Overcoming data scarcity in life-threatening arrhythmia detection through transfer learning.

Background: Life-threatening arrhythmias (LTAs) are a leading cause of death worldwide. Enhancing LTA detection in wearable monitoring systems is of great importance. One of the main challenges in building robust LTA detection algorithms is the limited availability of labeled LTA data.

Methods: We introduce an effective deep-learning algorithm for detecting LTAs from single-lead ECGs in out-of-hospital cardiac arrest applications. We address the data-scarcity issue by applying a transfer learning approach. The deep-learning model is pre-trained on a massive dataset (72'952 recordings) for rhythm classification and then fine-tuned on the target dataset with LTA events (102 recordings).

Results: Our model achieves a sensitivity of 92.68% and a specificity of 99.48%, with a granularity of 1.28 seconds, in detecting LTAs. Additionally, a confidence estimation procedure is introduced to enable emergency service pre-alerts in case of low-confidence detections.

Conclusions: Our transfer learning based approach has the potential to significantly mitigate the impact of data scarcity, advancing LTA detection in wearable monitoring systems, and supporting rapid, life-saving interventions in out-of-hospital cardiac arrest emergencies.