Model-based Prediction of Progression-Free Survival for Combination Therapies in Oncology

Conference poster, 2023

Objectives: 

Extend a joint modeling approach for predicting progression-free survival (PFS) for monotherapies [1] to combination therapies.
Test the model’s predictive capabilities by performing different cross-validations.
Methods: 

PFS is an important clinical metric for comparing and evaluating similar treatments for the same disease within oncology. Using the RECIST (version 1.1) guidelines all tumor lesions have to be accounted for by a combination of target and non-target lesions. A patient’s PFS time is set by target progression (TP) if there is at least a 20%- and 5-mm increase of the sum of the largest target diameters (SLD) compared to the nadir [2]. If the patient dies, a new lesion has appeared, or the non-target lesions are deemed unequivocal progressing the PFS time is instead set by non-target progression (NTP). If a patient leaves the trial before this occurs the patient is censored at that time point.

We present a joint modeling approach for predicting PFS for combination therapies where we link the risk of adverse events such as e.g., tumor metastasis or death with the derivative of SLD. Thus, the joint model consists of both a tumor growth inhibition (TGI) model, for the SLD time series, and a time-to-event (TTE) model to model the risk of adverse events. In addition, a Weibull TTE model is used to account for dropout.

Monolix [3] was used to calibrate the models with data coming from a clinical study comparing the efficacy of FOLFOX versus FOLFOX + panitumumab in metastatic colorectal cancer patients. The data were provided to us by ProjectDataSphere [4]. We did not have data for panitumumab given as a monotherapy and therefore assume that there were no interaction effects between the drugs. To adequately quantify the variability in the data the nonlinear mixed effects framework was used.

After the models were calibrated they were combined to make predictions of PFS. The algorithm below summarizes how the predictions are performed,

Generate artificial patients and simulate time series of SLD using the TGI model.
Estimate time when SLD has increased by 20% and at least 5 mm for each patient.
Construct individual survival curves and sample time of non-target progression event.
Sample dropout times using the estimated Weibull distribution.
Pick the time that occurs first for each patient, record the PFS trigger, and repeat it 1000 times. If dropout occurs first, the patient is censored that that time.
 

From this procedure, we obtain both a median prediction along with a 95% confidence interval for the prediction.

To both test the model’s predictive capabilities and validate the assumption of no interaction between the drugs we predicted the median PFS time for panitumumab given as a monotherapy and compared it with results from the ASPECCT study [5]. We also recalibrated the model with truncated data at 3,7, and 27 months and then made forward predictions of the remaining study.

Results: 

All models were successfully calibrated to the data and validated based on, e.g., the precision of parameter estimates, individual fits, distribution of Empirical Bayes Estimates (EBEs), and analysis of residuals. Furthermore, the combined (PFS) model was able to describe the PFS for both treatment arms of the study. When we recalibrated the model with truncated data, the forward predictions were very good for both the 7 and 27 months truncation points. The prediction for the median PFS time for patients given only panitumumab was similar to what was found in the ASPECCT study.

Conclusions: 

We successfully calibrated a joint model using clinical SLD and TTE data for a combination therapy. Using the model, we were able to first describe the PFS time of the same study well and then make model predictions. Predictions were performed on both truncated data sets and for data coming from a different study. In both cases, the model was shown to have good predictive capabilities.



 

References:
[1] Yu J, Wang N, Kågedal M. A New Method to Model and Predict Progression Free Survival Based on Tumor Growth Dynamics. CPT Pharmacomet Syst Pharmacol 2020;9:177–84. https://doi.org/10.1002/psp4.12499.
[2] Eisenhauer EA, Therasse P, Bogaerts J, Schwartz LH, Sargent D, Ford R, et al. New Response Evaluation Criteria in Solid Tumours: Revised RECIST Guideline (Version 1.1). Eur J Cancer 2009;45:228–47. https://doi.org/10.1016/j.ejca.2008.10.026.
[3] Monolix 2021R2 Lixoft SAS, a Simulations Plus company.
[4] Project Data Sphere 2022. https://www.projectdatasphere.org/.
[5] Kim TW, Peeters M, Thomas A, Gibbs P, Hool K, Zhang J, et al. Impact of Emergent Circulating Tumor DNA RAS Mutation in Panitumumab-Treated Chemoresistant Metastatic Colorectal Cancer. Clin Cancer Res 2018;24:5602–9. https://doi.org/10.1158/1078-0432.CCR-17-3377.