Crowdsourced Traffic Data As an Emerging Tool to Monitor Car Crashes.

Sean D. Young,Wei Wang,Bharath Chakravarthy

DOI: https://doi.org/10.1001/jamasurg.2019.1167

2019-01-01

JAMA Surgery

Abstract:Crowdsourced Traffic Data as an Emerging Tool to Monitor Car CrashesMore than 100 deaths daily and 2.5 million emergency department visits each year result from motor vehicle crashes, making motor vehicle crashes a leading cause of death in the United States. 1,2Reducing ambulance and emergency department treatment response time for crash patients could markedly save lives. 3We evaluated whether data from the crowdsourced traffic application Waze (Google) might be used to detect motor vehicle crashes faster than current reporting methods.Methods | We obtained data on California motor vehicle incident reports (eg, crashes, road hazards, and weather conditions) from June 12, 2018, to August 1, 2018, from the California Highway Patrol (CHP) (https://media.chp.ca.gov/) and the crowdsourced traffic reporting application Waze (https://www.waze.com/).Our analysis focused on the 406 559 Waze user incident reports and the 7776 CHP crash reports (of 18 582 total CHP reports).The University of California, Los Angeles Institutional Review Board waived this study because data are anonymous and aggregated.A support vector machine learning model was created based on a subset of the data (ie, training phase) to identify Waze user-reported alerts corresponding to each CHP-reported crash (ie, crash alerts).The features of the model included report time, location, type of incident reported, and user confidence in the reports.Using 10-fold cross-validation, we assessed precision (positive predictive value) and recall (sensitivity) on the remaining subset of the data (ie, test phase) to determine how well the model estimated CHP-reported crashes.We report results on the mean lead time that Waze reports occurred before CHP reports for each crash broken down by the order of Waze alerts received (eg, first, second, third Waze reports).Analyses were conducted using Python software, version 2.7 (Python Software Foundation).