Network memory in the movement of hospital patients carrying drug-resistant bacteria

Ashleigh Myall, Robert L. Peach, Andrea Y. Weiße, Frances Davies, Siddharth Mookerjee, Alison Holmes, Mauricio BarahonaarXiv 2020 · DOI · arXiv

Abstract

Hospitals constitute highly interconnected systems that bring into contact an abundance of infectious pathogens and susceptible individuals, thus making infection outbreaks both common and challenging. In recent years, there has been a sharp incidence of antimicrobial-resistance amongst healthcare-associated infections, a situation now considered endemic in many countries. Here we present network-based analyses of a data set capturing the movement of patients harbouring drug-resistant bacteria across three large London hospitals. We show that there are substantial memory effects in the movement of hospital patients colonised with drug-resistant bacteria. Such memory effects break first-order Markovian transitive assumptions and substantially alter the conclusions from the analysis, specifically on node rankings and the evolution of diffusive processes. We capture variable length memory effects by constructing a lumped-state memory network, which we then use to identify overlapping communities of wards. We find that these communities of wards display a quasi-hierarchical structure at different levels of granularity which is consistent with different aspects of patient flows related to hospital locations and medical specialties.

Introduction

To control the spread of antimicrobial resistance (AMR), understanding how patients—the primary vectors—move through hospital systems is critical. Most current epidemiological models treat these movements as a first-order Markov process, assuming the probability of moving to a new ward depends only on a patient's current location. However, medical care is structured into specific pathways (e.g., admission $\rightarrow$ surgery $\rightarrow$ recovery).

If a model ignores this "memory," it assumes a level of randomness that doesn't exist in reality. This leads to "inflated transitivity," where models suggest pathogens can reach any part of the hospital much faster than they actually do, potentially misdirecting resources for infection prevention and control (IPC). This paper demonstrates that patient movement is not a memoryless process; where a patient goes next depends heavily on where they have been.

Methodology: From First-Order to Memory Networks

The authors move beyond standard network analysis by using Higher-Order Memory Networks ($M_k$). They compare first-order models ($M_1$, memoryless) with second-order models ($M_2$), where a "state" represents a sequence of two wards.

Conceptual comparison of patient pathways, first-order networks, and memory networks.

To manage complexity and prevent data sparsity, the authors introduce a "state node lumping" technique that merges redundant state nodes based on the similarity of future visitation probabilities. They then apply Markov Stability (MS) for community detection to find ward clusters that "trap" patient flow at different time scales, and Multiscale Centrality (MSC) to identify "hub" wards that are important either locally or globally.

Data and Hierarchical Structures

The study utilizes anonymized electronic health records from a large 1000-bed London hospital Trust, comprising three distinct sites and 96 wards. The cohort included 967 patients who tested positive for Carbapenemase-producing Enterobacteriaceae (CPE) between 2018 and 2020.

Patient flows across hospital sites, buildings, wards, and specialties.

The analysis reveals that ward clusters are not just defined by physical buildings; they emerge from a mix of geography and medical specialties (e.g., Cardiology, Haematology), which act as functional "islands" within the hospital network. These communities display a quasi-hierarchical structure consistent with patient flows across different levels of granularity.

Key Findings and Results

The "memory" effect in patient movement is substantial. The entropy rate (a measure of uncertainty) dropped from 2.70 in the memoryless model to 0.57 in the second-order model, indicating that movement is far more predictable and constrained than standard models suggest.

Key results include:

  • Predictive Accuracy: The $M_2$ model was significantly more accurate at predicting ward rankings (Kendall-Tau correlation of 0.60) compared to the $M_1$ model (0.49).
  • Reachability Divergence: After three steps, a memoryless model predicts a patient can reach 79 different wards on average. The memory-aware model restricts this to just 5 wards.
  • Hub Identification: In memoryless models, general admission wards appear as the most important hubs. However, the memory model reveals that Renal wards are the true high-traffic centers for CPE patients, reflecting the specific clinical needs of this cohort.
  • Bridge Wards: The analysis identified 48 "bridge" wards—including Renal and Critical Care units—that belong to multiple communities, marking them as high-risk zones for cross-specialty transmission.

Conclusion

While data sparsity limited testing for memory effects beyond $k=2$, the findings have significant implications for hospital management. Integrating memory-aware networks into real-time IPC dashboards could help identify emerging "hotspots" and "bridge" wards before an outbreak spreads. Future work should address potential selection biases, as certain wards may test for CPE more frequently than others.

Summary updated

Abstract

Hospitals constitute highly interconnected systems that bring into contact an abundance of infectious pathogens and susceptible individuals, thus making infection outbreaks both common and challenging. In recent years, there has been a sharp incidence of antimicrobial-resistance amongst healthcare-associated infections, a situation now considered endemic in many countries. Here we present network-based analyses of a data set capturing the movement of patients harbouring drug-resistant bacteria across three large London hospitals. We show that there are substantial memory effects in the movement of hospital patients colonised with drug-resistant bacteria. Such memory effects break first-order Markovian transitive assumptions and substantially alter the conclusions from the analysis, specifically on node rankings and the evolution of diffusive processes. We capture variable length memory effects by constructing a lumped-state memory network, which we then use to identify overlapping communities of wards. We find that these communities of wards display a quasi-hierarchical structure at different levels of granularity which is consistent with different aspects of patient flows related to hospital locations and medical specialties.
BibTeX
@misc{myall2020network,
  author = {Ashleigh Myall and Robert L. Peach and Andrea Y. Weiße and Frances Davies and Siddharth Mookerjee and Alison Holmes and Mauricio Barahona},
  title = {{Network memory in the movement of hospital patients carrying drug-resistant bacteria}},
  year = {2020},
  eprint = {2009.14480},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2009.14480},
}
Figure