Probabilistic network models for cardiovascular monitoring

Abstract

While treating patients during their hospital stay, physicians must frequently take into consideration massive amounts of clinical data. This data can come in many forms, such as continuous blood pressure tracings, intermittent laboratory results, or simple qualitative observations on the patient's appearance. Although access to such a rich collection of information is beneficial for making diagnoses and treatment decisions, it can sometimes be difficult for clinicians to mentally keep track of everything, especially in hectic environments such as hospital intensive care units (ICUs). In addition, there are certain physiological variables that cannot be measured noninvasively, but are critical indicators of a patient's state of health. One such example in cardiology is cardiac output - the mean flow rate of blood from the heart. In this thesis, we explore probabilistic networks as a method for integrating different types of clinical data into a single model, and as a vehicle for summarizing population statistics from medical databases. These networks can then be used to estimate unobservable variables of interest. We propose and test several networks of varying complexity on both a set of experimental porcine data, and a set of real ICU patient data. We find that continuous estimation of cardiac output is possible using probabilistic networks, and that the errors produced are comparable to those obtained from deterministic methods that employ the same in:Formation. Furthermore, since this technique is purely statistical in nature, it can be easily reformulated for applications where deterministic methods do not exist.