Sensitivity Evaluation of HOS Parameters by Data Fusion from HD-sEMG Grid Mariam Al Harrach UMR CNRS 7338, Biomechanics and Bio-engineering University of Technology of Compiegne (UTC) Compiègne, France mariam.harrach@hotmail.com F.S. Ayachi Multimodal Interaction Laboratory, SIS-McGill University, Montréal, QC, Canada sofiane.ayachi@mcgill.ca Sofiane Boudaoud,Jeremy Laforet,Frederic Marin UMR CNRS 7338, Biomechanics and Bio-engineering University of Technology of Compiegne (UTC) Compiègne, France sofiane.boudaoud@utc.fr, jeremy.laforet@utc.fr, frederic.marin@utc.fr Abstract—the objective of this paper is to study the sensitivity of High Order statistics (HOS) parameters (the kurtosis and the Skewness) toward variation of the force intensity by applying different methods of data fusion. The data fusion allows us to obtain a single EMG signal or a single HOS parameter set from a 64 signals captured by an 8x8 High Density Surface EMG (HD-sEMG) grid. For this purpose, we started by calculating the HOS parameters (Kurtosis and Skewness) for the 64 monopolar signals for each one of three force intensities: 20%, 50% and 80% MVC. Then we applied two different data fusion procedures: Laplacian matrix coupled to Principle Component Analysis (PCA), and Laplacian matrix coupled with HOS parameter averaging. According to the obtained results, we noticed an important spatial sensitivity of the HOS parameters according to force variation for the monopolar grid. After data fusion, both studied techniques gave interesting results with better sensitivity for the Laplacian matrix combined to HOS parameter averaging method. Further studies are envisaged to assess the HOS parameter sensitivity to varying force and muscle anatomies. Keywords—Data Fusion, High Order Statistics, HD-sEMG, principal component analysis, Laplacian matrix, muscle force. I. INTRODUCTION he electromyogram (EMG) signal is a complex biomedical signal that measures electrical currents generated in muscles during contraction that represent neuromuscular activities. Therefore, a single monopolar electrode or even bipolar detection system is hardly sufficient to obtain a reliable signal that reflects the muscle, because of the variability of surface EMG that's caused by a number of factors such as: the timing and intensity of muscle contraction, the distance of the electrode from the active muscle area, the electrode and amplifier properties and the quality of contact between the electrodes and the skin [1], [2]. Since EMG signals constitute a summation of the motor unit (MU) action potentials, it occurs within the detection area of the electrode constructive and destructive superimpositions highly dependent on the MU spatial distribution, causing a change in the sEMG amplitude. Therefore the use of multiple, spatially distributed EMG channels, collecting independent information from separate sources, will improve the accuracy of the muscle activation analysis [4]. New sEMG recording methods have been lately developed; one of these methods is high-density sEMG. HD-sEMG is a non-invasive technique to measure electrical muscle activity with multiple closely spaced electrodes overlying a restricted area of the skin. In our study we used a simulated 64 electrode grid (8x8) by a developed sEMG-force model using parallel computing [1],[2]. The aim of the proposed study is to evaluate the sensitivity of HOS parameters according to contraction level variation by the data fusion from the simulated HD-sEMG grid. In fact, in precedent study, interesting results have been obtained with few electrodes [1], [2]. Multisensor data fusion is a technology that enables combining information from several sources in order to form a unified picture [4], [5]. This picture should be representative of principal modalities that act in the underlying process. The recorded monopolar signals from the grid can also be considered as a multidimensional dataset that probably contains redundant information to some degree [4]. As an unbiased statistical method, Principal Component Analysis (PCA) can be used to detect this type of redundancy in multivariate data by means of mode reduction [5]. In fact, after obtaining the Laplacian matrix by applying different combinations of Laplacian filters on the electrode grid, we attempted either to combine the Laplacian matrix with PCA and compute the HOS parameters on the obtained first mode or to calculate the average of the HOS parameters T 2013 2nd International Conference on Advances in Biomedical Engineering 978-1-4799-0251-4/13/$31.00 ©2013 IEEE 97