Space Time Trends of Community Onset Staphylococcus Aureus Infections in Children Living in Southeastern United States: 2002-2010

Background Staphylococcus aureus (S. aureus) remains a serious cause of infections in the U.S. and worldwide. Non antibiotic resistant Staphylococcus aureus (methicillin susceptible or MSSA) is the cause of half of all health care– associated staphylococcal infections, and methicillin resistant Staphylococcus aureus (MRSA) still is the leading cause of community onset skin and soft tissue infections in the U.S. This is the rst study to spatially look at trends of both community onset MRSA and MSSA infections over nine years and determine ‘best’ to ‘worst’ infection trends over a nine year period (2002-2010),which spanned when community onset MRSA infections were occurring in epidemic proportions across the U.S. Methods of trend, which differed signicantly from the trajectory patterns of surrounding or neighboring census tracts. This is the rst study to group S. aureus infection rates into community-onset MRSA and MSSA infection ‘trajectory’ group patterns at the community level and then, map spatially the trajectory patterns (ranging from ‘best’ to ‘worst’ trends of staphylococcal infections, stratied by their resistance to methicillin).


Background
Staphylococcus aureus (S. aureus) is a major cause of community onset and health care-associated infections, ranging from super cial skin and soft tissue infections (SSTI) to invasive infections, sepsis, and death (1). In 2017, U.S. Centers for Disease Control and Prevention reported the antibiotic resistant bacteria, methicillin resistant S. aureus (MRSA) was the cause of more than 11,000 deaths per year in the U.S. and the cause of over 80,000 invasive forms of infection (2). The non-antibiotic resistant form, methicillin susceptible S. aureus (MSSA), also is a serious cause of infections and is estimated to be the cause of half of all health care-associated S. aureus infections (1). Hospital onset MSSA infection rates have remained at a steady rate since 2012, whereas community onset MSSA (CO-MSSA) infection rates continue to increase (1). Given U.S. Department of Health and Human Services' goal for a fty percent reduction of S. aureus invasive infections by 2020 (3,4), community based primary and secondary prevention efforts to prevent spread of these infections seem more than ever important to put in place. Majority of efforts have focused on reducing infections within healthcare settings, rather than in community settings.
Risks associated with community onset staphylococcal infections have been widely reported over the last decade (5-10), but few studies have shown the relationship of 'place-based' or community level risks from a spatial or geographic lens (9,11). Spatial analyses factoring geographic areas allow for visualization of how socio-and built environments might contribute to occurrence of S. aureus infections. Little has been published on how rates of infections may differ between different geographic areas over a longitudinal time period, especially during a period when community onset MRSA (CO-MRSA) infections were occurring in epidemic proportions (1,12).
Analyses of trends in population health data have mostly been determined by comparing rate differences between two time points and changes in the linear or log linear rates over time (13). Another way to analyze trends of population is through group based trajectory models (GBTM), which incorporates information from all time points and examines non linear (quadratic, cubic, and other higher order) rate trends. In this manner, GBTM determine if groups of study units have similar trajectory patterns and can predict outcome trends of individual units which are grouped together into patterns (14)(15)(16). There are currently no studies to model or analyze multiyear trends of MRSA or MSSA rates over a longitudinal time period, identifying high or low infection trends based on location.
We sought to determine trends of CO-MRSA and CO-MSSA infections and speci cally, to identify particular trends (trajectory patterns) of infections rates which reveal areas with continued low rates of infections over nine years, relatively high rates of infection throughout this period, or 'deviant' trends (high rate with one type of S. aureus infection and low rate with the other.) We also sought to determine areas where clusters of trajectory patterns occurred and speci cally, to identify geographic deviants, where 'cluster' of census tracts had one type of trend, which differed signi cantly from the trajectory patterns of surrounding or neighboring census tracts. This is the rst study to group S. aureus infection rates into community-onset MRSA and MSSA infection 'trajectory' group patterns at the community level and then, map spatially the trajectory patterns (ranging from 'best' to 'worst' trends of staphylococcal infections, strati ed by their resistance to methicillin).

