The regional variation of a disease may have important implications for its diagnosis and care. For example, sudden cardiac arrest incidence varies two-fold across the US, potentially reflecting differences in population characteristics and public systems of emergency care . Stroke death is highest in the Southeast US, raising questions regarding the stroke susceptibility, health behaviors and systems of medical care [9–14]. Only limited data describe the regional distribution of sepsis [3, 5].
We observed a more than two-fold variation in the infection-attributed mortality in the US. The 11 states with the highest sepsis mortality comprised a contiguous cluster in the Southeastern and Mid-Atlantic US. Assuming the equivalence of sepsis and infection-attributed deaths, the increased sepsis mortality in this cluster (80.1 vs. 61.9 deaths per 100,000 in other regions) translates to over 8,500 excess adult sepsis deaths each year in the US.
The reasons for these observations remain unclear but may involve variations in the patients, environment or patterns of care. For example, the treatment of sepsis is often complex, involving the administration of intravenous fluids, antibiotics and vasopressors [4, 18]. Regional sepsis mortality variations could reflect differences in the execution of sepsis treatment protocols. Regional differences in medical comorbidities, health behaviors, diet, socioeconomic status, genetics or environmental exposures may potentially alter the risk of sepsis [11, 12]. Obesity is highest in the Southeastern US, and sepsis severity is higher in obese individuals, suggesting a potential contributory role [19–24]. Answers to these and other key questions could innovate sepsis treatment and prevention strategies, potentially reducing sepsis death and healthcare expenditures.
Our results provide interesting initial perspectives. For example, the regional distribution of infection deaths remained stable across age and sex strata, suggesting that age and sex are not contributors to regional variation. In contrast, the geographic distribution differed between African Americans and Whites, suggesting that racial differences may partially explain sepsis mortality variations. Prior studies of the sepsis epidemiology have used primarily hospital discharge data with inadequate scope or clinical detail to answer these questions [3, 5]. Appropriate answers would require study with a national population-based cohort encompassing knowledge of subjects' baseline characteristics and identification of subsequent sepsis events. Our observations also highlight that population-based studies limited to smaller regions may not result in nationally generalizable inferences.
The most unexpected observation was the similarity between the observed sepsis death cluster and the US "Stroke Belt." While defined in different ways, the Stroke Belt generally refers to a region of increased stroke mortality encompassing Mississippi, Alabama, Georgia, Tennessee, Kentucky, North Carolina and South Carolina [9–13]. Within the Belt a "Stroke Buckle" encompassing the North Carolina, South Carolina and Georgia costal regions contains the highest death rates. First identified in the 1930s, the pattern of excess deaths persists today despite secular trends in overall and race-stratified stroke mortality[9, 10, 13]. The Stroke Belt has spawned key hypotheses regarding the pathophysiology of and risk factors for cerebrovascular disease, including medical comorbidities, lifestyle, diet, socioeconomic status, genetics, differing responses to medications and environmental exposures [11, 12]. The overlap between the Stroke Belt and our observed sepsis cluster could point to unidentified similarities in the pathophysiology, patient characteristics or medical care of these conditions.
There are key differences between this study and prior sepsis epidemiology descriptions. Our estimate of 65.5 infection deaths per 100,000 contrasts with Martin, et al.'s estimate of 43.9 per 100,000. However, Martin, et al. used sampled data from National Hospital Discharge Survey and limited cases to those with ICD-9 sepsis diagnosis codes, potentially missing infection-related deaths not coded as sepsis . Angus, et al.'s study of combined statewide hospital discharge data (Florida, Maryland Massachusetts, New Jersey, New York, Virginia, and Washington) estimated a higher mortality (approximately 85.8 per 100,000), but their broad use of discharge diagnoses may have misattributed selected deaths to infection .
Melamed and Sorvillo examined secular trends in sepsis mortality using CMF multiple cause of death data set, classifying sepsis deaths as instances where any of the four causes of death included ICD-10 septicemia . They did not include other infection groups. Our approach differs in the use of a single underlying cause of death complemented by a broader sepsis definition. While our estimate of 65.5 sepsis deaths per 100,000 is higher than Melamed's estimate of 50.5 per 100,000, we included only individuals ≥ 15 years old. When we repeated our analysis using Melamed's approach but limited to individuals age ≥ 15 years, we observed a sepsis mortality of 62.6 per 100,000 as well as the same regional sepsis "belt." This observation supports the robustness of our approach.
Limitations of this analysis include the use of public death records. Listed causes of death are subject to classification or misattribution bias, which could affect our results [25–30]. We could not use conventional definitions of sepsis. We could not ascertain if secondary infections played prominent roles in the death of individual cases. While we used the CDC's existing cause of death categories, this taxonomy may have missed selected infections such as peritonitis, pyothorax, abscesses or unspecified infections. For analytic purposes we combined all infections together, but select patients may have responded differently to individual infections. It is unclear how these biases may have altered our observations. We did not include deaths of individuals <15 years or with unknown age. We note that there were only 56 deaths with unknown age.
We did not formally validate the accuracy of death records for identifying sepsis; this is the objective of a separate effort using adjudicated death records. However, when we repeated the analysis using Melamed's strategy with the CMF multiple cause of death data, we observed similar results, suggesting robustness of our approach. As discussed previously, we did not evaluate regional variations in sepsis hospitalizations because of the lack of appropriate data sets.
While our observed cluster appears to exclude South Carolina and Kentucky (two prominent representatives of the Stroke Belt) these states fell on the upper quartile cutoff (74.4 per 100,000) and could be included with this group. While stroke and infection may conceivably coexist in a patient, we used the CMF underlying cause of death data, precluding the possibility of confounding as the cause of stroke/sepsis geographic overlap.
We selected states as the unit of analysis in order to provide clearer national perspectives for this initial effort. Additional insights may have resulted from smaller geographic units (counties, census tracts). Also, geographic boundaries may not align with state boundaries. For example, heightened sepsis mortality in the Appalachian Mountains would have affected sepsis mortality estimates in many high risk states.
We included only deaths for individuals age ≥ 15 years in this study. We would expect different mortality patterns for children since the sepsis epidemiology differs in this age group. Due to their relatively sparse numbers, we did not separately examine sepsis patterns among Asians/Pacific Islanders and American Indians. Our study describes the regional distribution of those dying from sepsis but does not characterize survivors. We did not have sociodemographic or hospitalization information on each patient. While we did not formally evaluate longitudinal trends, we found similar regional patterns for each year of 1999-2005.