Peripheral Eosinopenia upon Admission Associates Independently with Mortality and Outcomes with Clostridium difficile Infection

Abstract

IMPORTANCE

The ability to identify high-risk patients with C. difficile infection (CDI) as early as the time of admission could improve outcomes by guiding management decisions.

OBJECTIVE

Recent evidence from an animal model suggests that peripheral loss of eosinophils in CDI is associated with severe disease. The primary aim of this study was to construct a model using clinical indices readily available at the time of hospital admission, including peripheral eosinophil counts, to predict inpatient mortality in CDI.

DESIGN

We evaluated 2,065 human subjects admitted for CDI from 2005 to 2015 at two academic institutions, forming a training and a validation cohort. The sample was stratified by admission eosinophil level (0 k/μL or >0 k/μL) and multivariable logistic regression was used to construct a predictive model for inpatient mortality as well as for other disease related outcomes.

SETTING

Two tertiary referral centers.

PARTICIPANTS

Patients admitted through the emergency department due to C. difficile infection.

MAIN OUTCOMES AND MEASURES

Inpatient mortality, with secondary outcomes including the need for a monitored care setting, need for vasopressors, and inpatient colectomy rates.

RESULTS

Patients with undetectable eosinophil levels at admission had increased in-hospital mortality in both the training (odds ratio [OR] 2.01, P=0.027) and validation (OR 2.26; P=0.002) cohorts, in both univariable and multivariable analysis. Undetectable eosinophil counts were also associated with indicators of severe sepsis such as admission to monitored care settings, the need for vasopressors and emergent total colectomy. Other significant predictors of mortality at admission included advanced age, increasing comorbidity burden and lower systolic blood pressures. In a subgroup analysis of patients presenting without initial tachycardia or hypotension, only patients with undetectable admission eosinophil counts, but not those with an elevated WBC count, had significantly increased odds of inpatient mortality.

CONCLUSIONS

This study describes a simple, widely available, inexpensive model predicting C. difficile severity and mortality to identify at-risk patients at the time of admission.

Introduction

Clostridium difficile infection (CDI) is the most common nosocomial infection in the United States,¹ being associated with mortality rates as high as 22% at 60 days after initial infection and 36% at 6 months among institutionalized patients 65 years of age or older.² The incidence of severe and even life-threatening forms of CDI are common³ among strains of this bacteria known to have virulent potential. These virulent strains are not only capable of producing the large Clostridial toxins, toxin A (tcdA) and toxin B (tcdB), which serve as virulence factors,⁴ but many will produce a third, antigenically distinct toxin known as binary toxin (CDT). Animal and human studies^5,6,7 suggest that CDT-producing strains are more frequently associated with severe colitis and higher disease-related mortality rates. Recently, two studies^7,8 using a mouse model of CDI have further described the relationship between eosinophils and CDI. These preclinical investigations suggest that eosinophilia protects against CDI mortality, and that CDT induces a peripheral eosinopenia associated with higher CDI mortality.

A current knowledge gap in CDI relates to a lack of validated prognostic indicators for disease course that are clinically available as well as reliable enough to guide care decisions. Despite evidence based consensus guidelines regarding the management of CDI,⁹ these guidelines focus on timely management once patients develop hemodynamic changes and/or significant laboratory abnormalities. The ability to identify high-risk CDI patients prior to these clinical changes would potentially improve patient outcomes by allowing for such interventions as admission to a monitored care setting and earlier surgical consultation. Our hypothesis was that peripheral eosinopenia at hospital admission for CDI would be associated with higher odds of mortality and other adverse events. We developed a clinical model to predict inpatient mortality in a large cohort of patients from an academic institution, with this model subsequently validated using a separate large cohort of patients from a second academic institution. The goal was to utilize well-established, inexpensive, easily obtainable clinical and laboratory indices, including peripheral eosinophil counts, to identify CDI patients at higher risk for mortality and other adverse events as early as the time of admission.

