No abstract

The issues of detailing recognition algorithms in order to increase the validity of their solutions in diagnosing patients are considered using the example of processing nephrology data. The training of algorithms with a teacher is implied. Procedures for detailing complexes of clinical signs and criteria for comparing such complexes in decision-making are proposed. This means dividing these objects into elements, extracting additional information for them from a priori and current data, and taking them into account in algorithms. Research in the work was focused on the development of software tools for detecting and evaluating additional reserves and opportunities for improving the quality of decisions of recognition procedures by extracting additional useful information from a priori and current data and using them in the process of detailing decision-making procedures. On a specific algorithm, various approaches to such detailing and to the study of its effectiveness were analyzed. Such detailing can be built on the basis of using the experience of clinical practice of observation of patients and their diagnosis in the form of training samples of symptom complexes and (or) observed signals in clinical cases with reliably confirmed diagnoses in the relevant databases. Detailing these algorithmic procedures can lead to the emergence of a multi variance of possible solutions for differently detailed algorithms and require the use of additional procedures for generating a generalizing conclusion based on the results of their mutual consultation. The order and results of detailing are demonstrated in the MatLab environment on two modifications of the proposed algorithm. The introduction reveals the relevance and content of the research. Section 1 reveals the composition of a priori patient data in demo examples and the information that is extracted from them at the training stage. Section 2 proposes two modifications of the algorithm to detalize it. Section 3 proposes software procedures for the statistical evaluation of the performance of the detalization of the algorithms under study. Section 4 describes the refinement of algorithms by introducing weights into the decision criterion, taking into account the spread of values of clinical signs. Section 5 demonstrates the detalization of the algorithms taking into account the information content of the features. The conclusions summarize the results of the work. In general, they are positive.

This article is aimed at analyzing and improving the methods of preprocessing ECG signals for the task of detecting low-amplitude regular components. This study analyzed the main advantages and disadvantages of existing ECG signal preprocessing methods for the detection of late ventricular and atrial potentials. Based on this analysis, a cardiac cycle averaging method was proposed in order to increase the accuracy of detection of late potentials by various algorithms and improve the quality of preprocessing of the ECG signal aimed at detection of low-amplitude components. The main feature of the proposed method is the division of a large number of cardiocycles for averaging into smaller aggregates (epochs), and the subsequent application of linear matrix decomposition to suppress irregular inclusions. Also, when dividing into epochs, it can be used overlapping. It can reduce the difference between epochs, and increase the number of cardiocycles for averaging. The use of this approach allows to minimize irregular inclusions in the ECG signal and increase the accuracy of the selection of low-amplitude late potentials. In addition, the division into epochs and overlapping makes possible to avoid blurring of low-amplitude high-frequency components during averaging as a result of heart rate variability, as well as to improve the quality of averaging with a reduced number of cardiocycles. To test the proposed method, various approaches were used to assess the ECG signal preprocessing. Mostly, we compared the cardiac cycles obtained as a result of different averaging algorithms and the proposed method with the template. To test the averaging method, an artificial ECG signal was developed with existing noise, late ventricular and atrial potentials, heart rate variability, and a high-amplitude component that occurs at a random location every two heartbeats. The template cardiac cycle was obtained from the original artificial signal without any distortion or noise. Firstly, we visually compared and evaluated different averaging methods with the template. Secondly, we calculated the similarity metrics of the late potentials on the averaged cardiac cycle with the late potentials on the template signal. Based on these metrics, the curves of dependence of the similarity values on the amplitude of late potentials on the ECG signal were calculated. Thirdly, we evaluated the impact of the proposed averaging method on the classification results of various machine learning algorithms on real ECG signals with available late potentials. The overall testing result showed that the proposed averaging method is able to reproduce the morphology of low-amplitude regular components by 10-30% more accurately and improve the classification accuracy by 5-12%.