Machine Learning Tool Accurately Predicted 86 Percent of Incidents, With 14 Percent False Positives, on I-270 in St. Louis.
A Predictive Analytics Pilot Focused on an Area of Construction on I-270 in Missouri to Inform Emergency Response.
St. Louis, Missouri, United States
Predictive Analytics for Traffic Management on I-270 in St. Louis
Summary Information
The Missouri Department of Transportation (MoDOT) uses predictive analytics to help plan for and react to traffic crashes. Predictive analytics is the integration of real-time and historical data sources into a single platform that can be processed using machine learning for analysis and decision making. When applied to traffic safety, it can identify circumstances that lead to crashes before they occur. MoDOT conducted a predictive analytics pilot project in a heavy construction area on I-270 near St. Louis. The goal of the pilot was to predict high crash risk areas up to 24 hours in the future, helping MoDOT better monitor these areas and proactively position its emergency response vehicles.
METHODOLOGY
Starting in August 2021, MoDOT evaluated the crash prediction and incident detection algorithms quarterly for their accuracy during peak hours (6 to 9 a.m. and 3 to 6 p.m.), for fatal and serious injury crashes only. The algorithms were compared to crash data captured by MDOT’s Transportation Management Center. The algorithms were improved over time (throughout 2022 and 2023) by including more connected vehicle data feeds, additional static cameras, and St. Louis County 911 Computer-Aided Dispatch (CAD) data. Video analytics were fully operational as of April 2022.
FINDINGS
An evaluation of the effectiveness of the algorithm found the following for the month of July 2023, the last reported month of the pilot:
- The algorithm predicted 86 percent of incidents correctly (790 of 922 predictions); 14 percent (132) were false positives. This percentage of false incidents increased compared to earlier months (April 2023 had 5 percent false positives) due to a change in camera settings.
- Of 320 total identified crashes, 49.1 percent (157) were detected by the algorithm prior to all other tools including police radio, Waze, and CCTV monitoring, among others. Incidents identified simultaneously with these other tools was 8.1 percent (26 incidents).
