10.1245/s10434-006-9134-5
Annals of Surgical Oncology 14:348-354 (2007)
© 2007 Society of Surgical Oncology
Evaluation of Intensive Adjuvant Chemotherapy in Gastric Cancer Using Life Expectancy Compared with Log-Rank Test as a Measure of Survival Benefit
Toshikuni Nishikawa, MD and
Shunzo Maetani, MD, PhD
Department of Abdominal Surgery, Tenri Institute of Medical Research and Tenri Hospital, , Tenri, Japan
Correspondence: Address correspondence and reprint requests to: Shunzo Maetani, MD, PhD; Tenri Institute of Medical Research, 200 Mishima-cho, Tenri, 632-8552, Japan, E-mail: maetani{at}tenriyorozu-hp.or.jp
 |
ABSTRACT
|
|---|
Background: The goal of radical cancer surgery with or without adjuvant therapy is to cure disease rather than to delay death. There is concern that the survival benefit of curative treatment may not be properly appreciated by the log-rank test (LRT), which is more sensitive to treatment that delays death than to treatment that achieves cure. To confirm this concern and to evaluate the survival benefit of adjuvant chemotherapy, the data from a previous randomized controlled trial are analyzed using both traditional and new methods.
Methods: In this trial, 1410 gastric cancer patients with serosal or subserosal invasion had been classified by nodal and serosal status into four strata and randomized to receive high-dose or low-dose adjuvant regimens (mitomycin and tegafur-uracil) after gastrectomy. The two treatment groups were compared using the LRT as well as the life expectancy (LE) derived from the Boag model and the competing risk model.
Results: The LRT showed no significant difference between the two groups, whereas the LE increased significantly with high-dose chemotherapy (1.4-year gain; 95% CI = 0.1–2.8). A greater gain of 4.4 years occurred exclusively in the serosa-negative node-positive stratum, associated with a 21% increase in cure rate. The gain in LE was particularly greater in younger patients.
Conclusions: Parametric LE analysis offers more relevant information about curative treatment than LRT. It suggests that high-dose chemotherapy may achieve cure in a subset of patients, eradicating residual malignancies left behind after gastrectomy and providing greater survival benefit than expected from LRT.
Key Words: Log-rank test • Life expectancy • Gastric cancer • Adjuvant chemotherapy • Cure rate • Boag model
|
INTRODUCTION
|
|---|
The primary goal of radical cancer surgery with or without adjuvant therapy is to cure patients by eradicating disease rather than to prolong time to recurrence and death. There is a great difference in expected survival benefit between cured and noncured patients. If cured, 50-year-old patients in developed countries are expected to live approximately 30 years longer than noncured patients, remaining free of disease. There is concern to both clinicians and statisticians that such a survival benefit of curative treatment may not be properly appreciated by nonparametric methods, including the log-rank test, the most popular survival test.1–3 The reasons may be that (1) this test is more sensitive to treatment that delays recurrence than to treatment that achieves cure,1,2 and (2) the log-rank test and other rank order statistics do not measure survival by its length but by its order, so that they do not always value a treatment resulting in longer survival length higher than a treatment resulting in shorter survival length.
According to Gamel et al.,2,4,5 who performed simulation studies of adjuvant chemotherapy for breast cancer, only when the followup data were analyzed using the Boag model6 could curative chemotherapy be reliably distinguished from palliative chemotherapy. Goldman,1 also simulating bone marrow transplantation, warned against a routine thoughtless use of the log-rank test, which is inappropriate for evaluating a difference in cure rate. Although the cure rate derived from the Boag model or from the disease-specific survival curve is a good measure of disease outcome, it is not an ideal measure of patient outcome because old patients cured of disease may die from intercurrent illness earlier than palliatively treated patients. As compared with these statistics, the usefulness of life expectancy (LE) or mean survival time as a measure of survival benefit has been appreciated by a number of investigators.3,7–10 It has also been confirmed by lifelong followup of 3600 gastric cancer patients3 that the LE is predictable from 5-year followup with reasonable accuracy using the Boag model combined with the competing risk model.11 We accordingly used the data from a previous randomized clinical trial (RCT) for two purposes: (1) to compare LRT and LE and (2) to re-evaluate the survival benefit of intensive adjuvant chemotherapy in gastric cancer.
