Attributable Outcomes of Endemic Clostridium difficile–associated Disease in Nonsurgical Patients

CDAD led to significantly worse outcomes in these patients.

For each patient, a modifi ed APACHE II Acute Physiology Score (APS) was calculated to adjust for severity of illness (15). The APS was based on laboratory results and vital signs collected within 24 hours of admission. The score was modifi ed because data for respiratory rates and Glasgow coma scores were unavailable electronically. In addition, the Charlson-Deyo method was used to classify co-existing conditions (16,17). Albumin levels within 24 hours of admission were collected and categorized into normal (>3.5 g/dL), low (2.5-3.5 g/dL), and very low (<2.5 g/dL). Multiple imputation methods were used to impute albumin levels for patients without recorded values (18). For CDAD case-patients, only medication and intensivecare unit exposures before the patient's fi rst positive stool toxin assay were included in analyses.

Case Defi nition
CDAD case-patients were defi ned as inpatients with positive C. diffi cile stool toxin assays (TechLab, Blacksburg, VA, USA). The microbiology laboratory only performs toxin tests on unformed stool, so all patients with a positive result for toxin were considered case-patients. Both community-onset and hospital-onset CDAD casepatients were included in the analyses.
Analyses were performed on the full cohort and a nested case-control population. The fi rst component was a retrospective cohort. For CDAD patients, the admission date when the patient's CDAD was fi rst identifi ed was used as the index admission. For noncases with >1 admission during the study period, 1 admission was randomly selected as the index admission. The nested case-control population consisted of propensity score matched cases and controls from patients identifi ed in the cohort.

Cohort Data Analysis
Survival was defi ned as the number of days from the index hospital admission until death. Survival was censored at 180 days. Time to readmission was calculated as the number of days between the index hospitalization discharge date and the date of the subsequent admission to BJH, if applicable. Days until readmission were censored at death or 180 days, whichever occurred fi rst.
Fisher exact, χ 2 , and Mann-Whitney U tests were used to compare characteristics of patients with and without CDAD. Time-to-event methods were used to estimate the effect of CDAD on 180-day survival and time-to-readmission. Patients who died during the index hospitalization were excluded from the time-to-readmission analysis. Kaplan-Meier analysis was used to evaluate the unadjusted relationships between CDAD and time-to-event outcomes.
Cox proportional hazards regression was used to estimate unadjusted and adjusted hazard ratios and 95% confi dence intervals (CIs). All variables with biologic plausibility or p<0.15 in the univariate analysis were considered in the multivariable Cox regression analysis by using backward stepwise selection. Variables were sequentially removed from the fi nal model, starting with the variable most weakly associated with the outcome. The signifi cance of individual covariates was determined by using a Wald statistic of p<0.05. The proportional hazards assumption was verifi ed by assessing the parallel nature of curves in log-log plots.

Propensity Score Matched-Pairs Analysis
The second component of this study was a propensity score matched-pairs analysis of outcomes attributable to CDAD. This study design complemented the cohort by enabling analyses that could not be conducted in the entire cohort, specifi cally hospital discharge status, attributable length of stay, attributable time-to-readmission, and attributable death. Hospital discharge status could not be analyzed for the entire cohort because manual review of medical records was required to determine the discharge location of each patient. The large size of the cohort prohibited this analysis. In addition, survival and time-to-readmission estimates generated in the cohort analysis were validated in the matched-pairs analysis.
Cases and a subset of controls were selected from the primary cohort for the matched-pairs analysis. CDAD casepatients were matched to controls based on their propensity for CDAD to develop. Patient-specifi c probabilities of developing CDAD were predicted by a full logistic regression model adjusted for all variables suspected to impact the risk of developing CDAD (online Appendix). Variables with p<0.05 in univariate analysis or biologic plausibility were included in the full logistic regression model. CDAD case-patients and controls were matched by a 1:1 ratio that used the nearest-neighbor method within calipers of 0.015 standard deviations (19). CDAD cases without an available nearest-neighbor control were excluded from the analysis. Chi-square, Fisher exact, and Mann Whitney U tests were used, as appropriate, to compare characteristics of CDAD case-patients and controls.
Medical records were reviewed for all CDAD casepatients and controls to determine hospital discharge location for each patient. Patients were categorized as being discharged to home, to a long-term-care facility, or to an outside hospital or dying in the hospital. Long-term-care facility was defi ned as a long-term-care facility, long-term acute-care facility/chronic ventilation facility, inpatient rehabilitation facility, skilled nursing facility, or nursing home. Outside hospital was defi ned as a non-BJH hospital or acute-care facility.

