Personalized risk prediction using longitudinal data: Applications in obstetrics and cancer

Friday, November 9, 2018 - 9:00am - 9:50am
Lind 305
Paul Albert (National Cancer Institute)
Assessing disease risk with repeated biomarker measurements has important applications in the screening and early detection for many diseases. In obstetrics, interest is on assessing the risk of a poor pregnancy outcome (i.e., small for gestational age or preterm birth) from longitudinal imaging and biomarker data. We will begin by presenting simple two-stage estimation procedure that approximates a full maximum-likelihood approach for predicting a binary event from multivariate longitudinal growth data using a shared random parameter model (Albert, Statistics in Medicine, 2012). We will present a pattern mixture approximation to this model (Liu and Albert, Biostatistics, 2014). Subsequently, we will present a class of joint models for multivariate growth curve data and a binary event that accommodates a flexible skewed error distribution for the ultrasound measurements and an asymmetric link function relating the longitudinal to the binary process (Kim and Albert, Biometrics, 2016). We will also present a tree-based approach for identifying subgroups of women who have an enhanced predictive accuracy for predicting a binary event from fetal growth data (Foster, et al., JRSS-A, 2016). Finally, we will discuss applications of these methods for early detection of cancer.