Materials and Methods

Patient Selection

This study was approved by the Institutional Review Boards (IRB) of both the Pennsylvania State University Milton S. Hershey Medical Center (PSU) and the University of Virginia Health System (UVA). Due to its retrospective nature, the requirement for informed consent was waived by IRB policy. The datasets analyzed during the current study, as well as the computer code utilized during the analysis, are available from the corresponding author upon reasonable request. Using each institution’s cost accounting database over an eleven-year period (2005 to 2015), patients with a positive C. difficile nucleic acid amplification test or ELISA associated with a hospital admission were identified. Only patients with C. difficile testing performed within 24 hours of hospital admission were included, in order to exclude patients who were involved in prolonged hospitalizations for alternate primary diagnoses. Patients were excluded if they were under 18 years of age, or if they did not have total white blood cell counts, eosinophil counts, or serum creatinine levels obtained within 24 hours of hospital admission. If multiple admissions for CDI were identified for a patient, then only the first admission associated with CDI was included in this study, in order to eliminate data pertaining to episodes of recurrent disease.

Covariates and Outcomes

Demographic data collected on admitted patients included age, sex and race. Comorbidity burden was characterized using the Charlson Comorbidity Index, using International Classification of Diseases, Ninth Revision (ICD-9) diagnosis codes.¹⁰ At PSU, use of mechanical ventilation was identified with Charge Description Master (CDM) code 511861. At UVA, mechanical ventilation was identified through either ICD-9 codes (96.7X), ICD-10 (5A1935Z) or Current Procedural Terminology codes (94002, 94003, 94656, 94657). Vital signs at the time of presentation to the hospital were available for the majority of patients (n = 1,675), allowing identification of initial hypotension (systolic blood pressure <90 mmHg) and tachycardia (heart rate >100 beats per minute). Laboratory parameters collected at the time of admission included white blood cell (WBC) count, eosinophil count and serum creatinine level.

The primary study outcome was inpatient mortality. Secondary outcomes included need for admission to a monitored care setting (intermediate or intensive care), need for vasopressors, and need for total abdominal colectomy during the index hospitalization.

Statistical Analysis

At both institutions, a range of eosinophil counts between and inclusive of 0.0 k/μL and 6.0 k/μL were reported, with the smallest non-zero values being 0.01 k/μL at both institutions. An eosinophil count reported as 0.0 k/μL, was the cutoff used to stratify the cohort. Binary and categorical variables were compared between the two groups using Chi-squared tests, and continuous data were analyzed using two-tailed Student’s t-tests. A logistic regression model was then developed using patients from PSU as the training cohort, modeling inpatient mortality as a function of the available patient covariates and admission laboratory parameters, including the markers of severity most commonly reported in the published literature,^11,12 which also reflect clinical practice. Additionally, an interaction term between low eosinophil count and high WBC count was created and tested within the initial model, but was not found to be significant, and was therefore excluded from the final model. We used multiple imputation to account for missing values for systolic blood pressure (SBP) and heart rate. We created 20 multiply imputed data sets and performed logistic regression on each of the data sets (described in Appendix 1). Coefficients were pooled using the methods suggested by Rubin.¹³ These same covariates were then validated in a separate dataset of patients from UVA, keeping patients at each institution separate for the construction of training and validation cohorts to avoid confounding influences from unaccounted institutional variables or differences in C. difficile strains between these institutions at different time points. Secondary outcomes were compared between cohorts using a similar methodology, using initial univariate comparisons and subsequent multivariable logistic regression analysis, utilizing the same covariates from the mortality model.

Several sensitivity and subgroup analyses were also performed in order to test the robustness of the model. To further characterize the relationship between eosinophil counts and mortality, eosinophil count was also examined as (1) a continuous variable, and as (2) a categorical variable by stratifying non-zero eosinophil values into quartiles and comparing them to a reference group of higher non-zero eosinophil quartile (counts >0.22 k/μL). Lastly, a subgroup analysis was performed to evaluate the findings of the model among patients who presented without hypotension or tachycardia (n = 551). The area under the receiver operating characteristic (ROC) curve was used to evaluate the predictive ability of the regression model. All statistical analyses were performed using Stata statistical software (version 12.1, StataCorp, College Station, TX). Statistical significance was set at P < 0.05.