Data Analysis
Median length of stay was determined for CDAD casepatients and controls. The difference in median pairwise length of stay was compared with the Wilcoxon signedrank test. Attributable length of stay was calculated as the median pairwise difference between CDAD case-patients and controls. Frequencies, adjusted odds ratios, and 95% CIs were calculated to determine if discharge location was associated with CDAD. CDAD-attributable 180-day readmission was calculated as the difference in readmission between CDAD case-patients and controls. Attributable deaths from 0-180 days, 0-60 days, and 61-180 days after admission were also calculated by using this method.
The primary survival endpoints of interest were death and readmission, which were both censored at 180 days or at death. Kaplan-Meier analyses, conducted by using log-rank tests, were used to determine relationships between the survival endpoints and CDAD. Cox proportional hazards regression stratifi ed by matched-pairs was used to obtain hazard ratios and 95% CIs. Violation of the proportional hazards assumption was verifi ed by the crossing nature of curves in the log-log plots. Therefore, we used an extended Cox regression model stratifi ed by matched-pairs for the periods <60 days and >60 days. The breakpoint of 60 days was chosen because the graph of survival curves for CDAD case-patients and controls diverged at ≈60 days. Violation of the proportional hazards assumption was confi rmed by the signifi cance of the coeffi cient for the product term between CDAD and <60 days and >60 days (20).
All tests were 2-tailed, and p<0.05 was considered signifi cant. Statistical analyses were performed with SPSS for Windows version 14.0 (SPSS, Inc., Chicago, IL, USA) and SAS version 9.1 (SAS Institute, Cary, NC, USA). The Washington University Human Studies Committee approved this project.

Results
Among 18,050 nonsurgical inpatients admitted during the 1-year study period, 390 had CDAD and 17,660 did not. Selected patient characteristics of the cohort are summarized in Table 1. CDAD patients were signifi cantly older (median 66.0 vs. 52.7 years, p<0.001) more likely to be men, and more likely to be Caucasian than were noncase-patients. CDAD case-patients had a higher severity of illness on admission than noncases, as indicated by the modifi ed APS. CDAD patients were more likely to have a diagnosis of congestive heart failure, chronic obstructive pulmonary disease, cancer, leukemia or lymphoma, and metastatic solid tumors. Of 17,492 patients alive at the index hospitalization discharge, 4,207 (24%) were readmitted to BJH within 180 days. Fifty-two percent of CDAD patients were readmitted within 180 days versus 23% of noncases (log-rank p<0.001). Univariate and multivariable Cox regression results for time to readmission are presented in Table 2. The adjusted hazard ratio (HR) for readmission within 180 days was signifi cantly higher for CDAD case-patients than noncases (HR 2.19, 95% CI 1.87-2.55) ( Table 2).
The propensity score matched-pairs analysis included 353 CDAD cases and 353 controls (N = 706). There were no signifi cant differences between the matched cases and controls after correcting for multiple testing with the Bonferroni procedure. Thirty-seven CDAD case-patients were  dropped because a nearest-neighbor control was not available. Unmatched CDAD patients had signifi cantly higher modifi ed APS (median = 7.0 vs. 5.0, p<0.001), longer median length of stay (13.6 days vs. 9.6 days, p = 0.01), and higher percentage of deaths at 180 days (59% vs. 36%, p = 0.01) than matched case-patients.
In the matched-pairs analysis, median length of stay was 9.6 days for CDAD patients compared with 5.8 days for controls, and the increased attributable length of stay for CDAD patients was 2.8 days (Wilcoxon signed-rank p<0.001). Among the 706 patients in the matched-pairs analysis, 445 (63%) were discharged to home and 188 (27%) were discharged to a long-term-care facility. Only 7 (1%) patients were discharged to an outside hospital; therefore, these patients were combined with patients discharged to a long-term-care facility in the analysis. CDAD patients were signifi cantly more likely than controls to be discharged to a long-term-care facility or outside hospital (32% vs. 23%, odds ratio 1.62, 95% CI 1.15-2.28, McNemar p = 0.01).
Among 290 matched-pairs with both patient and control alive at index hospitalization discharge, 148 CDAD patients were readmitted to BJH within 180 days compared with 92 controls, for an attributable readmission of 19.3% (11.4%-27.2%). In the Kaplan-Meier and Cox model analyses, CDAD patients were signifi cantly more likely than controls to be readmitted to the hospital within 180 days ( Figure 2, Table 4).
By 180 days after hospital admission, 127 CDAD patients died compared with 107 controls, for an attributable mortality of 5.7% (95% CI -1.3%-12.6%). Although CDAD case-patients were no more likely than controls to die within 60 days of hospital admission, a divergence in survival between CDAD case-patients and controls began 60 days after hospital admission ( Figure 3, Table 4). The HR for death from 61-180 days was signifi cantly higher for CDAD patients than controls (HR 2.00, 95% CI 1.47-2.72) ( Table 4). Among 223 matched-pairs with both casepatients and controls alive after day 60, 19.7% of CDAD patients and 12.6% of controls died within 180 days for an attributable mortality between 61-180 days of 7.2% (95% CI 0.4%-14.0%).

