RU2759069C1 - Method for noninvasive diagnosis of coronary heart disease - Google Patents

Abstract

FIELD: computer technology; medicine.SUBSTANCE: group of inventions relates to the field of computer technology in medicine, namely cardiology, in particular to methods and systems for non-invasive diagnosis of coronary heart disease (CHD), and can be used for the diagnosis of coronary heart disease, including in the field of predictive, diagnostic, preventive and rehabilitation medicine. A method for noninvasive computer diagnostics of coronary heart disease is proposed, including the following steps: ECG signals in the patient’s lead I at rest are received; ECG signals in the patient’s lead V5 during a test under load are received; the received ECG signals are processed, forming a vector of signs under load and a vector of signs of rest, the quality of the ECG signal is determined; the assessment of coronary heart disease at rest using a model for resting ECG based on the formed vector of signs of rest is determined; the assessment of coronary heart disease during exercise using a model for ECG under load based on the formed vector of signs under load is determined, and the choice of a specific model depends on the quality of the ECG signal; the integral indicator of coronary heart disease is determined using an assessment of coronary heart disease at rest and an assessment of coronary heart disease during a test under load. A method for noninvasive computer diagnostics of coronary heart disease is also proposed, including the following steps: receiving ECG signals in the patient’s lead I at rest; processing the received ECG signals, forming a vector of signs of rest; assessment of coronary heart disease at rest is determined using at least one model for a resting ECG based on the formed vector of signs of rest. A method for noninvasive computer diagnostics of coronary heart disease is also proposed, including the following steps: receiving ECG signals in the patient’s V5 lead during a test under load; processing the received ECG signals, forming a vector of signs under load, determining the quality of the ECG signal; the assessment of coronary heart disease during exercise is determined using an ECG model under load based on the formed vector of signs under load, and the choice of the model depends on the quality of the ECG signal.EFFECT: group of inventions provides an increase in the accuracy and efficiency of the determination of coronary heart disease by ECG.17 cl, 68 dwg

Description

FIELD OF TECHNOLOGY

This invention relates to the field of computer technology in medicine, namely to cardiology, in particular to methods and systems for non-invasive diagnosis of coronary heart disease (IHD), and can be used for the diagnosis of ischemic heart disease, including in the field of predictive, diagnostic, preventive and rehabilitative medicine.

LEVEL OF TECHNOLOGY

Currently, there are several common methods for diagnosing coronary artery disease: electrocardiography (ECG), exercise ECG test, echocardiography, myocardial scintigraphy, coronary angiography. Coronary angiography is an invasive procedure, myocardial scintigraphy requires the adoption of radioactive isotopes and tomographic equipment, echocardiography requires echocardiography, an ECG test with physical exertion is dangerous for the elderly, and the ECG technique is the simplest, but also the least accurate diagnostic method. In medical practice, the most common way to diagnose ischemic heart disease using a bicycle ergometric test / treadmill test with registration of an electrocardiogram (ECG).

ECG is a method of recording the electrical activity of the heart and can, for example, reveal a number of cardiological pathologies and give an idea of the functional state of the heart. Cardiac impulses are recorded on paper or digitally.

An important task of modern practical medicine is the anticipation and early detection of diseases. Non-invasive diagnostic methods are being developed, in which the largest number of indicators of the state of the human body is recorded in one measurement and their analysis and interpretation is carried out. Mathematical analysis of the data obtained by these methods makes it possible to predict and determine the nature of a cardiological or other disease.

The known method of non-invasive diagnosis of ischemic heart disease - myocardial scintigraphy with a radiopharmaceutical (RP) (Tolkachev Y. et al. Myocardial scintigraphy in the diagnosis of ischemic heart disease // News of radiation diagnostics. - 1998, No. 2. - S. 30-32. Lupanov V.P. The value of electrocardiographic stress tests and other modern instrumental methods in assessing the effectiveness of percutaneous coronary interventions and the detection of restenosis // Therapeutic Archives. - 2010, No. 4. - P. 67-74), in which the patient undergoes a bicycle ergometric test with registration of an electrocardiogram (ECG) sequentially at rest, at maximum load, at 1, 3, 5 and 7 minutes of the recovery period, according to the data obtained, a conclusion is made about the presence or absence of coronary artery disease in the patient, in the presence of ischemic heart disease, the localization of the ischemic focus is determined.

One of the disadvantages of this method is, at least, the lack of the possibility of repeated repeated research, the use of a radiopharmaceutical, the need for a specialized room and expensive equipment.

Also known from the prior art is a method for non-invasive diagnosis of coronary heart disease (see RU 2468742, publ. cells from I to IV intercostal space, synchronous registration of an electrocardiogram (ECG) from each of the grid electrodes, calculation of the integral of the ECG curve for each point of electrode placement, construction of an isointegral map, an index of difference (IR) for an isointegral map, IR = (P-N) / σ, where P is the value of the integral of the ECG curve at the corresponding points of application of the electrodes in the patient under study, N and σ are the mean value and standard deviation of the integral of the ECG curve at the corresponding points of application of electrodes in patients in the normal group, determination of the minimum IR and the region of negative IR values , as a percentage of the area of each isointegral map, the conclusion about the presence or absence of the patient Ischemic heart disease, and in the presence of ischemic heart disease, determining the localization of the ischemic focus by the coincidence of the area of negative IR values with the projection of the myocardial zones on the scan of the chest surface, where the projection of the anterior part of the interventricular septum (IVS) is located from the right to left parasternal line from I to VI intercostal space, the projection of the posterior parts of the IVS - from the level of the angle of the left scapula to the right paravertebral line at the level of the II-V intercostal space, the projection of the apex of the left ventricle (LV) - from the left parasternal line to the left anterior axillary line at the level of V-VI intercostal space, the projection of the anterior wall of the LV - from the left parasternal line to the left mid-axillary line at the level of I-VI intercostal space with the exclusion of the projection of the LV apex, the projection of the LV lateral wall - from the left mid-axillary line to the vertebral line at the level of I-VI intercostal space with the exception of the projection of the posterior part of the IVS, the projection of the posterior wall of the LV - from the vertebral line to the right parasternal line at the level of the V-VI intercostal space, the projection of the right th ventricle - from the vertebral line to the right posterior axillary line and from the right mid-axillary line to the right parasternal line at the level of I-IV intercostal space with the exclusion of the projection of the posterior part of the IVS, characterized in that synchronous ECG registration from each mesh electrode is carried out sequentially with a bicycle ergometric test at rest, at maximum load, at 1, 3, 5 and 7 minutes of the recovery period, while for each point of application of the grid electrodes, the integral of the ECG curve is calculated at the QRST interval and isointegral maps of rest, maximum load and 1, 3, 5 and 7 minutes of the recovery period, the calculation of IR is carried out according to the above formula, where P is the value of the integral of the ECG curve in the QRST interval at the corresponding points of application of the electrodes in the patient under study, N and o are the mean value and standard deviation of the integral of the ECG curve at the QRST interval in the corresponding points of application of electrodes in patients in the normal group, and with a decrease in min of the absolute IR by 10% or more and / or an increase in the area of negative IR values by 18% or more for any of the isointegral maps of the maximum load and 1, 3, 5 and 7 minutes of the recovery period relative to the isointegral rest map give a conclusion about the presence of the patient Ischemic heart disease. The method of non-invasive diagnosis of ischemic heart disease is carried out as follows. A regular grid of at least 64 electrodes is applied to the entire surface of the patient's chest, covering the anterior, posterior and lateral walls of the chest from I to VI intercostal space. At the beginning, the patient undergoes a bicycle ergometric test at rest and the ECG is synchronously recorded from each of the grid electrodes. Calculate the integral of the ECG curve at the QRST interval for each point of application of the grid electrodes. Points with equal values are connected by lines that form an iso-integral rest map. Then the patient is given a bicycle ergometric test at maximum load and the ECG is synchronously recorded from each of the grid electrodes. Calculate the integral of the ECG curve at the QRST interval for each point of application of the grid electrodes. Points with equal values are connected by lines forming an isointegral map of the maximum load. Then the patient undergoes a bicycle ergometric test at 1, 3, 5 and 7 minutes of the recovery period, and the ECG is synchronously recorded from each of the grid electrodes. Calculate the integral of the ECG curve at the QRST interval for each point of application of the grid electrodes. Points with equal values are connected by lines forming iso-integral maps of 1, 3, 5 and 7 minutes of the recovery period. For each of the isointegral maps of rest, maximum load, 1, 3, 5 and 7 minutes of the recovery period, the difference index is calculated, PI = (P-N) / σ, where P is the value of the integral of the ECG curve over the QRST interval at the corresponding points of overlap patient electrodes; N and σ - mean value and standard deviation of the integral of the ECG curve over the QRST interval at the corresponding points of electrode application in patients in the normal group. Determine the minimum IR and the area of negative IR values as a percentage of the area of each iso-integral map.

One of the disadvantages of this solution is, at least, the need for a grid of 64 electrodes, a multichannel electrocardiograph, the complexity and duration of the procedure.

SUMMARY OF THE INVENTION

This invention is aimed at eliminating the disadvantages inherent in existing solutions.

The technical problem (task) in this invention is to diagnose coronary heart disease with high accuracy using an electrocardiogram (ECG).

The proposed invention improves the accuracy of diagnosis of ischemic heart disease and reduce the labor costs of medical staff for the diagnosis.

The technical result achieved when solving a technical problem is to increase the accuracy and efficiency of determining coronary heart disease by ECG.

In some embodiments, a method for non-invasive computer diagnostics of coronary artery disease includes at least the following steps:

- receive ECG signals in lead I of the patient at rest;

- receive ECG signals in lead V5 of the patient during a stress test;

- process the received ECG signals forming a vector of signs under load and a vector of signs of rest, determine the quality of the ECG signal;

- determine the assessment of IHD at rest using a model for ECG at rest on the basis of the generated vector of signs of rest;

- determine the assessment of IHD during exercise using a model for ECG under load based on the generated vector of signs under load, and the choice of a particular model depends on the quality of the ECG signal;

- determine the integral indicator of IHD using the assessment of IHD at rest and the assessment of IHD during a test under load.

In some implementations, depending on the quality of the ECG signal, the following pairs of models are used:

(ST signs and ST signs + rest), (ST + T signs and ST + T signs + rest), (ST + HFQRS signs and ST + HFQRS signs + rest), (ST + T + HFQRS signs and ST + T signs + HFQRS + rest).

In some embodiments, the ECG signals include peak search, adaptive filtering, complex classification, construction and labeling of the ECG.

In some embodiments, a method for non-invasive computer diagnostics of coronary artery disease includes the following steps:

- receive ECG signals in lead I of the patient at rest;

- process the received ECG signals forming a vector of signs of rest;

- determine the assessment of IHD at rest using a model for ECG at rest on the basis of the generated vector of signs of rest.

In some embodiments, a method for non-invasive computer diagnostics of coronary artery disease includes the following steps:

- receive ECG signals in lead V5 of the patient during a stress test;

- the received ECG signals are processed, forming a vector of signs under load, the quality of the ECG signal is determined;

- determine the assessment of IHD during exercise using a model for ECG under load based on the generated vector of signs under load, and the choice of a specific model depends on the quality of the ECG signal.

In various implementations, the models are based on a combination of a linear or quadratic discriminant and a neural network.

In various implementations, some or all of the described steps are processed by a processor or other handler that executes the instructions.

