Automated Techniques for ECG Waveform Fiducial Point Recognition

Chishti, Parveen (2001) Automated Techniques for ECG Waveform Fiducial Point Recognition. PhD thesis, University of Glasgow.

Abstract

The aim of this project was to provide methodology to improve ECG wave measurements within the Glasgow Program both in the absence and presence of noise. To this end a database of 125 ECGs (including 'true' values for the fiducial points) was obtained from the Common Standards for Quantitative Electrocardiography (CSE) multilead reference library. The first technique investigated, effectively involved splitting an ECG several times and averaging the resulting waveforms (the Glasgow smoothing filter). This splitting technique was indeed tried and subsequently implemented as a filter. An initial set of 21 ECGs with 'true' duration and interval values provided was assessed. Standard deviations of the differences (estimated-'true' values) about zero were calculated for each of the P-duration, QRS-duration, PR and QT intervals. For these 21 ECGs, the standard deviations were considerably lower after appl5dng the Glasgow smoothing filter than using the Glasgow Program alone, with the greatest improvement in the P-duration. Five types of noise were added separately to each of these 21 ECGs and tested. Most estimation problems encountered by the conventional program were eliminated when the Glasgow smoothing filter was used. An alternative method investigated for fiducial point estimation was that of software based neural networks. Several networks were trained using a training set of 75 CSE ECGs to locate the corresponding QRS onsets. Each network was then tested using a validation set of 50 noisy CSE ECGs. A test set of 48 CSE ECGs was next tested. Using the validation ECGs, the best neural network accurately detected more QRS-onsets than either the Glasgow smoothing filter or the conventional Glasgow program. Based on the test set, the same networks performed poorly. It was clear that the Glasgow smoothing filter performed better than any of the neural networks. The final technique investigated was that of individualised linear templates. Individualised templates were constructed using 24 filtered (using the Glasgow smoothing filter) CSE ECGs for the P-onset, P-offset, QRS-onset, QRS-offset and T-end, respectively. Results from a test set of 88 CSE ECGs, showed that the individualised templates were extremely successful in estimating the P-onset and T-end. The Glasgow smoothing filter seemed to perform best when locating the P-offset and QRS-onset and QRS-offset. This also appeared true when the five noise types were added to these ECGs. The pattern of improvement was consistent across the different program versions for the different noise types. However, the addition of high frequency noise proved detrimental to the conventional program's ability to accurately estimate the P-onset, P-offset, QRS-onset and QRS-offset. This problem was alleviated by the use of the Glasgow smoothing filter as well as the template. Although the Glasgow Program is an invaluable ECG analysis program, the techniques described here can be used in conjunction with the Glasgow program to improve the accuracy of ECG wave measurement estimation. It is clear that the Glasgow smoothing filter is best at locating the P-offset, QRS-onset and QRS offset, while the template performs best when estimating the P-onset and T-end.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Additional Information: Adviser: P W MacFarlane
• Keywords: Medicine, Medical imaging
• Date of Award: 2001
• Depositing User: Enlighten Team
• Unique ID: glathesis:2001-76401
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 19 Nov 2019 14:44
• Last Modified: 19 Nov 2019 14:44
• URI: https://theses.gla.ac.uk/id/eprint/76401