РАЗРАБОТКА МЕТОДИКИ ДИАГНОСТИКИ СЕРДЕЧНЫХ ШУМОВ ПУТЁМ АНАЛИЗА ФОНОКАРДИОГРАММ

In modern medicine, automated diagnostic systems play an invaluable role in making a correct diagnosis to patients

With the help of various kinds of portable devices, you can read (and, if necessary, write to the database files) an electroencephalogram (EEG), an electrocardiogram (ECG) or sound signals of the tones and noises of the heart – a phonocardiogram (PKG), body temperature, systolic and diastolic pressure, photoplethysmogram and other indicators When organizing a signal processing procedure with a very small delay (practically "in real time"), it becomes possible to instantly respond to various deviations and malfunctions in the body and quickly stop them by changing the situation, taking medicines and using other modern medical methods and techniques

It should be noted that to date there are a number of commercial software packages that allow processing data files according to various physiological indicators after the fact and making an alleged diagnosis (for example, the Easy ECG program is known among Russian products

The purpose of this work is to create a methodology for diagnosing deviations of physiological indicators of a person from his individual physiological norm according to PKG based on the development of automated software using modern algorithms.

Thanks to the authors' study of the state of the problem and some preparatory work, a research strategy was developed aimed at achieving the described goal, consisting of several stages:

1. Collection of a database with readings of sensors of various human health indicators (impersonal files, but with the obligatory indication of a real confirmed diagnosis).

2. Choosing the most appropriate programming language for research purposes.

3. Select a method and create a program for primary analysis of files from the database.

4. The use of neural network programming to identify the physiological state without the participation of specialists.

5. Approbation of the above steps on volunteers.

At the moment, some encouraging results have been obtained up to and including paragraph 3 of the strategy.

Many reference databases were found on the Internet, containing files with readings of various medical sensors, indicating the diagnoses detected in patients

At the same time, the team of authors is intensively working on the creation of diagnostic software. This paper will describe the result achieved to date.

The multi-platform and freely distributed Python programming language, recognized as the most popular at the moment, is chosen as the language, which has high performance and almost unlimited possibilities for expanding its capabilities. In addition, for the Python language, there are libraries of special functions described below, which are necessary to implement the tasks set by the authors.

Let's consider the algorithm for Python developed by us, which performs processing of diagnostic files, consisting of the following blocks:

1. Import required function libraries: Scipy.io.wavfile to convert files downloaded from the wav database to data arrays

2. Writing to the data arrays of the results of reading the loaded wav file of the oscillogram of the analyzed signal: digitization frequency and sample amplitudes (measurements at each digitization).

3. Filtering extraneous noise (zeroing samples with amplitudes of less than a certain value) and trimming the created data array to a given number of samples (for example, the sample files used in this study were cut to a duration of 10 seconds).

4. Calculate and display the sample time interval (reference information for process control).

5. Create data arrays for the spectrogram from the oscillogram data using the FFT (Fast Fourier Transform) method for subsequent frequency analysis.

6. Primary analysis of the obtained spectrogram by calculating the trend line (approximating the dependence of a given order) using the Numpy.polyfit function and outputting this line to the spectrogram graph.

7. Secondary analysis of the spectrogram using the Scipy.ndimage.maximum_filter function (maximum detection filter). In the future, it is planned to test other automated analysis methods.

8. Using the functions of the Matplotlib.pyplot module to display on the screen and in a new file simultaneously the results of processing the signal spectrogram read in paragraph 2, as well as its oscillogram with the results of primary and secondary analysis.

Let's consider an example and prospects for using the developed program code using the example of processing several files of the PKG database containing phonocardiograms of 1568 subjects in wav format, as well as a description of the diagnosis for each file in a text file supplied by a doctor. The dataset was collected as part of two mass screening campaigns conducted in Northeast Brazil in July-August 2014 and June-July 2015

The mechanism of noise is explained by the appearance of turbulent blood flow in various areas of the heart muscle and the underlying vessels. They can occur in transient disorders of the blood flow, or in organic diseases of the heart. We have not yet considered this issue in more detail.

The task of demonstrating the developed method so far is only to identify the nature (form) of noise. To make a final medical diagnosis, additional large-scale studies will be carried out by a group of specialists, which includes the authors of this article.

From the described PKG database, 3 random files with different "shape" of noise were selected: decreschendo, plateau and diamond. All of them are measured in the same region of the human body – in the area of the tricuspid heart valve, located between the right ventricle and the right atrium. The function of the valve is to prevent reverse blood flow during the passage of systole. The noise in his work indicates an abnormal course of this process and may indicate a future, or an existing heart disease.

The program trims all PKG files to the same length of 10 s to ensure the correctness of the results obtained (initially the length of the signals was from 20 to 30 s or more).

Figure 1 shows the result of the program with recording the heartbeat with decreschendo noise.