BRIEF DESCRIPTION OF THE GRAPHIC MATERIALS

FIG. 1 illustrates an example of the result of a baseline distortion (LF noise) filter;

FIG. 2 illustrates an example of the result of the interference removal filter (notch);

FIG. 3 illustrates an example of the result of a low-pass filter;

FIG. 4 illustrates calculated features according to the present invention;

FIG. 5 illustrates the essence of the markup of cardio complexes using an approach based on a smoothed derivative of the signal;

FIG. 6 illustrates an exemplary embodiment of not recognizable ECG recordings in accordance with the present invention;

FIG. 7 illustrates an exemplary embodiment of moderately noisy electrode contact recordings in accordance with the present invention;

FIG. 8 illustrates an exemplary embodiment of recordings lying between non-ECG recordings and moderately noisy recordings in accordance with the present invention;

FIG. 9 illustrates an exemplary embodiment of a constructed “normal CCC” according to the present invention;

FIG. 10 illustrates an example of a bad ECG according to the present invention;

FIG. 11 illustrates a wavelet transform of one complex in accordance with the present invention;

FIG. 12 illustrates the partitioning of a wavelet transform result into frequency bands over which the power integrals are calculated in accordance with the present invention;

FIG. 13 illustrates a first derivative wavelet complex according to the present invention;

FIG. 14 illustrates an example for an R-wave leading edge according to the present invention;

FIG. 15 illustrates the area from the beginning of the R-wave (zero line) to the maximum line of the upward slope, the area between the maximum lines of the two slopes, the area from the maximum line of the downward slope to the end of the R-wave (zero line)? According to the present invention;

FIG. 16 illustrates a QRS complex with highlighted QRS angles according to the present invention;

FIG. 17 illustrates an exemplary embodiment of a single cardiac cycle transformation;

FIG. 18 illustrates an exemplary single cardiac lead ECG signal, in accordance with one embodiment of the present technical solution;

FIG. 19 illustrates the problem of not being able to fully utilize wavelet analysis for the ST segment according to the present invention;

FIG. 20 illustrates a sequence of T-waves from contiguous complexes, a derivative with a wide smoothing window and a derivative with a small smoothing window, in accordance with the present invention;

FIG. 21 illustrates an example of a recording in which the T-wave is weakly expressed according to the present invention;

FIG. 22 illustrates an extracted RF component against a background noise according to the present invention;

FIG. 23 illustrates an exemplary embodiment of a cardiac averaging-based parameter calculation approach in accordance with the present invention;

FIG. 24 illustrates groups of five CCs, synchronized for averaging, in accordance with the present invention;

FIG. 25 illustrates the result of an adaptive lowpass filter (using a wavelet transform) of a group of five CCs, synchronized for averaging, according to the present invention;

FIG. 26 illustrates an exemplary amplitude drawdown zones in accordance with the present invention;

FIG. 27 and FIG. 28 illustrate examples of CHD according to the present invention;

FIG. 29 and FIG. 30 illustrate examples of norms according to the present invention;

FIG. 31 illustrates an exemplary peak line profile view (R-wave, first and second derivatives);

FIG. 32 illustrates an exemplary ST tilt in overt ischemia: horizontal depression;

FIG. 33 illustrates an exemplary ST tilt in overt ischemia: descending depression;

FIG. 34 illustrates an exemplary use of the parameter: (D2 power at 15 Hz) / (D2 power at 6 Hz), according to the present invention;

FIG. 35 illustrates an exemplary neural network for diagnosing coronary artery disease;

FIG. 36 illustrates an exemplary frequency response of a notch filter cascade in accordance with the present invention;

FIG. 37 illustrates a variant of the difficulty in determining the points of the boundaries P and the end T on the UC, according to the present invention;

FIG. 38 illustrates an exemplary version of the constructed CCC for a ten-minute test under load (without a recovery phase);

FIG. 39 illustrates computed QRS start, QRS end ("J" point), peak and T-wave end points according to the present invention;

FIG. 40 illustrates a fragment from a satisfactory quality load test in accordance with the present invention;

FIG. 41-FIG. 45 illustrate the operation of the algorithm on the example of a sample under load of poor quality, according to the present invention;

FIG. 46 illustrates an exemplary embodiment of smoothing an amplitude trend at point J + 80 ms in accordance with the present invention;

FIG. 47 illustrates an exemplary embodiment of calculating value trends in accordance with the present invention;

FIG. 48 illustrates an example of poor quality T-wave;

FIG. 49 illustrates an example of a plane T-wave in lead V5;

FIG. 50 illustrates an exemplary ischemic index pattern;

FIG. 51 illustrates the trends of values throughout the record: pulse, amplitude at point J, amplitude at point J + 80 ms, asymmetry index T, HFQRS value for the case: angiography - negative, test - negative, algorithm - no coronary artery disease;

FIG. 52 illustrates fragments of the signal tape corresponding to the beginning of the recording, for the case: angiography - negative, test - negative, algorithm - no coronary heart disease, according to the present invention;

FIG. 53 illustrates fragments of the signal tape corresponding to the peak of the load, for the case: angiography - negative, test - negative, algorithm - no IHD, according to the present invention;

FIG. 54 illustrates fragments of the signal tape corresponding to the recovery phase for the case: angiography - negative, test - negative, algorithm - no coronary heart disease, according to the present invention;

FIG. 55 illustrates indicators with model responses: ST - based only on the shape of the ST segment for the case: angiography - negative, test - negative, algorithm - no coronary artery disease, according to the present invention;

FIG. 55B illustrates indicators with model responses: ST + T - ST-segment shape and T-wave parameters for the case: angiography - negative, test - negative, algorithm - no coronary artery disease, according to the present invention;

FIG. 55B illustrates indicators with model responses: ST + HF - the shape of the ST segment and the HF component of the QRS for the case: angiography - negative, test - negative, algorithm - no coronary artery disease, according to the present invention;

FIG. 55D illustrates indicators with model responses: Integral - the combined result of all models for the case: angiography - negative, test - negative, algorithm - no coronary artery disease, according to the present invention;

FIG. 56 illustrates the trends in the values throughout the record: pulse (HR), amplitude at point J (Jamp), amplitude at point J + 80 ms (J80amp), asymmetry index T, HFQRS value (HFQRS) for the case: angiography - positive, sample - positively, the algorithm is IHD according to the present invention;

FIG. 57 illustrates the fragments of the signal tape corresponding to the beginning of the recording, for the case: angiography - positive, test - positive, the algorithm - is IHD, according to the present invention;

FIG. 58 illustrates fragments of the signal tape corresponding to the peak of the load, for the case: angiography - positive, test - positive, the algorithm - is IHD, according to the present invention;

FIG. 59 illustrates fragments of the signal tape corresponding to the recovery phase for the case: angiography - positive, test - positive, algorithm - is IHD, according to the present invention;

FIG. 60A illustrates indicators with model responses: ST - based only on the shape of the ST segment for the case: angiography - positive, test - positive, algorithm - is IHD, according to the present invention;

FIG. 60B illustrates indicators with model responses: ST + T - ST-segment shape and T-wave parameters for the case: angiography - positive, test - positive, algorithm - is IHD, according to the present invention;

FIG. 60B illustrates indicators with model responses: ST + HF - the shape of the ST segment and the HF component of the QRS for the case: angiography - positive, test - positive, the algorithm - is IHD, according to the present invention;

FIG. 60G illustrates indicators with model responses: Integral - the combined result of all models for the case: angiography - positive, test - positive, algorithm - is IHD, according to the present invention;

FIG. 61 illustrates the trends of values throughout the record: heart rate (HR) / pulse, amplitude at the J point (Jamp), amplitude at the J point + 80 ms (J80amp), asymmetry index T, HFQRS value (HFQRS) for the case: angiography - positive, test - negative, algorithm - there is an ischemic heart disease;

FIG. 62 illustrates the fragments of the signal tape corresponding to the beginning of the recording, for the case: angiography - positive, test - negative, algorithm - there is an ischemic heart disease;

FIG. 63 illustrates the fragments of the signal tape corresponding to the peak of the load, for the case: angiography - positive, test - negative, algorithm - there is an ischemic heart disease;

FIG. 64 illustrates the fragments of the signal tape corresponding to the recovery phase, for the case: angiography - positive, test - negative, algorithm - there is an ischemic heart disease;

FIG. 65A illustrates indicators with model responses: ST - only based on the shape of the ST-segment, for the case: angiography - positive, test - negative, algorithm - there is IHD;

FIG. 65B illustrates indicators with model responses: ST + T - the shape of the ST-segment and the parameters of the T-wave, for the case: angiography - positive, test - negative, algorithm - there is IHD;

FIG. 65B illustrates indicators with model responses: ST + HF - the shape of the ST segment and the HF component of the QRS, for the case: angiography - positive, test - negative, algorithm - there is IHD;

FIG. 65G illustrates indicators with model responses: Integral - the combined result of all models for the case: angiography - positive, test - negative, algorithm - IHD;

FIG. 66 illustrates a flow diagram, in accordance with one embodiment.

FIG. 67 illustrates a block diagram of an exemplary embodiment of ECG processing (2210 and 2240, FIG. 66);

FIG. 68 illustrates an example of a general purpose computing (computer) system.

DESCRIPTION OF IMPLEMENTATION OPTIONS

In the present invention, a system means a computer system, a computer (electronic computer), a CNC (numerical control), a PLC (programmable logic controller), computerized control systems and any other devices capable of performing a given, well-defined sequence of operations (actions, instructions).

A command processing device means an electronic unit or an integrated circuit (microprocessor) that executes machine instructions (programs).

A command processor reads and executes machine instructions (programs) from one or more storage devices. The role of data storage devices can be, but are not limited to, hard disks (HDD), flash memory, ROM (read only memory), solid state drives (SSD), optical drives.

A program is a sequence of instructions intended for execution by a computer control device or a command processing device.

Below will be described the terms and concepts necessary for the implementation of the present invention.

Non-invasive - the term is used to characterize research or treatment methods during which no needles or various surgical instruments are applied to the skin.

Screening - in medicine (English screening sifting) is a method of active detection of persons with any pathology or risk factors for its development, based on the use of special diagnostic studies.

Electrocardiography is a technique for recording and studying the electric fields generated during the work of the heart.

Cardiological lead (lead) - an area on the electrocardiogram obtained from two or more electrodes placed in the corresponding lead of a part of the patient's body. Each of them, due to the different location of the electrodes on the body surface, reflects various aspects of the electrical activity of the heart and can facilitate the doctor's process of identifying the pathology and the nature of the pathology of the cardiovascular system.

An electrocardiogram (ECG) is a recording of the bioelectrical activity of the heart, made using a recorder on a moving strip of paper or recording on a data storage device with the ability to graphically display data on a display.

The ventricular complex of the electrocardiogram is a set of electrocardiogram teeth, reflecting the bioelectric processes that occur when excitation propagates through the myocardium of the ventricles of the heart.

Cardiograph - a medical device that records the bioelectric activity of the heart;

Clinical Document Architecture (CDA) is one of the HL7 standards designed to standardize the structure and ensure the semantic interoperability of medical systems when exchanging medical information and / or medical documents.

RR interval - the time interval between adjacent R waves (R-waves) of the electrocardiogram, equal to the duration of the cardiac cycle; used in determining the heart rate, in the diagnosis of arrhythmias.

The QRS complex (QRS) is a ventricular complex that is recorded during the excitation of the ventricles of the heart. This is the largest deviation on the ECG. The width of the QRS complex indicates the duration of intraventricular excitation.

Machine Learning is an extensive subset of artificial intelligence that studies methods for building learning-capable algorithms.

AFC - amplitude-frequency characteristic.

Phase-frequency characteristic - phase-frequency characteristic.

PTT (from the English Pulse transition time) - the time of propagation of the pulse wave, which coincides with the time from the R-peak to the maximum growth.

The R-peak (R-wave) is usually the main wave of the ECG, which is due to the excitation of the ventricles, and its amplitude in standard leads and in leads from the limbs depends on the position of the electrical axis of the heart.

HR- heart rate (pulse).

LF - low frequency, low frequency.

HF - high frequency, high frequency.

