Rodney Sparapani
Results A total of 1,842 patients were included; EP was used in 27% (n = 499). Treatment with EP was not associated with a survival advantage in a Cox proportional hazards model (hazard ratio HR, 0.97; 95% CI, 0.85 to 1.10), a propensity score matched cohort (HR, 1.07; 95% CI, 0.91 to 1.24), or a propensity score adjusted model (HR, 0.97; 95% CI, 0.85 to 1.10). In an instrumental variables analysis, there was no survival advantage for patients treated in centers where EP was used more than 50% of the time as compared with centers where EP was used in less than 10% of the patients (HR, 1.07; 95% CI, 0.90 to 1.26). Patients treated with EP, compared with patients treated with CP, had more hospitalizations (2.4 v 1.7 hospitalizations, respectively; P. INTRODUCTION Nearly one in four patients with non–small-cell lung cancer (NSCLC) has stage III disease at presentation. The outcome of these patients remains poor, with median survival times of only 15.3 to 21.7 months.
Patient Selection Using the VACCR, we identified patients diagnosed with stage III NSCLC (see Appendix, online only, for histologies included) between October 2001 and December 2010. The VACCR stages patients using the contemporaneous International Association for the Study of Lung Cancer/American Joint Committee on Cancer staging classification; thus, version 6 was used through 2009 and version 7 in 2010. We used the VA Decision Support System Pharmacy file to identify chemotherapy drugs received during the 120 days after diagnosis. We included patients who were classified by the VACCR as receiving radiotherapy as part of primary treatment within 7 days of the start of chemotherapy and received either CP or EP as the chemotherapy backbone. We excluded patients who the VACCR classified as receiving surgery as part of initial treatment or received other anticancer agents in addition to CP or EP as part of their initial treatment. We characterized chemotherapy as induction when it was prescribed as part of the initial 6 weeks of treatment and as consolidation when it was prescribed between weeks 7 and 16. We identified patients who received lung resection using International Classification of Diseases Ninth Revision (ICD-9) or Current Procedural Terminology procedure codes.
We augmented the VACCR database with demographic, diagnostic, and laboratory data from several other VA databases. We provide details of this process in the Appendix.
The study was approved by the institutional review boards of the Milwaukee and Durham VAs. Statistical Analysis The primary outcome was OS, defined by days from VACCR diagnosis until death. We censored OS at 5 years after diagnosis to avoid excessive statistical impact by patients with long survival diagnosed early in the study period. We grouped year of diagnosis into 2001 to 2004, 2005 to 2007, and 2008 to 2010. We used clinical records to obtain baseline values for estimated glomerular filtration rate (eGFR), serum albumin, platelet count, and hemoglobin, as well as weight loss before treatment. We defined anemia as hemoglobin less than 12 g/dL, hypoalbuminemia as serum albumin less than 3.5 g/dL, and chronic kidney disease as eGFR less than 60 mL/min/1.73 m 2.
We used ICD-9 codes from health care encounters during the year preceding diagnosis to calculate a summary burden of comorbidity using the National Cancer Institute (NCI) combined index. ICD-9 codes from encounters during the 4 months after chemotherapy were used to identify adverse effects of treatment.
Outcome Analysis We tested differences in baseline characteristics between groups using the Pearson χ 2 or t test for categorical and continuous variables, respectively. Survival curves were prepared using the Kaplan-Meier method. We used the following three complementary approaches to adjust our comparison of OS among patients receiving EP and CP for differences in baseline characteristics: a standard Cox proportional hazards model, a propensity score adjustment, and an instrumental variables technique. In our standard Cox model, we included variables that were associated with OS in univariable analysis with a significance level of P. Instrumental Variables and Near/Far Analysis Although propensity scores can address biases caused by observed variables, there may be unknown or unmeasured variables that also affect the choice of treatment and outcomes. Instrumental variable analyses address such unmeasured confounders by using an instrument that is strongly associated with the treatment but is not expected to affect outcomes.