Methods
Overview. A retrospective study was conducted of children < 19 years who were treated for S. aureus infections from January 1, 2002 through December 31, 2010 at two pediatric hospitals in Atlanta, Georgia (Scottish Rite Children's Hospital and Egleston Children's Hospital are part of Children's Healthcare of Atlanta (CHOA), a large pediatric healthcare system in Atlanta, Georgia, which consistently provides more than 80% of the pediatric hospitalizations for the Atlanta metropolitan statistical area (MSA)) (12). Patients < 19 years of age diagnosed with S. aureus infection who had emergency department visit (ED) and/or inpatient admission with residential addresses within the 20 counties of the Atlanta MSA were included. All study participants had an International Classi cation of Diseases, Clinical Modi cation (ICD-9-CM) code compatible with S. aureus infection and a positive S. aureus culture as previously described (12,17). We included only those patients who met the de nition for community-onset infection (patients with positive S. aureus culture within 48 hours from the time patient was initially evaluated or admitted) (18)(19)(20).
Individual Level Patient Data from Electronic Health Records. Demographic data (race/ethnicity, gender, type of health insurance, and place of residence); relevant laboratory information (S. aureus culture site(s), antibiotic susceptibility phenotype patterns); and relevant clinical information (clinical diagnoses, prior hospitalizations or ED visits within 12 months of the date of admission, number and source of cultures positive for S. aureus, recorded past medical history of chronic illnesses) were abstracted from patients' electronic health records (EHR). Georeferencing of the patients' U.S. postal addresses was performed after excluding addresses which were con rmed to be out-of-state or U.S. post o ce box numbers. All georeferencing was performed using ArcGIS 10.6 (ESRI, Redlands CA). See Fig. 1 . In group-based trajectory modeling, discrete underlying groups within the population are assumed to have their own case intercept, slope and possibly higher order terms. Proc Traj requires speci cation of the number of groups the model will t. We used a process of evaluating model t while simultaneously identifying informative similar trajectory groups as previously described by Baltrus et al (16). We estimated a zero-in ated Poisson model with a rst order, quadratic term for each outcome, and an independent variable of time (years) for each group. We continued adding groups and assessed the change in the Bayesian information criterion (BIC) as an evaluation of model t. Since Proc Traj has the limitation that does not allow for an offset variable in order to generate rates, we adjusted for each census tract's < 19 years of age population as a time varying covariate. (The CO-MSSA and CO-MRSA overall model t did not improve with the addition of more than two groups. The CO-MRSA model produced a small "very high infection" group of census tracts with the addition of a third group into the model with only a slight decline in the t of overall model.) We next added or removed second and higher-order terms from each group's model based on signi cance (p < 0.05). level variables, normality of distribution was determined using Shapiro-Wilk test. For normally distributed population, ANOVA was used when comparing three trend groups and T-test when comparing two trend groups. Kruskal-Wallis test was used for three groups, and Mann-Whitney test for two groups when population was not normally distributed. All statistical analyses were done by SAS version 9.4 (SAS Institute, Inc). Local 'clusters' of aggregated patients' point data were determined by applying the local indicators of spatial association (LISA) technique as previously described (16,24,25). From this, hot and cold spot analyses were performed using GetisOrd Gi* statistic. The output is a surface raster layer depicting the different degrees of 'hot spots' and 'cold spots' of CO-MRSA or CO-MSSA occurrences. Signi cant clusters were those areas, where an occurrence of infection (CO-MRSA or CO-MSSA) and its neighboring points all had high Getis-Ord Gi* values (hot spots). Geographic deviants were de ned as clusters that had much higher or much lower values than neighboring occurrence points (cold spots).

