Machine learning based risk analysis and predictive modeling of structure fire related casualties

Abstract

We analysed over 48,000 reported structure fire incidents in Oregon that occurred from January 2012 through August 2023. The dataset includes 2136 fires that led to civilian casualties including 317 confirmed fatalities. Bagged decision tree classifiers with random forest algorithm were used to quantify the importance of factors related to socioeconomic conditions, population characteristics, structural and behavioral incident details, and local infrastructure on the severity of injuries. Our results show that the age of victims, fire service response times, and availability of working smoke or fire detectors were among the most important parameters for predicting fatal outcomes of structure fires. Furthermore, a predictive Bayesian regularized neural network ensemble classifier was developed to model the severity of casualties and project a spatial risk classification on the census block level. The network model achieves a prediction accuracy of 92.5 % for the classification of structural fire-related casualty severities. With information aggregated to the census block scale and information related to specific fire incidents removed, the retrained model based solely on spatially available data reaches an 87.6 % severity classification accuracy. As the first statewide analysis of its kind, our spatial assessment provides a useful tool for resource allocation, risk factor reduction, and safety education efforts targeted to reduce the number of serious injuries or fatalities from structure fires.