Tremor Frequency Based Filter to Extract Voluntary Movement of Patients with Essential Tremor

Transcription

1 The Fourth IEEE RAS/EMBS International Conference on Biomedical Robotics and Biomechatronics Roma, Italy. June 24-27, 2012 Tremor Frequency Based Filter to Extract Voluntary Movement of Patients with Essential Tremor Yuya Matsumoto, Masatoshi Seki, Takeshi Ando, Member, IEEE, Yo Kobayashi, Member, IEEE Hiroshi Iijima, Masanori Nagaoka, Masakatsu G. Fujie, Fellow, IEEE Abstract Essential Tremor (ET) refers to involuntary oscillations of a part of the body. ET patients face serious difficulties in performing daily living activities. Our motivation is to develop a system that can enable ET patients to perform their daily living activities; hence we have been developing a myoelectric controlled exoskeletal robot for ET patients. However, the EMG signal of ET patients contains not only voluntary movement signals but also tremor signals. Accordingly, to control this robot correctly, tremor signals must be removed from the EMG signal of ET patients. To date, we have been developing a filter to remove tremor signals, which has been largely effective in this. However, tremor signals are generated both while voluntary movement is being performed and while a posture is being maintained, and the filter ended up attenuating both these signals. But, to control this robot accurately, the signal generated during performance of voluntary movement is expected not to be attenuated. Therefore, in this paper, we propose a method that attenuates only tremor signals arising during maintenance of a posture. To accomplish this objective, we focus on the frequency of tremor signals. From the experiment, we confirmed the characteristic that the frequency of tremor signals changed depending on the state of the patient s movement. We then used frequency as a switch to activate the previously proposed filter by setting a threshold. As an evaluation, signals processed by the proposed method were input to a time delay neural network. The proposed method succeeded in partly improving recognition due to reduction of attenuation during performance of voluntary movement. However, the proposed method failed recognition in cases where the frequency of tremor signals varied widely. As a future work we will review the method to calculate the frequency of tremor signals and improve recognition. Manuscript received January 31, This work was supported in part by the Global COE (Center of Excellence) Program Global Robot Academia, Waseda University, Tokyo, Japan and The Minister of Economy, Trade and industry R & D Project to Create Regional Innovation, Japan. Yuya Matsumoto is with the Graduate School of Creative Science and Engineering, Waseda University, Tokyo , Japan, ( yy-matsumoto@ruri.waseda.jp). Masatoshi Seki is with the Graduate School of Advanced Science and Engineering, Waseda University, Japan. Takeshi Ando is with the Robotics Development Center Cooperate Manufacturing Innovation Division, Panasonic Cooperation, Osaka, Japan, and the Faculty of Science and Engineering, Waseda University, Japan. Yo Kobayashi is with the Faculty of Science and Engineering, Waseda University, Japan. Hiroshi Iijima is with the Rehabilitation Engineering Dept., Yokohama Rehabilitation Foundation, Japan. Masanori Nagaoka is with the Department of Rehabilitation Medicine, Juntendo University Graduate School, Japan. Masakatsu G. Fujie is with the Faculty of Science and Engineering, Waseda University, Japan. A. Essential Tremor (ET) A I. INTRODUCTION tremor is an involuntary oscillation of part of the body; it is the most common of all involuntary movements [1]. The most accepted definition of tremor is an involuntary, approximately rhythmic and roughly sinusoidal movement. Tremors are classified in terms of symptoms. Tremors that occur while the body is at rest are called Rest Tremors. The most famous disorder involving rest tremors is Parkinson s disease. Tremors that occur while performing an action or maintaining a posture are called Action/Posture Tremors. Essential Tremor (ET) is the most widely known disorder involving an Action/Posture tremor, and elderly people are often ET patients including some 4 million in Japan (about 5% of the total population). While ET is not a life-threatening disorder, it may result in functional disability and social inconvenience. In most cases, the