Objective. Neonatal candidemia is often fatal. Empirical antifungal therapy is associated with improved survival in neonates and patients with fever and neutropenia. Although guidelines for empirical therapy exist for patients with fever and neutropenia, these do not exist for neonates.
Methods. A multicenter, retrospective, cohort study was conducted of neonatal intensive care unit patients (N = 6172) who had a blood culture (N = 21 233) after day of life 3 and whose birth weight was ≤1250 g. We performed multivariable conditional logistic regression of risk factors for candidemia. From the regression modeling coefficients, we developed a candidemia score.
Results. In multivariable modeling, thrombocytopenia (odds ratio [OR]: 3.56; 95% confidence interval [CI]: 2.68–4.74) and cephalosporin or carbapenem use in the 7 days before obtaining the blood culture (OR: 1.77; 95% CI: 1.33–2.29) were risk factors for subsequent candidemia. Children who were 25 to 27 weeks’ estimated gestational age (OR: 2.02; 95% CI: 1.52–3.05) and children who were born at <25 weeks (OR: 4.15; 95% CI: 3.12–6.29) were at higher risk of developing candidemia than were children who were born at ≥28 weeks. We developed a candidemia score on the basis of the ORs from the multivariable model. Children with a candidemia score ≥2 points were classified as having a “positive” score, and a score of ≥2 points had a sensitivity of 85% and a specificity of 47%.
Conclusions. We developed a clinical predictive model for neonatal candidemia with high sensitivity and moderate specificity for candidemia. On the basis of our model, when a physician obtains a blood culture, the physician should consider providing antifungal therapy to neonates who are <25 weeks’ estimated gestational age and to neonates who have thrombocytopenia at the time of blood culture. In addition, if a physician obtains a blood culture from a child who is 25 to 27 weeks’ estimated gestational age and is not thrombocytopenic but has a history of third-generation cephalosporin or carbapenem exposure in the 7 days before the blood culture, then the physician should consider administration of empirical antifungal therapy.
The incidence of candidemia is rising steadily, and Candida species are a leading cause of infectious mortality in the neonatal intensive care unit (NICU).1 Incidence of candidemia in extremely low birth weight ([ELBW]; <1000 g) infants is 4% to 15%.1–3 Candida bloodstream infection is associated with an attributable mortality of 38% and a crude mortality of 30% to 75%.4–6 Prompt diagnosis and initiation of antifungal therapy are crucial to survival across patient populations.2,7–9
Like other causes of neonatal sepsis, the current gold standard for the diagnosis of neonatal candidemia is the blood culture. Unfortunately, although the blood culture is sensitive for bacterial pathogens, it is a poor diagnostic tool for invasive fungal infections. From studies completed in animal models, adult patients, and children, the sensitivity of the blood culture to diagnose invasive candidiasis is poor.10–14 On the basis of autopsy studies, the sensitivity of the blood culture to diagnose invasive candidiasis is 29% when 1 vital organ is involved, and the sensitivity peaks at 80% when 4 or more vital organs have documented Candida invasion.
Early intervention in the therapy of invasive candidiasis in high-risk populations was designed to reduce morbidity and improve survival substantially. The rationale for empirical antifungal therapy in high-risk persistently febrile neutropenic patients is the early treatment of clinically occult disease, the prevention of disseminated candidiasis, and the prevention of invasive fungal infections. In response to the increasing incidence and inherent difficulty in diagnosing candidiasis and candidemia, investigators in the field of oncology developed a well-known clinical predictive model on which to initiate empirical antifungal therapy. Empirical antifungal therapy is initiated after 5 to 7 consecutive days of fever and neutropenia.7,8 This model was based on a complete cohort design and has been validated prospectively. Although there has been a concomitant rise in neonatal candidemia over the past 20 years, no such complete cohort model has been developed for neonates.2 We analyzed data from a complete multicenter cohort from 100 NICUs in an effort to develop a simple, clinically useful model to guide clinicians in the initiation of empirical antifungal therapy for the at-risk neonate.
