Application of Empirical Mode Decomposition Combined With Notch Filtering for Interpretation of Surface Electromyograms During Functional Electrical Stimulation


Pilkar R., Yarossi M., Ramanujam A., Rajagopalan V., BAYRAM M. S. B. , Mitchell M., et al.

IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, cilt.25, ss.1268-1277, 2017 (SCI İndekslerine Giren Dergi)

  • Cilt numarası: 25 Konu: 8
  • Basım Tarihi: 2017
  • Doi Numarası: 10.1109/tnsre.2016.2624763
  • Dergi Adı: IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING
  • Sayfa Sayısı: ss.1268-1277

Özet

The goal of this paper is to demonstrate a novel approach that combines Empirical Mode Decomposition (EMD) with Notch filtering to remove the electrical stimulation (ES) artifact from surface electromyogram (EMG) data for interpretation of muscle responses during functional electrical stimulation (FES) experiments. FES was applied to the rectus femoris (RF) muscle unilaterally of six able bodied (AB) and one individual with spinal cord injury (SCI). Each trial consisted of three repetitions of ES. We hypothesized that the EMD algorithm provides a suitable platform for decomposing the EMG signal into physically meaningful intrinsic mode functions (IMFs) which can be further used to isolate electrical stimulation (ES) artifact. A basic EMD algorithm was used to decompose the EMG signals collected during FES into IMFs for each repetition separately. IMFs most contaminated by ES were identified based on the standard deviation (SD) of each IMF. Each artifact IMF was Notch filtered to filter ES harmonics and added to remaining IMFs containing pure EMG data to get a version of a filtered EMG signal. Of all such versions of filtered signals generated from each artifact IMF, the one with maximum signal to noise ratio (SNR) was chosen as the final output. The validity of the filtered signal was assessed by quantitative metrics, 1) root mean squared error (RMSE) and signal to noise (SNR) ratio values obtained by comparing a clean EMG and EMD-Notch filtered signal from the combination of simulated ES and clean EMG and, 2) using EMG-force correlation analysis on the data collected from AB individuals. Finally, the potential applicability of this algorithm on a neurologically impaired population was shown by applying the algorithm on EMG data collected from an individual with SCI. EMD combined with Notch filtering successfully extracted the EMG signal buried under ES artifact. Filtering performance was validated by smaller RMSE values and greater SNR post filtering. The amplitude values of the filtered EMG signal were seen to be consistent for three repetitions of ES and there was no significant difference among the repetition for all subjects. For the individual with a SCI the algorithm was shown to successfully isolate the underlying bursts of muscle activations during FES. The data driven nature of EMD algorithm and its ability to act as a filter bank at different bandwidths make this method extremely suitable for dissecting ES induced EMG into IMFs. Such IMFs clearly show the presence of ES artifact at different intensities as well as pure artifact free EMG. This allows the application of Notch filters to IMFs containing ES artifact to further isolate the EMG. As a result of such stepwise approach, the extraction of EMG is achieved with minimal data loss. This study provides a unique approach to dissect and interpret the EMG signal during FES applications.