Cabozantinib

CheckMate 9ER is a phase 3, randomized, openlabel trial of nivolumab combined with cabozantinib as compared with sunitinib monotherapy. Patients underwent randomization in a 1:1 ratio and were stratified according to IMDC prognostic risk score (0 [favorable] vs. 1 or 2 [intermediate] vs. 3 to 6 [poor]),^{18 (link),19 (link)} geographic region (United States and Europe vs. the rest of the world), and tumor expression of the PD-1 ligand PD-L1 (≥1% vs. <1% or indeterminate). Specific risk factors that make up the IMDC score are included in the Supplementary Appendix, available at NEJM.org. Nivolumab was administered intravenously at a dose of 240 mg every 2 weeks, and cabozantinib was administered orally at a dose of 40 mg once daily. Sunitinib was administered orally at a dose of 50 mg once daily for 4 weeks, followed by 2 weeks off (6-week cycle). All trial treatment continued until disease progression or unacceptable toxic effects, with a maximum 2-year duration of nivolumab treatment. Crossover between groups was not permitted. Dose reductions were not allowed for nivolumab but were permitted for cabozantinib and sunitinib, according to the protocol. Dose delays for adverse events were permitted for all trial drugs. Discontinuation assessments for nivolumab and cabozantinib were made separately for each drug; if discontinuation criteria were met for only one drug, treatment could continue with the other drug that was not related to the observed toxic effect, according to the protocol. Dose-reduction specifications and discontinuation criteria for both groups are detailed in the trial protocol.

As detailed below, a five-step procedure was implemented including (1) elicitation of an initial composite grading algorithm from clinical investigators via a data collection exercise; (2) refinement of the algorithm through targeted questions administered to a broader audience of clinical investigators; (3) use of graphical approaches to ensure the directional consistency of composite grades; (4) quantitative testing of the algorithm’s measurement properties in a national study dataset; and (5) application of the algorithm to data from multi-site randomized cancer clinical trials.
In Step 1 (“Clinical Investigator Grade Assignment”), a data collection form was created with a table showing the 179 possible combinations of PRO-CTCAE item attributes each in a row (e.g., frequency “occasionally”, severity “mild”, interference “somewhat”), with a space for the investigator to select a single composite grade from 0-3 to which he/she felt that combination should map, corresponding to the CTCAE (Table 1, Supplemental Table S2). This form was administered anonymously to 20 clinical investigators with ≥10 years of experience serving as principal or co-investigators for NCI- and industry-sponsored cancer clinical trials. Composite grades assigned by the 20 investigators were tabulated for each combination, with consensus defined a priori as endorsement of a specific numerical grade by ≥75% of investigators for a given combination. Combinations that did not reach consensus were further investigated in Step 2.
In Step 2 (“Clinical Investigator Consensus”), input on the algorithm resulting from Step 1 was systematically elicited from clinical investigators attending the Alliance for Clinical Trials in Oncology biannual group meeting in Chicago, IL. During committee meetings for the Breast, Thoracic, Genitourinary, and Gastrointestinal disease committees, clinical scenarios were presented via PowerPoint slides (Supplemental Figure S1) which described patients with adverse events characterized by various individual attribute combinations. Investigators were asked to vote using electronic audience response units on a single composite grade to which they felt each combination should map, consistent with the CTCAE. No identifying information was collected from the investigators, although there was no overlap between investigators involved in Steps 1 and 2. Clinical scenarios included attribute item score combinations that did not reach consensus in Step 1. Clinical scenarios were also presented for 12 randomly selected item attribute combinations that had reached ≥75% consensus in Step 1 to confirm consensus. Agreement across the respondents was tabulated for each combination. The majority grades were used in the composite grading algorithm evaluated in Step 3.
Step 3 (“Directional Consistency Check”) entailed use of graphical and tabular approaches to ensure the directional consistency of composite grades. Contour plots of composite grades by frequency, severity, and interference were created in Matlab (method developed by co-author Claus Becker) and reviewed to identify whether situations existed such that increasing individual PRO-CTCAE attribute scores would lead to decreasing composite grades. Composite grades for adverse events with two-score combinations (e.g., frequency plus severity) were compared to the range of composite grades for adverse events with three-score combinations (frequency plus severity plus interference) to confirm consistency of composite grades. Directionally inconsistent composite scores were modified and the final composite grading algorithm was then quantitatively tested in Steps 4 and 5.
Step 4 (“Validation”) entailed comparing the measurement properties of the composite grades derived from Steps 1-3 to the previously published measurement properties of the individual items, including validity, reliability, and sensitivity to change. We sought to assess whether measurement properties were comparable. The algorithm was applied to PRO-CTCAE data collected at 2-3 visits in the primary validation study of the PRO-CTCAE,^{5 (link)} which enrolled 940 patients receiving active treatment for a variety of cancer types at 9 US-based centers. Convergent validity was assessed using Pearson correlations between PRO-CTCAE composite grades and European Organisation for the Research and Treatment of Cancer Quality of Life Questionnaire Core 30 (EORTC QLQ-C30) scales,^{11 (link)} including an overall summary score.^{12 (link)} EORTC QLQ-C30 health-related quality of life summary and functioning/global scales were reverse scored for analysis (original direction retained for graphical representations) such that higher scores represent inferior outcomes, matching the direction of PRO-CTCAE items. Comparison of mean EORTC QLQ-C30 health-related quality of life summary scores across increasing PRO-CTCAE composite grade groups were carried out using Jonckheere-Terpstra tests, which evaluate for monotonically decreasing health-related quality of life for increased PRO-CTCAE composite grade groups.¹³ Known-groups validity was assessed by comparing mean PRO-CTCAE composite grades between 66 previously defined^{5 (link)} groups of patients on the basis of Eastern Cooperative Oncology Group (ECOG) performance status (0-1 versus 2-4), cancer type, and treatment using two-sample t-tests with effect sizes computed as the differences between group means divided by the pooled standard deviation (i.e., Cohen’s d). Test-retest reliability was estimated using intraclass correlation coefficients based on a one-way analysis of variance model. Sensitivity to change was assessed by comparing changes from first to last visit between groups of patients reporting worsening, no change, or improvement via global impression of change items at the last visit using Jonckheere-Terpstra tests. Standardized response mean was computed per group as the mean change score divided by the standard deviation of the change scores. All methods and measurement properties of individual PRO-CTCAE items have previously been described.^{5 (link)}Finally, in Step 5 (“Clinical Trial Evaluation”), the grading algorithm was applied to PRO-CTCAE data collected in two double-blind placebo-controlled phase III trials: sorafenib vs. placebo in patients with advanced desmoid tumors (Alliance A091105, NCT02066181)^{14 (link)}; and cabozantinib vs. mitoxantrone plus prednisone in patients with metastatic castration-resistant prostate cancer (Exelixis COMET-2, NCT01522443).¹⁵ Details of patient consent are previously reported.^{14 (link),15} To assess the utility and meaningfulness of the composite grading approach, individual adverse event rates were compared between arms using individual PRO-CTCAE item scores vs. composite grades. Post-baseline PRO-CTCAE rates were tabulated and compared with Fisher’s exact tests using a previously described method for baseline adjustment,^{8 (link)} following application of the composite grading algorithm to scores at each assessment time point.