extent of the tremors for the ET patients is considerable. About 65% of ET patients have serious difficulties in performing daily living activities (e.g. eating, drinking, writing etc.) [2], [3]. B. Tremor Suppression Method Currently, approaches to suppressing tremor can be divided into two types. One is through medication to suppress the over-reaction of nerves, while the other comprises Deep Brain Stimulation (DBS). However, both these approaches have their respective problems in terms of medical side effects and invasiveness involved in implanting electrodes in the brain. Instead, some researchers use functional electric stimulation (FES) to suppress tremors [4] [6]. However, FES creates a feeling of overstimulation and causes muscle fatigue. Fig. 1 Exoskeletal robot for ET patients 1 (s) Fig. 2 EMG of ET patient [20] /12/$ IEEE 1415

2 As an alternative approach, we have been developing an exoskeletal wearable robot for the upper limbs to support voluntary movement among ET patients (Fig. 1). The exoskeletal robot suppresses oscillations and facilitates the patient s voluntary movement by the following two steps: 1. The exoskeletal robot constrains the patient s elbow joint mechanically and suppresses oscillations by restricting the patient s motion. 2. The exoskeletal robot estimates the patient s voluntary motion which is restricted by the constraint and follows the estimated movement. This paper especially focuses on the second step. To enable the patient s voluntary movement with mechanical constraint, the robot needs to estimate the patient s movement accurately. As a related study, Rocon et al. have developed a robotic exoskeleton for tremor patients [7]. Here, the angular velocity obtained by gyro sensors, which are implemented in each joint, is used as input data to control the exoskeleton. Consequently, this robot motion inevitably lags behind the user s motion due to inability to attain angular velocity ahead of the user s motion. Therefore, information obtainable before the start of movement should be used as the input signal to control the exoskeletal robot. In this study, electromyography (EMG) is used as the input signal to control the exoskeletal robot. C. EMG Signal of ET Patients Many researchers use surface EMG as the input signal to control exoskeletal robots [8] - [10]. In these studies, the targeted users of the robot are elderly persons with weakened muscle power, upper extremity amputees, or caregivers and nurses. The EMG signal of these users contains voluntary movement signals and high frequency noise. However, the EMG signal of ET patients contains not only voluntary movement signals and high frequency noise but also tremor signals. It was confirmed that tremor signals overlap with the EMG signals of ET patients within the range of 4 10Hz (Fig. 2) [11]. Tremor signals are generated involuntarily and cause an estimation error. To control the exoskeletal robot accurately through the EMG of ET patients, tremor signals must be removed from the EMG signal. Some approaches for removing tremor signals were proposed by past researchers. Yano et al. developed an adaptive filter for force sensor data which was used in admittance control for the meal-assist manipulator [12], [13]. This filter estimated the tremor frequency and attenuated the signal of the estimated frequency band by using a band stop filter. Riviere et al. have proposed a filtering algorithm for physiological tremor arising during micro surgery [14]. The above studies purposed on reduction of tremor noise in motion signals measured by a force sensor, position meter, etc. In these signals, effects of tremor are observed as an additive noise. Therefore, once the noise frequency has been detected, it is easy to design a real-time processing system. On the other hand, it can be anticipated that tremor effect in EMG signals is a multiplicative noise [15] because rhythmic muscle contractions and relaxations do not cause EMG offset ups and downs but cause amplitude fluctuation. For the reduction of multiplicative noise, the Cepstral Mean Normalization [16] (CMN) and maximum a posteriori estimation CMN (MAP-CMN) are widely recognized. CMN is not a real-time adapted method. MAP-CMN needs Cepstral Mean calculations from a signal of sufficient length to provide superior performance. D. Objectives To date, we have been developing a real-time tremor-removing filter [17] [19]. This filter was designed based on the hypothesis that rectified tremor signals are able to be approximated by a powered sine wave; that is, high correlation coefficient between both signals leads to high attenuation of tremor signals while low correlation coefficient leads to low attenuation (The details of this filter are provided in Section III C). This filter had a large effect on removing tremor signals from the EMG signals. However, tremor signals are generated both while a voluntary movement is being performed and while a posture is being maintained. Accordingly, this filter ended up attenuating not only tremor signals generated during maintenance of a posture but also those generated during performance of voluntary movement, and caused estimation error in the voluntary movement. To reduce estimation error, the attenuation of signals generated during performance of voluntary movement must be diminished. Consequently, the objectives of this paper are to propose a method to diminish the attenuation of signals generated during performance of voluntary movement. To accomplish this objective, we use the frequency of tremor signals to switch on/off the use of the powered sine filter. Through the experiment described in Section III, we confirm that the frequency of tremor signals is a parameter whose value during performance of voluntary movement is different from that during maintenance of a posture. That is, by setting a threshold for the frequency of tremor signals, the state of the patient s movement can be distinguished from each other. Therefore, we proposed a method in which the powered sine wave is activated only when the frequency of tremor signals is over the threshold. This paper is structured as follows: Section II describes the experiment to confirm the hypothesis that the frequency of tremor signals changes its value depending on the state of the patient s movement. Section III describes the experiment to evaluate the accuracy of the voluntary movement estimation by inputting signal processed by the proposed method to a 1416

3 time delay neural network (TDNN), and Section IV provides the conclusion of this paper and discusses future works. II. EXPERIMENT TO CALCULATE THE FREQUENCY OF TREMOR SIGNALS A. Objective As described in Section I, we use the frequency of tremor signals to switch on/off the use of the powered sine filter. Here, in the medical field, the frequency in each patient is said to be almost constant [21], [22]. However, from visual comparison, it seems that the frequency of tremor signals changes depending on the state of the patient s movement. Most of the earlier studies on the frequency of tremor signals reported only on the frequency during maintenance of a posture because ET shows especially strong oscillations while a posture is being maintained. Also, few studies reported the frequency measured in the experiment whose task contains both performance of voluntary movement and maintenance of a posture. Therefore, we hypothesized that the frequency of tremor signals changed depending on the state of the patient s movement. In this experiment, we confirm the hypothesis. B. Subject The subject of this experiment was one ET patient (male, 71 years old) who had tremor symptoms, especially in forearm rotation. Tremor signals are measured from the biceps brachii. We gave the subject a detailed account of our experimental objectives, made it clear that he was entitled to stop the experiment whenever he desired, and obtained his consent. This experiment was approved by the Institutional Review Board in Waseda University. C. Methodology The subject performed an elbow flexional movement while holding a bottle filled with water (weight: 550g), and also while holding the empty bottle (weight: less than 10g) to simulate the movement of drinking water. He practiced the elbow flexional movement in advance of the experiment to get used to the experimental movement. At this point, the target movement time was fixed at 1.25s throughout the practice, and this pace was maintained using a metronome. The subject performed the experimental movement 10 times with each bottle (Total 20 times). We set a short rest (about 3s) before the start of each motion. The EMG signals were obtained through surface electrodes (Biometrics Ltd) and DataLog (Biometrics Ltd) and sampled at the rate of 1000Hz. The electrodes were placed on the biceps