Geographical Variability in the Likelihood of Bloodstream Infections Due to Gram-Negative Bacteria: Correlation with Proximity to the Equator and Health Care Expenditure

Objective Infections due to Gram-negative bacteria exhibit seasonal trends, with peak infection rates during warmer months. We hypothesized that the likelihood of a bloodstream infection due to Gram-negative bacteria increases with proximity to the equator. We tested this hypothesis and identified geographical, climatic and social factors associated with this variability. Design We established a network of 23 international centers in 22 cities. Setting: De-identified results of positive blood cultures from 2007–2011 and data sources for geographic, climatic and socioeconomic factors were assembled for each center. Participants Patients at the 23 centers with positive blood cultures. Main outcome Due to variability in the availability of total culture volumes across sites, our primary outcome measure was the fraction of positive blood cultures that yielded Gram-negative bacteria; sources of variability in this outcome measure were explored using meta-regression techniques. Results The mean fraction of bacteremia associated with Gram-negative bacteria was 48.4% (range 26.4% to 61.8%). Although not all sites displayed significant seasonality, the overall P-value for seasonal oscillation was significant (P<0.001). In univariate meta-regression models, temperature, latitude, latitude squared, longitude, per capita gross domestic product and percent of gross domestic product spent on healthcare were all associated with the fraction of bacteremia due to Gram-negative bacteria. In multivariable models, only percent of gross domestic product spent on healthcare and distance from the equator (ie. latitude squared) were significantly associated with the fraction of bacteremia due to Gram-negative bacteria. Conclusions The likelihood of bacteremia due to Gram-negative bacteria varies markedly between cities, in a manner that appears to have both geographic (latitude) and socioeconomic (proportion gross domestic product devoted to health spending) determinants. Thus, the optimal approach to initial management of suspected bacteremia may be geographically specific. The rapid emergence of highly antibiotic-resistant Gram-negative pathogens may have geographically specific impacts.


Introduction
Gram-negative bacterial infections, including bloodstream infections, are increasingly recognized as exhibiting seasonal trends, suggesting that some variation in the predominance of infecting bacteria may be due in environmental conditions [1][2][3][4]. In the United States, data derived from adult and pediatric intensive care units demonstrated that Acinetobacter baumannii infections were significantly more common in July-October than in November-June, particularly bloodstream infections [3]. Another study demonstrated a dose-response increase in monthly incidence of infection by Pseudomonas aeruginosa and Acinetobacter baumanii [1]. In a study involving hospitals in four continents, Klebsiella pneumoniae bloodstream infections were 1.5 times more common during the warmest months of the year [5].
Such variability may provide clues regarding the potential role of the physical environment in the pathogenesis of bloodstream infections. Environmental contributions to pathogenesis could have important implications in the face of ongoing climate change [6]. Such findings also imply that the optimal approaches to the initial management of bacteremia would vary geographically and temporally, as optimal empirical management of bloodstream infections must balance the risks of inadequate coverage against risks to both individuals and the population associated with excessively broad-spectrum treatment [7,8]. However, a broad, comparative international assessment of bacteremia risk has not been available to guide such decisions.
We established a network of infectious disease physicians and microbiologists in 23 international medical centers (in 22 cities), and utilized assembled blood culture data to evaluate variability in the occurrence of Gram-negative bacteremia, and to identify geographical and social factors associated with between-region variability.

Study Design & Setting
An International Steering Committee was established. Hospital-based infectious disease and microbiology experts at different latitudes, and on different continents, were invited to participate in the study. Recruitment of the participating institutions was done by the International Steering Committee. An invitation email was initially sent to 89 centers and a detailed study proposal written by the International Steering Committee was subsequently shared with the contacts that expressed interest in participating in the study. Investigators representing 23 medical centers located in 22 cities (with two hospital groups participating from São Paolo, Brazil) agreed to participate in this study (

Selected datasets
All participating centers provided de-identified results of positive blood cultures from 2007 to 2011 (Table 2). Only the first bloodstream infection per pathogen was included in the analysis. Hence, if a patient had multiple positive blood cultures growing the same pathogen, this would be considered as one bacteremic episode. All pathogens were counted separately if more than one microorganism was isolated from a single blood culture.
The following microorganisms were considered as contaminants and excluded from the data collection process: Bacillus sp, Corynebacterium sp (except Corynebacterium jeikeium), Lactobacillus sp, and Propionibacterium sp. Coagulase-  negative staphylococci and viridans-group streptococci were included in the analysis if they were recovered from two separate blood cultures, regardless of the time interval between the two positive cultures. For each participating study center, data was obtained regarding latitude, longitude, mean annual precipitation, mean daily temperature, population density, per-capita gross domestic product and the percentage of gross domestic product in that country that was allocated for healthcare. Data sources for geographic, climatic, and socio-economic covariates are presented in Table 1.

