Classification of human hand movements based on EMG signals using nonlinear dimensionality reduction and data fusion techniques

Abstract Surface electromyography (EMG) is non-invasive signal acquisition technique that plays a central role in many application, including clinical diagnostics, control for prosthetic devices and for human-machine interactions. The processing typically begins with a feature extraction step, which may be followed by the application of a dimensionality reduction technique. The obtained reduced features are input for a machine learning classifier. The constructed machine learning model may then classify new recorded movements. The features extracted for EMG signals usually capture information both from the time and from the frequency domain. Short time Fourier transform (STFT) is commonly used for signal processing and in particular for EMG processing since it captures the temporal and the frequency characteristics of the data. Since the number of calculated STFT features is large, a common approach in signal processing and machine learning applications is to apply a linear or a nonlinear dimensionality reduction technique for simplifying the feature space. Another aspect that arises in medical applications in general and in EMG based hand classification in particular, is the large variability between subjects. Due to this variability, many studies focus on single subject classification. This requires acquiring a large training set for each tested participant which is not practical in real life application. The objectives of this study were first to compare between the performances of a nonlinear dimensionality technique to a standard linear dimensionality method when applied for single subject EMG based hand movement classification, and to examined their performances in case of limited amount of training data samples. The second objective was to propose an algorithm for multi-subjects classification that utilized a data alignment step for overcoming the large variability between subjects. The data set included EMG signals from 5 subjects who perform 6 different hand movements. STFT was calculated for feature extraction, principal component analysis (PCA) and diffusion maps (DM) were compared for dimension reductions. An affine transformation for aligning between the reduced feature spaces of two subjects, was investigated. K-nearest neighbors (KNN) was used for single and multi-subject classification. The results of this study clearly show that the DM outperformed the PCA in case of limited training data. In addition, the multi-subject classification approach, which utilizes dimension reduction methods along with an alignment algorithm enable robust classification of a new subject based on another subjects’ data sets. The proposed framework is general and can be adopted for many EMG classification task.