brachii and the positions of the electrodes were determined by an occupational therapist. The elbow angle was obtained with a goniometer (Biometrics Ltd.), which was also sampled at a rate of 1000Hz. The voluntary movement of the subject was determined by the elbow angle obtained with the goniometer. D. Auto-Correlation Function (ACF) The frequency of tremor signals was calculated with the Auto-Correlation Function (ACF), which is often used to calculate the frequency of a function with a periodic nature. ACF is a function used to calculate the correlation coefficient between a function f(t) and a function f(t+τ), that is shifted by τ (sec). ACF is calculated as (9): 1 T R ff ( ) lim f ( t) f ( t ) dt (9) T T 0 where R ff (τ) refers to the correlation coefficient. T refers to the window length of input signals and, in this study, T is fixed to 250ms to cover the tremor frequency band of 4 10Hz. The input data was the patient s rectified EMG signal processed by a low pass filter (LPF) with a cutoff frequency of 10Hz and damping coefficient of 0.7. E. Result and Discussion Fig. 3 shows one of the results of calculating the frequency of tremor signals. Fig. 3(a) and (b) shows the result of the experiment while holding the water-filled bottle and while holding the empty bottle respectively. As a graph, the frequency during maintenance of a posture in flexed position is longer than that during flexion. This characteristic was observed in all 20 movements. To quantitatively compare the difference of frequency between each state, an average of the frequency of the EMG signal in each state was calculated, and Fig. 4 shows the result. From the graph, we also confirmed quantitatively that the frequency of tremor signals during maintenance of a posture in the flexed position is longer than that during flexion. Therefore, it is highly likely that the frequency of tremor signals is a parameter whose value is changed depending on the state of the patient s movement. In accordance with this characteristic, we tried to determine the threshold to switch on/off the use of the powered sine filter. From Fig. 4, the average of the frequency during performance of voluntary movement was about ms and that during maintenance of a posture in flexed position was about ms. Therefore, by setting the threshold between each average of the frequency, it seems that we can activate the powered sine filter only when the patient maintains a posture in flexed position. Fig. 5 shows one of the results of each processing, Fig. 5 (a) and (b) shows the signal processed by the previous filter and by the proposed filter, respectively. Certainly, the proposed method succeeded in attenuating only the signal generated during maintenance of a posture in flexed position. In the next section, we evaluate the effect of the proposed method by calculating the recognition rate using a time delay neural network (TDNN). 1417

4 The average length of the period of tremor signals ms.... (a) The experiment with a water-filled bottle (a) Processed by the previous method (b) The experiment with an empty bottle Fig. 3 One of the results of calculating the frequency of tremor signals. (i) indicates the state of maintaining a posture in extended position, (ii) indicates that of performing voluntary movement, and (iii) indicates that of maintaining a posture in flexed position. Experiment using an empty bottle Threshold Experiment using a bottle filled with water (i) (ii) (iii) Fig. 4 The frequency of the EMG signal in each state of the patient s movement. (i), (ii) and (iii) respectively indicate the state of maintaining a posture in extended position, that of performing voluntary movement, and that of maintaining a posture in flexed position. III. EXPERIMENT TO EXTRACT VOLUNTARY MOVEMENT USING THE PROPOSED METHOD A. Objective In this section, we evaluate the recognition accuracy of the powered sine filter with threshold using time delay neural network. (b) Processed by the proposed method (Threshold 150ms) Fig. 5 One of the results of processing the EMG signal with each method. (i), (ii), (iii) indicate each state of the patient s movement just as Fig. 4. Current Input Past Input x(t) x( t 1) x( t 2) x( t n 1) Fig. 6 TDNN structure B. Time Delay Neural Network (TDNN) y(t) Since the recognition of voluntary movement is based on a noisy and complex EMG signal, a highly robust system that is unaffected