Identification of High-Risk Communities for Unattended Out-of-Hospital Cardiac Arrests Using GIS
Improving survival rates for out of hospital cardiac arrest (OHCA) at the neighborhood level is increasingly seen as priority in US cities. Since wide disparities exist in OHCA rates at the neighborhood level, it is necessary to locate neighborhoods where people are at elevated risk for cardiac arrest and target these for educational outreach and other mitigation strategies. This paper describes a GIS-based methodology that was used to identify communities with high risk for cardiac arrests in Franklin County, Ohio during the period 2004–2009. Prior work in this area used a single criterion, i.e., the density of OHCA events, to define the high-risk areas, and a single analytical technique, i.e., kernel density analysis, to identify the high-risk communities. In this paper, two criteria are used to identify the high-risk communities, the rate of OHCA incidents and the level of bystander CPR participation. We also used Local Moran’s I combined with traditional map overlay techniques to add robustness to the methodology for identifying high-risk communities for OHCA. Based on the criteria established for this study, we successfully identified several communities that were at higher risk for OHCA than neighboring communities. These communities had incidence rates of OHCA that were significantly higher than neighboring communities and bystander rates that were significantly lower than neighboring communities. Other risk factors for OHCA were also high in the selected communities. The methodology employed in this study provides for a measurement conceptualization of OHCA clusters that is much broader than what has been previously offered. It is also statistically reliable and can be easily executed using a GIS.
KeywordsOut-of-hospital cardiac arrest Bystander CPR Local Moran's I Single and multiple criteria clusters
Out of hospital cardiac arrest (OHCA) continues to be a major problem in the United States with an estimated 295,000 inividuals (quasi confidence intervals 236,000 to 325,000) dying per year as a result of the disease . The overall survival to hospital discharge rate across the country is just 7.9 %, a statistic that has largely remained relatively stagnant for over 30 years [25, 26]. Geographically, there is significant variation in survival rates across the country with rates ranging from 7.7 to 39.9 % in several cities studied by Nichol et al. . At the neighborhood level, researchers have observed wide variations in both OHCA and bystander-assisted CPR participation rates [25, 26]. Sasson and her colleagues found that the likelihood of a person receiving bystander CPR after a cardiac arrest event was nearly five times greater in higher income neighborhoods in Fulton, Georgia compared with lower income neighborhoods, even after controlling for individual level factors. Given the wide variability in OHCA survival rates and bystander CPR rates, an important task is identifying communities that have high risk for OHCA and targeting them for CPR education outreach, priority placement of AEDs, and other intervention activities, as a means to help to improve overall survival rates in these communities.
Various techniques exist in the spatial statistics literature for identifying areas with unusually high concentration of a phenomena, such as out of hospital cardiac arrest. Some techniques, such as the Moran’s I, Geary C or the General G statistic are called global tests for clustering because they measure spatial structure in the dataset as a whole. These measures indicate whether clustering exists in the data, but they do not identify the actual geographic location of the clusters, nor do they quantify how spatial dependency varies from one place to another. Other techniques, such as the Local Moran’s I and the Local Gi* Statistic are called local measures of clustering because they are based on subsets of the data and are able to detect spatial autocorrelation that exist within different parts of a geographic region [1, 9]. They also provide maps showing the actual location of the clusters. Univariate Local Moran’s I appears in most studies. This is a test of clustering that is based on the correlation that exists between a variable at a given location and the value of the same variable at neighboring locations. Less popular is Local Bivariate Local Moran’s I, which uses two different variables to test for clustering. The test is based on whether the value of a variable at a given location is correlated with neighboring values of another variable. Both global and local measures of clustering are widely used in disease mapping and analysis .
Kernel density estimation  is a popular technique used in GIS to identify disease clusters . This technique has the useful advantage of identifying clusters, which, similar to disease occurrences, do not follow administrative or geopolitical boundaries such as counties or electoral districts. This is a major attraction of the technique because maps produced by Local Moran’s I and Local G* statistic give the impression that disease clusters follow administrative units; are distributed homogeneously within spatial units; and that their intensity ends abruptly at the boundaries of these units. However, a major weakness of Kernel density mapping is that once the point intensity is calculated, the actual delineation of clusters is left partly to the cartographer, who must decide what level of density constitutes a cluster. Also, the technique does not calculate the statistical significance of clusters; there is no best bandwidth selection method; and, in some cases, the population bases of communities are not taken into consideration, so relative risks are not assessed .
