A multi-algorithm ordinal classification framework with SHAP interaction analysis for predicting construction accident injury severity

Abstract Construction remains one of the most hazardous industries worldwide, yet existing machine learning studies on construction injury severity have largely focused on single-factor importance rankings and have rarely quantified how risk factors combine to amplify fatality probability. This study developed a unified ordinal-classification framework benchmarking seven machine learning algorithms on 22,217 construction accident records extracted from the U.S. Occupational Safety and Health Administration (OSHA) database between 2015 and 2023. Inverse-frequency class weighting was applied across all models, and a custom focal-loss objective was additionally used in the three gradient-boosting learners to address class imbalance. All models were tuned through Bayesian optimization with five-fold cross-validation, and the best model was interpreted using TreeSHAP global importance, dependence, and pairwise interaction values. CatBoost achieved the best performance (Accuracy = 0.768, Macro-F1 = 0.721, AUC-ROC = 0.862, Fatal-Recall = 0.663), improving upon the Logistic Regression baseline by 6.1 percentage points (relative + 10.1%) in fatal recall. Height-related work, fall events, fall height, company size, and construction phase emerged as the top five predictors. Two pairwise interactions, company size × construction phase and height-related occupations × warm season (Spring/Summer), produced synergistic risk associations that are invisible to single-factor analyses. The strongest interaction (company size × construction phase) yielded a joint SHAP contribution 1.8 times the linear-additive expectation. These associations were translated into four threshold-based managerial rules with empirical fatality rates ranging from 12.8 to 21.3% (1.7 × to 2.8 × the sample baseline of 7.6%), supporting post-incident severity triage, targeted regulatory inspection prioritization, and risk-rated insurance underwriting. The analytical workflow is designed to be transferable to other national accident databases, though feature-specific thresholds require recalibration to account for differences in regulatory regimes, workforce composition, and climatic conditions.