Figure 1 - Oscillogram, spectrogram and PKG processing results with decreschendo noise

DOI:10.23670/IRJ.2023.130.1.1

The noise spectrograms presented in Figure 1 and all subsequent visualized ones are cut off (by setting the corresponding setting in the developed program) on the right at a frequency of 200 Hz, since preliminary analysis of the entire signal spectrum showed that above this frequency the components of the spectra of all the studied signals are almost zero. In addition, since the digitization frequency is 4000 Hz, according to Kotelnikov's theorem, only peaks explained by the imperfection of the FFT algorithm can be present at frequencies below 2000 Hz on the spectrogram.

Figure 1 clearly shows that the maximum spectrum amplitudes detected by the program are in the range of 25-75 Hz. Decreschendo noise, according to the theory, indicates that the patient has stenosis of the aortic mouth or pulmonary trunk.

Figure 2 shows the result of PKG processing with diamond noise.

Figure 2 - Oscillogram, spectrogram and PKG processing results with diamond noise

DOI:10.23670/IRJ.2023.130.1.2

Here we see a completely different picture of the analysis results – the noise frequency varies little in the frequency range of 20-200 Hz with the presence of some increase in amplitude at frequencies 40-60 and 70-90 Hz. But in general, in the entire presented frequency range of 25-200 Hz, the noise amplitude does not change much, a maximum of 2 times. Such noise may indicate that the patient has anemia; its cause is a decrease in blood viscosity and, as a result, an increase in blood flow.

The result of processing the third PKG file is with noise of the "plateau" type, shown in Figure 3.

Figure 3 - Oscillogram, spectrogram and PKG processing results with plateau noise

DOI:10.23670/IRJ.2023.130.1.3

In this case, in the range of 25-200 Hz, the amplitude of the frequency components of noise changes much more than in the case of the diamond noise study - up to 3 times. Also, at a frequency of 180 Hz, there is a pronounced peak of amplitude, and 3 narrow peaks are formed at frequencies of 35-45, 55-65 and 80-90 Hz. In general, the spectrogram of this type of noise from all those considered turned out to be the most complex, complex. This broadband type of heart noise is generally considered to be caused by transient functional reasons, for example, in adolescents.

Also, when analyzing all the presented records of cardiac noise, approximation dependencies of the spectrum of different orders were derived. Although they did not turn out to be useful in terms of identifying noise of different types, at least they made it possible to distinguish a spectrum range useful for diagnostics – up to 200 Hz. Interesting is the fact that the Numpy.polyfit function allows you to find the coefficients of the regression equation of very high orders - for example, in Figures 1-3, dashed yellow lines show graphs of 20th-order trend lines (with regression coefficients output to the program console), and this is not the limit. Also, in the case of insufficient order (maximum degree) of regression, i.e. with a large standard deviation, the program warns about this.

We will determine the range of issues solved in this study, as well as the problems that are planned to be worked on.

The described study can be considered some test of the capabilities of the Python programming environment, during the application of which it was possible to solve all the questions arising in the course of the described work, therefore, it is quite perfect to achieve the goals of this study. It should also be noted the convenience of using libraries and special Python functions

Further, the maximum amplitudes of frequency components obtained during the production of spectrograms of processed signals and identified programmatically are planned to be used as new signals supplied to the neurons of the input layer of the neural network (the number of these neurons should be equal to the number of frequency components – you can limit, for example, to ten maxima). Then, experimentally, it is necessary to select the number of intermediate neural layers and the number of neurons in them. The output layer must contain a number of neurons equal to the number of diagnoses. The authors have experience with neural networks of different types and topologies.

At the next stage, the developed method should be tested on different sets of initial data: PKG, ECG, EEG and others, with a choice of sets that are most suitable for diagnosing a certain group of diseases or deviations of the physiological state of the body. It is also planned to use different types and models of sensors and primary transducers. Tests are required to be carried out separately in different age groups.

Also, with the help of the developed methodology, it is possible to identify not only diagnoses, but also the prerequisites for their occurrence. To do this, the measurement and processing of physiological signals should be carried out dynamically, continuously, preferably in real time. Another interesting task, possible for solving by the method under consideration, is the identification of the emotional background of people; at the same time, sensors that take physiological readings can be a thermometer, a tonometer, a photoplethysmograph and other sensors built into a smartwatch. It is also possible to create your own domestic model of such a watch, which contains a heart rate sensor in the design (a wide range of such portable sensors are available on sale in online stores).

The following results were obtained during the described work:

- the purpose is defined, and the study plan is drawn up;

- review of literature (including Internet sources on the topic of research);

- the programming language for writing program codes is selected;

- a test database with phonocardiograms of persons with different manifestations of systolic noise was selected;

- the algorithm of the program for PKG analysis was created, the functions required for this library were selected;

- program code is written that allows to implement the tasks planned by this study;

- with the help of the written program, a large number of FKGs from the database were analyzed, the results of the analysis of some phonocardiograms are presented in the article; briefly described are the prospective diagnoses of patients with a different spectrum of identified cardiac murmurs.

In addition, the prospects for the development of the method described in the article and its use in a wider range of innovative applications are described.