Hospital hallway

Detecting and Controlling Network-based Spread of Hospital Acquired Infections


National Science Foundation

Hospital Acquired Infections (HAIs) are becoming a major challenge in today’s health systems worldwide; many of these infections are becoming resistant to antibiotics, which makes their treatment very difficult. Detection and control of HAIs are challenging and resource intensive logistics problem, because of the high costs of patient treatment and disinfection of hospital facilities, making them fundamental public health problems. Despite its huge importance for hospitals (which are financially penalized for high HAI rates), and the interest from both clinical and epidemiological researchers, these problems remain poorly understood, due to multiple reasons: there is limited data on outbreaks, and the dynamics of HAI spread are much more complex than other diseases, such as Influenza, due to many more factors and pathways involved. Therefore, a purely data-driven approach for detection and control of HAIs is unlikely to be very effective. Motivated by this, our project seeks to combine a novel model-based network science approach with techniques from data mining and machine learning to advance the science of hospital infection control.

Project Overview

Detection and control of HAIs are challenging both on the "ground" and from a methodological point of view. HAIs are relatively rare, but consume substantial resources when they do occur. Hand hygiene rates continue to be low among healthcare workers (HCWs) and it is also quite difficult to change workflows in a hospital. The problems of HAI detection and control also pose new challenges for data science. First, there is limited data on outbreaks—these are still relatively rare, and typically do not result in large outbreaks. Second, there are a host of interacting factors which are known to play a role in the spread of HAIs—this includes transmission within hospitals, transmission that occurs when patients are transferred between hospitals, and transmission in the community. Within the hospital, HAIs spread mainly through contacts with colonized or infected individuals and surfaces (e.g., beds, sinks and toilets). Therefore, a purely data-driven approach for detection and control of HAIs is unlikely to be very effective.

Motivated by this, our project seeks to develop a novel network-based approach to improve hospital infection control using models and data science. This will lead to new problems and techniques from data mining, network science and machine learning perspectives. HAI-spread can be represented by so-called "two-mode" models, in which the infection dynamics depend on both (i) person-person and person-location interactions and (ii) infection-load dynamics at locations (unlike well-known models like IC and SIR). Further, HAI models can be mathematically viewed as a combination of cascade and threshold style models, which haven’t been studied in data mining. Our project has the following aims.

Aim 1
Surveillance and inference of HAIs: this thrust involves constructing sensor sets (whom or what to monitor) so that an HAI outbreak is detected early, and identifying missing infections from positive test results. We use the two-model cascade model for formalizing these problems. Both the sensor set and missing infection problems in this setting are much more challenging than the corresponding problems for the SIR style models (which have been well studied), and new techniques are needed.

Aim 2
Designing interventions to control the spread of HAIs: such interventions are a com- bination of treating people, and sanitizing locations. These are expensive to implement, and medications have side effects, and this thrust will find effective interventions, subject to resource constraints. We also study new variations, such as multiple stage interventions, and adaptive control based on detected infections. Again, these problems are formalized using the two-mode model, and lead to challenging stochastic optimization problems, since temporal intervention design has not been studied and remains completely open even in the SIR model.

Aim 3
Exposure Risk Factors for HAIs: Individual-level risks for specific HAIs (e.g., being aged, having other comorbidities, being prescribed high-risk antibiotics, etc.) have been studied in the epidemiology literature. In contrast, risks due to exposure to other colonized or infected individuals and from exposure to surfaces with high pathogen loads are not as well understood. Modeling and estimating these “local” risks and quantifying the likelihood of different pathways for acquiring an HAI is the focus here. Results from this aim will not only increase our understanding of the epidemiology of these HAIs, but will also inform the surveillance and intervention aims of this project.


Development and analysis of new 2-mode models for HAI spread

New network science approaches for testing and inference of HAIs



Professor of Computer Science, School of Engineering and Applied Science

Research Associate Professor

Additional Team Members:

Aditya Prakash (Georgia Tech)
Sriram Pemmaraju (University of Iowa)