Evaluating Heterogeneity in Treatment Effects and Economic Value of Tumor-Agnostic Drugs

Abstract

Tumor-agnostic drugs (TAD), also known as histology-independent treatments, have the potential to benefit patients who currently have limited therapeutic options. TAD typically received accelerated approvals based on basket trials which included a small number of multi-cohort, single-arm studies. However, the evaluation of TAD poses major challenges for health technology assessment agencies, such as the potential heterogeneity in treatment effect by tumor type, the lack of comparative data due to single-arm studies, and variable standard of care (SoC) across tumor types. Consequently, these challenges create significant uncertainty regarding the expected clinical and economic impact of TAD. In Aim 1, I used Bayesian hierarchical models (BHM) to assess heterogeneity in treatment outcomes across tumor types and improve estimates of tumor-specific treatment outcomes from Phase II basket trials, which are crucial for healthcare decision-making. My analysis revealed high heterogeneity and uncertainty in survival endpoints, including median progression-free survival (PFS) and median overall survival (OS), although the treatment effects are more similar when judged by surrogate endpoint at approval. Metrics such as intra-class correlation can be used to quantify the variation between groups, which could inform a recommendation of TAD for use in all tumor types or a restricting subset of patients. The findings from our study are important because they demonstrated that BHM could reduce uncertainty of estimates derived from basket trial evidence, potentially improving confidence in tumor-agnostic decision making, despite small sample sizes in some tumor types. The methods presented in this study can be applied to the future assessment of TAD. In Aims 2 and 3, I address comparative effectiveness and economic value of TAD with single-arm trial evidence. To overcome the challenge of lacking comparators, I created eight external controls using observational data from the TriNetx electronic health databases. The Copula method was employed to simulate correlated trial samples while matching the dependence structure of the trial baseline covariates to that in the real-world population. Additionally, an inverse odds weighting approach was used to further balance the baseline characteristics between the trial and external control arms. Weighted Cox regressions showed that patients with MSI-H/dMMR advanced/metastatic colorectal and endometrial cancers receiving pembrolizumab were associated with significant prolonged PFS but not OS than real-world patients receiving chemotherapies. This analysis demonstrated that incorporating external control data in early phase trials may provide a more comprehensive understanding of treatment effects of tumor-agnostic drugs than relying solely on single-arm trials. Finally, using adjusted efficacy inputs from Aim 1 and external controls from Aim 2, I assessed the economic value of pembrolizumab compared to SoC across 8 tumor types to inform coverage and reimbursement decisions in the United States. A partitioned survival model with three health states (i.e., progression-free, post-progression, and death) was developed to evaluate the cost-effectiveness of pembrolizumab for previously treated patients with advanced or metastatic MSI-H/dMMR cancers. We found substantial variation in the economic value across tumor types, with pembrolizumab being a cost-effective strategy in treating colorectal and endometrial cancers at $150,000 willingness-to-pay per quality adjusted life years threshold, compared to SoC chemotherapies. However, pembrolizumab was not found to be cost-effective in treating other assessed cancers. The main findings from value of stratification estimates suggest that recommendations for using pembrolizumab in specific patient populations, based on comparative effectiveness or net health benefit, could result in greater overall value to the healthcare system compared to a tumor-aggregated recommendation.