Spatial Patterns for Community
Spatial Analyses. Optimized Hot Spot Analysis (OHPA) tool (Esri ArcGIS Pro, Redlands, CA) was performed to assess spatial patterns in our data. This allowed us to visualize statistically signi cant patterns, where clustering of cases (CO-MRSA) or controls (CO-MSSA) exist, which are not likely due to random occurrence. The results of the applying this tool showed us areas within the 20 counties where 'hot spots' occur within a 'cluster' in a non random way. (A simple 'hot spot' analysis will not tell necessarily inform "why" a particular event is occurring in a statistically signi cant cluster, only that it is occurring in a non random manner). We ran the analysis using all unique patients' point locations from 2002 to 2010, < 19 years, with CO-MRSA and then repeated these steps for CO-MSSA patients within the same time period and for the same ages. With the OHSA tool, we set our 'Study Area' to the boundaries of Atlanta's MSA 20 counties. We then assigned 'neighborhood' variables as 1-mile hexgon patterns. We chose the 1-mile hexagon neighborhood size based on Tober's First Law of Geography, which follows the spatial diffusion theory for infectious pathogens, whereby the risk for infection increases with closer distance to source of infection. In our model, we postulate that the closer an individual's neighborhood is geographically near to an occurrence of CO-MRSA or CO-MSSA, the more likely the persons within the neighborhood are likely to contract CO-MRSA or CO-MSSA. In this model, we accounted for various sizes of neighborhoods, apartment complexes, and other living or housing units. We estimated that a neighborhood is within approximately a mile of the residence of any CO-MRSA or CO-MSSA occurrence. This study was approved by Institutional Review Boards of hospitals and a liated academic institutions.  Table 2; no census tracts fell into the group-based trajectory pattern of Very High CO-MRSA -High CO-MSSA infections. Figure 2 shows the three CO-MRSA group-based trajectory trends and the two CO-MSSA group-based trajectory trends over this nine-year period: Of the three CO-MRSA group-based trajectory trends, only 0.8% of the census tracts showed a dramatic increase over the rst six years followed by a gradual decline in the last three years; over two thirds (67.2%) remained in the low infection trend across all years. In contrast, for CO-MSSA, 85.7% of census tracts fell into the high infection group-based trajectory pattern which lasted throughout the nine-year period, compared to 14.3% of census tracts which belonged to a low infection trend over this same period. (Supplemental Section: Table A shows the coe cients for the terms used in the three group-based CO-MRSA trajectory models and the two group-based CO-MSSA trajectory models. For the three CO-MRSA groupbased models, the equations all contained a quadratic term. For the two CO-MSSA group-based models, the equations contained a quadratic term, and a slope term that was relatively at.)   ** Other races include: Asian, Native Hawaii, Alaskan, and Multiple-races (self identi ed by patients as more than one race).

Results
***Other health insurance types include: Self-pay, no insurance, and those with both private and public.   There was a large band of cold spots, located in the northwestern portion of 20 counties boundaries, which ran along the perimeter of hot spots seen in Cobb and Cherokee counties (northwest of the most densely populated counties) (Figure 4). In order to discern what group-based trajectory patterns might fall within 'hot spots', we overlaid the group-based trajectory map with hot spots and found that the 'worst' group-based trajectory was located in areas found to be 'hot spots' for CO-MRSA and CO-MSSA, and the best trajectory groups tended to occur where there were neither cold nor hot spots. Interestingly, Bartow county (northwest of Atlanta's downtown area) had a number of census tracts which had a CO-MRSA-high -CO-MSSA-low designation, yet no CO-MSSA or CO-MRSA 'hot' or 'cold' spots.

Discussion
We applied group-based trajectory modeling to identify Atlanta MSA census tracts with 'worst' and 'best' temporal trends of community onset MRSA and MSSA infection rates over nine years. We identi ed three distinct CO-MRSA group-based trajectory infection patterns and two CO-MSSA group-based trajectory infection patterns over this period, which included the time when community associated MRSA infections were occurring in epidemic proportions around the U.S. We found ~ 94% of all the census tracts with S. aureus occurrence during this time period belonged to the best trend categories of low infection rates for both CO-MRSA and CO-MSSA. With groupbased trajectory models, we used all the rates generated over the nine year span instead of relying on change between just two rates at the beginning and end of this period. This allowed us to detect and examine the trajectory shape over the entire time period, including non-linear shapes which are common for the epidemic curves associated with communicable diseases. Given how this bacteria, S. aureus, can have many strains with various virulent factors that may contribute to infection, understanding these epidemic curves within any community is important. It provides some insight as to which communities may have higher proportion of its population, with risks for infection. A recent worldwide review of S. aureus nasal carriage estimated that the average prevalence of nasal colonization in the general population is 24%(26) but the possible place based risks which contribute to movement from colonized to infected population have not been elucidated. We have clearly delineated population level crowding is associated with worst trajectory trends for CO-MRSA infection in speci c communities at the census tract level. Furthermore, we also identi ed the groups which make up the trajectory patterns from both individual and area level data instead of de ning group categories a priori. This study highlights not only individual and area level risks which may contribute to why certain census tracts over time have either a high or low rate of infection, but also the location-based relationships which may be explanatory of these staphylococcal trend patterns of infection.
