영장류에서 뇌경막외피질뇌파(electrocorticogram) 신호를 이용한 양 손의 움직임 예측 연구

Abstract

전 세계적으로 뇌-컴퓨터 인터페이스(BCI : Brain-computer interface) 분야가 활발히 연구되면서 사람의 의도 파악이나 운동 기능에 대한 연구가 가속화되고 있다. 뇌피질뇌파는 대뇌 피질 위에 전극을 삽입하여 측정하는 방법으로 뇌전도보다 뇌파의 근원지에 가까워 신호가 선명하고 뇌 속으로 전극을 삽입하지 않아 뇌의 손상과 감염의 우려가 적어 BCI에서 활발한 연구가 진행 중이다. 예전부터 움직임에 대한 연구가 많이 이루어져 왔으며 이 중 팔 움직임에 대한 연구는 가장 많이 연구되고 있는 분야이다. 본 연구에서는 뇌피질뇌파 전극을 삽입한 영장류의 뇌파 분석을 통해 양손의 움직임을 예측하는 것을 목표로 하였다. 히말라야 원숭이에게 32채널의 패치 전극 2개를 각각 좌, 우반구의 운동영역을 덮을 수 있도록 부착하였다. 패치 삽입 후, 원숭이에게 한 손과 양손을 움직이는 훈련을 시켜 각 움직임에 대한 뇌피질뇌파를 측정하였다. 파란색은 왼손, 빨간색은 오른손을 움직이는 task로, 원숭이가 준비버튼을 누르면 trial이 시작되고 일정 시간 후 움직임에 해당하는 신호 버튼이 켜지면 해당 손을 움직여 버튼을 누르면서 trial이 끝난다. 실험의 분석은 MATLAB을 이용하였다. 전 처리 필터링은 0.3-200Hz 대역 통과 필터를 사용하였고 움직임 시작 신호의 인지 과정이 0.5초 정도 걸림을 고려하여 움직임 1초 전의 뇌파를 사용하여 해당 움직임에 대한 특징을 웨이블릿 변환을 이용하여 추출하였다. 추출한 특징과 움직임을 부분 최소 자승 회귀분석을 이용하여 뇌파와 움직임 간의 모델을 만들었다. 실험 결과 한 손의 움직임을 예측할 경우 실제 움직임과 예측치가 0.6 정도의 상관계수를 보였으며, 양손의 움직임을 예측할 경우 0.3 정도의 상관계수를 보였다. 또한, 한 쪽 반구의 패치만 가지고 양손의 움직임을 예측할 수 있었으나, 양쪽의 패치를 모두 사용하는 편이 성능이 더 좋았다. 움직임 간 모델의 계수 기여도를 비교하여 분석한 결과, 한 손을 움직일 때와 양손을 움직일 때는 M1(primary motor cortex), S1(primary somatosensory cortex), SMA(supplementary motor cortex), PMA(premotor cortex) 영역의 기여도가 달라지는 것을 확인하였고, 이 영역들이 한 손과 양손이 움직일 때 뇌의 상태가 차이 나는 영역일 가능성이 크다고 보여진다.| Recently, the interest for BCI(Brain Computer Interface) technology has been increased in research to understand human intent or to alternate motor function of motor disabilities. BCI can use non-invasive or invasive methods. Non-invasive BCI can use EEG(electroencephalogram) and it is widely used because of its safety and convenience. However, the spatial resolution of invasive methods, such as ECoG(electrocorticogram), is known to be much higher than EEG[15], and it provides more precise control and many degrees of freedom than EEG signals. In this study, we aimed movement prediction of both hands for BCI applications of ECoG activity recorded in primates. Two patches of 32 channel electrodes were implanted at each hemisphere of the rhesus monkey, covering from SMA(supplementary motor cortex) to the PPC(posterior parietal cortex. After implantation, the monkey was trained for unimanual and bimanual button task. The task started when the monkey pushed ready button. After a few seconds, the cue button was light on. For the left hand movements, the blue light was on, and for the right hand movements the red light was on. If the monkey pushed a cue button on time, the trial was finished. We used MATLAB for off-line analysis. A sampling frequency of ECoG Data was 1kHz and the data were bandpass filtered 0.3-200 Hz and notch filtered 60,120,and 180Hz. A sampling frequency of Motion Data was 20Hz. Then, the wavelet transform was performed for 1s ECoG data before movement to extract features of the movement. We made movement prediction model using extracted feature and observed motion data with PLS(partial least square) regression. To calculate prediction accuracy, we use correlation coefficient. As a result, the correlation coefficient of the unimanual task was 0.6 and the bimanual task was 0.3. Also, we could predict both hand movements using only one patch, but the performance using both patches was much better than using one patch. By analyzing PLS regression coefficients, we found the contribution difference in M1(primary motor cortex), S1(primary somatosensory cortex), SMA(supplementary motor cortex), and PMA(premotor cortex) area, and these areas would be likely to indicate change of brain status between bimanual and unimanual movements.; Recently, the interest for BCI(Brain Computer Interface) technology has been increased in research to understand human intent or to alternate motor function of motor disabilities. BCI can use non-invasive or invasive methods. Non-invasive BCI can use EEG(electroencephalogram) and it is widely used because of its safety and convenience. However, the spatial resolution of invasive methods, such as ECoG(electrocorticogram), is known to be much higher than EEG[15], and it provides more precise control and many degrees of freedom than EEG signals. In this study, we aimed movement prediction of both hands for BCI applications of ECoG activity recorded in primates. Two patches of 32 channel electrodes were implanted at each hemisphere of the rhesus monkey, covering from SMA(supplementary motor cortex) to the PPC(posterior parietal cortex. After implantation, the monkey was trained for unimanual and bimanual button task. The task started when the monkey pushed ready button. After a few seconds, the cue button was light on. For the left hand movements, the blue light was on, and for the right hand movements the red light was on. If the monkey pushed a cue button on time, the trial was finished. We used MATLAB for off-line analysis. A sampling frequency of ECoG Data was 1kHz and the data were bandpass filtered 0.3-200 Hz and notch filtered 60,120,and 180Hz. A sampling frequency of Motion Data was 20Hz. Then, the wavelet transform was performed for 1s ECoG data before movement to extract features of the movement. We made movement prediction model using extracted feature and observed motion data with PLS(partial least square) regression. To calculate prediction accuracy, we use correlation coefficient. As a result, the correlation coefficient of the unimanual task was 0.6 and the bimanual task was 0.3. Also, we could predict both hand movements using only one patch, but the performance using both patches was much better than using one patch. By analyzing PLS regression coefficients, we found the contribution difference in M1(primary motor cortex), S1(primary somatosensory cortex), SMA(supplementary motor cortex), and PMA(premotor cortex) area, and these areas would be likely to indicate change of brain status between bimanual and unimanual movements.