Introduction: Acute aluminum phosphide (AlP) poisoning is associated with high mortality. Early identification of high-risk patients remains challenging. Machine learning may enhance risk stratification by integrating complex clinical and laboratory patterns.

Methods: Two-year retrospective study included 252 adult symptomatic patients with acute AlP poisoning to develop and evaluate machine learning ensemble models for predicting mortality. Demographic, clinical, electrocardiographic, and laboratory data obtained at admission were extracted from medical records. Four ensemble machine learning models were developed: Gradient Boosting, HistGradient Boosting, Random Forest, and Voting classifier. Data preprocessing, feature selection, correlation, feature importance analysis, and K-fold stratified cross-validation were performed. Models’ performance was evaluated using accuracy, precision, recall, F1 score and Receiver Operator Characteristic curve analysis. The final model was calibrated, and decision curve analysis was performed.

Results: Mortality rate was 65.9%. Gradient Boosting demonstrated best overall predictive performance (accuracy 0.84, F1 score 0.88, AUC 0.86). The four models demonstrated comparable areas under the curve. Across models, mean arterial pressure, serum bicarbonate, oxygen saturation, and Glasgow Coma Scale were consistently the most influential predictors. Cross-validation of Gradient Boosting model showed performance plateauing after top 4 features. Evaluating final Gradient Boosting model based on top 4 features revealed accuracy was 0.86, precision 0.93, recall 0.85, F1 score 0.89, and AUC 0.94.

Conclusion: Our findings highlight the potential of utilizing interpretable machine learning ensemble models in risk stratification and decision making in acute AlP poisoning. Gradient Boosting offers an efficient tool linking hemodynamic instability, metabolic acidosis, impaired oxygenation, and neurological status to mortality.