DISSERTATION DEFENSE
TITLE: DEEP NEURAL NETWORK REPRESENTATIONS OF PHYSIOLOGICAL TIME-SERIES SENSOR DATA FOR IMPROVED RECOGNITION PERFORMANCE
Presenter: Lee Hinkle
Advisor: Dr. Metsis
Date: Wednesday, April 5, 2023
Time: 12:00 Noon
Location: Comal 212
Abstract:
Time-series data from non-invasive body-worn sensors offer valuable insights into human health and wellness, with applications in promoting healthy behaviors, monitoring activities of daily living, and determining emotional states. However, the classification of physiological data from sensors such as smartwatches, smartphones, EEG, and ECG devices faces numerous challenges, including poor signal quality, time-series nature of the data, lack of data format and model evaluation standards, difficulty in labeling data, and the need for multiple sensors with varying sampling rates and information content.
This dissertation addresses these challenges by presenting an end-to-end data pipeline, which includes a data collection structure, a novel extract-transform-label-model (ETLM) data pipeline for sensor data, and an evaluation methodology suited to the problem. A semi-supervised technique for improving labeling is introduced, along with the design of deep-learning neural network classifiers utilizing both engineered and learned features. Additionally, a data fusion approach is proposed to handle multi-modal sensor data effectively. The overall objective of this work is to enhance the accuracy and utility of physiological sensor data in various applications, ultimately contributing to improved human health and well-being.
Deadline: April 7, 2023, midnight