|
PATIENTS AND METHODS
|
|---|
Original Data
We used data from an RCT (T10)12 performed by the Japanese Research Foundation for Multidisciplinary Cancer Therapy. A total of 1410 patients from 178 institutions in Japan were enrolled in the trial between October 1, 1987 and September 30, 1990. This trial was to investigate the effect of high-dose adjuvant chemotherapy relative to low-dose therapy in gastric cancer patients undergoing histologically curative resection. Only those patients were eligible who, histologically, had tumor invasion either through the full thickness of the gastric wall (serosa-positive) or up to the subserosal level (serosa-negative) with or without lymph node involvement, and who had no distant metastases. The patients were stratified by serosal and nodal status into four strata, s(–)n(–), s(+)n(–), s(–)n(+), and s(+)n(+), according to the Japanese classification system13 published in 1981 in which s(+) includes ss
(invasion of the outermost area of the subserosa). Within each stratum the patients were randomized to receive a low-dose regimen or a high-dose regimen. The low-dose regimen consisted of 8 mg/m2 of intravenous mitomycin (MMC) on the day of surgery, followed by three capsules of uracil-tegafur (UFT; one capsule contains 100 mg of tegafur and 224 mg of uracil) daily for 6 months. The high-dose regimen consisted of 8 mg/m2 of intravenous MMC on the day of surgery and also at weeks 4, 10, 16, and 22, together with six capsules of UFT daily for 6 months.
The D1, D2, and D3 lymphadenectomies14 were performed on 5 patients, 569 patients, and 129 patients, respectively, in the low-dose group compared with 8 patients, 589 patients, and 110 patients in the high-dose group.
Statistical Method of Estimating Life Expectancy
When all patients sampled from a study population die during followup, the LE of the population is estimated simply as the mean of their survival times. When some of them are lost early, the LE may be estimated by creating a Kaplan-Meier survival curve and measuring the area under the curve (AUC). However, when some patients are still alive at the end of followup, as is usually the case, we need to "complete" the survival curve as described below before calculating the AUC.
For each group, the overall survival curve was first decomposed into two components: the disease-related (or disease-specific) survival curve and the disease-independent survival curve (Fig. 1
). In the first curve, only disease-related deaths were counted as events and all other deaths were censored; the converse applies to the second curve. The disease-related survival curve was then fitted by the Boag parametric model,6 by which the curve can be extrapolated outside the available followup period. As death from disease becomes rarer with increasing time, the disease-related survival curve approximates to a plateau (asymptote), which represents the Boag parameter c. The disease-independent curve was replaced by the survival curve of the general population matched for age, sex, and year of birth, using national life tables. This replacement serves to stabilize the later part of the disease-independent survival curve. The two simulated curves were then extended to 60 years and were recombined (multiplied) into a complete overall survival curve (Fig. 2), so that the LE of the group was estimated as AUC (competing risk model.3,11) The variance of LE was estimated by applying the Irwin method15 to the complete survival curve, for which the survival rate and variance were obtained by maximum likelihood estimation of the Boag parameters.
View larger version (14K):
[in this window]
[in a new window]
|
FIG. 1. Disease-related survival curve and disease-independent survival curve for each group. The disease-related curve is simulated by the Boag model, and the disease-independent curve is replaced by the life table of the age- and sex-matched contemporary population. Here the two disease-independent survival curves are simulated by a single common curve of both populations combined.
|
|
View larger version (13K):
[in this window]
[in a new window]
|
FIG. 2. Estimation of life expectancy as area under the overall survival curve (competing risk model). The shaded area represents additional life-years gained by the high-dose chemotherapy. The uppermost curve represents the survival curve for the age- and sex-matched contemporary population.
|
|
During this procedure, three clinically important Boag parameters were estimated, characterizing the survival pattern of the group. The first parameter (c) represents the fraction of patients cured (cure rate); the other two parameters characterize the survival distribution for the remaining noncured patients, i.e., the mean (m) and standard deviation (s) of log survival time in months (strictly, time to disease-related death). Note that em is the median survival time under the log-normal model.