We assembled a cohort of neonates from an administrative database. Clinical data for these neonates were recorded prospectively for the database and analyzed retrospectively for this article. Neonates who were eligible for inclusion in the study were born between September 16, 1996, and December 27, 2001; were ≤1250 g birth weight; and were discharged from 100 NICUs managed by Pediatrix Medical Group, Inc (n = 13 468). Infants who survived until at least the third day of life and had at least 1 blood culture on or after day of life 3 met inclusion criteria for this analysis (n = 6172). With a database from a computer-assisted tool that generates clinical progress notes on neonates cared for by the Pediatrix Group,15 we collected data on birth weight, estimated gestational age ([EGA]; based on the examination of the neonatologist), gender, race, and reported blood cultures obtained from each patient after day of life 3 until episode of candidemia, death, or discharge.
The unit of observation for this analysis was the blood culture. For each day that a blood culture was obtained, we evaluated the platelet count (lower limit of normal 150 000), presence of enteral feeding, ventilatory status, central vascular access, presence of intralipids, and use of vancomycin or other broad-spectrum antibiotics. We also evaluated the following from the 7 days before blood culture: vancomycin use, third-generation cephalosporin or carbapenem use, and previous blood cultures. Birth weight was evaluated as a continuous variable, and we evaluated race, gender, feeding, mechanical ventilation status, antibiotic use, intralipids, vascular access, and previous blood cultures as categorical variables. We evaluated platelet count and EGA as both categorical variables and continuous variables.
The outcome of interest was the first episode of candidemia. We excluded blood cultures that were obtained after an episode of candidemia in a child. These cultures included multiple positive cultures for Candida species over several days in a single episode of candidemia, recurrent episodes of candidemia, and cultures obtained after an episode of candidemia that were subsequently negative.
Because of a time sequence, whereby multiple blood cultures were obtained from the same individual, we used methodology specific to this longitudinal monitoring situation, as described in DeLong et al.16 This methodology implements a time-to-event analysis by means of conditional logistic regression approach in which patients are stratified according to the day of life that the culture was obtained. The daily values for platelet count, feeding and ventilation status, vancomycin, third-generation cephalosporin or carbapenem, intralipids, and vascular access were incorporated as time-dependent covariates.
For multivariable analysis, we conducted backward elimination (P value for retention <.1), and cases with missing values for any of the independent variables were excluded. From the final multivariable model, we developed a clinical predictive model. We based the predictive model on the β coefficients and what would be simple (and therefore useful) for clinicians.
For the predictive model, we assigned a score of 0, 1, or 2 for each of the 3 variables in the final multivariable model. Specifically, patients were assigned a score of 2 when they were thrombocytopenic on the day the blood culture was obtained (0 otherwise), a score of 1 when they had received third-generation cephalosporin or carbapenem in the 7 days before blood culture (0 otherwise), a score of 1 when they were EGA 25 to 27 weeks (0 otherwise), and a score of 2 when they were ≤24 weeks EGA (0 otherwise). So a child with estimated EGA of 25 weeks, with no history of third-generation cephalosporin or carbapenem use in the previous 7 days, and a platelet count <150 on the day of blood culture received a clinical predictive score of 3. The score for a child who had EGA 28 weeks, had received meropenem 2 days before the blood culture was obtained, and had a platelet count of ≥150 received a clinical score of 1.
Estimating Sensitivity and Specificity
We then used a value of 2 for the clinical predictive score as a cutoff value for a diagnostic test. When the clinical predictive score was ≥2, patients were categorized as having a positive candidemia score, and patients with a clinical predictive score <2 were categorized as having a negative candidemia score. We evaluated the ability of the candidemia score to predict candidemia by using the methodology of DeLong et al16 to calculate the sensitivity and specificity of this diagnostic test. We cross-tabulated the blood culture results versus the diagnostic test results by generating a 2 × 6 table based on all of the observations. Calculating sensitivity of the candidemia score from the 2 × 6 table is straightforward and is not inherently biased because each child can contribute, at most, 1 true positive test, but because patients can contribute multiple observations to nondiseased status (multiple negative blood cultures for Candida during their hospitalization), the specificity calculated from a simple 2 × 6 table will be inherently biased. To provide an unbiased estimate of the specificity, the methods presented by DeLong et al16 calculate the following: Specificity = (1 − sensitivity)*exp(β)/[(1 − sensitivity) *exp(β) + sensitivity].
