Published Biomedical Engineering, 2000,1,16-21 Diagnostic Rule Extraction Using Neural Networks Vitaly G. Schetinin 1 and Anatoly I. Brazhnikov 2 Penza State University, 40, Krasnaya str., Penza, 440017, Russia vschetinin@mail.ru 1 , anibra@yahoo.com 2 The neural networks have trained on incomplete sets that a doctor could collect. Trained neural networks have correctly classified all the presented instances. The number of intervals entered for encoding the quantitative variables is equal two. The number of features as well as the number of neurons and layers in trained neural net- works was minimal. Trained neural networks are adequately represented as a set of logical formulas that more comprehensible and easy-to-understand. These formulas are as the syndrome-complexes, which may be easily tabulated and represented as a diagnostic tables that the doctors usually use. Decision rules provide the evaluations of their confidence in which interested a doctor. Conducted clinical researches have shown that diagnostic decisions produced by symbolic rules have coincided with the doctor's conclusions. 1 Introduction In practice, a doctor-diagnostician applies the diagnostic rules that consist of subjective and objective features (called as symptoms) to accurately distinguish one disease ore state of the patient from others. Subjective features that reflect the complaints, the anamnesis, and the inquiry results of the patient have fuzzy, unquantitable evaluations. In contrast, objective features are the results of laboratory and tool researches that can be represented in quantitative, interval or nominal forms. A doctor interested in that confidence of diag- nostic rules would be maximal. Diagnostic rules should be not only accurate but also un- derstandable for a doctor, which wish to know how these rules work and why their usage brings the best decisions [1, 11]. For extraction and validation of diagnostic rules, a doctor must beforehand collect a representative data set involving the observations of the symptoms that occur in similar clinical cases. In practice, the data set is usually unrepresentative set because it is difficulty to collect a hundred and thousand of examples. Therefore the confidence of decision rules depends first on the size and the quality of data set a doctor classified and second on the structure of symptoms a doctor a prior suggested. In these real-world conditions, we as- sume that a doctor can not exactly evaluate the dividing ability or significance of each of symptoms and because we should estimate the contribution of each symptom to the deci- sion in order to find optimal structure and parameters of desirable diagnostic rules. Fur- ther, the extracted rules are validated on the testing set. If validation results are unsatisfac- tory, the rules extraction process is usually repeated under changed conditions by updating classified set (e.g. removing a contradictory example a user was able to recognize accu- rately, and extending a feature set). The extraction process is repeated until desirable rule of required accuracy would be found. Recently, the machine learning methods have been exploited to extract symbolic rules [2, 3, 16]. In particular, artificial neural networks have been trained to recognize the pathological states. A neural network typically consists of a number of units performing a logical function of formal neurons incorporated in a layer. The inputs to the unit in one layer have connected through weight synapses (synaptic links) with outputs of other units. Accordingly with the connectionist idea, the neural networks are fully connected and lay- ered - an output of each unit in one layer is connected to all inputs to units in the other layer. A neural network that consists of given number of the layers, the synaptic links, and the units is trained to minimize a network error by updating its synaptic weights. A neural