Computational methods for augmenting the echocardiographic assessment of low-flow, low-gradient aortic valve stenosis

Illyes, Marcell (2026) Computational methods for augmenting the echocardiographic assessment of low-flow, low-gradient aortic valve stenosis. PhD thesis, University of Glasgow.

Abstract

Aortic stenosis (AS) is the most common valvular disease requiring intervention in adults in developed countries. One of its most diagnostically challenging subtypes is low-flow low-gradient (LFLG) AS, characterised by discordant echocardiographic grading criteria. Current guidelines recommend dobutamine stress echocardiography (DSE) to resolve diagnostic uncertainty; however, DSE is often inconclusive and carries procedural risks. This research investigates whether resting echocardiographic recordings alone can distinguish true-severe from pseudo-severe LFLG AS.

First, statistical analyses and classical machine learning (ML) techniques were applied to standard and novel echocardiographic predictors to assess their ability to separate severity grades and flow states. The results show that the decision boundaries derived from concordant AS cases do not generalise well to LFLG AS. Although ML models achieved 79.3% accuracy in synthetic data generated from concordant AS, performance decreased when evaluated in the LFLG AS cohort in both rest and stress states.

Second, a novel strain imaging pipeline was developed to estimate three-dimensional myocardial deformation from 2D triplanar echocardiographic data. This enabled calculation of 3D strain parameters over time without full 3D acquisitions. The global and segmental longitudinal strain values did not show significant differences between the 3D and triplanar methods, confirming the validity of the pipeline. A qualitative analysis also suggested that regional strain values were lower in patients with severe LFLG AS.

Finally, deep learning (DL) models that combine transformer-based feature extractors with multiple-instance learning were trained on a public dataset for AS severity classification. The test accuracy showed marginal improvements in the public data set with an average improvement of 2.16% over the predefined test sets, and generalisation to clinical data from patients with concordant AS achieved a balanced accuracy of 67.5%, compared to 62.5% from previous models. Additionally, the model performance in LFLG AS was assessed under resting or both resting and stress states.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Subjects: Q Science > QA Mathematics > QA75 Electronic computers. Computer science; R Medicine > R Medicine (General); T Technology > TK Electrical engineering. Electronics Nuclear engineering
• Colleges/Schools: College of Science and Engineering > School of Engineering
• Supervisor's Name: Aggarwal, Dr. Ankush and McBride, Professor Andrew
• Date of Award: 2026
• Depositing User: Theses Team
• Unique ID: glathesis:2026-86089
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 10 Jul 2026 13:21
• Last Modified: 10 Jul 2026 13:21
• URI: https://theses.gla.ac.uk/id/eprint/86089