Alina Sofronova (1), Kirill Zhudenkov (1), Sergey Gavrilov (1), Oleg Stepanov (1), Kirill Peskov (1, 2)
(1) M&S Decisions LLC, Moscow, Russia (2) Computational Oncology Group, I.M. Sechenov First Moscow State Medical University, Moscow, Russia
Introduction: Joint modeling (JM) is a new and highly flexible tool in the analysis of the relationship between longitudinal and time-to-event data [1]. In non-small-cell lung cancer (NSCLC), dynamics of the sum of longest lesion diameters in target lesions (SLD) is frequently monitored and is assumed to be connected with overall survival (OS). However, estimation of SLD dynamics may be restricted by using linear models in the longitudinal part of JM [2,3]. In this research we are investigating the nonlinear joint models and assess these models’ predictive performance for selected NSCLC data.
Objectives: We tested baseline covariates and quantified the association of nonlinear SLD dynamics with OS in patients with NSCLC and compared model predictive efficacy using training and validation datasets. JM model performance in OS prediction was also compared with similar Cox models, that are capable to use baseline SLD only. Additionally, we checked whether the predictions of study OS can be carried through different lines of cancer therapy.
Methods: Data from control arms of three phase 3 clinical studies in second-line NSCLC treatment were obtained from Project Data Sphere repository and merged into one training dataset. It included 596 and 520 patients, treated with standard chemotherapy (docetaxel) from ZODIAC (NCT00312377) and INTEREST (NCT00076388) trials, respectively; and 512 patients with erlotinib treatment (NCT00364351). Another dataset, including the data for 606 patients with first-line chemotherapy from IPASS study (NCT00322452) was used as the dataset for model validation.
Cox models with different combinations of baseline biomarkers (SLD, alkaline phosphatase, aspartate aminotransferase, alanine aminotransferase, creatinine, neutrophils, white blood cells, EGFR mutation, WHO and smoking status) were tested. The final Cox model with Weibull baseline hazard for OS was selected based on the optimal log-likelihood as well as AIC and BIC values and relative standard errors of parameters less than 50%. Basing on this Cox model, JM was developed using empirical nonlinear bi-exponential model for SLD dynamics in the longitudinal part. Estimation of model parameters was conducted using the SAEM algorithm in Monolix 2019R1 software.
The quality of survival predictions of Cox and JM models were compared by assessment of the receiver-operating characteristic curve (ROC AUC) and Brier score (BS). ROC AUCs and BSs were calculated for patients from training and validation datasets with 90-day cut-off of longitudinal data. Additionally, multiple study-level simulated survival probability distributions were compared to observed KM curve using the logrank test.
Results: Significant difference between observed KM curves for patients with second and first lines of treatment was identified for training and validation datasets.
A significant (p<0.0001) association between longitudinal SLD and OS was quantified. Similar association constants for chemotherapy and erlotinib were identified.
Final Cox model included baseline values of SLD, alkaline phosphatase (ALP), white blood cells (WBC) continuous biomarkers and smoking and WHO status as categorical covariates. Corresponding JM model utilized nonlinear SLD dynamics and a similar set of other baseline biomarkers.
Accumulation of longitudinal data up to 90 days enhanced the survival prediction accuracy of JM on the training dataset in terms of ROC AUC and BS in comparison with Cox model. Moreover, JM outperformed Cox model on the validation dataset with another line of treatment.
Accounting of SLD dynamics up to 90 and 180 days in JM improved accuracy of OS predictions, as compared to survival models with only baseline covariates (Cox model).
Conclusions: JM with nonlinear SLD dynamics and longitudinal data accumulation for a selected cut-off provided higher individual (ROC AUC and BS) and overall (marginal OS simulation) survival predictive performance, in comparison with Cox proportional hazards model in considered patients with NSCLC.
References:
[1] Rizopoulos D. Joint Models for Longitudinal and Time-to-Event Data. CRC press,2012,ISBN 9781439872864.
[2] Gavrilov S et al. 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. PAGE Meeting, 2019.
[3] Gavrilov S et al. Longitudinal tumor size and NLR are more predictive of individual survival than their baseline values in patients with non-small cell lung cancer treated with durvalumab. 2019 ASCO Annual Meeting.
Reference: PAGE () Abstr 9489 [www.page-meeting.org/?abstract=9489]
Poster: Oral: Drug/Disease Modelling