Using spectral characterization to identify healthcare-associated infection (HAI) patients for clinical contact precaution

Jiaming Cui, Sungjun Cho, Methun Kamruzzaman, Matthew Bielskas, Anil Vullikanti, B. Aditya Prakash

Abstract

Healthcare-associated infections (HAIs) represent a significant challenge for hospital infection control, leading to increased patient mortality and substantial economic costs. While contact precautions—such as isolation and enhanced cleaning—are effective at suppressing spread, they are resource-intensive and often constrained by limited hospital budgets. This paper addresses two critical clinical problems: the Isolation Problem (IP), selecting the optimal patients for precaution to minimize spread, and the Clearance Problem (CP), identifying which patients to release when budgets are reduced. By leveraging non-linear system theory, we develop a novel spectral characterization of the 2-MODE-SIS model, which accounts for pathogen transmission between both people and locations. We introduce the GREEDY-SPECTRAL algorithm, which efficiently optimizes these precautions. Experimental results on real-world hospital datasets demonstrate that our approach significantly outperforms standard clinical and structural heuristics.

Introduction

In modern healthcare, infections like MRSA often spread through complex pathways involving not just patient-to-patient contact, but also environmental contamination and healthcare workers (HCWs) acting as vectors. Standard epidemiological models often fail to capture these "2-mode" dynamics where locations (rooms, units) accumulate pathogen loads. Because hospitals cannot afford to place every patient under strict contact precautions, they must strategically decide how to allocate limited resources. This research provides a mathematical framework to identify the most influential nodes in the hospital network to target for intervention.

Methodology: The 2-Mode-SIS Model

To accurately reflect hospital environments, the researchers utilized the 2-Mode-SIS model. Unlike traditional models that focus solely on human infection states, this model tracks the "load" of pathogens on both people and physical locations. This allows for the modeling of interventions like room sterilization or the use of personal protective equipment (PPE) by HCWs.

Illustration of the heterogeneous hospital contact network over time, incorporating patients, healthcare workers (HCWs), and locations.

Spectral Characterization and the Tipping Point

A key technical contribution of this work is the derivation of a system matrix ($S$) and its spectral radius $ ho(S)$. Using non-linear system theory, the authors proved that $ ho(S) = 1$ serves as a critical "tipping point" for the system. If $ ho(S) < 1$, the infection eventually dies out; if $ ho(S) > 1$, the infection persists within the population.

To overcome numerical precision challenges in large, time-varying networks, the authors demonstrated that $ ho(S)$ is upper-bounded by the spectral radius of a simpler "system-adjacency matrix" ($ ho(A)$). Minimizing this upper bound provides a stable and computationally efficient target for optimization.

The GREEDY-SPECTRAL Algorithm

The proposed GREEDY-SPECTRAL algorithm is a heuristic designed to select patients for isolation or clearance by greedily minimizing the spectral radius $ ho(A)$. By focusing on the structural properties of the hospital's contact network, the algorithm identifies patients whose isolation would most effectively disrupt the transmission chains across the entire facility.

Experimental Evaluation

The study utilized Electronic Health Records (EHR) from the University of Virginia (UVA) hospital to construct time-varying networks spanning two 294-day periods: one pre-COVID and one during the pandemic. These networks included patients, HCWs, and specific locations, with interactions inferred from medical notes and administration records.

Key Findings

  • Superior Performance: In the UVA-Precovid dataset, GREEDY-SPECTRAL reduced non-patient pathogen loads by 9.2% more than the Shorr score (a common clinical risk assessment) when isolating 5,000 patients.
  • Case Suppression: In scenarios with relaxed cleaning protocols (simulating resource shortages), the algorithm prevented 60–70 more MRSA cases than the Shorr baseline and 160–170 more than simple degree-based connectivity approaches.
  • Network as a Proxy for Risk: Analysis of the patients selected by the algorithm revealed that 75% had a history of surgery and 47% had invasive devices. This suggests that the network topology alone is a powerful proxy for clinical risk, identifying high-risk individuals without requiring access to detailed medical charts.

Conclusion

This work demonstrates that spectral characterization of 2-mode networks offers a robust and efficient way to manage hospital infection control under budget constraints. By identifying the "tipping point" of the system, hospital administrators can make data-driven decisions about patient isolation and clearance. Future work will look to extend this model beyond binary infection states to include colonization and more dynamic, time-sensitive isolation strategies.

Summary updated

Abstract

Healthcare-associated infections (HAIs) are a major problem in hospital infection control. Although HAIs can be suppressed using contact precautions, such precautions are expensive, and we can only apply them to a small fraction of patients (i.e., a limited budget). In this work, we focus on two clinical problems arising from the limited budget: (a) choosing the best patients to be placed under precaution given a limited budget to minimize the spread (the isolation problem), and (b) choosing the best patients to release when limited budget requires some of the patients to be cleared from precaution (the clearance problem). A critical challenge in addressing them is that HAIs have multiple transmission pathways such that locations can also accumulate 'load' and spread the disease. One of the most common practices when placing patients under contact precautions is the regular clearance of pathogen loads. However, standard propagation models like independent cascade (IC)/susceptible-infectious-susceptible (SIS) cannot capture such mechanisms directly. Hence to account for this challenge, using non-linear system theory, we develop a novel spectral characterization of a recently proposed pathogen load based model, 2-MODE-SIS model, on people/location networks to capture spread dynamics of HAIs. We formulate the two clinical problems using this spectral characterization and develop effective and efficient algorithms for them. Our experiments show that our methods outperform several natural structural and clinical approaches on real-world hospital testbeds and pick meaningful solutions.
BibTeX
@inproceedings{cui2023using,
  author = {Jiaming Cui and Sungjun Cho and Methun Kamruzzaman and Matthew Bielskas and Anil Vullikanti and B. Aditya Prakash},
  title = {{Using spectral characterization to identify healthcare-associated infection (HAI) patients for clinical contact precaution}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41598-023-41852-5},
}
Figure