left and right MI (40 trials per condition) and in the other sub-session on the day same day they performed right or bimanual MI (40 trials per condition).The order of sub-sessions varied between participants to counterbalance fatigue. The average accuracy was 69±3% and 66±3% for the left vs right and right vs both MI respectively, which was above chance level of 60% [5]. Discussion: This study shows that using CSPb it is possible to classify between uni and bimanual task above chance level. The same classification parameters were used in off and on-line sessions organised a week apart. Significance: Novel rehabilitation protocols and more natural assistive BCI can be created by combining un and bimanual tasks.

Item Type:Conference or Workshop Item
Glasgow Author(s) Enlighten ID:Pangaro, Miss Sara and Vuckovic, Dr Aleksandra
Authors: Vuckovic, A., and Pangaro, S.
College/School:College of Science and Engineering > School of Engineering
College of Science and Engineering > School of Engineering > Biomedical Engineering
Related URLs:

University Staff: Request a correction | Enlighten Editors: Update this record