Emg Based Diagnosis of Myopathy and Neuropathy Using Machine Learning Techniques

Abstract

Myopathy and Neuropathy are non-progressive and progressive neuromuscular disorders which weakens the muscles and nerves respectively. Electromyography (EMG) signals are bio signals obtained from the individual muscle cells. EMG based diagnosis for neuromuscular disorders is a safe and reliable method. Integrating the EMG signals with machine learning techniques improves the diagnostic accuracy. The proposed system performs analysis on the clinical raw EMG dataset which is obtained from the publicly available PhysioNet database. The two-channel raw EMG dataset of healthy, myopathy and neuropathy subjects are divided into samples. The Time Domain (TD) features are extracted from divided samples of each subject. The extracted features are annotated with the class label representing the state of the individual. The annotated features split into training and testing set in the standard ratio 70: 30. The comparative classification analysis on the complete annotated features set and prominent features set procured using Pearson correlation technique is performed. The features are scaled using standard scaler technique. The analysis on scaled annotated features set and scaled prominent features set is also implemented. The hyperparameter space of the classifiers are given by trial and error method. The hyperparameters of the classifiers are tuned using Bayesian optimization technique and the optimal parameters are obtained. and are fed to the tuned classifier. The classification algorithms considered in the analysis are Random Forest and Multi-Layer Perceptron Neural Network (MLPNN). The performance evaluation of the classifiers on the test data is computed using the Accuracy, Confusion Matrix, F1 Score, Precision and Recall metrics. The evaluation results of the classifiers states that Random Forest performs better than MLPNN wherein it provides an accuracy of 96 % with non-scaled Time Domain (TD) features and MLPNN outperforms better than Random Forest with an accuracy of 97% on scaled Time Domain (TD) features which is higher than the existing systems. The inferences from the evaluation results is that Bayesian optimization tuned classifiers improves the accuracy which provides a robust diagnostic model for neuromuscular disorder diagnosis.