Assessment of candidemia-attributable mortality in critically ill patients using propensity score matching analysis

Candida spp. are the fourth most common cause of bloodstream infection in hospitalized patients [1–4] and a leading cause of invasive fungal infection in the ICU setting [1]. Candidemia in critically ill patients is usually a severe and life-threatening condition with a high crude mortality ranging between 35% and 75% [5, 6]. In contrast, candidemia-attributable mortality is unclear, with studies showing high rates of 14.5% to 49% and others failing to demonstrate any significant increase in mortality [7].

Infection caused by Candida spp. develops in patients with multiple risk factors and severe underlying diseases, so that it is difficult to distinguish mortality attributed to candidemia from mortality related to the underlying illness. Thus, there is an interest in conducting case-control or matched cohort studies in which attributable mortality is obtained after matching and adjusting for confounding variables. Despite these techniques it seems almost impossible to match all factors (observable and unobservable) than can potentially have an influence upon mortality. This obstacle was overcome by Rosenbaum and Rubin [8] in 1983 who suggested matching 'cases' and 'controls' solely on their 'propensity score' - the estimated probability of being a 'case' given observable characteristics. In a scenario of ICU-acquired candidemia, each patient with candidemia is matched to a patient without candidemia who is most similar in terms of being a candidemic patient, where this probability is calculated on the basis of individual characteristics. Once the two groups are formed, the average effect is estimated for each outcome by simply computing the difference in means between the two groups. In recent years, the use of propensity score analysis in observational studies has increased considerably [9].

This study was designed to determine the attributable mortality of ICU-acquired candidemia in non-neutropenic critically ill patients using propensity score matching analysis.

The median duration from ICU admission to the onset of candidemia was 14 days (IQR 8.75 to 19.25 days). Abdominal surgery was the diagnosis on ICU admission in 31 (81.6%) patients. Infections were caused by C. albicans in 22 episodes, C. parapsilosis in nine, C. tropicalis in three, C. glabrata in two, C. krusei in one and Candida ssp. in one. The distribution of variables including severity indexes, ICU length of stay and mortality was similar in patients with candidemia caused by C. albicans and non-albicans Candida spp. As for the source of candidemia, infection of unknown origin was reported as the most frequent (n = 29, 76.3%) followed catheter-related infections (n = 8, 21.1%). One patient had candidemia and peritonitis concomitantly (2.6%). Multifocal colonization was a clinically relevant condition which was present in up to 91.4% of all patients before the development of candidemia. C. parapsilosis was the most frequently isolated species among the non-albicans spp in bloodstream infections and had a lower crude mortality (33%). Seven patients were not given antifungals and all of them died. Thirty-one patients were given antifungal treatment and in all cases the selection of the antifungal drug was appropriate according to the Candida spp. isolated. The following antifungal agents were used either as the sole agent throughout the course of treatment, or in a sequential pattern resulting in the use of multiple agents for a single episode (that is, de-escalation): fluconazole was used most often (n = 21, 67.7%), followed by amphotericin B-based (BL/CL) preparations (n = 7, 22.6%), caspofungin (n = 6, 19.4%) and voriconazole (n = 4, 12.9%). The median time for initiation of antifungal therapy was one day (IQR 0 to 4 days). In two patients with candidemia (5.7%), antifungal treatment was administered preemptively on the same day as material for blood culture was collected (day 0). The rest of the patients with candidemia were treated with targeted antifungal treatment after notification of a positive Candida blood culture result. Nearly one third of the episodes (n = 12, 31.6%) were treated between 24 and 48 hours (day one) after blood cultures were obtained (in some cases before final Candida species identification).

The crude ICU mortality rate was 52.6% (20/38) in patients with candidemia and 20.6% (220/1,069) in those without candidemia (P < 0.001). The crude hospital death rate was 55.3% (21/38) and 29.6% (314/1,069) among patients with and without candidemia, respectively (P = 0.01).

In the univariate analysis, age (hazard ratio (HR) 1.03, 95% CI 1.0 to 1.06; P = 0.032), APACHE II score on ICU admission (HR 1.09, 95% CI 1.12 to 1.81; P = 0.013), APACHE II score at diagnosis of candidemia (HR 1.13, 95% CI 1.03 to 1.24; P = 0.006), SOFA score at diagnosis of candidemia (HR 1.14, 95% CI 1.03 to 1.27; P = 0.011), and delay in starting antifungal therapy for more than 72 hours (HR 2.95, 95% CI 1.21 to 7.20; P = 0.017) were significantly associated with death. In the multivariate analysis, APACHE II score at the time of diagnosis of candidemia was the only predictor of death (adjusted HR 1.13, 95% CI 1.03 to 1.24; P = 0.006).

General findings obtained in this study, including an incidence rate of 34.3 episodes of candidemia per 1,000 ICU patients and a crude mortality rate of 52.6% are consistent with previously published data [5, 20]. Attributable ICU mortality to Candida-related bloodstream infection varies largely and has been estimated to be non-significantly different from that of patients without candidemia or to be higher by 5% to 40% [7, 21, 22]. For this reason, analysis in matched cohort samples is recommended, although only seven studies assessing attributable mortality of candidemia have been previously published [23]. Important limitations are related to methodological heterogeneity, including study design (usually retrospective), study population (patients hospitalized in ICUs and hospital wards and those undergoing transplantation), source of data (hospital or microbiological databases) and matching criteria mainly based on underlying disease and comorbidity, rather than on all possible factors that may influence mortality. Moreover, severity scores and length of ICU stay were not considered.

