Please use this identifier to cite or link to this item: http://localhost:8081/xmlui/handle/123456789/412
Title: ARTIFICIAL NEURAL NETWORK BASED ECG CLASSIFICATION
Authors: Vijaya, G.
Keywords: ARTIFICIAL NEURAL;NETWORK BASED;ECG CLASSIFICATION;NEURAL NETWORK
Issue Date: 1997
Abstract: Electrocardiogram(ECG) is either a unipolar or a bipolar non invasive recording of the electrical activity of heart obtained with the help of electrodes placed outside the cardiac cavity. An ECG recording, can be decomposed into several cycles with each ECG cycle consisting of several waves and deflections, typically in a sequence, termed as "P-wave", QRS-Complex" and "T-wave". The amplitudes &durations of these waves and the inter-wave intervals, estimated via the processing of the ECGs on all the standard 12- leads either visually or by computer methods, provide unique diagnostic information to the medical experts regarding the patient's cardiovascular system. Although a number of techniques for processing the ECGs have been reported in literature, no single technique is wholly satisfactory in realistic clinical situation, i.e., a situation in which an ECG signal is corrupted simultaneously by several sources of noise, since, most of them generally employ some form of an algorithmic approach. The success of the reported techniques for processing ECG signals were demonstrated mostly through their processing on single lead or using the ECGs which do not belong to a standard database. There has been emerging thrust for techniques that are suitable over multi-lead and their performance evaluation on ECGs of standard data bases. In addition to enable researchers in this area to develop & test more efficient & reliable algorithms, availability of an ECG diagnostic database has been felt very much vital. In the present work, the learn and generalize approach of the artificial neural networks has been well exploited through a suitable design of a predictive neural network (PNN) which can accurately and reliably identify several morphological variations in the fundamental ECG component wave, namely the QRS wave complex, on the results of which, entire superstructure namely the ECG analysis &classification problem is based upon. With the back propagation learning paradigm the PNN has been trained using more than a hundred QRS wave complexes ofdifferent morphology sandwiched between their corresponding non QRS segments from the ECGs of an internationally standard ECG library namely the CSE library (Data Set - 3). The CSE library consists of a few thousands of ECG recordings representing a variety of pathological events on all the 15 leads namely, 3 standard limb leads (1,11 &III), 3 augmented limb leads (aVR, aVL & aVF), 6precordial chest leads (V, to V6> and 3orthogonal leads (X, Yft Z,i AQRS wave detection sensitivity of 99.11% has been achieved, when tested on a heU«ECGs tf the CSE data base over the 12 standard leads constituting atotal of i*W ECG cycles^ ,„ ,ess than 1%of the cases the PNN conld not detect the QRS complexes for reasons sue as high noise content of the ECG recording or aQRS complex do not exist at all msuch ECG cycles. In addition, every QRS complex in all the standard 12-leads as been thoroughly analyzed to find the each QRS complex's morphologtcal variattons and the wave complex fiducials namely, the QRS-onset and QRS-offset. The other component waves, such as P-wave and T-wave have also been identify through asearch carried out between consecutive QRS wave complexes by ustng the PNN identified QRS wave complex locations as atime reference. Each component^wave,. amplttude, duration, morphology and the onsets &offsets were also found. In 97% of the ECGs of the standard database, the detection has been accurate. In the remammg %of the ECGs the detection could not be possible because of the very low P- &T- amphtudes of the order of less than 0.04 mV. I„ 99 01% of the ECG recordings of the standard library, all the component wave amplitudes, onsets and offsets been found to be accurate when inspected v—v inspected- The human referee estimates of Ponset, Poffset, QRSonset, QRSoffs <an Tend 725 out of 125 multilead ECG recordings are provided in by the atlas of the database. a of them, the program estimated onsets and offsets of all those ECGs Itave been compared with the median value of the human referee estimate and ,t has - ound te deviation of the program's estimates are we., within the toierances of the standard recommendations of the CSE working party. ,n this way about 256 fundamema, ECG parameters in all (that inc.udes all the amplitudes and durattons of vanot* ECG component waves) have been estimated out of one cycle of ECG reeordmg over all the .2 standard leads. Asuccessful attempt has been also made to c.assify agiven ECGI as^either^normal or abnormal. For this purpose, using the aforesaid program esttmated ***** ECG oarameters parameters of diagnostic significance namely: cardtac ate, P-wave Z7Z, P-an.pfiU.de and P-dura,ion, PR interval R-ampfi,ude«« presence/absence of Q-wave, amplitude ft duratton of Q-wave (,f present). progression, QRS interval, overall amplitude of QRS-wave complex, Frontal Plane Angle of QRS (FPA-QRS), Ventricular Activation Time (VAT), ST segment elevation, T-wave amplitude, and so on have been computed. Through a computer assisted inspection of these parameters over all the standard 12-leads, an ECG is declared as either an normal or abnormal. From 252 fundamental ECG parameters arrived at, that constitute amplitudes and durations of all the component waves, in all 168 parameters of diagnostic significance parameters, in all, have been thus estimated. An ECG acquisition system has been successfully implemented to acquire the ECGs on all standard 12-leads with an aim to create a ECG diagnostic database. Using this system, in the first phase, more than 500 ECGs have been collected from several sources such as: the Military Hospital, Roorkee, the public health camps conducted by a team of reputed cardiologists. While collecting the ECGs from the patients, other information such as patient's age, sex, clinical observations, habits, medication, if any, and so on have been documented as they help classify the ECGs into several disease categories. Out of the 500 12-lead ECGs, a limited set of 125 12-lead ECG records have been selected for inclusion into the ECG data base. Through a 2-step ECG review procedure diagnostic opinion were sought from a panel of medical experts. When the ECGs of this diagnostic data base are subjected to analysis by the ANN based ECG wave complex identification and classification software, in 98% of the cases the results were found satisfactory. Thus, the success of learn & generalize approach of Artificial Neural Network based wave complex identification, analysis of P- & T- waves & ECG classification have been demonstrated. An overall QRS wave complex detection sensitivity as high as 99.11% over all the standard 12-leads has been achieved when tested using 1500 multi-lead ECG recordings of a standard ECG data base. All the estimated wave boundaries namely the onsets and offsets of P- wave, QRS wave complex and the end of T-wave are found to be well within the tolerances set by the CSE standards. All the 125 multi-lead ECGs of a standard ECG data base have been analyzed and interpreted using the software. The analysis and interpretation findings of the developed software are found to be acceptable to clinicians in more than 97% if ECGs of the CSE library.
URI: http://hdl.handle.net/123456789/412
Other Identifiers: Ph.D
Research Supervisor/ Guide: Verma, H. K.
Kumar, Vinod
metadata.dc.type: Doctoral Thesis
Appears in Collections:DOCTORAL THESES (Electrical Engg)

Files in This Item:
File Description SizeFormat 
ARTIFICIAL NEURAL NETWORK BASED ECG CLASSIFICATION.pdf11.2 MBAdobe PDFView/Open


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.