The objects and features of the present invention, methods for achieving these objects and features will become apparent by reference to exemplary embodiments. However, the present invention is not limited to the exemplary embodiments disclosed below, but may be embodied in various forms. The essence recited in the description is nothing more than specific details provided to assist a person skilled in the art in a comprehensive understanding of the invention, and the present invention is defined only within the scope of the appended claims.

Used in the present description of the present invention, the terms "module", "component", "element" and the like are used to refer to computer entities that can be hardware / equipment (for example, a device, tool, apparatus, apparatus, part of a device, for example, processor, microprocessor, integrated circuit, printed circuit board, including electronic printed circuit board, breadboard, motherboard, etc., microcomputer and so on), software (for example, executable program code, compiled application, software module, part software or program code, and so on) and / or firmware (in particular, firmware). So, for example, a component can be a process running on a processor (processor), an object, executable code, program code, file, program / application, function, method, (software) library, subroutine, coroutine, and / or computing device (for example, microcomputer or computer) or a combination of software or hardware components.

The present invention makes it possible to diagnose coronary heart disease using an electrocardiograph, in particular a single-channel electrocardiograph, in the presence of a recorded (or recorded) ECG at rest and / or an ECG test with (under) stress.

For the implementation of the invention can be used electrocardiograph, for example, CARDIOQVARK, as well as (optionally) stress system, for example, stress system Schiller Cardiovit AT-104RS. Within the framework of the present invention, an annotated sample is collected from a given number of ECG records in the first lead with a given duration, for example, up to 900 seconds, some of which are with coronary artery disease documented by coronary angiography, a part of ECG records in the fifth lead with a given duration, for example, up to 900 seconds, part of them with coronary artery disease documented by coronary angiography. Further, preprocessing of the received ECGs is carried out, as a result of which a vector of signs is formed for each ECG. From the obtained vectors, a training and test sample is formed. The trained algorithm demonstrates high performance indicators on the test sample (for sensitivity and specificity). This algorithm, among other things, allows you to diagnose coronary heart disease using a single-channel electrocardiograph, without the need for echocardiography.

The user can perform a stress test on a treadmill or stationary bike using a modified Bruce protocol. Lead I ECG is recorded immediately before the test is performed at rest (the user applies fingers to the electrodes). During the execution of the load and in the recovery phase, an ECG is recorded in lead V5 (suction cups are used). The criteria for terminating the load are the occurrence of an attack of angina pectoris, ECG changes of an ischemic nature, the patient's refusal to continue the exercise due to fatigue of the leg muscles or a feeling of lack of air, the achievement of a submaximal age-related heart rate (HR), the occurrence of serious rhythm and conduction disturbances, a pronounced rise in blood pressure. In a particular case, the submaximal age-related heart rate is 85% of the maximum age-related heart rate, which is calculated by the formula:

208 - (product 0.7 by age).

Within the framework of the present invention, the diagnosis of coronary artery disease at rest is carried out as described below. The recorded (or recorded) ECG is marked and filtered, in particular, using the basic ECG marking and filtering algorithm, as described below.

So, to search for peaks, the ECG signal is digitally filtered in a given frequency band, for example, in a frequency band of 0.05 ... 100 Hz (this band contains a useful ECG signal sufficient for detecting complexes), after which the search for R-peaks is carried out using the method / the Pan-Tompkins algorithm (which is used for ECG processing and which is described, for example, in "The Pan-Tompkins algorithm in the HEARTBIT electrocardiograph system", Kerimbaev Nurassyl Nurymuly, Madieva Balnur Abdinabikyzy). Also, in addition to the Pan-Tompkins method / algorithm, any other (one of the known, for example, the Cuiwei Li, Romero Legaretta, etc. algorithms) method / algorithm for finding peaks can be used.

Further, with the already available information about the position of the R-peaks (positions of the peaks) by means of the system implementing the present invention, adaptive filtering of the ECG is carried out. In this case, three types of filters are used: a baseline distortion filter (low-noise), a noise removal filter (notch), and a low-pass filter.

FIG. 1 shows an example of the result of the baseline distortion filter (low-frequency noise).

Low-frequency noise that distorts the baseline can be caused by poor contact of the user's body parts with the electrodes (for example, a finger with a sensor), breathing patterns, etc. Baseline distortions can lead not only to incorrect values of the amplitude at the reference points, but also distort the shape of the QRS complex, and prevent the correct construction of the averaged cardiac cycle (UCC).

The principle of operation of the baseline distortion filter (low-frequency noise) is as follows. The search for the ECG isoline points located directly in front of the QRS complex is carried out. For this, an algorithm based on the first derivative can be applied: we move (move) from the found R-peak to the left until the modulus of the derivative becomes less than a given threshold, that is, the signal does not stop changing significantly. A value equal to 1/20 x the maximum value of the derivative at the leading edge of the R-wave is used as the threshold.

Further, according to the found points of the isoline, a smooth spline (cubic or higher order) is constructed, which models the distortion of the baseline. After subtracting it, the signal remains devoid of low frequency interference.

This method is used by an adaptive filter, i.e. its performance depends on the noise level. Since the frequency components of the noise and the useful signal intersect significantly (adaptability is fundamental): sharp peaks of the baseline can correspond to a frequency of up to 0.5-1 Hz, however, as a rule, the ST segment is used to diagnose ischemia, which occupies lower frequencies, therefore any classical filter, successfully removing low-frequency noise, can distort the desired signal.

The described filter is able to remove strong throws in the interval of just a couple of complexes.

FIG. 1 signal 101 is the original ECG signal with low frequency noise, and signal 111 is the signal after applying the described baseline distortion (low frequency noise) filter.

FIG. 2 shows an example of the result of the interference removal filter (notch).

As shown in FIG. 2, signal 122 is the original ECG signal, and signal 132 is the signal after applying the noise removal filter (notch).

A noise removal filter (notch) is used when harmonic interference is observed in the cardiogram at two basic frequencies: 50 Hz - interference from network equipment and 60 Hz - the refresh rate of the screen of a computing device, for example, a smartphone. To remove it, a cascade of the classic IIR Butterworth filter with a cut-out bandwidth of 0.5 Hz is used. This approach allows you to remove pickup with an amplitude that is much higher than the amplitude of the useful ECG signal, without damaging the useful signal.

FIG. 3 shows an example of the result of a low-pass filter.

A low-pass filter is used when there is a moderate myogram on the ECG signal, which can interfere with the accurate determination of the required parameters. As in the case of low frequencies, the complexity of filtering lies in the fact that the useful ECG signal (QRS complexes) extends to fairly high frequencies: 100 Hz and above. This HF information can be used to calculate some parameters for the diagnosis of coronary artery disease. The components of the myogram can be at frequencies less than 30 Hz.

One solution is to use a classic line filter. In this case, an infinite impulse response filter (so-called IIR filter) may be suitable, since it provides a sharp transition of the amplitude-frequency response (AFC) between the passband and the suppression band. This filter allows you to almost completely suppress high-frequency noise in a given frequency range and keep the useful signal in the passband. One option is a fourth order Butterworth filter, in particular with a cutoff frequency of 70 Hz. For the specified order of the Butterworth filter, the drop in the amplitude-frequency characteristic is already quite sharp, and the Gibbs effect (parasitic oscillations when sharp pulses are applied to the filter) is not yet very pronounced. Also, the Butterworth filter has a linear phase-frequency response in the passband, i.e. the frequency components of the signal when passing through the said filter have the same delay, therefore the signal will not be (severely) distorted.

Another option is to use an adaptive filter based on wavelets (from the English wavelet - a small wave, ripple is a mathematical function that allows you to analyze various frequency components of the data), which, in a particular case, avoids phase distortions and / or the Gibbs effect on artifacts. and allows you to save sharp R-peaks on the ECG (save a useful signal at the output of the lowpass filter).

As shown in FIG. 3 (A), signal 143 is the original ECG signal, and as shown in FIG. 3 (B) signal 153 - the signal after applying the low-pass filter. In the particular case of the implementation of the present invention, at least the ECG signal is decomposed into wavelet coefficients (i.e., in a particular case, signals are analyzed in the plane of the wavelet coefficients), small coefficients are equated to zero (in particular, using the procedure for determining the "smallness" threshold - "wavelet shrinkage") at levels corresponding to a high frequency (in particular, more than 70 Hz) and the application of the inverse wavelet transform is carried out. So, the mentioned zeroing of small coefficients (the coefficients of the decomposition of the original signal in the basis formed by the parent wavelet) is carried out in the time domain (in time, in particular on the time scale), moreover, it should be noted that the frequency of 70 Hz was selected experimentally and can vary, for example, depending on the RF noise level.

Based on the correlation analysis, all found cardiocycles are classified into groups according to the principle of similarity of the QRS-T shape. That is, pairwise correlations of all complexes are calculated (by QRS-T segments), then standard clustering takes place based on the specified proximity measure (for example, using agglomerative clustering algorithms).

In this case, for a normal ECG, an obvious dominant type should turn out. The remaining complexes are distorted by noise, as well as extrasystoles. A pathological cardiogram may contain groups of normal complexes and extrasystoles comparable in size.

Upon further analysis, all abnormal complexes are excluded from consideration. In this case, the recording can be completely excluded from consideration, for example, in the presence of a significant segment of the recording with artifacts, the user will be prompted to repeat the measurement, and in the presence of a large number of extrasystoles, it will be suggested to consult a doctor.

To build the CCC, a small number (predetermined, for example, about 10-15) of pairwise similar complexes of good quality with an approximately equal RR interval, as long as possible, is selected. The found complexes are combined according to the R-peak according to the criterion of minimality of pairwise correlations. UCC is calculated as the arithmetic mean of the indicated complexes.

Further, the CCC is subjected to marking - a search is carried out for markers of P, QRS, T-waves, their amplitudes, as described in more detail below. The calculated features are depicted in FIG. 4.

FIG. 4 shows calculated features according to the present invention.

FIG. 5 shows the essence of the marking of cardio complexes using the approach (method, method) based on the smoothed derivative of the signal.

Curve 165 is an ECG complex. The Savitsky-Golay method is used to calculate the first derivatives of the signal with varying degrees of smoothing. Curve 175 is a less smoothed derivative, the window width corresponds to the approximate frequency of the QRS complex, curve 185 is smoother, the window width corresponds to the approximate frequency of P and T waves, this derivative is not calculated for the QRS complex region.

It is carried out (using the algorithm) to find the point of the maxima and minima of the derivatives. The indicated points (markers 195) correspond to the points of the maximum steepness of the wave fronts of the complex. In the trend shown by curve 185, two lows correspond to the leading edges of the P and T-wave, two highs correspond to the trailing edges of the P and T-waves. On the trend shown on curve 175, the high corresponds to a falling edge from the R to the S-peak, and the lows to the rising edges to the R-peak and after the S-peak. The signal peaks themselves can be found at the points at which the derivative vanishes.

To determine the exact boundaries of waves, the algorithm moves from the found extrema of the derivatives, moving away from the peaks of the waves, until the absolute value of the derivative becomes less than a certain threshold. The found boundaries of P, QRS and the end of T are indicated in FIG. 5 markers 205.

Depending on the implementation of the algorithm, the indicated markup is performed for the averaged complex or for all good recording complexes (per cycle).

Thus, the time and amplitude parameters of the complex are determined, as well as additional parameters:

- PQ - time from the beginning of the P-wave to the beginning of the QRS complex;

- QT - time from the beginning of the QRS-complex to the end of the T-wave;

-QTc - corrected QTc interval (normalized to the pulse), Bazett's formula is used;

- TrTe - the time from the peak of the T-wave to its end;

- VAT - time from the beginning of the QRS to the R-peak;

- AJ - amplitude at point "J" (end of QRS);

- AJ80 - amplitude at point "J" +80 ms;

- QT / TQ (only with beat-by-cycle marking) - the ratio of the QT length to the time from the end of the T-wave of the previous complex to the beginning of the QRS complex of the next (TQ value).