Up to three analyses of the primary end point of overall survival were planned, when approximately 50%, 75%, and 100% of the expected deaths had occurred. We estimated that a sample size of 760 patients, with a total of 621 deaths, would provide the trial with 90% power to detect a hazard ratio of 0.76 favoring cabozantinib over placebo, with a two-sided log-rank test at a 5% level of significance. Assuming a median overall survival of 8.2 months in the placebo group (as shown in the Brivanib Study in HCC Patients at Risk Post Sorafenib [BRISK-PS]^{28 (link)}) and exponential distribution, this would correspond to 32% longer median overall survival (10.8 months) in the cabozantinib group. Inflation of the type 1 error associated with interim analyses was controlled with the use of the Lan–DeMets O’Brien–Fleming alpha spending function.^{29 (link)} If the null hypothesis of no difference in overall survival was rejected at either the first or second interim analysis, testing of secondary end points would proceed, and subsequent analyses of overall survival would not be performed.
Efficacy was assessed in all randomly assigned patients according to the intention-to-treat principle. Safety was assessed in all patients who received at least one dose of the trial regimen. For time-to-event end points, hypothesis testing was performed with the stratified log-rank test with adjustment for the stratification factors used at randomization; median durations and associated 95% confidence intervals were estimated with the Kaplan–Meier method. Hazard ratios were estimated with univariate Cox regression models, with the randomized group as the only predictor. Hazard ratios for overall analyses were calculated from models adjusted for the randomization stratification factors. Hypothesis testing of objective response was performed with the Cochran–Mantel–Haenszel method. All subgroup analyses of overall survival and progression-free survival were prespecified except those based on extrahepatic spread of disease or macrovascular invasion as separate factors and on sorafenib as the only previous therapy. For subgroup analyses, no adjustments were made for multiplicity, and confidence intervals are considered to be descriptive. Hazard ratios for subgroup analyses were calculated from unstratified models except those calculated for the subgroup of patients whose only previous therapy was sorafenib. All analyses were performed with SAS software, version 9.1 or higher (SAS Institute).