Towards evidence-based fire prevention policy: Uncovering drivers of urban residential fire spread via explainable machine learning

Urban residential fires pose growing threats to public safety due to rapid urbanization and increasing building density. This study analyzes key drivers of residential fire spread using over 100,000 UK fire records and employs Bayesian-optimized XGBoost modeling with SHAP interpretability analysis. Results indicate emergency response attributes, particularly alarm delays and detection times, dominate fire spread risk, contributing 51 % to predictive performance. Fire dynamics factors, including rapid growth and kitchen fires, contribute an additional 33 %. Although building and occupancy attributes carry less overall weight, they significantly affect spread under specific conditions. The Tomek links method best addresses class imbalance by accurately identifying major fire incidents. Based on these insights, recommendations include optimizing emergency resource allocation, enhancing alarm system technologies, and developing intelligent monitoring platforms. This research provides critical evidence-based guidance to enhance early warning capabilities and improve firefighting effectiveness in urban environments. • Over 100,000 UK residential fire cases analyzed using explainable machine learning. • Emergency response time and alarm delays are key drivers of fire spread risk. • Bayesian-optimized XGBoost with Tomek links effectively addresses class imbalance. • Multi-level SHAP analysis reveals and quantifies major fire spread influencing factors. • Findings support targeted, evidence-based policies for urban fire prevention and response.