Learning From Alarms: A Robust Learning Approach for Accurate Photoplethysmography-Based Atrial Fibrillation...

Area of Science:

• Biomedical Engineering
• Cardiology
• Machine Learning

Background:

• Atrial fibrillation (AF) is a prevalent cardiac arrhythmia requiring early detection and treatment to prevent severe health outcomes.
• Wearable devices with photoplethysmography (PPG) sensors show promise for AF detection, but large-scale labeled data is a significant challenge for continuous monitoring and population screening.
• Existing commercial solutions often rely on proprietary algorithms, limiting broader research and development.

Purpose of the Study:

• To address the lack of large-scale labeled PPG data for AF detection.
• To develop and validate a novel approach for generating large PPG-AF datasets using bedside monitor alarms.
• To introduce and evaluate a robust algorithm for handling label noise in PPG data for improved AF detection accuracy.

Main Methods:

• Leveraged AF alarms from bedside patient monitors to automatically label concurrent PPG signals, creating an 8.5 million record dataset from 24,100 patients.
• Introduced and open-sourced a novel cluster membership consistency (CMC) loss function designed to mitigate errors in noisy labels.
• Compared the CMC loss with state-of-the-art methods in a noisy label competition, focusing on performance with PPG data.

Main Results:

• Successfully generated the largest PPG-AF dataset to date, demonstrating a practical method for large-scale data labeling.
• The proposed CMC loss demonstrated superior performance in handling label noise compared to existing methods.
• The CMC loss showed resilience to poor-quality PPG signals and offered computational efficiency.

Conclusions:

• The developed approach provides a scalable solution for creating large, labeled PPG datasets crucial for advancing AF detection.
• The CMC loss is an effective and robust method for mitigating label noise in PPG-based AF detection, outperforming current techniques.
• This work facilitates the development of more accurate and reliable continuous AF monitoring systems for ambulatory settings and population-wide screening.