Another group of techniques for detecting clusters relies on space–time clustering methods, which deal with identifying locations where events occur close to each other in both space and time [14, 15, 18]. These techniques can be classified as either retrospective, where the focus is on analyzing past space–time clusters , or prospective, where the focus is on analyzing emerging space–time clusters . The Knox and Kulldorff methods of space–time analysis are two widely known approaches to space–time cluster analysis, but there are limitations to these methods. For example, when calculating the Knox Index, the selection of the critical time and distance bandwidths over which clustering occurs is subjectively determined. Furthermore, these distance bandwidths may remain constant even when there are considerable variations in underlying population densities. Kulldorff’s spatial scan statistic is a more robust approach to space–time clustering, but it has been criticized on the grounds that its use of circular-shaped scanning windows prevents it from detecting certain types of irregularly-shaped clusters . Various types of flexible scan statistic have been suggested as workarounds to this problem, including one proposed by Takahashi et al. .
To date, only minimal work has been done with respect to the identification of clusters of out of hospital cardiac arrest cases using spatial statistics . The study by Lerner et al.  used kernel density analysis to identify ‘hotspots’ or high-risk areas for cardiac arrest. Their method utilized a single criterion to define the high-risk areas, thus, clusters were defined as places that had high densities of OHCA cases over a given period of time. However, recent work by Sasson et al. [25, 26] has indicated that the level of bystander CPR participation in a community is also an important variable characterizing high risk areas for OHCA and that such a factor should be incorporated into a broader definition of areas that are at high risk for OHCA.
For this study, we defined high-risk communities as those that had high incidence risks of OHCA and low incidence risks of bystander CPR rates over a period of two consecutive years. We used Local Moran’s I to identify clusters of high OHCA high rates and clusters of low bystander participation rates. These clusters were then overlaid on each other to identify the high-risk communities. Experts with detailed knowledge of the OHCA situation in Columbus, Ohio were subsequently asked to evaluate the identified communities to determine whether they matched known source areas for large numbers of OHCA cases. Our overall goal was to identify communities at the census tract level that have significantly high rates of OHCA and a low incidence of bystander CPR in Columbus, Ohio over 2 year periods. These communities were labeled as high-risk communities for unattended OHCA and once identified would be targets for CPR education outreach, priority placement of AEDs, and other intervention activities.
Cardiac arrest data for Franklin County, Ohio were used for this study. The vast majority of cases came from Columbus, the largest urban area in the county. The data came from two separate data sources. Firstly, a cardiac arrest surveillance registry managed by the Columbus Fire Department (CFD) and based on the Utstein criteria  was used to obtain all case records for events from April 1, 2004 to August 31, 2007. The data was collected through a collaborative effort between the CFD and investigators at the Ohio State University Medical Center (OSUMC). The dataset included all cases of OHCA, regardless of etiology. It also included all cases that were responded to by EMS, regardless of whether they were treated or not.
The second data source was the Cardiac Arrest Registry to Enhance Survival (CARES) registry for Franklin County, Ohio, for the period January 1, 2008 to December 31, 2009. CARES is funded by the US Centers for Disease Control and Prevention and is housed at the Emory University Department of Emergency Medicine. CARES is also partially supported by the American Heart Association . This registry excludes patients if EMS personnel determined that arrest was due to a non-cardiac etiology or if out-of-hospital resuscitation was not attempted based on local EMS protocols. The total potential patients for the entire time frame (April 1, 2004 to December 31, 2009) was 4,553. Of this amount, 3,474 cases were obtained from the CFD registry for the period April 1, 2004 to August 31, 2007. The CFD joined CARES in August 2007 and data entry started in September 2007. For the period, September 1, 2007 to April 1, 2009, the data source was jointly CFD and CARES. A total of 1,079 cases were collected during this period. Between January 1, 2008 and April 30, 2009, only CARES data were used. A total of 678 cases were obtained for this period.
