Center for Global Research Data

PATIENT REPORTED OUTCOME (PRO) AS A PROGNOSTIC MARKER IN PATIENTS TREATED WITH TARGETED THERAPIES FOR ADVANCED HEPATOCELLULAR CARCINOMA (PRO²-HCC)

Lead Investigator: Julien Edeline, Centre Eugene Marquis
Title of Proposal Research: PATIENT REPORTED OUTCOME (PRO) AS A PROGNOSTIC MARKER IN PATIENTS TREATED WITH TARGETED THERAPIES FOR ADVANCED HEPATOCELLULAR CARCINOMA (PRO²-HCC)
Vivli Data Request:7062
Funding Source: None
Potential Conflicts of Interest: Consultant for BMS, not related to the subject of the current research

Summary of the Proposed Research:

Quality of life could be a marker of the aggressiveness of a disease. We will study whether patient-rated quality of life could replace the evaluation of the general health status performed by the physician as a way to evaluate the prognosis of hepatocellular carcinoma.

Statistical Analysis Plan:

Statistical analyses will be performed using R statistical software version 3.1.1.

Data collected for the present study will be completely anonymized, the patients will be identified by a code, different from the BRISK-FL study identification code, provided by the BRISK-FL sponsor. The correspondence table will be kept by the BRISK-FL sponsor. No date will be transmitted, but solely durations. Only aggregated results could be publically presented.

The following data will be extracted from the original database:

  • Patient’s characteristics at treatment initiation (or randomization):
    Demographics: gender, age;
  • PS;
  • HCC and OS prognostic parameters: Child-Pugh score, BCLC classification, region of origin, albumin, bilirubin, INR, tumor maximal diameter, Alpha-foeto-protein, portal vein invasion, metastasis, risk factors for hepatopathy, previous treatment for HCC, significant medical history or antecedents;
  • HR-QOL at treatment initiation: date of questionnaire, QLQ-C30 scores and atomic answers

Follow-up:

  • vital status (alive, death, lost to follow-up)
  • time from randomization to death or last news date
  • cause of death (related to cancer or not)
  • first antineoplastic treatment after the study
  • First progression and time from randomization
  • Best observed response and time from randomization

Protocol or procedure data:

  • Arm of treatment and treatment received
  • Deviation : yes or no
  • Type of deviations
  • Stratification parameters

Number of subjects:
In BRISK-FL study, 412 deaths in the Sorafenib arm and 425 deaths in the Brivanib arm were observed among the 1150 included patients. Thus, it could be expected about 280 and 557 events in data of our validation and training sets respectively. If the following assumptions are assumed: risks alpha and beta are 5% and 20% respectively, the new HR-QOL score is a binary variable and its repartition in the study population is well-balanced, therefore the number of events in the validation set appears sufficient to detect a HR around 1.4 (in Diouf et al. physical functioning subscales < 50 was independently associated with OS and HR was estimated to 2.00 [1.32 – 3.04]). Moreover, applying the empirical rules of 1 covariate for 10-20 events, the number of events found of the training set appears sufficient to allow evaluating concomitantly a high number of covariates with a Cox multivariable model.

Statistical analysis:
Statistical analyses will be performed using R statistical software version 3.1.1. Patients’ characteristics will be described by either medians and range, or means and standard deviation for continuous variables, and absolute frequencies and percentages for categorical variables. Survival curves will be estimated with the Kaplan-Meier approach, and compared using Log Rank test. A p-value < 0.05 will be considered as significant for all analyses.

Prognosis factors of overall survival will be evaluated using a multivariable Cox regression model. Prognosis factors will be transform (if necessary only) to assume proportional hazard assumptions which could be verified using the Cox model’s Schoenfeld residuals, and log-linearity for quantitative variables. A full model will be fitted by each variable individually associated with OS at a p value cut-off of 0.2 using Wald test. Then, covariate will be selected through a bidirectional stepwise elimination procedure using Akaike criterion to achieve parsimony of the final model. Variance inflation factor will be used to avoided multicolinearity.

The final model will be fitted from a patient sample consisting in two-thirds of our population study (training set), then the final model will be tested against first seen data which will consist in data from the last third of our study population (validation set). Both training and validation set will be built by a stratified random sampling on BRISK-FL study’s stratification variables, the study arm and the global QLQ-C30 HR-QOL score. Finally, adjusted and unadjusted hazard ratios (HR) will be estimated and presented with their 95% confidence interval.

Previously, using our training set a new QoL-based scoring classification will be derived from the EORTC QLQ-C30 questionnaires. Association with OS of each HR-QOL subscales as a continuous variable will be tested with a Cox univariable regression model. Only subscales associated with OS at a p value < 0.05 will be retained for the remainder of the analysis. The selected subscales will be then categorized by quartiles of distribution and the link with OS evaluated using survival curves and Log Rank test. According to this first visual evaluation, a reduction in the number of categories could be envisaged. Performance of these categorical prognosis scores will be also assessed with the Schemper statistic and Harrell’s C-Index, in order to find the best clinically-usable and efficient structure for the new HR-QOL prognosis factors. Higher the Schemper statistic is, better is the survival prediction; Higher the Harrell’s C-Index is, better is the discriminatory ability of the new score. Finally, the HR-QOL scores will be evaluated in association with the other prognosis factors as described above using a multivariable Cox proportional hazards modelling.

Moreover, association between the HR-QOL subscales, PS and other existing prognosis scores (as BCLC score) could be evaluated using scatter plots, Pearson or Spearman correlation coefficients, raw concordance rates and quadratic weighted kappa coefficients with bootstrapped 95% confidence intervals.

Requested Studies:

First Line Hepato Cellular Carcinoma (HCC) (BRISK FL)
Data Contributor: Bristol Myers Squibb
Study ID: NCT00858871
Sponsor ID: NCT00858871

Summary of Results:

The data shared due to the lack of common keys between the data tables are not usable, the analysis planned in this study could not be conducted.

Update: This data request was withdrawn on 20-Sep-2022 by the researcher.