USING DATA PREPROCESSING AND SINGLE-LAYER PERCEPTRON TO ANALYZE LABORATORY DATA

During daily work in hospitals a large amount of clinical data is prod uced each day. Totally computerized patient records are not yet widely used but a large part of essential information is already stored on c omputer files. These include laboratory test results, diagnoses, codes for operations, codes of histopathological diagnoses and maybe even t he patient's medication. Accordingly, these databases include much cli nical knowledge that would be useful for clinicians. Laboratories try to support clinicians by producing reference values for laboratory tes ts. It is, of course, necessary information but, however, it does not give very much information about the weight of evidence that an abnorm al laboratory test will give in special clinical settings. We have dev eloped a software package - DiagaiD - in order to build a smart link b etween patient databases and clinicians. It utilizes neural network-ba sed machine learning techniques and can produce decision support which meets the special needs of clinicians. From example cases it can lear n clinically relevant transformations from original numeric values to logical values. By using data transformation together with a single la yer perceptron it is possible to build nonlinear models from a set of preclassified example cases. In this paper, we use two small datasets to show how this scheme works in the diagnosis of acute appendicitis a nd in the diagnosis of myocardial infarction. Results are compared wit h those obtained using logistic regression or backpropagation neural n etworks. The performance of our neuro-fuzzy tool seemed to be slightly better in these two materials but the differences did not reach stati stical significance.