Chemotherapy in advanced ovarian cancer: four systematic meta-analyses of individual patient data from 37 randomized trials. Advanced Ovarian Cancer Trialists' Group.

The purpose of this systematic study was to provide an up to date and reliable quantitative summary of the relative benefits of various types of chemotherapy (non-platinum vs platinum, single-agent vs combination and carboplatin vs cisplatin) in the treatment of advanced ovarian cancer. Also, to investigate whether well-defined patient subgroups benefit more or less from cisplatin- or carboplatin-based therapy. Meta-analyses were based on updated individual patient data from all available randomized controlled trials (published and unpublished), including 37 trials, 5667 patients and 4664 deaths. The results suggest that platinum-based chemotherapy is better than non-platinum therapy, show a trend in favour of platinum combinations over single-agent platinum, and suggest that cisplatin and carboplatin are equally effective. There is no good evidence that cisplatin is more or less effective than carboplatin in any particular subgroup of patients.

Health care professionals and patients alike are becoming increasingly aware of the need to make medical decisions on the basis of up-to-date, objective and unbiased research (Chalmers and Haynes, 1994). The most reliable information results from randomized controlled trials (RCTs). Unfortunately, most RCTs, including those conducted in ovarian cancer, have been too small to demonstrate moderate treatment benefits with reliability, and many results have been inconclusive or contradictory. The Advanced Ovarian Cancer Trialists Group (AOCTG) recognized that the best means of synthesizing such randomized evidence is by systematic meta-analysis. In 1988, five meta-analyses of chemotherapy in advanced ovarian cancer using updated individual patient data were initiated. The first results were published in 1991 (AOCTG, 1991). The AOCTG recognized the importance of updating these results especially for the comparison of carboplatin and cisplatin, in which the data were relatively immature. The comparison of platinum analogues was considered of such clinical importance that further new investigations were initiated to identify whether any particular type of women or tumour would benefit more from either cisplatinor carboplatin-based chemotherapy.

PATIENTS AND METHODS
Trials were eligible for inclusion provided they examined first-line chemotherapy for advanced ovarian cancer, were properly randomized and made one of the treatment comparisons described below. Trials were identified by bibliographic searches using MEDLINE and CancerLit, by hand searching relevant meeting proceedings and by consulting trial registers (AOCTG, 1991). Both published and unpublished trials were included and updated data were sought for all randomized patients. All data were checked thoroughly and the final database entries for each trial were verified by the responsible trialist or data centre.
All analyses were based on intention to treat. Survival analyses were stratified by trial, and the log-rank expected number of deaths and variance was used to calculate individual and pooled hazard ratios (HRs) using the fixed-effect model (Yusuf et al, 1985). HRs (representing the overall chance of dying for those allocated treatment as compared with control) were also calculated for prespecified subgroups of patients using similar stratified methodology. Chi-squared tests were used to test for gross statistical heterogeneity over all trials in a comparison (hetX2) and between subsets of trials (interactionX2) (Early Breast Cancer Trialists' Group, 1990). These tests are aimed primarily at detecting differences in effect size rather than direction and were chosen because qualitative differences were not anticipated. Survival curves are presented as simple (non-stratified) Kaplan-Meier curves. Improvements or detriments to absolute survival rates were calculated by applying the HR to baseline survival (Freedman, 1982); proportional hazards are assumed. Baseline survivals of 45% at 2 years and 25% at

