AI Tools Combining Multiple Models Improved Prediction of Nighttime Pedestrian Crash Injury Severity in North Carolina by Five Percent Using Four Years of Crash Data.
Statewide Analysis Identified Variables Correlated with Pedestrian Injury Severity in North Carolina.
North Carolina, United States
Nighttime Pedestrian Safety in Different Communities: Application of Artificial Intelligence Techniques
Summary Information
National Highway Traffic Safety Administration reported that between 2009 and 2019, pedestrian fatalities in the US increased by over 50 percent, with nighttime conditions accounting for more than 85 percent of fatal incidents. The US Department of Transportation (USDOT) defined six unique socioeconomic indicators at the census tract level to identify areas with gaps. This study investigated the association of these six unique indicators and crash contributing factors (vehicle, driver, pedestrian, and roadway characteristics) with nighttime pedestrian crash injury severity. The study aimed to understand infrastructural, financial, and policy-related differences in communities associated with nighttime pedestrian crashes in North Carolina.
METHODOLOGY
The study integrated two datasets using census tracts: 1) pedestrian crash data for North Carolina from 2016 to 2019 (2,329 nighttime and 2,436 daytime pedestrian crashes), and 2) data compiled for the six indicators Based on multiple data sources, such as America Community Survey, Housing and Urban Development Location Affordability Index, Federal Emergency Management Agency (FEMA) Resilience Analysis & Planning Tool, and the 2018 FEMA National Risk Index. Collected data were aggregated to form a “burden indicator” for each indicator. The study used an analytical model (Ordered Logit model) to highlight the key variables correlated with pedestrian crash injury severity. For prediction, it adopted the “stacking” approach which combined the predictions of four base models (Ordered Logit, Decision Tree, Random Forest, and Gradient Boosting Model).
FINDINGS
- The stacking approach improved predictive accuracy by five percent points (78.85 percent) compared to the best-performing base model (73.56 percent).
- Variables associated with pedestrian crash injury severity were as follows:
- The transportation indicator and the economy indicator were both significant for nighttime and daytime crashes.
- A multilane road was significant for nighttime and daytime crashes.
- Driver/pedestrian alcohol use was significant for nighttime and daytime crashes, with notably stronger associations for nighttime crashes.
- Pedestrian crashes in rural localities, on roadways with posted speed limits between 60 and 75 MPH, involving pedestrians in travel lanes, male drivers, crashes with SUVs, and crashes with heavier vehicles showed significant effects in both models but had a stronger impact in nighttime crashes.
- Intersections and foggy weather were significant for the nighttime model.