Discussion
The results of this study indicate that CDAD is a major contributor to death even in nonoutbreak settings. In this CDAD-endemic setting, the disease was associated with a 23% higher hazard of death within 180 days after hospital admission in the multivariable cohort analysis and a 7.2% attributable mortality 61-180 days after hospital admission in the matched-pairs analysis. Historically, endemic CDAD has been reported to be associated with minimal increased risk in mortality although NAP1 strain CDAD outbreaks have been associated with much higher attributable mortality (10,11,13). Two studies of CDAD in endemic settings reported 1.2%-1.5% inhospital mortality rates from CDAD (10,13). Using a multivariable Cox proportional hazards model, Kyne et al. found no association between CDAD and 1-year mortality, although that study was quite small (47 CDAD patients) (11). In contrast, several studies have identifi ed increased deaths associated with outbreaks of the NAP1 strain. Pepin et al. estimated the 1-year attributable mortality of CDAD during an outbreak with the NAP1 strain to be 16.7% (9). Hubert et al. reported that CDAD was the attributable or contributive cause of death in 8.4% of patients infected with a strain of C. diffi cile that had the binary toxin and tcdC deletion (21). Loo et al. found CDAD to be the attributable cause of death within 30 days in 6.9% of CDAD patients and suspected that CDAD contributed to death in another 7.5% of CDAD patients (12). The estimate of 6.9% attributable mortality, however, was determined through chart review, not through multivariable analyses, and medical chart review may not be an adequate method to determine attributable mortality because of subjectivity (22).
Although the 5.7% 180-day attributable mortality determined in the propensity score matched-pairs analysis in our study was not statistically signifi cant, the estimate is substantially higher than estimates reported from other CDAD-endemic settings. The attributable mortality we report is more consistent with estimates from outbreaks of the NAP1 strain and is likely clinically signifi cant. The NAP1 strain was fi rst identifi ed at BJH during 2005, but the strain may have been present during the study period (23). During the years 2000-2006 at BJH, there were no apparent increases in hospital-onset CDAD incidence rates or severity of CDAD (as measured by the number of colectomies per CDAD case per year and by the percentage of patients with CDAD who died during hospitalization) (data not shown). Thus, the high attributable mortality found in this study has important implications for patients: CDAD remains an important cause of patient death even in a CDAD-endemic setting.
Our study showed that CDAD had a delayed impact on death. In the matched-pairs analysis, the divergence in survival between CDAD cases-patients and controls did not begin until >60 days after hospital admission. Within 60 days of admission, survival was not signifi cantly different between CDAD patients and controls, when all but 4 (1%) patients had been discharged from the hospital. This fi nding is consistent with those of 2 recent nested matched case-control studies in nonoutbreak settings, in which no signifi cant excess deaths were reported after 30 days (24) or at discharge (25). Although CDAD can be acutely lifethreatening, delayed death caused by CDAD may not be easily recognized as related to the initial CDAD episode. CDAD may contribute to a decline in patient function and overall illness over time, ultimately leading to death in many patients.