RESULTS
In the first cycle of analyses, five treatment comparisons were made. However, the comparison of single-agent and combination non-platinum drugs (AOCTG, 1991) was not included in this update as it was likely to yield minimal additional data and was primarily of historical interest. For the remaining four comparisons, data were available for all but three trials (256 women) and most were able to provide updated survival information. Four new trials have been identified and included in the analyses.
Results are based on data from 37 RCTs, 5667 patients and 4664 deaths (compared with 33 trials, 5043 patients and 4195 deaths previously). Tables providing details of the chemotherapy regimens and doses used are available on request; some of this supplementary information has been published previously (AOCTG, 1991).
Single-agent non-platinum vs platinum-based combination chemotherapy Data were available from a total of 11 trials including 1329 patients and 1169 deaths (Bell et al, 1982;Decker et al, 1982;Sturgeon et al, 1982;Williams et al, 1985; Gynaecological Group COSA, 1986;Wilbur et al, 1987;Leonard et al, 1989;Masding et al, 1990;Wadler et al, 1996;MRC Gynaecological Cancer Working Party, unpublished data;Crowther, unpublished data). Data were not available for two trials (99 patients) (Harvey et al, 1982;De Oliveira et al, 1990). One trial (Decker et al, 1982) of 42 patients showed a conventionally significant benefit for combination chemotherapy, the remainder had wide confidence intervals (CI) and were inconclusive. The overall results are inconclusive (P = 0.23) but favour combination chemotherapy with an HR of 0.93 (95% CI 0.83-1.05), equivalent to a 7% reduction in the overall risk of death. This translates to a suggested 3% benefit in absolute survival at both 2 and 5 years, improving survival from 45% to 48% and from 25% to 28% respectively (95% CI, 7% benefit to 2% detriment). There was no gross statistical heterogeneity between trials [hetX2 (10) = 16.42, P = 0.09]. Excluding the small trial with the positive result gives an overall HR of 0.96 (P = 0.51). HR plots and survival curves are available on request.
Addition of platinum to a regimen Data were available from all nine eligible trials including one new trial, that compared a non-platinum drug regimen with the same regimen plus cisplatin ( Figure IA). A total of 1704 patients and 1428 deaths were included. The overall HR of 0.88 favours the addition of platinum and is marginally significant (P = 0.02) (Figure lA and B). The suggested 12% reduction in the risk of death translates to a 5% improvement in survival at both 2 (45-50%) and 5 (25-30%) years (95% CI 1-8% benefit). Although the best evidence of a benefit is shown in the trials with a combination control arm (HR = 0.85), there is no clear evidence that the results between the two subsets of trials differ (interaction X2 = 0.64, P = 0.42).
There is no gross statistical heterogeneity between those trials with a combination control arm (P = 0.43), but for those with single-agent control arms there is evidence of statistical heterogeneity (P = 0.02). Excluding the small positive trial (Decker et al, 1982) reduces the heterogeneity within this subset of trials [X2 (3) = 2.18, P = 0.54] and does not alter the main results materially, with an overall HR of 0.90 (P = 0.05).