Our main ndings demonstrated that variations in spatial trend patterns of CO-MRSA infection and CO-MSSA infection rates across these 20 counties include both rural and urban communities. We mapped out the geographic clustering of the group-based trajectory patterns, with 'worst' trajectories clustering mainly in Dekalb and northwest counties of Bartow, Cherokee, Paulding and Cobb, and 'best' trajectories clustering in the northeast, southeast, and southwest of the inner-city area. Identi cation of these clusters of census tracts with similar rate trajectories over time will direct us to better understand what 'place based' conditions or factors might contribute to these ndings that are location speci c. For example, population density of the 'worst' trajectories, along with demographic factors including race and age distribution have been demonstrated to partially explain the unfavorable trajectory seen in DeKalb County. However, the southwest section of Bartow County has a signi cant group of census tracts which are geographically deviant from the surrounding 'low infection' census tracts. Socioeconomic factors (e.g., poverty, inequality index, employment status, etc.) may be more of a 'driver' than population density or household crowding in these rural communities.
When considering demographic characteristics of the 901 census tracts and the distribution based on ve distinct S. aureus infection group-based trajectories, we found that race differences were seen in census tracts where infection trends were high for both MRSA and MSSA infection trend patterns. Areas with low S. aureus infection trends had higher proportion of whites compared to blacks, whereas high infection trend patterns occurred in areas where there were higher proportion of blacks. (In Gwinnett county, 29.3% of the population are black and 54.5% are white, but the pockets with much higher trends of CO-MRSA correspond to census tracts with higher proportion of blacks.) Many of the census tracts found to have the 'worst' CO-MRSA infection trajectories were in areas with lower housing value, lower household income, and higher poverty rates. These ndings are consistent with ndings published from CDC researchers in their two year population surveillance study of 33 counties across nine states, which also demonstrated higher incidence of MRSA infection from census tracts with low-household income, persons living under the poverty level, and low education (27). They also found that 'crowding' was associated with increased MRSA rates. Our analyses, which focused on 'trend' patterns of MRSA infections over a nine-year period and not incidence rates, did not nd crowding to persist over this nine-year time period. This may be related to the fact that our data compared a longer time interval (nine years compared to two years) and was based on analyses of multi-level data, including individual patient level data. Another explanation maybe related to communities in our catchment area may have improved housing conditions between the early and late time periods, e.g., higher percentage of people living in less crowded housing situations over almost a decade of time.
Pro les of socioeconomic conditions in areas where CO-MSSA trajectories were high differed from the pro les where CO-MRSA trajectories were high: In areas where CO-MSSA infections trended high, the proportion of whites was higher than proportion of blacks. This relationship was reversed in areas of high or very high CO-MRSA groupbased infection trajectories, where proportions of blacks were higher than whites.
Higher household income and housing values and lower individual poverty levels were seen in areas with high CO-MSSA-infection trend patterns, compared to low CO-MSSA infection trend patterns. Whether or not this is a temporal effect related in part by the fact that some of these CO-MSSA-high infection census tracts crossover with CO-MRSA-low infection, is unclear.
