Xiaomin Ouyang

Research Abstract:

The prominence of mobile devices and recent breakthroughs in machine learning have enabled an emerging class of new AI-powered mobile health systems, which hold the promise of transforming today’s reactive healthcare practice to proactive, individualized care and well-being. My research focus is to build AI-powered cyber-physical systems for real-world smart health applications. My first research thrust is to develop efficient hardware sensor systems that can be rapidly deployed in real-world home environments for longitudinal health monitoring. In collaboration with the CUHK hospital, I built ADMarker (MobiSys’23), the first end-to-end system that leverages AI and multi-modal sensor devices to detect multidimensional (more than 20) digital biomarkers of Alzheimer’s Disease (AD), which has been deployed in a four-week clinical trial involving 91 elderly participants (31 with AD, 30 with mild cognitive impairment, and 30 cognitively normal subjects). My second research thrust is to design privacy-preserving machine learning systems for tackling real-world system challenges and dynamics in smart health, including limited data labels, multimodality, sensor variance, and limited computing resources. First, I focus on developing a multi-modal sensing system for characterizing complex human activities such as conversation and family meals, which are important AD digital biomarkers. We design Cosmo (MobiCom’22), a new multi-modal activity recognition system that can efficiently fuse heterogeneous sensor data with only small labeled data. Second, most of the previous health monitoring solutions are focused on the centralized learning approach, which imposes significant privacy concerns. I design several efficient Federated Learning (FL) systems for smart health applications to foster collaborative model training among distributed nodes while preserving users’ data privacy. For example, we propose ClusterFL (MobiSys’21, SenSys’21), a similarity-aware FL system for activity recognition, which automatically captures the intrinsic similarity among the data of different users to improve model accuracy and training efficiency in FL. We also propose Harmony (MobiSys’23), a new heterogeneous multi-modal FL system that harnesses the modality heterogeneity among nodes to enable distributed multi-modal sensor fusion. By addressing these challenges, my research seeks to detect digital biomarkers for personalized and private health monitoring and intervention.

Bio:

Xiaomin Ouyang is currently a postdoctoral researcher at UCLA, working with Prof. Mani Srivastava. She obtained her Ph.D. degree from The Chinese University of Hong Kong in 2023, advised by Prof. Guoliang Xing and Prof. Jianwei Huang. Her research interest is building Artificial Intelligence systems, with a primary focus on developing privacy-preserving machine learning systems for smart health applications. She also has extensive experience in deploying real-world smart health systems for Alzheimer’s Disease monitoring in a clinical trial. Her work has been published at top conferences in Mobile Computing and Internet of Things, including ACM MobiCom, MobiSys, and SenSys. She received ACM MobiSys 2023 Best Paper Award, ACM SIGBED China Best Doctoral Dissertation Award, Best Presentation Award at ACM MobiSys 2023 Rising Stars Forum, and was named as EECS Rising Stars in 2023.