Heart World Conference 2026

Abstract

Cardiac signals, including electrocardiograms (ECG) and cardiac vibration recordings, are central to diagnosis, patient monitoring, and risk assessment. Traditionally, these signals are analysed as independent time series, yet they are inherently structured. Temporal organization emerges from heartbeat morphology and rhythm dynamics, spatial structure arises from multi-lead acquisitions, and relational patterns exist across patients sharing physiological characteristics. Ignoring these intrinsic structures can reduce the effectiveness and interpretability of conventional signal processing and machine learning techniques.

This talk provides a comprehensive overview of graph-based methodologies designed to optimally exploit the structure of cardiac data for tasks such as denoising and classification. By representing relevant entities as nodes connected through similarity-based or physiologically meaningful edges, graph models offer a principled framework to capture relationships that are often overlooked in classical approaches. Within this framework, cardiac measurements are treated as signals defined over irregular domains, enabling the use of Graph Signal Processing (GSP) and graph-based learning strategies. This talk will demonstrate that explicitly leveraging the structural properties of cardiac data allows simple and fully explainable signal processing methods to achieve performance comparable to that of sophisticated deep learning models. By integrating temporal, spatial, and relational information, graph-based approaches offer a unified perspective on cardiac signal analysis, highlighting how exploiting data structure can lead to more accurate, reliable, and clinically meaningful outcomes.

Ultimately, graph-based techniques offer a flexible and powerful framework for cardiac signal analysis, delivering both interpretability and high performance while remaining closely aligned with underlying physiological mechanisms.