Also, within the framework of the present invention, the quality of the processed ECG is assessed. Due to the specifics of the removal, various types of artifacts are possible. On a bad ECG, due to noise, the calculated signs can be distorted, which, ultimately, can lead to a misdiagnosis. In this case, the user is prompted to repeat the measurement.

Each ECG recording by noise level belongs to one of the following groups (from bad to good):

1. Conditional "garbage" (an example of which is shown in FIG. 6) - records that are not recognized as ECG. These are large bursts / oscillations, exceeding the QRS in energy and amplitude for most of the recording. Also, this group includes strong noise from the contact of the electrodes. In this case, the QRS complexes are severely damaged ("twisted", tilted in different directions, etc.), visually there is a false impression of different QRS morphology. Such ECGs, in a particular case, are excluded from processing.

2. Moderate noise from electrode contact (an example of which is shown in FIG. 7). The QRS shape is visible, but baseline fluctuations are comparable to P and T waves.

For such records, it is difficult to analyze the ST-T segment by cycle. However, a SCC of satisfactory quality can be built and ischemia can be assessed by its parameters. Such records are of limited validity, depending on the implementation of the algorithm, they can be processed.

3. High RF noise. In this case, the signs of the ECG based on the HF components will be distorted. If the recording does not contain other artifacts, ischemia can be assessed using a special reduced model without information about high frequencies.

4. ECG of satisfactory quality - in total, more than 70% of the length does not meet the noises noted in paragraphs. 1-3.

FIG. 6 shows an exemplary version of "garbage", records that are not recognized as ECG, according to the present invention, in which it is impossible to determine even R-peaks.

FIG. 7 shows an exemplary version of moderate noise, recordings with moderate noise from electrode contact, according to the present invention, in which QRS areas are well viewed, there are no artifacts that are classified as false ectopic CCs, a normal CCC can be built (an example of which is shown in FIG. 9 ), use it to determine the lengths of the intervals. However, cycle-by-cycle assessment of the P-wave and ST-T segment, in a particular case, is impossible.

FIG. 8 shows the borderline case between "garbage" and moderate noise, in particular, an exemplary version of the records lying between the records not recognizable as ECG and records with moderate noise, according to the present invention, in which it is already possible to determine most of the CC of the dominant type. But the noise is so high that false LESs slip through.

FIG. 9 shows an exemplary embodiment of the above-mentioned "normal CCC" constructed according to the present invention.

The classification of ECG recordings according to the described groups is carried out as follows. The main indicators of poor quality are:

1) an increased number of different QRS forms per unit of recording length,

2) an increased number of bad CCs.

The first value is obtained on the basis of the classification procedure (complexes) as described above. With strong artifacts, there are many pairwise dissimilar CCs, as shown in FIG. ten.

FIG. 10 shows a typical example of a bad ECG according to the present invention.

Normally, one record should contain a small number of forms: one dominant, several types of extrasystoles + you can still reserve several types for artifacts.

The number of bad CCs for the second value is determined as follows. According to the procedure described above, the QRS boundaries are estimated for each complex, then the amplitudes before, during and after the QRS are measured, if they exceed the specified thresholds, then the complex is considered bad.

In a particular case, in the presence of artifacts in more than 30% of the recording, the ECG belongs to the first group and is rejected.

For the second group of recordings (moderate noise), a cycle-by-cycle ST-T evaluation is performed. In a particular case, the P-wave may not be assessed, since it may not be present (AF, supraventricular rhythms); The T-wave in this lead can be poorly defined, which can lead to relatively large dispersion in normal ECGs. Therefore, in a particular case, the main indicators for assessing the quality are the average value of the moving variance (window of 10 complexes) for the amplitudes at points J, J + 80 ms, R + 120 ms. Sliding variances are used to track rapid changes, otherwise all samples under load may be rejected. The thresholds are selected as follows: for J - 60 μV, for J + 80 ms and R + 120 ms - 50 μV.

RF noise for the third group is estimated as follows. Only a low-frequency interference filter is applied to the signal. Next, using a lowpass filter, as described above, the RF noise power in the signal is estimated. If the threshold is exceeded, the record belongs to the third group (group 3), and during subsequent processing, all signs containing information about high frequencies are excluded from consideration.

Further, according to the present invention, methods (including algorithms) for computing complex features based on wavelets are described.

Features based on the "first derivative" wavelet transform.

The signal f (t) is mapped into a function of two variables CWTf using a continuous wavelet transform according to the following formula:

Where ψ is a function called the mother wavelet, the parameter "a" is responsible for scaling, "b" - for the shift, i.e. these parameters are associated with frequency and time, respectively.

It can be seen that the described transformation resembles the time-dependent Fourier transform, only instead of a basis of exponentials, an expansion of the function f (t) in some basis based on the wavelet ψ is used. By choosing a specific type of wavelet, the required characteristics of the signal can be selected.

Thus, the described transformation maps the signal into two-dimensional space (time, frequency).

Consider the wavelet transform corresponding to the first derivative. For this, a function called a quadratic spline can be used as ψ.

FIG. 11 shows the aforementioned wavelet transform of one complex. FIG. 11 shows the complex itself with markers P, QRS, T and its two-dimensional wavelet transform. The horizontal axis is the time scale, it is the same for both figures, the vertical axis at the bottom of the figure is the frequency, from 4 to 64 Hz in a logarithmic scale (the axis is directed from top to bottom).

The color gradient encoded the value of the wavelet transform: red (311) - large positive values, blue (321) - negative, green (331) - small. It can be seen that each wave front forms a spot extended along the frequency axis. Rising fronts correspond to red spots (positive values of the derivative), to descending fronts - blue ones (negative values of the derivative).

You can also see that the QRS has the highest energy at high frequencies, and the P, T-waves at low frequencies.

By means of sections of the image at a fixed frequency, graphs of the first derivative with various degrees of smoothing can be obtained (the width of the smoothing window will be the reciprocal of the indicated frequency).

When calculating the QRS parameters, the considered frequency band is 4 ... 4 Hz. In this case, a logarithmic grid of frequencies of 4, 8, 16, 32, 64 Hz is used. In a particular case, there is no point in looking at higher frequencies (including, as shown by experiments): the R lines dissolve in noise, even on a filtered signal, the R-wave splitting begins, which greatly complicates the analysis. At frequencies below 4 Hz, the red spot corresponding to the leading edge of R merges with the red spot from the P-wave, which makes it difficult to find the beginning of R, and the position of the maximum line is distorted. A similar effect occurs at the trailing edge of R in the case of a negative T-wave.

For the same reasons, to calculate the parameters of the T-wave, a frequency grid of 2, 4, 6, 8, 10 Hz (in particular, not logarithmic) is used.

Further, the following parameters will be used:

1. Full integrals over QRS and T-wave spots (ie full powers of wave slopes).

2. Integrals over the QRS and T spots over the frequency ranges formed by cutting along the above frequency grids (as shown in FIG. 12).

FIG. 12 shows an exemplary embodiment of dividing the result of the wavelet transform into frequency ranges over which the power integrals are calculated, according to the present invention.

Maximum line profile.

For each spot corresponding to a certain slope and a fixed value of frequency f, the position of the absolute maximum can be found (for the red spot it will be the maximum, for the blue spot it will be the minimum). If you connect all such positions, following the spot along the frequency axis, a continuous line can be obtained, which is called the line of absolute maximum.

Likewise, lines of zeros can be drawn along the frequency axis, which are located between the spots (i.e., it is enough to simply mark the points where the wavelet transform vanishes). These lines will correspond to the peaks on the ECG. An illustration is shown in FIG. 13.

FIG. 13 shows a complex with a first derivative wavelet transform. For the QRS complex, the described lines of maxima and lines of zeros are drawn, highlighted in color and numbers.

Line 412 (line “2”, purple) - the maximum along the red spot, corresponding to the leading edge of the R-wave.

Line 414 (line “4”, white) is the minimum at the blue spot corresponding to the trailing edge of the R-wave.

Line 413 (line "3", black) - the zero line, corresponds to the R-peak.

Line 411 (line "1") and line 415 (line "5") are the lines of zeros corresponding to the boundaries of the QRS.

When passing along any line of the maximum, while comparing each frequency value to the corresponding value of the wavelet transform on the maximum line, the graph shown in FIG. fourteen.

FIG. 14 shows an example for an R-wave leading edge according to the present invention.

FIG. 14 it can be seen that the value along the line of the maximum of the R-wave front increases to a certain frequency, then begins to decrease. The maximum value and the frequency at which it is reached will be called Amax and Fmax, respectively.

For the T-wave, the maximum value will be at the lower frequency.

Signs of the maximum line used in the future:

1. apparent power (integral under the curve);

2. power by frequency ranges; integrals are indicated in FIG. 14: I1 (511), I2 (512), I3 (513), I4 (514);

3. Values on the frequency grid;

4. frequency of maximum power Fmax;

5. maximum power Аmax;

6. combined features - the ratio of integrals from different frequency ranges;

7. Integrals over the spots corresponding to the R-wave fronts separated by the maximum lines of the two fronts. The regions are illustrated in FIG. 15: area 1 (611) - from the beginning of the R-wave (zero line) to the maximum line of the upward slope, area 2 (612) - between the maximum lines of the two slopes, area 3 (613) - from the maximum line of the downward slope to the end of the R- waves (zero lines). That is, in fact, the integrals of the power of the middle part of the R-wave and tails are calculated.

QRS angles.

FIG. 16 shows the QRS complex (black curve) 701 with the QRS angles highlighted. Points A (711) and B (712) denote the points of maximum steepness of the leading and trailing edges of the R-wave (calculated as the maxima of the wavelet transform described above at a frequency of 32 Hz).

The parameters of interest are the angle a1 (721) and angle a2 (722) indicated in the figure, green and blue, respectively. These angles characterize the sharpness of the R-wave.

FIG. 17 shows an example of the transformation of one cardiocycle.

Second derivative.

Similar to the "first derivative" wavelet transform described above, the "second derivative" transform can be performed. For this, the mother's Mexican hat wavelet can be used, given by the formula:

ψ (t) = (1-t ²⁾ exp (-t ^2/2).

You can see that now the spots correspond not to the fronts, but to the peaks. Blue (807) - peaks with negative curvature (maxima of P, R, T-waves), red (817) - with positive curvature (beginning and ends of P, R, T-waves), green (827) - signal regions with small curvature (near-zero values of the second derivative).

The concept of a maximum line is defined in a similar way.

The parameters of the second derivative wavelet transform used in what follows:

1. Full integrals and integrals over frequency grids indicated in the description of features based on the "first derivative" wavelet transform above.

2. The parameters specified in the description of the high line profile above for the 3 high lines: R-peak, S-peak and T-peak.

ST-T-segment shape.

Another (in particular, the most important) group of parameters is associated with the ST-T segment. Its basic parameters are entered in the description of the marking algorithm: the amplitude at the "J" and "J" points plus (+) 80 ms. According to these values, one of the most important parameters can be determined - the slope of the ST segment.

Next, the complex parameters of the ST-T are determined.

FIG. 18 shows an exemplary ECG signal of one cardiac lead, in particular in FIG. 18 (A) shows a general view, and FIG. 18 (B) shows an enlarged fragment of an exemplary ECG signal with the features described in the present invention, in the ST region (848).

Unlike previous analyzes of R- and T-waves, in this particular case, the difficulty lies in the fact that the ST segment has a sufficiently large variation in the shape from user to user, for this reason it is rather difficult to find stable points for which the necessary ( required) parameters.

