Kidney obstruction is a serious disease which can lead to loss of renal function when not treated in a timely manner. Diuresis renography is widely used to detect obstruction in kidney. However, the diagnosis largely relies on experts' experiences, and there is no gold standard statistical approach designed to analyze renogram curves and clinical variables associated with patients. In this work, we propose an integrative Bayesian approach that models the triplet jointly: renogram curves, clinical variables of patients, and experts' ratings, conditional on the latent kidney obstruction status. In particular, we adopt a nonparametric approach for modeling renogram curves in which the coefficients of the basis functions are parameterized using latent factors that are dependent on the latent disease status. We develop an MCMC training algorithm and an associated prediction algorithm for kidney obstruction that are computationally efficient. We demonstrate the superior performance of our proposed method in comparison with several naïve approaches via extensive simulations as well as analysis of real data collected from a kidney obstruction study.