Spread of pathogens in the patient transfer network of US hospitals

TL;DR

The paper investigates how patient transfers among US hospitals shape the nationwide spread of nosocomial pathogens, using a two-year Medicare dataset to construct a directed, weighted transfer network. It demonstrates a positive correlation between hospital C. difficile incidence and the incidence among network neighbors in low-incidence regimes, supporting the network as a plausible transmission substrate. Through comparative sensor-placement analyses (in-degree, out-degree, random) under static and dynamic implementations, it shows that in-degree-based sensors are most efficient, enabling substantial coverage with a small sensor set, and that activation windows of a few days suffice to detect a large fraction of cases. These results imply that monitoring the hospital transfer network, not just individual hospitals, can enhance nationwide outbreak surveillance and inform public health responses, with practical implications for real-time network-based infection monitoring and rapid containment strategies.

Abstract

Emergent antibiotic-resistant bacterial infections are an increasingly significant source of morbidity and mortality. Antibiotic-resistant organisms have a natural reservoir in hospitals, and recent estimates suggest that almost 2 million people develop hospital-acquired infections each year in the US alone. We investigate a network induced by the transfer of Medicare patients across US hospitals over a 2-year period to learn about the possible role of hospital-to-hospital transfers of patients in the spread of infections. We analyze temporal, geographical, and topological properties of the transfer network and demonstrate, using C. Diff. as a case study, that this network may serve as a substrate for the spread of infections. Finally, we study different strategies for the early detection of incipient epidemics, finding that using approximately 2% of hospitals as sensors, chosen based on their network in-degree, results in optimal performance for this early warning system, enabling the early detection of 80% of the C. Diff. cases.

Figures (17)

• Figure 1: Network of hospital-to-hospital transfers of US Medicare patients. The network consists of hospitals that are connected by daily transfers of patients, here aggregated over the two-year period. Edge color encodes the number of patients transferred through each connection.
• Figure 2: Correlation between C.diff. incidence and transfer network structure. The horizontal axis represents the mean C. Diff. incidence at the focal hospital over time and the vertical axis is the mean C. Diff. incidence in the network neighborhood of that hospital (the mean taken first over time and then over all network neighbors). We exclude hospitals with fewer than 100 patients from subsequent correlation analyses, leading to exclusion of 7.5% (428) of all hospitals. The Pearson correlation coefficients are 0.47 and -0.01 for the low and high incidence regimes, respectively, which are separated by the vertical line.
• Figure 3: Finding the optimal sensor set for the static implementation of the surveillance system. Efficacy (a) and fraction of detected cases (b) on the static network as a function of the fraction of hospitals acting as sensors. The different curves represent different strategies for sensor selection: random selection (black), selection proportional to in-degree (red), and selection proportional to out-degree (blue).
• Figure 4: Finding the optimal sensor set for the dynamic implementation of the surveillance system. Heatmaps showing the efficacy (left column) and fraction of detected cases (right column) on the temporal transfer network. Results are shown as a function of the fraction of hospitals acting as sensors (horizontal axes) and the activity time that they implement (vertical axes). The rows of panels correspond to choosing the sensors randomly (top row), proportional to out-degree (middle row) and proportional to in-degree (bottom row).
• Figure 5: Efficacy of temporal sensor sets.a) Fraction of sensors for the most efficient sensor set from the temporal network for sensors chosen at random (black), proportional to in-degree (red), and proportional to out-degree (blue). We have smoothened the efficacy curves by averaging the results using a window of 5 sensors. b) Fraction of detected cases for the most efficient sensor set. c) Average fraction of time that a sensor stays in the active state (same color code as on the left).