In the following, the following five ST-T anchor points will be used:

- point J 908 - the point that is the end of the QRS;

- point (J + 80), which is the standard point used in the analysis of ECGs. 918 is a point 80 ms away from J (908); in a particular case, the energy at these points may not be calculated, since the ECG segment around the J (908) point has very large variability: the S-wave may be absent or dominant in the complex, the QRS may have a smooth or sharp ending (depending on the characteristics of intraventricular conduction , the presence of pathologies), etc., and for the point J + 80, the calculation of energies is not required, since this point is on a straight line;

- point D2 928 - the maximum point of the second derivative (center of the bend to the T-wave); the position of this point depends on the analyzed frequency (i.e., on the width of the window over which the derivative is calculated), therefore, in the context of energies, D2 is not a point, but a certain region; the described model assumes that from point J to D2 the ST-segment is an inclined straight line, In this case, D2 can be quite close to J, but (including as shown by experiments), for the case of rest, for almost all records, point D2 lies to the right of the point J + 80, thus, the slope of the initial part of the ST segment can be characterized by amplitudes at points J and J + 80, and the initial phase of the rising front to the T-wave - by the energy spectrum at point D2;

- point D1 938 - the maximum point of the first derivative; similarly to D2, when searching for the wavelet transform maxima at different frequencies, a certain region is obtained;

- point Tr 948 - the peak of the T-wave.

Thus, the ST-segment model is reduced to five points for which the features are calculated.

In a particular case, unlike the case of the R-wave, the wavelet analysis cannot be used to the full extent for the ST segment, which is due to the fact that the necessary information is located at sufficiently low frequencies and short segments in duration.

FIG. 19 illustrates this problem in accordance with the present invention.

If it is necessary to calculate the frequency distribution of energy at the marked point D2 928 (i.e. to understand how sharp the bend to the T-wave is), then it can be seen that the maximum energy at this point lies at the frequency corresponding to the sinusoid 979 (i.e., the curvature at point D2 is just equal to the curvature of this sinusoid). To implement a full-fledged wavelet transform, even with such a compact wavelet as "Mexican Hat", a signal segment with a length of at least 1 period of the specified sinusoid is required. Those. the value of the maximum energy of interest at point D2 is influenced by completely unnecessary parts of the ECG: the previous R peak, the subsequent T- and U-waves, etc. In addition, baseline fluctuations distort the result.

In other words, the problem arises from the fact that it is necessary to analyze and fully describe the wave, looking only at a small part of it.

Thus, the peak of the spectrum of the investigated point is not available, and as signs one can take only its falling slope at higher frequencies. For this reason, for points D1, D2, Tp, the group of features "frequency of maximum power" is excluded. In a particular case, experiments have shown that it is possible to reliably calculate the energies at the indicated points for the range [6, 15] Hz (the problem described above arises below, and above - the drop and "dissolution" of power in noise).

Signs of the ST-T segment used in the following:

1.ST slope (amplitudes J and J80);

2. the amplitude of the T-wave;

3.energy at points D1, D2, Tr on the frequency grid 6, 11.15 Hz;

4. integrals of power at points D1, D2, T over frequency bands [6, 11] Hz and [11, 15] Hz;

5. time from point D2 to the peak of the T-wave.

T-wave asymmetry.

This method (method) is based on the evaluation of the T-waveform. The presence of coronary artery disease can lead to some of its deformation. The ratio of the slope of the leading and trailing edges T.

Based on the "first derivative" wavelet transform, the derivative of the signal over the T-wave region with a certain smoothing parameter is calculated.

FIG. 20 above (FIG. 20 (A)) shows a sequence of T-waves from successive complexes. In the middle (FIG. 20 (B)) - the above derivative with a wide smoothing window, at the bottom (FIG. 20 (C)) - the derivative with a small smoothing window. Marker 1010 indicates the maximum of the derivative, that is, the point of maximum steepness of the leading edge, marker 1020 - the minimum, the point of maximum steepness of the trailing edge. It can be seen that small fluctuations in the T-wave caused by noise lead to instability in the position of the points.

It is impractical (including according to experiments) to smooth derivatives with a window of more than 100 ms, i.e. corresponding frequency of the wavelet transform below 10 Hz (in this case, the algorithm begins to capture the ST segment or the next P-wave at a high pulse), as well as above 100 Hz (the influence of noise is too strong, on the slopes of the T-wave the derivative breaks down into several peaks, as in FIG. 20).

The best anti-aliasing frequency is the minimum as far as the signal quality allows. That is, the following adaptive algorithm can be used: points are calculated at different degrees of smoothing, from strong to weak, while their stability is assessed. As soon as it is violated, we stop and take the last, least smoothed version.

Another possible implementation of the algorithm is adjustment for physiological characteristics. Some impaired users may experience a wide T-wave with long edges. With excessive smoothing, the derivative can split into several peaks corresponding to the features of the fronts. For such records, the algorithm can also adaptively adjust the window width.

The parameter that will be used in the future in determining the CHD is the ratio of the maximum of the modulus of the derivative of the leading edge to the maximum of the modulus of the posterior (T-wave asymmetry index).

There are two calculation options: according to the CCC and the median for all good complexes in the record. The latter is preferred for good quality recordings. The method may not be applicable, for example, for recordings in which T-waves are weakly expressed. An example of such a record in FIG. 21.

FIG. 21 shows an example of a recording in which the T-wave is weakly expressed according to the present invention.

In a particular case, an algorithm can be used that estimates the presence of a stable T-wave. After the cycle-by-cycle marking of the complexes specified in the description of the marking algorithm, the dispersion of the T-wave amplitude, the dispersion of the maximum and minimum values of the derivatives of the T-wave fronts is estimated. If the spread is large or the amplitude is less than the specified threshold, the described method for this record, in a particular case, is not applied.

Signs based on the HF component of the QRS.

When diagnosing coronary artery disease based on the HF component of the QRS, the most informative is the activity from the band [150, 250] Hz (including on the basis of experiments).

The difficulty in extracting this information lies in the fact that at high frequencies there is a sufficiently high noise level: myogram, pick-up harmonics, tails from strong surges of noise associated with poor electrode contact, etc. Direct selection of the band of interest using filters does not produce the desired result, as illustrated in FIG. 22: Even the R-peaks are not visible in the extracted RF component against the background noise.

FIG. 22 shows the extracted RF component against the background noise according to the present invention.

In a particular case, it is not possible to clear the received signal from noise: there is no data for the adaptive filter (the power and frequency of the useful signal and noise are approximately the same, there is no data that has a positive correlation with the noise), classical filters will destroy (level) the useful signal.

In a particular case, due to the indicated problems, it is not possible to calculate the parameters of interest per cycle, therefore, an approach based on averaging cardio complexes can be used, which (approach, algorithm, method, method), in a particular case, does not apply to records in which there was an excessively high RF noise was found (group 3 described for the data quality estimation algorithm).

The mentioned approach works as follows. Based on the data on R-peaks, adaptive filtering of low-frequency noise is carried out in the original ECG signal with a sampling frequency of 1000 Hz, according to the description of the low-pass filter circuit. At this point, baseline drift is completely removed, but the treble remains intact.

Further, using a high threshold for correlation, as well as the value of the RMS of the RF noise, the quality of each complex is assessed. At this stage, it is critically important to leave only ideal complexes (without strong noise, shape distortions, etc.) belonging to the dominant group. There are two reasons for this:

1. A rather subtle parameter is analyzed - the power of high frequencies. Any distortion (for example, a shifted complex due to noise from poor electrode contact or with a displaced position of the R-peak) even in one complex, when averaged, will immediately affect the result.

2. Averaging is performed over a sufficiently large segment of the signal (see below). If complexes of different morphology are included here (this is often the case on long stress-records), the result will be distorted at the CCC.

Further, on the basis of the described data on the quality of the complexes, averaged complexes are constructed using a 1000 Hz signal filtered from low-frequency interference. In this case, a special iterative procedure is used for the exact alignment of the complexes, called the Woody filter, since any, even a small error, will lead to a blurring of the resulting VCC, which is equivalent to lowpass filtering, that is, the destruction of the useful high-frequency component.

The permissible noise level in the CCC (including that established experimentally) is approximately 1 μV (RMS value), for this, about 100 cardiocycles are averaged, while the lowpass effect of the averaging operation remains insignificant, which allows keeping the HF component undistorted. In a particular case, the approach (method) can be used on one-minute records with a normal heart rate. With bradycardia, the length can be increased to minutes. On longer records, to increase the accuracy, the CVC can be calculated in a sliding manner over a window of 100 complexes wide in 10 seconds increments. After that, the medians of the parameters calculated for all CCCs can be taken. In this case, it is necessary to monitor the value of the RMS of the RF noise and discard the CCC of insufficient quality.

After constructing the UCC, a band-pass filter [150, 250] Hz is used. The result (including the experimental one) is quite sensitive to the choice of the filter type. A "soft" filter with a smooth transition of the frequency response works noticeably worse, on the other hand, a "hard" filter gives a "ringing" (Gibbs effect), which is undesirable.

In a particular case, the worst results were shown by filters based on a wavelet transform, and the best results were shown by a pair of lowpass and highpass fourth-order Butterworth filters (in this order, the required balance between the frequency response slope and the Gibbs effect is achieved), as well as a band-pass FIR filter with a symmetric impulse response. The latter works better for recordings with noise, and Butterworth filters for clean ones.

An illustration of how the approach works is shown in FIG. 23.

FIG. 23 illustrates an exemplary embodiment of a cardiac averaging-based parameter calculation approach according to the present invention. In the upper part, the CCC is shown for 100 seconds, in the lower part - an enlarged signal from the range [150, 250] Hz. The quantity of interest is the RF component amplitude (HFQRS) 1033.

For recordings with the presence of RF noise, especially for those obtained using electrodes installed on the user's fingers, an additional adaptive lowpass filter (using wavelet transform), described above, can be applied, which can significantly reduce the amount of RF noise, while not greatly distorting the desired signal, the result of which is illustrated in FIG. 24 and FIG. 25.

FIG. 24 shows a group of five CCs synchronized for averaging in accordance with the present invention.

FIG. 25 shows the result of an adaptive lowpass filter (using a wavelet transform) of a group of five CCs, synchronized for averaging, according to the present invention.

Thus, it can be seen that the HF noise significantly decreased without affecting the QRS (QRS was not affected). This approach allows you to weaken the criteria for the applicability of the considered method (method).

After the above-mentioned filtering, the high-frequency component slightly decreases, therefore, a weakened threshold is set for such records in the CAD model. In a particular case, for this reason, the method (method) may not be applied to recordings where HF noise is moderate.

The RAZ parameter (from the English reduced amplitude zone).

Another parameter related to RF signal information is RAZ.

The method for calculating it (the method for calculating it) is as follows. The HF component is extracted from the QRS wave by the method (method) of calculating features based on the HF component of the QRS described above, after which its envelope is constructed. CHD induces amplitude drawdown zones, well distinguishable by the shape of the envelope, as shown in FIG. 26.

FIG. 26 illustrates an exemplary amplitude drawdown zones in accordance with the present invention.

FIG. 27-FIG. 30 shows four examples: FIG. 27 and FIG. 28 - with IHD, arrows (1057, 1068) show the amplitude drawdown zones, in FIG. 29 and FIG. 30 is the norm.

Red curve (1067) - UKTs; blue (1077) - RF component obtained with an FIR filter; blue (1087) - its envelope; violet (1097) is the high-frequency component obtained with the Butterworth filter, black (1099) is its envelope.

Connection of parameters with the presence of ischemic heart disease.

The following are parameters that clearly correlate (including as a result of experiments) with IHD. In a particular case, complex combinations of features can be used in the model, which also show a good relationship with IHD.

- QTc - IHD has more;

- QT / TQ - IHD has more;

- asymmetry of the T-wave - in IHD, the front slope is steeper;

- the frequency of the maximum power of the R-wave in IHD is less (R is dumber);

- the power of the R-wave slopes in the IHD is higher;

- HFQRS in ischemic patients is less, RAZ is present.

FIG. 31 shows an approximate version of the profile of the maximum lines (R-wave, first and second derivatives): with ischemic heart disease - 1131, norm - 1141.

