Clinical prediction calculators and nomograms are becoming increasingly popular decision aids for use in predicting cancer risk, prevention, and therapeutic outcomes.21
There are a number of important cancer risk prediction models being used today for prostate,22–26
and other cancers.33
Clinical prediction tools are useful for individualizing therapeutic recommendations for a specific patient. Although prediction models can never substitute for evidence from prospective randomized clinical trials, these tools are useful adjuncts to clinical decision making in situations in which clinical trial data are not available, and optimal therapeutic management remains controversial.
In keeping with our findings, recent series have also suggested a survival benefit from adjuvant chemoradiotherapy, with encouraging 5-year survival rates over 30%,8–11
compared with historical reports of 10% to 30% after resection alone.34–36
Duke University reported its experience in 22 patients with resected gallbladder carcinoma treated with adjuvant therapy.8
Despite the locally advanced nature of patients' disease (86% of patients were T3/4 and/or node positive), 5-year survival was 37%. Median survival was 22.8 months, compared with 16 months in our study, which may be explained by the higher proportion of patients undergoing radical resection and lymphadenectomy in that series. Baeza et al9
reported their experience of treating 49 patients with resected gallbladder cancer with chemoradiotherapy. In this series, all patients underwent lymphadenectomy in addition to cholecystectomy, with a resultant 5-year overall survival of 52%. The Mayo Clinic10
published its experience of R0 resected gallbladder carcinomas treated with adjuvant chemoradiotherapy. As in our study, adjuvant chemoradiotherapy in this series significantly improved overall survival (hazard ratio for death, 0.30; 95% CI, 0.113 to 0.69; P
= .004). Also, in a recently published Korean study11
of a series of 100 patients, those with node-positive T2 or T3 disease experienced a survival benefit from adjuvant chemoradiotherapy.
In comparing adjuvant chemotherapy alone versus adjuvant CRT, our model found that CRT outperformed chemotherapy alone for virtually all patient subsets. This finding is consistent with what others have found for hepatobiliary cancers from SEER-Medicare. In fact, Davila et al37
found that SEER-Medicare patients with pancreatic cancer who received adjuvant chemotherapy had worse outcomes than those who received surgery alone. However, it is important to note that the majority of patients in these SEER-Medicare studies received fluorouracil alone in an era before gemcitabine was widely used. The outcomes predicted by our survival model are consistent with current National Comprehensive Cancer Network 2011 guidelines (http://www.nccn.org
) for gallbladder cancer, which state that one should consider a fluoropyrimidine-based adjuvant chemotherapy or CRT regimen for all patients, except those with T1b or N0 disease.
When using observational data to model treatment effects, there will always be inherent selection bias between treated and untreated groups, because patient selection for treatment can be influenced by patient or tumor characteristics. Propensity score methods can be used to reduce the impact of this treatment selection bias.17,38–41
The propensity score is defined as the probability of receiving treatment conditional on the patient's observed baseline covariates.38,39
There are several methods in which propensity scores have been incorporated into statistical modeling, including stratification, matching, covariate adjustment, and inverse probability of treatment weighting. Austin17
compared these four methods and found that matching and inverse treatment weighting performed better than the other two methods. We chose to implement the inverse treatment weighting approach, because this method yields a final survival model, the parameters of which can be readily incorporated into an interactive Web tool.
We used the AIC to compare the relative performance of the models. The AIC is a measure of the goodness of fit of regression models that is based on the concept of entropy.20
It can be viewed as the amount of information lost when a model is used to describe a set of observations. The AIC includes a penalty for number of model parameters and thus represents the tradeoff between bias and variance. Lower AIC values indicate a better model fit, and in our analysis, the LN model had the lowest AIC.