The results of the time-to-readmission and discharge location analyses further emphasize the negative impact of 1036 Emerging Infectious Diseases • www.cdc.gov/eid • Vol. 14, No. 7, July 2008  CDAD. CDAD patients were more than twice as likely to be readmitted to BJH within 180 days compared with controls. This fi nding is consistent with our prior fi ndings that CDAD contributes to an increase in hospital costs extending out to at least 180 days (26). CDAD patients were also signifi cantly more likely to be discharged to a long-termcare facility or outside hospital. Few data are available on the health of CDAD patients after hospital discharge, and future studies following CDAD patients as outpatients over an extended period are needed. Data on the excess length of hospital stay attributable to CDAD are limited. Wilcox et al. found that CDAD patients stayed in the hospital, on average, 21.3 days longer than non-CDAD patients; however, the attributable length of stay was not calculated (14). O'Brien et al. reported that the mean increase in hospitalization among CDAD patients was 2.9 days (27). Kyne et al. calculated the attributable length of stay at 3.6 days (11), which was comparable to the attributable length of stay estimate found in our study (2.8 days).
Our study has several limitations, including the retrospective study design. Use of electronic data from the hospital's Medical Informatics database has limitations, although use of these data made analysis of such a large cohort feasible. Differences seen in observational studies may be due to unmeasured confounders. We attempted to address this issue by using 2 methods to control for confounding: multivariable regression analyses and propensity score matched-pairs analyses. As evident from the Kaplan-Meier mortality analyses, the matched-pairs population is a more homogeneous population than the cohort. This design allows more precise effect estimation because the association between CDAD and the propensity score variables among the study participants is eliminated. A strength of the multivariable regression analyses is the use of all available data in the cohort. In the propensity score matched-pairs analyses, 37 CDAD cases were excluded because of lack of a suitable control. Unmatched case-patients were more severely ill than matched case-patients, and their exclusion is a limitation of the propensity-score matched-pairs analyses. In the time-to-readmission analyses, we were unable to identify readmissions to hospitals other than our institution. Finally, surgical patients were excluded from these analyses. Because of this exclusion, the most severely ill CDAD patients requiring colectomies (n = 3) were not represented in the dataset. The absence of these patients, as well as the 37 unmatched case-patients, may have resulted in estimates of attributable length of stay and death that are biased low.
Data on attributable outcomes associated with CDAD are scarce. As previously mentioned, some data on attributable mortality and length of stay exist; however, these fi ndings are limited by lack of adequate controls, small sample size, or outbreak settings. Our study provided detailed analysis on the effect of CDAD on time-to-readmission.
Another key strength of this study is the combination of 2 analytical methods: Cox proportional hazards regression in the primary cohort and propensity score matchedpairs analysis. Mortality and time-to-readmission analyses, which were conducted in both the cohort and matchedpairs populations, had remarkably similar results. The results of this study suggest that endemic CDAD can lead to signifi cantly poorer patient outcomes, including increased hospital length of stay, death, risk for admission to a longterm-care facility, and risk for hospital readmission. Even when the most severe CDAD cases are not considered, the detrimental effect of CDAD on patient health appears to extend beyond hospital discharge. Although prospective validation of these fi ndings is needed, proper allocation of healthcare resources toward prevention of this infection is necessary to prevent further illness and death attributable to CDAD.