In a particular case, in general, on ischemic ECGs, the LF component is higher (good indicators are the integrals over the bands [4, 8], [8, 16] Hz), the frequency of the maximum power is lower.

Parameter seven can be used (integrals over the spots corresponding to the R-wave fronts, separated by the maximum lines of the two fronts): the ratio of the power of the middle of the R-peak to the tails - for ischemics, the parameter is less.

The amplitudes at the points J and J + 80 with ischemic heart disease are less. In some cases, it can be, on the contrary, abnormally large (acute violation).

The parameters corresponding to the shape of the ST segment show a good relationship with the presence of coronary artery disease.

FIG. 32 shows an exemplary ST tilt in overt ischemia: horizontal depression (1152).

FIG. 33 shows an exemplary ST tilt in overt ischemia: descending depression (1163).

At point D2 (as described in the description of the ST-T-segment shape above) in ischemics, energy is concentrated at higher frequencies. In a particular case, the parameter can be applied (works well): (power D2 at 15 Hz) / (power D2 at 6 Hz), which is explained by the illustration in FIG. 34.

FIG. 34 shows an example of using the parameter: (D2 power at 15 Hz) / (D2 power at 6 Hz) for IS 1174 and norm 1184.

At point D1, on normal recordings, the drop in energy when moving to high frequencies is more pronounced than on ischemic ones. Those. the leading edge of the T-wave is steeper in ischemic recordings. In contrast to the T-wave asymmetry method, this result suggests that, in a particular case, only the leading edge of the T-wave can be used to diagnose IHD.

In general, on ischemic recordings, the T-wave amplitude is higher, as well as the energies at the Tp point in all frequency bands.

The segment length [D2, Tp] in the presence of coronary artery disease is increased.

Model.

A special model is used to make a decision on the presence of IHD in the record based on the parameters described above.

Depending on the implementation of the algorithm, the model can be one of the following architectures: linear / quadratic discriminant and / or neural network.

Linear / quadratic discriminant.

The linear discriminant is as follows:

Est = a0 + a1 * p1 + a2 * p2 + ... + an * pn,

where pi are the parameters described above, ai are coefficients, the value of which is selected during training, Est is a value called the ischemia index.

Est takes arbitrary values, if the value is negative, the model shows ischemia (the greater the negative value in modulus, the more pronounced ischemia). Likewise, the larger the positive score, the more normal the ECG.

The quadratic model can include terms of the form: ak * pi * pj (products of values-attributes). The introduction of such elements for some of the signs can improve the diagnostic accuracy.

Neural network.

A single layer neural network with a logistic activation function can be used. The network has one hidden layer of small neurons. Depending on the implementation of the algorithm, the size of the hidden layer can vary, as well as the type of the activation function (you can use a hyperbolic tangent or a linear function).

FIG. 35 shows a possible variant of a neural network for the diagnosis of coronary artery disease. The network under consideration has three layers: the first with the number of neurons equal to the number of features used in the model, a hidden layer of three neurons, and an output neuron.

Parameters 1214 are fed to the input of the neural network, and the output is the presence or absence of IHD 1224.

For training and testing of the neural network, the following sample was used: 63 patients, 31 of them were diagnosed with coronary artery disease (by the method of coronary angiography). The algorithm classified 6 records as not interpretable due to poor quality. In total, 57 records remain, 25 of them with coronary artery disease, 32 - the norm. Method / Algorithm correctly found 20 IHD (5 missed), identified 6 false IHD. Thus, the efficiency of the algorithm is: sensitivity = 80%, specificity = 81%.

Diagnosis of ischemic heart disease under load.

The user performs a stress test on a treadmill using a modified Bruce protocol or on a bicycle ergometer. Lead I ECG is recorded immediately prior to the test at rest (in particular, the user applies fingers to the electrodes). During the execution of the load and in the recovery phase, an ECG is recorded (in particular, recorded) in lead V5 (in particular, suction cup electrodes are used). The criteria for terminating the load are the occurrence of an attack of angina pectoris, ECG changes of an ischemic nature, the patient's refusal to continue the exercise due to fatigue of the leg muscles or a feeling of lack of air, the achievement of a submaximal age-related heart rate (HR), the occurrence of serious rhythm and conduction disturbances, a pronounced rise in blood pressure. The submaximal age-related heart rate is 85% of the maximum age-related heart rate, which is calculated using the formula 208 - (0.7 x age).

The recorded ECGs are processed by a data processing server, for example, on the CardioQVARK server.

Thus, the presence of coronary artery disease in the user is assessed by comparing the parameters of the cardiogram at rest and the dynamics of their development during exercise and in the recovery phase.

Basic layout and filtering algorithm.

To process the recording under load, a special case of the approach (method, method, algorithm) for ECG processing at rest is used, with the exception of the changes described below.

As with resting ECG processing, the Pan-Tompkins algorithm first searches for R-peaks.

For filtering, the same cascade of three filters is used: for high-frequency, low-frequency noise and a harmonic noise rejector. Since in this case there is a close interaction with network equipment (treadmill or treadmill), a fairly strong interference may occur. For this reason, the notch filter for loaded samples can be strengthened. So, this filter is tuned (such parameters are set) in such a way that, in addition to the base frequency of 50 Hz, cut (exclude) from the signal its (in a particular case, all) harmonics up to the Nyquist frequency: i.e. 50, 100, 150, 200, 250, 300, 350, 400, 450 Hz (moreover, in a particular case, the signal sampling frequency is 1000 Hz). In addition, the filter also cuts 60 Hz, which is the refresh rate of the screen of a computing device (eg a smartphone).

FIG. 36 shows an example of the frequency response of the constructed cascade of notch filters.

Due to the use of IIR filters with a narrow cut-out band (0.3 Hz), such a scheme (practically) does not distort the useful signal, in particular, information (data) about high QRS frequencies, which can be used when deciding on the presence of IHD.

In a particular case, if a system implementing the present invention (in particular, a device or several devices) system is used in a place where the network frequency is 60 Hz, the filter circuit can be changed accordingly.

In a particular case, another variation of the method (in particular, the algorithm) can be used, which allows you to save (in particular, it is better to save) information about high frequencies - a preliminary analysis of the pickup power. If its value is small or absent at multiple harmonics, a weaker version of the filter can be used.

Further, the classification of the complexes is carried out, as described above, with the allocation of the dominant type (in a particular case, the remaining complexes are either extrasystole or very noisy).

According to the algorithms described above, the construction and marking of the CCC is carried out using marking algorithms for a given time interval, in particular, for the first minute of recording. In this case, complexes with the largest possible RR interval are selected for averaging. The specified UCC will correspond to the state of rest of the user.

Next, the recording quality is assessed (using the data quality assessment algorithm described above), and all bad complexes are marked (marked, highlighted).

Per-minute UKTs, marking.

The resulting (all) record is divided into non-overlapping segments of a given length (in particular, one minute long), in each of which, according to the algorithm described above (for constructing and marking the CCC), CCCs are built. At the same time, for averaging, if possible, complexes with approximately equal RR are selected from the middle of the corresponding segment, in particular, in order to avoid blurring of the QRS-T region when averaging the complexes with a pronounced changing pulse. Also, at this stage, complexes marked as anomalous during classification and marked as corrupted by noise in the quality assessment procedure (in particular, according to the study scheme) may not be used.

Further, the marking of the constructed CCCs is carried out. This uses a modified version of the markup algorithm described above.

In a particular case, with a high pulse, the T-wave of the previous complex converges with the P-wave of the next complex, up to their intersection, therefore the previous version of the algorithm may have difficulties in determining the points of the boundaries of P and the end of T at the CCC. An illustration of the problem is shown in FIG. 37.

FIG. 37 shows a variant of the difficulties in determining the points of the boundaries P and the end T on the UC, according to the present invention.

The blue curve (1237) is the CCC signal, the red curve (1247) is the first derivative with smoothing corresponding to the frequency of the P / T wave (in the particular case, as before, it was calculated everywhere, except for the QRS complex). The P-wave is practically on the descending front of T, due to which the leading edge of P becomes invisible. FIG. 37 it can be seen that there may be no corresponding extremum of the derivative, which is marked with a purple marker (1257). There are difficulties both with the definition of the left boundary of P, and with the definition of the end of T (the exact position of the end is hidden under the layered P-wave). Problems are exacerbated by the presence of a U-wave as well as biphasic T / P.

This modified algorithm remembers the waveforms and marker positions for the CCC at rest (constructed using the RAZ parameter described above). Further, the algorithm uses the specified information when marking the remaining minute CCCs: first, it searches for regions on the derivative corresponding to the right front of the T-wave and the left front of the P-wave, then, as before, moves away from the wave peaks until the necessary condition is reached on the value of the derivative, while the algorithm, in a particular case, the boundary for the algorithm is the corresponding marker of the CCC at rest (the algorithm should not go beyond the corresponding marker of the CCC at rest).

Thus, using the algorithm, the marked CCCs are constructed with a given step (in particular, equal to one minute).

FIG. 38 shows an exemplary version of the constructed CCC for a ten-minute test under load (without a recovery phase).

Also, in some implementations of the algorithm for the minute CCC, the entire array of parameters described above is calculated using the appropriate algorithms (at least, features based on the "first derivative" wavelet transform, the profile of the maximum line, QRS angles, the second derivative, the ST- T-segment, T-wave asymmetry, features based on the HF component of QRS, RAZ parameter, marking algorithm, data quality assessment algorithm).

Per-cycle markup.

Some of the parameters for the stress test are calculated for each normal complex in the record.

The approach described above is used to find for each complex the points of the beginning of the QRS, the end of the QRS (point "J"), the peak and the end of the T-wave.

The calculated points are shown in FIG. 39.

FIG. 39 shows the calculated QRS start, QRS end ("J" point), peak and T-wave end points according to the present invention.

At the points QRSst (1319), J (1329), Tst (1349), Tpeak (1359), Tfi (1369), Tend (1379), as well as at the point J + 80 ms (1339), the signal amplitudes are calculated, while the zero level is considered to be the value at the point of the beginning of the QRS.

Additional quality assessment.

In contrast to the previous case, where the ECG at rest was considered and only the vector of parameters for the CCC was used, in this case the analysis of clock-wise parameters in dynamics is required, which imposes increased requirements on the signal quality.

During actual shooting (registration), fragments of the recording may be severely damaged. However, the remaining segments can still be diagnosed. A procedure is required to process the search and exclusion of ECG areas with poor quality.

Below, to illustrate the method, one parameter from the group of clock cycles (as described above) is considered - the amplitude at the point J + 80 ms.

FIG. 40 shows a fragment from a sample under load of satisfactory quality. Lines (markers) 1410 on complexes 1401 mark the amplitude at point J + 80 ms. So, even in this case, the parameter is subject to rather serious fluctuations: a jump of 1 mm (100 μV) is normal. If you collect all possible amplitude values from some clean ECG fragment, you get some bell of normal distribution. In the presence of more serious artifacts, the distribution may turn out to be less regular in shape.

The algorithm for constructing cycle-by-cycle trends allows you to correctly find the average value in the described normal distribution, and also a procedure is used that finds and rejects the segments of the record on which the distribution is obtained in such a form that it is not possible to restore the average. Scheme of the specified procedure:

1. For each complex in the record, the correlation is calculated (calculated) with the corresponding minute CCC in the QRS-T region. All complexes with a correlation less than a predetermined threshold are eliminated to filter out gross artifacts that distorted the shape of the QRS or ST segment.

2. Trends of clock-by-cycle parameters are built: points of the beginning and end of the QRS, the peak of the T-wave, the amplitudes at the indicated points, as well as the point J + 80 ms.

3. Trends are (slightly) smoothed by simple averaging (to remove only single outliers), statistical parameters are calculated in a sliding manner over a small window: standard deviations and ranges (max-min).

