Variations on a theme: Topic modeling of naturalistic driving data.

Area of Science:

• Data Science
• Transportation Safety
• Behavioral Science

Background:

• Naturalistic driving studies provide rich datasets for understanding driver behavior.
• Current event-based analysis methods may not capture the full complexity of driving patterns.
• Novel analytical strategies are needed to extract deeper insights from driving data.

Purpose of the Study:

• To introduce Probabilistic Topic Modeling (PTM) as an advanced analytical approach for naturalistic driving data.
• To demonstrate the utility of PTM in identifying distinct driving behavior patterns.
• To evaluate the effectiveness of PTM-derived features in a clinical context.

Main Methods:

• Probabilistic Topic Modeling (PTM), specifically Latent Dirichlet Allocation (LDA), was applied to naturalistic driving data.
• Drives were treated as documents, and speed/acceleration data were converted into symbolic words using Symbolic Aggregate approximation (SAX).
• A 20-topic LDA model was trained on data from 10,705 drives by 26 participants.

Main Results:

• The LDA model successfully clustered diverse driving behaviors into 20 distinct topics.
• Topic membership probabilities served as effective features for subsequent machine learning tasks.
• These features enabled accurate differentiation between healthy drivers and individuals with Obstructive Sleep Apnea.

Conclusions:

• PTM provides a powerful framework for analyzing complex naturalistic driving datasets.
• This approach offers a more comprehensive understanding of driver behavior compared to traditional methods.
• PTM holds potential for identifying behavioral markers of medical conditions like Obstructive Sleep Apnea.