03/23/2026
By Yidong Zhu

The Miner School of Computer and Information Sciences, Department of Computer Science, invites you to attend a doctoral dissertation defense by Yidong Zhu on “Investigating Wearable Sensor Signal Representation for Autonomous Health Monitoring.”

Name: Yidong Zhu
Degree: Doctoral
Date: Friday, April 3, 2026
Time: 12:30 - 2 p.m.
Location: Ball Hall 328, North Campus and via Zoom
Dissertation Title: Transforming Wearable Time-Series into Image Representations for Multimodal, Fair, and Scalable Healthcare Systems


Committee Members:

  • Mohammad Arif Ul Alam, Ph.D. (Advisor), Miner School of Computer and Information Sciences, UML
  • Yu Cao, Ph.D., Miner School of Computer and Information Sciences, UML
  • Pradeep U. Kurup, Ph.D., Department of Civil & Environmental Engineering, UML
  • Benyuan Liu, Ph.D., Miner School of Computer and Information Sciences, UML
  • Yuan Zhang, Ph.D., RN, FAAOHN, FAAN, Solomont School of Nursing, UML

Abstract: Wearable sensing technologies have the potential to transform healthcare through continuous, real-time physiological monitoring. However, their widespread adoption in machine learning pipelines is hindered by a fundamental label-scarcity problem: while wearable devices generate large volumes of multimodal time-series data (e.g., heart rate, activity, and other physiological signals), labeled annotations remain limited due to high cost, time constraints, and the need for domain expertise. Moreover, conventional transfer learning approaches based on computer vision and natural language processing models often fail to capture the unique temporal and spectral characteristics of physiological signals. 

This dissertation introduces a unified representation learning framework that bridges this gap by transforming multi-dimensional wearable time-series data into image representations. A key methodological contribution is the extension of Modified Recurrence Plots (MRP) into the frequency domain, enabling the capture of spectral structures and periodic patterns that are not observable in the time domain alone. These complementary time- and frequency-domain representations are further theoretically coupled with a modified MixUp-based augmentation strategy, yielding robust and generalizable features that consistently outperform single-domain baselines across multiple downstream tasks. 

The dissertation not only introduces scalable image-based representations for wearable data but also presents the design and deployment of a scalable mobile–cloud distributed computing infrastructure that enables real-time data acquisition, processing, and machine learning for wearable and portable sensing systems. The effectiveness of this platform is demonstrated through multiple IRB-approved real-world studies, including infant health monitoring (e.g., tummy time), behavioral sleep interventions comparing figurative and literal notification strategies, and community-driven monitoring of heavy metal exposure in drinking water, as well as multimodal OUD risk prediction through the integration of wearable and EHR data with large language models and fairness-aware wearable-based health assessment using a multi-attribute fairness scheme, collectively highlighting the system’s scalability and societal impact.

Wearable sensing technologies have the potential to transform healthcare through continuous, real-time physiological monitoring. However, their widespread adoption in machine learning pipelines is hindered by a fundamental label-scarcity problem: while wearable devices generate large volumes of multimodal time-series data (e.g., heart rate, activity, and other physiological signals), labeled annotations remain limited due to high cost, time constraints, and the need for domain expertise. Moreover, conventional transfer learning approaches based on computer vision and natural language processing models often fail to capture the unique temporal and spectral characteristics of physiological signals. This dissertation introduces a unified representation learning framework that bridges this gap by transforming multi-dimensional wearable time-series data into image representations. 

A key methodological contribution is the extension of Modified Recurrence Plots (MRP) into the frequency domain, enabling the capture of spectral structures and periodic patterns that are not observable in the time domain alone. These complementary time- and frequency-domain representations are further theoretically coupled with a modified MixUp-based augmentation strategy, yielding robust and generalizable features that consistently outperform single-domain baselines across multiple downstream tasks. The dissertation not only introduces scalable image-based representations for wearable data but also presents the design and deployment of a scalable mobile–cloud distributed computing infrastructure that enables real-time data acquisition, processing, and machine learning for wearable and portable sensing systems. The effectiveness of this platform is demonstrated through multiple IRB-approved real-world studies, including infant health monitoring (e.g., tummy time), behavioral sleep interventions comparing figurative and literal notification strategies, and community-driven monitoring of heavy metal exposure in drinking water, as well as multimodal OUD risk prediction through the integration of wearable and EHR data with large language models and fairness-aware wearable-based health assessment using a multi-attribute fairness scheme, collectively highlighting the system’s scalability and societal impact.