Secondary Logo

Journal Logo


Predicting bacteremia in children with fever and chemotherapy-induced neutropenia


Author Information
The Pediatric Infectious Disease Journal: January 2004 - Volume 23 - Issue 1 - p 61-67
doi: 10.1097/01.inf.0000106782.30100.4f
  • Free



Children and adolescents treated with chemotherapy for malignant disease frequently have episodes of fever during severe chemotherapy-induced neutropenia (fever and neutropenia). Compared with adults, children with fever and neutropenia present significantly less frequently with clinically apparent sites of infection, and the rates of serious complications and death are lower. In children and adults the proportion of episodes with detectable bacteremia is ∼20%, 1 with reported proportions ranging from 8 to 36%. 2–12 One-third to one-half of episodes are caused by Gram-negative bacteria. 2, 3, 5, 7–13 Immediate empiric broad spectrum antimicrobial therapy is mandatory in all episodes of fever and neutropenia in children. Case-fatality rates up to 80% in Gram-negative infections would be expected without effective antimicrobial therapy. 9 In contrast unnecessary antibiotic treatment should be avoided. It increases the risk of nosocomial infection, favors the development of antimicrobial resistances, causes adverse drug effects, elevates costs and impairs the quality of life of the concerned patients. Differentiation of episodes of fever and neutropenia caused by bacteremia or severe bacterial infection from other febrile episodes, however, can be difficult: In a large prospective study, Pizzo et al. 2 were unable to define factors predicting bacteremia at presentation. Subsequent prospective 3, 11 and retrospective 1, 8, 14 studies found specificities of predictive models around or below 20%, when aiming at a sensitivity of at least 90%. Decision-making tools for improved predictability of bacteremia might permit outpatient management of a subgroup of patients at low risk of bacteremia or severe bacterial infection. 15

We performed a retrospective analysis of all episodes of fever and neutropenia occurring in children treated in our institution between January 1,1993 and December 31, 2001. The aim was to define predictors of bacteremia in general and of Gram-negative bacteremia in particular, based on clinical, laboratory and radiologic information available to the physician in the emergency room situation. The prediction rule should reach a specificity within or above the range of previously published studies at a sensitivity of 95% or greater.


Type of study.

A retrospective, single site cohort study at the Division of Pediatric Hematology and Oncology, Department of Pediatrics, University of Bern, Bern, Switzerland, was performed.

Patient selection.

The charts of all children and adolescents diagnosed before the age of 17 years with a malignancy at our institution were screened for episodes of fever and neutropenia occurring between January 1, 1993 and December 31, 2001. During this period children presenting with fever and neutropenia were routinely hospitalized and treated with broad spectrum parenteral antimicrobial therapy, usually once daily ceftriaxone and amikacin, 16 and other measures of supportive care as needed. They were discharged from hospital early whenever possible, i.e. when they performed well after being afebrile for 48 h, blood cultures remained negative, leukocyte and/or absolute granulocyte counts (ANC) were rising and local infections were controlled.

Episodes of fever in neutropenia.

An episode of fever and neutropenia was defined as an axillary temperature of ≥38.5°C persisting for at least 2 h, or a single measurement of ≥39.0°C, in a child with severe neutropenia caused by myelosuppressive chemotherapy, i.e. ANC < 0.5 × 109/l or ANC < 1.0 × 109/l and expected to decline. 2, 9 Episodes of fever and neutropenia induced by myeloablative therapy were excluded from the study. Episodes of fever and neutropenia resulting from the initial bone marrow involvement by leukemia at the time of diagnosis, and not to chemotherapy, were excluded as well. Bone marrow involvement by lymphoma or solid tumors, or residual bone marrow involvement in leukemia after the time the ANC had risen beyond 0.5 × 109/l at least once, however, were not reasons for exclusion from the study.

Diagnostic information accessible in the emergency situation.

A total of 44 variables with possible relevance to bacteremia and accessible to the physician in charge within the first 2 h after fulfillment of the criteria of fever and neutropenia, were coded as categoric variables with up to 3 categories. A set of 39 variables has been defined as described, 17 consisting of 5 variables describing general history, 16 the history preceding presentation with fever and neutropenia, 9 results of clinical examination and 9 results of auxiliary tests. Besides 5 more variables were coded, 3 clinical ones and 2 summary assessments: (1) a clinically unequivocal diagnosis of bacterial infection; (2) a radiologically proven pneumonia; (3) a comorbidity requiring hospitalization independently from fever and neutropenia; (4) a severe bacterial infection (SBI) at presentation, which was defined as radiologically proven pneumonia, clinically unequivocal diagnosis of a bacterial infection or a serum C-reactive protein level of >150 mg/l as an indirect sign suggesting SBI, expanding the definition of Klaassen et al. 6; and (5) an a priori impossibility of outpatient management, defined as either the presence of an SBI or a comorbidity requiring hospitalization, or that patients were already hospitalized. An association of some of these items with the risk of bacteremia in fever and neutropenia had been established previously. 1–4, 8, 9, 11


