Identifying Importation and Asymptomatic Spreaders of Multi-drug Resistant Organisms in Hospital Settings.

Healthcare-associated infections (HAIs) due to multi-drug resistant organisms (MDROs) are a significant burden to the healthcare system. Patients are sometimes already infected at the time of admission to the hospital (referred to as "importation"), and additional patients might get infected in the hospital through transmission ("nosocomial infection"). Since many of these importation and nosocomial infection cases may present no symptoms (i.e., "asymptomatic"), rapidly identifying them is difficult since testing is limited and incurs significant delays. Although there has been a lot of work on examining the utility of both mathematical models of transmission and machine learning for identifying patients at risk of MDRO infections in recent years, these methods have limited performance and suffer from different drawbacks: Transmission modeling-based methods do not make full use of rich data contained in electronic health records (EHR), while machine learning-based methods typically lack information about mechanistic processes. In this work, we propose NEURABM, a new framework which integrates both neural networks and agent-based models (ABM) to combine the advantages of both modeling-based and machine learning-based methods. NEURABM simultaneously learns a neural network model for patient-level prediction of importation, as well as the ABM model which is used for identifying infections. Our results demonstrate that NEURABM identifies importation and nosocomial infection cases more accurately than existing methods.