Introduction Acute aluminum phosphide (AlP) poisoning has accounted for most of the reported lethality caused by acute poisoning. Machine learning (ML) is a rapidly evolving tool focusing on risk assessment in various medical conditions. This study aimed to establish ML-based models determining the risk of in-hospital mortality in patients exposed to AlP.

Methods This retrospective cohort study included 379 acutely AlP-poisoned patients, analyzing history, clinical manifestations, laboratory investigations, and outcome.

Results Four ML models were established [Logistic Regression, Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), and Random Forest]. Agitation, pH, serum bicarbonate, oxygen saturation, and central venous pressure were among the high-weight features across all models. Repeated five-fold cross-validation demonstrated good discrimination across all four models [area under curve (AUCs) of 0.84–0.91], while hold-out validation showed high discrimination (AUCs: 0.91–0.95). Although SVM achieved the highest observed hold-out AUC (0.95), model rankings differed between cross-validation and hold-out testing. The SVM model had the highest F1 score and accuracy on validation. It could correctly predict 54 deaths (out of 60) and 45 survivors (out of 52). However, the accuracy of Random Forest model on validation dropped from 99% to 86%. The SVM model offered good calibration at lower and very high probabilities.

Conclusion This study provides four ML-driven models that can be used to forecast the risk of AlP-induced in-hospital mortality, and thus facilitate decision-making and prioritize resource utilization, particularly in resource-constrained settings. However, external validation and prospective evaluation are required before their implementation in clinical practice.