Methods for Improving the Classification Accuracy of Biomedical Signals Based on Spectral Features

Biomedical signals are long records of electrical activity within the human body, and they faithfully represent the state of health of a person. Of the many biomedical signals, focus of this work is on Electro-encephalogram (EEG), Electro-cardiogram (ECG) and Electro-myogram (EMG). It is tiresome for physicians to visually examine the long records of biomedical signals to arrive at conclusions. Automated classification of these signals can largely assist the physicians in their diagnostic process. Classifying a biomedical signal is the process of attaching the signal to a disease state or healthy state. Classification Accuracy (CA) depends on the features extracted from the signal and the classification process involved. Certain critical information on the health of a person is usually hidden in the spectral content of the signal. In this paper, effort is made for the improvement in CA when spectral features are included in the classification process. Spectral features are extracted from EEG signal using Multi Wavelet Transform (MWT). Epileptic and Normal cases are classified using k-Nearest Neighbors (k-NN) classifier. Independent Component Analysis (ICA) and Discrete Wavelet Transform (DWT) are used to extract features from ECG signals. These features along with temporal features are used in the classification process. An Artificial Neural Network (ANN) with three hidden layers is used to classify the signal to Ventricular Fibrillation (VF) and non-VF. EMG signal is a train of Motor Unit Action Potentials (MUAP). Dominant MUAP is identified using temporal energy criterion and spectral features are extracted from this using DWT. This method reduces the computational complexity to a large extent. Classification of the signals in to Amyotrophic Lateral Sclerosis (ALS), Myopathy and Paul Thomas and Dr. R.S. Moni http://www.iaeme.com/IJARET/index.asp 106 [email protected] Normal is done with k-NN classifier. In all the three cases, CA is found to be better than those based on existing methods. Training data set for classification are selected as those closest to the mean feature vector. This step also contributed to the accuracy of the results.