Anaerobic and aerobic blood cultures were obtained before starting antimicrobial therapy, thereafter at intervals of at least 24 h when patients remained febrile, or when they had chills irrespective of the delay from the last blood cultures. Blood cultures were taken from all ports of existing central venous catheters, where applicable, or from existing peripheral lines. Peripheral cultures done in parallel with central cultures were not performed. 18 Bacteremia was defined as at least one positive blood culture using a qualitative automated culture system (BacT/ALERT; bioMérieux).


For univariate associations of variables with detected bacteremia and Gram-negative bacteremia, odds ratios (OR) and their exact conditional 95% confidence intervals (CI) were calculated. In this exploratory part of the study, no correction for the multiple tests performed was applied. A complete case approach was used for constructing multivariate decision trees predicting the risk of bacteremia and Gram-negative bacteremia. Variables with >10% of missing values and the variable “Year of fever and neutropenia” were discarded. Decision trees, regression type, were modeled applying binary recursive partitioning based on the least squares approach for node splitting. 19 Five percent of sample size was chosen as minimal node size, and pruning was performed by the alternative approach as described. 20 Terminal nodes were defined as low risk when the proportion of episodes with bacteremia was <5%. Aiming at an unbiased estimate of the predictive accuracy of the models, 100-fold cross-validation, 21, 22 a 10-fold extension of the usual 10-fold cross-validation without the drawbacks of n-fold cross-validation 23 was applied for evaluation of the performance of the models. Receiver-operating-characteristic (ROC) curves were calculated nonparametrically. 24 The details of variable coding, model construction and 100-fold cross-validation have been described elsewhere. 17 To determine the amount of bias introduced by the inclusion of multiple episodes per patient into the study, the analyses concerning bacteremia were rerun with the use of only the first episode per patient. Differences in patient characteristics between the first and the last episodes for patients with more than one episode were examined with exact conditional McNemar’s tests. Throughout the study, two sided statistical tests were performed, and results with a significance level of P < 0.05 were considered statistically significant. Correspondingly exact 95% CI were calculated. S-PLUS 2000 software (Insightful Corp., Seattle, WA) was used for model construction and cross-validation. StatXact 5.0.3 software (CYTEL Software Corp., Cambridge, MA) was used for exact statistics.


Patients, diagnoses and episodes of fever in neutropenia.

Between January 1993 and December 2001, 366 episodes of fever and neutropenia occurred in 133 patients. Data on 1 patient (0.8%) with 2 (0.5%) episodes were missing. The present study is thus based on the remaining 364 episodes of fever and neutropenia in 132 patients, with a median of 2 episodes (interquartile range, 1 to 4; maximum, 12) per patient. Patients were hospitalized in 361 (99.2%), treated with parenteral antimicrobial therapy in 356 (98%) and discharged early, i.e. within 3 days, in 132 (36%) episodes. Table 1 compares the characteristics of the first vs. last episodes of the children with more than one episode of fever and neutropenia.

Characteristics of 85 children with more than one episode of fever and neutropenia in first vs. last episodes


With a mean of 5.5 (median, 4; interquartile range, 2 to 6; maximum, 48) blood culture bottles obtained per episode of fever and neutropenia, bacteremia was detected in 87 of the 364 episodes (24%; 95% CI 20 to 29%). Gram-negative bacteremia was detected in 30 episodes (8%; 95% CI 6 to 11%). Table 2 lists the bacterial species isolated before or/and during empiric broad spectrum antibiotic therapy. In 14 episodes blood cultures taken both before and after antibiotic therapy were positive. In 18 episodes only cultures taken after start of antimicrobial therapy were positive. Two (0.5%) patients died because of infection in a nonpalliative treatment situation. One child died in irreversible shock caused by alpha-hemolytic streptococcal sepsis within 24 h of presentation. In an additional patient fungal sepsis was suspected, but no pathogen was identified. Of the 132 first episodes per patient, bacteremia was detected in 30 (23%), and Gram-negative was found bacteremia in 5 (4%).

Bacterial isolates causing bacteremia in children with fever and neutropenia

Prediction of bacteremia.

