Association between sepsis incidence and regional socioeconomic deprivation and health care capacity in Germany – an ecological study

We performed a retrospective ecological study based on the nationwide Diagnosis-related Groups (DRG) statistics of 2016. The DRG statistics is the largest all-payer inpatient database in Germany. Data collection is mandated by the Hospital Reimbursement Act §21 for all acute-care hospitals in Germany except for prison hospitals and psychiatric facilities. Each hospitalization is listed with primary and secondary International Classification of Diseases, Tenth Revision, German Modification (ICD-10-GM) codes, procedural codes, discharge disposition, patient demographics, and hospital length of stay. Furthermore, we used regional indicators provided in the INKAR (indicators and maps on spatial and urban development) database of the German Federal Institute for Research on Building, Urban Affairs and Spatial Development. Indicators were extracted on a district level and merged with the DRG statistics using the official municipality key of the patients’ residence included in both databases.

Among all hospitalizations in 2016 in Germany, we identified hospitalizations with explicitly coded sepsis using the ICD-10-GM codes R65.1 (severe sepsis) and R57.2 (septic shock). In 2016, these codes were defining sepsis according to the sepsis-1 definition as sepsis with organ dysfunction (severe sepsis) [18]. Additionally, we used an alternative implicit approach, which is known as Angus implementation [19]. It identifies sepsis by the combination of ICD-10-GM codes for infection and organ failures (supplementary file 1) and thereby captures cases in which sepsis was not explicitly coded at hospital discharge. This approach is considered less prone to external coding incentives such as reimbursement for higher patient complexity. We characterized sepsis patients by comorbidity as defined by the Charlson Comorbidty Index [20], surgical treatment (any procedural code from chapter 5 = surgical procedures), intensive care treatment (procedural codes 8–980, 8-98d, 8-98f for intensive care complex treatment), hospital length of stay, hospital death and discharge to hospice (discharge disposition code = 7 or 11 in the DRG statistics, respectively).

Germany has 16 federal states and 401 districts, which form its administrative units. Sepsis incidence was calculated according to the patients’ place of residence on federal state and district level, which are coded by the official municipality key in the DRG statistics.

We chose three commonly used district level socioeconomic deprivation indicators [21‐25] reflecting average occupation, income and education of the population for the analysis of contextual socioeconomic effects: The unemployment rate 2016 (proportion of unemployed among working age residents in %), the net household income 2016 (average household income in EUR per inhabitant), and the rate of school leavers without certificate 2016 (percent of students leaving school without having passed the lowest qualification certificate, the „Hauptschulabschluss “after 9 or 10 years of education, out of all students leaving school). Furthermore, we selected three indicators of inpatient and outpatient health care capacity and the density of medical services: hospital beds per 1000 inhabitants in 2016, general practitioner (GP) per 100,000 inhabitants in 2016, and straight line distance to the nearest pharmacy per inhabitant in meters as surrogate for the geographical proximity of medical services in 2017 (no 2016 data available). The definition of the airline distance to the nearest pharmacy is explained in the supplementary file 1.

Age has proved an important risk factor for sepsis [26]. To assess the relationship between age and sepsis incidence, we used the mean age of the districts’ population in 2016 provided in the INKAR database as predictor for the crude sepsis incidence.

Regional differences in the crude and age-standardized incidence of sepsis between German districts were tested in a first step using a χ²-Test proposed by Snijders & Bosker [27]. The intraclass correlation (ICC) was computed as an effect size measure of the between-district heterogeneity. Negative binomial (NB) regression models were used in a second step to explain between-district variance in the incidence proportions by indicators of socioeconomic deprivation and health care capacity. The NB regression model was preferred over the Poisson regression model due to significant overdispersion in our data (supplementary file 2 Tables 1 and 2), which was tested using likelihood ratio tests. The NB models were fitted to (a) the number of sepsis hospitalizations per district, and (b) the age-standardized expected numbers of sepsis hospitalizations per district calculated by the direct method with Germany’s overall population in 2016 as reference. The number of inhabitants per district was taken into account as an offset variable in the model to account for differences in numbers of inhabitants between districts. First, each district level predictor was analyzed individually in a simple regression model predicting crude and age-standardized sepsis incidences. Some of the predictor variables are substantially correlated (Table 1). In order to estimate the unique contribution of the socioeconomic and medical service indicators, we used three multiple NB regressions for each outcome in a second step: (a) the multiple NB regression with the three indicators of socioeconomic deprivation; (b) the multiple NB regression with the three indicators of health care capacity; and (c) the multiple NB regression with all predictor variables. The last model served for statistical testing of the uniquely explained variance by socioeconomic deprivation and health care capacity indicators using the likelihood ratio test. Nagelkerke’s Pseudo-R² was used as an effect size measure of the relationship between predictors and the incidence rates. The number of hospitals beds was unavailable for two German districts (Soemmerda and Fuerth), which were excluded from regression models that include this predictor variable.

