Related Experiment Video
Updated: Jul 3, 2025
09:17
High-Resolution Endocardial and Epicardial Optical Mapping in a Sheep Model of Stretch-Induced Atrial Fibrillation
Published on: July 29, 2011
14.8K
Continuous Atrial Fibrillation Monitoring From Photoplethysmography: Comparison Between Supervised Deep Learning and
Pavel Antiperovitch1, David Mortara1, Joshua Barrios2
1Division of Cardiology, Department of Medicine and Cardiovascular Research Institute, University of California-San Francisco, San Francisco, California, USA.
JACC. Clinical Electrophysiology
|February 10, 2024
Summary
Deep neural networks (DNNs) show promise for continuous atrial fibrillation (AF) monitoring using photoplethysmography (PPG) signals. DNNs analyze more data than traditional methods, even with motion artifacts, improving AF detection accuracy.
Area of Science:
- Cardiovascular Technology
- Biomedical Signal Processing
- Artificial Intelligence in Healthcare
Background:
- Continuous atrial fibrillation (AF) monitoring via photoplethysmography (PPG) is hindered by motion artifacts and infrequent sampling in wearables.
- Current consumer wearables often limit continuous AF analysis to rest periods, restricting monitoring capabilities.
Purpose of the Study:
- To compare a signal processing (SP) heuristic with a deep neural network (DNN) for continuous AF monitoring using PPG in free-living patients.
- To evaluate the performance and data classification capabilities of both SP and DNN methods under real-world conditions.
Main Methods:
- Collected 4 weeks of continuous PPG and electrocardiography data from 204 free-living patients.
- Developed and validated both SP and DNN models on holdout and external validation datasets.
Main Results:
- Both SP and DNN models achieved high accuracy (AUC ~0.97), with DNN showing slightly better performance (AUC 0.973 vs 0.972).
- DNN classified significantly more data (95% vs 62%), demonstrating superior motion artifact tolerance.
- External validation confirmed DNN's superior performance and data analysis rate across different populations and PPG sensors.
Conclusions:
- Deep neural networks (DNNs) offer comparable or superior performance to traditional signal processing (SP) for continuous AF monitoring.
- DNNs' ability to classify more data, including signals with motion artifacts, makes them a promising tool for enhanced wearable-based AF detection.
- The findings suggest DNNs may be better suited for robust, continuous PPG-based AF monitoring in real-world settings.
