|
Advertisement |
|
|
||
 |
|
|||||
|
|||||
|
 |
||||||
|
||||||
|
||||||
|
Journal of Clinical Oncology, Vol 26, No 27 (September 20), 2008: pp. 4524-4526 © 2008 American Society of Clinical Oncology. DOI: 10.1200/JCO.2008.18.2642
In ReplyDepartment of Radiation Medicine, Oregon Health & Science University, Portland, OR
Department of Radiation Oncology and Graduate Division of Radiological Sciences, University of Texas Health Science Center at San Antonio, San Antonio, TX
Department of Mathematics and Statistics, Portland State University, Portland, OR
Department of Radiation Medicine, Oregon Health & Science University, Portland, OR We appreciate the letters from Drs Yu, Arroyo, and Cleary, and thank them for their interest in our study. The respondents make several important points worthy of additional discussion. Regarding the broad range of surgical procedures included in the study, the inclusion of "curative intent" procedures was designed to cast a wide net for inclusion within the model. One of the values of Surveillance, Epidemiology, and End Results (SEER) analyses is the statistical power associated with large patient numbers, but this naturally comes at a cost of heterogeneity of patients and treatments given. As we explained in our discussion,1 such pooling may in fact result in more realistic estimates of actual heterogeneous population-based outcomes. To date, most existing retrospective institutional series have similarly included pooled surgical procedures.2,3 Several of the respondents mention additional information that they say would be desirable to have in the model, such as margin status. We agree that having margin status in SEER would have been helpful and would likely have been a significant prognostic factor. Unfortunately, as the respondents suggest that they are well aware, margin status is not available in the SEER database. This is a well-known limitation common to every published SEER analysis, and not unique to this study. We chose to proceed with building a model using the available data in SEER and found that the model performed reasonably well. If the respondents have access to a large database of patients with gallbladder disease with these additional data, we would encourage them to build a predictive model with it so that all can benefit from its use. Otherwise, we at least hope that the attention given to this issue by the multiple letters written herein will help raise awareness of the added value if SEER would consider adding these data fields in future versions of their registry. We agree that perioperative mortality may represent a bias in the model, and we thank Yu et al for pointing out this fact, which should have been mentioned in the discussion as another limitation. Attempts to manipulate the data set to address this issue of perioperative mortality, however, may prove to be more misleading than initially realized. We have concerns with the respondents’ approach of excluding patients based on an arbitrary survival cutoff of 4 months and attempting to make meaningful comparisons with the unadjusted survival curves in Figure 1B.1 Below are some issues to consider.
To our knowledge, there is not a standard definition for the time period during which deaths would be classified as perioperative mortality. A commonly used time period is 30 days, although others have used cutoffs of up to 6 months. In the setting of a cancer with median survival of only 10 months, the choice of a 4-month cutoff point as an exclusion criterion will obviously have a substantial effect on outcomes, with more effects than merely eliminating those who experienced perioperative mortality. Indeed, when some researchers have set similar survival cutoffs for other cancer sites, other respondents expressed equal concern about that approach.4 Given that patients survive longer periods of time from diagnosis and treatment, their survival probability is not static, but rather changes continuously with time. For many cancer sites, the greatest risks are during the initial period after diagnosis and treatment, independent of the treatment modality received. Thus, comparison of patients who have already survived 4 months necessarily results in a patient population that is distinct from those followed from diagnosis. This effect is called conditional survival (CS), a concept that we and others have written about for gallbladder5 and other cancers.6-11 For example, when we computed 12-month CS for the present series for patients who had already survived 1, 2, 3, 4, 5, and 6 months after diagnosis, we found that CS changed markedly during this initial time period for both groups of patients (Fig 1). A repeating theme that we have seen in this and other cancer sites is that patients who experience the greatest early hazard also tend to see the largest subsequent gains in CS over time. This is a form of selection bias where patients who are fortunate enough to survive this initial early hazard period have an improved prognosis when compared with other groups whose risk is more distributed over time. The improved prognosis for these selected survivors occurs independent of whether the prior early hazard in their group was due to perioperative mortality, failure to receive adjuvant radiotherapy (RT), or other causes. Coincidentally, the 4-month cutoff point selected by Dr Yu also happens to correspond with the time when the 12-month CS for the no RT (group crosses that for the RT group (Fig 1). Nevertheless, the better CS for the no RT group at extended times compared with the RT group should not be confused with better prognosis from the time of diagnosis. Perhaps even more importantly, we must remember that the two groups (RT and no RT) depicted in Figure 1B1 are not at all balanced with respect to stage or other covariates—an obvious limitation of all nonrandomized retrospective data