Computers in Biology and Medicine 37 (2007) 1617 – 1628 www.intl.elsevierhealth.com/journals/cobm Neuro-fuzzy classification of prostate cancer using NEFCLASS-J Ayturk Keles a , ∗ , A. Samet Hasiloglu b , Ali Keles c , Yilmaz Aksoy d a Department of Computer Engineering, Faculty of Engineering, Ataturk University, 25240-Erzurum, Turkey b Department of Computer Engineering, Faculty of Engineering, Ataturk University, 25240-Erzurum, Turkey c Department of Computer Education and Instructional Technology, Faculty of Kazim Karabekir Education, Ataturk University, 25240-Erzurum, Turkey d Department of Urology, Medical Faculty, Ataturk University, 25240-Erzurum, Turkey Received 15 July 2005; received in revised form 12 March 2007; accepted 21 March 2007 Abstract Medical diagnosis has been the most proper area for the implementations of artificial intelligence for approximately 20 years. In this paper, a new approach based on neuro-fuzzy classification (NEFCLASS) tool has been presented to classify prostate cancer. The tool has the features of batch learning, automatic cross validation, automatic determination of the rule base size, and handling of missing values to increase its interpretability.We have investigated how good medical data analysis could be done with NEFCLASS-J, and what effects selected parameters have on classifier performances. Medical data were obtained from patients with real prostate cancer and benign prostatic hyperplasia (BPH). The reason for the selection of these two illnesses was the fact that their symptoms are very similar yet their differentiation is very crucial. The results showed that, for creating high performance of classifier appropriate for the data used, firstly it is necessary to decide well on the membership type and the number of fuzzy sets and then validation procedure. After a good classifier has been found, other parameters should be investigated to improve this classifier. In the light of this study, we can present a foresight for the diagnosis of the patients with prostate cancer or BPH.  2007 Elsevier Ltd. All rights reserved. Keywords: Medical diagnosis; Data analysis; Neuro-fuzzy classification; Prostate cancer; Benign prostatic hyperplasia 1. Introduction Prostate cancer is one of the most common causes of death in men in most industrialized countries. The diagnosis of in- tra capsular prostate cancer is life saving because in this stage it is curable. However, differentiating it from benign prostatic hyperplasia (BPH) is difficult. Medical examinations include rectal examinations and laboratory tests such as the conven- tional prostate-specific antigen (PSA), ultrasound and biopsy. Our modern world is data-driven as many decisions are given based on the analysis of data. One important typical applica- tion area is medical diagnosis. Models obtained from data anal- ysis that are applied in practice usually require transparency and interpretability in terms of the attributes they process. This also requires small models, because models with many param- eters are not comprehensible to a user. An important aspect of ∗ Corresponding author. Tel.: +90 442 2314740; fax: +90 442 2360957. E-mail address: ayturkk@hotmail.com (A. Keles). 0010-4825/$ - see front matter  2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.compbiomed.2007.03.006 intelligent data analysis is to select an appropriate model with the application in mind. It may be necessary to sacrifice preci- sion for interpretability, i.e., a suitable balance between model complexity and comprehensibility, between precision and sim- plicity must exist. Intelligent data analysis also requires the se- lection of appropriate algorithms for the process of creating a model. There can be several algorithms available for creating the same kind of model and they may only differ not in computa- tional complexity, speed of convergence, ease of parameteri- zation, but also in the way they ensure certain features in the model they create from data. In order to use fuzzy systems in data analysis, it must be possible to learn them from examples. Learning in fuzzy systems are implemented most often by learning techniques derived from neural networks. The term neuro-fuzzy system (also neuro-fuzzy methods or models) refers to combinations of neural networks and fuzzy systems. This combination does not usually mean that a neural network and a fuzzy system are used together in some way. A neuro-