Date of Award

Summer 8-15-2022

Author's School

McKelvey School of Engineering

Author's Department

Computer Science & Engineering

Degree Name

Doctor of Philosophy (PhD)

Degree Type

Dissertation

Abstract

Wearable devices such as smartwatches and wristbands are gaining adoption. Recent advances in technology in wearables enable remote health monitoring. However, there are challenges in exploiting wearables in healthcare applications. First, sensor readings from wearables are vulnerable to motion and noise artifacts. A robust pipeline is needed to extract reliable measurements from noisy signals. Second, while wearables support an increasing number of sensing modalities, there is a significant need to generate more clinically meaningful measurements with wearables. Finally, to incorporate wearables into clinical practice, we need to establish the link between wearable measurements and clinical outcomes, thus supporting clinical decisions. To facilitate applications of wearables in healthcare, this dissertation research exploits wearables to predict a wide range of clinically relevant outcomes from physiological measurements to mental health disorders:

Measuring respiratory rate on a smartwatch with photoplethysmography: Modern smartwatches usually lack the ability to accurately measure the respiratory rate (RR) in ambulatory settings. We presented the \textit{RespWatch}, an application that can robustly measure RR and run completely on the smartwatch hardware. RespWatch directly reads the PPG waveform from the smartwatch, and utilizes a hybrid approach with both signal processing and deep learning techniques to handle the noisy sensor signals and generate robust RR measurements. A user study involving various activities showed that our hybrid method has advantages of both accuracy and efficiency over the previous approaches.

Detecting objective and subjective stress with a commercial smartwatch: In this work, we built stress detection models with commercial smartwatches, and we compared the objective stress detection (based on the objective marker of stressor tasks) with the subjective stress detection (based on the user's subjective responses). Results showed that the generic subjective stress models have worse performance than the objective stress models. To enhance the subjective stress detection, we proposed a personalized subjective model accounting for inter-individual differences via adaptive thresholds. Our personalized approach demonstrated better performance.

Multi-task learning for randomized controlled trials (RCTs) with wearables: In this work, we exploited machine learning models in conjunction with RCTs for personalized predictions of a depression treatment outcome in which patients were monitored with wearables. We formulated the predictions in different groups from an RCT as a multi-task learning (MTL) problem, and proposed a novel MTL model specifically designed for the RCT. The MTL approach was evaluated with an RCT involving 106 patients with depression, who were randomized in a 2:1 ratio to receive the integrated intervention. Our proposed MTL model outperformed both single-task models and existing multi-task models in predictive performance. Our approach represents a promising step in exploiting RCTs to develop predictive models for precision medicine.

Predicting mental disorders with wearables among a large cohort: Depressive and anxiety disorders are among the most prevalent mental disorders and are usually interconnected. We explored detecting those mental disorders with wearables in a large public dataset consisting of more than 11,000 participants. We proposed a novel deep model that combines a transformer encoder and convolutional neural network to directly learn from the raw daily activity time-series data from the wearables. Our method achieved an area Under the Receiver Operating Characteristic curve (AUROC) of 0.717 (S.D. 0.009), demonstrating the feasibility of utilizing wearables to assist in diagnoses of mental health disorders.

Language

English (en)

Chair

Chenyang Lu

Committee Members

Tao Ju, Thomas Kannampallil, Neal Patwari, Ning Zhang,

Share

COinS