Of 44 variables 11 were significantly associated with bacteremia when performing univariate analyses in all episodes of fever and neutropenia. When analyzing only the first episodes per patient, there were 3 variables significantly associated with bacteremia (Table 3).

Variables significantly associated with bacteremia by univariate analyses

There were 348 (96%) episodes without any missing values in the variables used for construction of the prediction model. Bacteremia was detected in 85 (24%) of these. The final decision tree model was based on four variables, see Figure 1 for the model and its non-cross-validated results of prediction. It reached a cross-validated specificity of 37% (95% CI 31 to 42%) and a negative predictive value of 96% (95% CI 91 to 99%) at the predefined sensitivity of ≥95%.

Fig. 1:
Decision tree model to predict low risk episodes for bacteremia. *, comorbidity requiring hospitalization independently from fever and neutropenia at presentation; G/l, × 109/l.

When constructing a decision tree based on only the first episodes per patient, the two variables “bone marrow involvement of malignancy” and “age at presentation with fever and neutropenia” contributed independently and in a statistically significant way to the model. The predictive accuracy was poor (data not shown).

Prediction of Gram-negative bacteremia.

Of 44 variables 6 had significant associations with Gram-negative bacteremia in univariate analyses when analyzing all episodes: (1) at least 3 past episodes of fever and neutropenia (OR 2.9; 95% CI 1.3 to 6.7); (2) at least 2 past episodes of fever and neutropenia (OR 2.9, 95% CI 1.2 to 6.7); (3) presence of central venous catheter (OR 4.4; 95% CI 1.7 to 13); (4) absence of clinical evidence of viral upper respiratory tract infection (OR 3.6; 95% CI 1.1 to 19); (5) leukocyte count ≤0.5 × 109/l (OR 5.2; 95% CI 1.9 to 18); and (6) a priori impossibility of outpatient management (OR 2.3; 95% CI 1.0 to 5.5).

The decision tree model was based on four variables (Fig. 2. In 2 groups a total of 239 (67%) of the 356 episodes were classified as low risk. In only 7 (3%) of these episodes, Gram-negative bacteremia was detected, whereas all others had a risk of at least 14%. These data are not cross-validated. Evaluating the decision tree predictive preformance of the model by cross-validation, it correctly classified 29 of 30 episodes with Gram-negative bacteremia (sensitivity, 97%; 95% CI 84 to 100%) and 139 of 326 episodes without Gram-negative bacteremia (specificity, 43%; 95% CI 37 to 48%) at the predefined level of sensitivity of ≥95%. Predicting 140 (39%) of all episodes as low risk, the negative predictive value was 99% (95% CI 96 to 100%).

Fig. 2:
Decision tree model to predict low risk episodes for Gram-negative (Gram −) bacteremia. *, comorbidity requiring hospitalization independently from fever and neutropenia at presentation; G/l, × 109/l.


This retrospective cohort study indicates that bacteremia in general, and Gram-negative bacteremia in particular, can be excluded in many episodes of fever in neutropenic pediatric cancer patients with high sensitivity, based on clinical information available in the emergency situation. Our findings of 24% prevalence of bacteremia, 8% prevalence of Gram-negative bacteremia and 0.5% deaths caused by infections are consistent with most previously reported studies. 1–8, 10, 11, 13, 14

Six variables available at presentation have been identified as risk factors for bacteremia in previous prospective 2, 3, 11 or retrospective 1, 4, 8 studies in children: (1) low absolute monocyte count 3, 8, 11; (2) high maximum admission temperature 1–3, 11; (3) chills 4; (4) hypotension or shock or requirement for fluid resuscitation 1, 4; (5) prolonged neutropenia before admission 1; and (6) diagnosis of leukemia or lymphoma as opposed to solid tumors. 4 The first three variables, or factors similar to them, were significantly associated with bacteremia in our retrospective study by univariate and multivariate analyses. Variable 1 was partially matched as a low leukocyte count in our study, because no differentiation of leukocytes was performed in leukopenia at or below 0.5 × 109/l. The well-known association of long estimated duration of neutropenia after presentation with fever and neutropenia 14 is reflected in our study by bone marrow involvement by malignancy, induction chemotherapy or more intensive chemotherapy than maintenance therapy in pre-B cell acute lymphoblastic leukemia (ALL). Comorbidity requiring hospitalization independently from fever and neutropenia is a known risk factor for complications in fever and neutropenia in adults 25 and children. 14 Besides confirming these factors, our study found two additional factors available in the emergency situation to be significantly linked to bacteremia. These are absence of a clinically (focus of viral upper respiratory tract infection, significant bacterial infection) or radiologically (pneumonia) evident source of infection and of previous past episodes of fever and neutropenia without or with severe bacterial infection and/or bacteremia.

