Semester Award Granted

Spring 2025

Submission Date

May 2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Thesis/Dissertation Advisor [Chair]

Jinwoo Jang

Abstract

Creating unobtrusive, intelligent sensing systems that can be embedded into daily life and assess changes in the physical and cognitive functions of people has garnered significant attention. This dissertation presents a data-driven pattern recognition framework that integrates human-AI interaction and machine learning to analyze and quantify cognitive decline in older drivers based on the data stream of their driving data. This study harnesses real in-vehicle sensing data collected from +65 older drivers. Various AI-enabled data analytics modeling techniques are developed and tested to understand the driving patterns of older drivers, the relations between driving data and their cognitive functions, and detect cognitively impaired drivers. First, an in-vehicle sensing platform is employed to collect and preprocess telematics data from older drivers, capturing critical driving metrics such as speed, acceleration, braking, and steering behavior. Second, pattern mining techniques using unsupervised learning methods, including Self-Organizing Maps (SOM) and Deep Embedded Clustering (DEC), are applied to identify driving behavior patterns. These patterns are then analyzed to differentiate between normal aging-related behavior and behavior indicative of MCI. Third, pattern recognition and quantitative analysis are conducted to examine the relationship between cognitive decline and driving features. Statistical and machine learning models are used to assess how driving behaviors are altered due to cognitive impairment. Finally, a novel two-stage self-supervised deep contrastive learning framework is developed to detect MCI from telematics data. This framework first leverages self-supervised learning to extract meaningful driving behavior representations, followed by supervised classifiers to identify MCI cases with uncertainty quantification.

This study identifies critical driving behaviors—particularly long-term patterns involving trip frequency, nighttime and peak-hour driving exposure, and throttle control—as significant digital biomarkers of MCI, contrasting with traditional assumptions that short-term driving variability is the primary indicator of cognitive decline. This approach significantly contributes to non-invasive, scalable, and privacy-preserving cognitive health assessments, highlighting how AI-driven behavioral analytics can complement traditional clinical assessments effectively. It establishes a pipeline for early detection and personalized interventions in healthcare, transportation safety, behavioral science, and aging research.

Available for download on Saturday, April 18, 2026

Share

COinS