Abstract: To address the limited physical interpretability of black-box machine-learning models in geohazard prediction, this study integrates extreme rainfall data to construct an interpretable debris-flow hazard prediction framework and quantitatively identify sensitive thresholds of key predisposing factors. The national key debris-flow prevention and control area in Dongchuan, Yunnan Province, was selected as the study area. A genetic algorithm was used for feature selection, and the Isolation Forest algorithm was used for sample purification. Three gradient-boosting models (XGBoost, LightGBM, and CatBoost) were constructed, and Bayesian optimization was applied to determine optimal hyperparameters. The SHAP framework was then introduced for global and local interpretability analysis to examine the nonlinear coupling effects of extreme rainfall factors with topographic and surface-deformation factors on debris-flow risk. All three models achieved test-set AUC values above 0.84. LightGBM performed best (AUC = 0.860), followed by XGBoost (0.858) and CatBoost (0.842). SHAP analysis showed that: (1) rainfall frequency (heavy-rainfall days > 18 d/a; consecutive 5-day rainfall events > 28 times/a) is the dominant driver, with a marginal contribution substantially greater than that of rainfall intensity; (2) slope angles of 22°-28° form a gravity-sensitive interval with pronounced topographic amplification under high-frequency rainfall; and (3) a surface deformation rate > 9 mm/a can serve as an early-warning threshold for source-area instability. This study transforms black-box prediction models into mechanism-interpretable tools. The extracted long-term climatic background thresholds provide quantitative support for regional background risk assessment and a methodological basis for building a hierarchical prevention and control system combining "background risk assessment" with "event-triggered early warning".