Comparison of Deep Neural Network Models and Effectiveness of EMG Signal Feature Value for Estimating Dorsiflexion

Abstract

Robotic ankle–foot orthoses (AFO) are often used for gait rehabilitation. Our research focuses on the design and development of a robotic AFO with minimum number of sensor inputs. However, this leads to degradation of gait estimation accuracy. To prevent degradation of accuracy, we compared a few neural network models in order to determine the best network when only two input channels are being used. Further, the EMG signal feature value of average rate of change was used as input. LSTM showed the highest accuracy. However, MLP with a small number of hidden layers showed results similar to LSTM. Moreover, the accuracy for all models, with the exception of LSTM for one subject (SD), increased with the addition of feature value (average rate of change) as input. In conclusion, time-series networks work best with a small number of sensor inputs. However, depending on the optimizer being used, even a simple network can outrun a deep learning network. Furthermore, our results show that applying EMG signal feature value as an input tends to increase the estimation accuracy of the network.

High-Performance Surface Electromyography Armband Design for Gesture Recognition

“…For instance, the peak value reaching a certain threshold can be used as the judgment criterion for a corresponding gesture, while the root-mean-square (RMS) value reflects the average level of sEMG over a period of time. Frequency-domain features help to reveal components of the signal, including the median frequency, mean power frequency, and spectral area [ 15 , 16 ]. Angkoon Phinyomark et al extracted eight feature sets from the time–frequency domain and achieved a maximum accuracy of 89.7%, leaving ample room for further improvement [ 4 ].…”

The use of nonnormalized surface EMG and feature inputs for LSTM-based powered ankle prosthesis control algorithm development

“…(2) sEMG features, calculated over sEMG amplitudes of leg muscles, were used as LSTM inputs to predict ankle position and moment (see Figure 1 for a block diagram). A similar input integration into LSTM has been shown to be highly efficient in estimating joint angles ( Ma et al, 2020 ; Zaffir et al, 2021 ; Foroutannia et al, 2022 ). Presently, two separate LSTMs were developed and trained for predicting ankle joint position and moment.…”

“…In different studies, the number of sEMG amplitudes utilized for predicting joint kinematics and kinetics varies from eleven (e.g., Huang et al, 2009 , 2011 ; Du et al, 2013 ; Hargrove et al, 2013 ; Young et al, 2014 ; Spanias et al, 2016 ; Liu et al, 2017 ) to four (e.g., Hoover et al, 2012 ; Spanias et al, 2018 ). Zaffir et al (2021) developed and compared different types of neural network models for estimating dorsiflexion for robotic ankle-foot orthoses with similar concerns of minimizing the number of muscle inputs and eliminating mechanical sensors. They utilized sEMG data of four leg muscles of healthy participants.…”