In the present prospective and multicenter study, candidemia was not associated with an increase in either ICU or hospital mortality. This finding may be explained by the use of propensity matching score analysis to control for all potential confounding variables related to the development of Candida spp. bloodstream infection. Matching was highly effective and both candidemic and non-candidemic groups had similar characteristics, differing only in the development of bloodstream infection.

These favorable results could be attributed to earlier treatment of bloodstream infection and better monitoring (weekly surveillance sampling). An early start of appropriate antifungal therapy (within the critical time-frame of the first 24 to 48 hours) is crucial for the reduction of bloodstream infection-related mortality [24, 25]. This observation confirms the importance of an empirical and preemptive strategy. Another important factor is the rapid reporting of the microbiological results (positive bloodstream infection and species identification) in order to initiate targeted antifungal treatment or to modify previous empirical antifungal agents. In our study, the median time for initiation of antifungal therapy was one day (IQR 0 to 4 days). In 36.9% of patients it was initiated within 48 hours of obtaining the first positive blood culture (5.2% of them as preemptive treatment). Another important problem preventing the earlier recognition and treatment of candidemia is the lack of specific clinical findings. Different prediction rules based on a variety of risk factors, including Candida species colonization are recommended to identify patients at high risk for fungal bloodstream infections [26, 27]. High density colonization can be used to identify patients who may benefit from preemptive antifungal treatment in the appropriate clinical setting [28]. In our study multifocal colonization was a clinically relevant condition which was present in up to 92.1% of all patients before the development of candidemia and there was a high similarity between causative and colonized strains; this could guide appropriate antifungal treatment. Additionally, an appropriate antifungal agent was administered in all treated patients and this may also contribute to increased survival. Moreover, earlier replacement of central venous catheters, common practice in ICU's, could also reduce mortality. On the other hand, antifungal therapy was initiated in 180 non-candidemic high-risk patients (n = 180/1,069, 16.8%) as an empiric or preemptive strategy. This antifungal strategy has not been evaluated in previous articles estimating candidemia-attributable mortality. As this can be a major confounding factor, empirical or preemptive treatment was introduced in the logistic regression model. There was no difference in this variable in the propensity score-matched patients as we list in Table 2. It has been shown that ICU patients with candidemia have a substantially higher severity of illness on the day of diagnosis compared with the ICU admission day [29]. This is also an important disadvantage for matching because severity scores are determined on ICU admission but not at the time of developing candidemia. However, after matching by the propensity score, APACHE II and SOFA severity scores recorded weekly did not vary significantly between cases and controls.

The crude ICU mortality in patients with candidemia in our study is 52.6% with a median of five days from the diagnosis until death. Due to the high mortality and poor prognosis of patients with Candida spp. bloodstream infection, there is a strong need to identify predictors of death. Factors associated with fatal outcome, such as age, malnutrition, delay in removal of central venous catheter, candidemia caused by non-albicans Candida spp. and delay in starting antifungal therapy have been extensively recognized [25, 30–32]. However, in our study only APACHEII score at the time of diagnosis of candidemia was a predictor of mortality.

The lack of differences in mortality rates between the groups of C. albicans and non-albicans Candida spp. may be explained by different reasons: i) Adequacy of antifungal therapy with all candidemias caused by Candida spp. with a possible decreased sensitivity or resistance to azoles being treated with candins or different formulations of amphotericin B; ii) a relatively high percentage of cases caused by C. parapsilosis (23.7%), associated with lower mortality [1, 33, 34]; and iii) a systematic removal policy for central venous catheters.

Two frequently cited studies [25, 29] have shown an association between mortality and delay in start antifungal treatment. It should be noted that in both studies, the majority of episodes of candidemia occurred in patients admitted to hospital wards, with an in-hospital mortality rate of about 30%. In our study, all patients had ICU-acquired candidemia, with a higher crude mortality. It may be possible that antifungal treatment is started earlier in critically ill patients admitted to the ICU than in less severely ill patients in the hospital wards where strict protocols of microbiological surveillance are not established.

Despite the fact that more than 1,000 patients were recruited, the small number of candidemias related to the incidence of invasive candidiasis in the critically ill patient is a limitation of the study and might have compromised our power to detect differences in mortality between cases and controls. Additionally, the sensitivity of blood cultures for candidiasis is low (approximately 60%). This could impact the actual incidence of candidemia and also the overall mortality in the ICU. Therefore, there is a possibility that patients with candidemia that may have had identical matching risk factors but negative blood culture were considered matching controls instead of cases.

In summary, ICU-acquired candidemia in critically ill patients was not associated with an increase of either ICU or hospital mortality. APACHE II at the time of diagnosis of candidemia was the only predictor of death in patients with candidemia.