Census tracts were used as proxies for neighborhoods as they tend to represent social and economically homogenous groups of approximately 4,000–7,000 people. Census tract shapefiles from the 2000 US population census were obtained from the US Census Bureau (http://www.census.gov/) along with critical socio-economic data for Franklin County such as population by age and ethnicity, median income, level of education, etc. US Census Bureau’s TIGER files (street centerline shapefiles) were also downloaded from the Bureau’s website.
Identification of High-Risks Census Tracts
A total of 4,553 cardiac arrest cases were eligible for initial consideration. Of this total, 1,998 cases were excluded for not being treated by EMS; 329 cases were excluded for presumed non-cardiac etiology; 34 cases were excluded for missing follow-up information, and 20 cases for not being in Franklin County. We also excluded an additional 403 records to remove bias in census tracts that had large number of OHCA events due to the existence of certain facilities. Thus, OHCA events that occurred in hospitals and other medical facilities (n = 51) and nursing homes/extended care facilities (n = 352) were removed from the dataset, as they tended to non-randomly inflate the cardiac arrest counts of the census tracts.
The address associated with each record was geocoded using ArcGIS 9.3 . An additional 6 cases were unable to be geocoded due to: (a) incorrectly entered zip codes in the CARES registry, (b) incorrect spelling of street names; and (c) street names mentioned by patients, but not existing in the street map used for geocoding (TIGER file). Finally, 93 cases were removed for being less than 18 years old, because the analysis is limited to adult cases of OHCA. In all, a total of 1,670 cases were eligible for inclusion into the final dataset for the calculation of OHCA CPR rates.
Our target communities were defined as those having a higher than expected OHCA incidence risks and lower than expected incidence risks of bystander CPR over a period of two consecutive years. For the calculation of bystander CPR rates, patients who experienced a cardiac arrest after the arrival of EMS personnel, or who arrested in the presence of EMS personnel were not included in the rate calculations, as these patients were not eligible for bystander CPR (n = 187). A total of 1483 cases were found to be eligible for inclusion into the final dataset for the calculation of bystander CPR rates.
After geocoding, the resultant point features were categorized by the year of the event and then, for each year, the GIS software was used to summarize the number of points that fell within each census tract. GeoDa , a dedicated spatial statistics software package, was then used to calculate Spatial Empirical Bayesian incidence risks for OHCA and Bystander CPR rates for each community. Rates were calculated based on Franklin County’s 18-year and older population at the census tract level. Spatial Empirical Bayesian incidence risks were calculated to stabilize rates in communities with small populations  Spatial Empirical Bayesian rates differ from the better-known Empirical Bayesian rates in that the calculated rates are adjusted towards the mean rate of the surrounding communities rather than the overall mean of the study area .
After calculating incidence risks, we used the ArcGIS routine for Local Moran’s I to identify communities that had higher than expected OHCA rates and lower than expected bystander CPR rates for each year for the period, April 1, 2004 to March 31, 2009. For any given census tract, Local Moran’s I was used to separately examine the OHCA and bystander CPR rates to determine whether the rates of the tract and all of its neighbors were high or low relative to the overall mean rate of the county. If the rate of a given census tract was high and the rates of the surrounding tracts were also high, then a “hot spot” or cluster of communities with high rates was identified. Alternatively, if the rate of a given census tract was low and the rates of neighboring tracts were also low, then a cold spot or cluster of communities with low rates was identified. After Local Moran's I clusters of OHCA and Bystander CPR were separately identified, they were overlaid on each other to identify overlapping areas that met both criteria.
The approach described above was selected over Bivariate Local Moran’s I which is available in the Geoda software because clusters, as defined in this study, are different from how they are defined in Bivariate Local Moran's I. In Bivariate Local Moran’s I, a cluster is defined as an area with a high (or low) value of one variable at a given location surrounded by similarly high (or low) values of a different variable at neighboring locations. Since Bivariate Local Moran’s I is based on high-high or low-low combinations of two different variables at neighboring locations, this technique was deemed unsuitable for this study because our definition of an OHCA cluster required high values of variable at a given location and low values of another variable at the same location as well as neighboring locations.
