A Two-Stage Deep Contrastive and Ensemble Framework for Early Cognitive Impairment Recognition Using Vehicle Telematics
Year: 2025
Type: preprint
Abstract: The rising number of licensed older drivers aged 65 and above and the high rates of fatal crashes for this age group underscores the importance of studying their driving behavior. Additionally, growing evidence indicates that neurocognitive disorders, such as mild cognitive impairment (MCI), can impair driving abilities. Therefore, analyzing the driving behavior of older drivers with MCI can help elicit important cognitive patterns to assess MCI in the preclinical stage. The current study proposes a two-stage approach that harnesses telematics data to infer the MCI status of older drivers. In the first stage, cognitive deep contrastive learning (CDCL) extracted driving patterns from known MCI drivers, distinguishing them from non-MCI drivers. The second stage applied the model to a larger, unlabeled dataset to identify drivers potentially showing early cognitive decline. Monte Carlo dropout was incorporated to quantify uncertainty in the model's predictions, enhancing the robustness of pattern extraction. The reported mean and maximum uncertainty values for both stages are relatively low (≈0.05), indicating high confidence in the model's ability to identify MCI driving patterns. This research implemented a stacking classifier comprising random forest, XGBoost, and support vector machine as base classifiers, with logistic regression as the final estimator, where XGBoost emerged as the most influential. Shapley additive explanations analysis was subsequently applied to the XGBoost model to elucidate the impact of specific driving features on MCI predictions, where RPM and throttle positioning emerged as the most influential features. The results indicate that telematicsbased modeling with self supervised CDCL, enhanced with uncertainty quantification and supervised machine learning techniques, can improve pattern recognition and prediction between MCI and non-MCI drivers.
Institution: Florida Atlantic University
Language: English
Topic: Older Adults Driving Studies
Field: Health Professions
Domain: Health Sciences
Open Access status: gold
OpenAlex ID: W4414138404