Probabilistic modeling of kidney dynamics for renal failure prediction

Abstract

The large quantity of clinical data collected from the Intensive Care Unit (ICU) has made clinical investigation by a data-driven approach more effective. In this thesis, we developed probabilistic models for modeling variable kinetics and temporal dynamics of states. We applied the models to the prediction of acute kidney injury (AKI), but the models are applicable to other medical conditions as well. It is known that serum creatinine follows first-order clearance kinetics. We developed a stochastic kinetic model for first-order clearance and used it to model creatinine kinetics. Some properties implied by the model that are verifiable with the available data are consistent with the empirical results. Those properties are mean-reversion, variation with linear standard deviation, and convergence of variance to a finite value. Based on the stochastic kinetic model, creatinine can be treated as a lognormal random variable with state-dependent parameters. We model the temporal dynamics of kidney states and creatinine using a Hidden Markov Model. Observations of creatinine are assumed to be random variables, with baseline creatinine as mean. Each individual baseline is itself a random variable sampled from a population distribution. Baseline for each patient can be estimated by combining the population distribution and all creatinine observations of the patient using techniques similar to Bayesian inference. Prediction of acute kidney injury with this generative model gives an AUC of 0.8259 and 0.8497 for female and male population respectively.