Prospective Detection of Foodborne Illness Outbreaks Using Machine Learning Approaches

Abstract

Despite advances in food safety regulations, food-borne illness imposes a heavy health burden, with nearly 50 million estimated incident cases of illness each year. Having a prospective foodborne illness outbreak detection mechanism for more accurate and timely triggering of outbreak control measures would offer notable public health dividends, but is challenging due to the subclinical character of most foodborne illnesses. Within this work, collected synthetic datasets of incident illness cases and vendor contamination records from a previously contributed and empirically grounded model of foodborne illness, are used to study the efficacy of Hidden Markov Models (HMMs) for syndromic surveillance monitoring and disease outbreak detection under two data collection regimes, one involving a sentinel population using smartphone-based app for tracing location of food consumption and subclinical reporting. A support vector machine (SVM) approach was applied to compare the results to the HMM. Findings suggest that while reliance on clinical data offers poor potential for automatic outbreak detection, the use of HMMs offer excellent potential for detecting foodborne illness outbreak when informed by subclinical reporting by even a very small (4% of population) sentinel group. By contrast, SVM offers relatively poor prospects for detection. Furthermore, experiments with an empirically grounded agent-based model suggest that use of an HMM may be advantageous for triggering outbreak investigations among public health inspectors.