Single-agent platinum vs platinum combination
Data were available from all eligible trials ( Figure 2A) which compared cisplatin and carboplatin either as single agents or each in combination with the same drugs in multidrug regimens. This included two new trials (Athanassiou et al, 1990;Skarlos et al, 1996) bringing the total number to nine and total patients and deaths to 1095 and 894 respectively. One trial (GICOG, 1992)   Trials are ordered with the oldest at the top and most recent at the bottom. The HR is given along the horizontal axis, with the vertical line drawn through unity indicating equivalence or no difference between treatments. HRs to the right of this line favour the _ single-drug regimens, whereas those to the left favour combination chemotherapy. Each individual trial is represented by a square, the centre of which denotes the HR for that trial, with horizontal bars whose extremities denote the 99% Cl and the inner tick marks the 95% Cl. The size of the square is directly proportional to the amount of 48 60 information in the trial. The black diamond gives the overall HR when the results of all trials are combined, the centre denoting the HR and the extremities the 95% Cl. showed a conventionally significant result at the 5% level, the remainder were inconclusive. Overall, the results favour the use of combination chemotherapy with a HR of 0.91 suggesting a 9% reduction in the overall risk of death, although this is inconclusive (P = 0.21) (Figure 2A and B). This is equivalent to a 3% benefit in survival at both 2 (45-48%) and 5 years (25-28%) (95% CI, 8% benefit to 2% detriment). There is no evidence of gross statistical heterogeneity between trials. There is, perhaps, some visual suggestion of a qualitative interaction, that cisplatin-based trials favour combination chemotherapy (HR = 0.86, P = 0.07), whereas carboplatin-based trials favour singledrug therapy (HR = 1.05, P = 0.21). However, the carboplatin result is based on a relatively small number of events and CIs are wide such that there is no clear evidence of a difference in effect between the results for these groups of trials (interaction X2 = 1.76, P = 0.18) If the Royal Marsden trial (Wiltshaw et al 1996) is excluded from the analysis as was done previously (AOCTG, 1991), because it compared high-dose cisplatin on its own with low-dose cisplatin plus chlorambucil, the overall HR is 0.88 (P = 0.08) and the HR for cisplatin-based trials is 0.80 (P = 0.02). Carboplatin versus cisplatin Data were available from 12 trials ( Figure 3A) trial (Gennatas et al, 1992), in total accountin, and 1745 deaths. Data from one further trial 1992) including 157 women were not availa regimens and drug doses used in these trials ar The results of individual trials are very consist evidence of statistical heterogeneity. There is n any difference between cisplatin and carbopla B) when given either as a single drug (HR = 1 combination (HR = 1.02, P cisplatin, but the confidence intervals are such that it could be including one new consistent with modest benefits of either drug. In terms of absolute o9 niews survival at both 2 and 5 years, the 95% CI is consistent with g (BePpomm eta improvements in overall survival of 3% benefit for cisplatin and ble. DBetailse of 4% benefit for carboplatin.
Lble. Details of the re given in Table 1.
Treatment effects in different subgroups tent and there is no Io good evidence of Different patient subgroups were analysed using data provided for tin ( Figure 3A and 11 of the trials included in the carboplatin/cisplatin comparison. No .01, P = 0.92) or in such analyses had been carried out previously. Figure 4A-C indi-Dn X2 = 0.003, P = cates that there is no good evidence that any group of women spec--sts a 2% benefit of ified by age, stage, performance status, residual tumour bulk, extent  cisplatin vs carboplatin-based chemotherapy of operation, histology or grade will do any better or worse when treated with either cisplatin or carboplatin. There is perhaps some suggestion that stage II tumours may benefit more from cisplatin. However, very few stage II tumours were included, the CIs are wide and it is difficult to draw any conclusions from the result.

