Kahori Kita

Control strategy for a myoelectric hand: Measuring acceptable time delay in human intention discrimination

In order to enhance controllability of a myoelectric hand, we focus on a gap between the time when a human intends to move a myoelectric hand and the time when the hand actually moves (i.e., time delay). Normally, the myoelectric hand users dislike the time delay because it makes them feel uncomfortable. However, the users learn the time delay within some time ranges and, eventually, get feel comfortable to operate the hand. Thus, we assume, if we reveal the acceptable delay time (i.e., the time the users accept the gap with their learning ability), we can provide more time in a human intention discrimination process, and enhance its success rate. Therefore, we developed a mobile myoelectric hand system with an embedded linux computer, and conducted a ball catch experiment: we investigate the acceptable delay time by adding the delay time (i.e., 120[ms], 170[ms], 220[ms], 270[ms], 320[ms]) into the human intention discrimination process. As a result, we confirmed that the max accept delay time was approximately 170 [ms] that achieves 61% success rate.

Motion classification using epidural electrodes for low-invasive brain-machine interface

Brain-machine interfaces (BMIs) are expected to be used to assist seriously disabled persons’ communications and reintegrate their motor functions. One of the difficult problems to realize practical BMI is how to record neural activity clearly and safely. Conventional invasive methods require electrodes inside the dura mater, and noninvasive methods do not involve surgery but have poor signal quality. Thus a low-invasive method of recording is important for safe and practical BMI. In this study, the authors used epidural electrodes placed between the skull and dura mater to record a rat’s neural activity for low-invasive BMI. The signals were analyzed using a short-time Fourier transform, and the power spectra were classified into rat motions by a support vector machine. Classification accuracies were up to 96% in two-class discrimination, including that when the rat stopped, walked, and rested. The feasibility of a low-invasive BMI based on an epidural neural recording was shown in this study.

Development of Autonomous Assistive Devices -Analysis of change of human motion patterns-

The purpose of our research is to build a system for mutual adaptation between a user and assistive devices for restoration of motor function. To build such system, it is necessary to know humans motion patterns. In this paper, as the first step, we investigated human motion characteristic on human-machine system like EMG (electromyogram) prosthetic hand and EMG to motion classifier system. In the experiment, we measured the EMG signals and investigated a difference between motion patterns of teaching motion, i.e. users intended motion, and that of actual motion using the proposed criteria. As results, it is clear that these criteria are useful to analyze changes of human motion patterns

Self-organized clustering approach for motion discrimination using EMG signal

In order to control a myoelectric hand, it is necessary to discriminate among motions using electromyography (EMG) signals. One of the biggest problems in doing so is that EMG feature patterns of different motions overlap, and a classifier cannot discriminate clearly between them. Therefore, we propose a motion discrimination method to solve this problem. In this method, representative feature patterns are extracted from the EMG signals by using a self-organized clustering method, and users intended motions are assigned as class labels to these feature patterns on the basis of the joint angles of the hand and fingers. The classifier learns using training data that consists of feature patterns and class labels, and then discriminates motions. In an experiment, we compared the discrimination rates of the proposed and conventional methods. The results indicate that the discrimination rate obtained with the former is 5–30% higher than that obtained with the latter; this result verifies the effectiveness of our method.

Analysis of skill acquisition process: A case study of arm reaching task

Analysis of continuous process of motor learning gives a lot of useful knowledge for the recovery of human motion activities, and functional adaptability in new environment. This paper proposes a valuation index for degree of proficiency, and shows results of motor skill analysis for several arm reaching tasks. The motor skills were evaluated by using the reproducibility of muscle activation patterns, which were represented by using the variance value of the Electromyographic (EMG) signal patterns, and the motion accuracy. We confirm that the reproducibility is high when the motion accuracy is high, and the various skill acquisition processes exist due to individual difference. We conclude that, the reproducibility is one of the important indices for evaluating the degree of proficiency.

Enhanced Kapandji test evaluation of a soft robotic thumb rehabilitation device by developing a fiber-reinforced elastomer-actuator based 5-digit assist system

Abstract The main function of human hands is to grasp and manipulate objects, to which the thumb contributes the most. Various robotic hand rehabilitation devices have been developed for providing efficient hand function training. However, there have been few studies on thumb rehabilitation devices. Previously, we proposed a soft thumb rehabilitation device which is based on a parallel-link mechanism, driven by two different types of soft actuators. In this study, the device was integrated into a 5-digit assist system, in which fiber-reinforced elastomer actuators with improved bending angles, forces, and degrees of freedom were assembled onto a forearm socket. The device was evaluated by an enhanced Kapandji-Test, which included also a pressing force measurement in addition to the reachable positions of the thumb on the opposing fingers. The results showed that with the proposed approach, thumb functions for hand rehabilitation could be realized, which paves the way towards a full hand rehabilitation package with the 5-digit soft robotic hand rehabilitation system.

