Using principal component analysis to aid bayesian network development for prediction of critical care patient outcomes.

Background: Predicting an intensive care unit patient's outcome is highly desirable. An end goal is for computational techniques to provide updated, accurate predictions about changing patient condition using a manageable number of physiologic parameters.

Methods: Principal component analysis was used to select input parameters for critical care patient outcome models. Vital signs and laboratory values from each patient's hospital stay along with outcomes ("Discharged" vs. "Deceased") were collected retrospectively at a Level I Trauma-Military Medical Center in the southwest; intensive care unit patients were included if they had been admitted for burn, infection, or hypovolemia during a 5-year period ending October 2007. Principal component analysis was used to determine which of the 24 parameters would serve as inputs in a bayesian network developed for outcome prediction.

Results: Data for 581 patients were collected. Pulse pressure, heart rate, temperature, respiratory rate, sodium, and chloride were found to have statistically significant differences between Discharged and Deceased groups for "Hypovolemia" patients. For "Burn" patients, pulse pressure, hemoglobin, hematocrit, and potassium were found to have statistically significant differences. For a "Combined" group, heart rate, temperature, respiratory rate, sodium, and chloride had statistically significant differences. A bayesian network based on these results, developed for the Combined group, achieved an accuracy of 75% when predicting patient outcome.

Conclusions: Outcome prediction for critical care patients is possible. Future work should explore model development using additional temporal data and should include prospective validation. Such technology could serve as the basis of real-time intelligent monitoring systems for critical patients.