Toward accurate health monitoring through large-scale Photoplethysmography signal from wearable devices
Author(s)
Ding, Cheng
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
This dissertation presents a comprehensive study on enhancing health monitoring through advanced analysis of photoplethysmography (PPG) signals from wearable devices. As wearable technologies proliferate, there is a significant opportunity to leverage the PPG technology embedded in devices like smartwatches for continuous health monitoring. This research addresses critical challenges in PPG signal utilization for health diagnostics, such as atrial fibrillation (AF) detection, through a combination of novel data augmentation techniques, robust machine learning models, and the development of a large-scale labeled PPG dataset. Firstly, the dissertation introduces a Generative Adversarial Network (GAN)-based technique for data augmentation to tackle the inherent class imbalance in PPG datasets used for AF detection. This approach enhances the generation of synthetic PPG signals by incorporating spectral loss adjustments, which in turn improves the performance of AF classifiers. Secondly, recognizing the limitations of synthesized data, this study compiles a novel dataset of over 8 million real-world PPG records labeled using bedside monitor alarms for AF. A robust learning method tailored to handle the label noise in this large dataset is developed, significantly boosting the accuracy of AF detection. Finally, the dissertation introduces SiamQuality, a foundational model based on a Siamese network architecture that addresses signal quality issues in PPG data. By ensuring that signals of similar physiological states yield similar feature representations, irrespective of their quality, the model sets new standards for reliability in PPG-based health monitoring applications. Collectively, these advancements represent a significant step forward in the utilization of PPG technology for health monitoring, promising to enhance diagnostic capabilities and patient outcomes in real-world settings.
Sponsor
Date
2024-07-17
Extent
Resource Type
Text
Resource Subtype
Dissertation