An ultrahigh frequency electrocardiogram mapping analysis method and system

Abstract

The application provides an ultrahigh frequency electrocardiogram mapping analysis method and system, through a Hilbert transform algorithm, time-frequency characteristics of high-frequency electrocardiogram signals can be retained, and influence of complex clutter on the high-frequency electrocardiogram signals can be reduced to the maximum extent; through drawing of an electrical activation map, time-space distribution of electrical activity is described in a graphical manner, and atrium / ventricle asynchrony or myocardial ischemia problems can be more accurately revealed; through transfer learning, a model can better understand and process new data, and the problem of poor model training effect caused by less high-frequency electrocardiogram data is effectively solved; through construction of an intelligent diagnosis system, the automation level of diagnosis is improved, and auxiliary diagnosis of inputting an electrical activation map and outputting a classification result is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of biomedical electronics, and particularly relates to an ultra-high frequency electrocardiogram mapping analysis method and system. BACKGROUND

[0002] Electrocardiogram technology is a basic tool for clinical diagnosis of cardiovascular diseases, and the technology has been developed for more than a hundred years. The frequency response range of an electrocardio collection device is usually 0-150 Hz. Since the 1980s, Goldberg et al. have collected high-frequency electrocardio with a frequency response of 150-250 Hz, proposed a corresponding evaluation index, root mean square (RMS), to judge myocardial ischemia, and improved the accuracy of myocardial ischemia diagnosis. In the following more than 20 years, Abboud S et al. enriched the detection means of high-frequency QRS, studied the high-frequency electrocardio information with a frequency upper limit of 250 Hz or 300 Hz, proposed a reduced amplitude zone (RAZ) index, and explored its application in the field of myocardial ischemia detection. With the development of digital signal processing and electronic technology, ultra-high frequency electrocardio mapping technology (frequency response exceeding 300 Hz) has begun to attract attention, and is expected to provide a potential non-invasive and quantitative new mapping method for clinical detection of cardiac electrophysiology, myocardial ischemia and atrial / ventricular dyssynchrony. SUMMARY

[0003] In view of the current lack of ultra-high frequency electrocardio mapping analysis devices, especially the lack of non-invasive and quantitative analysis of cardiac dyssynchrony and myocardial ischemia, the application provides an ultra-high frequency electrocardio mapping analysis method and system. The method is non-invasive and can provide quantitative and positioning analysis to assist doctors in evaluating the applicability of myocardial ischemia and cardiac resynchronization therapy. Meanwhile, machine learning and deep learning techniques are combined to improve the diagnostic accuracy and automation level.

[0004] The specific technical solutions are as follows:

[0005] An ultra-high frequency electrocardio mapping analysis method,

[0006] S1: acquiring high-frequency electrocardio data: collecting high-frequency electrocardio signal data by using a high-frequency electrocardio collection device; the high-frequency electrocardio signal data refers to electrocardio waveform data collected by a high-frequency electrocardio signal collection device with an ADC precision of greater than or equal to 24 bits and 6-18 leads, and a frequency response range of 0-1000 Hz; the leads include chest leads, limb leads and pressurized limb leads;

[0007] S2: data preprocessing: using a Butterworth IIR filter and a notch filter to preprocess the high-frequency electrocardiogram signal data collected in S1;

[0008] S3: R-wave-based beat division: using the Pan-Tompkins algorithm to find the R wave in the high-frequency electrocardiogram signal data collected by the limb leads or the compressed limb leads after preprocessing in S2; and recording the time nodes of the waveforms at the specified time before and after the peak value of the R wave; according to the recorded time nodes, the high-frequency electrocardiogram signal collected by the chest leads is cut to obtain a high-frequency electrocardiogram signal of a heart beat, i.e. a heart beat;

[0009] S4: feature extraction: using the Hilbert transform algorithm to calculate the amplitude envelope of the QRS wave in the high-frequency electrocardiogram signal in each frequency band in the continuous frequency band for the heart beat obtained in S3; the continuous frequency band refers to a series of continuous frequency ranges, one frequency range is one frequency band, and the difference between the upper limit and the lower limit of each frequency range is the same;

