Bayesian Data Augmentation for Recurrent Events in Electronic Medical Records of Patients with Cancer

点击翻译按钮获取中文摘要

PROJECT SUMMARY/ABSTRACT Cancer and its treatment frequently result in sequelae that are not pro-actively reported but rather intermittently assessed. For example, patients with cancer experience a higher rate of falls compared to the general population, but falls are commonly reported only when elicited. Although electronic medical record (EMR) databases capture these elicited reports, the assessments are intermittent and their intervals often overlap, e.g., if patients are asked “have you fallen within the last three months” at two outpatient visits one month apart. However, current methods for analyzing event counts within intervals when exact event times are unknown (“interval count data”) require assessment intervals to be non-overlapping. This project addresses this critical gap by developing Bayesian statistical methods and software for analyzing interval count data with overlapping intervals, as arise from fall reports and other intermittently assessed EMR data. These methods apply a Gibbs sampler in which one step uses Bayesian data augmentation to impute full event histories (including event times) to which other steps may apply a broad toolkit of models for more fully observed recurrent event data. In Aim 1 we will develop and apply Bayesian data augmentation for intermittently assessed recurrent events following Poisson processes. Event histories will be imputed using specialized rejection samplers optimized for high computational efficiency in our data setting. In Aim 2 we will develop and apply Bayesian data augmentation for intermittently assessed recurrent events following renewal processes. Event histories will be imputed using random walk samplers with specialized perturbation proposals. The performance of both methods will be assessed via simulation study, and an R software package will be developed and distributed to CRAN. In both aims we will evaluate incidence and risk factors for falls via EMR data from an NCCN comprehensive cancer center using the developed statistical methods. Through our proposed Bayesian data augmentation approach and software developed, this project will provide uniquely capable and innovative tools to integrate clinical, demographic, and recurrent outcome data as commonly recorded in EMR databases to assess incidence and risk, allowing for risk-stratified interventions. The tools for recurrent events, though originally conceived of to address falls among patients with cancer and survivors, will be broadly applicable to both other types of patient-reported sequelae such as occurrences of nausea and pain, and other health-related fields that collect recurrent event data in EMR databases.