Rodney Sparapani Greater Milwaukee Area
The variability among VA medical centers in the use of EP versus CP provided an instrument for our analysis. We used a novel technique, near/far matching, that unites propensity score matching with classical instrumental variables analysis., We identified EP-encouraging centers, in which more than 50% of patients received EP, and EP-discouraging centers, where less than 10% of the patients received EP. We matched an EP-encouraging center to one or more EP-discouraging centers based on overall hospital volume, oncology service volume, and the number of facility oncologists. Using the propensity score obtained in the prior analysis and treatment year, we matched each patient treated in an EP-encouraging center with a patient treated in a matching EP-discouraging center.
Thus, the patients were a near match (nearest match) on patient and facility characteristics and a far match (farthest match) based on the instrument. The analysis proceeds as a Cox proportional hazards model with a single predictor variable (encouraged v discouraged); matched patient pairs define the strata. Patient Characteristics From an initial sample of 17,010 patients with stage III disease, 1,842 patients fulfilled our eligibility criteria. Most of our patients (98%) were men. EP was used in 27% of the patients (n = 499) and CP in 73% (n = 1,343; ). Patients treated with EP, compared with patients treated with CP, were younger (mean age, 61.3 v 65.5 years, respectively; P. Characteristic Observational Data Set (n = 1,842) Propensity Score–Matched Data Set (n = 762) EP (n = 499) CP (n = 1,343) Standard Difference P EP (n = 381) CP (n = 381) Standard Difference P No.
Of Patients% No. Of Patients% No.
Of Patients% No. Of Patients% Age, years 0.50 2 46 9.3 230 17.2 0.24. Odds ratio plot. This plot represents characteristics associated with the use of carboplatin-paclitaxel (CP) versus etoposide-cisplatin (EP). EGFR, estimated glomerular filtration rate; NCI, National Cancer Institute. (.) Multiples of 10. Patients treated with CP received a median of five cycles (interquartile range, three to six cycles), with 46.8% of patients receiving six or seven cycles.
In the EP group, 81.8% of patients received two cycles, and 18.2% received only one cycle. Patients who received CP were more likely to receive consolidation chemotherapy than patients who received EP (67.5% v 46.1%, respectively; P =.0026). The need for consolidation chemotherapy was put in question after the results of the Hoosier Oncology Group study were presented in 2007.
In our database, before 2007, the rates of consolidation chemotherapy among patients receiving CP were similar to rates among those receiving EP. However, after 2007, the rate of consolidation in the CP group increased from 65.2% to 71.7%, whereas in the EP group, it decreased from 53.6% to 32.4%. Surgical resection after induction chemotherapy was uncommon but was more frequently used in the EP group 7% (n = 35) in contrast to only 2.4% (n = 32) in the CP group. Survival Outcomes In an unadjusted two-group analysis, patients treated with EP had a better outcome compared with those treated with CP (median OS, 17.3 v 14.6 months, respectively; HR, 0.88; 95% CI, 0.79 to 0.99; P =.0209; A). In a Cox proportional hazards model, patient age (HR, 1.08; 95% CI, 1.01 to 1.15; P =.0258), percentage of weight loss (HR, 1.04; 95% CI, 1.03 to 1.05; P. Of Patients Median Survival (months) Cox Univariable Analysis Cox Multivariable Analysis Hazard Ratio 95% CI P Hazard Ratio 95% CI P (n = 1,437) Chemotherapy regimen.0209.6327 Carboplatin and paclitaxel 1,343 14.6 — — Cisplatin and etoposide 499 17.3 0.88 0.79 to 0.98 0.97 0.85 to 1.10 Baseline hemoglobin, g/dL.
Propensity Score For the matched analysis, 381 patients treated with EP were identified and matched based on their propensity score and era of treatment with an equal number of patients receiving CP. This analysis eliminated differences in age, hemoglobin, albumin level, percentage of weight loss, and comorbidity scores seen in the larger cohort and revealed no survival advantage for EP (HR, 1.07; 95% CI, 0.91 to 1.24; P =.4264; C).
