Comparative Analysis of Classification Algorithms in the Analysis of Medical Images From Speckle Tracking Echocardiography Video Data

Abstract

Background. Machine learning allows applying various intelligent algorithms to produce diagnostic and/or prognostic models. Such models can be used to determine the functional state of the heart, which is diagnosed by speckle-tracking echocardiography. To determine the patient's heart condition in detail, a classification approach is used in machine learning. Each of the classification algorithms has a different performance when applied to certain situations. Therefore, the actual task is to determine the most efficient algorithm for solving a specific task of classifying the patient's heart condition when applying the same speckle-tracking echocardiography data set. Objective. We are aimed to evaluate the effectiveness of the application of prognostic models of logistic regression, the group method of data handling (GMDH), random forest, and adaptive boosting (AdaBoost) in the construction of algorithms to support medical decision-making on the diagnosis of coronary heart disease. Methods. Video data from speckle-tracking echocardiography of 40 patients with coronary heart disease and 16 patients without cardiac pathology were used for the study. Echocardiography was recorded in B-mode in three positions: long axis, 4-chamber, and 2-chamber. Echocardiography frames that reflect the systole and diastole of the heart (308 samples in total) were taken as objects for classification. To obtain informative features of the selected objects, the genetic GMDH approach was applied to identify the best structure of harmonic textural features. We compared the efficiency of the following classification algorithms: logistic regression method, GMDH classifier, random forest method, and AdaBoost method. Results. Four classification models were constructed for each of the three B-mode echocardiography positions. For this purpose, the data samples were divided into 3: training sample (60%), validation sample (20%), and test sample (20%). Objective evaluation of the models on the test sample showed that the best classification method was random forest (90.3% accuracy on the 4-chamber echocardiography position, 74.2% on the 2-chamber, and 77.4% on the long axis). This was also confirmed by ROC analysis, wherein in all cases, the random forest was the most effective in classifying cardiac conditions. Conclusions. The best classification algorithm for cardiac diagnostics by speckle-tracking echocardiography was determined. It turned out to be a random forest, which can be explained by the ensemble approach of begging, which is inherent in this classification method. It will be the mainstay of further research, which is planned to be performed to develop a full-fledged decision support system for cardiac diagnostics.