Deep hybrid neural networks for prediction missing segments in sEMG time series data

Surface electromyography (sEMG) has illustrated noteworthy findings over different disciplines; however, it suffers from several issues like signal interference, noise, and interruptions. In this research, the Savitzky-Golay filter was first used to extract meaningful data while preserving the overall shape of the data, and then two hybrid neural network models based on Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) were performed to predict the missing sEMG data. Their performance was compared with independent LSTM and GRU models using the coefficient of determination (R-squared), Root Mean Square Error (RMSE), and correlation coefficient (ρ). All models were trained, validated, and tested on extended and limited datasets. In addition, the optimal number of hidden neurons was determined experimentally for each condition. The outcomes indicated that the deep learning architecture based on sequential GRU and LSTM models outperformed all competitors with a prediction accuracy of 99.25% and 98.91% for the long and short training datasets, respectively.