by potential electrode misalignment, individual differences or the surrounding electrical condition is necessary to recognize EMG signals accurately. A neural network (NN) is a learning machine that uses EMG signals to detect motion [23] [25]. NNs are capable of nonlinear mapping, generalization, and adaptive learning. There are two kinds of NNs that recognize a time-series signal. One of these is the time delay neural network (TDNN) [26]. The TDNN treats time-series signals in terms of not only the input of present signals but also of past signals to the network, as shown in Fig. 6. The other is a recurrent neural network (RNN) [27], [28]. RNN holds past signals by feeding back the output of the output layer or hidden layer into the input layer. 1418

5 Electromyograph [mv] To avoid needless time-stretching property, a 3-layer TDNN was selected as a learning machine. The relationships between each pair of units in the TDNN are shown as follows: 200 Windowed EMG Squared Sin Correlation 1 net m i n m 1 j1 m m x (10) ij m 1 j x i m = f(net i m ) (11) f(net) = 1/ (1+exp(-u 0 net)) (12) where m = 2, 3, i = 1,., n m, n m is the number of the m th layer unit, ω ij m is the weight between the (m-1) th layer s i th unit and the m th layer s j th unit, x i m is the output of the m th layer s i th unit, θ i m is the threshold in the m th layer s i th unit, and u 0 is the constant to decide the gradient of the sigmoid function. The TDNN was able to recognize the subject s voluntary movement. The TDNN outputting ON meant that voluntary movement had been performed; while OFF meant that no voluntary movement had been performed, that is, resting or maintaining a posture. C. Powered Sine Wave Filter [17] [19] As described briefly in Section I, our previously proposed filter was designed based on the hypothesis that rectified tremor signals are able to be approximated by a powered sine wave. The formula for this filter is as (1). s n 1 Ae n, e n 1, e n 2,..., e n N 1 e n (1) where s n is the filtered signal and n is the number of samples. A is the attenuation ratio of the input signal; it was calculated from only present and past surface EMG data e n, e n-1, e n-2,...,e n-n+1. These EMG data are preprocessed by rectification and band-pass filtering (30 300Hz). The BPF is of Chebyshev II type with damping coefficient 0.7. The number of windowed EMG data units N is detected by the tremor frequency F and data sampling period as (2). N 1 Ft (2) Attenuation ratio A is calculated with correlation coefficient C MAX, which is the correlation coefficient between measured EMG data and pth powered sine function at phase (3): E p C MAX sin E A f (3) To estimate the phase of the present tremor signal, the E algorithm searches for the maximum correlation C MAX between the powered sine wave vectors B m (m=1,2,...,n) and the windowed EMG vector E n. Fig. 7 presents an example when the base wave is a squared sine wave. The correlation function is expressed as (4) (5) (6) (7) below: En Bm C (4) E B n m i 100 C MAX : Maximum Correlation Phase E : Estimated Phase Fig. 7 Phase estimation algorithm. This algorithm searches for the maximum correlation between the base wave and the windowed EMG. n e n N, e n N 1 e n b m N 1, bm N 2 bm p bx xft E,..., (5) B m,..., (6) sin (7) B m is the powered sine wave whose phase ranges from m N 1 Ft to mf t. E n is the vector of the present and past 1/F s sampled EMG data. The maximum correlation C MAX is also a parameter of the gain function. The gain of the attenuation ratio is defined by the sigmoid function (8). 1 f CMAX (8) 1 exp a C offset The gain a of the sigmoid function was set at 100. The offset value was set at 0.8. D. Methodology The data used in this experiment was the same as that of the experiment described in Section III. The subject was one ET patient (male, 71 years old). The experimental task was an elbow flexional movement while holding a water filled bottle and while holding an empty bottle. The patient s EMG signal and elbow angle are measured through an EMG electrode (Biometrics Ltd) and by a goniometer (Biometrics Ltd), respectively. Each of the data was sampled at a rate of 1000Hz. To make instruction signal for the TDNN, we divided EMG data into three phases of the elbow joint state: the stable "OFF" phase on the extended position, the flexing "ON" phase, and the stable "OFF" phase in flexed position. The two state phases were defined at one second from the zero cross point of the elbow joint angular velocity. The angular velocity v n was calculated by the five-point method for the first derivative as follows: gn2 8gn 1 8gn 1 gn2 vn (13) 12t Where g n is the elbow joint angle passed LPF which is of Chebyshev II type with cutoff frequency 10Hz and damping coefficient 0.7. MAX 1419