Aiming at a reasonable sensitivity of ≥95%, clinically important proportions of 29 and 67% of the episodes were classified as low risk, respectively. For all episodes in which a low risk of bacteremia was predicted, a low risk for Gram-negative bacteremia was predicted as well. The performance of the tree model predicting bacteremia based on all episodes compares well with published prospective 1, 3, 11, 14 and retrospective 8 studies (Table 4).

Performance of several published models to predict bacteremia

This study has several limitations. The retrospective study design has inherent drawbacks compared with prospective studies: Some data were missing, e.g. the systematic differentiation of leukocytes. This is relevant for the construction of multivariate prediction models. The number of missing cases was very small, however. The problem of retrospective studies of overestimating the predictive performance 20 has been minimized by an extended version of cross-validation. Repeated episodes of fever and neutropenia per patient were used for the principal analyses. The median number of episodes per child was small. The differences between the first and the last episodes per patient reflect the natural course of the disease and its treatment (Table 1). Comparing the univariate analyses of bacteremia in only the first episodes per patient vs. all episodes demonstrated the reduced power to detect associations resulting from the lower number of cases, but not a fundamental bias (Table 3). The decision tree based on the first episodes of fever and neutropenia per patient was not useful because of the small number of variables included and of bacteremias detected in these episodes. Analyses of the first episodes as to Gram-negative bacteremia were not performed because of the small number of positive cases. In sum the bias introduced by the analysis of multiple episodes per patient appears to be small and clinically not relevant, as it has been shown in a similar study. 7 Instead information from past episodes of fever and neutropenia was useful for the prediction of bacteremia (Table 3) and Gram-negative bacteremia (Fig. 2). The predefined level of sensitivity of ≥95%, reflecting the important difference in the effects of the two kinds of false classifications, is arbitrary. Direct comparison with most other studies is hampered by the temperature limits for defining fever, which are 0.5°C higher than in most other studies, but have proved to be safe at our institution for more than a decade, i.e. were not associated with an increased rate of children presenting with severe illness or of children dying from fever and neutropenia. Because of the relatively low incidence of bacteremia and especially Gram-negative bacteremia, we did not apply the highly conservative Bonferroni correction in the exploratory univariate analyses despite multiple comparisons performed. 26

Despite the limitations discussed, the major study findings are robust in the context of previous studies with similar questions, and both study aims were reached. First a decision tree model predicting the risk of bacteremia in episodes of fever and neutropenia, combining a small number of variables available in the emergency situation has been constructed. The specificity at a sensitivity of 95% or greater is within or above the range of previously published studies, which allows safe attribution of a low risk of bacteremia to a clinically important proportion of patients. Second the basic knowledge on this problem has been extended by identifying two new factors associated with bacteremia, namely the absence of a clinically or radiologically evident source of infection, and an increased number of previous episodes of fever and neutropenia. These findings merit inclusion into future study designs.

The results of this study need confirmation and refinement in prospective studies, extending the pilot study by Aquino et al. 15 An international scoring system for prediction of the risk of bacteremia or other severe bacterial infections in fever and neutropenia, comparable with the risk index established in adult cancer patients, 27 would be an important step toward broad application in clinical practice. Once established, such models might lead to outpatient management of selected low risk episodes of fever and neutropenia in children and adolescents, reducing hospitalization rate, the risk of nosocomial infections, side effects of antimicrobial therapy, costs and increasing patients’ and parents’ satisfaction.


