Network Based Prediction and Control Models for Healthcare Associated Infection

Healthcare associated infections (HAIs) have become a major challenge to public healthcare and pose a heavy burden on modern healthcare systems. Some nosocomial infections such as Methicillin-resistant Staphylococcus aureus (MRSA) have a long history of threatening patients and healthcare workers in a hospital. It can even spread to communities near a contaminated hospital. Other infectious diseases like SARS may turn into a severe epidemic spread around the world from a single hospital outbreak. The most important and frequent mode of the transmission of healthcare associated infections is person-to-person contact among patients and healthcare workers. With a better understanding of human interaction pattern in a hospital environment, healthcare systems can control hospital acquired diseases more effectively and efficiently. Furthermore, in-depth investigation on transmission mechanism of healthcare associated infections is necessary to reduce the risk of hospital outbreaks as far as possible. This work studies the problem of predicting and controlling the transmission of healthcare associated infections over dynamic human contact networks. Three main problems are investigated.;The first problem is to track the transmission of healthcare associated infections over time-varying contact networks at an individual level. We present a method combining the ideas of network epidemiology and coupled hidden Markov models to realize individual-level tracking of HAI transmission. A time-varying contact network is first constructed to characterize person-to-person contacts among patients and healthcare workers from their location data collected in a hospital in Hong Kong. Given an outbreak occurs over the network, we provide an approach to monitor illness evolution of an individual and identify the hidden health state of any individual at any time. Moreover, an estimation method is used to learn unknown HAI parameters in the case that the knowledge of the hospital outbreak is incomplete.;The second problem is to predict the macro-level phenomena of the spreading process of nosocomial infections, such as the total infected number over time and epidemic thresholds. Traditional epidemiology study generally focuses on city-level epidemic outbreaks. However, disease propagation in a healthcare setting is different due to the modular and hierarchical structure of a hospital. In this work we use a differential equation based model combining with the technique of probability generating functions to study the macro-level transmission of healthcare associated infections, which is considered to spread on a static projection of the time-varying contact network constructed from human positioning data collected from the hospital.;The third problem is to develop efficient control strategies against healthcare associated infections with the goal of reducing the impact of disease outbreaks as far as possible. In this work we study a common approach of targeted immunization for intervening epidemic outbreaks and show its equivalence to the influence maximization problem which is extensively studied in social network analysis. We formulate the targeted immunization problem, which involves identifying "super spreaders" who play a key role in spreading disease over human contact networks, as a concise optimization problem under the Linear Threshold (LT) propagation model. In this way the targeted immunization problem as well as the influence maximization problem can be solved through an optimization approach efficiently and effectively.;We conduct various experiments on four-month human tracking data collected from a hospital to evaluate the performance of proposed prediction and control models. Results demonstrate that the prediction models is capable of tracking the illness evolution at an individual level with high accuracy and capturing macro-level epidemic dynamics of healthcare associated infections. The performance and scalability of the control strategies are also evaluated on several real-world large-scale networks. Results show that the proposed approaches achieve better solutions for selecting targeted seeds compared to existing methods.