Cardiac arrest (CA) is one of the most prominent causes of mortality identified worldwide. The number of years of potential life lost is estimated to be more than three million in the USA . In France, we estimate that the number of out-of-hospital CA events is 46 000 with a 5% survival rate .
Several studies have been dedicated to the spatial analysis of CA incidence variations, generally to understand and improve the automated external defibrillator location , bystanders and emergency medical team interventions [5–7] and survival rates [8–10]. Some studies considered socioeconomic status as an important factor for explaning differences in CA incidence [569–14]. The aim of these studies was to adapt public health measures within a territorial and sociodemographic context. They generally found a link between these variables, some of which were significant. However, to the best of our knowledge, only two studies have used scan statistics to analyse CA epidemiology  and to identify CA high-risk geographic areas (census tracts). This method has already been used in the analysis of other pathologies such as Crohn’s disease . In 2012, Sasson et al.  concluded that ‘future research will be conducted to better understand how race, socioeconomic status, and educated attainment affect the likelihood of being a high-risk census tract’.
In this framework, the main aim of our study is to implement an original two-step approach that first aims at identifying the distribution of CA incidence clusters using a Bayesian approach. Second, we perform the determination of subincidence and over-incidence clusters using scan statistics. The secondary aim is to socioeconomically characterize these clusters.
Patients and methods
In 2014, the French population was ~66 million, spread across 101 administrative counties (http://www.insee.fr/). In each county, the emergency medical service (EMS) system is a two-tiered, physician-based system including a fire department ambulance for prompt intervention and basic life support (BLS), a single ‘SAMU’ medical dispatch centre and several prehospital emergency departments that include one or more Mobile Medical Team (MMT) mandatorily composed of an emergency physician, a nurse and an emergency medical technician .
Cardiac arrest data collection
CA data were extracted from RéAC. RéAC is the French national CA registry created in June 2011. RéAC includes OHCA patients of any age irrespective of the aetiology and, when an MMT was involved, irrespective of resuscitation attempt. All EMS involved in the registry used a specific data paper form: the RéAC form. This form meets the requirements of French EMS organizations and is structured according to the Utstein universal style. It contains six categories of information: sociodemographic data, schedules and time intervals, CA history, BLS description, ACLS description and the immediate patient outcome. If the patient is alive at hospital admission, a 30-day follow-up record sheet must also be filled in and entered into the database .
To collect and compile data, the RéAC uses a secure internet-based data collection system. Currently, more than 80% of French EMS participate in the registry. The three EMS involved in our study (SAMU 92 ‘Hauts-de-Seine’, SAMU 93 ‘Seine-Saint-Denis’, SAMU 94 ‘Val-de-Marne’) were selected according to their volunteered, exhaustive and complete participation during the study period (August 2013–August 2015).
This study was approved by the French advisory committee on information processing in health research (CCTIRS) and the French National Data Protection Commission (CNIL, authorization no. 910946). It was approved as a medical assessment registry without the requirement for patient consent.
Data collection for spatial analysis
This study was carried out in the ‘small crown’ area around Paris, which is composed of three counties in the immediate vicinity: Hauts-de-Seine, Seine-Saint-Denis and Val-de-Marne. It has a surface area of 657 km2, with ~4.533 million inhabitants in 2014. The 123 municipalities in this region were used to represent distinct spatial units. Patients were ranked by age groups of 5-year interval lengths.
Data collection for socioeconomic status
For the first level of analysis, we used a deprivation index, the Human Development Index 2 (HDI 2) , derived from the HDI developed by the United Nations (HDI). Both HDI and HDI 2 are three-dimensional composite indexes with different components: life and healthy life (indicators: life expectancy at birth for HDI and HDI 2); knowledge or education (indicators: expect years of schooling and mean years of schooling for HDI; percentage of the population aged 15 and older and graduated from the school system for HDI 2); suitable standard of living (indicators: gross national income per capita for HDI; median household income per unit of consumption for HDI 2). We used the 2013 HDI 2 data. They were computed at a municipality level from 2011 data provided by the French National Institute of Statistics (INSEE).
