The limitations of decision trees and automatic learning in real world medical decision making.

Abstract

The decision tree approach is one of the most common approaches in automatic learning and decision making. The automatic learning of decision trees and their use usually show very good results in various "theoretical" environments. But in real life it is often impossible to find the desired number of representative training objects for various reasons. The lack of possibilities to measure attribute values, high cost and complexity of such measurements, and unavailability of all attributes at the same time are the typical representatives. For this reason we decided to use the decision trees not for their primary task--the decision making--but for outlining the most important attributes. This was possible by using a well-known property of the decision trees--their knowledge representation, which can be easily understood by humans. In a delicate field of medical decision making, we cannot allow ourselves to make any inaccurate decisions and the "tips," provided by the decision trees, can be of a great assistance. Our main interest was to discover a predisposition to two forms of acidosis: the metabolic acidosis and respiratory acidosis, which can both have serious effects on child's health. We decided to construct different decision trees from a set of training objects. Instead of using a test set for evaluation of a decision tree, we asked medical experts to take a closer look at the generated trees. They examined and evaluated the decision trees branch by branch. Their comments show that trees generated from the available training set mainly have surprisingly good branches, but on the other hand, for some, no medical explanation could be found.