In the above equation, exp(β) is the exponentiated value of the β coefficient estimated by the conditional logistic regression equation that has the outcome (candidemia) as the dependent variable and the diagnostic test score (positive or negative) as the independent variable.
In this article, we present likelihood ratios. A likelihood ratio is equal to the probability of the observed test result among the diseased divided by the probability of the same test result among the nondiseased. This ratio may be expressed mathematically as is equal to (1 − sensitivity) divided by the specificity. So a likelihood ratio for a negative test with 85% sensitivity and 47% specificity is equal to (1 − 0.85)/0.47 = 0.31.
The physician uses the likelihood in a 3-step process:
Convert the pretest probability (prevalence or risk of vertical transmission) to pretest odds. Pretest odds = probability/(1 − probability)
Determine the posttest odd by multiplying the pretest odds by the likelihood ratio. Posttest odds = pretest odd*likelihood ratio.
Convert the posttest odds to posttest probability. Posttest probability = odds/(1 + odds)
The sample size presented in the final model is less than the number of blood cultures analyzed because in 5870 of the 21 233 blood cultures analyzed, no platelet count was reported. These analyses were repeated with 3 different methods of imputation for the platelet count—the median platelet value was used, the platelet count the day before acquisition of the blood culture was used, and the infants were assumed to have a platelet count >150. These analyses produced results very similar to the analysis presented in this paper (data not shown).
We conducted the analysis using SAS 8.02 (SAS Institute, Cary, NC); reported P values are 2-tailed, and we used Wald 95% confidence intervals (CIs). Permission to conduct the analysis was provided by the Duke University Institutional Review Board. Because the analysis was completed on data without identifiers, written informed consent was not obtained.
The Pediatrix group cared for 13 468 infants who were born at <1250 g birth weight; 7296 neonates did not have a blood culture after day of life 3, and 6172 did. We therefore evaluated 6172 infants from 100 NICUs, and of the infants in this cohort, 348 (5.6%; 95% CI: 0.05–0.06) developed candidemia during their time in the NICU. Almost half of the infants (47%) were nonwhite, and almost half (48%) were female. EGA breakdown was as follows: 37% infants were 28 weeks’ EGA or older, 49% were 25 to 27 weeks’ EGA, and almost 15% were <25 weeks’ EGA.
From these 6172 premature infants, 23 369 blood cultures were obtained after day of life 3. We excluded 2136 blood cultures that were obtained after an initial episode of candidemia. The cohort of presumed sepsis episodes analyzed in this article was 21 233 blood cultures.
We conducted bivariable analysis of basic demographics and previously described risk factors for neonatal candidemia. From the conditional logistic analyses, decreased EGA was strongly associated with candidemia (odds ratio [OR]: 5.29; 95% CI: 3.97–7.03) for children with EGA <25 weeks compared with children ≥28 weeks. Children who were thrombocytopenic when their cultures were drawn were more likely to develop candidemia (OR: 4.58; 95% CI: 3.44–6.06). Children who required mechanical ventilation (OR: 1.76; 95% CI: 1.32–2.19) or intralipids (OR: 1.33; 95% CI: 1.00–1.67) or were on corticosteroids (OR: 1.72; 95% CI: 1.29–2.31) were also at increased risk of candidemia. Race and previous exposure to vancomycin were associated with increased candidemia (Table 1). Children who received a third-generation cephalosporin or carbapenem in the 7 days before the blood culture were more likely to have candidemia (OR: 1.98; 95% CI: 1.49–2.46), but children who received vancomycin in the week before the blood culture was obtained or required central vascular access were not at substantially increased risk of candidemia in bivariable analysis (Table 1).
We then included all of the independent variables in Table 1 into a conditional logistic regression with backward elimination, and the final multivariable model is presented in Table 2. Only 3 risk factors were significantly associated with candidemia in this cohort: thrombocytopenia (OR: 3.56; 95% CI: 2.68–4.74), third-generation cephalosporin and carbapenem use (OR: 1.77; 95% CI: 1.33–2.29), and extreme prematurity. Compared with children who were ≥28 weeks’ EGA, children with an EGA of 25 to 27 weeks were at greater risk of developing candidemia (OR: 2.02; 95% CI: 1.52–3.05), as were children with an EGA <25 weeks (OR: 4.15; 95% CI: 3.12–6.29). From this multivariable analysis, we developed a clinical predictive model for subsequent candidemia.
