Multimodal machine learning framework for outcome prediction in congenital heart disease

Alkan, Muhammet (2025) Multimodal machine learning framework for outcome prediction in congenital heart disease. PhD thesis, University of Glasgow.

Abstract

Congenital Heart Disease (CHD) affects approximately 1.2 million newborns annually world wide, with around 4,600 cases occurring in the UK each year. CHD encompasses a complex set of structural heart defects that pose challenges in early diagnosis, risk stratification, and treatment planning. Traditional methods employed for predicting clinical outcomes constrained by the pronounced anatomical and functional heterogeneity, limited number of datasets, and single-modal clinical markers, which often hinders the development of generalisable models in congenital heart diseases. Recent advancements in the field of Machine Learning (ML) and Deep Learning (DL) offer opportunities to integrate multi-modal data sources, thereby enabling a more comprehensive understanding of patient health. This thesis explores a multi-modal machine learning framework designed to improve CHD classification and outcome prediction, by integrating multi-modal data and geometric learning.

A significant challenge encountered during the course of this research is the heterogeneity characteristic of clinical data sources. Patient records contain Electrocardiogram (ECG) signals, cardiopulmonary exercise testing metrics and unstructured clinical documentation, each with different formats and level of completeness. Furthermore, the inherent anatomical and physiological heterogeneity of CHD increases the complexity of predictive performance. It is important to note that a model trained on one subtype may exhibit suboptimal performance when applied to a different CHD presentation, making generalisation across diverse patient populations a challenge. This thesis attempts to bridge these gaps by leveraging Riemannian geometry for the purpose of feature extraction, employing covariance augmentations to generate more data, and utilising multi-modal data integration to maximise predictive potential.

Risk prediction models are statistical or machine learning-based frameworks designed to estimate the likelihood of future adverse events for a given patient or population. In the domain of cardiology, these models facilitate predictions about a variety of outcomes, including the risk of mortality and the progression of the disease. This, in turn, serves to inform the development of early intervention and treatment strategies. They often rely on features extracted from clinical data, including ECGs, laboratory results, imaging data, and patient demographics to generate meaningful insights. However, developing accurate risk prediction models with small sample sizes presents several challenges such as limited generalisation, high variance, reduced reliability, and an insufficient representation of rare cases, particularly due to the low prevalence of related events and the inherent imbalances in datasets. Furthermore, models constructed solely on mortality data often suffer from significant imbalances, which can compromise their predictive performance. To address these challenges, this thesis explores the use of Cardiopulmonary Exercise Testing (CPET) as a surrogate for mortality, providing a novel approach to enhance model accuracy even with limited data. This key contribution not only aims to improve the reliability of risk predictions but also demonstrates the potential for developing robust predictive models that can better inform clinical decisions and improve patient outcomes in the CHD population.

Geometric deep learning can be defined as a subfield of machine learning that involves the utilisation of manifold-based, or topology-aware methodologies, for the extraction of features from structured data. Unlike conventional deep learning models, which assume inputs are organised in a regular format, such as image or text, geometric deep learning preserves spatio-temporal relationships and dependencies inherent in medical signals like ECGs. In this thesis, the covariance structure of ECG signals plays a fundamental role in enhancing risk prediction models, given that ECG readings exhibit correlated variations across different leads. The utilisation of covariance matrices to represent signals in Riemannian space ensures the preservation of higher-order relationships and can generate more stable and generalisable features, thereby reducing the impact of small sample sizes.

Machine learning applications in CHD research have traditionally focused on heartbeat classification, arrhythmia detection, and patient risk stratification based primarily on ECGs interpretation. While deep learning architectures have demonstrated promising results, challenges remain in model generalisability, dataset diversity, and clinical utility. This thesis explores the development of a multi-modal machine learning framework designed to incorporate a variety of clinical indices. The framework utilises multiple data modalities including medical health records and ECGs, with the objective of enhancing the precision and reliability of outcome prediction models. Furthermore, regression models are employed to assess cardiopulmonary exercise test results, providing insights into cardiac function of the patients. Text-mining techniques are also applied to extract meaningful clinical information from physician notes, enabling richer data-driven assessments of patient conditions.

By leveraging multiple data modalities, including medical health records and ECGs, this research aims to enhance the precision and reliability of outcome prediction models by providing a more comprehensive understanding of patient health. The scope encompasses the identification and digitisation of multiple data sources, the design and implementation of relevant machine learning models, and the evaluation of the framework’s performance in clinical settings. The integration of multi-modal data enhances the ability to capture complex cardiac abnormalities, thus offering a more comprehensive approach to diagnosis. The findings from this thesis contribute to the growing research on machine learning and congenital heart disease outcomes. We present a data-driven pathway for improving classification and outcome prediction, addressing key challenges such as imbalanced datasets, model generalisability and multi-modal data integration. By expanding dataset accessibility, future research can enhance the application of machine learning models in CHD, thus supporting clinical decision-making and improving patient care.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Subjects: Q Science > QA Mathematics > QA75 Electronic computers. Computer science; Q Science > QA Mathematics > QA76 Computer software; R Medicine > R Medicine (General)
• Colleges/Schools: College of Science and Engineering > School of Computing Science
• Supervisor's Name: Deligianni, Dr. Fani
• Date of Award: 2025
• Depositing User: Theses Team
• Unique ID: glathesis:2025-85428
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 03 Sep 2025 10:31
• Last Modified: 03 Sep 2025 10:33
• Thesis DOI: 10.5525/gla.thesis.85428
• URI: https://theses.gla.ac.uk/id/eprint/85428

Related URLs

• Pre-print
• Pre-print
• Conference item
• Article DOI
• Article DOI