6 Two types of signals were input to the TDNN. One was the signal processed with the powered sine filter and passed LPF, which was of Chebyshev II type with cutoff frequency10hz and damping coefficient 0.7. The other was the signal processed with the powered sine filter with the threshold switch and the LPF. The base wave of the powered sine filter was the 4 th power sine wave at 5.3Hz, which was the optimal condition for this subject in our previous study [17] [19]. The threshold for the powered sine filter was set as in Table 1 by trial and error. The length of the signal input to the TDNN was set at 500 ms by trial and error. The number of units of each layer was 500 at the input layer, 250 at the hidden layer, 1 at the output layer. As learning data for every condition, 2 out of 10 movements were selected randomly while the remaining 8 movements were selected as test data. To compensate for the shortage of test data, k-fold cross validation (k=5) was used. The TDNN learned times. The recognition rate is calculated in each movement state. The definition of the recognition rate, R state, in this study is as follow: R 1T( state other ) (14) state T state where T(state other) is the time when recognition result is false in the state, and T state is the time length of the state. E. Result and Discussion Fig. 8 shows the result of the calculated recognition rate, R state. The result was almost the same except for the two characteristic results. One is that the proposed method was especially superior to the previous method at the state of maintaining posture in extended position in the experiment while holding the empty bottle (Result 1). The other is that the proposed method was inferior to the previous method at the state of maintaining a posture in flexed position in the experiment while holding the bottle filled with water (Result 2). To discuss these two results, we focus on the following two factors: 1) Reduction of the attenuation of signals generated during performance of voluntary movement 2) Variability of the frequency of tremor signals Regarding the first factor, as stated before, the proposed method can prevent attenuation during performance of voluntary movement. Fig. 9 shows the total activation of the EMG signal processed with each method. The total activation, EMG Total, was calculated as follows: it End EMG EMG( i) (15) Total itstart where T Start and T End respectively refer to the time when the state starts and that when the state ends. From the graph, the total activation of the EMG signal generated during performance of voluntary movement in the experiment while holding the empty bottle was quite different between the proposed method and the previous method: 73.0V vs 54.2V. The proposed method succeeded in reducing attenuation during voluntary movement. This reduction seems to be the reason for the good recognition in Result 1. Focusing on the difference between total activation during maintenance of a posture in extended position and that of performance of voluntary movement in each method, the difference of the proposed method is much larger than that of the previous method. The large difference makes it easier for the TDNN to recognize each state. From this point of view, it can be said that the proposed method was able to progress the recognition. The second factor is variability of the frequency of tremor signals. Here, we make the same consideration as the first factor for Result 2; that is, we consider the difference between total activation during performance of voluntary movement and that during maintenance of a posture in flexed position while holding the water filled bottle. The difference of the proposed method was also larger than that of the previous method. However, the recognition rate does not match the difference. The recognition rate of the previous method is higher than that of the proposed method. The cause of this mismatch seems to be variability of the frequency of tremor signals. Fig. 10 (a) and (b) shows the desirable and undesirable case