[0010] S5: signal averaging: performing an averaging operation on the amplitude envelope of the QRS wave obtained in S4 by signal averaging technology; performing low-pass filtering on the averaged amplitude envelope to form a feature vector corresponding to the amplitude envelope corresponding to each lead;

[0011] S6: drawing an electrical activation map: assembling a feature matrix, i.e. an electrical activity conduction matrix, from the feature vectors of all chest leads in the lead used in S1; calculating the time difference between lead activation, i.e. the dyssynchrony time difference, from the electrical activity conduction matrix; normalizing the electrical activity conduction matrix by a normalization algorithm so that its amplitude is between 0 and 1; using the colormap algorithm to draw a heat map from the normalized electrical activity conduction matrix to obtain an electrical activation map;

[0012] S7: pattern recognition: inputting the electrical activation map in S6 into a VGG19 model based on transfer learning, and outputting a preliminary auxiliary diagnosis result of detecting atrial / ventricular dyssynchrony or myocardial ischemia or no detection;

[0013] S8: result visualization: displaying the high-frequency electrocardiogram signal collected by the 6-18 leads in S1, the electrical activation map obtained in S6, the dyssynchrony time difference, and the preliminary auxiliary diagnosis result obtained in S7 on the user interface.

[0014] Preferably, the 6-18 leads include: 6, 9, 12, 14, and 18 leads; wherein:

[0015] The 6 leads are six chest leads V1-V6;

[0016] The 9 leads are I, II, and III limb leads based on the 6 leads;

[0017] The 12 leads are based on the 9 leads and have three pressurized limb leads added, i.e., the standard 12 leads.

[0018] The 14 leads are based on the 12 leads and have V7 and V8 chest leads added;

[0019] The 18-lead system is based on the 14-lead system and adds four more leads, namely V9, V3R, V4R, and V5R. The V3R lead is located on the right chest, symmetrical to V3; the V4R lead is located on the right chest, symmetrical to V4; the V5R lead is located on the right chest, symmetrical to V5; and V9 is the left posterior chest wall lead.

[0020] Preferably, the data preprocessing method described in S2 is: using a Butterworth IIR filter to filter out low-frequency baseline drift of the high-frequency ECG signal data; using a notch filter to filter out power frequency interference of a specified frequency in the high-frequency ECG signal data.

[0021] Preferably, the Hilbert transform algorithm described in S4 is to convert the divided high-frequency ECG signal into an amplitude envelope signal, and the specific steps are as follows:

[0022] S41: Calculate Fourier transform: Set the Fourier transform of the high-frequency ECG signal x(t) after heart beat division to X(jw), and according to the Fourier transform formula, we get:

[0023]

[0024] In formula (1), x * (t) is the conjugate of x(t), j is the imaginary unit; if the phase is changed by 90°, that is, ω is -ω, then x can be calculated * The Fourier transform of (t) is X * (-jω);

[0025] S42: Calculate the Hilbert transform: The Fourier transform of a real signal contains negative frequencies. If you want to remove the negative frequencies but keep the total power unchanged, multiply X(jw) by 2u(t), where u(t) is a step function, to obtain the high-frequency ECG signal after the Hilbert transform:

[0026]

[0027] The frequency response characteristics are:

[0028]

[0029]

[0030] It can be seen from formula (4) that the Hilbert transform shifts the phase of all positive frequency components by 90 degrees and shifts the phase of all negative frequency components by 90 degrees;

[0031] S43: Constructing the analysis signal: turning the real signal into a complex signal, taking the high-frequency ECG signal x(t) after the beat division as the real part, and taking the high-frequency ECG signal after the Hilbert transform as the imaginary part, to obtain:

[0032]

[0033] The high-frequency ECG signal x(t) after the beat division has both amplitude information and phase information, and x(t) is set as:

[0034] x(t) = A(t)cos(ωt + θ(t)) (6)

[0035] wherein A(t) represents the amplitude information of the signal, and θ(t) represents the phase information.

