HYBRID EXPERT SYSTEM FOR DECISION SUPPORT IN THE MEDICAL AREA L. M. Brasil*, F. M. de Azevedo* and J. M. Barreto** * Biomedical Engineering Laboratory (GPEB), Dept. of Electrical Engineering, ** Dept. of Informatics and Statistics Federal University of Santa Catarina Florianópolis, Brazil, 88040-900, Phone (048)331-9594, Fax (048)331-9770 {lourdes, azevedo}@gpeb.ufsc.br, barreto@inf.ufsc.br Abstract: The Knowledge Acquisition (KA) process consists on extracting and representing knowledge of a domain expert. In this work, one of the main goals was to minimise the intrinsic difficulties of the KA process. We have obtained all possible rules from the domain expert in a short time and also a set of examples. Moreover, we developed a Hybrid Expert System (HES) to minimise the problems of the KA task using a new methodology. Building this kind of hybrid architecture has led us to use many tools: symbolic paradigm, connectionist paradigm, fuzzy logic and, Genetic Algorithm (GA). The methodology developed to HES was tested for two cases: toy and real problems (e.g., a medical domain area). Introduction Hybrid architectures for intelligent systems are a new field of Artificial Intelligence (AI) research concerned with the development of the next generation of intelligent systems. Current research interests in this field focus on integrating the computational paradigms of Symbolic Manipulation and Artificial Neural Networks (ANN) [1]. In this work we developed a Hybrid Expert System (HES) to aide in the medical domain, where Knowledge Elicitation task is one of the most hard processes in AI. Materials and Methods The domain expert has difficulty in specifying all rules mainly when imprecision is pervasive to the problem and fuzzy techniques are to be used [2]. In this case, it is often difficult to choose the membership function. Nevertheless, he is able to supply examples of real cases [3]-[8]. So, the knowledge engineer use the rules that were supplied by the domain expert to implement a basic structure of a Neural Network Based Expert System (NNES). After, the NNES is refined through a training algorithm, e.g., Genetic-Back-propagation Based Learning Algorithm (GENBACK) that uses the set of available examples [7]. This algorithm was inspired in the classical Back-propagation one [6]. The NNES also foresees the possibility of different kinds of variables in its input as quantitative, linguistic, or Boolean valued. Another problem with relation to trained ANN is the difficulty that it owns in explaining how it arrived at a solution [6]. We developed a technique for extracting rules of a trained fuzzy NNES called Fuzzy Rule Extraction (FUZZYRULEX) [8]. It was also possible to implement another system, e.g., a Rule Based Expert System (RBES) [7], which provides explanations of the results achieved by the NNES. Steps of the GENBACK Algorithm 1 - Creation - individuals of a population. It was applied to the Gauss distribution; 2 - Codification: chain of chromosomes. It was used a fixed chain of chromosomes with 8 bits; 3 - Training: based on back-propagation algorithm; 4 - Evaluation: used the fitness functions, e.g., F 1 = N CH /ERROR (1) F 2 = 1/(N CH + ERROR ) (2) where F 1 and F 2 = fitness function or objective function or cost function, N CH = neurons number in the hidden layer of the NNES, ERROR = error in the NNES output; 5 - Application of the roulette wheel process; 6 - Ordering: each of individuals in a population; 7 - Application: selection, crossover, mutation; 8 - New generation; 9 - Come back to pass 3: winner NNES. Steps of the FUZZYRULEXT Algorithm 1 – To choose an input-output pattern; 2 – To select those neurons i in the preceding layer that have a positive impact on the conclusion at output neuron j; 3 – Let the set of m i neurons of the hidden layer, so selected, be denoted as {a 1 , a 2 , ..., a mi } and let their connection weights to neuron j in the output layer be given by the set: wet (n)ak = {w (n)ja1 , w (n)ja2 , ..., w (n)jamj } (3) 4 – To determine the set of accumulative link weights wet (n-1)i to neuron i in the hidden layer along the maximum weighted path be as O (n)i > 0 and w (n-1)aki > 0 (4) wet (n-1)i = max[wet (n)ak + w (n-1)aki ] (5) 5 – To select the set of input neurons as m l ={a 1 , a 2 , ..., a ml };