Colonization and Succession of Hospital-Associated Microbiota

The indoor environment has become our most intimate ecosystem, though its reduced microbial diversity relative to the outside world may be linked to an increased incidence of immunological diseases, asthma, and allergies^1–4. A strong link has been observed between the microbial communities associated with human skin and those recovered from buildings^5–7. However, the implication of this interaction on the incidence of hospital-acquired infections as a leading cause of patient death^8–10 has not been investigated. Culture-dependent analysis of HAIs and pathogen genotype tracking allows for the near-real time post hoc characterization of potential transmission routes¹¹. However, outside of a small number of studies limited to ICUs and neonatal care rooms^12–14 no detailed longitudinal culture-independent analyses of the hospital microbiome have been performed¹⁵.

Here, we present a yearlong microbial survey of the microbiota associated with the patients, staff, and surfaces of the newly constructed Center for Care and Discovery (University of Chicago). Sampling began 2 months prior to the hospital becoming operational on February 23^rd 2013, and continued for nearly a year post-opening. We collected 6,523 microbial samples from multiple sites (Table S1) in 10 patient care rooms and two nursing stations across two hospital levels. At least 5,000 high quality 16S rRNA V4 amplicons were generated per sample. One patient room on each level was sampled daily while all other environments where sampled weekly. Environmental conditions, including temperature, humidity, illuminance, CO₂ concentrations and infrared doorway beam breaks were continuously monitored^16–17.

A soon as the hospital became operational, the floor and nursing station surfaces saw a significant increase in the relative abundance of the human-skin associated genera Corynebacterium, Staphylococcus, and Streptococcus, and a decrease in Acinetobacter and Pseudomonas, which dominated pre-opening samples (Figure S1A,B,C; Figure S2). OTU-level Shannon diversity significantly increased in nursing station surfaces that commonly interact with human skin (Figure S1 D), but not in floor samples.

Skin samples were generally the least diverse of all sample types (although they had greater phylogenetic diversity than expected given their low Shannon index), while surfaces most likely to interact with the outdoors, such as air filters, shoes, and floors, were the most diverse (Figure 1A). Samples taken from the phone, chair armrest, countertop, computer mouse, and floor of the nurse stations, as well as the floor and air filters from the patient rooms, all had significantly lower beta diversity compared to samples taken from patient skin, staff noses, faucet handles, and unused latex gloves (Figure 1B). Staff hand microbial communities were significantly more similar to the microbiota of built hospital surfaces than were patient hands, likely as a result of greater staff mobility within the hospital. Supervised learning models could successfully differentiate nose and hand samples taken from staff members and patients (error ratios of 2.5 and 3.9, respectively) and hand samples could even be differentiated using genus-level data (error ratio = 3.3), with the genera Micrococcus (staff-associated) and Prevotella (patient-associated) having the highest feature importance scores. Pre-opening room and station floor samples had highly similar microbial communities, but were dissimilar to other surfaces; while post-opening floor samples had a greater degree of similarity to all surfaces (Figure 1B).

To determine the strength of microbial interaction between different hospital surfaces, we calculated the degree to which samples taken from two different surfaces on the same day and in the same room or building level, resembled each other using a method we term PC space correlation (Figure 2). Patient hands and bedrails had a strong interaction (ρ = 0.5); but all pairwise correlations within patient rooms were significant, suggesting a degree of microbial homogenization on the same day within each room. The strongest observed correlations were between the hand microbiota of hospital staff and their personal cell phones and pagers (ρ = 0.52 and 0.50 respectively), which has been observed previously^18–19. Correlations within the nurse station environment were all significant but comparatively weak, probably due to the diversity of people using these environments daily. Correlations between environments were generally much weaker than those within environments, likely because of the reduced overlap in personnel occupancy. The station and patient floors had the greatest correlations between environments, likely due to homogenization by shoe traffic between locations. However, staff shoe and floor samples did not seem to interact significantly despite previous observations¹⁸. This may be due to regular cleaning of the hospital floors or, more likely, the diversity of other individuals walking throughout the hospital who were not sampled in this study.

To investigate the degree to which the microbiota of individual patient rooms varied over the year, we calculated the variability of the core microbiome (Figure 3A). For all surface types, the average percent of 16S rRNA reads detected in every sample of a given room varied between 15 and 35%, being greatest for patient noses and lowest for patient axillae. The trends for all 5 surfaces were remarkably similar, suggesting that variability in the microbiota of room bedrails and floors could be attributable to variation in the skin microbiota of patients. We used a Bayesian source-tracking approach²⁰ to match microbial profiles of samples taken from the second day of a patients stay, to those taken on their first day (Figure 3B). For the same surface type (e.g. hand to hand), the model was highly predictive for matching the patient on day 2 to the correct patient on day 1. Models using the microbial profile of day 1 hands to predict day 2 bedrails, and day 1 bedrails to predict day 2 hands, were also highly accurate. Floor, nose, and axilla samples, by contrast, had much weaker predictive accuracy, though with the exception of nose or floor predicting axilla, were significantly better than random.

Within a patient room on the same day higher temperatures and higher illuminance was consistently associated with greater microbial dissimilarity between patient and surface microbial communities, while higher relative humidity and humidity ratio consistently correlated with greater microbial similarity (Figure 4A). Skin-associated microbial similarity between staff members working on the same floor showed a seasonal trend, with the greatest similarity in late summer/early fall, and the least similarity in the winter (Figure 4B). The greater the humidity, the greater the nose and hand-associated microbial similarity between staff; while staff hands became less similar with increasing temperatures (Figure 4C). Hand, nose, and floor samples taken on the same week were generally more similar than those taken on different weeks for both patient and staff environments (Figure S3).

The four dominant genera identified in the hospital comprised 303 unique strains (Oligotypes): with 116, Staphylococcus, 83 Streptococcus, 77 Corynebacterium, and 27 Acinetobacter strains (Figure S4). When we applied Bayesian source tracking to the two patient rooms that were sampled daily, using the microbial profiles on hand or axilla samples from each patient as a source, it was often possible to accurately predict the correct patient based on microbial profiles on the bedrails, and to a lesser extent on floors and faucet handles (Figure S5; Figure S6). This suggests that individual patients can harbor a unique strain-level skin microbial profile, though the fact that many oligotypes were found in a large number of samples (Figure S4) limits predictive accuracy.

The complex dynamic environment of a hospital is an extreme environment for microbial life. With frequent cleaning and constant turnover of occupants this environment is purpose built to reduce the microbial interaction between occupants. We have demonstrated that the building parameters shape the microbial sharing between occupants, but that these drivers are linked to seasonal humidity trends, driven by local climate.