Implementation of machine learning models bridges the prognostic gap in Aluminum Phosphide poisoning

Abstract Background Pesticide poisoning, including aluminum phosphide poisoning, is a common cause of death worldwide. Aluminum phosphide (AlP) is a highly hazardous, inexpensive pesticide with no specific antidote. Early predictions of prognosis can play a crucial role in management. This study aims to identify the key predictors of mortality in cases of acute aluminum phosphide poisoning and to explore the application of machine learning models in predicting patient outcomes. Results A prospective study was conducted on 117 patients poisoned with aluminum phosphide who were admitted to the Alexandria Poison Center (APC) at Alexandria Main University Hospital over six months. Data regarding history, clinical presentation, lab investigations, and management were collected. The male-to-female ratio was nearly 1:1. Most cases were suicidal and were administered through the oral route. 54.7% of patients survived. All non-survivors had a poison severity score (PSS) of three. Receiver Operating Characteristics (ROC) curve analysis of mortality predictors showed that the mean blood pressure had the best predictive power, with the highest accuracy (94.9%). On implementing machine learning models to predict mortality, the Artificial Neural Network (ANN) and Random Forest Metaclassifier demonstrated the lowest mean absolute error (MAE), (0.06% and 0.1%, respectively). Conclusion Mean blood pressure had the most excellent predictive power for mortality. This is the first study to demonstrate the efficacy of the Random Forest model in mortality prediction of acute aluminum phosphide-poisoned patients, achieving a high accuracy (94%) in mortality prediction.