Predicting overall survival (OS) and overall response (OR) following durvalumab treatment in patients with multiple cancer types using a hybrid modeling strategy.

Abstract

e15161 Background: This study aimed to predict OS/OR and identify key predictors in patients with diverse cancer types treated with durvalumab, a PD-L1 targeting monoclonal antibody, using a hybrid modeling strategy that combines population pharmacodynamic (PD) modeling and machine learning (ML) algorithms. Methods: Individual longitudinal tumor size measurements and OS/OR data were available for 855 patients who received durvalumab therapy (10 mg/kg Q2W or 20 mg/kg Q4W; NCT01693562). Nine cancer types included non-small cell lung cancer (NSCLC), bladder cancer (BC), microsatellite instability-high (MSI-H) cancer, hepatocellular carcinoma (HCC), squamous cell carcinoma of the head and neck (SCCHN), gastroesophageal cancer (GEC), ovarian cancer (OC), pancreatic adenocarcinoma (PDAC) and triple-negative breast cancer (TNBC). A tumor kinetic model was developed to characterize diverse temporal profiles using a population-based modeling approach. Individual estimated tumor kinetic model parameters and patient demographic/physiological factors were used as inputs for predicting OS/OR using several ML approaches. Results: The final tumor kinetic model with liver metastasis (LM), neutrophil/lymphocyte ratio (NLR), tumor size at baseline (TBSL) and cancer types as covariates characterized the temporal tumor size data well. HCC and MSI-H cancer have the slowest tumor growth rate constant (kg), while GEC, SCCHN and TNBC have the fastest kg. BC, NSCLC and OC have the highest tumor killing rate constant. The most important predictors of OS identified by ML approach were tumor kinetic parameters (kg, fraction of drug-sensitive cells, time-delay in immune response), along with baseline disease factors, including hemoglobin (HGBBL), albumin (ALB), and NLR. Decision tree-based algorithms showed the best performance in predicting OR with accuracy above 90%. In addition to tumor kinetic parameters, PD-L1 expression on tumor cells (TC) and ALB were the most important predictors of OR. Conclusions: A combined population PD/ML approach showed good predictions of OS/OR in patients with different cancer types treated with durvalumab. LM, NLR,TBSL and cancer types were found to be important factors for tumor kinetics. In addition to tumor kinetic parameters, HGBBL, ALB, and NLR were found to be important predictors of OS, and TC and ALB were found to be important predictors of OR. These findings could provide a guidance on patient selection in future clinical trials.