Assessing the risk of traffic accidents in lisbon using a gradient boosting algorithm with a hybrid classification/regression approach

Traffic accidents significantly impact public health and economy through injuries, fatalities, and property damage. Effective emergency response planning requires sophisticated risk prediction tools with precise spatial and temporal resolution. While previous studies have assessed accident risk, they typically employed coarse spatial grids that lack the street-level detail crucial for emergency operations. This research presents a novel two-stage gradient-boosting predictive model, using tree-based learning algorithms to analyze traffic accidents requiring firefighter intervention in Lisbon, Portugal. To address the inherently unbalanced nature of accident data, we developed a sequential approach: first, a classification model identifies locations with non-zero accident probability; second, a regression model quantifies accident probabilities at street level across different time periods. The resulting risk simulator enables emergency planners to recalculate accident probabilities when street characteristics or weather conditions change, providing actionable insights for resource allocation and response planning. This research contributes both methodologically, through its innovative handling of spatially imbalanced data, and practically, by delivering an operational tool that supports evidence-based emergency service management. Validation results demonstrate the model’s effectiveness in predicting high-risk locations and times, allowing for proactive deployment of emergency resources.