Enhanced Pesticide Exposure Diagnosis: A Machine Learning Approach

In response to a growing human population, more attempts to correspondingly increase agricultural production become necessary. To boost crop production, pesticides have been widely applied against pests and plant diseases to attain enhanced crop efficiency. Unfortunately, routine application of pesticides brings about adverse effects unintentionally not only on human health but to the environment as well. Identifying pesticide intoxication is important for early intervention efforts to be implemented. The study aims to predict whether the patient is intoxicated due to pesticide exposure and identify which methodological enhancements as to the feature selection step and class imbalance handling method would yield the highest predictive capability for pesticide intoxication. Several machine learning algorithms (k-Nearest Neighbors, support vector machine, random forest and adaptive boosting) in various model configurations representing the methodological enhancements were evaluated on the predictive capability. Results showed random forest with the optimum predictive performance generating the highest Matthews correlation coefficient and F1-score across all model configurations. The incorporation of a feature selection step has significantly helped k-Nearest Neighbors in predicting pesticide intoxication while the improvement in support vector machine can be attributed to the class imbalance handling method. Further optimization with hyperparameter tuning generally showed additional improvement in the prediction capability for all machine learning algorithms. The integration of these tools in decision support systems for clinical observation of patients suspected of pesticide intoxication can be useful for swift diagnosis. This is a significant contribution to the health of agricultural workers while ensuring agricultural productivity.