Results

Patient Sample

A total of 2,428 patients met inclusion criteria; 363 patients had missing laboratory values (white blood cell counts or eosinophil counts) and were excluded from the study, leaving a final study cohort of 2,065 patients which met criteria for inclusion: 1,064 patients from PSU and 1,001 patients from UVA. Of the study population, 799 (38.7%) presented with an admission eosinophil count of 0 k/μL. Patients with eosinophil counts of 0 k/μL had a higher mean age (64.2 years vs. 61.6 years; p=0.002), but they otherwise did not differ significantly by sex, race or comorbidity burden from those with measurable admission eosinophil counts. Patients with admission eosinophil counts of 0 k/μL were more likely to be hypotensive (p=0.001) and tachycardic (p<0.001) and were more likely to require mechanical ventilation (7.3% versus 4.3%; P = 0.005), with higher admission WBC counts (11 k/μL vs 16 k/μL; p<0.001). (Table 1).

Table 1.

Comparison of Patient Demographics, Comorbidities, and Laboratory Parameters Between Eosinophil Level Groups

	Eosinophil > 0 k/μl	Eosinophil = 0 k/μl
Variable	(n = 1,266)	(n = 799)	P-value
Age	61.6	64.2	0.002
Age 18–45	248 (19.6%)	109 (13.6%)
Age 45–60	289 (22.8%)	195 (24.4%)
Age 60–75	410 (32.4%)	256 (32%)
Age >75	319 (25.2%)	239 (29.9%)
Sex			0.257
Male	584 (46.1%)	389 (48.7%)
Female	682 (53.9%)	410 (51.3%)
Race/ethnicity			0.138
White, non-Hispanic	1071 (84.6%)	700 (87.6%)
Black, non-Hispanic	135 (10.7%)	60 (7.5%)
Hispanic	5 (0.4%)	3 (0.4%)
Asian	6 (0.5%)	3 (0.4%)
Other	31 (2.4%)	8 (1%)
Charlson comorbidity index*	2 (0–6)	2 (1–6)	0.242
Clinical markers of severity
Mechanical ventilation	55 (4.3%)	58 (7.3%)	0.005
Hypotension (SBP<90 mmHg)	58 (4.6%)	64 (8.0%)	0.001
Tachycardia (>100 beats/min)	275 (21.7%)	254 (31.8%)	<0.001
Laboratory parameters
Eosinophil count, k/μl*	0.10 (0.05–0.21)	0.0 (0.0–0.0)	<0.001
Total WBC count, k/μl*	11.05 (7.5–15.7)	16.1 (9.5–24.5)	<0.001
Creatinine level, mg/dL*	1.05 (0.8–1.8)	1.2 (0.8–2.2)	0.071

^*Reported as median (IQR)

SBP=systolic blood pressure

WBC=white blood cell

Training Cohort

Using data from PSU (n = 1,064), inpatient mortality was first compared between cohorts. Unadjusted comparisons of mortality rates between patients with admission eosinophil counts of 0 k/μL (n=454) and those with eosinophil counts > 0 k/μL (n=610) in the training cohort are shown in Figure 1, demonstrating that mortality rates were significantly higher among patients admitted with eosinophil counts of 0 k/μL than among patients with counts > 0 k/μL (10.1% versus 3.3%; P < 0.001).

Figure 1.

Comparisons of primary and secondary outcomes between the eosinophil level groups. Two-sided univariate χ² tests were used to compare outcomes of patients with eosinophil count>0 k/μL (n = 610) and patients with eosinophil count=0 k/μL (n = 454). For all four outcomes, P < 0.05 between groups. Error bars represent 95% confidence intervals.

A multivariable logistic regression model was then constructed using the training cohort. After controlling for other patient characteristics to isolate the effect of eosinophil count, an admission eosinophil count of 0 k/μL was an independent predictor of inpatient mortality (odds ratio [OR] 2.01; 95% confidence interval [CI] 1.08 – 3.73). Similarly, WBC ≥15 k/μL was associated with over twice the odds of inpatient mortality compared to WBC <15 k/μL (OR 2.69; 95% CI 1.44 – 5.05). Serum creatinine levels were not significantly associated with inpatient mortality (OR 1.11 per 1-unit increase in serum creatinine, 95% CI 0.97 – 1.28). Results for other model parameters are shown in Table 2. The area under the ROC curve associated with this model for mortality was 0.821, suggesting that this model had good ability to predict the occurrence of mortality.

