auditory evoked potentials (AEP) and visually evoked potentials (VEP). To find out the characteristics of these signals, we compared the frequency-based and channel-based pattern of pure motor imagery and EEG contamination tasks of healthy and stroke groups of subjects. Using the discovered characteristics of EEG contaminations, we designed a three-phase algorithm using a single region of interest (ROI) channel to obtain a motor imagery signal, and three non-ROI channels on the source of the EEG contamination signals to filter out the false positives. Four healthy subjects and nine stroke patient subjects participated in the online BCI trial. The algorithm has shown a sensitivity of 35% and 34.81%, and selectivity of 73.68% and 71.76% for each group. The false-positive rate were 10% for the healthy group, 13.7% for the stroke group. Compared to other MI-BCI studies applied to stroke patients, our study has outperformed all studies with false positive rate and false per minute, while using the least number of channels (4). Also, our system was robust to day-by-day difference for prolonged intervals (28.5 days for the healthy group, 11.78 days for the stroke group). These results suggest that our proposed system could be practically used for the clinical environment. For future work, we need to increase the sensitivity while maintaining the advantages of the system. Evaluating the difference of brain plasticity between intensity focused BCI system and specificity focused BCI system would guide our further studies. The follow-up evaluation study of long term BCI therapy applied to stroke patients might follow.; Many studies use motor imagery-based brain-computer interface systems on stroke rehabilitation to induce brain plasticity. However, many systems only focused on detecting true-positives but ignored the false positive. False-positive can be a threat to stroke patients since the false positive can induce wrong-directed brain plasticity results in wrong directed rehabilitation. Therefore, we propose a motor imagery detection algorithm focusing on rejecting false positives. We categorized the cause of false positives into two different EEG contaminations
Table Of Contents
ABSTRACT - 1 - List of contents - 2 - I. Introduction - 3 - II. Method - 8 - 1. Definition and selection of the non-ROI - 8 - A. Participants - 9 - B. Concept of non-ROI - 10 - C. Apparatus and procedure of calibration - 11 - D. Selecting ROI and non-ROI candidate channels - 15 - E. Characteristics of ROI and non-ROI - 19 - 2. Algorithm of proposing MI-BCI system - 22 - A. Preprocessing of the signal and moving window - 22 - B. The three-phase algorithm - 23 - C. Extraction of the training data - 27 - D. Offline simulation for specific channel selection - 30 - 3. Structure of proposing MI-BCI system - 31 - 4. Experiment - 33 - A. Calibration session - 33 - B. Interval of sessions - 33 - C. MI-BCI session - 34 - 5. Data analysis - 36 - A. Group analysis of paradigms during the calibration session - 37 - B. Offline simulation during the calibration session - 37 - C. MI-BCI session - 38 - III. Results - 40 - 1. Group analysis of paradigms during the calibration session - 40 - 2. Offline simulation during the calibration session - 42 - 3. Results of MI-BCI session - 44 - IV. Discussions - 50 - V. Conclusion - 59 - References - 60 -