A Novel 1D Generative Adversarial Network-based Framework for Atrial Fibrillation Detection using Restored Wrist Photoplethysmography Signals

TL;DR

This study introduces a novel 1D CycleGAN-based framework for reconstructing wrist PPG signals to improve atrial fibrillation detection accuracy, demonstrating high reliability comparable to ECG-based methods.

Contribution

First application of 1D CycleGAN for wrist PPG signal reconstruction to enhance AF detection from wearable devices.

Findings

Achieved over 96% accuracy in AF detection using reconstructed PPG signals.

Reconstructed signals improved AF detection accuracy by up to 5%.

Performance comparable to ECG-based AF detection methods.

Abstract

Atrial fibrillation (AF) increases the risk of stroke. Electrocardiogram (ECG) is used for AF detection, while photoplethysmography (PPG) is simple to use and appropriate for long-term monitoring. We have developed a novel approach to detect AF from smartwatch-based wrist PPG signals. To the best of our knowledge, this is the first study to employ 1D CycleGAN for the reconstruction of 1D wrist PPG signals, since the quality of wrist PPG signals is very poor due to motion artifacts and acquisition site limitations. Our proposed approach is validated on a dataset of 21,278 10s long wrist PPG segments. The dataset was divided into two halves and two experiments were conducted. One half was used for training and the other half for testing and vice versa. Our classification model (Self-AFNet) with 1D-CycleGAN for restoration achieves an accuracy of 96.41% and 97.09% for both splits, respectively. There is a noteworthy improvement in accuracy (2.94% and 5.08% for test splits, respectively) for restored signals as compared to unrestored PPG. AF detection task was also performed using ECG signals with the matched PPG signals to assess the validity of employing reconstructed PPG-based AF classification. Self-AFNet managed to achieve an accuracy of 98.07% and 98.97%, respectively using two ECG splits which is comparable to the performance of AF detection utilizing reconstructed PPG segments. This study reflects that reconstructed wrist PPG signals acquired from wearable devices can be used for reliable detection of atrial fibrillation, which can significantly help in the reduction of the risk of stroke.