On the basis of the coefficients that generated the odds ratios in Table 2, we gave children who were thrombocytopenic a score of 2 (0 otherwise), children who had received a third-generation cephalosporin or carbapenem 1 (0 otherwise), children who had an EGA of 25 to 27 weeks a 1 (0 otherwise), and children who had an EGA of <25 weeks a 2 (0 otherwise). Children with a combined score (minimum = 0, maximum = 5) of 2 or more were classified as having a positive candidemia score.
We then developed a 2 × 6 table to evaluate the sensitivity of the candidemia score (Table 3) and as a “test” for candidemia. A score ≥2 had 85% sensitivity, and we present the sensitivities for each score. The specificity for the candidemia score should not be derived from the 2 × 6 table presented in Table 3 but by using the respective score as a cutpoint and applying the equation provided in the Methods section. The specificity of a candidemia score of 2 was 47%. We also developed a second scoring system whereby any points (clinical score ≥1) were considered a positive candidemia score. This resulted in a model with sensitivity of 99%, but the specificity was only 14%.
We evaluated data from a multicenter cohort to develop a multivariable model and subsequent clinical predictive model. The goal of designing the clinical predictive model was to provide guidance to neonatologists as to when to consider empirical antifungal therapy. Our goal was to provide a model analogous to the well-known clinical predictive model of initiating empirical antifungal therapy for patients with prolonged fever and neutropenia.7,8 Empirical antifungal therapy in high-risk populations such as ELBW infants and children with neutropenia is warranted because invasive fungal infections are common and extremely difficult to diagnose: the sensitivity of the blood culture for invasive candidiasis is <50%.
Empirical antifungal therapy has been shown to reduce mortality in a wide variety of immunocompromised patients, and although there is an algorithm for empirical antifungal therapy in children with neutropenia, there is no such model for neonates. The lack of a simple predictive model in neonatal candidemia is understandable; neonates with invasive candidiasis do not simply present with fever and neutropenia. The presentation of neonatal candidiasis is more complex, and development of a protocol is challenging. Although they are not neutropenic, neonates are clearly an immunocompromised host, and empirical antifungal therapy in neonatal candidemia has been linked to improved patient outcomes and reduced mortality.2,6
Most of the epidemiology of neonatal candidiasis has been limited to single-center investigations of risk factors and incidence rates. Risk factors that have had particularly strong associations or have been found in multiple studies have included2,17–23 thrombocytopenia, central vascular access, birth weight, EGA, mechanical ventilation, feeding intolerance, and antibiotic use (specifically, third-generation cephalosporin use).
Recently, a prospective multicenter investigation of risk factors for candidemia was reported. However, this study is limited because only 35 cases occurred during the study period.21 The authors provided confirmation of several risk factors for neonatal candidemia, including the observation that even among ELBW infants EGA was still an important risk factor. It is interesting that colonization was not a significant risk factor for candidiasis in the multivariable analysis. Although the Saiman study was notable both for its collaborative effort and solid fundamental design, the results were not translated into a clinically useful tool as we have done.
Like previous investigators, we found that (even among ELBW infants) extreme prematurity, thrombocytopenia, and recent third-generation and carbapenem use predicted subsequent candidemia. It is not so important for the purposes of this analysis whether such use of antibiotics causes candidemia; in fact, that should not be inferred from these data. Rather, recent exposure to third-generation cephalosporins and carbapenems is clearly a marker for candidemia. Rather than (or in addition to) inducing candidemia, these drugs may simply be an effective mechanism to measure how ill-appearing a patient is despite routine (ampicillin and aminoglycoside) antibiotics. Moreover, thrombocytopenia is much more likely to be a symptom than a cause of candidemia, much like fever, which is a symptom rather than a cause of invasive fungal infection in the neutropenic host. Although the factors that we present are likely to be a mixture of cause, marker, and symptom, they can be used to guide therapy; in fact, empirical therapy for the symptomatic immunocompromised host is a hallmark of antifungal management.
