Automated decomposition of needle EMG signal using STFT and wavelet transforms

We present an automated method for decomposing EMG signals into their components, motor-unit action-potential (MUAP) trains based on short time Fourier transform STFT and wavelet transform. Since the number of MUAP classes composing the EMG signal, the number of MUAP's per class, their firing pattern, and the expected shape of the MUAP waveforms are unknown, the decomposition of real EMG signals into their constituent MUAP's and their classification into groups of similar shapes is a typical case of an unsupervised learning pattern recognition problem. The method is able to handle single-or multi-channel signals, recorded by concentric needle electrodes during low and moderate levels of muscular contraction. The method uses empirical features in STFT transform, shape and template of MU and CWT in order to decompose the signal to its original MUAP. Also the discrete wavelet transform has been acquired in early steps in order to eliminate the level of low amplitude noise in signal. We compare the output of the automated algorithm with manual decomposition and results seems quiet acceptable. The average success rate for the FCM with wavelet coefficients as features was 91.01 %.