One of the main limitations of the Local Moran’s I routine in ArcGIS is that it does not adjust for the multiplicity problem. This problem arises because the reference distribution used to calculate the statistical significance of Local Moran's I for each census tract is created from repeated sampling of the same population, thereby making the local probabilities dependent on each other. Thus, if the probability of incorrectly detecting spatial autocorrelation for the study is set at 0.05, then, for totally independent tests, the probability of incorrectly detecting local spatial autocorrelation for at least one of the n locations in the distribution is 1 − 0.95n . Clearly, as n increases, the likelihood of incorrectly detecting spatial autocorrelation at a given location increases. Without adjustment to the alpha level, it is likely that some locations with elevated risks may be identified as clusters when they are actually not clusters. We used the ClusterSeer software from Biomedware to adjust for multiple comparisons. This software uses both Bonferroni and Sidak adjustments to correct the alpha levels when several locations are considered simultaneously. Instead of testing at alpha (α) = 0.05 level of significance, Bonferonni tests at the αc = α/n where n is the number of polygons in the distribution, and αc is the adjusted significance level. The Sidac adjustment uses the following formula: αc = 1 − (1 − α)1/n. The Bonferroni method is that said to be overly conservative in that while it successfully controls for the possibility of making false positives, it increases the probability of overlooking locations that are significant for spatial autocorrelation, thereby creating false negatives. The Sidac adjustment compensates for some of the weaknesses of the Bonferroni method.
Following the initial map overlays a major issue was deciding whether the areas identified as having high rates of OHCA and low rates of bystander CPR were stable enough over time to warrant policy intervention efforts. Many studies suggest that cluster communities tend to shift geographically over time so that the presence of a cluster at time t0 does not mean that the cluster will remain at that location at time t1. For example, a study of domestic burglaries in Britain by Bowers  showed that burglary clusters do not remain at the same locations for more than 3 months but rather move in a ‘slippery manner’ to nearby areas at successive points in time. For cardiac arrest clusters, it is likely that the neighborhoods that make up the clusters change their location at a much slower pace, as the demographic profiles of neighborhoods with high OHCA rates change at a much slower rate than the rate at which criminals are expected to change foraging grounds. Furthermore, although some amount of space–time clustering of cardiac arrest cases is expected, there is no evidence that suggests that OHCA clusters would change as quickly as sociologically driven events such as burglaries. Therefore, for this study, we assumed that if a cluster remained in the same location for at least 2 years, then it would be considered a “persistent” cluster.
To locate the “persistent” high-risk areas for OHCA, we resorted to a second round of map overlay. By overlaying census tracts with high OHCA rates and low bystander CPR participation rates for 1 year on similarly labeled tracts for the next consecutive year and finding the areas which spatially overlap, it was possible to identify areas that were persistently at higher risk for cardiac arrest and which could serve as potential targets for community-based interventions (Fig. 2). Since, we were interested in locating clusters that were stable over 2-year periods in which they were no temporal overlaps, we utilized data for the following periods: April 1, 2005 to March 31, 2007, and April 1, 2007 to March 31, 2009. We omitted the 2004 dataset as we wanted to have equal number of years in each time period.
Expert Assessment of Identified Clusters
After the persistent high-risk census tracts for OHCA were identified, a group of experts that included emergency response personnel and public health officials familiar with cardiac arrests patterns in Columbus were assembled to evaluate the outcome of the GIS process. A Google Map Mashup complete with the use of street view to help identify specific areas was built and the persistent high-risk census tracts were superimposed on the Google Map to make it easy for people to recognize the location of these tracts. The panel of experts did not recommend any change of designation for any of the communities identified by the GIS as being at high risk for unattended out of hospital cardiac arrests. However, they expressed the view that the GIS analysis identified a slightly more spatially concentrated set of communities compared to their knowledge of the geographic distribution of these communities. Following consensus about the high-risk tracts, summaries were made of the basic OHCA and demographic characteristics of the each tract including the number and rate of OHCA cases, median household income, and the proportions of whites, black and other races.
