Gunter Kanitz

Distance and mutual information methods for EMG feature and channel subset selection for classification of hand movements

Abstract Different approaches have been proposed to select features and channels for pattern recognition classification of myoelectric upper-limb prostheses. The goal of this work is to use deterministic methods to select the feature-channels pairs that best classify the hand postures at different limb positions. Two selection methods were tried. One is a distance-based feature selection (DFSS) that determines a separability index using the Mahalanobis distance between classes. The second method is a correlation-based feature selection (CFSS) that measures the amount of mutual information between features and classes. To evaluate the performance of these selection methods, EMG data from 10 able-bodied subjects were acquired when performing 5 hand postures at 9 different arm positions and 10 time-domain and frequency-domain features were extracted. Classification accuracy using both methods was always higher than including all the features and channels and showed slight improvement over classification using the state-of-art TD features when evaluated against limb variation. The CFSS method always used less feature-channel pairs compared to the DFSS method. Using both methods, selection of channels placed on the posterior side of the forearm was significantly higher than anterior side. Such methods could be used as fast screening filters to select features and channels that best classify different hand postures at different arm positions.

Decoding of individuated finger movements using surface EMG and input optimization applying a genetic algorithm

In this paper we present surface electromyo-graphic (EMG) data collected from 16 channels on five unimpaired subjects and one transradial amputee performing 12 individual finger movements and a rest class. EMG were processed using a traditional Time Domain feature-set and classifiers: a Linear Discriminant Analysis (LDA) a k-Nearest Neighbors (k-NN) and Support Vector Machine (SVM). Using continuous datasets we show that it is possible to achieve an accuracy up to 80% across subjects. Thereafter possibilities to reduce the numbers of channels physically required, as well as the number of features have been investigated by means of a developed Genetic Algorithm (GA) that included a bonus system to reward eliminated features and channels. The classification was performed firstly on the full datasets and in later runs using the GA. The GA demonstrated high redundancy in the recorded 16 channel data as well as the insignificance of certain features. Although the GA optimization yielded to reduce 8 to 11 channels depending on the subject, such reduction had little to no effect on the classification accuracies.

The Effects of Weight and Inertia of the Prosthesis on the Sensitivity of Electromyographic Pattern Recognition in Relax State

ABSTRACT For transradial amputees, the muscles in the residual forearm physiologically used for flexing/extending the hand fingers are the most appropriate targets for multifingered prostheses control. However, once the prosthetic socket is manufactured and fitted on the residual forearm, the electromyographic (EMG) signals recorded from the residual limb might not be originated only by the intention of performing finger movements but also by the muscular activity needed to sustain the prosthesis itself. In this work, we show that in eight healthy subjects wearing a prosthetic socket emulator, 1) variations in the weight of the prosthesis and 2) upper arm movements significantly influence the robustness of a traditional classifier based on a k-nn algorithm, causing a significant drop in performance. We demonstrate in simulated conditions that traditional pattern recognition does not allow the separation of the effects of the weight of the prosthesis because a surface recorded EMG pattern due only to the lifting or moving of the prosthesis is misclassified into a hand control movement. This suggests that a robust classifier should add to myoelectric signals, inertial transducers like multi-axes position, acceleration sensors, or sensors able to monitor the interaction forces between the socket and the end-effector.

Feature and Channel Selection Using Correlation Based Method for Hand Posture Classification in Multiple Arm Positions

In this work we propose a method based on correlation-based feature selection (CFS) to select features and channels for pattern recognition control of upper-limb prostheses. This method was applied on features extracted from myoelectric signals acquired from two able-bodied subjects and one individual with transradial amputation while contracting the muscles as to perform five functional hand postures in nine arm positions. The classification accuracy increased by using CFS for the able-bodied, while no statistical improvements has been highlighted for the amputee subject. The channels selected by this approach were mainly placed on the posterior side of the forearm which might reflect importance role of the extensor muscles over the flexor muscle when performing these hand postures. Further analysis with bigger dataset will be conducted to validate these preliminary findings.

Vector Autoregressive Hierarchical Hidden Markov Models for Extracting Finger Movements Using Multichannel Surface EMG Signals

We present a novel computational technique intended for the robust and adaptable control of a multifunctional prosthetic hand using multichannel surface electromyography. The initial processing of the input data was oriented towards extracting relevant time domain features of the EMG signal. Following the feature calculation, a piecewise modeling of the multidimensional EMG feature dynamics using vector autoregressive models was performed. The next step included the implementation of hierarchical hidden semi-Markov models to capture transitions between piecewise segments of movements and between different movements. Lastly, inversion of the model using an approximate Bayesian inference scheme served as the classifier. The effectiveness of the novel algorithms was assessed versus methods commonly used for real-time classification of EMGs in a prosthesis control application. The obtained results show that using hidden semi-Markov models as the top layer, instead of the hidden Markov models, ranks top in all the relevant metrics among the tested combinations. The choice of the presented methodology for the control of prosthetic hand is also supported by the equal or lower computational complexity required, compared to other algorithms, which enables the implementation on low-power microcontrollers, and the ability to adapt to user preferences of executing individual movements during activities of daily living.

Decoding of individual finger movements from surface EMG signals using vector autoregressive hierarchical hidden Markov models (VARHHMM)

In this paper we present a novel method for predicting individual fingers movements from surface electromyography (EMG). The method is intended for real-time dexterous control of a multifunctional prosthetic hand device. The EMG data was recorded using 16 single-ended channels positioned on the forearm of healthy participants. Synchronously with the EMG recording, the subjects performed consecutive finger movements based on the visual cues. Our algorithm could be described in following steps: extracting mean average value (MAV) of the EMG to be used as the feature for classification, piece-wise linear modeling of EMG feature dynamics, implementation of hierarchical hidden Markov models (HHMM) to capture transitions between linear models, and implementation of Bayesian inference as the classifier. The performance of our classifier was evaluated against commonly used real-time classifiers. The results show that the current algorithm setup classifies EMG data similarly to the best among tested classifiers but with equal or less computational complexity.

A Simulation Study to Characterize the Effects of Frequency Modulation during Epidural Electrical Stimulation

Epidural electrical stimulation (EES) is a well-known technique used to generate functional states in the networks of subjects with spinal cord injury. Recently, promising results have been achieved in both animals and humans showing that a voluntary control of lower limbs, or even of the complete locomotion is possible if EES is coupled with intense trainings, drugs, and with a proper tuning of the stimulation parameters. In this work we tested the effect of EES stimulation frequency onto the building block system of the spinal cord neural networks, i.e., the monosynaptic pathway. Single and compound firing rates of alpha motorneurons were computed using a realistic finite element model of the spinal cord coupled with a realistic fiber cable model of the monosynaptic reflex. Our results show that there is a typical resonating frequency for the system at 40 Hz, which is the optimal frequency found in experimental works to achieve the best walking performances. Moreover, the compound effect of the single cell firing rates on the muscle twitches has been also investigated.