Detection of amyotrophic lateral sclerosis using signal processing and machine learning

Abstract

Electromyography(EMG) signals provide signi cant information for the diagnosis of neuromuscular disorders like Amyotrophic Lateral Sclerosis(ALS) which is a form of Motor Neuron Disease(MND). Due to the stochastic nature of EMG signals di erent preprocessing and feature extraction techniques need to be applied in order to extract useful information from the raw noisy signals. Time-Frequency analysis and EMG Decomposition are two of the widely implemented techniques for feature extraction from EMG signals. However, due to extrinsic and intrinsic artifacts any one feature extraction technique alone does not provide enough information in order to show a consistent performance of classi cation across a variety of dataset. EMG signal data set acquired from di erent sources provide varying outcome when passed through the same classi cation technique. This is a major problem while creating software which is able to perform automated classi cation and analysis of EMG signals on a wide variety of data set with minimum human intervention. This paper proposes a method for classi cation of ALS based on evaluation of multiple features extracted from three domains of EMG signal: time domain representation, frequency domain representation and Muscle Unit Action Potential(MUAP) waveform acquired via EMG decomposition of the signal. 43 features were evaluated using feature selection techniques like chi-squared test and recursive feature elimination. Our experimental results show that amplitude, duration and area of the MUAP waveform estimated for each motor unit, inter-spike-intervals of the motor units, variance, zero crossings, zero lag of autocorrelation, waveform length and slope sign change of the time domain representation, average spectral amplitude, total power, variance of central and mean frequency from feature domain representation of the signal provides the best accuracy at an average rate of 85%, a true positive rate(TPR) of 86% and a false positive rate(FPR) of 20% approximately.