12/13/2023
By Yidong Zhu
The Kennedy College of Sciences, Miner School of Computer & Information Sciences, invites you to attend a doctoral dissertation defense by Yidong Zhu on "Investigating Wearable Sensor Signal Representation for Autonomous Health Monitoring."
Candidate Name: Yidong Zhu
Time: Thursday, Dec 21, 2023, 10 a.m. to noon
Location: DAN 309 and via Zoom
Committee Members:
- Mohammad Arif Ul Alam (Advisor), Assistant Professor, Miner School of Computer & Information Sciences, Biomedical Engineering and Biotechnology
- Benyuan Liu (Member), Professor, Director, Miner School of Computer & Information Sciences, UMass Center for Digital Health (CDH), Computer Networking Lab, CHORDS
- Yuan Zhang(Member), Associate Professor, Susan and Alan Solomont School of Nursing
Abstract:
Wearable sensor-based health monitoring automation holds the promise of transforming healthcare by providing continuous and real-time insights into an individual's well-being. These devices offer a non-intrusive and unobtrusive means of collecting vital health data, enabling proactive interventions and personalized care. With the ability to monitor various physiological parameters such as heart rate, activity levels, and sleep patterns, wearable sensors empower individuals to take an active role in managing their health. The standard protocol for leveraging wearable sensor data to predict health monitoring vitals through machine learning involves extensive data collection, encompassing a diverse range of individuals and scenarios. Following data acquisition, a rigorous process of preprocessing, feature representation, annotation, and the application of deep learning techniques is undertaken to extract meaningful insights and establish predictive models for health-related vital parameters. However, the implementation of deep learning necessitates a substantial volume of data that must be properly labeled by experts, referring the process both costly and time-consuming. In addressing the challenges of data collection and annotation for time-series wearable sensor data, computer vision and NLP scientists employ a pre-training technique. This technique involves training powerful computers on specific classification or embedding tasks using exceptionally large datasets. Subsequently, these pretrained models are publicly accessible, allowing Computer Vision and NLP researchers to employ transfer learning for new tasks, achieving higher efficiency and accuracy with a reduced need for annotation. However, unlike established methods in computer vision and NLP, there is currently no such pretraining approach or models available for time-series wearable sensor data. This absence is attributed to the heterogeneous nature of wearable sensors, variations in sensor data types, calibration issues, and the presence of noise, which pose unique challenges not encountered in the realms of NLP and computer vision.
This doctoral thesis proposal aims to tackle the challenges in wearable sensor processing by exploring the potential of transforming wearable sensor signals into image representations. By doing so, scientists in the field can leverage existing pretrained Computer Vision models to address health monitoring problems related to wearables more efficiently and effectively, requiring less annotated data. The first step involves the development of a novel spatiotemporal time-series data to image representation technique. This technique incorporates spatial and temporal recurrent plots into a mix-up data augmentation approach for enhanced image representation. The proposed method is then applied across various wearable sensors, including Electrodermal Activity, EEG, Heart Rate, and accelerometer. Subsequently, this representation technique is utilized for predicting wearable-based activity recognition and blood glucose levels, making use of existing pretrained computer vision-based deep learning models and transfer learning methodologies.
Moving forward, our focus is directed towards the development of a more efficient image representation platform tailored for wearable sensors, catering to a diverse array of researchers across various domains. This includes applications in water contamination prediction from cyclic voltammogram, sleep monitoring, infants' energy expenditure, and pain assessment. Moreover, our objective extends beyond mere algorithmic efficiency to include considerations of fairness in the classification processes within these domains. Our overarching aim is to contribute to advancements in wearable sensor technology, fostering innovations that hold significant implications for critical areas such as environmental monitoring, healthcare, and infant well-being.