1. Hann I, Viscoli C, Paesmans M, Gaya H, Glauser M. A comparison of outcome from febrile neutropenic episodes in children compared with adults: Results from four EORTC studies, International Antimicrobial Therapy Cooperative Group (IATCG) of the European Organization for Research and Treatment of Cancer (EORTC). Br J Haematol 1997; 99: 580–8.
2. Pizzo PA, Robichaud KJ, Wesley R, Commers JR. Fever in the pediatric and young adult patient with cancer: a prospective study of 1001 episodes. Medicine (Baltimore) 1982; 61: 153–65.
3. Rackoff WR, Gonin R, Robinson C, Kreissman SG, Breitfeld PB. Predicting the risk of bacteremia in children with fever and neutropenia. J Clin Oncol 1996; 14: 919–24.
4. Lucas KG, Brown AE, Armstrong D, Chapman D, Heller G. The identification of febrile, neutropenic children with neoplastic disease at low risk for bacteremia and complications of sepsis. Cancer 1996; 77: 791–8.
5. Chiu HH, Huang LM, Lee PI, Lee CY. Bacteremia and fungemia in hematological and oncological children with neutropenic fever: two-year study in a medical center. J Microbiol Immunol Infect 1998; 31: 101–6.
6. Klaassen RJ, Goodmann TR, Pham B, Doyle JJ. “Low-risk” prediction rule for pediatric oncology patients presenting with fever and neutropenia. J Clin Oncol 2000; 18: 1012–19.
7. Santolaya ME, Alvarez AM, Becker A, et al. Prospective, multicenter evaluation of risk factors associated with invasive bacterial infection in children with cancer, neutropenia, and fever. J Clin Oncol 2001; 19: 3415–21.
8. Baorto EP, Aquino VM, Mullen CA, Buchanan GR, DeBaun MR. Clinical parameters associated with low bacteremia risk in 1100 pediatric oncology patients with fever and neutropenia. Cancer 2001; 92: 909–13.
9. Alexander SW, Walsh TJ, Freifeld AG, Pizzo PA. Infectious complications in pediatric cancer patients. In: Pizzo PA, Poplack DG, eds. Principles and practice of pediatric oncology. 4th ed. Philadelphia: Lippincott Williams & Wilkins, 2002: 1239–83.
10. Ariffin H, Navaratnam P, Lin HP. Surveillance study of bacteraemic episodes in febrile neutropenic children. Int J Clin Pract 2002; 56: 237–40.
11. Madsen K, Rosenman M, Hui S, Breitfeld PP. Value of electronic data for model validation and refinement: bacteremia risk in children with fever and neutropenia. J Pediatr Hematol Oncol 2002; 24: 256–62.
12. Santolaya ME, Alvarez AM, Aviles CL, et al. Prospective evaluation of a model of prediction of invasive bacterial infection risk among children with cancer, fever, and neutropenia. Clin Infect Dis 2002; 35: 678–83.
13. Aquino VM, Pappo A, Buchanan GR, Tkaczewski I, Mustafa MM. The changing epidemiology of bacteremia in neutropenic children with cancer. Pediatr Infect Dis J 1995; 14: 140–3.
14. Alexander SW, Wade KC, Hibberd PL, Parsons SK. Evaluation of risk prediction criteria for episodes of febrile neutropenia in children with cancer. J Pediatr Hematol Oncol 2002; 24: 38–42.
15. Aquino VM, Herrera L, Sandler ES, Buchanan GR. Feasibility of oral ciprofloxacin for the outpatient management of febrile neutropenia in selected children with cancer. Cancer 2000; 88: 1710–14.
16. Charnas R, Ridolfi Lüthy A, Ruch W. Once daily ceftriaxone plus amikacin vs. three times daily ceftazidime plus amikacin for treatment of febrile neutropenic children with cancer. Pediatr Infect Dis J 1997; 16: 346–53.
17. Ammann RA, Hirt A, Ridolfi Lüthy A, Aebi C. Identification of children presenting with fever in chemotherapy-induced neutropenia at low risk for severe bacterial infection. Med Pediatr Oncol 2003; 41: 436–43.
18. Adamkiewicz TV, Lorenzana A, Doyle J, Richardson S. Peripheral vs. central blood cultures in patients admitted to a pediatric oncology ward. Pediatr Infect Dis J 1999; 18: 556–8.
19. Breiman L, Friedman JH, Olshen RA, Stone CJ. Classification and regression trees. Monterey, CA: Wadsworth and Brooks/Cole, 1984.
20. Zhang H, Singer B. Recursive partitioning in the health sciences. Berlin: Springer, 1999.
21. Stone M. Cross-validation choice and assessment of statistical predictions. J Royal Statist Soc B 1974; 36: 111–47.
22. Breiman L. Bagging predictors. Machine Learning 1996; 24: 123–40.
23. Efron B, Tibshirani RJ. An introduction to the bootstrap. Boca Raton: Chapman, Hall, 1993.
24. Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology 1982; 143: 29–36.
25. Talcott JA, Siegel RD, Finberg R, Goldman L. Risk assessment in cancer patients with fever and neutropenia: a prospective, two-center validation of a prediction rule. J Clin Oncol 1992; 10: 316–22.
26. Altman DG. Practical statistics for medical research. London: Chapman and Hall, 1991.
27. Klastersky J, Paesmans M, Rubenstein EB, et al. The Multinational Association for Supportive Care in Cancer risk index: a multinational scoring system for identifying low-risk febrile neutropenic cancer patients. J Clin Oncol 2000; 18: 3038–51.

Children and adolescents; chemotherapy; fever and neutropenia; bacteremia; Gram-negative bacteremia; prediction

© 2004 Lippincott Williams & Wilkins, Inc.