Statistical methods
Several possible measures of local propensity towards bacteremia due to Gramnegative bacteria are possible; however, the geographic and socioeconomic heterogeneity among participating sites led us to suspect that apparent incidence of bacteremia due to Gram-negative bacteria might be affected by different acuity thresholds for admission to hospital (if patient-days were used as a denominator) and propensity to culture (if total blood cultures were used as a denominator). Thus, we decided to estimate the incidence of bloodstream infection due to Gramnegative bacteria in two ways: first, as all hospitals provided data on the fraction of blood cultures that yielded Gram-negative bacteria and non-Gram-negative bacterial pathogens, we evaluated the Gram-negative bacteremia fraction, defined as Gram-negative bacteremia divided by all bacteremias. Second, as total culture volumes (positive and negative) were available only for a subset of thirteen hospitals, we performed exploratory analyses on the incidence of Gram-negative bacteremia using monthly admissions (available from all participating sites) as model denominators (''offsets''), and adjusting for total culture volumes as model covariates.
For the fraction of bloodstream infections due to Gram-negative bacteria estimates, we converted monthly proportions of positive blood cultures found to contain Gram-negative bacilli to their logits (natural logs of odds), with odds estimated as P GNB P GNB , where proportion of bacteremias caused by Gram-negative bacilli is denoted as P GNB and P GNB~1 {P GNB . Variance of the logit is approximated as , where N is the number of isolates [9].
As propensity to culture appeared to be a strong predictor of crude Gramnegative bacteremia incidence, we estimated culture-adjusted incidence of Gramnegative bacteremia per 100 admissions using Poisson regression analysis for the 13 centers for which monthly culture volumes were available [10]. Cultureadjusted baseline incidence density of Gram-negative bacteremia was estimated as the exponentiated intercepts from models that included total culture volume (negative or positive) as a covariate. This quantity can be interpreted as the baseline incidence of bloodstream infections due to Gram-negative bacteria as culture volumes approach zero. We also utilized these models to evaluate cultureadjusted seasonality and waveforms for Gram-negative bacteremia, by adding Fourier transforms to our models such that the estimated incidence of Gramnegative bacteremiaÎ GNB À Á was: I GNB~e a Ã zb 1 sin 2pÃmonth 12 À Á À Á zb 2 cos 2pÃmonth 12 is a Fourier transform representing annual seasonal oscillation, and a Ã is a composite intercept term that includes both the expected baseline Gram-negative bactereremia culture count, and the monthly admissions offset (denominator) [10,11]. The b 3 adjusts for total culture submissions. Nonlinear variance estimates were obtained via the delta method. Coefficients from Fourier transforms were utilized to estimate phase as 52 1z . We approximated standard error for phase and amplitude terms as the sum of standard errors of Fourier terms from Poisson regression models, minus covariance [10,11].
A significant between-site heterogeneity was identified in the fraction and incidence of bacteremia due to Gram-negative bacteria, we used meta-regression methods [10] to identify geographical, ecological and economic characteristics of cities (latitude, longitude, mean annual temperature and precipitation, population density, national per capita gross domestic product and per capita healthcare spending as a fraction of gross domestic product) that we postulated a priori might explain between city differences in Gram-negative bacteremia. We also evaluated income inequality based on Gini coefficient, a widely used quantitative metric of income distribution [13]. As we postulated that distance from the equator, north or south, might explain changing Gram-negative bacteremia risk, we modeled latitude as a quadratic term [14]. We evaluated the extent to which between-site heterogeneity in waves was explained by model covariates using the tau-squared statistic [15,16].

