MODELS BASED ON CONFORMAL PREDICTORS FOR DIAGNOSTIC SYSTEMS IN MEDICINE

A disadvantage of many diagnostic systems is the inability to sufficiently assess the decisions reliability. While solving the problem of classification, each example may be classified with different degree of quality. So, a measure of the quality of an example classification was used (a non-conformity measure). The goal of the research is to improve evaluation of the diagnostics reliability in medicine based on conformal predictors which allow carrying out a probabilistic classification, as well as identifying abnormal cases when either the classifier is unable to determine the class for a particular object, or assigns one object to several classes at once. The paper describes the constructing and testing of various probabilistic binary classification models based on machine learning, particularly, the SVM method and conformal predictors using a non-conformity measure. For learning and testing the medicine Breast Cancer Wisconsin (Diagnostic) Data Set was used to construct linear, polynomial of different degrees and RBF models. We assessed the prediction results for every example from the test set as well as the integral characteristics of the quality of the models, taking into account both the correctness of the predictions for each class and the number of different types of anomalies. On the basis of the best selected models (linear, polynomial model of the 2nd degree and RBF), we developed an intelligent diagnostic system in medicine, which allows automating the model's construction, as well as carrying out the diagnostics and displaying the confidence of the received diagnosis or a message about the impossibility of making a diagnosis. The program also allows multiple doctors to log in to the system, adding new patients and editing information about them; every patient has their medical record with the results of the examination and the diagnoses given. The results of the research can be applied in the diagnostic systems for various diseases. This can be done by using the data with the symptoms and the corresponding diagnoses and constructing the appropriate models on this basis.