Leveraging Self-Supervised Learning Methods for Remote Screening of Subjects with Paroxysmal Atrial Fibrillation.

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Research Methodology
• Cardiology and Electrophysiology

Background:

• Clinical research often requires extensive labeled data, posing challenges due to privacy regulations and expert annotation needs.
• This data bottleneck hinders the investigation of novel clinical questions, particularly with limited patient cohorts.
• Artificial Intelligence (AI) methods, while powerful, are constrained by these data limitations.

Purpose of the Study:

• To explore the application of Self-Supervised Learning (SSL) for preliminary clinical study results with small cohorts.
• To assess SSL's efficacy in screening for Paroxysmal Atrial Fibrillation (P-AF) using single-lead ECG signals during normal sinus rhythm.
• To compare SSL performance against supervised learning in limited data scenarios for P-AF detection.

Main Methods:

• Utilized state-of-the-art Self-Supervised Learning (SSL) techniques.
• Applied SSL to single-lead electrocardiogram (ECG) data from remote monitoring during normal sinus rhythm.
• Compared SSL performance with traditional supervised learning approaches on a limited cohort dataset.

Main Results:

• Self-Supervised Learning (SSL) outperformed supervised learning in detecting Paroxysmal Atrial Fibrillation (P-AF) from limited ECG data.
• SSL prevented misleading conclusions often associated with supervised methods in small cohort settings.
• Demonstrated effective P-AF detection from normal sinus rhythm recordings captured by wearable devices.

Conclusions:

• SSL offers a robust solution for preliminary clinical studies requiring minimal labeled data.
• Researchers can leverage SSL to assess clinical question feasibility before large-scale data collection.
• The findings support scalable population screening for P-AF via wearable ECG monitoring, enhancing early diagnosis and intervention.