Detection of Sleep Apnea from ECG signal using WT and ANN Classifiers

ABSTRACT The paper discusses the methodology followed for detection of Sleep apnea from ECG signal using Wavelet transform (WT)based Preprocessing, QRS detection and Feature extraction, followed by Feature Reduction process using Principal component Analysis(PCA). Finally classification of the input ECG signal is performed using ANN classifiers trained using back propagation algorithms namely, Levenberg-Marquardt (ANN_LM) and Scaled Conjugate algorithms (ANN_SCG) and compare their performance. The benchmark dataset from MIT's Physionet.org, ECG Apnea database consisting of 70 ECG signal recordings are used for the experimentation. The performance measures of preprocessing and QRS detection using WT are presented in terms of improved Signal to Noise Ratio (SNR) of 15%, Accuracy(Acc)=99.7%, Sensitivity(Se) =99.5 and Specificity (Sp)= 99.6% along with reduced computation time. PCA acts on the time and frequency domain features of ECG signal and enables the selection of dominant features to be given as input to ANN classifiers. K means clustering used with ANN_SCG classifier could achieve the improved classification measures as Acc=Se=Sp=100%, as well the reduced computation time as compared with ANN_LM classifier used without clustering. The saving of computation cost and higher performance measures are achieved mainly by the WT based denoising and QRS detection , followed by PCA, k-means clustering and the fast convergent SCG algorithm, making ANN_SCG classifier have an edge over the performance of ANN_LM classifier. The performances of these classifiers are further compared with other two published methods. Keywords: ECG-Apnea database, Wavelet Transform, Principal Component Analysis, Artificial Neural Networks, Levenberg –Marquardt algorithm, Scaled Conjugate Gradient Algorithm