The development and implementation of automated methods for the reproducible assessment of aortic stenosis and mitral regurgitation

Abstract

Valve disease is the commonest reason for a patient to undergo repeated echocardiography, which is itself by far the commonest type of cardiac imaging. Of the various valve diseases, aortic stenosis and mitral regurgitation are the two that most commonly require intervention, typically with surgery. Echocardiographic evaluation, alongside symptoms, is the most powerful determinant of the timing of surgery for aortic stenosis and mitral regurgitation. The purpose of sequential follow-up echocardiography is to make a major contribution by accurately defining the severity of the valve lesion at any point in time. However, under blinded conditions conventional echocardiographic evaluation of these conditions has surprisingly poor test-retest reproducibility. In clinical practice this leads to indecision and wasteful clinical activity. In clinical trials this leads to the need for larger (i.e. more expensive) trials to detect the same effect size. Improvements in the precision of quantification would allow clinical services to improve the certainty of diagnostic staging and reduce the frequency of follow-up. It would allow clinical trials to successfully test treatments with fewer patients. In this thesis I demonstrate in Chapter 2 how the reproducibility of existing techniques has been misleadingly reported in the literature, compared with what is achieved even in the most stringent real-world scenarios. I then develop methods for segregating and then displaying to clinicians the sources of variability in echocardiographic measurements (Chapter 3). I apply this to the valve condition to where this is most important, namely aortic stenosis. I then develop in Chapter 4 a theoretical algorithm which should improve the test-retest reproducibility of echocardiographic assessment of aortic stenosis severity. I create an open-source software suite to perform the analysis automatically. To make it maximally applicable I designed this to operate on image streams from any model and any manufacturer of echocardiograph. Having developed the software suite, I enrolled 93 patients in clinical study of validity and test-retest reproducibility with an approximately 12-month interval between scans. To minimise the possibility of me biasing the results favourably, I engaged a colleague in another university to check the validity in a cohort of 10 patients in her hospital. Having completed my initial ambition to improve reproducibility of aortic stenosis evaluation, I explored the application of what I had learnt to mitral regurgitation (Chapter 5). This is a very different problem of physics to solve, and it seems to have a different solution. I have proposed a method of using colour M-mode traces to improve the reproducibility of measurements of effective regurgitant orifice area. I then perform a pilot validation study of the improved method at my own institution. In my thesis overall, I detect a systematic problem in the way reproducibility is described, and then focus on one valve condition (aortic stenosis). I bring together insights from cardiology, statistics, and computer science to design and build a method of improving precision evaluation of stenosis. The subsequent testing suggests that this method does indeed reduce variance about 4-fold. The method I present for decomposing and displaying the variably may have general applicability throughout clinical measurement science. The approach I have taken to these valve diseases could be taken with other conditions in the future.