A Bayesian population-based framework for detecting hyperprogressive disease on cancer immunotherapies.

Abstract

2665 Background: Hyperprogressive Disease (HPD), defined as an unexpected treatment-induced rapid increase in tumor growth rate relative to the tumor burden pretreatment growth rate, has been reported in 9% of patients receiving immune checkpoint inhibitor (ICI) therapies (Champiat et al, Clin Cancer Res (2017)). The definition of HPD used in that work was an increase in the on-treatment “Tumor Growth Rate” (TGR) by a factor of 2 or greater over the pre-treatment TGR as calculated from three consecutive CT scans: pre-baseline, baseline and on-treatment. This method of directly calculating TGR from slopes between successive tumor measurements, however, does not consider uncertainty in the CT-assessed sum of diameters (SoD) of target lesions (~8% per Zhao et al, Radiology (2009)), nor prior knowledge of the distribution of responses under ICI treatment. We sought to develop and test a Bayesian approach to HPD detection and compare its receiver operating characteristics (ROC) to the published TGR ratio threshold of 2-fold as well as the TGR ratio considered as a continuous classifier of HPD. Methods: We represented the prior distribution of exponential TGR (eTGR) pre-treatment based on population modeling of historical data of untreated NSCLC tumor dynamics. We then calibrated the on-ICI-treatment prior distribution based on reported rates of HPD and tumor regression. Next, we developed a Bayesian parameter estimation system to take three successive SoD assessments to calculate a patient’s posterior probability of being in a state of HPD, attenuated tumor growth (ATG) or tumor regression (REG). We then simulated a cohort of 1000 virtual patients (VPs) with known state and tested the ability of the Bayesian method, the TGR ratio method, and the TGR ratio > 2 method to correctly classify each VP. We additionally developed a user-friendly web-based prototype tool ( https://deanbot1.shinyapps.io/GRICalc/ ) to solicit feedback from a potential future user community. Results: ROC analysis estimated the area under ROC curve (AUROC) to be 0.94 for the Bayesian method, a significant improvement in classification accuracy over the TGR ratio method (AUROC = 0.77) as well as better maximal sensitivity (Se) and specificity (Sp) than the TGR ratio > 2 method (Se = 80% & Sp = 90% vs Se = 90% & Sp = 40%). Conclusions: We have developed a Bayesian population-based methodology and prototyped a web-based tool which under simulated conditions consistently outperformed previous methods for detecting HPD. We believe that this approach, trained on a larger patient-level dataset, can potentially support and improve clinical management and decision-making for cancer patients taking immune checkpoint inhibitor therapies.