Locating Public Facilities for Targeted Intervention
After identifying census tracts that were at high risk for out of hospital cardiac arrest and their associated socio-economic characteristics, the next step was to locate specific public facilities, such as schools, churches, and hair dressing salons, situated either within or in close proximity to the identified tracts, and which could be potential locations for CPR educational outreach efforts. For example, to identify schools and churches, detailed lists of elementary/middle/high schools and churches were obtained from the yellow pages of Columbus, Ohio. The addresses in the lists were geocoded and points that fell in the high-risk census tracts were selected. Points that fell within 200-m buffer of the high-risk census tract boundaries were also selected. The inclusion of geocoded points that fell within the buffers was done in recognition of the fact that phenomenon such as high risk areas for OHCA occurrences are more likely to have fuzzy boundaries rather than the crisp or abrupt boundaries of census tracts . Therefore, facilities in close proximity to the high-risk tracts might themselves be in need of targeting for intervention activities.
Summary statistics for out of hospital cardiac arrest (OHCA) hot and cold spots in Franklin County, Ohio compared to the entire Franklin County
Hot spot census tracts
Cold spot census tracts
Population >65 years
9.0 % (2.44)
Median household income
Proportion below poverty level
25.32 % (7.8)
3.2 % (1.72)
Proportion of African Americans
1.61 % (1.0)
Proportion of White
51.9. % (28.4)
90.3 % (5.9)
OHCA Rate per 100,000
Bystander CPR Rate (%)
75 % (38.2)
Compared to the entire county, the high-risk communities as a whole had a slightly larger proportion of older inhabitants, i.e., people aged 65 years and older (10.34 % high-risk communities; 11.3 % Franklin County); lower median household income ($21,618 high-risk communities; $37,897 Franklin County) and a larger percentage of inhabitants living below the poverty level (28 % high-risk communities; 14.8 % Franklin County) Table 1. Table 1 also reveals that 40.29 % of persons identified themselves as being black or African Americans in the high-risk communities compared to 16.46 %, which is the proportion of African Americans living in the county. The proportion of whites in the high-risk communities was 54.4 % compared to 77.2 % for Franklin County as a whole. Summary demographic statistics for the cold spots versus Franklin County as a whole are also presented in Table 1.
Identifying areas with elevated incidences of a phenomenon is a complex task and no single method is ideal for every cluster investigation. By integrating the strengths of Local Moran’s I, map overlay, and expert knowledge of the study area, we identified areas with high rates of OHCA coupled with low rates of bystander CPR as well as areas where this relationship persisted in Franklin County. To our knowledge, this is the first project that has employed such a methodology to the OHCA population.
The census tracts identified in this study as being persistently at high risk for unattended OHCA appear plausible from a spatial statistical standpoint as well as from the point of view of experts with local knowledge of the area. The location of the high-risk tracts is similar to the general location of census tracts having high incidence risks of OHCA found in previous studies [7, 10, 13, 28, 30]. Also, the demographic profile of these communities was similar to places found in other studies that were identified as having high rates of out of hospital cardiac arrest. Such communities tend to be composed mostly of lower income individuals, frequently dominated by minorities [10, 13, 28, 30].
The practical benefit of using GIS to identify areas of OHCA clusters in Franklin County was that the technology provided speed coupled with statistical rigor in the identification of neighborhoods where public health officials could target efforts to reduce high rates of OHCA, or improve the bystander participation rates. With its spatial query features, the GIS also made it possible to quickly identify schools, churches and other community facilities located either within or close to high-risk areas that could be used for community-based OHCA intervention activities. In Franklyn County, focused groups were organized in the high-risk areas and these groups were used to understand neighborhood dynamics, particularly as it relates to social capital. Knowledge gained from focus groups and other sources were then used to better plan educational interventions that fit the characteristics of communities. These interventions focused mostly on teaching CPR, but by ascertaining the social dynamics of the communities beforehand, teaching material and strategies were designed to more directly fit the needs of the participants.
