Longitudinal assessment of tumor size and neutrophil count in multivariate joint models are more predictive of survival than their baseline values in patients with non-small cell lung cancer
Sergey Gavrilov (1), Kirill Zhudenkov (1), Kirill Peskov (1, 2), Gabriel Helmlinger (3), Sergey Aksenov (3)
(1) M&S Decisions LLC, Moscow, Russia, (2) Computational Oncology Group, I.M. Sechenov First Moscow State Medical University, Moscow, Russia, (3) Clinical Pharmacology and Safety Sciences, R&D BioPharmaceuticals, AstraZeneca, Boston, USA
Introduction: Tumor size, quantified by the Sum of Longest Diameters of target lesions (SLD), has been established as a prognostic marker for NSCLC survival of patients with non-small cell lung cancer (NSCLC), while neutrophil count (NEU) and other hemogram measurements have recently been investigated. Joint models of survival and longitudinal biomarkers, e.g. SLD [1] are a useful framework for survival prediction. Gain in prognostic accuracy of SLD and NEU when measured longitudinally has not been established.
Objectives: The aim of this work was to quantify the association of selected clinical biomarkers ECOG, SLD and NEU with overall survival (OS) in NSCLC data, and systematically test joint model prediction performance when using baseline vs. longitudinal biomarkers SLD and NEU.
Methods: We used two datasets: one set (studies NCT02087423 [2] and NCT01693562 [3]) included 679 NSCLC patients (53% with PDL1 expression >=25%, 41% - <25%, 6% - unknown) from clinical studies of durvalumab, an anti PD-L1 inhibitor (10 mg/kg intravenously every two weeks). Another set (study NCT00322452 [4]) included 354 NSCLC patients (27% mutant EGFR, 17% - wild-type, 56% - unknown) from a clinical study of gefitinib, an EGFR inhibitor (250 mg oral daily). We used the first set to develop three models of OS, all with ECOG as a baseline covariate: a Cox proportional hazards model with SLD and NEU as baseline covariates (COX); a joint model of OS and longitudinal SLD and baseline NEU (JM SLD); and a joint model of OS and longitudinal SLD and NEU (JM SLD&NEU).
In order to estimate and compare the predictive accuracy of these models, we evaluated model performance using area under the receiver-operating characteristic curves (ROC AUC) and Brier scores (BS). Time-dependent ROC AUCs and BSs were calculated for patients from the first, training durvalumab set and the second, prediction set for gefitinib using longitudinal data with different cut-offs in time. Marginal survival was calculated by simulating times of death or event-free survival using the joint models qualified using the data from durvalumab dataset. For each simulation, baseline and longitudinal data were picked from 300 resampled patients from gefitinib dataset. There were 500 simulated datasets. We calculated the median and range over the simulated survival distributions and compared them with the observed Kaplan-Meier (KM) curve.
Results: Patients in the durvalumab and the gefitinib datasets were similar in demographic and disease characteristics and baseline SLD; however, SLD declined at a faster rate in EGFR mutant than EGFR wild-type on gefitinib or patients on durvalumab; NEU were uniformly lower on gefitinib than durvalumab; and OS was better on gefitinib than durvalumab.
Starting from a 2-month cut-off of using longitudinal data, JM SLD and JM SLD&NEU outperformed the COX model on both the training durvalumab and the gefitinib datasets. Moreover, multivariate JM SLD&NEU showed better performance in comparison to JM SLD.
For instance, the following results were obtained, using a 4-month cut-off of longitudinal data and for up to 12-month individual patient survival discrimination after start of treatment: the COX model had a ROC AUC = 0.66, BS = 0.22 based on the training durvalumab dataset, and ROC AUC = 0.61, BS = 0.18 for the gefitinib dataset. JM SLD scored ROC AUC = 0.74, BS = 0.19 for the durvalumab dataset, and ROC AUC = 0.77, BS = 0.14 for the gefitinib dataset. JM SLD&NEU scored ROC AUC = 0.81, BS = 0.17 for the durvalumab, and ROC AUC = 0.82, BS = 0.13 for the gefitinib dataset. Noninformative models would have ROC AUC 0.5 and BS 0.25.
OS predictions agreed with the observed KM estimate better using JM SLD and JM SLD&NEU than COX, for both durvalumab and gefitinib datasets.
Conclusions: Using longitudinal data for SLD and NEU in joint models increased individual discrimination performance and marginal survival predictions vs. baseline data in COX models. Different marginal survival in durvalumab and gefitinib treated patients was explained by differences of SLD and NEU kinetics in these different disease entities and a common modeled relationship between SLD and NEU and survival in the models.
References:
[1] Brilleman, S. L., et al. (2018). Statistical Methods in Medical Research. https://doi.org/10.1177/0962280218808821
[2] Garassino MC, et.al. Lancet Oncol. 2018 Apr;19(4):521-536;
[3] Powles T., et.al. JAMA Oncol. 2017 Sep 14;3(9):e172411;
[4] Wu YL, et al. Asia Pac J Clin Oncol. 2012 Sep;8(3):232-43.