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

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 Information about trials not registered in Medline, including unpublished trials, were located by contacting the pharmaceutical industry and the authors and by scanning reference lists of articles and reviews.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.

Statistics

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.

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.

Rationale

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

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