TIMELY at CinC 2024: Advancements in Atrial Fibrillation Detection

At Computing in Cardiology (CinC) 2024, the TIMELY project presented two important studies that focus on improving atrial fibrillation (AF) detection using wearable ECG devices and deep learning techniques. Here’s a look at their key contributions.

1. Detecting Atrial Fibrillation from Reduced-Lead Electrocardiograms of Mobile Patches Using Interpretable Features

Authors: Alexander Hammer, Boris Schmitz, Hagen Malberg, Martin Schmidt

Atrial fibrillation (AF) is a common heart arrhythmia that can be difficult to detect due to its intermittent nature. Traditional 12-lead ECGs are often impractical for continuous monitoring, so the TIMELY team explored the potential of mobile ECG patches with fewer leads.

The study analyzed 2,478 12-lead ECGs and tested how well AF could be detected with fewer leads. The results showed that even with just one lead (lead III), they could achieve an F1 score of 0.918—almost as good as the 0.961 score from the full 12-lead ECG. The study found that key features related to the P wave were most important for distinguishing AF from normal rhythms.

This research suggests that mobile ECG patches with fewer leads could offer a viable and accurate solution for long-term AF monitoring.

2. Morphology Features Self-Learned by Explainable Deep Learning for Atrial Fibrillation Detection Correspond to Fibrillatory Waves

Authors: Alexander Hammer, Hagen Malberg, Martin Schmidt

In this poster, Alexander Hammer and colleagues presented a novel approach to improving the interpretability of deep learning models used for AF detection. One of the challenges of deep learning in healthcare is understanding how models make their predictions. To address this, the team used an explainable deep learning architecture (xECGArch) that self-learns features from ECG data.

They focused on F waves, irregular waves characteristic of AF, and found that the deep learning model gave higher relevance to these waves, improving both accuracy and explainability. By making the model’s decision-making process more transparent, this approach helps to make deep learning models more suitable for clinical use.

Conclusion

The TIMELY project’s poster presentations at CinC 2024 highlight promising developments in AF detection using mobile ECG technology and explainable deep learning. Their findings suggest that reduced-lead ECGs can be highly effective for monitoring AF, and that improving the transparency of deep learning models can make them more useful in clinical practice. These innovations point to a future where heart health monitoring is more accurate, accessible, and interpretable.