At the second level of analysis, we added a number of variables derived from the detailed statistics collected in the census carried out periodically by the INSEE. These variables were selected and grouped into the following categories (Table 3): housing, education level, equipment of territories, household situation, income and poverty, and unemployment.
Quantitative variables were described by median, first and third quartiles (Q1–Q3) because of skewed distributions.
The aim of this study was two-fold. First, the spatial distribution of CA incidence was evaluated. We used a standardized incidence ratio (SIR) to depict the spatial distribution of CA incidence. The SIR was smoothed using the hierarchical Bayesian model with three levels as proposed by Besag et al. . Low-incidence or high-incidence municipality was characterized by a smoothed SIR value, respectively, lower or greater than 1. A smoothed SIR was considered significant if its credible interval did not contain the value 1.
Second, the detection of significant clusters of municipalities with a high or a low incidence of CA was performed. We used spatial scan statistics on the basis of a Poisson model to test for the presence of CA clusters and identify their location without preselection bias. These methods present a very good power and are well known in the field of spatial cluster detection . Moreover, the cluster detection was adjusted through indirect standardization for the age of patient, which is a well-known confounding factor of CA incidence . The calculations were carried out using the SaTScan (Harvard Medical School, Boston, Massachusetts, USA) software and a cluster was defined as significant when associated with a one-sided P-value less than 0.05. All clusters, including primary clusters associated with the ‘most likely cluster’ (MLC), and secondary clusters associated with a high likelihood function, but not overlapping with the MLC, were taken into account. Their significance was evaluated by computing the P-value adjusted for MLC according to Zhang’s method. An RR was associated with each significant cluster. This measure was interpreted as the risk of observing CA in the highlighted cluster compared with the whole area excluding the identified clusters.
Comparison of group of municipalities
All municipalities in the studied area were divided into three groups: municipalities included in clusters with a high CA incidence, those in clusters with a low CA incidence and those not included in a cluster. Because the socioeconomic characteristics in the three groups of municipalities did not follow a normal distribution, a nonparametric test, the Kruskal–Wallis test, was used for comparison. The nonparametric test of Dunn with Bonferroni correction was used for post-hoc comparisons among the three groups.
The statistical comparative analyses were carried out using SAS software, version 9.3, SAS Institute Inc., Cary, NC, USA. All statistical analyses were considered significant at type 1 error 5%.
From August 2013 to August 2015, we studied 3414 CAs: 1032 in the Hauts-de-Seine county, 1439 in the Seine-Saint-Denis county and 943 in the Val-de-Marne county. The population was mainly male (65.4%) and the median age was 70 years. CA occurred mainly at home (70.2%) and 69.3% were witnessed. Bystander BLS was started in 42.0% of patients. The no-flow duration (delay between CA and first cardiopulmonary resuscitation) was 5 min; 79.1% of the patients had a first monitored rhythm asystole, 5.1% had ventricular fibrillation, 6.3% had pulseless electrical activity and 9.5% had return of spontaneous circulation at MMT arrival. The median MMT response time was 20 min. The rates of return of spontaneous circulation, survival to hospital admission and 30-day survival were low (23.4, 21.8 and 5.5%, respectively). Finally, 89.0% of the 30-day survival patients achieved a good neurologic recovery. Other characteristics of the general population are detailed in Table 1.
During the period from August 2013 to August 2015, the mean annual age-standardized incidence rate of CA for our population was 76.5/100 000 inhabitants. The Bayesian hierarchical model showed strong geographical variations in the age-adjusted smoothed SIR of CA in the 123 municipalities of the studied area (Fig. 1). The smoothed SIR of CA ranged from 0.15 to 2.51. Among the 123 smoothed SIRs, 33 presented a significantly increased risk (value > 1) and 37 presented a significantly low risk (value < 1).
The age-adjusted spatial scan statistics provided two types of clusters: clusters with low incidence and clusters with high incidence (Fig. 2). The descriptions of significant cluster characteristics (geographical location, radius, population, observed and expected cases, relative risk, and P-value) are presented in Table 2. The 123 municipalities were grouped into seven significant spatial clusters of CA incidence, including three clusters with low incidence (422 CA, 1 487 975 inhabitants, mean incidence 30/100 000 inhabitants) and four clusters with high incidence (1699 CA, 1 536 363 inhabitants, mean incidence 132/100 000 inhabitants). The relative risk varied from 0.23 to 0.54 for clusters with low incidence and from 1.43 to 2 for clusters with high incidence.
