Advancing cardiac health: trustworthy and practical approaches to deep 12-lead ECG analysis

The electrocardiogram (ECG) is a common diagnostic tool that is crucial for the assessment of cardiac status and is usually analyzed manually with basic rule-based algorithm support. In recent years, however, a growing number of studies have highlighted the enormous diagnostic potential of combining ECG with machine learning and deep learning techniques. Despite remarkable progress, their application still faces significant challenges and limitations, including the lack of transparency of the inner workings of these algorithms, insufficient robustness against physiological noise, or the scarcity of (labeled) data in the field of cardiology. In this work, we present approaches and methods to address these challenges. We test machine learning models, both those that operate on raw ECGs and those that take pre-extracted features as input, by analyzing their decision-making process and then comparing it to cardiologists’ diagnosis rules to ensure correct model behavior. By extracting and modeling physiological noise, we evaluate and improve the robustness of deep learning models to realistic noise. In addition, we propose methods to utilize unlabeled data via self-supervised learning paradigms for enhanced performance and improved label efficiency. Finally, we investigate state-space deep learning models that are well suited for time series data and show that incorporating patient metadata into the decision-making process leads to an astonishing increase in accuracy in the diagnosis of heart disease. Together, these approaches promote a unified vision of deep learning for 12-lead ECG classification, representing an important first step toward the real-world application of deep learning for cardiology in clinical settings.