The LN survival is an accelerated failure time parametric survival model that has a long history of usage in cancer survival.42
Although not as popular as the semiparametric CPH model, in many settings in which the proportionality assumption does not hold, the LN model has been shown to be a more appropriate survival model in, for example, breast42–45
and lung cancers.46
Gamel et al47
developed an extension to the original Boag model that allows prognostic covariates to be incorporated into the LN model. In this LN survival model, the log of survival time has a normal distribution and is a linear function of covariates. In this setting, the hazard function is not constant over time but instead rises quickly to a peak and then declines over time. We have previously demonstrated that this LN model performs well in modeling extrahepatic cholangiocarcinoma,48
and the current study indicates that an LN model also demonstrates a good fit for gallbladder cancer.
Our current findings are consistent with the overall conclusions from our original SEER-based gallbladder nomogram13
(ie, most patients with T2 or N+ gallbladder cancer or greater would be predicted to benefit from adjuvant therapy). Chemotherapy was not included in the original model, because this information is not available in SEER, but our current SEER-Medicare analysis confirms that the majority of these patients also received chemotherapy. Differences between the two nomograms in the actual predicted survival estimates are mainly the result of the incorporation of more recent data and use of improved survival modeling methods.
There are several limitations to this study. This study was performed using SEER-Medicare data and was limited to predictive factors available in this database. SEER does not include information on margins or performance status, so these prognostic factors could not be included. Patients who received both radiotherapy and chemotherapy within a 6-month time window were assumed to have received concurrent adjuvant CRT. We also examined a shorter 4-month time window and found similar results. Because SEER does not capture cancer recurrence, this approach may have also inadvertently captured patients who received therapy for an early recurrence within 6 months and those who received sequential and not concurrent therapy, and it would have missed adjuvant therapy administered after 6 months.
Perioperative mortality can bias the apparent effect of adjuvant therapy in nonrandomized observational studies. To partially compensate for this bias, we excluded all patients who died within 2 months of surgery. However, it is important to note that this type of exclusion may have subjected the results to a different type of bias resulting from conditional survival,49
in which all patients' prognoses improve when they are presumed to have already survived a period of time since treatment.
To capture the largest relevant data set, we included all patients who underwent at least a total cholecystectomy. In looking at extent of resection, several studies have established that gallbladder cancer survival outcomes are improved with radical resection and lymphadenectomy.50–54
In fact, some series have demonstrated that patients who incidentally discover T2 gallbladder cancer after simple cholecystectomy have better outcomes if they undergo reresection with radical surgery and lymphadenectomy.55
Unfortunately, the number of SEER-Medicare patients coded as having undergone these extended procedures is low (6% to 7%), which precluded our ability to incorporate these variables in our final nomogram. However, our preliminary analysis indicated that these patients generally had better survival outcomes compared with those who did not, even after adjuvant CRT, suggesting that patients with gallbladder disease should have these extended procedures performed whenever possible. Interestingly, our preliminary analysis suggests that patients who underwent extended lymphadenectomy did not derive as large a benefit from adjuvant chemotherapy or CRT. In the future, when more of these patient cases have accumulated in SEER-Medicare, we plan to incorporate radical resection and lymphadenectomy as additional covariates in the next version of our nomogram.
In some cases, the model predicted only a small-percentage improvement from the addition of adjuvant therapy, such as in certain cases of node-negative disease. We did not specify a specific threshold at which adjuvant therapy should be recommended. We believe that the final decision of whether adjuvant therapy should be administered is a decision that should be made after thoughtful discussion between clinician and patient, taking into account multiple factors, many of which cannot be accounted for in a prediction model. Quality of life and specific patient preferences are also important considerations in treatment decision making.
Recently, there has been a movement toward personalized medicine, in which specific information about an individual patient is used to optimize the patient's care. We believe that these types of predictive models will become increasingly important in the future, as we attempt to improve outcomes by individualizing therapeutic recommendations.
In summary, we have built an interactive survival prediction model that can make an individualized estimate of the net survival benefit of adjuvant therapy for patients with gallbladder cancer. This tool can assist clinicians and patients in quantifying the potential benefit of adjuvant chemotherapy or CRT after surgical resection of gallbladder cancer.