SEMG Onset Detection via Bidirectional Recurrent Neural Networks With Applications to Sports Science

Abstract

Surface electromyography (sEMG) provides physiological information that can be used in sports science. In many applications, sEMG signal activity, i.e., contractions, needs to be detected in the stream of sensor recordings. During sports exercises, the impact of any collision on the body due to an athlete’s movement (e.g., jump) forms an additive noise called motion-induced artifact (MIA) in sEMG recordings. This study proposes a bidirectional long short-term memory recurrent neural network (BLSTM-RNN) to automatically identify sEMG signal activity in measurements that include MIA. The proposed model is compared with the state-of-the-art techniques that are envelope, sample entropy (SampEn), modified adaptive linear energy detector (M-ALED), and adaptive contraction detection (ACD). As hamstring strain injuries (HSIs) are the most frequent and recurring injuries in professional football, this article uses sEMG data of different hamstring exercises performed by first-team players of the Leeds United Football Club. On data recorded using state-of-the-art sensors, the classification accuracy of the proposed solution is 96.73%, while the other methods reach 61.41% (sEMG envelope), 84.95% (SampEn), 58.86% (M-ALED), and 65.54% (ACD).