[0036] After substituting formula (6) into formula (5), the following is obtained:

[0037]

[0038] S44: Calculating the amplitude envelope signal: taking the absolute value of to obtain the amplitude envelope signal:

[0039]

[0040] Preferably, the specific steps of the signal averaging of S5 are as follows:

[0041] S51: averaging the amplitude envelope of the QRS wave corresponding to all the beats in the same lead and the same frequency band to obtain the amplitude envelope of the simplified QRS wave;

[0042] S52: calculating the median of each time point in the amplitude envelope of the simplified QRS wave in all frequency bands, i.e., the feature vector of each lead; the median is a one-dimensional feature vector of 1 x the division time length.

[0043] Preferably, the dyssynchrony time difference DYS of S6 is calculated by calculating the time difference between the lead with the earliest excitation and all the leads in the electrical activity conduction matrix; the group of adjacent leads with the largest time difference is taken as the main delay lead group MDL, and the time difference is the main delay time difference δDYS; the numerical value of the dyssynchrony time difference represents the degree of atrial / ventricular dyssynchrony, and the larger the numerical value, the stronger the degree of atrial / ventricular dyssynchrony.

[0044] Preferably, the specific method for obtaining the VGG19 model based on transfer learning in S7 is as follows:

[0045] ​S61: Pre-training the VGG19 model using the EGC dataset to obtain the pre-trained model weights; the intermediate variable of the VGG19 model is a replaceable feature variable, which is the key carrier of information transmission and processing between layers of the VGG19 model;

[0046] S62: Establishing a pattern recognition model, i.e., a VGG19 model; using a transfer learning algorithm, migrating the pre-trained model weights of S61 to the VGG19 model, and fine-tuning the VGG19 model using the EAG dataset to obtain a VGG19 model for detecting atrial / ventricular dyssynchrony or myocardial ischemia; the fine-tuning is to use the EAG as the input of the model, and the output is the preliminary auxiliary diagnostic result of detecting atrial / ventricular dyssynchrony or myocardial ischemia or no detection;

[0047] S63: Verifying the fine-tuned model, and the indicators for verifying the performance of the model include accuracy, F1 score, precision, and recall.

[0048] An ultrahigh frequency electrocardiogram mapping analysis system, comprising:

[0049] A data acquisition module: using a high-frequency electrocardiogram acquisition device to acquire high-frequency electrocardiogram signal data;

[0050] A data preprocessing module: filtering out low-frequency baseline drift in the acquired high-frequency electrocardiogram signal data through a Butterworth IIR filter unit; and filtering out power frequency interference of a specified frequency in the acquired electrocardiogram signal data using a notch filter;

[0051] A feature extraction module: calculating the amplitude envelope of the electrocardiogram signal data processed by the data preprocessing module within a specified frequency range using a Hilbert transform unit; performing averaging operation on the amplitude envelope obtained by the Hilbert transform unit through a signal averaging technique unit; and filtering the amplitude envelope averaged by the signal averaging technique unit through a low-pass filter unit to form a waveform corresponding to the amplitude envelope corresponding to each lead;

[0052] A visualization module: constructing a feature matrix, i.e., an electrical activity conduction matrix, by constructing a matrix unit using the feature vector of a specified lead; inputting the electrical activity conduction matrix into a calculation dyssynchrony time difference unit to calculate the time difference between lead activations; and using a colormap algorithm unit to draw the electrical activity conduction matrix into a heat map to obtain an EAG;

[0053] A pattern recognition module: inputting the EAG of the visualization module into a VGG19 model unit to obtain a preliminary auxiliary diagnostic result of detecting atrial / ventricular dyssynchrony or myocardial ischemia or no detection;

[0054] A user interface: displaying the high-frequency electrocardiogram signal of the data acquisition module, the EAG and dyssynchrony time difference of the visualization module, the preliminary auxiliary diagnostic result of the pattern recognition module, and operation options.

