Enabling Accurate Cardiopulmonary Monitoring using Machine Learning and a Chest-Worn Wearable Patch

Date: 5/17/2023

Time: 12pm-2pm

Location / Metting Link: TSRB 523a / https://gatech.zoom.us/j/92139753425?pwd=NEwyZ0tmOU1VVHVldm1kNzl3YzNvQT09&from=addon#success

Michael Chan, MSE

Machine Learning PhD Student

Department of Biomedical Engineering

Georgia Institute of Technology

Committee

1 Omer T. Inan, PhD (Advisor, School of Electrical and Computer Engineering and Department of Biomedical Engineering, Georgia Tech)

2 Rishi Kamaleswaran, PhD (Department of Biomedical Informatics, Emory University)

3 Thomas Plötz, PhD (School of Interactive Computing, Georgia Tech)

4 Mozziyar Etemadi, MD, PhD (Department of Biomedical Engineering, Northwestern University)

5 Mark A. Davenport, PhD (School of Electrical & Computer Engineering, Georgia Tech)

Abstract

Non-invasive physiological signals, particularly Seismocardiogram, Photoplethysmogram, and Electrocardiogram, measured at the chest provide diagnostic and prognostic value for disease monitoring out of the clinic. Conventionally, cardiopulmonary parameters indicative of our health status such as SpO2, respiratory rate (RR), heart rate (HR), and oxygen uptake (VO2) can be estimated from these signals via algorithms based on digital signal processing (DSP) and physiological knowledge. More recently, the emergence of sophisticated machine learning (ML) models brings exciting improvements in this field. By applying black box models that work in natural language processing, computer vision, or automatic speech recognition on the target estimation tasks, we may improve the accuracy of the algorithms quickly. However, doing so may also result in a missed opportunity to leverage the conventional DSP techniques and the accumulated physiological knowledge toward the estimation of cardiopulmonary parameters. To address this, we attempt to merge DSP, physiological knowledge, and ML gradually in this dissertation. First, we start by conservatively replacing one functional block of a conventional approach with a data-driven model to estimate SpO2. Next, we further replace multiple functional blocks of conventional approaches with ML to estimate RR and HR. Lastly, we merge DSP, physiological knowledge, and ML cohesively to estimate VO2, harnessing the representation capability of convolutional neural network, leveraging the interpretability of DSP, and exploiting the additional labels available at hand and known physiological knowledge through multi-task learning. Collectively, this work demonstrates an alternative algorithmic direction to improve the accuracy of cardiopulmonary parameters estimated from non-invasive physiological signals.