Using Topic Modeling to Develop Multi-level Descriptions of Naturalistic Driving Data from Drivers with and without...

Area of Science:

• Data Science
• Transportation Safety
• Sleep Medicine

Background:

• Naturalistic driving data analysis often focuses on isolated events, neglecting continuous driving patterns.
• Holistic analysis of highly variable driving datasets presents a significant challenge.
• Identifying subtle variations in driver behavior is crucial for safety and diagnostics.

Purpose of the Study:

• To develop and validate a novel approach for analyzing naturalistic driving data using topic modeling.
• To identify and characterize patterns in driving behavior associated with untreated obstructive sleep apnea (OSA).
• To assess the efficacy of topic modeling in predicting driver condition from driving data.

Main Methods:

• Converted continuous driving data (speed, acceleration) into a discrete text representation ("driving words").
• Applied topic modeling to identify recurring patterns (topics) in driving behavior across 5000 trips.
• Utilized identified topics in random forest models to predict the presence of OSA in drivers.

Main Results:

• Topic modeling successfully reduced data dimensionality and captured non-linear patterns in driving behavior.
• Models incorporating 10, 15, and 20 topics demonstrated improved accuracy in predicting driver condition, with a maximum AUC of 0.73.
• Drivers with untreated OSA exhibited distinct patterns, characterized by smaller lateral accelerations at lower speeds.

Conclusions:

• Topic modeling offers a powerful and effective method for extracting meaningful insights from complex naturalistic driving datasets.
• The identified driving patterns provide a potential biomarker for detecting untreated obstructive sleep apnea.
• This approach enhances the understanding of driver behavior and its relationship to medical conditions.