Temporal Trends in Sepsis Incidence and Mortality in Patients With Cancer in the US Population

We obtained data for this analysis from the Healthcare Cost and Utilization Project National (Nationwide) Inpatient Sample (NIS), the largest publicly available all-payer inpatient health care database in the United States. The NIS is managed by the Agency for Healthcare Research and Quality and includes a 20% sample of patient records from 4378 hospitals.¹⁶  The database provides information on patient demographics, diagnoses, and outcomes for each hospitalization. It includes data from more than 7 million hospital stays, weighted according to the survey’s sampling design to estimate more than 35 million hospitalizations nationally. Weights are calculated by dividing the number of universe discharges (estimated from the American Hospital Association annual survey) by sampled discharges within each NIS stratum. Strata are defined by hospital characteristics including census division, location (rural or urban), bed size, teaching status, and ownership.

The board of the National Taiwan University Hospital approved this study. Because the study used a deidentified database for research and involved no patient interventions, registration was not required and institutional review board approval was waived.

We identified hospitalizations with a validated approach using the International Classification of Diseases, Ninth Revision (ICD-9) diagnostic codes defined by Martin et al.¹⁷  We identified hospitalizations of adult patients (≥20 years of age) with sepsis from 2006 through 2014 by selecting all patients with an explicit ICD-9, Clinical Modification (ICD-9-CM) code for sepsis or systemic fungal infection (038, septicemia; 020.0, septicemic; 790.7, bacteremia; 117.9, disseminated fungal infection; 112.5, disseminated Candida infection; or 112.81, disseminated fungal endocarditis) and a diagnosis of acute organ dysfunction. Acute organ or system dysfunctions identified for this study were cardiovascular, respiratory, central nervous, hematologic, hepatic, and renal system dysfunctions. Shock was included as a form of cardiovascular dysfunction. Sources of infection were categorized as lower respiratory tract, genitourinary tract, skin and skin structure, catheter-related bloodstream, intra-abdominal, biliary tract, or musculoskeletal infection. We excluded hospitalizations with missing values, transfers to short-term hospitals, and patients discharged against medical advice. We recorded sources of infection and microbiological profiles by using ICD-9-CM pathogen codes (Supplemental Table 1 and Supplemental Table 2). Eligible patients were classified as having or not having cancer on the basis of the presence or absence of ICD-9-CM cancer codes (Supplemental Table 3).

We used diagnostic codes to identify admitted patients who had a diagnosis of lung, breast, gastrointestinal tract, hematologic, gynecologic, or unspecified site cancer (Supplemental Table 3). According to the Uniform Hospital Discharge Data Set guidelines, we used diagnostic codes to identify conditions that coexisted at the time of admission, that developed subsequently, or that affected the treatment received and/or the length of stay.

Annual incidence of sepsis events was presented as events per 100 000 hospitalizations, stratified by time periods (2006-2008, 2009-2011, and 2012-2014). For 13 specific pathogens, we assessed 2 longitudinal outcomes: mean annual change in incidence and change in hospital 30-day mortality. We stratified these outcomes according to the presence or absence of cancer. General mortality data for broad subgroups of pathogens (gram-positive bacteria, gram-negative bacteria, anaerobes, and fungi) included percentage changes in 30-day mortality rates and crude and propensity score–matched hazard ratios (HRs) to calculate overall death rates.

We estimated the frequency of sepsis hospitalizations with infections due to specific pathogens following the recommendations of the Agency for Healthcare Research and Quality. By using survey-specific statements (SURVEYMEANS, SAS Institute Inc), we weighted the patient data using the weights provided in the NIS database. Continuous variables are presented as means and SEs. Categorical variables are reported as numbers and percentages. We calculated the overall and mean annual percentage changes in incidence and 30-day mortality for sepsis and specific sources of infection from 2006 through 2014. To examine the significance of incidence and mortality trends, we performed a generalized linear model analysis. We compared trends of sepsis incidence and mortality between patients with cancer and those without cancer by using the Cochran-Armitage test for trends.^18,19  We calculated adjusted and unadjusted HRs by using Cox regression models to estimate the mean mortality rate. To account for imbalanced covariates between patients with cancer and patients without cancer, we constructed a propensity score model²⁰  on the basis of age, sex, and combined Charlson score.²¹  We then calculated a propensity score–adjusted HR to compare the mortality rates of patients with cancer and those without cancer. We also used Cox regression models to analyze propensity score– adjusted sepsis mortality rates in patients with different types of cancer. Two-sided P values of less than .05 were considered significant for all analyses. We conducted data management and statistical analyses with SAS software (SAS Institute Inc).

