Multi-modal machine learning for predicting amyloid positivity using on-ramp driving.

Abstract

Introduction: Early detection of amyloid p is critical for Alzheimer's disease (AD) risk identification. This study leverages machine learning of multi-modal attributes, including vehicular, physiological, and demographic data, to classify older adults with and without amyloid positivity.

Methods: Driving data and physiological responses from 53 cognitively normal older drivers with known positron emission tomography amyloid status were collected during freeway on-ramp, merging, and post-merge stages of a fixed-course drive. Statistically significant features (P ≤ 0.05) were used to train random forest and XGBoost classifiers to classify amyloid-positive and -negative participants, with feature importance evaluated based on model performance.

Results: Integrating multiple data modalities (demographics, vehicular, and physiological features) improved classification performance, distinguishing amyloid status. XGBoost with all statistically significant features achieved the highest accuracy (85.1%). Vehicular data provided the most predictive power, highlighting driving behavior relevance for classification.

Discussion: Results underscore the importance of complementary insights from on-ramp multi-modal data to predict amyloid status and potential early AD detection.

Highlights: We analyzed driving behavior and physiological signals for cognitive decline detection.Artificial intelligence (AI) models (random forest, XGBoost) effectively classified amyloid beta positive and negative participants.Interpretable AI identified on-ramp driving, that is, ZOI_1, as key for classification.Multi-modal analysis during on-ramp driving aids early cognitive decline detection.Challenging traffic environments enable non-invasive cognitive health monitoring.