To illustrate the strength of the relationship between predictors and outcomes in common metrics, the parameters of the NB regressions were used to estimate the expected percentage change (EPC) as well as the expected change (EC) in the number of sepsis hospitalizations associated with an increase in the respective predictor variable (see supplementary file 3). In case of multiple NB regressions, EPC and EC are estimated under statistical control of the other predictors in the model and therefore can be interpreted as adjusted EPC and EC.

We report point estimates and interval estimates with 95% coverage in addition to two-sided p-values. The significance level was α = 0.05. All statistical analyses and computations were conducted via remote data processing using R [28], including the R functions ‘nagelkerke’ from the r package ‘rcompanion’ [29] and ‘glm.nb’ from the r package MASS [30]. Maps were created using the ‘spplot’ function from the ‘sp’ package [31, 32].

In the simple regression analyses, we found that sepsis incidence was significantly associated with the mean population age at the district level (Table 3). The expected change was EC = 12.6 [95% CI, 9.4, 15.8] (p < .001). Hence, two randomly selected districts that differ by 1 year in the mean age have an expected difference of 12.6 sepsis hospitalizations per 100,000 population. Furthermore, all socioeconomic indicators and the distance to the nearest pharmacy were statistically significant predictors of sepsis incidence in the simple regression analyses, whereas we observed no evidence for such an association to the number of hospital beds and the number of GPs. The mean age was found to be the strongest single predictor for the number of sepsis hospitalizations in terms of Pseudo-R² (explicit: Pseudo-R² = 0.140, implicit: Pseudo-R² = 0.242).

Adjusting for all other predictors in the multiple regression analyses, effects of mean household income and rate of school leavers without certificate on sepsis incidence were attenuated, but still significant. Unemployment rate was not significantly associated with the sepsis incidence given the other two predictors. Positive associations were also found between age-adjusted sepsis incidence and the rate of school leavers without certificate, although they were weaker compared to the associations with unadjusted sepsis incidence (Table 3). The adjusted EC in the age-standardized sepsis incidence for the rate of school leavers without certificate was EC = 3.3 [95% CI, 0.1, 6.5] (p = .041). Hence, two randomly selected districts with equal age distributions that differ by 1% in the proportion of school leavers, but no differences in the other predictors in the model have an expected difference of 3.3 sepsis hospitalizations per 100,000 population. The adjusted EC of the crude sepsis incidence: EC = 5.9 [95% CI, 2.4, 9.4] (p < .001).

Among indicators of medical infrastructure, an increase of the mean distance to the nearest pharmacy by 1000 m was found to be associated with an expected increase in sepsis incidence by 21.6 [95% CI, 10.1, 33.0] sepsis hospitalizations per 100,000 inhabitants, multiple regression analyses (p < 0.001). A statistically significant, but weaker association was found between mean pharmacy distance and age-standardized sepsis incidence in the multiple regression analyses (Table 3, EC =  11.1 [95% CI, 1.0, 21.2], p = .026). All other indicators of medical capacity were not significantly associated with crude- or age-standardized sepsis incidence.

Socioeconomic indicators and health capacity indicators explained 6 and 10%, respectively, of the variance in sepsis incidence between districts in the multiple regressions including each set of indicators separately. The Pseudo-R² of the full multiple NB regression with both sets of indicator variables (i.e., the indicators of socioeconomic deprivation and medical capacity, Table 3) was 0.078. Dropping one set of indicators resulted in significantly lower proportions of explained regional variation (omitting socioeconomic deprivation indicators: Δ Pseudo-R² = 0.033, p = 0.004; omitting medical capacity indicators: Δ Pseudo-R² = 0.024, p = 0.022). Hence, both sets of variables seem to address unique proportions of the regional variation of the sepsis incidence (supplementary file 2 Table 5). The full model with age-adjusted sepsis incidence as outcome failed statistical significance (χ²(6) = 12.27, p = 0.056). Accordingly, a unique proportion of regional variance was neither explained by socioeconomic deprivation nor by health care capacity.

In comparison to hospitalizations with explicitly coded sepsis, the incidence of hospital-treated sepsis identified by implicit coding was higher (overall incidence = 1498, median = 1494 (IQR = 1306 to 1737) per 100,000 population, 6.5% of hospitalizations, supplementary file 2 Figure 8). Implicit and explicit sepsis incidence were positively correlated (r = 0.603). We found a similar result pattern for socioeconomic indicators as predictors for implicitly defined sepsis in the simple and multiple regression analyses, but higher Pseudo-R² values indicate a stronger stochastic relationship than for the explicitly defined sepsis (Table 4). Contrary to our findings on explicit sepsis, mean household income and hospital bed capacity was positively associated with crude and age-standardized sepsis incidence in the simple and multiple regression analyses. The Pseudo-R² of the full multiple NB regression including all indicators of socioeconomic deprivation and medical capacity was R² = 0.209 for the crude sepsis incidence and R² = 0.030 for the age-standardized sepsis incidence.