Enabling Accurate Cardiopulmonary Monitoring Using Machine Learning and a Chest-Worn Wearable Patch
Loading...
Author(s)
Chan, Michael
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
The development of non-invasive instruments to monitor health parameters such as HR, RR, SpO2, VO2, blood pressure, body temperature, etc. have had a significant impact in the history of medicine. Without much pain and infection risk, doctors may gather information about a patient’s body and make better informed make informed decisions to treat the patient. The emergence of these instruments ever since the 19th and 20th century have already saved many lives and revealed many mysteries of our body. Advancements in miniaturization and digitization technologies in recent decades have further encouraged public’s interests to measure health parameters anywhere and anytime, most notably in the form factor of a wrist-worn watch. However, this convenient form factor also presents an inevitable constraint—it cannot capture cardiac vibrations continuously and hence would result in a missed opportunity to examine cardiac function more comprehensively. In contrast, another convenient form factor of a chest-worn wearable patch is a more fitting option.
The chest-worn wearable patch has been designed to capture several physiological signals such as ECG, SCG, and PPG simultaneously. However, since this measurement site is not well-perfused and less studied, it remains uncertain whether it is accurate enough for estimating health parameters. Accordingly, in this dissertation, we have developed a series of algorithms to improve and validate the accuracy of the health parameters estimated. To estimate these health parameters, DSP pipelines that consider the sensing principles and the physiology have been employed for denoising, demodulation, and feature extraction while ML models were used capture the complex relationship between the extracted features and the target variables. Conventionally, physiological knowledge was used with DSP pipelines to obtain signal representation that could be easily used with ML models sequentially. More recently, advanced ML models such as those based on DL architectures have emerged and outperformed the conventional methods owning to their expressivity to learn more salient, scalable representations of the signals, together with regression/classification simultaneously and synergistically. Nevertheless, we have noticed that these DL methods were also less interpretable and analytical and hence may not deepen our understanding of the physiological signals, the physiology, and the diseases while they become more powerful and advanced. Note that the term interpretability is referred to the ability to interpret a ML model using engineering concepts rather than to interpret it using clinical terms. To avoid this “hidden trap” of ML, we have reconsidered and tested different ways to combine DSP, physiological knowledge, and ML.
Such observation has motivated the three aims presented in this dissertation, each has followed an implicit approach to combine DSP, physiological knowledge, and ML for the estimation tasks. In the first aim, we have constructed the classical DSP pipelines for deriving health parameters and replaced one functional module with ML models for SpO2 estimation and RR estimation. In the second aim, we have replaced multiple functional modules with ML models for RR estimation and HR estimation. In the last aim, we have designed a DL architecture with DSP-inspired functional modules. Collectively, we have developed algorithms that harness the interpretability of DSP, leverage the flexibility, expressive power of DL, and exploit the data available at hand and the physiological knowledge through MTL. Overall, this work has addressed the accuracy problem of the chest-worn wearable patch by proposing a new algorithmic direction for merging DSP, physiological knowledge, and ML in a cohesive manner.
Sponsor
Date
2023-07-30
Extent
Resource Type
Text
Resource Subtype
Dissertation