Simulated Effects of Curative and Palliative Chemotherapies
The effects of curative and palliative types of chemotherapy on 5-year survival, hazard ratio, and LE were compared using the competing risk model. It is assumed that purely curative chemotherapy will increase only the Boag parameter c without changing the Boag parameter m, and that the converse is true for purely palliative chemotherapy. To simulate the effect of curative chemotherapy, the survival curve for the low-dose group (Fig. 2
) was changed by increasing the parameter c stepwise from its starting value of 0.519 (Table 1), while the other two parameters were kept unchanged. At each step we calculated three statistics for the group with new parameters: the LE, 5-year survival, and hazard ratio. The c value was increased until the gain in LE was about 10 years (from 12.9 to 22.9 years). To end the whole process in about 20 steps, the increment of c was adjusted to 0.02. A total of 20 groups thus created were named C1 to C20 in the order of c values. The hazard ratio16 relative to the original value was obtained assuming that a total of 700 patients were followed for 5 years without being censored. To simulate the effect of palliative chemotherapy, the same procedures were repeated except that the parameter m (instead of c) was increased stepwise from its starting value of 0.31 with an increment of 0.19. The groups thus created were named M1 to M20. For each chemotherapy, the increase in the 5-year survival was plotted against the gain in LE, and the relative hazard reduction (1 – hazard ratio) was plotted against the gain in LE.
|
RESULTS
|
|---|
Nonparametric Analyses
Five-year follow-up data were analyzed on the intention-to-treat principle. Although the high-dose group tended to exhibit higher survival than the low-dose group, comparison of the survival rate and the log-rank test (including a stratified test) at 5 years found no statistically significant difference between the two groups (Table 1). Subset analysis revealed that high-dose chemotherapy was associated with significantly better survival for the s(–)n(+) stratum (log-rank test: P = 0.04). This therapy resulted in reduced compliance with the regimen and a significant increase in grade as well as frequency of toxicity. However, there were no chemotherapy-related deaths.
Parametric Analysis
The overall comparison between the two groups showed that in contrast to the nonsignificant difference in 5-year survival, the LE of the high-dose group is significantly longer than that of the low-dose group (Table 1). The disease-related survival curve of the high-dose group continued to diverge from that of the low-dose group during the 5-year period (Fig. 1), and the overall survival curves of both groups did not converge until they approached the time axis (Fig. 2). Nevertheless, the average gain in LE from the intensive chemotherapy (the shaded areas in Fig. 2) was small compared to the total loss of life-years resulting from disease.
Subset analysis revealed a striking difference in response to intensive chemotherapy between the s(–)n(+) stratum and the other three strata (Table 2). In the s(–)n(+) stratum, intensive chemotherapy increased the LE by 4.4 years (P < 0.01) and increased the cure rate by 21.4% (P = 0.02) without significantly changing the median survival time for noncured patients. This gain in LE was more significant than the increase in 5-year survival or the reduction in hazard. In contrast, the other strata, including those with serosal invasion, showed essentially no response to the enhanced dosage.
Figure 3 illustrates the remarkable gains in LE from the high-dose regimen with decreasing age in male patients with subserosal invasion (serosa-negative disease) and nodal metastasis. For example, the expected gain in LE is nearly 10 years for young patients in their twenties, whereas it is only 2 years for 80-year-old men, reflecting the wide difference in LE between young and old people.