Evaluation method for the proficiency level of an operating myoelectric hand using EMG signals

To evaluate the proficiency level of an operating myoelectric hand, we proposed an evaluation index consisting of the accuracy and the reproducibility of electromyography (EMG) signal patterns. Our proposed method is not an absolute evaluation because we use bio-signals, so it is necessary to verify the correlation between the proposed index and performance evaluation to confirm the usefulness of the index. Therefore, we conducted classification tests on eight forearm motions and verified the correlation between the proposed method and the classification rate. There was a strong correlation between the accuracy and the classification rate. In addition, if the accuracy was high, high reproducibility led to an increase in the classification rate. We conclude that the proposed method can evaluate the proficiency level of a myoelectric hand.

Analysis of skill acquisition process for myoelectric hand control

s / Neuroscience Research 68S (2010) e335–e446 e437 P3-q23 Purpose or Strategy? Reconsideration of temporal discount in non-Markov situation Yoshiya Yamaguchi 1 , Yutaka Sakai 2 1 Grad. Sch. of Brain Sciences, Tamagawa Univ, Tokyo 2 Brain Sci Inst, Tamagawa Univ, Tokyo Temporal discount on subjective value of future return is an important issue in behavioral economics and neuroscience of decision-making. In reinforcement leaning theory, the temporal discount is introduced as a parameter in the purpose of learning. ‘State-value’ is defined as the expectation value of infinite summation of temporally discounted returns after each state. The purpose is to learn behavioral policy to maximize the state-values in all possible states. Existence of an optimal policy for this purpose is guaranteed in Markov decision process, but not in non-Markov decision process. In general, a policy to maximize state-value in a state does not always maximize state-value in another. We designed and analyzed a simple choice task in which the reward probability depends on the past choices as well as the current cue and choice, and we found that the optimal policy for the discounted value problem does not exist. This shows that the discounted value problem is not appropriate in the task. Whether the decision process is Markov or not depends on the definition of the state. In natural situation, the subject is not explicitly told how to define the state from available information. The subject should learn an appropriate state definition, as well as the action choice. In such a situation, the purpose depending on the state definition is inappropriate. A simple purpose independent of the state definition is to maximize the temporal average of net returns. We considered the temporal discount as a parameter of the learning strategy to maximize the temporal average, and analyzed the role in the proposed task. doi:10.1016/j.neures.2010.07.1935 P3-q24 Emerging direction sensitivity through STDP in a model of Xenopus visual system Minoru Honda 1,2,3 , Hidetoshi Urakubo 3,4, Shinya Kuroda 3,4 1 Dept Comp Biol, Univ of Tokyo, Tokyo, Japan 2 JSPS Research Fellow, Japan 3 Brain and Neural Systems Team, Comp Sci Research Program, RIKEN, Tokyo, Japan 4 Dept Biophys Biochem, Univ of Tokyo, Tokyo, Japan Neurons in visual systems can be characterized by direction sensitivity to moving objects, and the direction sensitivity is acquired by visual experience. In a Xenopus tadpole, the tectal neurons can be trained to become directionsensitive after repetitive exposure of retina to the moving bars in a particular direction. Although the acquirement of direction sensitivity seems to involve spike-timing-dependent plasticity (STDP), the overall process has not been revealed. Here, we have developed a computational model of a retinotectal circuit of a Xenopus tadpole to reveal the scenario about the acquirement of direction sensitivity. We have incorporated the simpler version of a biophysical STDP model into the retinotectal synapses of the model circuit as learning rule. The model circuit successfully reproduced the observed experimental results, the direction sensitivity acquirement of the tectal neurons, including training efficacies in different bar speeds. In the model circuit, the response of the retinal neurons depends simply on light intensity in their receptive fields, not on direction of moving objects. Therefore, the integrated retinal signals to the tectum are the same, regardless of the direction of moving bars. Through the analyses of the model circuit, we found that the direction sensitivity is acquired by concert of two mechanisms: timing shift of the retinal inputs by STDP, and amplification of the tectal inputs. STDP induces strong and transient retinal signal in a trained direction, and the intratectal connections amplifies the signal as a circuit activity. doi:10.1016/j.neures.2010.07.1936 P3-q25 Analysis of skill acquisition process for myoelectric hand control Kahori Kita 1,2 , Hiroshi Yokoi 1,3 1 Interfaculty Initiative in Information Studies, The University of Tokyo, Tokyo, Japan 2 Computational Neuroscience Laboratories, Advanced Telecommunications Research Institute International, Kyoto, Japan 3 Department of Mechanical Engineering and Intelligent Systems, University of Electro-Communications, Tokyo, Japan We have been proposed an index to evaluate the degree of proficiency of myoelectric hand control. Our index consists of accuracy