A data science approach to understanding heart failure

Abstract

Background: Heart failure (HF) is a complex clinical syndrome with increasing global prevalence. Despite advances in evidence-based pharmacotherapy, HF remains associated with high morbidity and mortality. Aim: This thesis examines the epidemiology of HF, factors influencing its prevalence and outcomes, and the application and limitations of machine learning (ML) methods in HF datasets. Methods: A systematic literature review of ML applications in HF was conducted following PRISMA guidelines. The Cross-Industry Standard Process for Data Mining (CRISP-DM) framework guided four studies incorporating exploratory data analysis(EDA), unsupervised ML (k-means clustering), survival analysis, and supervised ML. Analyses were performed on aggregated open data, patient-level datasets, and electronic health records from the Southern Health and Social Care Trust, Northern Ireland. Results: The literature review identified limited clinical expert involvement and frequent overstatement of ML utility. Study 1 demonstrated increasing HF prevalence in Northern Ireland alongside declining coronary artery disease, with higher rates in rural areas and regions distant from percutaneous coronary intervention centres. Study 2 highlighted the importance of domain experts in feature selection and interpretation during clustering analyses and proposed a checklist to support clinician engagement. Study 3 (n=5121) showed 1-, 5-, and 10-year survival of 81%, 53%, and 35%, respectively. Patients referred to specialist HF outpatient services by non-cardiology teams had significantly worse survival after adjustment for age and comorbidities (HR 1.40, 95% CI 1.28–1.53) and lower uptake of evidence-based pharmacotherapy. Study 4 demonstrated limitations of ML in mortality prediction, identifying a performance ceiling of approximately 70% and challenges related to overfitting and model interpretability. Conclusion: This thesis demonstrates the value of clinically curated EHR data and domain-led analytics, while highlighting limitations of current ML approaches in HF and the importance of Learning Health Systems, interdisciplinary collaboration, and data literacy. The findings support a call for change in HF service design and delivery to bridge the gap between primary and secondary care.

Date of Award	May 2026
Original language	English
Sponsors	Public Health Agency Northern Ireland
Supervisor	Raymond Bond (Supervisor), Pardis Biglarbeigi (Supervisor), David McEneaney (Supervisor), Patricia Campbell (Supervisor) & Dewar Finlay (Supervisor)