View larger version (16K):
[in this window]
[in a new window]
|
FIG. 3. Effect of age on years of life gained from the high-dose chemotherapy with 99% confidence interval in male patients with subserosal invasion (serosa-negative disease and lymph node metastasis. It is postulated that the disease-related survival is not affected by age.17
|
|
Figure 4 illustrates the effect of purely curative therapy and purely palliative therapy on three survival statistics: gains in LE, increase in 5-year survival, and relative hazard reduction. Regardless of the type of chemotherapy used, increases in one statistic are associated with increases in the other two statistics. However, the curative therapy produces less change in the 5-year survival or hazard ratio for the same gain in LE than the palliative therapy. In Fig. 4, the C10 group treated by curative chemotherapy shows a greater gain in LE but a smaller increase in 5-year survival than the M6–M9 groups treated by palliative chemotherapy. The same applies to the relative hazard reduction.
View larger version (9K):
[in this window]
[in a new window]
|
FIG. 4. Effect of two types of chemotherapy on improvement in 5-year survival (top) and hazard ratio (bottom) for equal gain in mean survival.
|
|
Peritoneal Relapse
The most common type of relapse was peritoneal dissemination (265), followed by hepatic metastasis (68) and others (211). The incidence of peritoneal relapse in the low-dose group compared with that in the high dose group for each stratum was 2/69 vs. 0/71 for s(–)n(–), 30/155 vs. 19/164 for s(+)n(–), 11/118 vs. 4/119 for s(–)n(+), and 100/361 vs. 99/353 for s(+)n(+). Overall comparison found no significant difference in the risk of peritoneal relapse (Table 3). However, the high-dose regimen produced a significant reduction of peritoneal relapse (from 7% to 2%) in patients with no serosal invasion, in contrast to serosa-positive patients who derived no benefit from intensive chemotherapy.
|
DISCUSSION
|
|---|
Intensive Adjuvant Chemotherapy
The efficacy of adjuvant chemotherapy in gastric cancer is still controversial. However, it is generally accepted that at best it benefits only a particular subset of patients (responders)18; it is useless or even harmful to patients who are early in the course of the disease when they are unlikely to harbor residual cancer cells. The same can also apply to patients at a very late stage in the disease, who have excessive tumor burden not susceptible to chemotherapy. This view is consistent with reports that, provided curative resection is performed, node-positive patients have a more favorable response than node-negative patients.19–22 No consensus yet exists on the optimal depth of invasion (T) for adjuvant chemotherapy. The present trial is the first to histologically focus on subserosal invasion, separating it from muscular invasion, although both belong to the same T2 category according to the UICC classification.23 Our subset analysis showed that only node-positive patients who were spared from serosal invasion responded to the high-dose regimen. If this s(–)n(+) stratum had been expanded to include patients with muscular invasion, our results might have been negative, as were the results of the two RCTs in which similar regimens were used in T1 plus T2 patients,18 and in T2 patients,24 respectively. The present survival analysis is consistent with the observation that peritoneal dissemination, which is the most common form of relapse, occurred in more than 20% of serosa-positive patients after curative resection. Although the high risk of peritoneal recurrence in serosa-positive patients cannot be reduced by increasing the dose of MMC and UFT, it significantly benefited patients with subserosal invasion. It is tempting to suppose from these results that patients remain good responders so long as cancer invasion is limited to the subserosal level, but no response is expected once the serosal surface is involved. This assumption needs to be tested in other trials using the same subset analyses. Although more recent chemotherapeutics may be efficacious in some of serosa-positive patients, it is likely that responders comprise only a small subset of gastric cancer patients. This may partly explain the controversial results of RCTs; the study populations are "diluted" with varying proportions of nonresponders, reducing to various extents the statistical power to detect the advantages of adjuvant chemotherapy.
Parametric analysis gives further insight into the mechanism of the high-dose therapy. This therapy tends to enhance the Boag parameter c (cure rate) rather than the parameter m in the s(–)n(+) stratum, so that its action appears to be curative rather than palliative, saving the lives of patients who would otherwise have died of residual malignancy. Importantly, our simulation studies as well as those of Gamel et al.2,4,5 indicate that curative chemotherapy compared with palliative chemotherapy produces relatively small improvements in 5-year survival or hazard ratio. These paradoxical results may explain why nonparametric tests did not detect the overall benefit of curative intensive chemotherapy in this trial.
