Research Abstracts - 2006

Tracking Hemodynamic Changes in Intensive Care Patients

William Long & ShuChyng You

Problem

Patients in the intensive care unit typically have continuous monitoring of their heart rate, blood pressure, and often pulmonary and central venous pressures as well to enable the clinicians to track and recognize the important changes taking place in their hemodynamic state. Current monitoring equipment allows alarms to be given when individual parameters are too high or too low but it is up to the clinician to integrate the mountain of data to determine what is happening in terms of the pathophysiologic mechanisms producing these data and therefore what response is appropriate. What is needed is a way of automatically integrating the data using a hemodynamic model of the cardiovascular system identifying changes in the changes in the hemodynamic state such as cardiogenic or septic shock rather than just bounds checks on parameters.

Approach

We have a simple model of the cardiovascular system that has been used to estimate the effects of common cardiovascular medications in a variety of pathophysiologic states. It relates the measurable parameters of the system and can approximately represent the state of a patient given measurements that are often available in a monitored patient. The model has been successful in estimating the effects of several medications in patients with three different valvular diseases by comparing the program to published studies in the medical literature.

The problem in the intensive care unit is somewhat different from before and after data from a paper. The patients often have more than one problem and the data is noisier. On the other hand, we have more data. We are concentrating on the data verified by a nurse entered at intervals of 15 minutes up to an hour, but even that is noisy. The first step is to verify that the model adequately captures the important physiologic mechanisms causing the changes in the patients. To do this we are using the patient data that has been collected to characterize the model, first section by section and then as a whole. The challenge is to distinguish among model parameters that are constant across patients, those that vary from patient but can be individualized with a few measurements, those that reflect unmodeled but important mechanisms, biases in measurement techniques, and noise due to lack of precision in the sensors.

We have all of the monitoring data on about 2,000 patients seen in several intensive care units with a variety of conditions and a variety of different treatments. The have developed several tools to help in the process of picking patients and time periods to test the model. We need patients on a small number of well characterized medications with known pathiophysiologic states appropriate to the part of the model being tested. The cardiovascular model assumes the patient goes from one steady state to another and does not try to model the transients in between, so we need data sufficiently after any medications that the system has stabilized. Fortunately, most of the medications used in the intensive care unit act rapidly so this is not usually a problem.

We have the tools necessary to identify appropriate sets of data are just starting to determine how well different parts of the model characterize the data. For example, according to the model, changes in the blood pressure cause changes in the heart rate through a combination of vagal and sympathetic stimulation. These are modified by a number of medications plus exercise, but without changes in these factors, the relationship should be constant. We are now accumulating data to test this.