Taro Shibanoki

A Quasi-Optimal Channel Selection Method for Bioelectric Signal Classification Using a Partial Kullback–Leibler Information Measure

This paper proposes a novel variable selection method involving the use of a newly defined metric called the partial Kullback-Leibler (KL) information measure to evaluate the contribution of each variable (dimension) in the data. In this method, the probability density functions of recorded data are estimated through a multidimensional probabilistic neural network trained on the basis of KL information theory. The partial KL information measure is then defined as the ratio of the values before and after dimension elimination in the data. The effective dimensions for classification can be selected eliminating ineffective ones based on the partial KL information in a one-by-one manner. In the experiments, the proposed method was applied to channel selection with nine subjects (including an amputee), and effective channels were selected from all channels attached to each subjects forearm. The results showed that the number of channels was reduced by 54.3 ±19.1%, and the average classification rate for evaluation data using selected three or four channels was 96.6 ±2.8% (min: 92.1%, max: 100%). These outcomes indicate that the proposed method can be used to select effective channels (optimal or quasi-optimal) for accurate classification.

A Novel Channel Selection Method Based on Partial KL Information Measure for EMG-based Motion Classification

To control machines using electromyograms (EMGs), subjects’ intentions have to be correctly estimated and classified. However, the accuracy of classification is greatly influenced by individual physical abilities and measuring positions, making it necessary to select optimal channel positions for each subject. This paper proposes a novel online channel selection method using probabilistic neural networks (PNNs). In this method, measured data are regarded as probability variables, and data dimensions are evaluated by a partial KL information measure that is newly defined as a metric of effective dimensions. In the experiments, channels were selected using this method, and EMGs measured from the forearm were classified. The results showed that the number of channels is reduced with 33.33 ± 11.8%, and the average classification rate using the selected channels is almost the same as that using all channels. This demonstrates that the method is capable of selecting effective channels for classification.

Bioelectric signal classification using a recurrent probabilistic neural network with time-series discriminant component analysis

This paper outlines a probabilistic neural network developed on the basis of time-series discriminant component analysis (TSDCA) that can be used to classify high-dimensional time-series patterns. TSDCA involves the compression of high-dimensional time series into a lower-dimensional space using a set of orthogonal transformations and the calculation of posterior probabilities based on a continuous-density hidden Markov model that incorporates a Gaussian mixture model expressed in the reduced-dimensional space. The analysis can be incorporated into a neural network so that parameters can be obtained appropriately as network coefficients according to backpropagation-through-time-based training algorithm. The network is considered to enable high-accuracy classification of high-dimensional time-series patterns and to reduce the computation time taken for network training. In the experiments conducted during the study, the validity of the proposed network was demonstrated for EEG signals.

A class selection method based on a partial Kullback-Leibler information measure for biological signal classification

This paper proposes a novel class selection method based on the Kullback-Leibler (KL) information measure and outlines its application to optimal motion selection for bioelectric signal classification. When a user has no experience of controlling devices using bioelectric signals, for instance controlling a prosthetic hand using EMG signals, it is well known that voluntary generation of such signals might be difficult, so that the classification issue of multiple motions thus becomes problematic as the number of motions increases. An effective selection method for motions (classes) is needed for accurate classification. In the proposed method, the probability density functions (pdfs) of measured data are estimated through learning involving a multidimensional probabilistic neural network (PNN) based on the KL information theory. A partial KL information measure is then defined to evaluate the contribution of each class for classification. Effective classes can be selected by eliminating ineffective ones based on the partial KL information one by one. In the experiments performed, the proposed method was applied to motion selection with three subjects, and effective classes were selected from all motions measured in advance. The average classification rate using selected motions under the proposed method was 92.5 ± 0.9 %. These outcomes indicate that the proposed method can be used to select effective motions for accurate classification.

A training system for the MyoBock hand in a virtual reality environment

This paper proposes a novel EMG-based MyoBock training system that consistently provides a variety of functions ranging from EMG signal control training to task training. Using the proposed training sytem, a trainee controls a virtual hand (VH) in a 3D virtual reality (VR) environment using EMG signals and position/posture information recorded from the trainee. The trainee can also perform tasks such as holding and moving virtual objects using the system. In the experiments of this study, virtual task training developed with reference to the Box and Block Test (BBT) used to evaluate myoelectric prostheses was conducted with two healthy subjects, who repeatedly performed 10 one-minute tasks involving grasping a ball in one box and transporting it to another. The BBT experiments were also conducted in a real environment before and after the virtual training, with results showing an improvement in the number of tasks successfully completed. It was therefore confirmed that the proposed system could be used for myoelectric prosthesis control training.

Multi-channel surface EMG classification based on a quasi-optimal selection of motions and channels

This paper introduces a motion and channel selection method based on a partial Kullback-Leibler (KL) information measure. In the proposed method, the probability density functions of recorded data are estimated through learning involving a probabilistic neural network based on the KL information theory. Partial KL information is defined to support evaluation of the contribution of each dimension and class for classification. Effective dimensions and classes can then be selected by eliminating ineffective choices one by one based on this information, respectively. In the experiments, effective channels for classification were first selected for each of the six subjects, and the number of channels was reduced by 32.1±25.5%. After channel selection, appropriate motions for classification were chosen, and the average classification rate for the motions selected using the proposed method was found to be 91.7 ± 2.5%. These outcomes indicate that the proposed method can be used to select effective channels and motions for accurate classification.