Recently, 2 randomized trials evaluated the utility of antifungal prophylaxis. These studies3,24 differed in their assessment of the utility of prophylaxis. Prophylaxis exposes far more patients with much greater cumulative doses of antimicrobial agents than empirical therapy; but even if prophylaxis ultimately achieves widespread acceptance, we believe that these data will be useful. Antifungal prophylaxis in populations in which it is currently indicated (eg, patients who have received stem cell transplantation) reduces but does not eliminate invasive fungal infections.25 In children who receive antifungal prophylaxis, empirical antifungal therapy continues to show benefit.
The critical questions posed in the studies of prophylaxis, however, are to determine which patients may benefit from early antifungal intervention. Empirical therapy based on a statistically sound risk assessment ultimately exposes fewer patients to antifungal compounds with possibly similar or even better outcomes.
These data do not support a specific length of empirical therapy, but if empirical therapy is initiated, then we recommend starting therapy when the blood culture is obtained and continuing antifungal therapy until blood cultures are formally interpreted as negative. This recommendation is based on the upper limit of the expected length of time for Candida to grow in the blood culture bottle. Specific choice of antifungal therapy (fluconazole, amphotericin B, lipid complex amphotericin) should be based on the epidemiology of the local NICU and should incorporate the patient’s clinical status and other medications with hepatic or renal toxicity.
A fundamental limitation to the analysis and predictive model that we provide is that these recommendations should be validated prospectively. Models of high performance can be developed for 1 cohort or data set but when tested in a second population do not perform well.
Our results may be applied as follows in the neonate who is older than day of life 3. Physicians should consider starting empirical antifungal therapy in the following circumstances: when the physician is obtaining a blood culture from a neonate with unexplained thrombocytopenia (regardless of gestational age); when the physician is obtaining a blood culture from a neonate with EGA <25 weeks (regardless of platelet count); or when the physician is obtaining a blood culture from a neonate who is EGA 25 to 27 weeks, is not thrombocytopenic, but has had third-generation cephalosporin or carbapenem coverage in the past 7 days.
Just as the model for empirical antifungal therapy for prolonged fever and neutropenia required repeated prospective validation before widespread use, these data will need to be validated prospectively in a complete cohort design. Before widespread use, these guidelines should be validated prospectively in a trial that includes adverse event rates, costs, and improved outcomes in the neonate at risk for candidemia.
Dr Benjamin received support from the National Institute of Child Health and Human Development (R03HD42940-01).
- Received December 19, 2002.
- Accepted March 6, 2003.
- Address correspondence to Daniel K. Benjamin, Jr, MD, MPH, PhD, Duke Clinical Research Institute, Box 17969, Durham, NC 27715. E-mail:
- ↵Stoll BJ, Hansen N, Fanaroff A, et al. Late-onset sepsis (LOS) in very low birth weight neonates: the experience of the NICHD Neonatal Research Network. Pediatr Res.2001;49 :228A
- ↵Benjamin DK Jr, Ross K, McKinney RE Jr, Benjamin DK, Auten R, Fisher RG. When to suspect fungal infection in neonates: a clinical comparison of Candida albicans and Candida parapsilosis fungemia with coagulase-negative staphylococcal bacteremia. Pediatrics.2000;106 :712– 718
- ↵Makhoul IR, Sujov P, Smolkin T, Lusky A, Reichman B. Epidemiological, clinical, and microbiological characteristics of late-onset sepsis among very low birth weight infants in Israel: a national survey. Pediatrics.2002;109 :34– 39
- Hughes WT. Systemic candidiasis: a study of 109 fatal cases. Pediatr Infect Dis J.1982;1 :11– 18
- Faix RG, Kovarik SM, Shaw TR, Johnson RV. Mucocutaneous and invasive candidiasis among very low birth weight (less than 1,500 grams) infants in intensive care nurseries: a prospective study. Pediatrics.1989;83 :101– 107
- ↵Kicklighter SD, Springer SC, Cox T, Hulsey TC, Turner RB. Fluconazole for prophylaxis against candidal rectal colonization in the very low birth weight infant. Pediatrics.2001;107 :293– 298
- Copyright © 2003 by the American Academy of Pediatrics