[0055] Beneficial effects:

[0056] The present invention proposes an ultra-high frequency electrocardiogram (ECG) mapping and analysis method and system, which provides quantitative and localized analysis to assist doctors in assessing the suitability of myocardial ischemia and cardiac resynchronization therapy; at the same time, it combines machine learning and deep learning technologies to improve diagnostic accuracy and automation. Specifically,

[0057] By applying the Hilbert transform algorithm, the signal-to-noise ratio of ultra-high frequency signals can be significantly improved, effectively suppressing noise interference in the ultra-high frequency band, thereby protecting valid signals from being affected. This algorithm not only preserves the time-frequency characteristics of the signal, but also minimizes the interference of complex clutter on the signal, ensuring the accuracy and reliability of signal analysis.

[0058] Electrical activation mapping accurately reveals delays or blocks in cardiac electrical conduction and identifies the leads in which these problems occur. This mapping not only provides qualitative information on the extent of myocardial ischemia and cardiac dyssynchrony, but also enables quantitative and localized analysis, assisting physicians in assessing patient suitability for cardiac resynchronization therapy. By calculating the delay times of individual leads in the electrical activation map and the differences in delays between them, physicians can determine the degree of electrical dyssynchrony in specific ventricular regions.

[0059] By building an intelligent diagnostic system and pre-training the model, rich feature representations can be learned. Transfer learning can then help the model better understand and process new data, effectively addressing the issue of poor model training due to limited high-frequency ECG data and significantly improving model recognition accuracy. Machine learning and deep learning techniques are also used to improve diagnostic accuracy and automation. Deep learning based on VGG19, based on the presented electrical activation maps, enables automated and intelligent diagnosis of atrial / ventricular dyssynchrony or myocardial ischemia. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 A flow chart of an ultra-high frequency electrocardiogram mapping and analysis method.

[0061] Figure 2 Schematic diagram of the results of each step in the process of ultra-high frequency electrocardiogram mapping analysis method.

[0062] Figure 3 Schematic diagram of each step of the results of a specific example of an ultra-high frequency electrocardiogram mapping and analysis method.

[0063] Figure 4 Schematic diagram of electrical activation diagram of atrial / ventricular dyssynchrony group.

[0064] Figure 5 Schematic diagram of the electrical activation map of the normal group.

[0065] Figure 6 An ultra-high frequency electrocardiogram mapping analysis system architecture diagram. DETAILED DESCRIPTION

[0066] As Figure 1 shown, an ultra-high frequency electrocardiogram mapping analysis method comprises:

[0067] S1: Obtain high-frequency electrocardiogram data: use a high-frequency electrocardiogram acquisition device to collect high-frequency electrocardiogram signal data; the high-frequency electrocardiogram signal data refers to electrocardiogram waveform data collected by a high-frequency electrocardiogram signal acquisition device with a frequency response range of 0-1000 Hz, an ADC precision of greater than or equal to 24 bits, and 6-18 leads; the leads include chest leads, limb leads, and pressurized limb leads;

[0068] S2: Data preprocessing: use a Butterworth IIR filter and a notch filter to preprocess the high-frequency electrocardiogram signal data collected in S1;

[0069] S3: Heartbeat division based on R waves: use the Pan-Tompkins algorithm to find the R waves in the high-frequency electrocardiogram signal data collected by the limb leads or pressurized limb leads after preprocessing in S2; record the time nodes of the waveforms at a specified time length before and after the R wave peak; cut the high-frequency electrocardiogram signal collected by the chest leads according to the recorded time nodes to obtain a heartbeat of high-frequency electrocardiogram signal, i.e., a heart beat;

