Results 1. Armband rotation simulation 2. Classification Spatial Registration of Hand Muscle Electromyography Signals Contact and more information: [email protected] Manfredo Atzori 1 , Claudio Castellini 2 and Henning Müller 1 1 Department of Business Information Systems at the HES- SO Valais, Sierre, Switzerland 2 Institute of Robotics and Mechatronics, German Aerospace Research Center, Wessling, Germany Summary 1.sEMG hand prostheses have low control capabilities compared to recent advances in mechatronics 2.the Ninapro database (http://ninapro.hevs.ch ) has been realized to test sEMG prosthetics control algorithms 3.inter-subject differences in the positioning of the electrodes can determine substantial signal differences 4.increasing inter-subject similarity with spatial registration improves the classification results with statistical significance Introduction 1. Hand robotic prosthetics • usually controlled via surface electromyography (sEMG) • low control capabilities • long training times 2. Ninapro sEMG acquisition setup • 8 electrodes equally spaced around the forearm • 2 electrodes placed on fingers extensor and flexor muscles 3. Inter-subject differences in the positioning of the electrodes • have strong impact on the signal and on the classification task • are analyzed in few papers • can be diminished with spatial registration algorithms Methods 1. Datasets • 4 datasets of 27 subjects, 10 repetitions of 4 & 12 movements, acquired with 8 & 10 electrodes. 2. Armband rotation simulation • linear interpolation of signals from subsequent electrodes at steps of 1/10 of the distance between each of the electrodes • evaluation of the rotated sEMG signal similarity through the mean value of the Euclidean distance along the timeline • identification of the signals that minimize the distance between the inter-subject sEMG signals 3. Classification • preprocessing: synchronization; low-pass filtering at 1Hz (zero-phase second order Butterworth); normalization (0 mean and unitary standard deviation); averaging. • classifier: multi class Least-Squares support vector machines (LS-SVM) with RBF kernel • training: ten repetitions of each movement • test: on all the other movements, subjects and repetitions Conclusions • we performed spatial registration simulating the rotation of equally spaced electrodes • spatial registration of the sEMG signal of finger movements augments inter-subject LS-SVM classification with statistical significance • the difference between the average improvement of the similarity of the signals (8.12 ± 7.02)% and the average improvement in the classification (2.99 ± 1.43)% encourages to do further analyses about spatial registration and the features used for the classification 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 Subject 1 time (seconds/100) sEMG Activity (V) electrode 1 electrode 2 electrode 3 electrode 4 electrode 5 electrode 6 electrode 7 electrode 8 0 50 100 150 200 250 300 0 1 2 3 4 5 Subject 2 time (seconds/100) sEMG Activity (V) electrode 1 electrode 2 electrode 3 electrode 4 electrode 5 electrode 6 electrode 7 electrode 8 0 50 100 150 200 250 300 0 0.5 1 1.5 2 2.5 Subject 3 time (seconds/100) sEMG Activity (V) electrode 1 electrode 2 electrode 3 electrode 4 electrode 5 electrode 6 electrode 7 electrode 8 Figure 1: Inter-individual variability of sEMG signal patterns. Example of sEMG signal patterns from three subjects doing the same movement. Figure 2: Maximum inter- subject signal similarity improvement due to rotation simulation in percentage (top). Correspondent rotations (bottom). • the average similarity increase over all of the subjects due to rotation simulation is 8.12%, with standard deviation 7.02 % • the maximal rotation simulation increase in the similarity of signals across subjects is 33% 8 equally spaced electrodes •4 movements: classification improved of 4.69 percentage points, with an increase of 8.09% (p=0.01) •12 movements: classification improved of 1.21 percentage points, with an increase of 2.09% (p=0.04) 10 electrodes: •4 movements: classification improved by 2.87 percent, with an increase of 4.74% (p=0.01) •12 movements: classification improved of 1.94 percent, with an increase of 3.2% (p=0.02) Figure 4. Mean and standard deviations of the LS-SVM classification errors on the signal of the 8 electrodes equally spaced on the elastic armband (left) and from all the electrodes (right). 4 4 12 12