Overall, multi-level factors contribute to the explanation of geographically 'deviant' communities whose trajectory group is surrounded by very different types of infection trend patterns. Our study did not look at the changes of demographic composition as a dependent variable or other factors, e.g., changes in socioeconomic conditions, housing conditions, etc. which may potentially contribute to the geographic deviation found. Interestingly, we did not see a 'Hispanic paradox' (28), where census tracts with higher proportion of Hispanics were in areas with the best infection trajectories. In our analyses, we did not see signi cant differences among Hispanics for any of the CO-MRSA group based trajectories and in fact, found worsening infection trend rates among Hispanics in the low and high CO-MSSA group based trajectories over the nine-year period.
The use of this novel method of group-based trajectory analysis (GBTA) of communicable disease trends holds some advantages over traditional hot-spot analysis. Hot spot analysis is good for spotting where concentrations of cases have occurred over time. However, hot spot analyses may not always identify geographic areas where residents are at highest risk, since the analyses does not take into account the population size where cases are actually occurring. For instance, a tract that has 5 cases arising over the nine year period may not show up in a hot spot analysis; however, if that tract has a relatively low population it may show up as a high infection rate trajectory tract in GBTA. This may be why some tracts in the more suburban areas in our study were found to be high infection rate trajectory tracts, yet did not show up as 'hot spots'. The practical clinical application of using GBTA can be demonstrated, when estimating whether or not a patient who presents with a skin or soft tissue infection has CO-MRSA or CO-MSSA, based on the patient's assigned group based trajectory infection category. In this scenario, the treating healthcare provider could factor location based CO-MRSA or CO-MSSA infection trends as part of the management strategy. Another advantage is that with GBTA, we can ascertain a general idea of the rate pattern experienced over the study period by tracts without having to examine the hot spot maps for each year.
Limitations. There are several limitations to this analyses: First we were not able to determine causal association (relationship) between the population characteristics of an area and CO-MRSA and CO-MSSA infections, since the study draws from area level data which were collected in two different time periods (2000 and 2010). Although we have included children seen for all staphylococcal infections within one healthcare system, our ndings may not necessarily be generalizable to children who did not access care from this single pediatric healthcare system or applicable to other geographically distinct locations where the place based factors are different from this catchment area.
Future Directions. We plan to look at geographic in uences from changes which are place based, using particular 'neighborhoods' as boundaries, and not census tracts. These 'neighborhood' boundaries tend to outline areas where the sociocultural in uences are similar among the residents of the neighborhood, e.g., types of healthcare access, density and types of daycare centers, mode and availability of transportation, and racial and ethnic groups including various immigrant populations or cultural in uences.
Our study clearly outlines which census tracts have the greatest disparities in temporal S. aureus infection trends, and speci cally, which census tracts are located in the 'worst' or 'best' group-based S. aureus infection trajectory areas. It is these 'geographic deviants' that deserve more attention. Identifying the potential remediable factors may pave the path towards eliminating spread of these types of infections or at least, mitigate contributing factors to the spread. Similarly, looking at why certain census tracts perform consistently better than their surrounding areas will provide insights on how to prevent the spread of this bacterial infection, and thereby, improve outcomes for other census tracts which have failed to have a positive infection trajectory trends over time. Deciphering the primary multi-level factors, driven by place and time elements, for both antibiotic resistant and non-antibiotic resistant S. aureus infections is paramount in the effort to develop effective prevention models that effectively and e ciently decrease transmission and infections at the neighborhood or community level.
Conclusions. We demonstrated S. aureus infection trends over a nine-year period, factoring in time and space for both community onset MRSA and MSSA. We identi ed speci c areas which are unique or overlapping at the census tract level for these staphylococcal infections, and speci cally, identi ed unique areas which consistently proved to have high infection rates and performed the 'worst' over the nine year period. Group-based trajectory modelling allows us to not only identify areas which perform better, worse, or remain essentially, unchanged, but also identify the spatial relatedness of these areas. There is a continued need to develop strategies aimed at primary and secondary prevention, especially at the community level. Therefore understanding 'trends' of infection patterns, strati ed by antibiotic resistance, over geographic areas and time can provide evidence-based information, necessary for community-speci c prevention guidelines.   territory, city or area or of its authorities, or concerning the delimitation of its frontiers or boundaries. This map has been provided by the authors.