[0070] S4: Feature extraction: use the Hilbert transform algorithm to calculate the amplitude envelope of the QRS wave in the high-frequency electrocardiogram signal in each frequency band set artificially in the heart beat obtained in S3; the continuous frequency band refers to a series of continuous frequency ranges, one frequency range is one frequency band, and the difference between the upper limit and the lower limit of each frequency range is the same;

[0071] S5: Signal averaging: perform an averaging operation on the amplitude envelope of the QRS wave obtained in S4 through signal averaging technology; perform low-pass filtering on the averaged amplitude envelope to form a feature vector corresponding to the amplitude envelope corresponding to each lead;

[0072] S6: Draw an electrical activation map: assemble the feature vectors of all chest leads in S1 into a feature matrix, i.e., an electrical activity conduction matrix; calculate the time difference between lead activations, i.e., the asynchrony time difference, through the electrical activity conduction matrix; perform normalization processing on the electrical activity conduction matrix using a normalization algorithm so that the amplitude values are all between 0 and 1; use the colormap algorithm to draw a heat map of the normalized electrical activity conduction matrix to obtain an electrical activation map;

[0073] S7: Pattern recognition: input the S6 electrical activation map into the VGG19 model based on transfer learning, output the preliminary auxiliary diagnosis results of detecting atrial / ventricular dyssynchrony or myocardial ischemia or no detection;

[0074] S8: Result visualization: display the high-frequency electrocardiogram signal of 6-18 leads obtained in S1, the electrical activation map obtained in S6, and the preliminary auxiliary diagnosis results obtained in S7 on the user interface.

[0075] An ultrahigh-frequency electrocardiogram mapping analysis system, comprising:

[0076] Data acquisition module: high-frequency electrocardiogram signal data is collected by a high-frequency electrocardiogram acquisition device;

[0077] Data preprocessing module: low-frequency baseline drift in the collected high-frequency electrocardiogram signal data is filtered out by a Butterworth IIR filter unit; and power frequency interference at a specified frequency in the collected electrocardiogram signal data is filtered out by a notch filter;

[0078] Feature extraction module: the amplitude envelope of the electrocardiogram signal data processed by the data preprocessing module within a specified frequency range is calculated by a Hilbert transform unit; the amplitude envelope obtained by the Hilbert transform unit is averaged by a signal averaging technique unit; and the amplitude envelope averaged by the signal averaging technique unit is filtered by a low-pass filter unit to form a waveform corresponding to the amplitude envelope corresponding to each lead;

[0079] Visualization module: the feature vector of a specified lead is used to construct a feature matrix, i.e., an electrical activity conduction matrix, by a matrix construction unit; the electrical activity conduction matrix is input into a calculation of dyssynchrony time difference unit to calculate the time difference between lead activations; and the electrical activity conduction matrix is plotted into a heat map by a colormap algorithm unit to obtain an electrical activation map;

[0080] Pattern recognition module: the electrical activation map of the visualization module is input into a VGG19 model unit to obtain preliminary auxiliary diagnosis results of detecting atrial / ventricular dyssynchrony or myocardial ischemia or no detection;

[0081] User interface: display the high-frequency electrocardiogram signal of the data acquisition module, the electrical activation map and the dyssynchrony time difference of the visualization module, the preliminary auxiliary diagnosis results of the pattern recognition module, and the operation options.

[0082] Specific method:

[0083] As shown in Figure 2 and Figure 3 , a specific process of an ultrahigh-frequency electrocardiogram mapping analysis method, wherein Figure 2 the electrical dyssynchrony map and the electrical activation map in are different names for the same kind of map:

[0084] 1. Data acquisition:

[0085] The 6-18 lead ECG collection device with a sampling rate of not less than 4000 Hz, a frequency response range of 0-1000 Hz, and an ADC precision of not less than 24 bits is used to complete at least 1 minute of collection to obtain sufficient data. Taking 14 leads as an example, the leads include standard 12 leads (I lead, II lead, III lead, aVL lead, aVR lead, aVF lead, and six chest leads V1-V6) and two additional chest leads (V7, V8).