Results
The locations of 23 participating centers are displayed in Fig. 1 (with two centers located in Sao Paolo, Brazil), and their clinical, geographic, climatic, and socioeconomic characteristics are displayed in Table 1. Cities were arrayed across a range of latitudes spanning 91 degrees (with each degree of latitude equal to approximately 111 km), and a range of longitudes spanning 264 degrees (with each degree signifying a variable distance, depending on latitude, but equal to 85 km at 40 degrees north or south latitude). A total of 1074 months of blood culture data were collected from a total of 23 participating centers in Table 2. Data for two centers located in Sao Paolo were aggregated, such that analyses were based on 22 cities.
The mean fraction of bacteremia associated with Gram-negative pathogens was 48.4% with a range from 26.4% (Iowa City) to 64.4% (Porto Alegre) (Fig. 2). When we assessed the fraction of bloodstream infections due to Gram-negative bacteria via evaluation of logits, we identified extreme heterogeneity (Q-statistic 3547.8 on 21 d.f., P,0.0001).
Culture volumes were highly variable and closely correlated with crude incidence of Gram-negative bacteremia per 100 hospital admissions as well as crude incidence of Gram-negative bacteremia (Spearman's, P50.002). Logits for the fraction of bloodstream infections due to Gram-negative bacteria were most closely correlated with intercepts from culture-adjusted models than with crude incidence of Gram-negative bacteria estimates (Table 3). Thus we used fraction of bloodstream infections due to Gram-negative bacteria (primarily, due to completeness of data) and culture-adjusted incidence (secondarily, due to lack of total culture volumes for nine of 23 sites) as our indices of Gram-negative bacterial bloodstream infection risk.
We evaluated the seasonality of culture-adjusted incidence of Gram-negative bacteremia using Fourier transforms from culture-adjusted incidence models (Fig. 3); although not all sites displayed significant seasonality, the overall P-value for seasonal oscillation was significant (P,0.001). However, the amplitudes of   There was a linear relationship between phase (peak month of occurrence) and latitude (P50.013) but this relationship simply reflects the inversion of summer and winter months in northern and summer hemispheres, and no relationship was seen between phase and latitude in analyses restricted to the northern or southern hemispheres (Fig. 4).
In univariable meta-regression models, several site characteristics, including temperature, longitude, latitude, latitude-squared, per-capita gross domestic product, and percent of gross domestic product spent on healthcare, were associated with the fraction of bloodstream infections due to Gram-negative bacteria. However, in multivariable models, only percent of gross domestic product spent on healthcare and latitude-squared (ie. distance from the equator) were associated with fraction of bloodstream infections due to Gram-negative bacteria at the P,0.05 level (Table 4). Greater explanatory power was seen when mean annual temperature was retained in the model, though this association was not statistically significant (P50.135). The resultant multivariable model explained approximately 64% of between-site variation in fraction of bloodstream infections due to Gram-negative bacteria (Fig. 5). No factors explained betweensite variation in culture-adjusted intercepts from incidence of Gram-negative bacteremia models (Table 5). It should be noted that the interpretation of Fig. 4. Relationship between peak month of occurrence of bloodstream infection due to Gram-negative bacteria (Y-axis) and latitude (X-axis) (circles). It can be seen that while there is a strong linear overall relationship between month of peak occurrence and latitude, this largely reflects the inversion of summer and winter months in the northern and southern hemispheres (fitted solid black line). Within the northern hemisphere (fitted dashed line) and the southern hemisphere (fitted gray line) there is no relationship between distance from the equator and month of peak incidence.  Geographic Variability of Bacteremia latitude-squared coefficients is complex; in our best-fit model for the fraction of bloodstream infection due to Gram-negative bacteria, there is not a constant odds ratio associated with a given change in latitude. Rather, the reduction in Gram-  negative bacteremia risk associated with moving from 0 to 10 degrees latitude would be less than that seen moving from 10 to 20 degrees latitude. This relation is presented graphically in Fig. 6.