Characteristics of municipalities with a high or a low cardiac arrest incidence
The 123 municipalities of the Parisian ‘small crown’ were distributed as follows: 37 belonged to a cluster with a low CA incidence, 33 belonged to a cluster with a high CA incidence and 53 did not belong to a cluster. A comparison of socioeconomic characteristics among the three groups of municipalities is shown in Table 3. The group of municipalities with a high CA incidence was characterized by a lower HDI 2 than the municipalities with a low CA incidence and municipalities with a normal CA incidence [0.418 (0.349–0.342) vs. 0.630 (0.592–0.727) vs. 0.666 (0.552–0.720); P < 0.001]. This result was confirmed when we studied each HDI 2 component (health, education, income). We found similar results when we examined other socioeconomic variables for all of those describing income and poverty, education, unemployment rate, households with single parents and households where the reference person is a worker or an employee. However, we did not find a significant difference between high-incidence and low-incidence clusters for equipment territories. Finally, for the housing variables, the results were more complex. There was no significant difference for the HLM main residence (rent-controlled housing for individuals with low incomes) variable, whereas we found a significant difference for ‘number of inhabitants per main residence’.
The analysis of the smoothed standardized incidence reports of cases of CA in the Parisian ‘small crown’ area over a 2-year period shows the existence of heterogeneity in terms of CA incidence. We notably observe that out of the 123 municipalities studied in total over three counties, the SIRs are greater than 1.5 for around 20 municipalities and less than 0.5 for nearly 25 municipalities. The use of the scan statistics enabled us to identify seven significant clusters, of which four showed over incidence and three showed underincidence. These results confirm those of existing studies that show spatial heterogeneity in the incidence of CA [5911–131920]. These studies highlighted the existence of high-risk areas. Nevertheless, these studies used different methodological approaches.
Our study results are based on the use of a method combining the use of Bayesian analysis and spatial scans , whereas the studies mentioned above only used Bayesian methods or scan statistics. The advantage of using these two methods together is that the spatial heterogeneity (Bayesian methods) can be studied and then significant areas of over incidence or underincidence (scan statistics) can be highlighted. These areas can then be characterized by socioeconomic variables. In the international literature, we have only been able to identify a single study specifically addressing the incidence of CA using a combination of Bayesian methods and scan statistics . In this study, Sasson et al.  analysed 1632 CAs that occurred in Denver between April 2004 and April 2009. To the best of our knowledge, our study offers the most important contribution to this topic over a 2-year period on a population representing more than 4.5 million inhabitants. Moreover, our study is the first one carried out in France on the subject of CA incidence. We can also mention a combined study  (using both Bayesian analysis and spatial scans), but the aim of this study was to look for spatial variability in bystander BLS levels. This study, focusing on the cities of Houston and Austin, highlighted the income and ethnic origin criteria to explain the geographic variability in the implementation of BLS.
The literature characterizing areas of over incidence have generally shown strong links between incidence and ethnicity, educational attainment and income. Some studies  have shown a correlation between income level, ethnicity, level of education and CA incidence. Semple et al.  highlighted the primary role played by socioeconomic criteria, similar to our study. Other studies  have focused on the impact of ethnicity. Our study can neither confirm nor deny these results because statistics about ethnicity are forbidden in France. However, it may appear achievable and relevant in the USA or Asia. Ong and colleagues  found a strong relationship between high education level and underincidence. However, Ong and colleagues did not find a significant relationship between income level and incidence. These results can be explained by the high proportion of pensioners in certain territories and the characteristics of Singaporean pension plans. These results differ from those found in our study, where low education and low incomes were associated with CA over incidence.
Ong and colleagues  were also interested in the composition of the family structure and household situation. Thus, they highlighted the strong link between over incidence and isolation (defined as a ‘no family nucleus’ household). This result was also found in our study, where we have also shown a strong link between the absence of household stability and CA over incidence, and especially for households with a single parent. However, we know that in industrialized countries, there is a relationship between the absence of a family nucleus and socioeconomic precariousness.
