Combining SHAP and iBreakDown to Interpret XGBoost-Based Future Condition Predictions of Steel–Concrete Composite Box-Girder Bridges

Abstract As the number of road bridges increases to meet transportation demands, maintaining this aging network requires a substantial financial burden. Although modern inspection technologies enable precise diagnostics, their costs limit scalability. Thus, cost-effective models that estimate bridge damage indices using minimal data are required. This study presents machine-learning models that predict the future conditions of steel–concrete composite box-girder bridges by integrating publicly accessible datasets, including meteorological observation data and road traffic information, with essential bridge specifications. Among several candidate machine-learning models, the extreme gradient boosting model was identified as the best-performing model owing to its superior predictive capability. To enhance model transparency, an integrated interpretability framework combining Shapley additive explanations and iBreakDown was established to capture both individual feature contributions and their interactions. Finally, the model applicability was evaluated using three bridge data samples. The findings highlight the potential of interpretable machine-learning models to support informed bridge maintenance planning.