4. If the range and standard deviation do not exceed the specified thresholds, the segment is considered good.

To optimize the calculations in the last point, in a particular case, it is sufficient to estimate only (at least) two parameters: the amplitudes at points J and J + 80 (i.e., in fact, the difference between the signal values at the indicated points with the amplitude of the QRS onset). The time points themselves in the complexes that have passed the test for correlation from item 1 are determined quite stably.

FIG. 41-FIG. 45 is an illustration of the operation of the algorithm using an example of a sample under load of poor quality.

FIG. 41 shows three trends: green (1421) - heart rate, blue (1431) - amplitude J + 80, red (1441) - standard deviation, according to the value of which a decision on the suitability of a segment is made. ECG plots (1451, 1452, 1453, 1454) are shown in FIG. 42-FIG. 45.

FIG. 42 shows a section of the ECG (1451), in particular, shows the initial part of the recording in which a small deviation value is recorded, i.e. this fragment is "good" and is suitable for recording the initial values of the ECG parameters (before exercise).

FIG. 43 shows a section of the ECG (1452), in particular, shows a noisy section during the stress phase. In a particular case, the parameters are unsuitable for analysis.

FIG. 44 shows a section of the ECG (1453), in particular, in which short clean fragments slip among the noise, which can be used in a model for determining coronary heart disease.

FIG. 45 shows an ECG section (1454), in particular, in which the quality is good in the recovery phase with the exception of a small fragment.

Thus, the proposed algorithm makes it possible to successfully find ECG fragments that can be (and, in a particular case, should) be excluded from the ischemia analysis.

In a particular case, when noise damage is critical: for example, too many fragments are damaged or an important area, for example, a load peak, where parameter deviations are most informative. In this case, the algorithm gives the doctor (user of the system) an appropriate notification (alarm) and notification, in particular, a message about the absence of the need to evaluate IHD.

Construction of cycle-by-cycle trends of the parameters of interest.

To assess the dynamics of parameters, one-to-one trends can be built throughout the entire recording. When generating a graph containing complex-counted values, the result can be a fairly noisy curve (an illustration of the spread of values is shown in FIG. 41-FIG. 45).

To smooth the trend, a method / method / algorithm can be used, for example, a moving median, in which the median is calculated using a window of a fixed width, and the window itself is shifted by one complex each time. On the one hand, the median is not sensitive to outliers, on the other hand, the choice of a suitable window width can be used to obtain the desired smoothing result. A fast algorithm can be used for the calculation. At the same time, complexes that were not marked as abnormal as a result of the procedures described above are not considered (per-minute CCC, marking and additional quality assessment).

An example of smoothing the amplitude trend at point J + 80 ms is shown in FIG. 46.

FIG. 46 shows an example of smoothing the amplitude trend at point J + 80 ms.

In a particular case, using the above-mentioned algorithm, for each position of the sliding window, a search for the center described above (additional quality assessment) of the distribution bell is carried out.

This approach can be used to build trends of all parameters necessary for diagnosing IHD, which were measured in a cycle-by-cycle manner, as described above (cycle-by-cycle marking). At the same time, simple interpolation can be carried out on the segments that were evaluated by the algorithm as bad.

FIG. 47 shows an example of calculating trend values. Red (1517) - pulse, purple (1527) - amplitude at point J, green (1537) - amplitude at point J + 80 ms, blue (1547) - T-wave asymmetry index.

Assessment of the applicability of various methods (approaches, methods).

Depending on the presence of noise in the signal, some groups of parameters may be distorted. However, for the rest, it is still possible to reliably assess the presence of coronary artery disease. To determine the possibility of using different groups of parameters, an appropriate algorithm can be used.

The minimum requirement for assessing ischemia is the ability to reliably track the dynamics of the ST segment. Records where it is impossible to do this are eliminated as a result of the procedure described above (a rolling estimate of the variance and spread of the parameters - the amplitude at points J and J + 80 ms).

One of the possible implementations of the mentioned algorithm for highly noisy recordings is the weakening of the thresholds for variance and range, described above (additional quality assessment), but at the same time, to assess the dynamics, not per-cycle trends, but a sequence of ST-segment parameters of per-minute CCCs can be used. In a particular case, the averaging procedure effectively suppresses high-frequency noise, and the information regarding ST is at low frequencies, which are stored without distortion. Therefore, as a result of averaging, the indicated useful information can be obtained even from very noisy recordings.

Another group of parameters, which in a particular case may be uninformative.

- signs of a T-wave, which may be the result of a strong noise and / or the absence of a pronounced T-wave in the observed lead (physiological cause). Both situations are depicted in FIG. 48 and FIG. 49.

FIG. 48 shows an example of a poor quality T-wave.

FIG. 49 shows an example of a plane T-wave in lead V5.

The following parameters can be used to track the described cases:

- large variability of the positions of the points of maximum steepness of the T-wave fronts;

- large variability of values of the maximum steepness of the T-wave fronts;

- too small values of the maximum steepness of the T-wave fronts.

In a particular case, if at least one point of the condition is met, then the parameters of the T-wave are not used further (the first two conditions indicate a high noise level, the last - the lack of expression of the T-wave in the removed lead).

The last group of parameters that can be distorted due to poor quality is the HF QRS information (features based on the HF QRS component, RAZ parameter). The following RF noise estimation procedure can be used to diagnose this event.

In a particular case, all RF-based features are extremely sensitive to noise, as described above (QRS-based features), which is often present in recordings. In this case, in order to avoid misdiagnosis, these signs can be ignored. Unlike the case of an ECG at rest, where, in a particular case, only one value averaged over the entire record is important, in this case the dynamics of the parameters is important. But at the same time, only a part of the recording can be damaged by HF noise (in the particular case, most often in the area of the load peak), therefore, the noise assessment can be carried out on a cycle-by-cycle basis.

In a particular case, the critical parameter is precisely the ratio of the level of the useful RF QRS signal to the noise level. And since the physiological variation of the HF component is very large, the threshold for the noise level acceptable for the analysis of the recording also "floats" a lot. So, for example, if during IHD under load there is a drop in the HFQRS amplitude from 100 μV to 50 μV, then the noise level of 20 μV can be quite satisfactory - a pathological HFQRS drawdown will be visible. But with the amplitude of the physiological RF component of 15 μV, nothing will be recognized (seen) under 20 μV of noise.

In connection with the indicated fact, a clock-by-clock parameter can be added to the algorithm - the signal-to-noise ratio for the RF component of the QRS, which is calculated as follows. High frequencies are extracted from the signal using a low-pass filter (i.e., the difference between the original signal and the filter output is taken - an anti-filter circuit). The extracted RF component in the QRS region is calculated as the sum of the useful signal and noise, and outside the QRS region it is pure noise (P, T waves do not have their own RF components).

In a particular case, the areas where the calculated signal-to-noise ratio is small are not used to estimate the HFQRS parameter, and if there are many such areas, then the HF information may not be used to diagnose IHD.

Models for the diagnosis of ischemic heart disease.

Based on the parameters calculated as described above (basic algorithm for ECG marking and filtering, including peak search, filtering, classification of complexes, construction and marking of EKC, marking algorithm, as well as an algorithm for evaluating data quality and description and algorithms for calculating complex features based on wavelets , including features based on the first derivative wavelet transform, maximum line profile, QRS angles, second derivative, ST-T segment shape, T-wave asymmetry, as well as features based on the high-frequency QRS component and RAZ parameter (reduced amplitude zone),), a model is built to assess IHD.

Depending on the type of model (which are below) and the quality of the data, different subsets of the parameters can be used. The minimum set is the amplitudes at points J and J + 80mc. They allow you to evaluate the rise / depression of the ST-segment, and also (in particular, in the most general terms) describe its shape. In a particular case, the main working parameters for stress tests are the following 3 groups: ST parameters, T-wave signs, QRS HF information.

The model uses a vector of initial parameter values, as well as maxima / minima during the load and recovery phase.

In particular, in the group of basic parameters, the following quantities are of interest:

- parameters ST-T, HFQRS before load;

- minimums / maxima of amplitudes at points J and J + 80 ms in the load / recovery phase;

- maximum T-wave asymmetry for the entire recording time (load phase + recovery);

- maximum and minimum HFQRS in the load phase.

At the same time, depending on the version of the algorithm, one-to-one trends or minute-by-minute CCTs can be used.

So, the following basic models are built:

1. Model 1 - works only according to the characteristics of the "ST" group: change in the amplitude at points J and J + 80 ms during the period of the load phase. This model is used when the recording quality is poor: if the T-wave is not pronounced or is corrupted by noise, there is a high level of RF noise.

2. Model 2 - signs from the "ST" group + T-wave asymmetry. This model can be used with high RF noise levels.

3. Model 3 - traits from the "ST" group + HFQRS trait. A model for recordings with a bad T-wave.

4. Model 4 - all signs: ST + asymmetry T + HFQRS. Basic model for recordings of satisfactory quality.

5. Any of the models can be supplemented with an extended set of features that are described above (the basic algorithm for marking and filtering the ECG, including peak search, filtering, classification of complexes, construction and marking of the ECG, a marking algorithm, as well as an algorithm for assessing data quality and description and algorithms computation of complex features based on wavelets, including features based on the first derivative wavelet transform, maximum line profile, QRS angles, second derivative, ST-T segment shape, T-wave asymmetry, and features based on the high-frequency component of the QRS and the RAZ parameter (reduced amplitude zone), as well as the relationship of the parameters with the presence of IHD): in detail the ST-T parameters, the QRS wavelet energy, the QRS / T energy ratio, etc.

The structure of the models (depending on the algorithm and its modification, in particular, the version):

- classical linear discriminant.

- a neural network with one hidden layer of three neurons with a logistic activation function (as shown in FIG. 35).

The models provide a numerical ischemia index shown in FIG. 50.

FIG. 50 shows an exemplary ischemic index model.

If this value is positive, automatic diagnosis - no coronary artery disease, negative - there is. The absolute value indicates the probability (the more, the higher the confidence coefficient).

The following are exemplary options for the operation of the method / method / algorithm described above, (in particular, the entire sequence: ECG processing, calculation of parameters, application of the model).

FIG. 51-FIG. 55 shows an example of data for: angiography - negative, test - negative, algorithm - no ischemic heart disease.

FIG. 51 shows the trends of values throughout the record: red (1811) - pulse (HR), violet (1821) - amplitude at point J (Jamp), green (1831) - amplitude at point J + 80 ms (J80amp), blue (1841 ) - the index of asymmetry T (the ratio of the absolute values of the derivatives at the points of maximum steepness of the leading and trailing edges of the T-wave, BetaSm), brown (1851) - the value of HFQRS (HFQRS). The timeline is signed in minutes.

FIG. 52 shows the fragments of the signal tape corresponding to the beginning of the recording.

FIG. 53 shows the fragments of the signal tape corresponding to the peak of the load.

FIG. 54 shows the fragments of the signal tape corresponding to the recovery phase.

FIG. 55A shows indicators with model responses: ST - only based on the shape of the ST segment.

FIG. 55B shows indicators with model responses: ST + T - ST-segment shape and T-wave parameters.

FIG. 55B shows indicators with model responses: ST + HF - the shape of the ST segment and the HF component of the QRS.

FIG. 55G shows the indicators with the responses of the models: Integral - the combined result of all models (considered the final response of the algorithm).

FIG. 56-FIG. 60 shows an example of data for: angiography - positive, test - positive, algorithm - there is an ischemic heart disease.

According to the amplitude trends at points J and J + 80 ms, a pronounced depression of the ST segment (flat shape) is observed, thus a positive test is determined. According to the BetaSm and HFQRS parameters, the presence of ischemic heart disease is also observed: under load, the former increased sharply (the T-wave form became pronounced asymmetric with steeper leading edges), the latter decreased. The answers of all models are unanimous.

