Comparison of machine learning algorithms to predict intentional and unintentional poisoning risk factors

A major share of poisoning cases are perpetrated intentionally, but this varies depending on different geographical regions, age spectrums, and gender distribution. The present study was conducted to determine the most important factors affecting intentional and unintentional poisonings using machine learning algorithms.

The current cross-sectional study was conducted on 658 people hospitalized due to poisoning. The enrollment and follow-up of patients were carried out during 2020-2021. The data obtained from patients' files and during follow-up were recorded by a physician and entered into SPSS software by the registration expert. Different machine learning algorithms were used to analyze the data. Fit models of the training data were assessed by determining accuracy, sensitivity, specificity, F-measure, and the area under the rock curve (AUC). Finally, after analyzing the models, the data of the Gradient boosted trees (GBT) model were finalized.

The GBT model rendered the highest accuracy (91.5 ± 3.4) among other models tested. Also, the GBT model had significantly higher sensitivity (94.7 ± 1.7) and specificity (93.2 ± 4.1) compared to other models (P < 0.001). The most prominent predictors based on the GBT model were the route of poison entry (weight = 0.583), place of residence (weight = 0.137), history of psychiatric diseases (weight = 0.087), and age (weight = 0.085).

The present study suggests the GBT model as a reliable predictor model for identifying the factors affecting intentional and unintentional poisoning. According to our results, the determinants of intentional poisoning included the route of poison entry into the body, place of residence, and the heart rate. The most important predictors of unintentional poisoning were age, exposure to benzodiazepine, creatinine levels, and occupation.