Motion is one of the most important ways for human to communicate with environments such as other people and objects. A moving process mainly bases on motion commands, which is generated from neurons in motor cortex, transferred by motor nervous system, and finally executed by voluntary muscles. Among them, motor cortex is obviously the essential part while the other two are important as well. There are patients that have clear consciousness but are not able to move their body because of injuries to their peripheral motor nerves or the muscles for execution. In these cases, Brain Computer Interface (BCI) is a possible solution. In BCI, a computer is used to replace the pathway of motion command, and artificial devices such as robotic arms are for the purpose of replacing the injured body parts. Thus BCI system can directly link brain activity to artificial devices and enable paralyzed patients to move as healthy people.

In recent years, many researchers have focused on motion system and presented several methods to control man-made devices by using motor activities. These techniques range from invasive methods such as implanted microelectrode arrays [1] and electrocorticography (ECoG) [2] to non-invasive methods such as electroencephalography (EEG) and magnetoencephalography (MEG). In invasive studies, electrodes are directly placed in certain part of brain or on the surface of brain, thus the recorded data has a very high signal to noise ratio (SNR), which is effective for investigating complicated motion. However, a surgery is needed in invasive technique so that it is risky and unstable for real application. Non-invasive studies are safer and more convenient, but the recorded brain activities are always contaminated by environmental noises and thus making it difficult to be used for motion pattern extraction. Currently, non-invasive studies on continuous motion mainly concentrate on the prediction of motion trajectory and present relatively high prediction performances [3], [4]. However, the characteristics of motor activities during this procedure remain unclear. Thus current predictions in non-invasive studies are ineffective, which is represented by large feature number and large training data set.

In this study, we developed a noise reduction method which can improve SNR of MEG single-trial data and applied it on motion prediction. Then, we used preprocessed data to investigate temporal, spectral and spatial characteristics of motion related activities. Finally, we confirmed that selected subject-dependent frequency features are really generated by motor and sensorimotor cortices and thus are effective motion-related features.

Chapter 1 introduces the background of motion researches, which includes motor system and current status of BCI. Then we talk about our measurement device (MEG), noise conditions, and data analysis method adopted in motion feature selection and prediction. Moreover, we explain the significance of our study and briefly review the whole thesis.

Chapter 2 offers an effective noise reduction method with almost no brain activity loss, which is vital for the accurate prediction of motion trajectory from single trial data. It should be noted that this method can be applied to all kinds of MEG systems, regardless of having magnetometer or not.

In this chapter, we first describe tSSS method [5] developed by Dr. Samu Taulu. This method is developed for Elekta system with both magnetometer and gradiometer, while the applications to other systems with only gradiometer type sensors are still not fully tested. We implemented tSSS algorithm on Matlab and applied it to our system (Yokogawa PQ2440R). However, signal leakage problem occurred and brain activities reduced to 1/3~1/2 after using tSSS method on our system. We discuss the possible reasons for the signal leakage and provide a solution by discriminating brain activity and interference noise from temporal information. Utilizing this process, interference noise leakage can be further suppressed and brain activity can be recovered from the leakage by signal projection.

In all of computer simulation, phantom simulation, and real brain signals application, compensation tSSS method work well as is shown in Figure 1. Compared to tSSS method, our compensation method preserves signals well and has very small reconstruction error. This suggests that the compensation tSSS is a valuable noise reduction method for single trial analysis in our gradiometer only system.

Chapter 3 presents a study on 1-D continuous motion using a tool bar. This study offers a successful feature selection method for arm motion trajectory prediction and confirms the efficiency of compensation tSSS method in continuous motion prediction.

In this study, we applied a non-invasive MEG study on 4 BCI-untrained subjects' continuous motion. In the task, BCI-untrained subjects were asked to perform continuous motions using toolbar and both subjects' brain activity and motion position were recorded simultaneously. We calculated the spectrum of brain activity below 100 Hz and investigated the correlation between brain activity spectrum and motion parameters. From the correlation results shown in Figure 2, we considered 9-14 Hz which has a relatively high correlation value as motion related frequency feature. Then, we adopted different channel selection models and time-windows, and determined proper features using prediction performance evaluated by multivariate linear regression. By using these motion related features, we offered an effective motion trajectory prediction with an acceptable correlation coefficients (average value across all subjects is 0.32, p < 0.001) on the single trial data preprocessed by compensation tSSS method. This indicates that using well-selected motion related feature, non-invasive methods can also achieve accurate prediction on continuous untrained limb motion as invasive methods. Moreover, compensation tSSS preprocessed data provides a significant higher prediction performance than original tSSS preprocessed data which specifies that compensation tSSS provides a higher SNR and performs better than original tSSS method on our MEG system.

Chapter 4 reconsiders the experiment in Chapter 3 and investigates correlation between motion trajectory and each frequency below 100 Hz for each subject. The difference between subjects is compared and subject-dependent frequency bands are selected for single trial prediction.

Table 1. Comparison of prediction performance results using tSSS and compensation tSSS method, fixed model, main subject-dependant model and combined model. The result is shown as mean ± standard deviation (SD) of all subjects.

In this study, we concentrated on spectral amplitudes of MEG and tested several ways of frequency band selection to further improve the prediction. From the correlation between spectral amplitude and motion trajectory, we extracted several subject-dependent frequency bands, which range from μ (8-16Hz) and β rhythm (18-24Hz) to low frequency δ rhythm (5-7Hz) and some part of high frequency γ bands (30-50Hz, 60-70Hz). Compared to fixed frequency band (9-14 Hz) mentioned in Chapter 3, subject-dependent frequency band offers a better prediction performance, as is listed in Table 1. In addition, the combination of two or three subject-dependent frequency bands further improves the prediction performance and this improvement is significant for most of subjects. From the prediction performance of all subjects, we concluded that using correlation based feature selection method, single-trial MEG data can also predict continuous motion well (r=0.47) with few features (less than 100).