We identified 13 996 374 hospitalizations of patients with sepsis from 2006 through 2014 in the NIS database. Of these patients, 1 901 847 (13.6%) had cancer. The Figure shows the patient inclusion criteria, the proportion of cancer patients, and the 30-day mortality in both groups. The demographic characteristics of patients in the study are shown in Table 1. Compared with patients without cancer, patients with cancer were slightly older, were more likely to be female, and had a slightly higher burden of comorbidities. The lower respiratory and genitourinary tracts were the most common sources of infection in both groups. Patients with cancer had a significantly higher incidence of lower respiratory tract infection (35.0% vs 31.6%) and intra-abdominal infection (5.5% vs 4.6%) than did patients without cancer. Gram-positive and gram-negative bacteria were the most common pathogens in both groups. Anaerobes (1.2% vs 0.9%) and fungi (4.8% vs 2.9%) were more common pathogens in patients with cancer than in patients without cancer (Table 1).

The trend analysis revealed a steady increase in sepsis incidence from 2006 through 2014 in patients with cancer and in those without cancer (Tables 2 and 3). Escherichia coli (77.1 cases per 100 000 hospitalizations), Staphylococcus aureus (56.2 cases per 100 000 hospitalizations), and Streptococcus species (53.9 cases per 100 000 hospitalizations) were the 3 most common microorganisms causing sepsis in patients with cancer from 2012 through 2014. Among these 3 organisms, the mean increase in annual incidence was highest for E coli (19.1%), followed by Streptococcus species (11.9%) and S aureus (6.4%). The incidence of anaerobic sepsis increased by 90.8% among patients with cancer during the study period, with an annual increase of 16.5%.

Compared with patients without cancer, patients with cancer had significantly increased rates of sepsis caused by gram-negative, gram-positive, and anaerobic organisms. The temporal incidence trends for the 2 groups are shown in Table 3 and Supplemental Figure 1. Despite the different trends of sepsis incidence in patients with cancer and those without cancer, the trends of sepsis mortality were similar in both groups of patients. The sepsis mortality rate decreased dramatically during the study period, particularly for infections caused by gram-positive bacteria, gram-negative bacteria, and anaerobes (Table 4, Supplemental Figure 2).

Patients with cancer had a higher crude 30-day mortality rate (20.0%) than did patients without cancer (13.9%). To evaluate the independent impact of cancer on sepsis mortality, we created a propensity score for matching. The standardized differences in baseline covariates after propensity score matching were minimal (Supplemental Table 4). In the propensity score–matched cohort, the mortality rate remained higher for patients with cancer than for patients without cancer overall (HR, 1.25; 95% CI, 1.24-1.26). The mortality rate differences in patients with cancer and patients without cancer were the most pronounced for sepsis caused by gram-negative organisms, followed by fungi, anaerobic organisms, and gram-positive organisms (Table 5). We further evaluated the independent impact of different cancer sites on survival of patients with sepsis by using a propensity score–adjusted Cox regression model. Compared with patients with breast cancer, who had the highest survival rate, patients with lung cancer had the lowest survival rate (HR, 1.65; 95% CI, 1.60-1.69), followed by those with gastrointestinal cancer (HR, 1.09; 95% CI, 1.06-1.12) and gynecologic cancer (HR, 1.06; 95% CI, 1.03-1.10). The mortality rate for patients with hematologic cancer was similar to that of patients with breast cancer (HR, 1.03; 95% CI, 1.00-1.06).

Although several previous studies have focused on the epidemiology of sepsis at the population level,^22,23  few have specifically evaluated sepsis in patients with cancer. A prospective analysis of cancer survivors suggested that this population had a 2-fold increase in the risk of community-acquired sepsis, even after adjustments for common confounders.⁶  The authors postulated that physiologic mechanisms to account for this association include cytokine release and the resulting chronic inflammatory state^6,24  and the destruction of healthy cells during cytotoxic treatments, all leading to compromised immunity.^6,24,25 

A small retrospective study of patients with sepsis at a tertiary hospital emergency department suggested that patients with sepsis and cancer were younger, had fewer comorbidities, and had more hemodynamic instability during presentation than patients with sepsis and no cancer.⁷  Our study, which used a larger, nationally representative sample, conversely showed that patients with sepsis and cancer were older, were more likely to be female, and had more comorbidities than patients with sepsis and no cancer.

