Intended for healthcare professionals


Meta-analysis of prophylactic or empirical antifungal treatment versus placebo or no treatment in patients with cancer complicated by neutropenia

BMJ 1997; 314 doi: (Published 26 April 1997) Cite this as: BMJ 1997;314:1238
  1. Peter C Gøtzsche, directora,
  2. Helle Krogh Johansen, senior researchera
  1. a The Nordic Cochrane Centre, Rigshospitalet, Department 7112, DK-2200 Copenhagen N, Denmark
  1. Correspondence to: Dr Gøtzsche gotzsche{at}
  • Accepted 24 January 1997


Objective:To determine whether antifungal agents given prophylactically or empirically decrease morbidity and mortality in patients with cancer complicated by neutropenia.

Design:Meta-analysis of randomised trials of amphotericin B, various lipid soluble formulations of amphotericin B (for example, AmBisome), fluconazole, ketoconazole, miconazole, or itraconazole compared with placebo or no treatment.

Setting:Trials conducted anywhere in the world.

Subjects:Patients with cancer complicated by neutropenia.

Main outcome measures:Mortality, invasive fungal infection (defined as positive blood culture, oesophageal candidiasis, or lung or deep tissue infection), and colonisation.

Results:24 trials with 2758 randomised patients were reviewed; the total number of deaths was 434. Prophylactic or empirical treatment with antifungals as a group had no effect on mortality (odds ratio 0.92; 95% confidence interval 0.74 to 1.14). Amphotericin B decreased mortality significantly (0.58; 0.37 to 0.93) but the studies were small and the difference in number of deaths was only 15. Antifungal treatment decreased the incidence of invasive fungal infection (0.47; 0.35 to 0.64) and fungal colonisation (0.45; 0.30 to 0.69). For every 73 patients treated (95% confidence interval 48 to 158) one case of fungal invasion was prevented in surviving patients.

Conclusions:There seems to be no survival benefit of antifungal agents given prophylactically or empirically to patients with cancer complicated by neutropenia. These agents should be restricted to patients with proved infection and those in randomised trials. A large, definitive placebo controlled trial of amphotericin B is needed.

Key messages

  • Prophylactic or empirical treatment with antifungal agents is often recommended for patients with cancer complicated by neutropenia for fear that they might die if treatment is delayed

  • A meta-analysis has failed to show a convincing survival benefit of antifungal agents in patients with cancer complicated by neutropenia

  • Amphotericin B and fluconazole decrease the incidence of invasive fungal infections

  • Seventy three patients need to be treated to prevent one case of fungal invasion in surviving patients

  • The use of antifungal agents should be restricted to patients with proved infections


Bacterial infections are a major cause of death in patients with cancer,1 particularly when they have neutropenia.2 Disseminated fungal infection, most often caused by Candida and Aspergillus species,3 is also considered to be an important cause of morbidity and mortality.4 5 The death rate in patients with candida sepsis or deep tissue infection is around 75%,4 5 and a positive blood culture or histological evidence of invasion was found before death in 37% of patients in one series.2

Antifungal agents are often given prophylactically in conjunction with chemotherapy or bone marrow transplantation or empirically to patients without documented fungal infection but with persistent fever despite antibiotics. The rationale is to start treatment before it is too late—that is, before death is inevitable—as it is difficult to diagnose an invasive fungal infection with certainty.4 6 Studies with historical controls have shown a positive effect of antifungal agents on mortality,3 7 but such non-randomised comparisons substantially overestimate the effect of cancer treatments.8 Randomised studies have yielded varying results, and the power of most of them has been very low. We conducted a meta-analysis of trials in which a commonly used antifungal agent was compared with non-treatment in a control group.


The primary aim was to determine whether commonly used antifungal agents decrease mortality. Secondary variables were the effect on invasive fungal infection and colonisation.

Search strategy

All randomised trials, irrespective of language, of amphotericin B, various lipid soluble formulations of amphotericin B (for example, AmBisome), fluconazole, ketoconazole, miconazole, or itraconazole compared with placebo or no treatment in patients with cancer complicated by neutropenia were eligible. We excluded studies which solely concerned treatment or prevention of oral candidiasis. A Medline search from 1966 on SilverPlatter was developed iteratively. The final strategy could retrieve all relevant trials from whatever source if they were registered in Medline. The search was most recently updated in February 1996.