Graphically, the possibility that the test treatment is curative should be contemplated when the disease-related survival curve diverges progressively from the control curve (Fig. 1).10 A similar pattern was seen when the cumulative risk of relapse (1 - disease-related survival) was compared between extended lymphadenectomy (D2) and limited lymphadenectomy (D1) by the Dutch gastric cancer group.25 The diverging pattern should be an indication for further followup even though nonparametric tests show no difference.
Advantages and Limitations of LE
An advantage of LE is that it is easily appreciated by patients and clinicians because it shows how many years of life are lost as a result of the disease and how many of the lost years are recovered by the treatment. Graphic representation of LE as the area under the curve, together with its upper and lower limits (Fig. 2), can further help patients and clinicians to assess the survival benefit of treatment. By contrast, if the treatment benefit is presented as a log-rank chi squared or P value to a patient experiencing severe but nonfatal toxicity of adjuvant chemotherapy, it will be very difficult for the patient to weigh the expected benefit of the treatment against its cost. Another advantage of LE over traditional nonparametric tests is that even within the s(–)n(+) subgroup, the survival benefit is shown to vary widely depending on the age of the patient. Information that is tailored to the individual by age is very helpful for patients of different ages in making their decision.
On the other hand, our parametric LE analysis is based on several hypotheses. It is assumed that for a group of cancer patients the LE is a more reliable measure of survival benefit than the 5-year survival or median survival, because the LE is estimated from the total survival curve whereas the latter two are derived from single points on the curve. However, estimation of the total survival curve involves errors, particularly when the followup time is shorter than the modal value of the distribution,2 the sample size is small,3 or information about the cause of death is inaccurate. Furthermore, it has yet to be determined whether patients who appear to be cured by adjuvant chemotherapy are actually free of disease or harbor dormant cancer cells capable of regrowing. Although very late recurrence is rare in gastric cancer, there is a case report that a patient with advanced gastric cancer remained disease-free until 18 years after intensive adjuvant chemotherapy when he relapsed with multiple bone metastases (the longest time to recurrence after adjuvant chemotherapy in the literature).26 Lifelong followup may be needed to confirm that the models used here have an equally good fit to survival of patients treated by adjuvant chemotherapy, including those with late recurrence.
|
ACKNOWLEDGMENTS
|
|---|
The authors are indebted to Dr. Inokuchi, former director of the Japanese Foundation for Multidisciplinary Cancer Therapy and Dr. Danno, representative of the T10 Group for providing the data.
Received for publication June 8, 2006.
Accepted for publication June 14, 2006.
|
REFERENCES
|
|---|
- Goldman AI. The cure model and time confounded risk in the analysis of survival and other timed events. J Clin Epidemiol 1991; 44:1327–1340.[CrossRef][Medline]
- Gamel JW, Vogel R. Comparison of parametric and non-parametric survival methods using simulated clinical data. Stat Med 1997; 16:1629–1643.[CrossRef][Medline]
- Maetani S, Nakajima T, Nishikawa T. Parametric mean survival time analysis in gastric cancer patients. Med Decis Making 2004; 24:131–141.[Abstract/Free Full Text]
- Gamel JW, Vogel RL, McLean IW. Assessing the impact of adjuvant therapy on cure rate for stage 2 breast carcinoma. Br J Cancer 1993; 68:115–118.[Medline]
- Gamel JW, Vogel RL. A model of long-term survival following adjuvant therapy for stage 2 breast cancer. Br J Cancer 1993; 68:1167–1170.[Medline]
- Boag JW. Maximum likelihood estimates of the proportion of patients cured by cancer therapy. J R Stat Soc 1949; B11:15–53.