One of the main limitations of this study was that although the ArcGIS software package successfully identified and mapped clusters of communities with significantly high rates of OHCA (p < 0.05) for each year, it failed to detect clusters of communities with significant rates of low bystander CPR (p < 0.05) for at least 2 years. To overcome this problem, the bystander CPR rates were classified into quartiles and census tracts that fell into the lowest quartile were used as proxies for tracts that had statistically significant low rates of bystander CPR rates. While this a reasonable proxy, it may have resulted in slightly more communities being included in the final set of target communities. This is because selecting the lowest quartile of bystander CPR rates is a less rigorous procedure compared to a Local Moran’s I test for statistical significance. The issue raises questions as to whether Local Moran’s I is an efficient test for detecting clusters on rates that tend to be marginally close to 0. This issue needs to be explored in future studies.
A second limitation of this study is that the methods devised have only been applied to one location, i.e., Franklin County, Ohio, and therefore their generalizability is uncertain or limited. Further work needs to be done to verify how these methods would work with respect to other cities.
A third issue is that the polygons identified by Local Moran’s I routine in ArcGIS as being part of clusters represent only the core of clusters. Neighboring polygons that may also have high rates of the events being analyzed and which are also part of the cluster are not mapped. This situation also occurs with the GeoDa Spatial Analysis software. While it is true that one can investigate the map and attribute table to locate the neighborboring polygons with high rates, it is not clear in all cases, which of the polygons should be included in the cluster and many researchers simply leave out the neighboring polygons and work only with the polygons that form the core of the cluster. This latter practice has been followed in this paper, hence it should be understood, that the polygons highlighted on our maps are polygons that form the core of cardiac arrest clusters.
Finally, the persistent high-risk hot spots in this analysis are based on 2-year periods. This period was arbitrarily chosen, however, it was thought that this was a sufficient length of time for OHCA clusters to stabilize. Varying this period could result in different set of target communities being selected.
This study succeeded in identifying a number of census tracts which matched our definition of areas that comprise clusters or “hotspots” of OHCA events in Columbus, Ohio. The hotspots were located close to downtown Columbus and, compared to the rest of Franklin County, they had slightly higher proportions of people aged 65 years and older; lower median household income; a larger percentage of inhabitants living below the poverty line; and a higher proportion of persons identifying themselves as being black or African Americans. The socio-demographic characteristics of census tracts identified as hotspots in this study were very similar to the profile of census tracts described in other studies as having high rates of OHCA, and thus added validity to our findings.
An important contribution of this study is the presentation of a methodology for detecting OHCA hotspots. These hotspots can be the focus for siting educational and intervention programs targeted at improving survival of the OHCA. The distinctive aspect of the methodology employed in this study is that it allows more than one factors to define the clusters or hot spots and not just a single variable, as is customarily done. The use of Local Moran’s I coupled with traditional map overlay techniques, provided a simple, straightforward way of accommodating the two factors used to define the clusters. Overall, this approach is not only statistically robust, but the results are easy to interpret and can be used for intervention planning by the local civic and public health authorities.
- 3.Anselin, L. (2005). Exploring spatial data with GeoDaTM: A workbook. Urbana-Champaign: Spatial Analysis Laboratory Department of Geography, University of Illinois.Google Scholar
- 4.Bonham-Carter, G. F. (1994). Geographic information systems for geoscientists—Modeling with GIS. New York: Elsevier Science.Google Scholar
- 6.ESRI. (2007). ArcView 9.3 Service Pack 1. Redlands, CA: Environmental Systems Research Institute.Google Scholar
- 16.Kulldorff, M. (2001). Prospective time periodic geographical disease surveillance using a scan statistic. Journal of the Royal Statistical Society: Series A, 164, 61–72.Google Scholar
- 24.Roger, V. L., et al. (2011). Heart disease and stroke statistics—2011 update: A report from the American Heart Association. Circulation, 123, e18–e209.Google Scholar
- 27.Schuurman, N., et al. (2008). Are obesity and physical activity clustered? A spatial analysis linked to residential density. Obesity, 17(12), 2002–2209.Google Scholar