Fractional Order Filtering Approach Towards ECG Non-Stationary Biomedical Signal Processing

Abstract

Due to the increase of desk-bound activities and sedentary lifestyle, cardiac diseases are increasing at an alarming rate especially in developing countries like India. According to the WHO data, one in four Indians died, because of cardiovascular diseases. Electrocardiogram (ECG) is the signal originated due to human heart activities. It is a record of the electrical commotion caused by depolarization and repolarization of the atria and ventricles of the heart muscles. ECGs are used to find anomalies in the heart beat which may be indicative of various cardiovascular diseases. Accurately detecting the anomalies in an ECG is the relevant issue of the medical field. Each beat of an ECG is composed of several pulses of different bandwidths (known as waves P, Q, R, S and T), and an iso-electric period which corresponds to the lapse of time between two consecutive beats. As the behaviour of an ECG waveform changes with time, so it is non-stationary, pseudo periodic in nature. ECG is a powerful tool in determining the health and functioning of the heart. Faster detection and diagnosis of the cardiovascular conditions would aid physicians to provide appropriate treatment to the patients. Proper processing of an ECG signal and its accurate detection is very much essential as it determines the condition of the heart. The analysis of an ECG signal requires the information both in time and frequency, for clinical diagnosis. The non-stationarity of an ECG is often corrupted by low and high frequency noise components like power line interference (PLI), baseline wander, electromyogram, motion artifacts, etc. These different artifacts affect the morphology of the ECG waveform, thus making its analysis a difficult job. There are many digital filtering techniques which are used in its processing in order to perform this task on computer-aided diagnosis. Finite Impulse Response (FIR) has been extensively used for ECG filtering. However, there seems to be an improvement for designing filter based upon self-convolution window concept. Hence, initial research work in this thesis aims at establishing a new filter based on Hamming window referred as Hamming Self-Convolution window (HSCW) to remove various artifacts in an ECG waveform and further authenticate the efficacy of the proposed design through simulation results as compared to conventional window-based methods, such as Hamming and Kaiser window. In the last decades, considerable focus has been paid by research community on the study of fractional-order digital differentiator (FODD). The FODD is concerned with estimating the fractional order derivatives of a signal or an unknown signal from its noisy observed data. It finds usage in fractional order control systems, signal processing, chaos and fractals, electrical networks, electromagnetic field theory. In signal processing, FODD proves to be an important mathematical tool that can give more peculiar characteristics as compared to the integer-order differentiator. Moreover, it provides an extra degree of freedom which helps in optimizing the performance based upon traditional integer-order calculus. Based on the concept of fractional order differentiation, the fractional order digital differentiator (FODD) could be designed for suitable signal processing applications. Therefore, Riesz fractional-order digital differentiator (RFODD) is explored in this thesis in context of ECG signal denoising and R-peak detection. ECG waveforms are analysed by varying fractional-order RFODD and their performance is estimated. The simulation studies conducted on MIT-BIH arrhythmia database (MIT BIH-AD) indicates the superiority of the proposed method against well-established state of-the-art methods. After establishing an efficient method based on RFODD for pre-processing of an ECG signal, further research work aims at virtual-based implementation of the proposed fractional-order filter. For this, a classification system that includes the acquisition and pre-processing of real-time ECG signals, feature extraction, and classification of ECG beats into normal (subjects who have never had any heart ailment history) and abnormal (subjects who have had heart ailment history) classes using adaptive K nearest neighbor (ADkNN). The various real-time ECG subjects’ waveforms acquired are analyzed in the proposed research. After preliminary processing based on RFODD, features are extracted and classification is performed. The performance parameters like sensitivity, positive predictivity, and accuracy of ADkNN are calculated in the proposed method. From the simulation studies conducted, it is seen that the proposed method performed well in terms of sensitivity, accuracy, and positive predictivity of the classifier. Furthermore, to make a proposed classification system more generalized, robust and include more number of ECG beats classes, it is extended to the classification of MIT BIH-AD. It is used to classify ECG beats of MIT-BIH-AD into six categories. The traditional methods for arrhythmia classification mainly includes intra-patient criterion that may not befit the inter-patient criterion. The research carried out in the thesis focuses both on intra and inter-patient criteria for classification. The simulation studies conducted on MIT-BIH-AD for classification proves that the obtained results in both criteria have outperformed other methods in the literature and the proposed work exhibits better results for minority class of MITBIH-AD in both the scheme. Hence, the study carried out in this thesis has unfolded significant advantages of self-convolution window, fractional-order calculus for ECG processing applications and classification of real-time subjects and MIT-BIH-AD. Hence, the study opens up a new paradigm of research to include more number of classes for real-time ECG classification.