Spatial Crash Typology and Risk Prediction

2024–2025 · Interpretable ML to classify crash typologies across census tracts and forecast risk

This project applies UMAP and Gaussian mixture models to classify 2,480 census tracts into crash typologies using state-level crash records in New England.

Research Objectives

The goal is to develop a data-driven framework for understanding spatial heterogeneity in crash patterns and their relationship to roadway characteristics and socioeconomic factors. This work enables more targeted and effective transportation safety interventions by identifying distinct crash typologies across census tracts.

Methodology

Data Collection and Processing

  • Analyzed one year of state-level crash records across Massachusetts, Connecticut, and Vermont
  • Examined 2,480 census tracts across the region
  • Integrated crash data with demographic, roadway, and built environment variables

Dimensionality Reduction

  • Applied UMAP (Uniform Manifold Approximation and Projection) to extract latent spatial structures from high-dimensional crash data
  • Preserved important local and global relationships in crash patterns
  • Reduced feature space while maintaining interpretability

Classification and Prediction

  • Used Gaussian Mixture Modeling (GMM) to classify census tracts into distinct crash typologies
  • Built interpretable predictive models using gradient boosting to forecast crash risk by typology
  • Generated spatial visualizations showing geographic distribution of crash patterns
  • Identified roadway and demographic characteristics associated with each typology

Key Results

  • Identified distinct crash typologies linked to specific roadway characteristics and socioeconomic factors
  • Enabled targeted interventions by classifying tracts into actionable risk categories
  • Demonstrated scalability of framework for transportation agencies nationwide
  • Provided interpretable predictions that inform evidence-based safety planning

Impact

This research moves transportation safety analysis beyond one-size-fits-all approaches toward context-specific, data-driven interventions. The framework helps agencies prioritize safety investments in high-risk areas based on crash typology-specific characteristics.

Funding

New England University Transportation Center (NEUTC), 2024-2025 (PI: J. Oke)

Publications

A Roadway Crash Typology of Census Tracts Enables Targeted Interventions via Interpretable Machine Learning (2025)

  • Published in Data Science for Transportation, Volume 7, Article 14
  • doi: 10.1007/s42421-025-00128-2

Conference Presentations

  • “Enhancing Road Safety: A Data-Driven Spatial Typology of Crashes in New England” - INFORMS Annual Meeting, Seattle, WA (October 2024)
  • “A Spatial Typology Analysis of Crash Characteristics across 2480 Census Tracts” - TRB Annual Meeting, Washington, D.C. (January 2025, poster)
  • “A Roadway Crash Typology Enables Targeted Interventions” - NEUTC Symposium & Leadership Summit, Norwich University, VT (February 2025, poster)