Aurélie Lombard (1), Hitesh Mistry (1), Sonya Tate (2), Ivelina Gueoguieva (2), Leon Aarons (1), Kayode Ogungbenro (1)
(1) University of Manchester, (2) Eli Lilly and Company, Erl Wood Manor, Windlesham
Objectives: Being readily available during oncology clinical trials, tumour size (TS) measurements are commonly used as a biomarker for drug efficacy. Indeed, TS represents a key variable of the treatment phase of a study that determines disease progression and consequently, whether patients will stay on treatment for another cycle. Variability in TS measurements could lead to drug effect misinterpretation and therefore affect patient’s performance within the trial. Although inter-observer variability has been investigated in the past [1], there is limited knowledge available about its impact on response-to-treatment assessment using modelling approaches despite their use to determine patient benefit during clinical trials. Here, we explore the inter-operator variability of tumour size measurements in a selected population of a phase III clinical trial and its impact on model-based drug effect evaluation at individual lesion level.
Methods: TS measurements (longest individual lesion diameter) were obtained from a randomized phase III clinical trial [2] where metastatic non-small cell lung cancer patients were treated with cisplatin alone or in combination with gemcitabine. 122 lesions from 62 patients (out of 522) were selected according to imaging methods (Computed Tomography-scan) and based on the availability of two lesion measurements of the same CT-scan at each time point; measurement 1 (M1), performed at the hospital; measurement 2 (M2), performed at the centralised centre. Firstly, a graphical exploration was performed to identify trends within TS kinetics. The correlation between M1 and M2 was investigated by using linear regression and the relative error ratios (RER) were derived. Secondly, a tumour growth inhibition (TGI) model was applied separately to the M1 and the M2 data (NONMEM 7.3, FOCE I). The correlation between M1 and M2 individual estimates of TS at baseline and the drug effect estimate were assessed by using linear regression. The relative error ratios of population parameter estimates were also derived.
Results: We found three different patterns by visually comparing M1 and M2 measurements: (i) M1 and M2 were similar over time (23%) and so no impact will be observed on drug effect assessment; (ii) M1 and M2 were different but follow the same trend (42%) and the lack of agreement would mostly affect estimation of the tumour size at baseline; (iii) M1 and M2 were different and did not follow the same trend (35%) and the discrepancies would directly affect drug efficacy evaluation. The linear regression analysis showed that M1 and M2 of the same lesion were correlated with an r2=0.72; however, a higher correlation would have been expected, as M1 and M2 are based on the same “true value” and the variability only relies on the operator. The RERs of M1 compared to M2 were widely distributed from -91.7% and 1,200.0%. Extreme ratios were mostly observed when one radiologist (e.g. at the hospital) measured a tumour which was considered to be non-existent by the other radiologist (e.g. at the centralised centre). The 1st and the 3rd quantiles were distributed from -13.4% to 28.5%, close to the 15% variation observed by Hopper et al. [1]. The analysis of the TGI model parameters showed that the correlation between M1 and M2 individual estimates of TS at baseline was comparable to the raw measurements (r2 = 0.71), contrary to the individual estimates of drug effect, which appeared not to be correlated (r2 = 0.27). However, M1 and M2 population parameters were similar (less than 15% variation), indicating that interpretation of outcomes for a typical patient will be close.
Conclusion: This analysis revealed that the operator is an important variable to consider which can induce a wide variability in tumour size measurements. This could affect the model-based interpretation of drug response, especially at the individual level, as no correlation was observed between M1 and M2 drug effect estimates. However, drug-response interpretation for a typical patient will be similar as population parameters were comparable; suggesting that the global evaluation of drug efficacy by modelling approaches might not be affected.
References:
[1] Hopper et al. (1996);167(4):851-4.
[2] Sandler et al. J Clin Oncol. (2000);18(1):122-30.
Reference: PAGE 28 (2019) Abstr 8853 [www.page-meeting.org/?abstract=8853]
Poster: Drug/Disease Modelling - Oncology