DISCUSSION
The results for the comparison of single non-platinum drugs versus platinum-based combinations, which is undoubtedly the most clinically heterogeneous comparison, tend to favour platinum combination chemotherapy. However, the confidence limits are such that the results remain inconclusive. For the comparison of the addition of platinum to otherwise similar drug regimens, with the inclusion of one new trial and additional follow-up, the results are now marginally significant (at conventional levels) in favour of plat- inum. An absolute benefit of around 5% at 2 and 5 years is suggested. Given that there are now few patients 'at risk' for whom additional follow-up will be possible in either of these comparisons, it is unlikely that these results will change over time unless further large trials emerge. Thus, these results will probably remain the best and least biased estimates of the benefits of platinum-based therapy over non-platinum regimens (which were mostly based on alkylating agents). When interpreting these British Journal of Cancer (1998) 78 (11) results, however, it should be appreciated that many women in these trials are likely to have received platinum on relapse from which they may have derived a late benefit, and that drugs were sometimes administered at doses and schedules that would not be considered adequate today. Thus, in effect these comparisons are probably comparing a policy of immediate versus delayed platinum-based therapy. The results, therefore, suggest that the policy of giving immediate platinum-based treatment results in better overall survival than delaying such treatment until relapse.
The results for the comparison of single-agent platinum with platinum in combination are inconclusive, but for the cisplatinbased trials there is a strong trend in favour of combination chemotherapy. These results are driven largely by the GICOG multicentre Italian trial, which contributed 50% of the total information. In most of these studies, the dose of platinum used as a single agent was lower than is currently standard, and the difference could be attributable to the higher total drug dose rather than combination chemotherapy per se. For these trials, there are Advanced Ovarian Cancer Trialists ' Group 1485 reasonable numbers of patients for whom further follow-up is possible. It may, therefore, be important to update this analysis in future and to incorporate data from currently ongoing trials. If, as these results suggest, there may be a modest advantage of combination chemotherapy, then it is important to have a reliable estimate of effect with tight confidence intervals as the trade-offs involved and the subsequent choice between the two types of treatment is not necessarily straightforward. In such circumstances, precise estimates of any survival differences are essential. The comparison of cisplatin and carboplatin shows no obvious advantage of one compound over the other in terms of survival. These results appear very consistent across trials. Data were not available for one trial (Belpomme et al, 1992) whose preliminary results showed a significant prolongation of median survival for cisplatin. As far as is known, this trial, which prohibited crossover to cisplatin, has never been published in full. As in other comparisons, this meta-analysis compares treatment policy, in this case the policy of immediate cisplatin versus immediate carboplatin. The individual patient data collected for this meta-analysis show that crossover rates during the treatment period were not excessive and are comparable on each treatment arm. With the exception of two trials (Taylor et al, 1994;Edmonson et al, 1989), comprising 10.6% of the total patients, such crossover rates were less than 10%. However, it remains likely that patients may have been treated with the alternative platinum analogue on relapse if this happened outside the period of primary treatment. Thus, the comparison could, in fact, be one of immediate versus delayed treatment with the two platinum compounds. It will be important to update this analysis in future, looking at long-term survival, especially as the results are somewhat inconsistent with those found in testicular cancer, in which cisplatin has been shown to be superior to carboplatin (Bajorin et al, 1993;Horwich et al, 1994). However, the consistency of the results in the subgroup analyses lends support to the interpretation that neither drug is superior in terms of improving overall survival in advanced ovarian cancer. There was no good evidence that cisplatin was more or less effective in any particular predefined subgroup of patients and, therefore, no good grounds for selecting women on the basis of age, performance status, extent of resection, tumour stage, residual bulk, grade or histology to receive one or other treatment. The somewhat extreme HR in favour of cisplatin in stage II tumours is based on very small numbers of patients. Owing to this and the increased possibility of false-positive results because of multiplicity of subgroup analysis, this result should certainly not be regarded as anything more than hypothesis generating. It should be noted that trials included in the meta-analysis do not do not include recent and ongoing randomized trials using taxanes as a component of combination chemotherapy. Future updates will aim to include data from these trials. Implications for research Currently, much research effort is focused on paclitaxel, but it is not yet clear what should be used as the appropriate 'control' arm in these trials. These results suggest that this should be either platinum as a single agent or in combination. If the latter, this should probably be the CAP regimen which a separate meta-analysis has shown to be superior to CP (Ovarian Cancer Meta-analysis Project, 1991). However, in that meta-analysis, the doses of cisplatin and cyclophosphamide were similar in the two treatment arms and the observed difference could, therefore, have been C) Cancer Research Campaign 1998 attributable to either the addition of doxorubicin or to higher total doses of drug on the CAP arm. The full results of ICON2 (Torri et al, 1996), comparing CAP with single-agent carboplatin, are awaited with interest. When these results are available, taken together with the results presented here, the best 'standard' therapy may be identified which can be used as the baseline against which to measure current and future drug development.

Implications for practice
Just as no clinical trial can provide prescriptions of how to treat individual cases, neither can a meta-analysis. Although not conclusive, the results suggest that platinum-based chemotherapy is better than non-platinum therapy, that platinum combinations may offer improved survival over single-agent platinum and that cisplatin and carboplatin are equally effective. However, patients are not uniform in their preferences and the trade-offs between choosing more and less intensive therapy are not always straightforward.
Ultimately, the treatment chosen is to be decided by the patient and clinician and will depend on many factors including toxicity and quality of life in addition to survival estimates. However, the results of this meta-analysis provide the current most reliable estimates of the relative survival benefits of the treatments studied to be used as part of this decision-making process.