Efficacy of machine learning models for the prediction of death occurrence and counts associated with foodborne illnesses and hospitalizations in the United States

Abstract

Foodborne outbreak data released through national surveillance systems provides essential information about the results of investigations. This study evaluates the efficacy of machine learning (ML) models for the prediction of death occurrence and counts associated with foodborne illnesses and hospitalizations in the United States. Confirmed foodborne outbreaks were obtained from the Centers for Disease Control and Prevention's National Outbreak Reporting System (NORS). Foodborne pathogens causing at least 10 deaths in total were selected for analysis. The binary classification performance (accuracy, %) and prediction efficacy of ML models (mean absolute errors, MAE) were used for evaluation. A total of 10,069 foodborne outbreaks with confirmed single etiology resulted in 275,827 illnesses, 18,579 hospitalizations, and 458 deaths. Salmonella was the leading causative agent (54.23 %) of bacterial foodborne outbreaks, followed by pathogenic Escherichia coli (12.13 %). Norovirus (96.69 %) and Cyclospora cayetanensis (60.76 %) represented major causes of viral and protozoan/parasite foodborne outbreaks, respectively. The classification performance of ML models ranged from 88.9 to 94.5 % for the overall prediction of death occurrence associated with foodborne illnesses and hospitalizations. Prediction efficacy of ML models for death counts remained <0.9 with MAE, except for Listeria monocytogenes with an average MAE of 134.1 ± 11.1. This study indicates the potential use and performance of ML algorithms for the prediction of death occurrence or counts caused by foodborne etiological agents to improve public health safety based on the numbers of illnesses and hospitalizations.