A Machine Learning and Explainable AI Framework for Detecting Preclinical Alzheimer’s Disease from Naturalistic Driving Behavior

Sai Santosh Reddy Danda, Alper Kursat Uysal, Yi Murphey, Amanda Maher, Carol Persad, Savannah Rose, Robert Koeppe, Bruno Giordani

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Alzheimer’s disease, amyloid status, naturalistic driving, Bayesian hyperparameter optimization, explainable artificial intelligence

TL;DR: This study uses machine learning and explainable AI to classify cognitively normal older adults as amyloid-positive or negative based on driving patterns in their naturalistic driving behavior.

Abstract: Alzheimer's disease (AD) is a progressive neurological disorder that primarily affects older adults, with amyloid-beta accumulation serving as a key biomarker of early pathological change. Timely detection of amyloid positivity is critical for enabling early interventions and improving clinical outcomes. However, current diagnostic tools such as PET imaging or CSF analysis are expensive, invasive, and not scalable for widespread screening. In this paper we present TADEC (Trip And Day Level Explainable Classification), a novel data-driven classification framework for automatically identifying amyloid-positive and negative individuals based on their naturalistic driving behaviors. TADEC employs two levels of behavioral modeling Trip-Level, based on individual driving trips, and Day-Level, which aggregates driving behaviors among all the trips taken in the same day, and an explainable AI (XAI) component for generating explanations of prediction results. A comprehensive set of vehicular features categorized into acceleration and jerk patterns, speed behavior, and turn dynamics were extracted. The algorithms in TADEC were implemented, trained, and evaluated using a data collected from 30 amyloid-positive and 35 amyloid-negative consensus-diagnosed cognitively normal older adults. The Trip-Level model achieved the highest classification accuracy of 89.23%, while the Day-Level model reached 86.15%, indicating strong predictive performance and demonstrating that both fine-grained and broader behavioral patterns contribute effectively to classification performances. For the Day-Level approach, average lateral and longitudinal acceleration emerged as the most impactful predictors, whereas trip distance and the number of large negative jerk events were most influential in the Trip-Level approach. TADEC offers a promising, non-invasive, and scalable tool for early detection of preclinical Alzheimer’s disease using naturalistic driving behavior.

Track: 1. Biomedical Sensor Informatics

Registration Id: Q6N34WT4SQ7

Submission Number: 163