Reading between the lines: Artificial intelligence for ECG interpretation

Abstract

This thesis explores the development, clinical application, and implementation of artificial intelligence-based electrocardiogram interpretation (AI-ECG) in cardiovascular care. The work demonstrates that successful AI-ECG development depends not only on model architecture, but also on reliable outcome labels, harmonised data infrastructure, waveform standardisation, and integration into clinical workflows. The first part of the thesis focuses on the data and methodological foundations required for robust AI-ECG development. Structured cardiovascular outcomes were extracted from free-text echocardiography reports using natural language processing, improving label quality for supervised learning. In parallel, a FHIR-based harmonization framework was developed to align heterogeneous clinical datasets across institutions, addressing challenges in interoperability and reproducibility. To reduce variability between ECG systems and vendors, a semantic segmentation model was created for waveform delineation, beat classification, fiducial point detection, and standardized median-beat generation. This model enables more consistent downstream AI-ECG analysis and supports open benchmarking through publicly released tools and datasets. The second part of the thesis evaluates clinical applications of AI-ECG. In outpatient cardiology, AI-ECG accurately identified patients at very low risk of structural heart disease, suggesting that up to one-third of echocardiograms could potentially be deferred while missing very few clinically relevant abnormalities. These findings support the role of AI-ECG as a triage tool to improve diagnostic efficiency and reduce pressure on imaging services. In patients with muscular dystrophy, AI-ECG detected reduced left ventricular ejection fraction and predicted future decline, illustrating its potential for personalised longitudinal surveillance and risk-based follow-up. In electrophysiology, AI-ECG improved localisation of accessory pathways in Wolff-Parkinson-White syndrome compared with conventional ECG algorithms, demonstrating potential value in procedural planning and workflow optimisation. The thesis further examines the challenges of translating AI-ECG from research into routine clinical care. Explainability methods were systematically evaluated and found to provide inconsistent and potentially misleading visual explanations. The findings suggest that explainability should primarily serve as a diagnostic tool for developers to detect bias and unintended model behaviour, rather than as proof of model validity. Clinician interviews highlighted that trust in AI depends not only on transparency, but also on demonstrated clinical value, workflow integration, professional endorsement, and clear governance. Overall, this thesis demonstrates that AI-ECG can improve diagnostic efficiency, support longitudinal monitoring, and assist procedural planning when developed for clearly defined clinical use cases. Future progress will likely involve multimodal and foundation models, interoperable data infrastructures, and scalable deployment strategies. With responsible development and implementation, AI-ECG has the potential to become an integrated component of routine cardiovascular care.