The complex link between socioeconomic variables argues in favour of the use of indices of social disadvantage or deprivation. Ong et al.  proposed the use (but did not actually use) of the Carstairs Index , which includes variables such as unemployment, the lack of a personal vehicle, overcrowded housing and low social class. To the best of our knowledge, only one British study  used an index of social disadvantage, the Townsend Index , which includes variables such as unemployment, car ownership and an indicator of housing overcrowding. The main result of this study showed a significant link between the Townsend Index and the incidence of CA in Nottinghamshire. In our study, we used the HDI 2 . This index is adapted from an index developed by the United Nations within the framework of the United Nations Development Program . This index that we used is composed of three variables: life expectancy at birth, percentage of the population older than 15 years old graduating from the school system and the median income per household. Although these indices of social disfavour or deprivation differ in their content, our study confirms the trends reported in the study of Soo et al. . In fact, there was a strong link between low HDI 2 and CA over incidence.
In our study, we used additional variables to more finely characterize the notion of income. We introduced the share of social benefits in global income and the proportion of individuals living below the poverty threshold. Our results show that the statistically significant relationship between these parameters and over incidence indicates a strong link between impoverishment and CA over incidence. The use of median income alone was insufficient for analysis. The composition of this income is at least as important as the income itself. In addition, we examined other criteria describing the level of poverty. We found a statistically significant link between the level of unemployment or socio-occupational category (workers) and CA incidence.
Ong et al.  raised the issue of the possible link between the economic activity of the territory and the CA incidence by evoking a possible bias linked to the existence of a concentration of firms in certain territories. In our study, we analysed the relationship between incidence and economic development (number of businesses created). Our results showed that there was no statistically significant relationship between these variables and incidence, which led us to speculate whether the economic activity of the territory always benefits the inhabitants of this same territory.
The first limitation of this study was the use of Bayesian methods, which have been the subject of detailed comments in the literature . Scan statistics, which are less used, are powerful tools. Scan statistics are an ideal complement to Bayesian methods as they enable the characterization of territories of over incidence and underincidence. In our study, we aggregated the individual variables that characterize CA and the use of population variables that characterize the socioeconomic status of territories. The accuracy of studies involving the processing of personal data and individual-level measures should also be questioned, but this approach refers to ethical, legal and technical issues. Personal data processing, especially in health issues, often requires special authorization. Other than completely revising the framework of the registries, or that of the censuses, we do not yet have a simple solution to match CA and SES individual data, except to include a dedicated and authorized longitudinal cohort. Finally, for ethical reasons, the collection and use of ethnic statistics is prohibited by law in France. Therefore, we could not compare our results with the results of several studies in the literature.
Our study shows that there is spatial heterogeneity in the incidence of CA. This heterogeneity makes it possible to identify socioeconomically contrasting territories with respect to the incidence of CA. This observation can be an effective lever for action that is both more relevant and more adapted to the realities of the field in the implementation of public health policies in the context of CA. Public health policies can be differentiated by adapting the level of prevention actions or the organization of care to the characteristics of the territories identified. This will provide the keys for targeting tangible objectives in public health actions territorialization.
The RéAC registry was supported by the French Society of Emergency Medicine (SFMU), a patient foundation – Fédération Française de Cardiologie, the Mutuelle Générale de l’Education Nationale (MGEN), the University of Lille and the Institute of Health Engineering of Lille.
Conflicts of interest
There are no conflicts of interest.
1. Benjamin EJ, Blaha MJ, Chiuve SE, Cushman M, Das SR, Deo R, et al. Heart Disease and Stroke Statistics-2017 update: a report from the American Heart Association. Circulation 2017; 135:e146–e603.
2. Stecker EC, Reinier K, Marijon E, Narayanan K, Teodorescu C, Uy-Evanado A, et al. Public health burden of sudden cardiac death in the United States. Circ Arrhythm Electrophysiol 2014; 7:212–217.