of the EMG signal processed with the proposed method, respectively. The desirable processing is no attenuation during voluntary movement and large attenuation during maintenance of a posture. However, in several cases, the EMG signal generated during performance of voluntary movement was attenuated, or the EMG signal generated during maintenance of a posture was not attenuated, because of variability of the calculated frequency of tremor signals. This failure caused the misrecognition. To organize these two factors, the proposed method was able to progress the recognition accuracy by reducing the attenuation during voluntary movement, but was not able to recognize accurately in the cases where the frequency of tremor signals varied widely between each movement. Therefore, as a future work, we need to develop a method that satisfies both these factors. To accomplish this, we will review the method to calculate the frequency of tremor signals and enable the filter to accommodate to the variability of the frequency of tremor signals. IV. CONCLUSION ET is a disorder that causes involuntary oscillations in patients both while they are engaged in actions and while maintaining a posture. ET patients face serious difficulties in performing daily living activities, for example, eating food, drinking water, and writing. We have thus been developing an EMG-controlled exoskeletal robot to suppress tremors. In this paper, we described the signal processing method used to estimate voluntary movement from the EMG signal of ET patients, which involved a mixture of voluntary movement 1420

7 Table 1 The preset threshold of each experimental condition Experimental condition Preset threshold ms Holding a bottle filled with water 165 Holding an empty bottle 150 (a) Desirable attenuation (a) The experiment with an empty bottle (b) The experiment with a bottle filled with water Fig. 8 The recognition rate of each state of the patient s movement. (i), (ii),and (iii) indicate the state of the patient s movement of maintaining posture in extended position, that of voluntary movement, and that of maintaining posture in flexed position, respectively. (a) The experiment with an empty bottle (b) The experiment with a bottle filled with water Fig. 9 The total EMG activation of each state of the patient s movement. (i), (ii),and (iii) indicate the state of the patient s movement of maintaining posture in extended position, that of voluntary movement, and that of maintaining posture in flexed position, respectively. (b) Undesirable attenuation Fig. 10 The desirable attenuation and undesirable attenuation of the proposed method accurately, tremor signals must be removed from the patient s EMG signal. To date, we have been developing a filter that removes tremor signals from the patient s EMG. This filter was designed based on the hypothesis that the rectified tremor signals are able to be approximated by a powered sine wave. This filter had large effect on removing tremor signals. However, tremor signals are generated both while voluntary movements are being performed and while a posture is maintained, and the filter ended up attenuating both signals. But, to control this robot accurately, the signal generated during performance of voluntary movement is expected not to be attenuated. Therefore, in this paper, we proposed a method that attenuates only tremor signals during maintenance of a posture. To accomplish this objective, we focus on the frequency of tremor signals. From the experiment, we confirmed the characteristic that the frequency of tremor signals varied depending on the state of the patient s movement. We then used the frequency as a switch to activate the previously proposed filter by setting a threshold. As an evaluation, signals processed by the proposed method and by the previous method were input to a time delay neural network. The proposed method succeeded in partly improving recognition due to reduction of attenuation during 1421

