The role of ECG analysis in the diagnosis of cardio-vascular ailments has been significant in recent times. Although effective, the present computational algorithms lack accuracy, and no technique till date is capable of predicting the onset of a CVD condition with precision. In this chapter, the authors attempt to formulate a novel mapping technique based on feature extraction using fractional Fourier transform (FrFT) and map generation using self-organizing maps (SOM). FrFT feature extraction from the ECG data has been performed in a manner reminiscent of short time Fourier transform (STFT). Results show capability to generate maps from the isolated ECG wavetrains with better prediction capability to ascertain the onset of CVDs, which is not possible using conventional algorithms. Promising results provide the ability to visualize the data in a time evolution manner with the help of maps and histograms to predict onset of different CVD conditions and the ability to generate the required output with unsupervised training helping in greater generalization than previous reported techniques.
TopIntroduction
ECG analysis is an ongoing domain of research since past several decades for two primary reasons. It helps to improve our ability to understand the operational mechanism of the human heart (Davie, A., Francis, C., et. al. 1996) and also help in accurate and early prediction of the onset of conditions which may give rise to the different Cardio-Vascular Diseases (CVD). The second requirement is of paramount importance as one of the most severe medical conditions in present times as it is the presence of CVDs which could lead to the death of a person(Zipes D. P. and Wellens H. J., 1998). The aforementioned domain of research has implemented a significant number of mathematical techniques that have been developed for other engineering and scientific applications(Thakor N. V. and Zhu Y.S., 1991). However, their use is to convert the pictorial description and/or time series data of the ECG to a numerical format that can be interpreted using a computer. Present research has focused on the aspect of automating the classification of the different CVDs (Stamkopoulos T., Diamantaras K., et. al. 1998; Mitra S., Mitra M., and Chaudhuri B. B., 2006; Jambukia S. H., Dabhi V. K., and Prajapati H. B., 2015), than predicting the condition(s) that lead to them. Recently, the use of Pattern Recognition (PR) techniques has shown to be invaluable in the development of these algorithms(Kelwade J. and Salankar S., 2015; Gautam M. K. and Giri V. K., 2016; Acharya U. R., Fujita H., et al. 2017). PR techniques based on supervised and unsupervised algorithms, have been demonstrated with varying degrees of success, their individual capabilities to isolate and classify the necessary CVD(Asadi F., Mollakazemi M. J., et al. 2015; Yang T., Yu L., et. al. 2017; Liao K. Y.K., Chiu C.C. and Yeh S.J., 2015). however, in order to condense the greater amount of information prevalent in an ECG graph, the preliminary Feature Extraction (FE) algorithms always tend to reduce the dimensionality of the problem, before the PR technique is invoked. These algorithms include the use of Wavelet transforms(Polania L. F., Carrillo R. E., et. al., 2015), Fourier Transforms(Lee M.-L., Nien C., et. al. 2015), fast ICAs(Barhatte A. S., Ghongade R., and Tekale S. V. (2016)), Pan-Tomkins algorithms(Agostinelli A., Marcantoni I., et. al. 2017), Hilbert-Huang transformations etc. Reduction in the dimensionality improves the computational efficiency of the algorithm to perform the desired classification although it may reduce the prediction efficiency of the same. The aforementioned FE algorithms, are best suited for use with supervised PR techniques, e.g. Feedforward Artificial Neural Networks (ANNs) and Support Vector Machines (SVMs), which requires pre-ordained target outputs in order to train the system to its required operational efficiency. The use of the same in conjunction with non-supervised techniques, e.g. K-means clustering, Self-Organizing Maps etc. is extremely limited(Garg A. R., Mathur M. K., and Singh M., 2016; Xia Y., Han J., and Wang K., 2015). Non-supervised techniques learn inherent patterns embedded within the data, without the requirements of any target outputs. Thus the use of the aforementioned FE algorithms for non-supervised techniques reduces the latters’ effectiveness, as there is a loss of information, which is required to be learnt.
Till date the majority of the research into the domain of ECG signal processing, aims to automate the classification of the different CVDs encountered till date. To aid in this research, several databases have been made available online that can be used extensively. Unfortunately, the recordings of a majority of the online databases were originally performed on tape and then digitized at a later date. This procedure has caused a loss of information, relegating these databases for morphological ECG study, analysis and development in which they have excelled. In contrast, the development of algorithms for the prediction of CVDs, has been quite limited. Medical research has shown that critical CVDs, resulting in sudden cardiac death can be predicted by studying the variations in circadian rhythms.