Photoplethysmography-Based Machine Learning Approaches for Atrial Fibrillation Burden: Algorithm Development and Validation.

Abstract

Background: Atrial fibrillation (AF) burden is associated with cardiovascular events such as stroke and heart failure. Recent advancements in photoplethysmography (PPG) technology have provided new insights into noninvasive and convenient AF burden detection.

Objective: This study aimed to establish an AF burden model based on smartwatch-monitored PPG technology to track the progression of AF.

Methods: This prospective pilot study (January 2024 to January 2025) at the Chinese PLA General Hospital enrolled patients with paroxysmal AF. Participants underwent simultaneous rhythm monitoring using smartwatch PPG and 24-hour Holter electrocardiogram monitoring (the gold standard). Five PPG-derived AF burden metrics were defined: (1) ratio of AF episode duration to total monitoring time (M1), (2) ratio of AF episode frequency to total measurements (M2), (3) AF episode density (M3), (4) AF episode variability (M4), and (5) proportion of rapid ventricular rate in AF episodes (>120 beats per minute; M5). Smartwatch PPG signals were collected once per minute. Sensitivity, specificity, accuracy, precision, and F1 score were used to evaluate the PPG algorithm's AF detection capability through comparison with the gold standard (24-hour Holter monitoring). The mean absolute error (MAE) and Spearman rank correlation coefficient (rs) were used to assess the correlation between the PPG-based AF burden metrics and the gold standard.

Results: A total of 145 participants with paroxysmal AF (n=96, 66.2% male; mean age 63.28, SD 14.23 years) were included. Compared to the gold standard, the PPG-based AF burden model demonstrated a sensitivity of 91.5% (95% CI 87.9%-95.1%), specificity of 97.2% (95% CI 95.9%-98.5%), precision of 92.9% (95% CI 88.6%-97.3%), accuracy of 93.3% (95% CI 88.2%-98.5%), and F1 score of 90.5% (95% CI 86.3%-94.7%). The AF burden model exhibited strong discriminatory power in the test cohort (area under the curve=89.5%, 95% CI 89.4%-89.7%). For M1, the MAE for the model of AF episode duration as a proportion of total monitoring time was 0.0400 (P=.008), with a correlation coefficient (rs) of 0.8788 (P<.001). For M4, the MAE for the AF episode variability model was 3.9967 (P<.001), with a correlation coefficient (rs) of 0.7876 (P<.001). The MAE for the average real variability model was 4.6436 (P<.001), with a correlation coefficient (rs) of 0.8127 (P<.001). The MAE for the average AF change model was 0.3893 (P=.27), with a correlation coefficient (rs) of 0.7246 (P<.001).

Conclusions: The PPG-based AF burden model demonstrated high concordance with the gold standard of 24-hour Holter monitoring in tracking AF episode duration and variability, providing new perspectives for exploring AF progression dynamics.