Table 3 shows the sensitivity, specificity, and accuracy of the model provided in Table 2 in predicting mortality. For example, predicting a patient to experience mortality at a cutoff of the probability of mortality ≥10%, 56.1% of deceased patients were correctly predicted to experience mortality, while 87.4% of patients who were not predicted to experience mortality were correctly predicted to survive. At a higher threshold for predicted mortality, ≥70%, only 4.6% of deceased patients were correctly predicted to experience mortality, whereas 99.9% of patients who were not predicted to experience mortality were correctly predicted to survive. Since the number of patients who experience mortality was low relative to those who survived, this model has > 80% accuracy for any patient whose predicted probability of mortality exceeds approximately 7.5%, and has an accuracy >90% for any patient whose predicted probability of mortality exceeds 20%.

Secondary Outcomes

In examining secondary outcomes of interest among the training cohort, patients with admission eosinophil counts of 0 k/μL more frequently required admission to monitored care settings, more frequently required the use vasopressors for the treatment of septic shock, and were more likely to require total colectomy for severe, medically refractory disease when compared to those with admission eosinophil counts >0 k/μL (all P < 0.05; Figure 1).

Validation Cohort

In the validation cohort of UVA patients (n = 1,001), inpatient mortality was significantly higher on univariate analysis comparing the group with admission eosinophil counts of 0 k/μL to those >0 k/μL (14.2% versus 6.6%; P < 0.001). As observed in the multivariable model for the training cohort, mortality was significantly higher when comparing the group with admission eosinophil counts of 0 k/μL to those with counts >0 k/μL (OR 2.26; 95% CI 1.33 – 3.83). However, total WBC counts ≥15 k/μL did not show a significant association with mortality among the UVA patient cohort (OR 1.33; 95% CI 0.78 – 2.26). The area under the ROC curve fit with UVA data was 0.863.

Sensitivity and Subgroup Analyses

Considering eosinophil count as a continuous variable within the model rather than as a dichotomized variable, there was no significant linear association between eosinophil count and mortality (OR 1.03, 95% CI 0.54 – 1.98). Eosinophil count was then modeled as a categorical variable, with quartiles of non-zero counts stratified into (1) eosinophils >0 & ≤0.06 k/μL, (2) eosinophils >0.06 & ≤0.1 k/μL, (3) eosinophils >0.1 & ≤0.22 k/μL, and (4) eosinophils >0.22 k/μL. The only category of eosinophil count that was significantly associated with mortality remained the subgroup with eosinophil counts of 0 k/μL (OR 2.91, 95% CI 1.08 to 7.86). Odds ratios for all pairwise comparisons are presented in Appendix 2, with inferences adjusted for multiple comparisons.

Importantly, in the subgroup analysis of patients known to present without vital sign derangements (tachycardia or hypotension), patients with eosinophil counts of 0 k/μL had significantly higher odds of mortality (OR 5.76, 95% CI 1.99 – 16.64), whereas those with WBC ≥15 k/μL did not (OR 1.62, 95% CI 0.64 – 4.10).

Discussion

The present study finds that admission eosinophil counts allow for an immediate assessment of mortality risk at admission, one which is inexpensive and part of a differential for a standard complete blood count available at any hospital. The absence of eosinophils (as opposed to a range of eosinophils, which may vary slightly between hospital laboratories) is not only an independent predictor of inpatient mortality but it is also associated with higher odds of severe disease requiring intensive care, vasopressor use, and need for surgery. These clinical findings correlate well with the recent discovery in a murine model that peripheral blood eosinophils are depleted in connection with binary toxin produced by certain strains of C difficile, likely through accelerated apoptosis rather than reduced eosinopoeisis.⁸ Although certain ribotypes are known to exhibit increased toxigenicity,¹⁴ the data correlating specific ribotypes with worsened clinical outcomes are divergent,^5,15–16 and tests for binary toxin detection and ribotyping are not available as a part of clinical care.⁵ The measurement of eosinophil levels is a widely available measure that serves as a marker for key adverse outcomes, which in turn affect length of stay, cost of care and mortality.