One or more of random*, control*, blind*, clinical-trial in pt, clinical-trials/all subheadings, placebo*, and tg=comparative-study were combined with one or more of amphotericin, AmBisome, fluconazol*, itraconazol*, ketoconazol*, and miconazol* and with one or more of bone?marrow*, transplant*, cancer*, fungemia, h?ematologic*, malignanc*, neoplas*, neutropeni*, granulocytopeni*, leuk?emi*, lymphom*, sepsis, septic*, intensive-care/all subheadings, intensive care, and immunodeficiency. Information about trials not registered in Medline, including unpublished trials, were located by contacting the pharmaceutical industry and the authors and by scanning reference lists of articles and reviews. We also scanned selected conference proceedings—namely, the interscience conference of antimicrobial agents and chemotherapy, 1990-5; the general meeting of the American Society for Microbiology, 1990-5; and the seventh European congress of clinical microbiology and infectious diseases, 1995.

Data extraction

Decisions on which trials to include and which variables to use when more options were available for the same outcome were based on the methods sections of the trials only. Details on diagnosis, drug, dose, rules for use of additional (rescue) antifungal agents, average length of treatment with placebo, length of follow up, randomisation9 and blinding methods, number of randomised patients, number of patients excluded from analysis, deaths, invasive fungal infections, colonisation, and use of rescue drugs were extracted by each of us independently. Disagreements were resolved by discussion.

We defined invasive fungal infection as a positive blood culture, oesophageal candidiasis, lung infection, or microscopically confirmed deep tissue infection. We excluded cases of oropharyngeal and vulvovaginal candidiasis, skin infections, Candida in the urine, and vaguely described infections. To check the robustness of our findings we also analysed fungal infection according to the authors' own definitions.

Authors were asked to confirm the extracted information and answer additional questions. For six trials27 30 31 35 36 39 we obtained additional outcome data. Numbers in the tables therefore differ from those in the published articles. To increase the response rate we used Medline to obtain authors' most recent addresses. In an attempt to increase the power of the meta-analysis and avoid reporting bias we specifically asked authors for three months' mortality data for all randomised patients, including those subjected to secondary exclusion. We also sought details on the randomisation process, especially whether randomisation was concealed and irreversible so that an allocation could not be known beforehand or changed later. This question, however, seemed not always to have been well understood. We considered randomisation to have been concealed when central randomisation, sealed envelopes, or a code provided by a pharmacy or a company was described. On one occasion what seemed to have been sound randomisation provided by a pharmacy proved on further questioning to be medicine packages labelled A and B, which we would not have expected in a trial published in 1993.39 With such a procedure, should the code be broken for just one patient it would be possible to predict all future allocations.


Outcomes were weighted by inverse variance. As the studies were expected to be heterogeneous because of the various designs, diagnoses, drugs, doses, routes of administration, and criteria for fungal invasion, colonisation, and use of rescue drugs a random effects model was used.10 When the P value for heterogeneity exceeded 0.10, however, a fixed effects analysis was preferred.11 We used the Meta-analyst 0.975 program.12 Odds ratios were calculated with 95% confidence intervals.


We identified 31 trials,13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 of which 29 were reported in English and two in Japanese. Seven trials were excluded: one was not truly randomised41; in another only 14 of 146 patients had neutropenia and only data on oropharyngeal candidiasis were provided42; a third trial was concerned only with a subgroup of 72 of 298 randomised patients who came to necropsy13; and four trials were unpublished29 35 38 (H Brincker, personal communication). Of the 24 trials reviewed, two were published only as abstracts28 29 and one was published as an interim analysis.22

An antifungal agent was given prophylactically in 21 trials and empirically in three (table 1). Acute leukaemia was the most common indication in 14 trials and bone marrow transplantation in 10. The total number of randomised patients was 2758 (table 2). Duration of follow up was given in only 13 trials (54%; median 56 days). Probably it varied for different patients within the same study, as several authors stated explicitly that the trial drugs had been given till the neutropenia had resolved.15 16 17 18 19 20 21 24 25 30 35 37 39 The average number of days on placebo was reported in 16 trials (67%), in which the median was 20 days. Use of rescue antifungal treatment was more common in the untreated groups (odds ratio 0.68; 95% confidence interval 0.52 to 0.90).

