Prediction tree for severely head-injured patients

Abstract

✓ Prediction tree techniques are employed in the analysis of data from 555 patients admitted to the Medical College of Virginia hospitals with severe head injuries. Twenty-three prognostic indicators are examined to predict the distribution of 12-month outcomes among the five Glasgow Outcome Scale categories. A tree diagram, illustrating the prognostic pattern, provides critical threshold levels that split the patients into subgroups with varying degrees of risk. It is a visually useful way to look at the prognosis of head-injured patients. In previous analyses addressing this prediction problem, the same set of prognostic factors (age, motor score, and pupillary response) was used for all patients. These approaches might be considered inflexible because more informative prediction may be achieved by somewhat different combinations of factors for different patients. Tree analysis reveals that the pattern of important prognostic factors differs among various patient subgroups, although the three previously mentioned factors are still of primary importance. For example, it is noted that information concerning intracerebral lesions is useful in predicting outcome for certain patients. The overall predictive accuracy of the tree technique for these data is 77.7%, which is somewhat higher than that obtained via standard prediction methods. The predictive accuracy is highest among patients who have a good recovery or die; it is lower for patients having intermediate outcomes.