and reproducibility of motion patterns. We use feature vectors consisting of amplitude and frequency spectra extracted from EMG as motion patterns. In myoelectric hand control, such user is judged as skilled: both overlap of feature vectors among different motion and variance of feature vectors in each motion are small. Thus, we defined the former as the accuracy d and the later as the reproducibility e. Here, d and e values show degree of overlap and variance, respectively. Thus, the accuracy and the reproducibility are higher when d and e are lower, and vice versa. In the previous research, we determined the threshold of proficiency: d < 5.0 and e < 0.005. In this study, we used them to judge whether subjects were skilled for myoelectric hand control, and also evaluated the change of the index value during training. In the experiment, healthy subject A and B trained to control these motions: relax, flexion, extension, grasp, open, thumb flexion, thumb extension, 4-5th finger flexion and pinching. We measured EMG on flexor carpi ulnaris, extensor carpi ulnaris and flexor pollicis longus. As a result, d and e of both subjects were lower than 5.0 and 0.005 at seven motions, and they were considered as skilled. Subject A also acquired proficiency in eight motions but B was not. When the subjects became skilled, the accuracy of all motions was increased through the training. On the other hand, the reproducibility was also finally increased, however those processes were varied with kind of motions. For instance, some motions had a maximum value, and it means that the subjects use different EMG patterns in each time. It is considered that the subjects try a lot of motion patterns to search suitable patterns. The results suggest that our proposed method is able to evaluate the degree of proficiency and also analyze the changes in more detail. doi:10.1016/j.neures.2010.07.1937 P3-q26 Abstract category learning scheme combining matching problem andvector quantization Atsushi Hashimoto , Haruo Hosoya Graduate School of Information Science and Technology, The University of Tokyo We study a novel pattern recognition problem of abstract categories. In this, input vectors composed of concrete values are classified according to templates composed of abstract values such that the input matches the template with its each abstract value replaced with a distinct concrete one. For example, both inputs AABB and BBCC match the template XXYY, while ABBA and CCCC do not. Moreover, a set of templates are learned from a set of inputs such that all inputs can match some templates with minimal errors. This kind of learning seems prevalent. Indeed, closely related neural activities have been discovered in prefrontal cortex of monkey (Shima et al. 2007), where some neurons responded selectively not to each action but to each abstract category of action sequence. Naive approaches extending existing techniques are insufficient for this problem. For example, vector quantization can bring together only close values in the input space, while abstract category learning must categorize distant inputs, e.g. AABB and CCAA. Also, simply assigning a different abstract value to each newly appearing concrete value in the input will not be robust against noises in inputs. We propose a new learning scheme for abstract categories. In this, we repeatedly and alternatingly perform matching and learning as follows. In matching, based on a given set of templates, we find an assignment of abstract values to concrete values for each input, which can be done by solving the maximum weight perfect matching problem. In learning, we update the set of templates minimizing matching errors, which can be done by vector quantization. We conducted a simulation of the learning scheme using the same task as used in the above-mentioned neurophysiological experiment, where we used exhaustive search for matching and Self-Organizing Maps for learning. After training the model with various input sequences, we found some units that successfully learned to response selectively to each abstract category. doi:10.1016/j.neures.2010.07.1938 P3-q27 Comparison of different outcome-based models for switching behavior of numerical device use by monkeys Junichi Iwata , Sumito Okuyama, Jun Tanji, Hajime Mushiake Physiology, Tohoku University School of Medicine, Sendai Our previous study showed that monkeys were able to change the number of dots shown on a display to a target number by using devices that increased (left external device) or decreased (right external device) the dot number by one dot in some trial blocks (i.e., device use (1) and vice versa

A self-organizing approach to generate raining data for EMG signal classification

We propose a method for generating training data by using a self-organized clustering technique for electromyography (EMG) signal classification. In this method, EMG signals are measured during motions, and representative feature patterns are extracted from the EMG signals by using the self-organized clustering method. A user determines the connections between feature patterns and motions, and training data are generated. These training data are employed for the classification of the users intended motions. It is necessary to determine the number of feature patterns required for motion classification. Therefore, we verify appropriate thresholds which determine the number of feature patterns with consideration of classification rate and learning time.