|
Abstract:
|
Accurate risk prediction models for CRC or other relevant neoplastic lesions rely on detailed longitudinal colonoscopic findings. However, to fit seamlessly into available software, polyp-level data are summarized. This results in a potential loss of information and forces investigators to rely on summary covariates. For longitudinal risk prediction models, the summarization choice is particularly challenging. Time is routinely discretized in these models, treating the complex relationship between disease progression and time as simple and distinct. To overcome these limitations, we develop a joint model incorporating a patient’s entire history of colonoscopic findings. We jointly model the longitudinal detection of polyps on colonoscopy and the longitudinal development of advanced adenomas. In the latter outcome we use a continuous time-dependent weighting function to weight contributions of past colonoscopic findings, enabling us to incorporate a patient’s entire colonoscopy record to model the current risk. We test performance of our model using a simulation study and then apply our model to a prospective cohort of asymptomatic Veterans under colonoscopy surveillance.
|