Subsequently, a Cox proportional hazards model weighted on the inverse propensity for being treated with EP was fitted. This analysis also showed no survival advantage for EP (HR, 0.97; 95% CI, 0.85 to 1.10; P =.6212). Instrumental Variables In facilities that treated more than 20 patients, there were marked differences between facilities in EP and CP use: eight facilities used EP more than 50% of the time (range, 55% to 81%), whereas 11 centers used EP in fewer than 10% of the cases (range, 0% to 9%). Among these facilities, which were all affiliated with an academic medical center, there were no significant differences in oncologist full-time employment equivalent (median, 2.7 v 2.3; P =.8043), unique patients seen annually in oncology clinics (median, 1,898 v 1,714; P =.5915), or overall facility volume (mean, 58,037 v 65,151; P =.5357). For the instrumental variable analysis, patients treated at EP-encouraging centers, compared with patients treated at EP-discouraging centers, were younger (mean age, 62.1 v 65 years, respectively; P =.001) and had higher baseline albumin (mean, 3.8 v 3.5 g/dL, respectively; P =.0032) and hemoglobin levels (mean, 12.9 v 12.6, g/dL, respectively; P =.0346; Appendix, online only).
Rodney Sparapani
Despite having patients who had better prognostic characteristics, no survival advantage was seen for patients treated in EP-encouraging centers in a univariable analysis (HR, 1.07; 95% CI, 0.90 to 1.26; P =.4653; C). Furthermore, the near/far analysis, which equilibrated prognostic variables, did not demonstrate a survival advantage for matched patients treated in EP-encouraging centers (HR, 1.03; 95% CI, 0.75 to 1.40; P =.8736; D). DISCUSSION In a large cohort of patients with stage III NSCLC who received concurrent chemoradiotherapy at academically affiliated tertiary care VA hospitals, we found that EP offered no survival advantage over CP and was associated with more toxicity. This finding was consistent across multiple analytic approaches in a database that included a wide range of potential confounders.
After the establishment of concurrent chemoradiotherapy as the standard of care, little research has addressed the choice of chemotherapy. Laboratory values.
We defined baseline hemoglobin, platelet, creatinine, and albumin levels as the mean of tests obtained within 45 days before and 7 days after initial treatment. We excluded biologically implausible values (hemoglobin 22 g/dL or 8 g/dL), which made up less than 0.1% of observations. No implausible values for creatinine or platelets were identified. Plausible hemoglobin, albumin, and platelet levels were identified in 98.2% (n = 1,822), 89.8% (n = 1,666), and 92% (n = 1,878) of eligible patients, respectively.
We determined the estimated glomerular filtration rate using the Modification of Diet in Renal Disease equation (Levey AS, et al: Ann Intern Med 145:247-254, 2006); estimated glomerular filtration rate was obtained in 97% of patients (n = 1,801). The Decision Support System pharmacy records were used to obtain chemotherapy information. We defined induction chemotherapy as that prescribed in the first 6 weeks after the initiation of treatment and consolidation as that administered after 6 weeks and until 16 weeks after initiation of treatment. During the induction phase, for patients treated with CP, all chemotherapy drugs identified every 7 days were counted as a cycle. For patients treated with EP, all chemotherapy drugs identified every 19 days were counted as a cycle. Download car customizing programs for mac free.
For the consolidation phase in both groups, cycles were identified as all drugs administered every 19 days. Comorbidities We measured comorbidities using ICD-9 diagnosis and procedure codes from inpatient and outpatient encounters occurring between 13 months and 1 month before non–small-cell lung cancer diagnosis. We applied the Deyo algorithm for generating the Charlson comorbidity index from administrative data. We then excluded cancer diagnoses and applied the lung cancer–specific weights developed by Klabunde et al (Klabunde CN, et al: J Clin Epidemiol 53:1258-1267, 2000) to obtain the National Cancer Institute combined index. We identified VA hospitalizations during the year before the start of treatment, because prior hospitalizations are known to predict future hospitalizations and death (Anderson G, et al: JAMA 263:967-972, 1990). We used the 2009 Area Resource File to identify area-level surrogates for socioeconomic status (ie, proportion of adults with high school education, median household income, urban/rural status).
