A Novel Outlier Detection Method for Identifying Torque-related Transient Patterns of in vivo Muscle Behavior

Abstract

This paper proposed a novel outlier detection method, named l1-regularized outlier isolation and regression (LOIRE), to examine torque-related transient patterns of in vivo muscle behavior from multimodal signals, including electromyography (EMG), mechanomyography (MMG) and ultrasonography (US), during isometric muscle contraction. Eight subjects performed isometric ramp contraction of knee up to 90% of the maximal voluntary contraction, and EMG, MMG and US were simultaneously recorded from the rectus femoris muscle. Five features, including two root mean square amplitudes from EMG and MMG, muscle cross sectional area, muscle thickness and width from US were extracted. Then, local polynomial regression was used to obtain the signal-to-torque relationships and their derivatives. By assuming the signal-to-torque functions are basically quadratic, the LOIRE method is applied to identify transient torque-related patterns of EMG, MMG and US features as outliers of the linear derivative-to-torque functions. The results show that the LOIRE method can effectively reveal transient patterns in the signal-to-torque relationships (for example, sudden changes around 20% MVC can be observed from all features), providing important information about in vivo muscle behavior.