Explainable artificial intelligence for the electrocardiogram

Abstract

The 12-lead electrocardiogram (ECG) is a crucial diagnostic tool in clinical practice, used to identify various cardiac abnormalities. However, interpreting ECGs is complex, with significant variability between and within clinicians, and expert knowledge is not always accessible. This has led to the development of computerized interpretation algorithms (CIE), though they have not yet matched the accuracy of physicians. Recent advancements, particularly the use of deep neural networks (DNNs), have shown potential to process raw ECG data more effectively without relying on hand-crafted features. With the availability of large labeled ECG datasets, ECGs are ideal for developing AI-based algorithms. This research aimed to bridge the gap between AI advancements and their clinical application, bringing these technologies closer to real-world use. Four key prerequisites for developing and implementing ECG-AI were identified: asking the right clinical questions, ensuring rigorous quality control, addressing potential bias using explainability techniques, and conducting appropriate implementation studies. Additionally, we highlighted four key opportunities for ECG-AI applications: disease screening, diagnostic workflow optimization, decision support in treatment, and detecting new ECG features. The first step in exploring DNNs for ECG interpretation involved developing an algorithm to optimize diagnostic workflows. This algorithm efficiently triaged ECGs, identifying which required expert consultation and when, ensuring safe and effective clinical outcomes. An implementation study in a hospital setting demonstrated the algorithm’s practical benefits. Another important aspect of AI implementation is ensuring the algorithms can express uncertainty, as DNNs tend to make predictions even with noisy or unfamiliar data. We developed a method to quantify the algorithm's confidence in its predictions, ensuring that only certain ECGs undergo automatic analysis, while others are referred to an expert. Addressing challenges in DNN usage, including the lack of explainability and the need for large datasets, we created a novel method to interpret ECGs in an explainable manner. By training a variational auto-encoder on over one million ECGs, we decomposed the ECG morphology into 32 explainable factors, termed FactorECG. This method performed as effectively as traditional "black box" models but with added transparency, offering promise in applications like detecting reduced ejection fraction. FactorECG also surpassed heatmap-based techniques in improving explainability. Given that large ECG datasets are often unavailable for certain clinical questions, we explored the transferability of DNN knowledge from large to small datasets. We developed algorithms to detect specific genetic mutations and predict disease progression in patients with various heart conditions. Remarkably, these algorithms outperformed traditional prediction models, relying only on the 12-lead ECG for input. In conclusion, while significant progress has been made in developing ECG-AI, challenges remain before it can be fully implemented in clinical practice. Future work must focus on refining the FactorECG algorithm, conducting large-scale trials, and developing software platforms to integrate these algorithms into clinical workflows. Importantly, further research is needed to assess whether ECG-AI improves patient outcomes and reduces healthcare costs and burdens.