Table 1

Details of studies reviewed

View this table:
Table 2

Numbers of randomised patients and exclusions, numbers of patients who received rescue antifungal treatment, durations of follow up, and durations of treatment in controls

View this table:

A total of 434 deaths were reported, corresponding to a mortality of 18% when studies without mortality figures were excluded. Prophylactic or empirical treatment with antifungals had no effect on mortality (table 3); 210 patients given antifungal treatment died compared with 224 controls (odds ratio 0.92; 95% confidence interval 0.74 to 1.14 (P=0.44)). There was no heterogeneity between the trials (P=0.60) and no tendency for newer drugs or regimens to be more effective than older ones. The odds of dying were lowest with the oldest drug, amphotericin B, which had a significant effect on mortality (odds ratio 0.58; 0.37 to 0.93) (table 3). As expected, studies with a concealed randomisation method gave a more conservative estimate of the treatment effect than other studies9 (odds ratio 1.01 (0.78 to 1.31) v 0.75 (0.51 to 1.09)). The odds ratios were similar for patients with acute leukaemia and those who had bone marrow transplantation (0.91 (0.67 to 1.24) v 0.93 (0.69 to 1.25)). Odds ratios were also independent of sample size and the design of the study—that is, the odds ratio was 0.67 (0.33 to 1.38) for studies based on empirical treatment, but the number of deaths (15 v 20) was too small for meaningful comparison with the studies of prophylactic treatment.

Table 3

Numbers of deaths and numbers of patients with invasive fungal infections and colonisation

View this table:

Antifungal treatment decreased the incidence of invasive fungal infection significantly (P<0.00001), both by our definition (odds ratio 0.47; 95% confidence interval 0.35 to 0.64) and by the wider definitions used by the primary authors (0.43; 0.32 to 0.56). There was no heterogeneity between the trials (P=0.16). However, the use of ketoconazole was associated with an increase rather than a decrease in the incidence of infections (1.36; 0.67 to 2.78) (table 3).

Authors' definitions of fungal colonisation and their methods varied widely and there was considerable heterogeneity between the trials for the effect of the drugs (P<0.00001). However, the overall effect of treatment on fungal colonisation was very similar to the effect on invasive infection (odds ratio 0.45; 0.30 to 0.69 (P=0.0002)) (table 3).


This meta-analysis failed to show an effect of prophylactic or empirical treatment with azoles on mortality, though there was a significant effect with amphotericin. The difference between treatment with amphotericin and no treatment was only 15 deaths, however, of which 12 occurred in two studies using low dose amphotericin (0.1 mg/kg intravenously/24 h).15 16 Further, unreported interim analysis was used in one study (see below),16 and small studies often exaggerate the treatment effect.8 Unfortunately, the largest study of amphotericin has not been reported apart from the subset of patients who came to necropsy.13 It is therefore difficult to judge whether the effect seen with amphotericin was real or a result of bias.

It may be very difficult to define the cause of death in patients severely ill with cancer. Therefore, we did not use death attributed to fungal infection as an outcome measure, as this information may be unreliable and prone to bias. Two studies reported only infection related deaths,30 33 but we succeeded in obtaining data on total mortality for one of them.30 The advantage of using total mortality as the outcome measure is not only that it is unbiased; it may also be the most relevant, as the drugs could have important adverse effects leading to drug related mortality. Ketoconazole, for example, is immunosuppressive, and in all three trials in which bacterial infections were reported these were more common with ketoconazole than with placebo (37 v 21,30 33 v 24,33 and 20% v 15% of neutropenic courses34). Fluconazole was associated with an excess of graft versus host disease or organ failure, or both, in the two large studies of bone marrow transplant recipients—namely, 29 v 16 deaths23 (or 44 v 24 from later correspondence43) and 102 v 85 cases of graft versus host disease.26

A positive effect on mortality may be overlooked if the trial drug is given for too short a period to allow the granulocyte count to rise, as the effect might be less during granulocytopenia. However, the odds ratio was 0.86 (95% confidence interval 0.63 to 1.17) in trials in which treatment was continued till the neutrophil count recovered, which was similar to the effect seen in other trials.

It may also be difficult to show an effect if rescue antifungal treatment is instituted too quickly in controls. Three large studies of fluconazole 400 mg daily in which roughly half of all deaths in the meta-analysis were recorded illustrate this possibility. In two studies the trial drugs were continued until the neutrophil count had increased to more than 1.0 x 109/l and had remained at that level for seven days or until systemic fungal infection was suspected or proved.23 24 In the third study trial drugs were given for 75 days or until systemic fungal infection was proved26; if fungal infection was only suspected amphotericin B was added empirically to the trial drugs. This difference in design led to placebo being given for an average of only 20 and 14 days, respectively, in the first two studies whereas it was given for 55 days in the third study (table 2).