- Haybittle JL. Life expectancy as a measurement of the benefit shown by clinical trials of treatment for early breast cancer. Clin Oncol 1998; 10:92–94.[CrossRef]
- Wright JC, Weinstein MC. Gains in life expectancy from medical interventions: standardizing data on outcomes. N Engl J Med 1998; 339:380–286.[Abstract/Free Full Text]
- Tan LB, Murphy R. Shifts in mortality curves: saving or extending lives?. Lancet 1999; 554:1378–1381.
- Gamel JW, Bonaddona G, Valagussa P, Edwards MJ. Refined measurement of outcome for adjuvant breast carcinoma therapy. Cancer 2003; 97:1139–1146.[CrossRef][Medline]
- Gross A, Clark VA. Survival distributions: reliability applications in the biomedical sciences. New York: John Wiley 1975,.
- Danno M, Shiroto H, Kuni Y, et al. Study on the intensity of MMC and UFT in postoperative adjuvant chemotherapy for gastric cancer: study report of JFMTC study No. 10. Gan To Kagaku Ryoho 2001; 28:195–203.[Medline]
- Kajitani T. The general rules for gastric cancer study in surgery and pathology. Jpn J Surg 1981; 11:127–145.[CrossRef][Medline]
- Working committee of the Japanese Society for Gastric Cancer. Japanese classification of gastric carcinoma, 1st ed. Tokyo: Kanehara & Co., 1995.
- Irwin JO. The standard error of an estimate of expectation of life, with special reference to expectation of tumourless life in experiments with mice. J Hygiene 1949; 47:188–189.
- Simon R. Confidence intervals for reporting results of clinical trials. Ann Int Med 1986; 105:429–435.[Abstract/Free Full Text]
- Choi JH, Chung HC, Yoo NC, et al. Gastric cancer in young patients who underwent curative resection. Comparative study with older patients. Am J Clin Oncol 1996; 19:45–48.[CrossRef][Medline]
- Nakajima T, Nashimoto A, Kitamura M, et al. Adjuvant mitomycin and fluorouracil followed by oral uracil plus tegafur in serosa-negative gastric cancer: a randomized trial. Lancet 1999; 354:273–277.[CrossRef][Medline]
- Neri B, de Leonardis V, Romano V. Adjuvant chemotherapy after gastric resection in node-positive cancer patients: a multicenter randomized study. Br J Cancer 1996; 73:549–552.[Medline]
- Tsavaris N, Tentas J, Kosmidis P, et al. A randomized trial comparing adjuvant fluorouracil, epirubicin and mitomycin with no treatment in operable gastric cancer. Chemotherapy 1996; 42:220–226.[Medline]
- Earle CC, Maroun JA. Adjuvant chemotherapy after curative resection for gastric cancer in non-Asian patients. Revisiting a meta-analysis of randomized trials. Eur J Cancer 1999; 35:1059–1064.[CrossRef][Medline]
- Bajetta E, Buzzoni R, Mariani L, et al. Adjuvant chemotherapy in gastric cancer: 5-year results of a randomized study by the Italian Trials in Medical Oncology (ITMO) Group. Ann Oncco 2002; 13:299–307.
- UICC. TNM Classification of Malignant Tumours. New York: Wiley-Liss; 1997, pp 59–62.
- Nashimoto A, Nakajima T, Furukawa H, et al. Randomized trial of adjuvant chemotherapy with mitomycin, fluorouracil, and cytosine arabinoside followed by oral fluouracil in serosa-negative gastric cancer: Japan Clinical Oncology Group 9206–1. J Clin Oncol 2003; 21:2282–2297.[Abstract/Free Full Text]
- Bonenkamp JJ, Hermans J, Sasako M, van de Velde CJH. Extended lymph-node dissection for gastric cancer. N Engl J Med 1999; 340:908–914.[Abstract/Free Full Text]
- Maetani Y, Murakami M, Kuroda Y, Kobashi Y, Maetani S. Recurrence of advanced gastric cancer presenting bone marrow carcinosis 18 years after surgery: report of a case. Tenri Med Bull 1998; 1:79–88.
TOP
ABSTRACT
INTRODUCTION
