Detection of Ischemic Episodes with a combination of unsupervised and supervised learning S.Papadimitriou, L. Vladutu, S. Mavroudi , A. Bezerianos Department of Medical Physics, School of Medicine, University of Patras, 26500 Patras, Greece, tel: +30-61-996115, fax: +30-61-992496 email: bezer@patreas.upatras.gr Summary – The paper presents a novel neural network architecture for the effective detection of ischemic episodes. This architecture combines unsupervised learning for the simple regions and supervised for the difficult ones in a two stage learning process. The unsupervised learning approach extends and adapts the Self-Organizing Map (SOM) algorithm of Kohonen. The supervised learning has the objective of improving the decision boundaries at some parts of the state space (i.e. at the ambiguous regions). This combination of learning paradigms allows to realize the classification performances of advanced supervising learning approaches with the use of significantly less computational resources. Keywords –Myocardial Ischemia, Self-Organizing Maps (SOM), Principal Component Analysis, Radial Basis Functions (RBF), Support Vector Machines (SVM). I. INTRODUCTION In this paper a new neural network architecture for the classification of ischemic episodes is presented. This architecture combines unsupervised and supervised learning in a two stage learning process. The unsupervised learning is based on a modified Self Organizing Map (SOM) algorithm of Kohonen [1,3] with dynamic insertion/deletion of nodes in a bidimensional lattice of neurons. This unsupervised algorithm performs well at discriminating the ischemic from the normals over a large part of the state space while at the same time is computationally efficient. However, there remain some parts of the state space that require the enforcement of complex decision boundaries. These parts are detected at the unsupervised phase and a supervised learning paradigm is explored in order to cope with the difficult regions of the state space. Since the new architecture combines the modified SOM algorithm with supervised neural network models is called the Network Self-Organized Map (NetSOM).