NeuralPALS - learning exposure functions and infection probabilities for contagion spread

Abstract

Understanding how contagions spread is an important task, particularly when considering infectious diseases. An individual's likelihood of getting infected by a contagion is determined by a combination of inherent susceptibility and exposure to other individuals who may spread the disease. In a real world setting, an individual's infection status may be directly observable, but it is difficult to identify whether an individual is spreading the disease. As a result, the exact influence function by which disease is transmitted is difficult to understand as well. We present a neural network based method, NeuralPALS, to learn the spreader, exposure, and infection status of individuals in a network. Unlike previously developed methods, we do not assume an exposure function and instead devise methods to learn this function. Through experiments on synthetic data we illustrate our method's efficacy in determining both spreader and infection states. We also demonstrate NeuralPALS's ability to learn different exposure functions. In addition, we utilize a dataset of patients from a large urban hospital and demonstrate our preliminary results in determining the spread of Clostridioides difficile.