Supervised Classification with Associative SOM Rafael del-Hoyo 1 , David Buldain 2 , Alvaro Marco 1 Instituto Tecnológico de Aragón, C/ María Luna nº6, 50005 Zaragoza, Spain rdelhoyo@ita.es http://www.ita.es 2 Departamento de Electrónica y Comunicaciones, University of Zaragossa, C/ María Luna 1, 50005 Zaragoza, Spain buldain@posta.unizar.es Abstract. This paper presents an extension of the Self Organizing Map model called Associative SOM that is able to process different types of input data in separated data-paths. The ASOM model can easily deal with situations of incomplete data-patterns and incorporate class labels for supervisory purposes. The ASOM is successfully compared with Multilayer Perceptrons in the incremental classification of six erythemato–squamous diseases, where only partial data is available in successive steps. 1 Introduction Supervised neural models as Multilayer Perceptrons (MLP) and Radial Basis Functions (RBF) have demonstrated in many applications their capability for resolving classification problems. However in a real situation, the necessary information for a classification task often is not completely obtained or is not acquired at the same time. The incomplete data and the great heterogeneity of information are usual situations in real data, a very common problem found in medical diagnosis. The correct diagnosis is a classification problem done in two steps of incremental data acquisition: a compilation of symptoms and, if it is necessary, several laboratory analyses. The first step is critical in many diseases, and physicians usually can decide the medication using only the clinical inspection. However, the second step sometimes is necessary to obtain a correct diagnosis. This second data acquisition is modified when new laboratory tests are discovered and old ones are abandoned. This is a good example of what we call an incremental classification problem. This variability and heterogeneity in the data sources is not well handled by the most known neural models as MLP and RBF networks supervised with the mean squared erro (MSE). An automated classification system must present high flexibility in its structural implementation to deal with the continued apparition of new data sources or the deletion of obsolete ones. A neural classification system should be able to deal with the different data sources in an associative process, where each data source could be easily included or excluded during the classification process. The solution consists