Only the third study found a significant difference in mortality (31 v 52 deaths after 110 days). However, a biased decision on length of follow up may have been taken. In a conference abstract describing all 301 patients and the same mean time on the study drug as in the final paper only 14 versus 25 deaths were mentioned.44 The abstract notes a follow up period of 75 days (maximum length of treatment) plus an additional two weeks. The final paper gave no explanation why the follow up period was extended to 110 days. We used the data after 89 days (21 v 28 deaths), which came closest to the three month follow up that we aimed at in the meta-analysis and which we assume were also those stipulated in the trial protocol for the study. The primary author did not respond to our request for further information.

Stopping rules and ethics

Another bias which may have occurred relates to informal interim analyses. Concern has been raised about bias in cancer trials caused by loose stopping rules,45 and a recent survey showed that most cancer cooperative groups perform annual interim analyses of their trials without formal stopping rules at all.46 One author informed us that one of his studies was stopped prematurely after 30 patients when an interim analysis showed a significant effect, but the trial report did not mention this or that the study was planned to include 50 patients.36 The final report of another study did not mention any interim analysis,16 though an interim analysis was reported in a conference abstract.47 A third study described interim analyses but gave no rules,33 and a fourth study was published only as an interim analysis.22

Loose stopping rules are an important source of bias only in small meta-analyses48 49—for example, the subgroup analysis of mortality with amphotericin mentioned above. The possibility of publication bias,50 which is also a well documented phenomenon in cancer trials,8 51 is of greater concern. As an example, a study of fluconazole was stopped by the manufacturer of the drug in 1990 when 32 patients had been entered; the investigator never learnt the reason or the results (H Brincker, personal communication). Our contacts with the pharmaceutical companies were not successful. Pfizer and Janssen-Cilag refused access to unpublished reports or even just to list them so that we could approach the investigators. This behaviour, which was recently repeated when other meta-analysts approached Janssen,52 must be changed.53 Clinical trial data can be assembled only through patients' willingness to contribute to science for the benefit of future patients. These data should therefore be regarded as public property to be used for the public good. Pharmaceutical companies must therefore not obstruct researchers in their attempts to contact each other or try to prevent the medical community, patient organisations, or society from obtaining as unbiased information as possible on the effectiveness of treatments.


Amphotericin and fluconazole had a clear effect on invasive fungal infection. It is of concern, however, that the diagnosis of systemic infection is difficult and somewhat arbitrary and therefore open to bias. The blinding of some studies may have been less than desired because of the effect of the drugs on oral candidiasis or their side effects or because the final evaluation of the infections might have taken place after the code was broken. Only two reports mentioned that possible end points were evaluated without knowledge of the treatment assignment.24 37 However, the effect was so large and consistent that it would be unreasonable to ascribe it to bias alone.

We were not surprised that the use of rescue antifungal agents was more common in controls, as this may not only reflect a difference in the incidence of systemic fungal infection. The decision to use another drug may also have been influenced by the presence of oral candidiasis or access to culture data showing superficial colonisation. As we could not detect an unequivocal effect of prophylactic or empirical antifungal treatment on survival, we dismiss the common rationale for the widespread use of these drugs in neutropenic patients—namely, the fear that the patient might die if treatment is instituted too late.

Another rationale for prophylaxis might be to prevent invasive mycosis in sites which are difficult to treat. The number of patients who would need to be treated to prevent one case of fungal invasion may be calculated as the inverse of the risk difference.54 The overall risk difference was 0.034 (95% confidence interval 0.016 to 0.053). The number of patients who would need to be treated to prevent one case of fungal invasion is therefore 29 (19 to 63). A more interesting figure, however, is the number who would need to be treated to prevent one case among surviving patients. In those studies in which the fate of patients with fungal invasion was reported 60% of the patients died. Thus to prevent one case of fungal invasion in a surviving patient we should need to treat 73 patients (48 to 158). However, many infections described in the reports were fungaemias or oesophageal candidiasis. Therefore, the number of patients who would need to be treated to prevent one difficult to treat fungal invasion must be even higher. To elucidate this further we looked at a subset of seven trials in which both the fungal diagnosis and the ultimate fate of the affected patients were reported.17 18 19 20 21 23 31 36 37 These trials, which included 907 patients, were representative of the whole sample of trials. The odds ratio for fungal invasion by our criteria was 0.32 (0.19 to 0.52).

