Development of a recurrent neural network based estimation algorithms for ankle power by using surface EMG

Estimation of ankle power can be used in identification of gait abnormalities and establishing timings of net power generation in powered prosthetic devices. Current inverse dynamics calculations of ankle power rely on gait analysis data collected in specialized, expensive laboratories, which limits its applicability and accessibility for prosthetic device users. The aim of this study is to develop a Recurrent Neural Network system to estimate ankle power during level walking by using only surface electromyography (sEMG) as algorithm inputs. For this purpose, an open access data set which includes 50 participants with 25 males and 25 females aged between 6 to 72. In the dataset there are sEMG data from upper leg muscles: Biceps Femoris (BF), Gastrocnemius Medialis (GM), Gluteus Maximum (GMax), Rectus Femoris (RF), Vastus Medialis (VM) and lower body muscles; Peroneus Longus (PL), Soleus (SO), Tibialis Anterior (TA). Algorithms for combinations of all these muscles have been developed. A correlation coefficient of 0.90 between the actual (result of gait analysis) and predicted ankle power is considered to perform successfully. 25 muscle combinations yielded successful correlations with 1 set of 1 muscle, 3 sets of 2 muscles, 5 sets of 3 muscles, 9 sets of 4 muscles, 5 sets of 5 muscles, 1 set of 6 muscles, 1 set of 7 muscles. Note that, all successful muscle groups include either PL or GM muscle. Our findings suggest that our system can be used in powered prosthetics control and detection of gait abnormalities.