FIG. 56 shows the trends of values throughout the record: red (1911) - pulse (HR), purple (1921) - amplitude at point J (Jamp), green (1931) - amplitude at point J + 80 ms (J80amp), blue (1941 ) - the index of asymmetry T (the ratio of the absolute values of the derivatives at the points of maximum steepness of the leading and trailing edges of the T-wave, BetaSm), brown (1951) - the value of HFQRS (HFQRS). The timeline is signed in minutes.

FIG. 57 shows the fragments of the signal tape corresponding to the beginning of the recording.

FIG. 58 shows the fragments of the signal tape corresponding to the peak of the load.

FIG. 59 shows the fragments of the signal tape corresponding to the recovery phase.

FIG. 60A shows indicators with model responses: ST - only based on the shape of the ST segment.

FIG. 60B shows the indicators with the answers of the models: ST + T - the shape of the ST segment and the parameters of the T-wave.

FIG. 60B shows the indicators with the responses of the models: ST + HF - the shape of the ST segment and the HF component of the QRS.

FIG. 60G shows the indicators with the responses of the models: Integral - the combined result of all models (considered the final response of the algorithm).

FIG. 61-FIG. 65 shows an example of data for: angiography - positive, test - negative, algorithm - there is an ischemic heart disease.

In this example, the depression of the ST-segment is insufficient for the positiveness of the sample. However, the integral model, based on the set of parameters, correctly finds IHD.

FIG. 61 shows the trends of values throughout the record: red (2011) - pulse (HR), violet (2021) - amplitude at point J (Jamp), green (2031) - amplitude at point J + 80 ms (J80amp), blue (2041 ) - the index of asymmetry T (the ratio of the absolute values of the derivatives at the points of maximum steepness of the leading and trailing edges of the T-wave, BetaSm), brown (2051) - the value of HFQRS (HFQRS). The timeline is signed in minutes.

FIG. 62 shows the fragments of the signal tape corresponding to the beginning of the recording.

FIG. 63 shows the fragments of the signal tape corresponding to the peak load

FIG. 64 shows the fragments of the signal tape corresponding to the recovery phase.

FIG. 65A shows indicators with model responses: ST - only based on the shape of the ST segment.

FIG. 65B shows the indicators with the answers of the models: ST + T - the shape of the ST segment and the parameters of the T-wave.

FIG. 65B shows the indicators with the responses of the models: ST + HF - the shape of the ST segment and the HF component of the QRS.

FIG. 65G shows the indicators with the responses of the models: Integral - the combined result of all models (considered the final response of the algorithm).

FIG. 66 is a block diagram of the above algorithm.

In a particular case, one of the objectives of the present invention is to improve the reliability of automatic diagnosis of coronary artery disease by stress tests using ECG records recorded (obtained) at rest before and after testing (I standard lead), while using the approaches and models described above, in in particular, for cases of ECG at rest + under stress.

At step 2206, a sample under load is obtained as described in the context of the present invention.

Next, in step 2210, processing of the obtained samples under load is carried out, as described in the framework of the present invention. The processing includes the following phases (as shown in FIG. 67): peak search, filtering, classification of complexes, construction and labeling of CCCs, recording quality assessment, and parameter computation. The result is the parameters supplied to the input of the model, as well as the quality assessment, on which the choice of a particular model depends.

Next, in step 2216, the “ST feature model” is applied to the processing results depending on the signal quality in step 2216, or the “ST + T feature model” is applied in step 2020, or the “ST + HFQRS feature model” is applied in step 2026, or in step 2030 apply the "ST + T + HFQRS feature model" as described in the framework of the present invention.

Thus, the four models (2216, 2220, 2226, 2230) on the left side of FIG. 66 calculate the numerical index of ischemia (in fact, the likelihood of having an ischemic heart disease) from the test under load. The arrows from left to right indicate the transfer from them of the value of the indicated numerical index to one of the integral models shown on the right (2246, 2250, 2256, 2260).

And the result of applying one of the models is the value of the IHD assessment under load.

At step 2236, a resting ECG is recorded (recorded) just prior to exercise, as described in the context of the present invention.

Next, in step 2240, pre-load write processing (including the above-described phases as shown in FIG. 66) is performed as described in the context of the present invention.

At step 2242, a "resting ECG model" is applied, resulting in an IHD-rest estimate value.

Depending on the signal quality, Step 2246 applies ST Model + Rest, or Step 2250 applies Model ST + T + Rest, or Step 2256 applies Model ST + HFQRS + Rest, or step 2260 applies Model ST + T + HFQRS + rest. "

And the result of using one of the models is the value of the integral indicator of coronary heart disease.

FIG. 67 shows a block diagram of an exemplary ECG processing step that is equally applied in two branches of the algorithm: for stress tests (2210, FIG. 66) and recordings at rest (2242, FIG. 66).

Step 3010 searches for peaks as described in the context of the present invention.

At 3014, adaptive filtering is performed as described in the context of the present invention.

In step 3018, the classification of the complexes is carried out as described in the framework of the present invention.

In step 3022, the construction of the SCC is carried out, as described in the framework of the present invention.

At step 3026, CCC markup is performed as described in the context of the present invention.

Step 3030 evaluates signal quality as described in the context of the present invention.

In step 3034, the parameters are calculated as described in the context of the present invention.

FIG. 68 shows an example of a general-purpose computer system that includes a general purpose computing device in the form of a computer 20 or server including a processor 21, system memory 22, and a system bus 23 that links various system components including system memory to processor 21.

The system bus 23 can be any of various types of bus structures including a memory bus or memory controller, a peripheral bus, and a local bus using any of a variety of bus architectures. System memory includes read only memory (ROM) 24 and random access memory (RAM) 25. ROM 24 stores a basic input / output system (BIOS) 26, consisting of basic routines that help communicate information between elements within a computer 20, for example, at the time of launch.

Computer 20 may also include a hard disk drive 27 for reading from and writing to a hard disk not shown, a magnetic disk drive 28 for reading from or writing to a removable magnetic disk 29, and an optical disk drive 30 for reading from or recording on a removable optical disc 31 such as a compact disc, digital video disc, and other optical media. The hard disk drive 27, the magnetic disk drive 28, and the optical disk drive 30 are connected to the system bus 23 via a hard disk drive interface 32, a magnetic disk drive interface 33, and an optical drive interface 34, respectively. The storage devices and their corresponding computer-readable means provide non-volatile storage of computer-readable instructions, data structures, program modules, and other data for the computer 20.

Although the typical configuration described herein uses a hard disk, a removable magnetic disk 29, and a removable optical disk 31, one skilled in the art will appreciate that other types of computer-readable media may also be used in a typical operating environment that can store data that is accessible by a computer. such as magnetic cassettes, flash memory cards, digital video discs, Bernoulli cartridges, random access memory (RAM), read only memory (ROM), and the like.

Various software modules, including the operating system 35, may be stored on a hard disk, magnetic disk 29, optical disk 31, ROM 24, or RAM 25. Computer 20 includes a file system 36 associated with or included in operating system 35, one or more software application 37, other program modules 38, and program data 39. A user can enter commands and information into computer 20 using input devices such as keyboard 40 and pointing device 42. Other input devices (not shown) may include a microphone , joystick, gamepad, satellite dish, scanner or any other.

These and other input devices are connected to the processor 21, often through a serial port interface 46 that is connected to the system bus, but may be connected through other interfaces such as a parallel port, game port, or universal serial bus (USB) Monitor 47 or other type of device the visual display is also connected to the system bus 23 via an interface such as a video adapter 48. In addition to a monitor 47, personal computers typically include other peripheral output devices (not shown) such as speakers and printers.

Computer 20 can operate in a networked environment through logical connections to one or more remote computers 49. The remote computer (or computers) 49 can be another computer, server, router, networked PC, peer-to-peer device, or other node on a single network, and also typically includes includes most or all of the elements described above in relation to computer 20, although only storage device 50 is shown. Logical connections include a local area network (LAN) 51 and a wide area computer network (WAN) 52. Such networking environments are common in institutions , corporate computer networks, the Internet.

A computer 20 used in a LAN environment connects to a local area network 51 via a network interface or adapter 53. A computer 20 used in a GCS network environment typically uses a modem 54 or other means to communicate with a global computer network 52 such as the Internet.

Modem 54, which can be internal or external, is connected to the system bus 23 via a serial port interface 46. In a networked environment, program modules or portions thereof described with respect to computer 20 may be stored on a remote storage device. It will be appreciated that the network connections shown are typical and other means may be used to establish communication links between computers.

In conclusion, it should be noted that the information given in the description are examples, which do not limit the scope of the present invention defined by the claims. One skilled in the art will appreciate that there may be other embodiments of the present invention consistent with the spirit and scope of the present invention.

Claims (31)

1. A method for non-invasive computer diagnostics of coronary heart disease includes the following steps: - receive ECG signals in lead I of the patient at rest; - receive ECG signals in lead V5 of the patient during a stress test; - process the received ECG signals, forming a vector of signs under load and a vector of signs of rest, determine the quality of the ECG signal; - determine the assessment of coronary heart disease at rest using a model for ECG at rest on the basis of the generated vector of signs of rest; - determine the assessment of coronary heart disease during exercise using a model for ECG under load on the basis of the generated vector of signs under load, and the choice of a specific model depends on the quality of the ECG signal; - determine the integral indicator of coronary heart disease using the assessment of coronary heart disease at rest and the assessment of coronary heart disease during a test under stress. 2. The method according to claim 1, in which, depending on the quality of the ECG signal, the following pairs of models are used: (ST signs and ST signs + rest), (ST + T signs and ST + T signs + rest), (ST + HFQRS signs and ST + HFQRS signs + rest), (ST + T + HFQRS signs and ST + T signs + HFQRS + rest). 3. The method according to claim 1, in which the processing of ECG signals includes the search for peaks, adaptive filtering, classification of complexes, construction and labeling of the ECG. 4. The method of claim 1, wherein the models are based on linear or quadratic discriminant. 5. The method of claim 1, wherein the models are based on a neural network. 6. The method of claim 1, wherein the models are based on a combination of a linear or quadratic discriminant and a neural network. 7. A method for non-invasive computer diagnostics of coronary heart disease includes the following steps: - receive ECG signals in lead I of the patient at rest; - process the received ECG signals, forming a vector of signs of rest; - determine the assessment of coronary heart disease at rest using at least one model for ECG resting on the basis of the generated vector of signs of rest. 8. The method according to claim 7, in which, depending on the quality of the ECG signal, the following pairs of models are used: (ST signs and ST signs + rest), (ST + T signs and ST + T signs + rest), (ST + HFQRS signs and ST + HFQRS signs + rest), (ST + T + HFQRS signs and ST + T signs + HFQRS + rest). 9. The method according to claim. 7, in which the processing of ECG signals includes the search for peaks, adaptive filtering, classification of complexes, construction and labeling of the ECG. 10. The method of claim 7, wherein the models are based on linear or quadratic discriminant. 11. The method of claim 7, wherein the models are based on a neural network. 12. The method of claim 7, wherein the models are based on a combination of a linear or quadratic discriminant and a neural network. 13. The method of non-invasive computer diagnostics of coronary heart disease includes the following steps: - receive ECG signals in lead V5 of the patient during a stress test; - the received ECG signals are processed, forming a vector of signs under load, the quality of the ECG signal is determined; - determine the assessment of coronary heart disease during exercise using a model for ECG under load on the basis of the generated vector of signs under load, and the choice of the model depends on the quality of the ECG signal. 14. The method according to claim 13, wherein the processing of the ECG signals includes finding peaks, adaptive filtering, classification of complexes, construction and marking of the ECG. 15. The method of claim 13, wherein the models are based on linear or quadratic discriminant. 16. The method of claim 13, wherein the models are based on a neural network. 17. The method of claim 13, wherein the models are based on a combination of a linear or quadratic discriminant and a neural network.

Images