Logo image
Back
Healthcare associated infections: computational modeling, inference, and prediction
Dissertation   Open access

Healthcare associated infections: computational modeling, inference, and prediction

Hankyu Jang
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Autumn 2023
DOI: 10.25820/etd.006964
pdf
HankyuJang_Thesis13.89 MBDownloadView
Open Access Free to read and download

Abstract

Healthcare associated infections (HAIs) are infections that occur during care in a healthcare facility (e.g., hospital ICU or nursing home). An example of a common HAI is Methicillinresistant Staphylococcus aureus (MRSA) infection. Approximately 5% of patients in the United States’ hospitals are colonized with MRSA. MRSA commonly causes skin infections, and when MRSA gets into the bloodstream, it could cause infections in other organs, becoming fatal. Another common HAI is Clostridioides difficile (C. diff) infection (CDI). Nearly half a million CDI cases occur yearly in the United States alone. CDI causes diarrhea and inflammation of the colon that can be life-threatening. Efforts to model HAIs date from the traditional compartmental models (e.g., SEIR model) to more complex and realistic models that layer agent-based simulations on top of contact networks. HAIs threaten patients, so hospitals, clinics, and long-term care facilities are interested in preventing and mitigating the spread of HAIs. This dissertation explores various computational problems associated with HAIs, including the design of realistic and tractable models (computational modeling), inference problems associated with HAIs, and prediction to mitigate the spread of HAIs. The dissertation describes seven research projects: the first two projects on the computational modeling, the next four projects on the inference, and the last project on the prediction. The transmission mode of MRSA and CDI includes an infected agent shedding pathogens on the nearby environments, where other agents pick up pathogens and drop them off at different locations. Susceptible agents get infected due to the pathogen load on their skin. We propose a load sharing model (Mload) to model these dynamics and track pathogen load on patients, healthcare providers (HCPs), and locations. We run Mload on top of the contact networks extracted via instrumentation of HCPs at a dialysis unit, where the load transfers bidirectionally whenever two agents have contact. We further explore non-pharmaceutical interventions (NPIs) that involve architectural changes to slow down the infection spread. We observe that the behavior of Mload is very different than compartmental models likes susceptible-exposed-infected-recovered (SEIR): more contacts do not necessarily increase the infection counts, which suggests interventions on environment-mediated HAIs should be done with care. Other HAIs spread with different modes of transmission. COVID-19, for example, spreads when an infected person sneezes respiratory droplets that travel within 6 feet and land on a susceptible person’s nose, eye, or mouth. We propose a viral shedding model for COVID-19: the shedding level increases exponentially since exposure, reaches the maximum on the day of the symptom, and then reduces exponentially. We run this model on top of the dialysis unit contact networks. We show that a combination of low-cost NPIs, such as surgical masks, social distancing, distancing dialysis chairs, and additional interventions (e.g., short-term N95 mask usage) upon detection of the first symptomatic patient reduces the spread significantly. We would like to note that designing effective NPIs is challenging due to asymptomatic cases. For instance, nearly half of those with COVID-19 were asymptomatic yet infectious. There is a need to infer those contributing to the spread of HAIs. Hand hygiene and environment cleaning are widely adapted intervention strategies in controlling the spread of CDI or MRSA. Such strategies are costly to implement across the board, so interventions should be prioritized to those that contribute to the spread of HAIs more than others. We pose this problem as finding central nodes (e.g., patients, HCPs, or fomites) in a temporal contact network to which HAI spread can be attributed. We propose centrality measures to identify nodes that are (i) likely to be contaminated or (ii) transfer most pathogens. We impose interventions on central nodes and observe the reduction in the infection counts of Mload simulation on contact networks from four healthcare facilities. Results show that our measures are computationally efficient yet outperform all the baselines in reducing HAI spread on all contact networks. When some HAIs are detected in a healthcare setting, much effort, such as contact tracing or additional testing, is invested in rapidly identifying the infection’s sources. This corresponds to the source detection problem. We formulate the source detection problem for Mload. The natural formulation is computationally very hard, so we present tractable formulations, where the tractability of our problems depends crucially on the submodularity of the expected number of infections as a function of the source set. We propose algorithms that extend existing algorithmic results from submodular optimization. We evaluate our method by detecting sources from synthetically generated outbreaks on contact networks from three hospitals and show that our algorithms outperform baselines. Furthermore, our algorithms are able to detect clinically meaningful sources from a real CDI outbreak. Aside from sources, detecting any missing HAIs (e.g., asymptomatic infections) is also important because missing infections are accountable for the spread of HAIs. We propose a 2-stage classification model for inferring asymptomatic CDI. The Stage 1 model detects asymptomatic C. diff carriers by training on symptomatic CDI cases using known risk factors from CDI as features. The Stage 2 model evaluates those detected cases using an additional feature set on the exposure to those detected in Stage 1. Our results imply that asymptomatic C. diff carriers contribute to CDI spread, confirming an important conjecture from the CDI literature. We propose another method to detect asymptomatic cases by more carefully taking into account both individual risk and exposure to pathogens - previous research has largely ignored the interplay between these dual aspects influencing disease spread. We formulate the asymptomatic case detection problem as a directed prize-collecting Steiner tree problem. We present an approximation-preserving reduction from this problem to the directed Steiner tree problem and use this reduction to obtain scalable algorithms. We demonstrate that our methods outperform various baselines. Accurately predicting the onset of HAIs allows healthcare facilities to monitor high-risk patients and prevent subsequent infections. When a patient goes through a sequence of high-risk events, such as visiting a contaminated room, getting a prescription for high-risk antibiotics, etc., the risk of getting HAI increases. We propose to capture the information in a series of events occurring at the hospital for each node (e.g., patients, HCPs, medications, and rooms) as embeddings. We propose DECEnt, an auto-encoding heterogeneous coevolving dynamic neural network, for learning heterogeneous dynamic embeddings of patients, doctors, rooms, and medications. We evaluate the learned embeddings in their performance of predicting the onset of CDI and other preventive modeling tasks. Results show that DECEnt outperforms all the baselines in all evaluation tasks.

Details

Metrics

4 File views/ downloads
160 Record Views
Logo image