Modeling Multi-Source Data in Hodgkin Lymphoma

点击翻译按钮获取中文摘要

ABSTRACT Hodgkin lymphoma (HL) is associated with excellent cure rates. Despite early disease control, this success comes at considerable cost, with treatment-induced morbidity (“late effects” [LE]) that manifests as early as a year after therapy (e.g., 2nd cancers, cardiovascular [CVD] disease), compromised health-related quality of life (HRQL), and early mortality. Over the past 20 years, while important HL clinical trials have been conducted, there has been a lack of consensus about the preferred approach to treatment, considering both short- and longer term outcomes. While clinical trials provide rich information about short term disease outcomes, they do not follow efficacy beyond 5 years and rarely track post-therapy morbidity. In contrast, HL registries capture longitudinal outcomes and can be linked to other sources including healthcare utilization via administrative claims and cancer registries to characterize LE and survival. This proposal represents an unprecedented opportunity to draw on the collective expertise of International clinical HL oncology experts, epidemiologists, imaging experts, and methods & modeling experts with access to individual patient data from modern HL clinical trials and real-world registries from across the world. In preparation for the project, we harmonized individual patient data (IPD) for >12,000 HL patients from 12 clinical trials and 4 large HL registries. Building on data science principles, we established a common data model & unifying data dictionary, which allows us to expand the current database over time, incorporating future data as they become available. We will harness these multi- source data to first, create a modern prediction model via predictive modeling of pre-treatment clinical factors; second, we will estimate and validate disease progression and early emergent late effects, enhanced by the addition of interim PET imaging data and alternative treatment options via multi-state Cox proportional hazards models to estimate transition probabilities; and third, we will generate a dynamic decision model to project precalculated outcomes of interest, such as short term and longer term outcomes, including HRQL, The aggregated, “standardized” precalculated visualizations/projections (life expectance and quality-adjusted life years) will provide a basis for comparing alternative treatments for a broad range of disease subgroups and patient characteristics “on the fly.” Furthermore, both the results of the statistical models & parameter estimates from the simulation model will be validated and re-calibrated as needed in large external cohorts (i.e., German Hodgkin Study Group (GHSG) clinical trials data and Dutch HL registry). The simulation model will then be converted to open source, which will allow local analyses that go beyond our precalculated standard scenarios. Our three-step approach (predictive modeling to multi-state modeling to simulation/decision modeling) will be designed to incorporate future prospective data as new/novel treatments, and knowledge emerges over time. Overall, this project’s results will be significant as they are expected to delineate precise outcomes & optimal therapy for individual HL patients and inform the development of models and paradigms for other diseases.