3. Hubert H, Tazarourte K, Wiel E, Zitouni D, Vilhelm C, Escutnaire J, et al. Rationale, methodology, implementation, and first results of the French out-of-hospital cardiac arrest
registry. Prehosp Emerg Care 2014; 18:511–519.
4. Lin BC, Chen CW, Chen CC, Kuo CL, Fan IC, Ho CK, et al. Spatial decision on allocating automated external defibrillators (AED) in communities by multi-criterion two-step floating catchment area (MC2SFCA). Int J Health Geogr 2016; 15:17.
5. Ong ME, Wah W, Hsu LY, Ng YY, Leong BS, Goh ES, et al. Geographic factors are associated with increased risk for out-of hospital cardiac arrests and provision of bystander cardio-pulmonary resuscitation in Singapore. Resuscitation 2014; 85:1153–1160.
6. Root ED, Gonzales L, Persse DE, Hinchey PR, McNally B, Sasson C. A tale of two cities: the role of neighborhood socioeconomic status in spatial clustering of bystander CPR in Austin and Houston. Resuscitation 2013; 84:752–759.
7. Dahan B, Jabre P, Karam N, Misslin R, Tafflet M, Bougouin W, et al. Impact of neighbourhood socio-economic status on bystander cardiopulmonary resuscitation in Paris. Resuscitation 2017; 110:107–113.
8. Chen CC, Chen CW, Ho CK, Liu IC, Lin BC, Chan TC. Spatial variation and resuscitation process affecting survival after out-of-hospital
cardiac arrests (OHCA). PLoS One 2015; 10:e0144882.
9. Nichol G, Thomas E, Callaway CW, Hedges J, Powell JL, Aufderheide TP, et al. Regional variation in out-of-hospital cardiac arrest incidence
and outcome. JAMA 2008; 300:1423–1431.
10. Wells DM, White LL, Fahrenbruch CE, Rea TD. Socioeconomic status and survival from ventricular fibrillation out-of-hospital cardiac arrest
. Ann Epidemiol 2016; 26:418–423.e1.
11. Nassel AF, Root ED, Haukoos JS, McVaney K, Colwell C, Robinson J, et al. Multiple cluster analysis for the identification of high-risk census tracts for out-of-hospital cardiac arrest
(OHCA) in Denver, Colorado. Resuscitation 2014; 85:1667–1673.
12. Ong ME, Earnest A, Shahidah N, Ng WM, Foo C, Nott DJ. Spatial variation and geographic-demographic determinants of out-of-hospital
cardiac arrests in the city-state of Singapore. Ann Emerg Med 2011; 58:343–351.
13. Sasson C, Cudnik MT, Nassel A, Semple H, Magid DJ, Sayre M, et al. Identifying high-risk geographic areas for cardiac arrest
using three methods for cluster analysis. Acad Emerg Med 2012; 19:139–146.
14. Vaillancourt C, Lui A, de Maio VJ, Wells GA, Stiell IG. Socioeconomic status influences bystander CPR and survival rates for out-of-hospital cardiac arrest
victims. Resuscitation 2008; 79:417–423.
15. Genin M, Duhamel A, Preda C, Fumery M, Savoye G, Peyrin-Biroulet L, et al. Space-time clusters of Crohn’s disease in northern France
. J Public Health (Bangkok) 2013; 21:497–504.
17. Jahan S, Jespersen E, Mukherjee S, Kovacevic M, Bonini A, Calderon C, et al. Human development report 2015: work for human development. 2015.New York, NY: UNDP.
18. Besag J, York J, Mollie A. Bayesian image-restoration, with 2 applications in spatial statistics. Ann Inst Stat Math 1991; 43:1–20.
19. Semple HM, Cudnik MT, Sayre M, Keseg D, Warden CR, Sasson C, et al. Identification of high-risk communities for unattended out-of-hospital
cardiac arrests using GIS. J Community Health 2013; 38:277–284.
20. Soo L, Huff N, Gray D, Hampton JR. Geographical distribution of cardiac arrest
in Nottinghamshire. Resuscitation 2001; 48:137–147.
21. Carstairs V. Small area analysis and health service research. Community Med 1981; 3:131–139.
22. Townsend P. Deprivation. J Soc Policy 1987; 16:125–146.