Can methods based on existing models really aid decision making in non-small-cell lung cancer (NSCLC) trials?
Jonathan L French (1), Daniel G Polhamus (1), and Marc R Gastonguay (1)
(1) Metrum Research Group, Tariffville, CT, USA
Objectives: The need for more efficient drug development in oncology is widely recognized. Wang et al [1] propose the use of % change in tumor size at 8 weeks (PTR8) as a marker of efficacy to aid decision making in NSCLC drug development. Sharma et al. [2] advocate M&S in oncology drug development through the use of adaptive Phase II-III trials. In this work we compare 3 approaches to using M&S for making decisions based on accruing data in a Phase II clinical trial.
Methods:We simulated clinical trials with 400 NSCLC patients randomized 1:1 to 2 groups receiving first-line treatment. We assume a recruitment period of 6 months with an additional 9 month follow-up period. Overall survival (OS) and PTR8 data were simulated using the NSCLC model of Wang et al. [1]. For each simulated trial, 3 interim analyses (IAs) were performed: 8 weeks after (1) 80 pts enrolled (~10 events), (2) 280 pts enrolled (~50 events), and (3) 400 pts enrolled (~90 events).
Two drug effect scenarios were evaluated. The first had a median difference in PTR8 of 40%, a difference in median OS of 100 days, and expected hazard ratio of 0.67. The second had no difference in PTR8 or OS.
We compared decision rules based on difference in PTR8 to Bayesian rules based on the posterior predictive distribution for the log hazard ratio (logHR) at the end of the study. For the Bayesian rules, the relationship between PTR8 and OS was updated using the accruing data in the trial and prior distributions centered at the estimates from [1]. We also evaluated rules based on the estimated logHR using a Cox model. Decision rules were compared using ROC analysis (true positive rate, TPR, and false positive rate, FPR).
Simulation and analysis were performed using R 2.15.1 and OpenBUGS 3.2.1.
Results: Under scenario 1, all rules performed poorly at IA1. At IA2, the best Bayes rule performed notably better (TPR=.67, FPR=.29) than either the best PTR8 rule (TPR=.56, FPR=.46) or Cox rule (TPR=.65, FPR=.37). At IA3, the best Bayes and Cox rules (TPR~.75, FPR~.25) performed better than the best PTR8 rule (TPR=.55, FPR=.33). Scenario 2 showed similar results.
Conclusions: Model-based approaches can aid decision making in NSCLC trials. However, IA decision rules based on PTR8 alone have little ability to predict the outcome of positive or negative studies, consistent with recently published results [3]. The model-based Bayesian decision rules evaluated here performed notably better.
References:
[1] Wang Y. et al. Elucidation of relationship between tumor size and survival in non-small-cell lung cancer patients can aid early decision making in clinical drug development. Clin Pharmacol Ther. 86, 167--174 (2009).
[2] Sharma MR, Maitland ML, and Ratain MJ. Models of Excellence: Improving Oncology Drug Development. Clin Pharmacol Ther. 92, 548--550 (2012).
[3] Claret L. et al. Simulations using a drug-disease modeling framework and phase II data predict phase III survival outcome in first-line non-small-cell lung cancer. Clin. Pharmacol. Ther. 92, 631--634 (2012).
