Session P71.4

New Feature Selection Methods for Qualification of Patients for Cardiac Pacemaker Implantation

R Mlynarski*, G Ilczuk, A Wakulicz-Deja, W Kargul

Medical University of Silesia
Katowice, Poland

An important element of life-threatening heart rhythm disorders is pacemaker implantation. The success of implantation depends on the proper classification of patients and the choice of the type of pacing. The Decision Support Method can be very useful in this area.
PURPOSE: to create and analyze feature selection methods for pacemaker type selection (as a key element in our complex decision system).
METHODS: Data of 4316 pts., hospitalized due to heart rhythm disorders, were imported from an information system and transformed for the data mining appliance. Finally, 13 attributes, characterizing the patients’ health status were extracted. These datasets were used as input for the attribute selection methods. In our research we used 2 attribute selection algorithms: CFS and the Chi-square. For synthetic verification our implementation of a rule induction system based on the Rough Set MLEM2 algorithm was used. Practical validation of the selected attributes was done by experts from the Electrocardiology Department by using decision trees (J48 DT C4.5 release 8) - integrated part of our software.
RESULTS are presented for types of pacing as accuracy [%] vs. experts in reduced / unreduced sets: AAI: 81, 7 / 78, 1; VVI: 80, 2 / 76, 4; VDD: 77, 6 / 74, 3; DDD: 75, 1 / 71, 1. We obtained better accuracy for reduced set in shorter time, which was very surprising. High recognition accuracy for VVI type was the result of an over-fitting effect where due to class distribution (noticeable in more patients without VVI pacemaker) the generated decision rules classified more new cases into the non-VVI category. Therefore, synthetic recognition accuracy for this case must be considered carefully. Such effects limit the value of synthetic tests in the medical domain and this is why we always validate the results with domain experts. Decision tree algorithms showed their practical usefulness in helping experts verify results in 100% of the cases.
CONCLUSIONS: 1. A hybrid method which combines the advantages of both algorithms can be an interesting solution for feature selection in the cardiological domain. 2. The usefulness of the presented method in clinical practice was confirmed.

(Abstract Control Number: 111)