Building-specific seismic damage prediction of Italian historical churches: A gradient boosting approach on a large empirical database

Italian historical masonry churches represent an extraordinary cultural heritage exposed to significant seismic hazard and requires suitable protection. In this context, while fragility curves and vulnerability indices remain the standard tools for territorial risk assessment, they rely on population-based classifications and cannot predict damage at the individual building scale. This work presents a data-driven framework for building-specific seismic damage prediction based on a LightGBM regressor trained on post-earthquake A-DC survey records from seven seismic sequences spanning ten Italian regions. Two damage targets are addressed in parallel: a continuous damage index BE and a discrete EMS-98 damage level BG. The model achieves satisfactory predictive performance on a stratified held-out test set, representing a substantial improvement over a baseline predictor. A SHapely Additive exPlanation SHAP interpretability analysis reveals a physically coherent feature importance hierarchy dominated by peak ground acceleration, followed by maintenance state, apse presence, nave height, and pre-existing damage. The persistent importance of the apse across the full multi-event dataset is an essential outcome that extends the qualitative observations of post-earthquake field surveys; moreover, regional residual analysis confirms generalisation across most represented regions. The proposed approach is complementary to traditional fragility-based methods: while the latter remain appropriate for territorial risk assessment, the data-driven model provides building-specific damage estimates that can support inspection prioritisation, preventive maintenance planning, and resource allocation for seismic risk mitigation.