Outcomes We obtained data regarding date of death from the VA's vital status file updated in February of 2014. This is created by combining mortality data from various VA databases and the Social Security Administration and provides data quality similar to that available from the National Death Index for veterans who use the VA for health care (Sohn MW, et al: Popul Health Metr 4:2, 2006). We summarized the number of VA hospitalizations and outpatient visits during the 4 months after the start of treatment. We identified complications commonly associated with chemotherapy using the ICD-9 diagnosis codes associated with these encounters. When available, previously defined coding algorithms were applied; for other complications, we developed algorithms based on appropriate coding documents. Complications were categorized as infection or neutropenia; acute kidney injury or dehydration; nausea or vomiting; hemorrhage; or mucositis/esophagitis.
We did not attempt to grade the severity of these complications, but note that they were significant enough to be coded. Complications of Treatment Hospital encounters and outpatients records were used to obtain coding information to identify complications from treatment that occurred within 4 months of the start of treatment.
Infections or neutropenic complications were defined as having an inpatient or outpatient visit with a primary or secondary diagnosis of neutropenia (ICD-9 Clinical Modification 288.0), fever (780.6), or a list of infections previously used by other authors (Weycker D, et al: BMC Health Serv Res 13:60, 2013). Acute kidney disease or dehydration was defined as having a primary or secondary diagnosis of dehydration (276.5), hypernatremia (276.0), acute kidney failure (584.0), or renal failure, unspecified (586.0). Nausea/vomiting complications were identified with diagnosis for nausea and vomiting (787.0) or persisting vomiting (536.2). Hemorrhage was defined as having ICD-9 codes for extracranial hemorrhages (ie, GI, genitourinary, retroperitoneal) or primary and secondary diagnoses of intracranial hemorrhage, including intracerebral, subarachnoid, or subdural hemorrhages, as done by Fang et al (Fang MC, et al: J Am Coll Cardiol 58:395-401, 2011). Characteristic Comparing All Patients Near/Far Match Analysis EP-Encouraging Center (n = 252) EP-Discouraging Center (n = 431) Standard Difference P EP-Encouraging Center (n = 163) EP-Discouraging Center (n = 163) Standard Difference P No. Of Patients% No. Of Patients% No.
Of Patients% No. Of Patients% Treatment 1.83 2 28 11.2 66 15.3 0.12.1259 21 12.9 19 11.7 0.04.7357. AUTHORS' DISCLOSURES OF POTENTIAL CONFLICTS OF INTEREST Although all authors completed the disclosure declaration, the following author(s) and/or an author's immediate family member(s) indicated a financial or other interest that is relevant to the subject matter under consideration in this article.
John Sparapani
Certain relationships marked with a “U” are those for which no compensation was received; those relationships marked with a “C” were compensated. For a detailed description of the disclosure categories, or for more information about ASCO's conflict of interest policy, please refer to the Author Disclosure Declaration and the Disclosures of Potential Conflicts of Interest section in Information for Contributors. Employment or Leadership Position: None Consultant or Advisory Role: None Stock Ownership: Kiran Devisetty, Abbott Laboratories, AbbVie Honoraria: None Research Funding: None Expert Testimony: None Patents, Royalties, and Licenses: None Other Remuneration: None.
Bayesian additive regression trees (BART) provide a framework for flexible nonparametric modeling of relationships of covariates to outcomes. Recently, BART models have been shown to provide excellent predictive performance, for both continuous and binary outcomes, and exceeding that of its competitors. Software is also readily available for such outcomes.
In this article, we introduce modeling that extends the usefulness of BART in medical applications by addressing needs arising in survival analysis. Simulation studies of one‐sample and two‐sample scenarios, in comparison with long‐standing traditional methods, establish face validity of the new approach. We then demonstrate the model's ability to accommodate data from complex regression models with a simulation study of a nonproportional hazards scenario with crossing survival functions and survival function estimation in a scenario where hazards are multiplicatively modified by a highly nonlinear function of the covariates.
Using data from a recently published study of patients undergoing hematopoietic stem cell transplantation, we illustrate the use and some advantages of the proposed method in medical investigations. Copyright © 2016 John Wiley & Sons, Ltd.