Convolutional Neural Network Learning Approaches for Driver Injury Severity Classification and Application in Single-Vehicle Crashes in RITI Communities

Zhang, Guohui; Yang, Hanyi; Yu, Hao; Yuan, Runze · 2024 · ROSA P / University of Alaska Fairbanks. Center for Safety Equity in Transportation (CSET)

archive: archived pipeline: cataloged verified

Get this paper ↗ (full text — opens at the source; we link to it, we don't host it)

Summary

This study addresses the challenge of analyzing traffic crash characteristics in Rural, Isolated, Tribal, or Indigenous (RITI) communities, where fatality rates are significantly higher than in urban areas. Traditional statistical methods often fail to account for unobserved heterogeneities and temporal instability inherent in crash data. To overcome these limitations, the researchers developed two advanced modeling approaches: a Fusion Convolutional Neural Network with Random Term (FCNN-R) for driver injury severity classification, and hierarchical Bayesian random parameters models to address spatiotemporal heterogeneity in crash frequency. The FCNN-R model was designed to handle both categorical and continuous variables by combining sub-neural networks with a multi-layer convolutional neural network. The researchers applied this model to seven years (2010–2016) of single-vehicle crash data. They tested various model layouts, including different CNN layer depths and the inclusion of dropout layers or regularization terms, to mitigate overfitting and improve stability. Additionally, the study employed hierarchical Bayesian random parameters models with spatiotemporal interactions using yearly county-level data on alcohol/drug-impaired driving crashes in Idaho from 2010 to 2015. This dataset covered three injury severities: minor, major, and fatal. The results demonstrated that the proposed FCNN-R model outperformed five other typical approaches in predictive accuracy. While deeper CNN layers led to premature training convergence with limited data, the inclusion of dropout layers and regularization improved the stability of variable effects. Marginal effect analysis confirmed that the FCNN-R model could uncover underlying correlations similar to traditional statistical models, offering interpretability alongside high predictive power. In the Bayesian analysis, variables such as daily vehicle miles traveled, proportion of male drivers, unemployment rate, and educational attainment significantly impacted crash frequency. Crucially, significant temporal and spatial heterogeneous effects were detected across all three injury severities, validating the necessity of incorporating spatiotemporal interactions into random parameters models. The significance of this work lies in its advancement of traffic safety analysis for RITI communities. By integrating deep learning with traditional statistical insights, the study provides a robust framework for identifying risk factors and understanding the complex, non-linear relationships in crash data. The findings support the use of spatiotemporal heterogeneity in crash modeling, offering transportation planners and policymakers more accurate tools to design targeted countermeasures. This approach enhances the ability to predict injury severity and crash frequency, ultimately aiding in the development of proactive strategies to reduce risks and fatalities in rural and isolated areas.

Key finding

The proposed fusion convolutional neural network with random term model outperformed five typical traditional approaches in predicting driver injury severity, while hierarchical Bayesian models confirmed significant spatiotemporal heterogeneity in crash frequencies across different injury severities.

Methodology

modeling

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via bulk_ingest_rosap on 2026-05-23 (6 acquisition events logged).

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• telematics crash prediction
• demographic disparities
• incidence prevalence
• induced exposure
• fatality injury trends

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Empirical Findings: crash risk outcomes
• Theoretical Contribution: computational model