Toward Vehicle-Agnostic Driving Signatures for Cognitive Impairment Prediction from Naturalistic Driving Data

This thesis, completed as part of the DRIVES Project at Washington University in St. Louis under the advisement of Dr. Alvitta Ottley (chair), Dr. Ganesh Babulal, and Dr. Nathan Jacobs, asks whether everyday driving behavior — captured passively through in-vehicle telematics — can help predict cognitive impairment in older adults, while accounting for differences across the many makes and models of vehicles people actually drive. Prior driving-based screening approaches often rely on signatures tuned to a single vehicle or sensor configuration, which limits their use across a real-world population.

The analysis used a final dataset of 26,968 participant-weeks of telematics data from 304 participants, combined with demographic covariates, evaluated under leave-one-participant-out cross-validation across six modeling approaches — including baseline classifiers, a domain-adversarial neural network (DANN), and a sequence-based GRU-DANN model. The best model (GRU-DANN) reached a participant-level ROC AUC of 0.599, with Random Forest close behind as the strongest non-deep baseline. The thesis's main contribution is a rigorous, vehicle-agnostic comparison of these modeling choices, alongside a candid account of the limits of naturalistic driving data alone as a cognitive-status signal — tempering optimistic claims in prior driving-based dementia-screening work while laying out a framework for future, likely multimodal, research.

Highlights

• Final dataset of 26,968 participant-weeks from 304 participants, with demographic covariates
• Six modeling approaches compared under leave-one-participant-out cross-validation
• Best model (GRU-DANN) achieved a participant-level ROC AUC of 0.599
• Vehicle-agnostic framing designed to generalize across heterogeneous vehicles and sensors
• Conducted with the DRIVES Lab (PI: Dr. Ganesh Babulal), advised by Dr. Alvitta Ottley and Dr. Nathan Jacobs

Read the full thesis on WashU Open Scholarship (opens in a new tab).