analyses. Thus, it is difficult to draw meaningful inferences from unadjusted comparisons of these unbalanced groups in aggregate. Of course, this is the reason for performing the multivariate regression analysis in the remainder of our study—to control for multiple covariates. Perhaps a more fair comparison would be to look at specific patient subsets to evaluate whether adjuvant RT may be beneficial. For example, if we examine the subset of patients with stage T3,N1 disease, even if we excluded patients not surviving at least 4 months, we found that adjuvant RT was still associated with a decreased risk of death, with a hazard ratio of 0.84 (95% CI, 0.73 to 0.96). We appreciate Yu et al pointing out that the Schoenfeld residuals suggest that the hazard ratio for RT contains a time-varying component. Since the publication of our original article, we have been exploring the use of other modeling techniques that do not rely on the proportional hazards assumption, such as random survival forests,12 and the Cox-Aalen survival model.13 Interestingly, our initial analyses suggest that using these more advanced models does not necessarily result in improved performance, suggesting that the other, non–time-varying covariates may be the predominant determinant of model performance. Others have also found that the Cox proportional hazards model often works as well as other, more advanced models.14 All three respondents, Yu, Arroyo, and Cleary, make a similar point regarding the extent of surgical resection in predicting outcomes for gallbladder cancer. This is an excellent point, as multiple prior studies have shown that extent of surgical resection has prognostic significance. We must emphasize that we never meant to suggest that adjuvant RT can ever replace optimal surgical resection, and we are puzzled as to how Cleary et al arrived at this conclusion from reading our study. We had originally omitted these data from our model because of the small numbers of patients in SEER who were coded as having had a "radical" (en bloc) resection (< 12%), or a lymphadenectomy with at least three nodes resected (< 10%). These are both lower rates than we would have expected to see for this series of patients, but it is well known that information in SEER regarding extent of surgery may be inaccurate or incomplete because of the way SEER data are collected and recorded.15 Nevertheless, to address the respondents’ concerns, we constructed a new model with the addition of these two new binary covariates (radical resection and lymphadenectomy) to investigate their effect on outcomes. Our results were consistent with Cleary's analysis above. Overall, in multivariate analysis, we found that having had a lymphadenectomy (with at least three nodes resected) was a significant independent predictor of improved performance (hazard ratio, 0.67; 95% CI, 0.58 to 0.77), but radical resection (as coded by SEER) was not (hazard ratio, 0.97; 95% CI, 0.85 to 1.11). More importantly, even after adjusting for extent of resection with these new covariates, we found that adjuvant RT was still a favorable independent prognostic factor. With the revised model, we found that some younger T2N0 patients no longer derived a survival benefit from adjuvant RT after a full lymphadenectomy, but there was still an additional survival benefit for the remaining subsets of patients noted in our original study. For example, for a 70-year-old white female with a T3N1 gallbladder cancer status postsurgical resection including a lymphadenectomy, the revised model predicts that 2-year survival would improve from 28% (95% CI, 22% to 34%) without adjuvant RT, to 44% (95% CI, 36% to 52%) with adjuvant RT. The revised prediction model with these additional covariates is available on our Web site (www.ohsu.edu/radmedicine/predict.cfm). Our overall summary findings, that adjuvant RT may be beneficial for patients with gallbladder cancer with node-positive or more than T2 disease, is consistent with the recommendations of the 2007/2008 National Comprehensive Cancer Network Clinical Practice Guidelines for adjuvant treatment of gallbladder cancer. One objective validation measure of the performance of prediction models is the C-index for discrimination. Our model's C-index of 0.71, although still leaving room for improvement, is comparable to other published cancer prediction models.14 Modeling retrospective databases will always have inherent limitations and one can never account for all biases. Statistical modeling will never serve as a replacement for prospective randomized controlled clinical trials. It is important to note that a nomogram is merely a convenient graphical representation of a multivariate regression analysis, and is no more or less valid than other multivariate regression analysis studies that may report results in other forms (eg, as hazard ratios).16 As we stated in our discussion, we believe that the final decision of whether adjuvant RT should be administered remains a decision that should be made after careful discussion between the clinician and patient, accounting for multiple factors, many of which cannot be considered in a prediction model. But until such time as a large-scale prospective randomized controlled clinical trial can be conducted for this rare cancer, providers will be left in the dark when faced with making treatment decisions for patients with this disease. We encourage the respondents (and others) to continue to build on and improve our original study idea and submit full peer-reviewed manuscripts for publication so that we can continue to raise awareness about this rare but aggressive malignancy and assist clinical providers who must make treatment recommendations for these patients on a daily basis. AUTHORS’ DISCLOSURES OF POTENTIAL CONFLICTS OF INTEREST The author(s) indicated no potential conflicts of interest.
