[PAST EVENT] Minglong Sun, Computer Science - PhD Defense

December 14, 2023
10am - 12pm
McGlothlin-Street Hall, zoom
251 Jamestown Rd
Williamsburg, VA 23185Map this location

In the field of ubiquitous computing, health-related problem analysis has gained increasing
attention. Collaborations between domain doctors and computing researchers have been
established to recognize and address health-related issues. However, accurate detection and
recognition of health-related problems remain a major challenge that attracts extensive
research efforts. Among all the research works, wearable sensors-based health-related problem
recognition is promising as it is low cost, low power, and easy to carry. This dissertation focuses
on utilizing wearable sensors to study health-related problems.
First, we introduce in this dissertation is TremorSense, a PD tremor detection system designed
to classify Parkinson's Disease hand tremors. PD hand tremors are common symptoms in all
stages of PD and have a severe influence on patients' daily quality of life. TremorSense utilizes
accelerometers and gyroscopes as wearable sensors on patients' wrists to collect data from 30
PD patients. An 8-layer Convolutional Neural Network (CNN) was designed to classify PD rest,
postural, and action tremors. The CNN model was evaluated with self-evaluation, crossevaluation,
and leave-one-out evaluation, with accuracies greater than 94% for all three
The second project introduces a PD action tremor detection method to recognize PD tremors
from regular activities. The method uses a dataset from 30 PD patients wearing accelerometers
and gyroscope sensors on their wrists. Hand-crafted time-domain and frequency-domain
features were selected and compared with existing CNN data-driven features. Multiple
supervised machine learning models were trained, including Logistic Regression (LR), K-Nearest
Neighbors (KNNs), Support Vector Machines (SVMs), and Convolutional Neural Networks
(CNNs), for detecting PD action tremors. The performance of all models using the hand-crafted
features achieved more than 90% F1 scores in five-fold cross-validations and 88% F1 scores in
the leave-one-out evaluation. Specifically, SVMs performed the best in both evaluations with
over 90% F1 scores.
Next, we review the previous research on Freezing of Gait (FoG) computing, which refers to a
sudden and short event in which a patient loses the ability to step forward, commonly
experienced by advanced Parkinson's Disease patients. Falling is possible when FoG occurs,
which has a severe influence on patients' quality of life. Wearable devices have been explored
for detecting and predicting FoG and falls in PD, but a systematic survey is still lacking in this
area. This project discusses a series of FoG challenges and future research trends, which will
contribute to further research advancement.
Finally, Automatic Dietary Monitoring (ADM) is an essential tool for health management. To
automate food tracking, we present AudioPalate, a food recognition system designed to
automatically classify different types of food based on audio data using commercialized device
Apple AirPods Pro. The system is developed and evaluated using data collected from four users,
including a variety of food types with different cooking methods. We outline the data
preprocessing steps, which involve data normalization, outlier removal, and the application of
Short-Time Fourier Transform (STFT) to extract relevant features within the frequency range of
0-500Hz. Additionally, we employ data augmentation techniques to enhance classification
performance. The core of our system is a Long Short-Term Memory (LSTM) model designed for
the classification of 20 food types. To evaluate the performance of AudioPalate, we conduct
several evaluations, including self-evaluation, leave-one-user-out evaluation, and leave-intakeout
cross-evaluation. The results demonstrate high accuracy and F1-scores exceeding 85%
across all three evaluation scenarios. These findings indicate the effectiveness of the
AudioPalate system in accurately recognizing and classifying different types of food, even when
considering variations in cooking methods. Overall, AudioPalate represents a significant
advancement in the field of automatic food recognition. The system's robust performance and
high classification accuracy make it a valuable tool for dietary monitoring and promoting
healthier eating habits.
Minglong Sun is a PhD Candidate in the Department of Computer Science at William & Mary.
His Ph.D. advisor is Prof. Gang Zhou. His research focuses on solving health-related problems
via wearable sensing leveraging machine learning and deep learning models. His Ph.D. research
has appeared in CHASE 19, CHASE 21, CHASE 23, Ubicomp 21, Smart Health 15, Smart Health
18. Previously, he received his Bachelor’s Degree in Electrical Engineering and Business
Management from Tianjin University, China in 2015.