Learning From Alarms: A Robust Learning Approach for Accurate Photoplethysmography-Based Atrial Fibrillation Detection using Eight Million Samples Labeled with Imprecise Arrhythmia Alarms
Authors:
Cheng Ding,
Zhicheng Guo,
Cynthia Rudin,
Ran Xiao,
Amit Shah,
Duc H. Do,
Randall J Lee,
Gari Clifford,
Fadi B Nahab,
Xiao Hu
Abstract:
Atrial fibrillation (AF) is a common cardiac arrhythmia with serious health consequences if not detected and treated early. Detecting AF using wearable devices with photoplethysmography (PPG) sensors and deep neural networks has demonstrated some success using proprietary algorithms in commercial solutions. However, further advancement of this paradigm of continuous AF detection in ambulatory sett…
▽ More
Atrial fibrillation (AF) is a common cardiac arrhythmia with serious health consequences if not detected and treated early. Detecting AF using wearable devices with photoplethysmography (PPG) sensors and deep neural networks has demonstrated some success using proprietary algorithms in commercial solutions. However, further advancement of this paradigm of continuous AF detection in ambulatory settings, towards a population-wide screening use case, still faces several challenges, one of which is the lack of large-scale labeled training data. To address this challenge, in this study, we propose to leverage AF alarms from bedside patient monitors to label concurrent PPG signals, resulting in the largest PPG-AF dataset so far (8.5M 30-second records from 24100 patients) and demonstrating a practical approach to build large labeled PPG datasets. Furthermore, we recognize that the AF labels thus obtained contain errors because of false AF alarms generated from imperfect built-in algorithms from bedside monitors. Dealing with label noise with unknown distribution characteristics in this case requires advanced algorithms. We, therefore, introduce and open source a novel loss design, the cluster membership consistency (CMC) loss, to mitigate label errors. By comparing CMC with state-of-the-art methods selected from a noisy label competition, we demonstrate its superiority in multiple aspects including handling label noise in PPG data, resilience to poor-quality signals, and computational efficiency.
△ Less
Submitted 12 November, 2023; v1 submitted 7 November, 2022;
originally announced November 2022.
