Authors:
Emad Ahmadi
1
;
Hoda Javadi
2
;
Amin Khansefid
3
;
Atousa Asadi
4
;
Mohammad Mehdi Ebadzadeh
5
and
Dirk Timmerman
6
Affiliations:
1
Tehran University of Medical Sciences, Iran, Islamic Republic of
;
2
Iran University of Medical Sciences, Iran, Islamic Republic of
;
3
University of Tehran, Iran, Islamic Republic of
;
4
Amirkabir University of Technology, Iran, Islamic Republic of
;
5
Amirkabir University of Technojavascript:WebForm_DoPostBackWithOptions(new%20WebForm_PostBackOptions("ctl00$ctl00$Main$bodyContent$ManageAuthorsEditEditingEnds$RadGridAuthors$ctl00$ctl18$RadGridNesteEmails$ctl00$ctl04$LinkButtonUpdateEmail",%20"",%20true,%20"2274",%20"",%20false and%20true))logy, Iran, Islamic Republic of
;
6
University Hospitals Leuven, Belgium
Keyword(s):
Decision tree learning, Fuzzy logic, Adnexal diseases/diagnosis.
Related
Ontology
Subjects/Areas/Topics:
Biomedical Engineering
;
Cardiovascular Technologies
;
Clinical Problems and Applications
;
Computing and Telecommunications in Cardiology
;
Health Engineering and Technology Applications
;
Health Information Systems
;
Medical and Nursing Informatics
;
Pattern Recognition and Machine Learning
Abstract:
The study problem was learning a fuzzy decision tree to classify patients with adnexal mass into either of benign or malignant class prior to surgery using patients’ medical history, physical exam, laboratory tests, and ultrasonography. A learning algorithm was developed to learn a fuzzy decision tree in three steps. In the growing step, a binary decision tree was learned from a dataset of patients while fuzzy discretization was used in decision nodes testing continuous attributes. The best degree of fuzziness was automatically found by an algorithm based on optimization procedures. In the pruning step, the overfitted nodes were removed by an algorithm based on critical value post-pruning method. In the refitting step, the labels of the leaf nodes were optimized. The final resulted tree had 10 decision nodes and 11 leaf nodes. Performance testing of the tree gave AUC of ROC of 0.91 and mean squared error of 0.1. The tree was translated into a set of 11 fuzzy if-then rules and the cli
nical plausibility of the rules was assessed by domain experts. All rules were verified to be in agreement with medical knowledge in the domain. Despite the small learning set and the lack of some important input variables, this method gave accurate and, more importantly, clinically interpretable results.
(More)