OPTIMIZING ELECTRODE POSITIONS FOR TELE-CARDIOLOGY

Abstract

First of all, we obtain 64-channel ECG signals from patients. This is accompanied by mapping of electrode positions by FASTRAK. We store the ECG signals in *.bdf format, using ActiView interface provided by BioSemi. We can visualize the data using BioSemi, as well as save it, select electrodes, etc. But, we cannot read the data. For this, we utilize the MCECGreader developed by Antoun. For analysis, we first import the data using the MCECGreader. While importing, we simultaneously down-sample 64-channel data to make the whole procedure faster as well as decrease the size of the data, while retaining the whole information. After this, we perform wavelet threshold based denoising and baseline wander removal using wavelets. Then, we find out the channel having wave shape closest to the Wilson Central Terminal (WCT). We call such a channel as equivalent WCT. Now, we change the reference from CMS/DRL to equivalent WCT, by subtracting it from all other channels. This is done in order to increase the CMRR and SNR. We then detect the Fiducial points using Wavelet Transform Modulus-Maxima based approach. After this, we select the QT intervals of beats from each channel and arrange them in matrices, so that their centers are all in the same column. Finally, we align the beats using correlation of each shifted position of a beat with the average of all beats in that channel (also called as template) as a measure. For this purpose, we have also developed a faster method of performing this correlation. Rather than shifting the beats by every next sample, we now shift the beats by n positions and hence reduce net time taken for alignment. We finally remove outliers, using Hotelling's T2 technique. Here, in Hotelling's T2, we find out the distance of each point from the origin. The good points(characteristic of a particular phenomenon), form a cluster. While, the unwanted points (the outliers, not characteristic of that phenomenon), are away from this cluster. In this way, depending upon the distance of each point from the cluster, we can annotate and finally remove the outliers. This will give us a matrix for each channel, with beats aligned according to R peak and outliers removed. We, then, apply PCA on beats of each channel, to select the most varying beat from each channel, based upon the results of PCA. These will be the representative beats for the corresponding channel. Finally, based upon the standard deviation values of the channels, we do a two-dimensional interpolation throughout both sides of the torso. We do a contour plotting based upon this interpolation result. This will show the degree of variation of each region in different colors. So, in this way, we can easily identify the regions showing the most of the variation and hence the maximum change in morphology.