It is unsurprising that a recent systematic review¹⁷ of risk factors for mortality in CDI did not include the association of eosinophil levels and disease-related adverse events, as there is little published in the literature on this topic. Two previous studies^18,19 examining mortality in CDI patients used a wide variety of laboratory parameters (C-reactive protein, alkaline phosphatase, cholesterol, sodium, calcium, lactate dehydrogenase) as well as eosinophil counts in their analyses. These studies found that lower eosinophil counts obtained at various time points following hospital admission were associated with higher mortality; however, these studies examine eosinophil counts as a continuous biomarker, which as we previously noted may be subject to variations in laboratories between hospitals. Our study also provided a subgroup analysis of patients who initially presented to the hospital without any derangement in vital signs (heart rate >100 beats per minute or systolic blood pressure <90 mmHg), and demonstrated that eosinophil counts of 0 k/μL were predictive of inpatient mortality; conversely, the more commonly used indicator of severe disease—WBC >15 k/μL—was not significantly associated with mortality among this subgroup. Since there are data that suggest that inflammation from the host immune system can actually be deleterious to the host in situations where the inflammatory response is overactive,^20,21 an undetectable admission eosinophil may be a more reliable marker for ribotypes of C. difficile that promote the patterns of host immune response that are most associated with adverse outcomes. By contrast, a significantly greater degree of attention has been given to the prominent leukocytosis frequently associated with CDI. Interestingly, the finding that numerous toxin dependent and toxin independent mechanisms²² in CDI are a likely cause for this leukocytosis raises issues as to whether the absolute value of the white blood cell count (as opposed to its trend) is as much a measure of severity of colitis as it is a combination of colitis severity and immune response. This may help to explain the finding that for patients with normal vital signs upon admission, eosinophil count correlated with mortality while white blood cell count did not.

This study has several limitations. The findings from data between the two participating institutions were consistent; although questions regarding generalizability given different strains of C. difficile could be raised, the use of training and validation cohorts suggests the results are generalizable. This study is also retrospective, and prospective studies that better control for potential confounders, such as baseline medication usage, immune status, and CDI therapies, are required. A retrospective approach may also introduce issues related to selection bias; our inclusion criteria would tend to favor more severe forms of CDI as these require inpatient hospitalization. However, since eosinophil counts are not currently used to guide management decisions at either institution, the focus of this study on admission eosinophil counts is unlikely to have introduced bias into our observed results. Prospective studies will be necessary to demonstrate that the information provided by our model would lead to improved outcomes in CDI patients, as well as to answer questions regarding whether eosinopenia, or recovery of a detectable eosinophil count, at time points after admission are predictive of disease outcomes. Additionally, although the data regarding the association between ribotyping and CDI severity is conflicted in the literature, the lack of availability of ribotyping data as well as any other characterization of the predominant strain of C. difficile among subjects in this study are important data that were unavailable in the current work since this analysis is not a part of routine clinical care. Lastly, although the findings of Cowardin et al. suggest a biologically plausible mechanism for the observed association between the absence of peripheral eosinophils and increased C. difficile virulence, further studies are required to better elucidate this mechanism in human subjects.

Our group recently published results for a scoring system predicting postoperative mortality in CDI patients who require total colectomy.²³ Our ongoing research in the area of the microbiome in CDI strongly indicates that a host immune reaction associated with an exaggerated inflammasome response is potentially more predictive of severity than many of the commonly assessed clinical indices. Peripheral eosinopenia is a potential marker for this exaggerated inflammasome response, and our group is in the process of prospectively evaluating an at-admission prognostic score which includes eosinopenia. This future study will not only further evaluate the prognostic value of eosinopenia, but it will address the question of whether improved prognostication at the time of admission will result in decreased mortality by guiding treatment decisions at the very beginning of inpatient care.