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

Area of Science:

• Cardiology
• Artificial Intelligence
• Wearable Technology

Background:

• Life-threatening arrhythmias (LTAs) are a major global cause of mortality.
• Accurate detection of LTAs in wearable monitoring systems is crucial.
• Limited labeled LTA data presents a significant challenge for algorithm development.

Purpose of the Study:

• To develop an effective deep-learning algorithm for LTA detection from single-lead ECGs.
• To overcome data scarcity issues in LTA detection using transfer learning.
• To enhance the reliability of LTA detection in out-of-hospital cardiac arrest scenarios.

Main Methods:

• Implemented a deep-learning model for LTA detection from single-lead ECGs.
• Utilized transfer learning by pre-training on a large dataset (72,952 recordings) and fine-tuning on a smaller LTA dataset (102 recordings).
• Incorporated a confidence estimation procedure for improved detection reliability.

Main Results:

• Achieved high performance in LTA detection with 92.68% sensitivity and 99.48% specificity.
• Demonstrated detection granularity of 1.28 seconds.
• The confidence estimation procedure enables pre-alerts for emergency services.

Conclusions:

• The transfer learning approach effectively addresses data scarcity in LTA detection.
• This method significantly advances LTA detection capabilities in wearable monitoring systems.
• The algorithm supports timely, life-saving interventions for out-of-hospital cardiac arrest.