8 voluntary movement. However, the proposed method failed recognition in the cases where the frequency of tremor signals varied widely. As a future work we will review the method to calculate the frequency of tremor signals and improve recognition. ACKNOWLEDGEMENT We sincerely thank the subject for participating in our experiment. REFERENCES [1] A. Anouti and W. Koller, Tremor disorders : Diagnosis and management, Western J. Med., vol. 162, no. 6, , [2] R. Elble, and W. Koller, Tremor. Baltimore, MD : Johns Hopkins Univ. Press, 1990 [3] E. Rocon, J. Belda-Lois, J. Sanchez-Lancuesta, and J. L. Pons, Pathological tremor management : Modeling, compensatory technology and evaluation, Technol. Disability, vol. 16, 13-18, [4] A. P. L. Bo, P. Poignet, D. Zhang, W. T. Ang, FES-controlled co-contraction strategies for pathological tremor compensation, The 2009 IEEE/RSJ IROS, [5] A. P. L. Bo, P. Poignet, Tremor attenuation using FES-based joint stiffness control, IEEE International Conference on Robotics and Automation, , [6] J. A. Gallego, E. Rocon, J. L. Pons, Estimation of instantaneous tremor parameters for FES-based tremor suppression, IEEE International Conference on Robotics and Automation, , 2010 [7] E. Rocon, J. M. Belda-Lois, A. F. Ruiz, M.Manto, J. C. Moreno, and J. L. Pons, Design and Validation of a Rehabilitation Robotic Exoskeleton for Tremor Assessment and Suppression, IEEE TRANS. NEU. SYS. REH. ENG., 15(3), 2007 [8] Y. Hasegawa, Y. Mikami, K. Watanabe, Z. Firouzimehr, Y. Sankai, Wearable handling support system for paralyzed patient, IEEE/RSJ IROS, , [9] K. Kiguchi, Y. Imada, M. Liyanage, EMG-Based Neuro-Fuzzy Control of a 4DOF Upper-Limb Power-Assist Exoskelton, 29 th IEEE EMBS, , 2007 [10] T. Ando, J. Okamoto, M. G. Fujie, Intelligent corset to support rollover of cancer bone metastasis patients, The 2008 IEEE/RSJ IROS, , 2008 [11] M. Oosawa, Clinical Application of Surface Electromyography, Tokyo Women s Medical Univ. Press, 59(6), , 1989 [12] Eiichi Ohara, Kenichi Yano, Satoshi Horihata, Takaaki Aoki and Yutaka Nishimoto, Tremor suppression control of Meal-Assist Robot with adaptive filter, The 2009 IEEE ICORR, , 2009 [13] E. Ohara, K. Yano, S. Horihata, T. Aoki, Y. Nishimoto, Development of Tremor-Suppression Filter for Meal-Assist Robot, Eurohaptics Conference/IEEE the third WorldHAptics, , 2009 [14] C. N. Riviere, W.-T. Ang, and P. K. Khosla, Toward active tremor canceling in handheld microsurgical instruments, IEEE Trans. on Robotics and Automation, vol. 15, , 2003 [15] Journre, H.L. Demodulation of amplitude modulated noise: a mathematical evaluation of a demodulator for pathological tremor EMGs. IEEE Trans. Biomed. Eng., 1983, BME-30: [16] F. Liu, R. Stern, X. Huang and A. Acero Efficient Cepstral Normalization for Robust Speech Recognition. Proceedings of ARPA Human Language Technology Workshop, 1993 [17] M. Seki, Y. Matsumoto, T. Ando, Y. Kobayashi, H. Iijima, M. Nagaoka, M. G. Fujie, Development of Essential Tremor Noise Suppression Filter for Voluntary Movement Extraction from surface EMG, 2011 IEEE EMBC [18] M. Seki, Y. Matsumoto, T. Ando, Y. Kobayashi, H. Iijima, M. Nagaoka, M. G. Fujie, Development of Robotic Upper Limb Orthosis with Tremor Suppressiblity and Elbow Joint Movability, 2011 IEEE SMC [19] M. Seki, Y. Matsumoto, T. Ando, Y. Kobayashi, H. Iijima, M. Nagaoka, M. G. Fujie, The weight load inconsistency effect on voluntary movement recognition of essential tremor patient, International conference on Robotics and Biomimetics, 2011 [20] K. Hirose, Text of Electromyography, Bunkoudou, [21] G Deuschl, J Raethjen, M Lindermann, The pathophysiology of tremor, Muscle Nerve, 24, pp , 2001 [22] M. E. Heroux, G. Pari, K. E. Norman, The effect of inertial loading on wrist postural tremor in essential tremor, Clinical Neurophysiology, 120, , 2009 [23] K. Kuribayashi, et al., A discrimination system using neural network for EMG-controlled prostheses, IEEE International Workshop on Robot and Human Communication, 63-68, 1992 [24] O. Fukuda, T. Tsuji et al. An EMG controlled human supporting robot using neural network, The 1999 IEEE/RSJ IROS, [25] M. Zecca et al., Control of Multifunctional Prosthetic Hands by Electromyographic Signal, Critical Rev. in BIO. Eng., 30(4-6), , [26] Alexander Waibel, Phoneme Recognition Using Time-Delay Neural Network, IEEE Transaction on Acoustic, Speech, and Signal processing, Vol. 37, No. 3, , [27] M. F. Kelly, P.A. Parker, R. N. Schott, The Application of Neural Networks to Myoelectric Signal Analysis, A preliminary Study, IEEE Transactions on Biomedical Engineering, Vol. 37, No. 3, , [28] T. Tsuji, O. Fukuda, H. Ichinobe, M. Kaneko, A log-linearized Gaussian mixture network and its application to EEG pattern classification, IEEE Trans. Systems, Man and Cybernetics Part C: Application and Reviews, Vol. 29, No. 1, 60-72,