Session P92.3

Visualization of Decision Rules: From the Cardiologist's Point of View

A Wlodyka*, R Mlynarski, G Ilczuk, E Pilat, W Kargul

Upper Silesian Cardiology Centre
Katowice, Poland

A lot of decision systems work internally using different forms of decision rules. In our experiments on large medical datasets, we found that when the number of conditions in a decision rule increases and the overall number of rules is greater than 20-50, it is really difficult to analyze and manage the stored knowledge. In this situation, methods that allow the visualization of the induced decision rules are of great importance. Our research concentrated on two methods of the visualization of decision rules: decision trees (AQDT-2 algorithm) and the so-called rule-diagrams, which present conditional parts of decision rules in a 3D matrix (2D layers are stacked in 3D cube). But the question is which of these methods is the best from a cardiologist's point of view.
Methods: Sets of decision rules from 5 425 medical records were generated (our implementation of rough set based MLEM2) for 3 decision attributes: a decision about pacemaker implantation (15 rules), a decision about implantation of DDD pacemaker (21 rules) and a decision about B-blocker treatment (70 rules). These rules were generated and afterwards visualized using both decision trees and rule-diagrams. The results were validated by 3 experienced cardiologists and 1 allied professional using the following two criteria: ease of understanding and interactive reasoning.
Results: Experts agreed that decision trees present an attractive possibility of data visualization for small sets of rules (up to 100). They are intuitively understandable and with our own extensions, they provide a quick method for checking several what-if scenarios. For larger sets of rules (more then 100) rule-diagrams are definitely the better methods for analyzing patterns within data. The visualization of rules containing attributes with a large number of possible distinct values is an especially strong advantage of rule-diagrams. Nevertheless they require a learning process at the beginning.
Conclusions: In the case of mid-range and large medical databases, rule-diagrams seems to be more suitable for the task, whereas, in the case of small sets of rules when compared to the larger sets, decision trees were shown to have more advantages.

(Abstract Control Number: 301)