Table 4) shows the numbers of invasive fungal infections in patients after the exclusion of those who died. Clinically we could argue that, as fungal infection reflects the poor general condition of patients, the most seriously affected patients might be expected to die irrespective of any antifungal treatment whereas many less seriously affected patients might be expected to combat the fungus with their own immune system once the immune depression caused by cytotoxic drugs was abolished. Accordingly, in the study by Goodman et al 11 of 13 patients with tissue infection died whereas all 17 whom we defined as “invasive” cases because of a positive blood culture alone survived.23

Table 4

Invasive fungal infections in surviving patients in trials in which diagnosis and fate of affected patients were reported

View this table:

Among the remaining 17 trials, in which the fate of individual patients was not reported, three cases of cerebral mucormycosis were mentioned, in two of which the patients were receiving placebo; no mortality data were given for these patients.26 33 Thus it seems difficult to identify any type of difficult to treat patients among the 2758 randomised patients in these studies who would benefit definitively from antifungal prophylaxis. For example, oesophageal candidiasis is usually avoided by treating the clinical symptoms of oral candidiasis with fluconazole, and candidaemias may be treated when they occur. A final concern is that widespread use of antifungal agents could lead to resistance or to infection with inherently resistant species of fungi—for example, C krusei and glabrata, which may be a problem when fluconazole is used.55

Some patients have a particularly high risk of invasive infection—for example, because of prolonged neutropenia, heavy colonisation or colonisation with Aspergillus, previous invasion, concomitant bacteraemia, treatment with steroids, and graft versus host disease.56 57 58 These patients, however, were also represented in the trials we reviewed, and it has not been shown that it would be worthwhile to treat this subgroup prophylactically or empirically.

We therefore question the current widespread practice of routinely giving either prophylactic or empirical treatment with antifungal agents to patients with cancer complicated by neutropenia. We suggest that these drugs should be restricted to patients with proved infections and to patients participating in randomised trials. As a beneficial effect of amphotericin on mortality was suggested by the meta-analysis a large, definitive placebo controlled trial of this drug should be performed. Further, the effect of empirical treatment, for which the data were very sparse, could be addressed in future studies. The studies should include data on length of hospital stay and similar measures which will allow cost-benefit analyses to be performed. Our review will be published in the Cochrane Database of Systematic Reviews, where it will be updated when new trials appear.59


We thank the following investigators for additional information on their trials: Dr Ellen Benhamou, Dr Hans Brincker, Dr Masataka Fukuda, Dr Jesse L Goodman, Dr Richard M Hansen, Dr Jan Palmblad, Dr John R Perfect, Professor Philip A Pizzo, Dr Andrew T Pavia, Dr Jan Tollemar, Dr Ben E de Pauw, Professor Gérard Schaison, Professor John R Wingard, and Dr Drew J Winston. Dr Hiroto Takada translated the Japanese reports. We also thank Bristol-Myers Squibb and Janssen-Cilag for supplementary literature searches, the Cochrane Infectious Diseases Group for useful comments on the meta-analysis protocol, Dr Hans Brincker for useful discussion, and Professor Peter Skinhøj for comments on the manuscript.

Funding: Research grant from Rigshospitalet.

Conflict of interest: None.


  1. 1.
  2. 2.
  3. 3.
  4. 4.
  5. 5.
  6. 6.
  7. 7.
  8. 8.
  9. 9.
  10. 10.
  11. 11.
  12. 12.
  13. 13.
  14. 14.
  15. 15.
  16. 16.
  17. 17.
  18. 18.
  19. 19.
  20. 20.
  21. 21.
  22. 22.
  23. 23.
  24. 24.
  25. 25.
  26. 26.
  27. 27.
  28. 28.
  29. 29.
  30. 30.
  31. 31.
  32. 32.
  33. 33.
  34. 34.
  35. 35.
  36. 36.
  37. 37.
  38. 38.
  39. 39.
  40. 40.
  41. 41.
  42. 42.
  43. 43.
  44. 44.
  45. 45.
  46. 46.
  47. 47.
  48. 48.
  49. 49.
  50. 50.
  51. 51.
  52. 52.
  53. 53.
  54. 54.
  55. 55.
  56. 56.
  57. 57.
  58. 58.
  59. 59.
View Abstract