Classification of hand movements via electromyogram and auxiliary sensor data

CLASSIFICATION OF HAND MOVEMENTS VIA ELECTROMYOGRAM AND AUXILIARY SENSOR DATA SUMMARY In everyday life, human hand performs wide variety of functional movements via complex hierarchical control and sensor systems. Loss of natural hand, partial or complete loss of hand function due to neuromuscular disorders, impairs the quality of human life. For many years, numerous studies in the light of information from different disciplines such as control engineering, biomedical engineering, medical sciences, etc. have been carried out in order to improve or restore natural hand functions of patients who suffer from this negative condition. These studies have mainly focused on the exoskeleton support systems that help patients who have partial loss of hand function and the prostheses that perform movement by providing motion command from the existing healthy limb or existing nerves for the patients who have lost their natural limb. In general, these systems which aim to achieve natural hand function are called Myoelectric Control System (MCS). MCSs aim to use the signals which are acquired from the human body in order to recognize the hand motion intention and to convert recognized movements into the control commands. For the recognition of human motion intention, MCSs mainly benefit from the EMG signals which are the result of Action Potentials (AP) during nerve conduction. EMG signals, which can be obtained on the skin surface via surface electrodes, within the muscle fiber or subcutaneously via needle electrodes, have the highest determination characteristic to recognize hand movements among other bioelectric signals. EMG-based MCSs studies have started in the 1940's and are still in existence today. Natural human hand has a great number of Degree-of-Freedom (DoF), can provide a sensory feedback while interacting and manipulating different types of objects and perform wide diversity of functional movements. Despite the significant developments in EMG-based control systems, systems have not achieved the aforementioned capabilities of natural hand and as a consequence of this have not been produced commercially yet. Especially the absence of sensory feedback results in 'unnatural' hand sense which is the major obstacle in order to be approved by the patients. It is anticipated that integrating modern mechatronics and sensor systems technologies will eliminate a significant portion of the aforementioned problems. The recognition of the movement intention through the help of natural signals from the body retains its existence as a problem. Supervised pattern recognition techniques are commonly used in the recognition of movement intention. In the control scheme based on these methods; discriminative features of movements are extracted from the EMG signals obtained over the active muscle groups during movements and these features along with the movement class that corresponds to the features constitute the feature vector. Features extracted from EMG signals can be divided into 3 categories. Time domain features are extracted by using signal amplitude and does not require any signal domain transform, complex computational operations and long computational time. Frequency domain features requires signal domain transformation provides frequency content information by using power spectrum density of the signal. Time-frequency domain features provides more accurate description of physical phenomenon through time-frequency localization and requires transformation which analyzes signal through changeable-sized time windows depending on the frequency characteristics of the signal. The selection of the appropriate properties that represent the movement is the most decisive stage in the classification accuracy. Different learning algorithms are trained with the feature matrix including feature vectors and class of each observation and classifier models are produced as a result of training. The classification performance of the model is measured by the rate of classification accuracy obtained from the test data presented to the classifier without specifying movement classes. In order to increase the classification accuracy; the development of different classification algorithms and the use of multiple classifier systems, the identification of new features representing the movements with high discrimination and the creation of appropriate feature subsets, the creation of algorithms that reduce the dimension of signal in order to reduce operational time and complexity by preserving the large part of the information content, the use of the features obtained from auxiliary sensor data are recommended. In this study; The NinaPro database, which contains EMG, accelerometer, inclinometer and strain gauge signals of 50 movement frequently performed in daily life that are obtained from amputee and healthy individuals, is used. Studies with the highest number of movements classified with an acceptable classification success have been used as a starting point. The effect of ReliefF algorithm that weigths the features depending on the contributions to the classification accuracy, the effect of axuillary sensor data to the classification process, the effect of feature subsets which are different from the ones within existing studies have been examined. As a result of the study, it has been observed that reducing the dimension of the feature matrix via Relieff algorithm increased the clasification accuracy of kNN (k=1), QSVM, cubic SVM and ensemble claffier. It is also shown that using the time-domain features provides a higher classification accuracy than time-domain features or frequency domain features standalone. The last and most significant contribution of the study is that classification accuracy is increased proportional to the auxillary sensor data when appropriate features are selected.