Interpretable prediction of mortality in zinc phosphide poisoning using combined statistical and machine learning models: a retrospective observational study

Abstract

Abstract Zinc phosphide poisoning is associated with high mortality due to rapidly evolving multi-organ dysfunction, yet reliable early risk stratification remains limited. This retrospective observational study evaluated the feasibility of combining conventional statistical modelling with interpretable machine learning (ML) for prediction of in-hospital mortality. A total of 290 adults with confirmed zinc phosphide poisoning admitted to a tertiary care hospital between January 2016 and November 2022 were included. In-hospital mortality occurred in 88 patients (30.3%). Demographic, physiological, and laboratory variables recorded within the first 48 h of hospitalization were analysed. Multivariable logistic regression was performed to identify independent predictors of mortality. Subsequently, ML models (logistic regression, Random Forest, and Extreme Gradient Boosting [XGBoost]) were developed using repeated stratified cross-validation, and model discrimination was assessed using the area under the receiver operating characteristic curve (AUROC), calibration analysis, and bootstrap-derived confidence intervals. Explainable artificial intelligence using SHapley Additive exPlanations (SHAP) was applied to the best-performing model. Serum lactate remained the only independent predictor of in-hospital mortality in the multivariable logistic regression analysis. Among the ML models, XGBoost achieved the highest numerical AUROC (0.745; 95% confidence interval, 0.680–0.807), although differences between models (DeLong’s test) were not statistically significant. SHAP analysis identified bicarbonate, lactate, potassium, creatinine, and activated partial thromboplastin time as the largest contributors to the mortality predictions generated by the optimized XGBoost model. These findings suggest that interpretable ML may complement conventional regression by providing transparent, individualized mortality prediction while preserving clinical interpretability. However, these exploratory findings require prospective multicentre validation before routine clinical implementation.