This paper describes the construction of various combinations of features from neural networks to solve the problem of determining the presence of heart disease in order to use them as data for machine learning methods. Next, for machine methods, after selecting optimal hyperparameters, various types of ensembles are built. Afterwards, various techniques for improving target metrics are considered using machine methods, neural networks and certain models, which together can give a higher predictive ability of the final program. The result is the repeated construction of ensembles from various models until the target metrics under consideration increase. The resulting solution is of interest because it is universal both for ECG diagnostics and, in principle, for solving classification problems in general.