Machine Learning Methods in Interpretation of Results of Stress Myocardial Perfusion Scintigraphy
Coronary artery disease (CAD) is the main cause of (premature) mortality in Slovenia and elsewhere in developed countries. The CAD diagnostic process is stepwise. An increase in the accuracy of stress myocardial perfusion scintigraphy (SMPS), one of the main steps in the diagnostics of CAD, would lead to an earlier diagnosis of CAD, while at the same time a decreased need for further diagnostic process would increase its availability. The decision making capability of machine learning methods, which are capable of independent decisions after initial learning on cases with implemented solutions, already achieves and in some fields even exceeds that of the humans. Our study is an attempt to use machine learning methods in interpretation of SMPS results in CAD diagnostics under the supposition that this could increase the diagnostic accuracy of this diagnostic method. Our intention was to choose and to learn adequate machine learning method to make an expert system. In addition, we wanted to objectify a value of its application in interpretation of SMPS results. Our study encompassed the patients who were treated during the course of the diagnostic process of CAD at the Department of Nuclear Medicine of Clinical Centre in Ljubljana in the period from 1 January 2001 to 31 December 2004. All encompassed patients had passed through the entire stepwise diagnostic process of CAD and also fulfilled all other criteria for the study. In the first part of the study we copied out 55 crucial data for the CAD diagnostics of each 120 patient that were gained in the process of the stepwise CAD diagnostics. We calculated the diagnostic values (sensitivity, specificity and accuracy) of SMPS by means of the standard model of decision making. In the second part we, after the selection and training of machine learning method called Naïve Bayes Classifier, obtained two expert systems for SMPS that we named »Naivni bajsi I« and »Naivni bajsi II«. After that we calculated the diagnostic values for the SMPS with the aforementioned kinds of decision making and made a comparison between both expert systems and the standard model of decision making. In the final analysis we included 350 patients. With the standard model of decision making the sensitivity of SMPS was 68%, specificity 70% and accuracy 69%. In decision making with the expert system »Naivni bajsi I« sensitivity of SMPS was 67%, specificity 70% and accuracy 68%. In decision making with the use of expert system »Naivni bajsi II« sensitivity of SMPS was 69%, specificity 80% and accuracy 74%. The higher specificity and accuracy of the interpretation of SMPS results with the expert system »Naivni bajsi II« compared to the standard decision making model was found to be statistically significant (p < 0,05, McNemara’s test). Only the second hypothesis was confirmed: The interpretation of image results of SMPS by decision making with the expert system »Naivni bajsi II« has a higher diagnostic accuracy compared to the model of standard decision making. We could not confirm the first hypothesis. The interpretation of numeric results of SMPS by decision making with the expert system »Naivni bajsi I« in our study does not have a higher diagnostic accuracy compared to the standard decision making model, and it only had comparable diagnostic accuracy.