Center for Biomedical Informatics in Critical Care (C-BICC)
Norman Aleks, Fei Sha1, Stuart J. Russell and Geoffrey Manley2
CBICC is a joint project between UC Berkeley and UC San Francisco, funded currently by Intel and UC Discovery. The aim is to develop new methods of collecting and processing sensor data to improve clinical outcomes in intensive care.
It has become possible to collect a wide range of physiological measurements in the intensive care unit (ICU). It has not yet proved possible to take full advantage of all these measurements, for two main reasons. First, noise and sensor failures make data analysis by traditional means difficult, cause false alarms, and sometimes lead to treatment errors. Second, because much of the important patient state is not directly measurable, a sensor snapshot is insufficient for reliable decisions; observing all sensor measurements over a substantial period can compensate for this to a large extent, but this gives far too much data for a human to interpret. This problem is exacerbated by the fact that the underlying physiological dynamics is complex and varies from patient to patient.
To address these problems, we propose to develop probabilistic models of patient physiology and ICU sensors in the form of dynamic Bayesian networks (DBNs) [1-3] and to use these models to obtain more reliable estimates of the true patient state. In the project's first phase, we will develop relatively simple physiological transition models with parameters fitted by statistical methods applied to the datasets obtained from traumatic brain injury patients at San Francisco General Hospital. We will test whether these models can reduce the rate of false alarms and overcome problems with noise and missing data. Our initial results are promising: even our initial "toy" models (like the one shown here) combine knowledge of physiology and sensor function at a level that comparable prior work has not , and our model's inferences appear to be of good quality.
In the project's second phase, we will develop more sophisticated physiological models that allow for pathophysiological states and for uncertainty in the model parameters. We will demonstrate the ability of the monitoring software to identify such states and to adapt these parameters to the individual patient in real time, as sensor data arrives. We will show that this adaptation process leads to better out-of-sample prediction and improved ability to estimate patient state and hence better clinical decisions. Finally, we will investigate algorithms for modifying the DBN structure  by adding hidden state variables that "explain" divergent patterns of physiological behavior in patients--potentially yielding new insights into physiological processes in ICU patients.
Figure 1: A dynamic Bayesian network modeling true current heart rate, two of its monitors, and three sources of monitor artifact, with an atemporal variable for the patient's resting/baseline heart rate. The transition model for true-HR is defined as a linear Gaussian: at time 0, it is centered at 80 with standard deviation 30, and at other time steps t it is centered at a weighted sum of true-HR(t-1) and resting-HR, with standard deviation 5. The sensor models for ECG-HR and pleth-HR are defined with Gaussians centered at true-HR, or at 0 if disconnected; their standard deviations range from 3 (for the ECG on a still patient) to 20 (for the pleth on a moving patient). The sensor model for SpO2 is a Gaussian centered at 97 if it is attached to a non-moving patient, at 80 if it is attached to a significantly moving patient, and at 0 if it is detached. The transition models for ECG-detached, pleth-detached, and patient-movement assign a high probability to persisting in the current state.
Figure 2: The model of Figure 1 applied to actual monitor data collected from a healthy 40-year-old male volunteer in the SFGH ICU. The volunteer was still for minutes 1 through 4, made moderate hand movements minutes 5-10, was still again minutes 11-14, thrashed on the bed minutes 15-19, and was still again through the end of the recording. Note that the small hand movements were insufficient to perturb the data or cause inference of movement artifact, but that movement artifact was inferred during the thrashing phase. Also note that the final resting heart-rate estimate was in the low 80's, which is correct for this person.
- T. Dean and K. Kanazawa, "A Model for Projection and Action," Proc. Int. Joint Conference on Artificial Intelligence (IJCAI-89), pp. 985-990, Detroit, MI, Morgan Kaufmann, 1988.
- T. Dean and K. Kanazawa, "A Model for Reasoning and Persistence and Causation," Computational Intelligence, Vol. 5, No. 3, pp. 142-150, 1988.
- S. J. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, Second Edition, Upper Saddle River, NJ, Prentice Hall, 2003.
- C. K. I. Williams, J. Quinn, and N. McIntosh, "Factorial Switching Kalman Filters for Condition Monitoring in Neonatal Intensive Care," Advances in Neural Information Processing Systems, Vol. 18, ed. Y. Weiss, Y. Schölkopf, and J. Platt, pp. 1513-1520, Cambridge, MA, MIT Press, 2005.
- N. Friedman, "The Bayesian Structural EM Algorithm," Proc. Conf. Uncertainty in Artificial Intelligence, Madison, WI, Morgan Kaufmann, 1998.
2Neurosurgery, San Francisco General Hospital