Discussion
Understanding the epidemiology of bloodstream infection due to Gram-negative bacteria is of great importance as the global spread of highly antimicrobial resistant strains is becoming more prevalent [17]. Previous single state or single country studies have demonstrated a greater incidence of bloodstream infection due to Gram-negative bacteria during warmer months [1,2,4,18,19,20,21]. This observation may reflect optimal growth conditions for many Gram-negative bacteria at 32-36˚C. Based on this data, one would suspect that the likelihood of a bloodstream infection due to Gram-negative bacteria compared to Gram-positive bacteria would generally correlate with distance from the equator. Thus, we hypothesized that bacteremia due to Gram-negative bacteria would be most common in medical centers closest to the equator. We assessed this possibility among medical centers at different latitudes in Europe, South America, North  Table 4. The log-odds of bloodstream infections due to Gram-negative bacteria (logit, blue curve) is calculated as 20.0003 x latitude 2 , while the odds ratio for GNB (relative to the odds at the equator) is this quantity exponentiated (red curve). The change in odds ratio per 10 degree increment is non-constant by latitude, and is presented as the green curve. America, Australia, and Asia. We confirmed seasonal variability of Gram-negative bacteremia as demonstrated by other investigators. More strikingly, using multivariable, meta-regressive modeling, we demonstrated that the likelihood of bloodstream infection due to Gram-negative bacteria inversely correlated with the percentage of gross domestic product spent on healthcare and distance from the equator measured as latitude-squared. Thus, for bacteremic patients, the likelihood that the infection is due to Gram-negative bacteria is significantly greater in locations closest to the equator, during warmer months of the year, and in locales with a lower amount of gross domestic product spent on healthcare. Lastly, our findings are supported by a recent study that detected a direct correlation between the proportion of the human microbiome made up of Firmucutes (i.e., Gram-positive bacteria) and latitude [22]. What distinguishes our study from previous studies? We included 23 medical centers from different continents across five calendar years, we used a common approach in defining bacteremia, we used logits as a common metric and we controlled for a number of confounding variables in our models. Our data suggests that multiple mechanisms may be at play regarding the likelihood that a bloodstream infection is due to Gram-negative bacteria versus other pathogens, some environmentally-based and others that have economic underpinnings related to healthcare spending. It is hoped that future studies will allow a better understanding of how variables interact leading to the findings we observed. Additionally, one might interpret our seasonal waveform analysis to suggest that bloodstream infection due to Gram-negative bacteria could be less common further from the equator if such infections only predominate during seasonal surges. However, in aggregate, we found oscillatory patterns in Gram-negative bacteremia at all latitudes. We did, on average, observe a flip in waveforms but very heterogeneous and no clear patterns in amplitude that would explain bloodstream infections due to Gram-negative bacteria near the equator via ''all year round persistence''. Thus, additional factors such as the physical environment, weather, or topography may have an as yet unexplained impact on this phenomenon.
Our study has limitations. The effects we evaluated were ecological in nature and consequently may be subject to bias. We did not differentiate communityacquired from hospital-acquired bloodstream infection or whether or not they were primary from secondary bloodstream infections. We did not control for inhospital temperature and humidity and we did not collect data regarding whether or not participating centers had air conditioning for their inpatient wards which could have impacted hospital-acquired bloodstream infections. We also did not control for altitude of the hospitals that participated in this study. The fact that the data collection process did not provide information regarding the frequency with which potential contaminant organisms such as Bacillus sp. or a Propionibacterium sp. were recovered and excluded or the frequency with which coagulase-negative staphylococci and viridans-group streptococci were excluded from the data collection, in the case that they were isolated only in a single blood culture. It is possible that variations in the blood culture collection practices over time and between centers may have biased our study results by underestimating the true frequency of Gram-positive bacteremias. However, we did address this limitation in data collection by using monthly admissions that were available from all participating sites as a denominator. Also, some Gram-positive bacterial infections, such as those due to Staphylococcus aureus, follow a seasonal pattern and they may vary geographically according to distance from the equator; however, we did not assess for that possibility in this study.
Although 23 centers is the largest cohort to date to study this issue, it is not a random sample of geographically distributed hospitals and it still represents a small proportion of all hospitals globally. Additionally, use of blood culture data limits the interpretability of the findings since risk factors for bloodstream infection could represent increased risk of acquisition of the Gram-negative bacteria, increased risk of infection once already colonized, or a combination of both. Lastly, we did not determine which of the many species or groups of Gramnegative bacteria were responsible for the changes noted in this study.
In summary, the likelihood that a bloodstream infection is due to Gramnegative bacteria compared to other bacteria correlates with distance from the equator and healthcare spending. We hope that future studies will build on these observations adding to the existing knowledge regarding the epidemiology of Gram-negative bacteremia and assist in the control of such infections in different parts of the world.