Abstract 5761: Bayesian estimation of a semi-parametric recurrent event model with competing outcomes for personalized risk prediction among cancer survivors

Abstract

Abstract Multiple primary cancer (MPC) is becoming more common in the general population, accounting for approximately 20% of incidence cases in the United States. Epidemiological studies indicate a strong dependence of the subsequent primary on the characteristics of the first primary. For example, there is evidence that individuals with a first primary breast cancer is more likely to develop a second primary lung cancer. These observations highlight the need to develop a statistical model that characterizes age-to-onset of cancers beyond the first primary, while accounting for the complex relationships between the cancer occurrences. Such a modeling framework was previously lacking. We hereby propose a Bayesian semiparametric framework, where the occurrences of each cancer type follow a non-homogeneous Poisson process. The time-varying intensity of this process is conditioned on genetic and demographic covariates, such as status of genetic mutation and sex, as well as on a patient’s cancer history, such as type and timing of the first primary, thus allowing our model to capture the heterogeneity in cancer risks across individuals. Since our model requires a dataset that is enriched with MPC cases, we utilize data collected from families affected with Li-Fraumeni Syndrome (LFS), which is a genetic disorder characterized by germline mutations in the tumor-suppressor gene TP53. People with LFS are at higher risks of certain cancer types, and many cancer survivors develop additional primary malignancy. We train and cross-validate our model on a patient cohort selected according to clinical LFS criteria at MD Anderson Cancer Center from year 2000 to 2015. The cohort consists of 11,186 individuals across 429 families, out of which 2,286 were diagnosed with at least one primary cancer and 335 were tested positive for germline mutations in TP53. Upon model training, we construct cancer-specific penetrance curves for the second primary cancer, which vary considerably among patients with different covariates and cancer history, thus highlighting the utility of our model for personalized risk prediction. Our penetrance estimates display good performance when being used to make cancer-specific predictions of the second primary among cancer survivors, achieving AUCs (Areas under the Receiver Operating Characteristic curves) of 0.91, 0.76 and 0.68 respectively for sarcoma, breast cancer, and all other cancers combined. While we apply our model to an LFS dataset, the statistical framework is general, and can be tailored to analyze any time-to-event datasets with suitably selected sets of covariates. Future applications include population-based characterization of specific cancer type combinations over time to impact public health policy making. Citation Format: Nam H. Nguyen, Elissa B. Dodd-Eaton, Seung Jun Shin, Jing Ning, Wenyi Wang. Bayesian estimation of a semi-parametric recurrent event model with competing outcomes for personalized risk prediction among cancer survivors. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 5761.