Ocenjevanje pravilnosti razpoznave posameznih impulzov iz večkanalnih konvolutivnih mešanic impulznih izvorov : doktorska disertacija

slovenian

doktorska dela

2021

convolutive mixtures of pulse sources   decomposition   decomposition accuracy   high-density surface electromyogram   identification accuracy assessment   individual firing   motor unit   pulse trains

Univerza v Mariboru, Fakulteta za elektrotehniko računalništvo in informatiko

In this doctoral thesis, we develop and present the Witness Based Segmentation (WBS) algorithm for accuracy assessment of individual pulse identification from multichannel convolutive mixtures of pulse sources. We limit ourselves to the High-Density surface Electromyograms (HDsEMG), whose individual motor unit (MU) pulse trains are estimated by using previously established Convolution Kernel Compensation (CKC) decomposition method.%Evaluating the performance of HDsEMG decomposition is of great importance when we use the decomposition results to support the solutions in the fields of neurorehabilitation, prosthetics, orthopedics, ergonomics, ambient intelligence, sports sciences and in the video game industry. Several accuracy assessment techniques have already been developed, each with its advantages and disadvantages. But none of them supports the accuracy assessment for every individual MU firing identification. Instead, they mostly provide a joint accuracy assessment for identification of all the MU firings of each MU.%The developed WBS algorithm builds on the already established Pulse-to-Noise Ratio (PNR) metric, which is upgraded in the first step of the developed algorithm. For this purpose, we use witnesses, selected from MU pulses that are identified by the CKC method. We demonstrate that the value of the PNR metric calculated on these witnesses decreases statistically significantly when an incorrectly identified pulse is included in the calculation of the MU filter used to estimate the MU pulse train. We then use this upgraded PNR metric to automatically estimate whether an individual pulse that is not a witness belongs to the set of correctly or incorrectly identified MU firings. The witness selection strategy, the number of witnesses, and the number of added pulses, which were not identified as MU firings by the CKC decomposition method, represent the internal parameters of the algorithm. Therefore, we thoroughly analyze their impact on the performance of the WBS algorithm.%The performance of the developed WBS algorithm is tested on MU pulse trains identified by the CKC decomposition method from both synthetic and experimental HDsEMG signals. Gaussian and Laplacian noise with a signal-to-noise ratio of 30, 20 and 10 dB was added to the synthetic signals of the simulated biceps brachii muscle. The set of experimental signals consisted of signals from biceps brachii and tibialis anterior muscles and signals from simultaneously active triceps surae muscles. In the case of biceps brachii and tibialis anterior muscle, we also had intramuscular EMG signals acquired simultaneously with the HDsEMG signals. Therefore, we used the previously established two-source validation method to identify the currently and incorrectly identified MU firings from HDsEMG signals and validate our WBS algorithm. %The developed WBS algorithm exhibits linear computation complexity O(n), which means that its complexity increases linearly with the number of tested MU pulses. As such, the WBS algorithm is time and memory efficient and suitable also for use in real time MU firings identification.%The efficiency of the WBS algorithm is well demonstrated by the results presented in this thesis. In almost all the tested cases, there were significantly fewer incorrectly identified MU firings after the application of the algorithm than before its use. The results did not improve only in the case of experimental signals without or with very little noise, where the CKC decomposition method already returned high quality results. The efficiency of the algorithm was most evident in the cases of MUs that were identified by the CKC method with poor quality, that is in MUs that were identified by CKC method with PNR value lower than 28 dB.

07.08.2023