REFERENCES
1. Wang SJ, Fuller CD, Kim JS, et al: Prediction model for estimating the survival benefit of adjuvant radiotherapy for gallbladder cancer. J Clin Oncol 26:2112-2117, 2008 2. Kresl JJ, Schild SE, Henning GT, et al: Adjuvant external beam radiation therapy with concurrent chemotherapy in the management of gallbladder carcinoma. Int J Radiat Oncol Biol Phys 52:167-175, 2002[CrossRef][Medline] 3. Czito BG, Hurwitz HI, Clough RW, et al: Adjuvant external-beam radiotherapy with concurrent chemotherapy after resection of primary gallbladder carcinoma: A 23-year experience. Int J Oncol Biol Phys 62:1030-1034, 2005 4. Hanna NH, Einhorn LH: Postoperative chemotherapy for N2 non–small-cell lung cancer: Conclusions are not black and white, but gray. J Clin Oncol 24:5611-5612, 2006 5. Fuller CD, Wang SJ, Thomas CR: Conditional survival of gallbladder adenocarcinoma treated with radiotherapy: Analysis from the SEER Database. Presented at the 47th Annual Meeting of the American Society for Therapeutic Radiology and Oncology, October 16-20, 2005, Denver, CO 6. Henson DE, Ries LA: On the estimation of survival. Semin Surg Oncol 10:2-6, 1994[Medline] 7. Fuller CD, Wang SJ, Thomas CR, et al: Conditional survival in head and neck squamous cell carcinoma: Results from the SEER Dataset 1973-1998. Cancer 109:1331-1343, 2007[CrossRef][Medline] 8. Wang SJ, Fuller CD, Thomas CR: Ethnic disparities in conditional survival of patients with non–small-cell lung cancer. J Thorac Oncol 2:180-190, 2007[CrossRef][Medline] 9. Wang SJ, Fuller CD, Emery R, et al: Conditional survival in rectal cancer: A SEER Database Analysis. Gastrointest Cancer Res 1:84-89, 2007 10. Wang SJ, Emery R, Fuller CD, et al: Conditional survival in gastric cancer: A SEER database analysis. Gastric Cancer 10:153-158, 2007[CrossRef][Medline] 11. Choi M, Fuller CD, Thomas CR Jr, et al: Conditional survival in ovarian cancer: Results from the SEER dataset 1988-2001. Gynecol Oncol 109:203-209, 2008[CrossRef][Medline] 12. Breiman L: Random forests. Machine Learning 45:5-32, 2001[CrossRef] 13. Scheike TH, Zhang MJ: Extensions and applications of the Cox-Aalen survival model. Biometrics 59:1036-1045, 2003[Medline] 14. Kattan MW: Comparison of Cox regression with other methods for determining prediction models and nomograms. J Urol 170:S6-S9, 2003[Medline] 15. Cooper GS, Yuan Z, Stange KC, et al: Agreement of Medicare claims and tumor registry data for assessment of cancer-related treatment. Med Care 38:411-421, 2000[CrossRef][Medline] 16. Coburn NG, Cleary SP, Tan JCC, et al: Surgery for gallbladder cancer: A population-based analysis. J Am Coll Surg (in press)
Related Correspondence
|
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
|
|||||||||||
|
|||||||||||
|
|
Copyright © 2008 by the American Society of Clinical Oncology, Online ISSN: 1527-7755. Print ISSN: 0732-183X
|
