Associative Memories in Medical Diagnostic

Neural networks are used as associative memories to build an expert system for diagnostic. Similarly to expert systems implemented using symbolic manipulation, here the knowledge is introduced by a knowledge engineer using a collection of known cases. Fuzzy sets are used as interpretation for connection values and/or excitation state of units. The main result is that the proposed neural network allows not only to find a solution in some cases, but also to suggest to obtain more clinical data if the data available is insufficient to conclude. To illustrate the approach the case of two diseases with similar symptoms (difficult diagnostic) is used. INTRODUCTION Traditional expert systems based on two value logics, are often not very well suited to deal with real world problems, due to its inconsistencies, exceptions to the rules and incomplete specifications. Medical diagnosis involves these kinds of difficulties and more: a good diagnostic is often achieved by similarity with previous studied cases, where imprecision is pervasive. It is possible to identify two types of domains where an expert system can actuate: man made (artificial) domains and natural domains. Examples are respectively computer aided design (XCON specifies the configuration of computers) and medical diagnosis. In the first case the functioning of the system is known because it is man made. In the other the functioning is known only partially, as result of research and most knowledge derives from particular cases observed. If in the first case, rules of type IF...THAN...ELSE are natural, in the other, they are not: it is necessary to extract the structure of the causal reasoning from examples and to arrive at conclusions by analogy. As pointed out in [Barreto, 1990], neural networks can be considered as a good programming paradigm in these ill defined cases, where the knowledge is available by examples and inferences taken by analogy. Some previous neural networks were used in the medical diagnostic field. Without being exhaustive, it is possible to mention the work of Gallant (1988) that considered 6 symptoms and two diseases and using a feedforward network showed how explanations can be obtained. He trained the network to learn associations between symptoms and diagnostic. After presenting a set of symptoms it is possible to get the corresponding diagnostic This approach was followed by some other researchers, for example [Peng, 1989] and [Kosho, 1987]. Here a different approach is followed. The paradigm of associative memories in medical diagnosis is considered. The available cases are represented directly by weights of connections between units representing symptoms, diseases and names of the corresponding patients. The intensity of connections is a number in the interval [0 1] representing the fuzzy value [Zadeh, 65] of importance of symptom or degree of illness. A consultation is done by exciting a particular element and corresponding symptoms. § Presently in “Institut d’Informatique, FUNDP, Namur, Belgium, sponsored by CNPq (Conselho Nacional de Desenvolvimento Científico e Tecnológico”), Brazil. §§ Presently in "Clinique Universitaire Mont Godine" UCL, Belgium, sponsored by UNIRIO in Brazil. 10th Int. Cong. on Medical Informatics 2 To test the approach a difficult case was chosen: a case where it is difficult even for a physician to arrive easily to a conclusion. STRUCTURE OF THE NEURAL NETWORK There are several good references on neural networks and their possibilities. Arbib (1989) presents the main basic concepts and the physiological motivation. Rumelhart & all. (1986) contains several classical works on the subject. Here we limit ourselves to the presentation of the used network. The network (Fig.1) has three pools representing respectively the diseases, the symptoms and particular cases known. This knowledge and the convenient values of connections constitutes the knowledge base with which inferences about a particular patient will be done. The units of all three pools are supposed visible. All the connections are bi-directionals, excitatory or inhibitory, with weights varying between 0 and 1 and representing the fuzzy degree or the relation between the concepts represented by the units. It is then similar to the network consided in [Kosko,1987]. Between the elements of the diseases pool there are inhibitory connections. As it is possible to have a patient with two diseases no connection is completely inhibitory (weight -1) and diseases that occur often simultaneously have low inhibitory connection between them. Between the units of the patients there is maximum inhibition and this means that each patient is an isolated case. This leads to a competition and the system is driven to find the more similar case to the case whose diagnostic is desired. In the pool of symptoms, as they are not mutually exclusif, the connections are set to 0 and this represents ignorance. There is no hidden units. The particular case used in this work to exemplify and study the approach considers two diseases, 14 symptoms and 12 known cases. • The diseases are: Rheumato Arthritis Adult Form (RA) and Systemic Lupus Erythematosus (SLE). RA is a chronic inflammatory disease of unknown ethiology. It is a common disease affecting about 1.5% of the population in North America. It may occur at any age, but most often strikes between the ages of 20 and 60, with a peak incidence in women from 40 to 60. Women are affected more frequently than men, the ratio being about 3:1. SLE is a chronic, inflammatory disease of unknown cause affecting skin, joints, kidneys, nervous system and often other organs of the body. This one, if less frequent than RA, affects essentially the same population. As they have similar symptomatology in the initial phase of disease development, a correct diagnosis is difficult if based exclusively in clinic data, except if symptoms characteristic and specific of each disease is present. However these specific symptoms appear generally in a latter phase. However the use of laboratory examinations generally allows to distinguish between them. • The following symptoms were considered: Fever, Arthralgia, Arthritis, Morning Stiffness, Myalgia, Subcutaneous Nodules, Butterfly Rash, Raynaud's Phenomenon, Photosensitivity, Alopecia, Renals Manifestations, Central Nervous System Manifestations, Pulmonary Manifestations and Rheumatoid Hand. No speculation about possible relation between symptoms is made • As set of known case 12 real cases were considered, 6 having being diagnostically as RA and 6 as SLE. The data described was used to build a synaptic matrix representing the relations between units of a pool and between two different pools. In these synaptic matrix, the membership values representing the several fuzzy 10th Int. Cong. on Medical Informatics 3 membership were considered, expressing a clinical classification. All connections were symmetric and no unit connected to itself. this implies a symmetric matrix with null elements in the main diagonal. The choice of this matrix is interesting because, as shown in [Cohen & Grossberg, 83], if wij = wji, i ≠ j and wii = 0, the equilibrium points of the neural network are stable. RESULTS To illustrate the possibilities of the approach have two examples are presented. They were programmed taking as starting point the collection of programs in McClelland & all., (1988). Example 1: The goal of this example is to show the structure of the data. Fig.2 shows the result obtained after 40 cycles when the subject of a particular case is excited. Fig.2 Final screen for a particular case of the data base It is possible to remark: • In the data base “Pat” is associated with Arthralgia, Raynaud's Phenomenon, Alopecia e Butterfly Rash. However, the result indicated by the stable point found by the neural network indicate other symptoms of the universe of considered symptoms that are not related with this disease (negative excitationreverse video). • In the pool of the known cases three other cases are excited, Sue, Mag & Beth. This indicate a similar symptomatology, and the degree of excitation is the fuzzy membership value of this degree of similarity. • The diagnosis given by the network is really an stable point. In fact, Fig.3, showing the evolution of the activation of the neuron representing the illness indicates clearly that the steady state was attained.