Chapter 5 focuses on the robustness of different motion cycles and devices, and reveals the spatial patterns of motion related frequency features. The source level analysis provides the evidence that such selected frequency features are motion related activities come from motor cortex and sensorimotor cortex.

To test the robustness of our feature selection method, we performed similar motion using a different device (trackball) in task 1. The prediction result confirms that our feature selection method works equally well on different devices which indicates a robustness of different devices. In task 2, we considered a different motion cycle without visual guidance and confirmed the efficiency of our feature selection method on different motion cycle, which shows a robustness of different motion. As there is no visual guidance, the selected features are verified to be from motion brain activities. From further contour map and source estimation studies, it is also confirmed that the sources of frequency features selected by our method are really located in the contralateral motor cortex and sensorimotor cortex.

Chapter 6 summarizes all the studies in this paper and discusses the future research directions.

In this paper, we firstly discussed the interference noise condition and developed compensation tSSS method to effectively suppress the noise in single trial MEG data. Then we investigated on continuous motion and designed a motion related feature selection method which greatly reduces feature number in motion trajectory prediction. By combining several subject-dependent frequency bands, a successful prediction is provided with single trial MEG data. Further contour map and source analysis confirm these features come from motor cortex and thus are motion related features.

Our study reveals detailed characteristics of motion related activities which are consistent to ECoG and EEG studies. It also provides a guidance to select features and achieves a successful single trial motion prediction. The high quality prediction demonstrates that non-invasive measurement can predict motion comparably well as invasive measurement such as ECoG. Also, the prediction of arm movement trajectory in our study provides a possibility of controlling external prosthetic devices.

本論文は全6章よりなり、tSSS(temporal Signal Space Separation)というMEGデータのノイズ除去ソフトの改善を行った上で、MEGデータを用いて、BCI(Brain computer Interaction)において重要な連続的な運動の予測法の開発を行った。

第1章では,運動システムを含むBCI研究の現状を述べ、MEG計測装置とそのノイズ状態,また動き特徴選択とその予測に関するデータ解析法について記した.また研究の意義と論文全体の概要を述べた.

第2章では、脳活動の情報損失がほぼ発生しない高精度のノイズ除去法(compensation tSSS)を提案した.これは単一試行のデータから,手の動きを高精度で予測する上で非常に重要である.Tauluらによって開発された方法は、マグネトメータ及びグラディオメータを共に備えるElekta製のMEGシステムを主眼に開発され,グラディオメータのみのシステムにおいては、有効性は示されていない.Tauluらの方法の検討の結果,脳活動が三分の一から二分の一まで減衰することが明らかになったので,時間情報を用いて脳活動とノイズを認識して分離する手法を提案し,ノイズ抑制を実現し脳活動の損失を回復する手法を提案した.

第3章では,棒を用いた一次元運動の予測に関する研究を示した.四人の非熟練BCI被験者からMEGを用いて取得したデータ使用した.被験者が棒を左右に運動させている時の脳活動と手の位置の同時計測を行った.運動予測に効果的な周波数帯域を探索するために,100Hz以下の帯域で脳活動のスペクトルを求め,このスペクトルと実際の運動の相関関係を調べた.ここで得られた最適な特徴集合を用いた動きの予測を行い,提案手法によるノイズ除去の影響と,予測の効率性について確かめた.提案手法によるノイズの前処理を行った単一試行データを用い、実際の手の動きと脳活動から予測された手の動きの相関を求めたところ、平均が0.32 (p < 0.001)となり,選択された特徴の有効性と,ノイズ除去に用いた提案手法の効率性が統計的に示された.

第4章では,第3章の解析をさらに改良するため,被験者に依存した周波数帯域を用いることでより高い精度の運動予測を実現できることを示した.第3章の結果と比べて予測性能の向上が見られたことから、予測性能における最適周波数帯が被験者に依存することが示された.さらに,被験者ごとに選択する周波数帯を複数にすること(combined model)により予測性能が顕著に改善することも確認された.全被験者における予測性能から,相関に基づく特徴選択を用いることで,単一試行MEGデータからでも,100に満たない少ない特徴数で優れた動きの予測(r=0.47)が可能であること結論付けた.

第5章では,我々が提案した特徴選択法の頑健さを確かめるために,異なる運動サイクルや入力デバイス(トラックボール)用いて実験を行った.さらに,活動源解析の結果から,我々の特徴選択法で選択される周波数帯が,運動野及び体性感覚野の動き関連の活動から生じることを確認した.本実験では,運動軌道の予測が視覚関連反応に基づく可能性を排除するため,視覚刺激の提示を行わず,セルフペースで運動を行うよう被験者に教示した.このような場合でも、有効に本手法が適用可能であることが示された.

第6章では,一連の研究を総括し,将来の展望について議論した.本研究は、ECoGやEEGを用いた先行研究と一貫する腕の動き予測に関連する脳活動の特徴を明らかにした.その特徴の選択法を提供し,単一試行データからの運動軌道予測を実現した。本手法を用いることで,非侵襲計測でも侵襲計測を用いたものに劣らない高精度の運動軌道予測が実現できることを示した.

なお、本論文第3章は、武田常広との共同研究であるが、論文提出者が主体となって分析及び検証を行ったもので、論文提出者の寄与が十分であると判断する.したがって、博士(科学)の学位を授与できると認める.