Our trend analyses showed that the incidence of sepsis increased over time but that the sepsis mortality rate decreased in patients with cancer and in those without cancer. The general increase in sepsis incidence may be attributed partly to the aging population; older patients are at higher risk for infection because of frailty and a higher incidence of comorbidities.²⁶  However, changes in the coding of sepsis over the years may be a confounding factor.²⁷  Much of the improvement in sepsis mortality has been attributed to the Surviving Sepsis Campaign²⁸  and early goal-directed therapy. Early goal-directed therapy emphasizes the importance of early recognition and antibiotic administration to improve sepsis care in the emergency department.^29-32  Although this approach has become more controversial, with recent clinical trials questioning the need for every component of early goal-directed therapy,³³  following sepsis guidelines has been shown to improve mortality rates.³⁴  Many studies have shown improvements in sepsis mortality rates over the years. However, a large single-institution study found that sepsis mortality rates declined from 2003 to 2014 in patients with cancer but not in patients without cancer.⁸  Unlike our study, that study used clinical criteria instead of administrative codes to define sepsis. This difference may be pertinent because the results of one study suggested that clinical data paint a different picture of sepsis trends than do claims-based analyses, showing that the sepsis-related mortality rate was stable from 2009 through 2014.³⁵  Cooper et al⁸  theorized that cancer-specific reductions in sepsis-related mortality indicate that universal improvements in sepsis care may be overstated and that cancer-specific trends, such as the growing availability of targeted, less-toxic cancer treatments, should be examined instead.^8,36 

A systematic review by Montassier et al¹²  revealed that gram-negative bacteria are now the most dominant pathogens in sepsis. Although sepsis caused by gram-positive bacteria was more common in our study, the differences have become smaller over time. This epidemiologic shift may have resulted from the decreased use of quinolone prophylaxis, the nature of chemotherapy (myeloablative vs nonmyeloablative chemotherapy), the decreased use of indwelling catheters, and the emergence of antibiotic-resistant pathogens. In our study, E coli was the most prevalent causative organism, although sepsis caused by anaerobes increased at the fastest rate (16.5% per year). Concern has been growing over the increase in extended-spectrum β-lactamase–producing E coli infections among immunocompromised patients with cancer.³⁷  This increase appears to have a strong link to prior antibiotic use. The use of quinolone prophylaxis in patients with cancer has decreased,³⁸  which may be one explanation for the trend of increased gram-negative over gram-positive bacteremia over the years. Another potential explanation is the emergence of more antibiotic-resistant organisms, for example, the observed increase in bacteremia caused by extended-spectrum β-lactamase–producing E coli and vancomycin-resistant enterococci among cancer patients with neutropenia.^37,39  The use of quinolone prophylaxis has also been linked to the rise in antimicrobial-resistant gram-negative bacteria. In the future, careful antimicrobial stewardship must be tailored both to individual cancer patients and to the general population.

The sepsis mortality rate is higher in patients with underlying cancer than in patients without cancer.^6,34  One study demonstrated that in-hospital mortality caused by sepsis was 29% higher in patients with cancer than in patients without cancer.⁷  The authors postulated that the higher sepsis-related mortality rate in patients with cancer may stem from a large tumor burden, cytokine release, and cytotoxic agents in patients with immunosuppression. In our study, we also found that patients with cancer had a higher mortality risk in both the crude analysis and in the propensity score–matched cohorts. One study of sepsis incidence and mortality in patients with different cancer types showed that the incidence of sepsis was higher in patients with hematologic malignant neoplasms than in patients with solid tumors but that the mortality rate was similar in both groups.⁴⁰  Lung cancer was associated with the highest mortality rate in that study,⁴⁰  but gastrointestinal tumors were associated with the highest sepsis mortality rate in another.⁷  Our findings were fairly consistent with the findings of these 2 studies, showing that lung cancer followed by gastrointestinal cancer had the highest rate of mortality resulting from sepsis.

The advantage of our study is the use of the largest administrative database of hospital admissions in the United States, with a nationally representative weighted sample, to examine incidence and mortality trends across time. These data are inherently subject to many limitations. Despite the robust propensity score matching that we performed, unmeasured confounders remain a potential issue given the observational nature of the study. All available data are from patients’ hospitalizations, and the NIS does not include long-term outpatient data, so our analysis was limited to representing the short-term impact of cancer on event outcomes. Some data may be incomplete, misclassified, or omitted because of the reliance on medical coders. Changes in coding practices and quality over time may also affect the results. The codes from the NIS database used to define sepsis have been internally and externally validated¹⁷  and did not change significantly during the study period. Nevertheless, the use of ICD data rather than clinical data for identifying sepsis is an increasing concern.³⁵ 

The epidemiology of sepsis has undergone substantial changes over the years in patients with cancer and in patients without cancer, so our study has widespread clinical and public health implications. Our study shows that patients with cancer are at high risk of sepsis incidence and sepsis mortality and provides pathogen-specific data relevant to clinicians. Guideline-based therapy may improve sepsis mortality rates. Antibiotic stewardship programs should identify patients with cancer as a high-risk group. Future studies are needed to tailor surveillance and antimicrobial prophylaxis appropriately according to the latest epidemiologic data.

We thank the staff of the Eighth Core Lab, Department of Medical Research, National Taiwan University Hospital, for technical support and Medical Wizdom, LLC (Greater Boston) for technical assistance in statistical analysis.