[0086] 2. Data preprocessing:

[0087] A Butterworth IIR filter is used for preprocessing to filter out low-frequency baseline drift; a notch filter is used to filter out 50 Hz power frequency interference. The Pan-Tompkins algorithm is used for R-wave-based heartbeat division, and the waveform of 120 ms before and after the R-wave peak is cut.

[0088] 3. Feature matrix solving, i.e., feature extraction:

[0089] The Hilbert transform is used to calculate the amplitude envelope in the frequency ranges of 100-150 Hz, 150-250 Hz, 200-300 Hz, 250-350 Hz, 300-400 Hz, 350-450 Hz, 400-500 Hz, …, 850-950 Hz, and 900-1000 Hz; the signal averaging technique is used to improve the signal-to-noise ratio, the waveform of 120 ms before and after each QRS wave divided in the time period is averaged, and the signal is low-pass filtered (passband 0-40 Hz) to form a waveform corresponding to each lead, which is arranged from top to bottom according to the V1-V8 leads to form a 2D matrix representing the distribution of electrical activity in time (relative to the R-wave time) and space (V lead).

[0090] Specifically, the signal averaging technique is specifically exemplified by extracting 10 s of 2-lead ECG, dividing it into 10 heartbeats, and then dividing it into 9 frequency bands. The 10 heartbeats in each frequency band are averaged, so that each frequency band is simplified to 1 heartbeat. The simplified heartbeats are then median filtered, that is, for each point in the heartbeat, the median value in the 9 frequency bands is taken.

[0091] After constructing the electrical activity conduction matrix, the time difference between the earliest excited lead and the time of all leads is taken as the dyssynchrony time difference DYS, and the group with the maximum time difference (dyssynchrony time difference) between adjacent leads is taken as the major delayed lead group (MDL), and the time difference is the major delayed time difference (Major Delayed Duration, denoted as δDYS).

[0092] Specifically, in the index calculation process, due to the ultra-high frequency signal signal-to-noise ratio is too low, the amplitude of the noise in the ultra-high frequency band is much larger than the amplitude of the effective signal, therefore, an algorithm is needed which can not only retain the time-frequency characteristics, but also can reduce the influence of complex clutter on the signal to the greatest extent - Hilbert transform algorithm.

[0093] Let the Fourier transform of x(t) be X(jw), then according to the Fourier transform formula:

[0094]

[0095] In formula (1), * (t) is the conjugate of x(t), j is the imaginary unit, if the phase is changed by 90 degrees, that is, ω is taken as -ω, then the Fourier transform of x * (t) can be calculated as X * (-jω); the Fourier transform of the real signal contains negative frequencies, if you want to remove the negative frequencies but the total power remains unchanged, multiply X(jw) by 2u(t), u(t) is a step function, and the Hilbert transform is obtained:

[0096]

[0097] The frequency response characteristics of the Hilbert transform are:

[0098]

[0099] As can be seen from formula (4), the Hilbert transform shifts the phase of all positive frequency components by -90 degrees, and shifts the phase of all negative frequency components by +90 degrees; construct an analytic signal, convert the real signal into a complex signal, let the original signal x(t) as the real part, and the signal after Hilbert transform as the imaginary part, get:

[0100]

[0101] The signal x(t) has both amplitude information and phase information, let

[0102]

[0103] Where, A(t) represents the amplitude information of the signal, and θ(t) represents the phase information.

[0104] Substituting formula (5) into formula (5) gives:

[0105]

[0106] Taking the absolute value of , the envelope signal can be obtained:

[0107]

[0108] It should be noted that in practical applications, in order to calculate the envelope of the ultra-high frequency band alone, band-pass filtering is required. However, the envelope signal still has different random errors due to noise, so the divided heart beat signals within 5 minutes are superimposed and divided by the total to obtain an average signal, and filtered below 40 Hz to ensure that the waveform is not affected by noise. The matrix obtained after processing the 8 leads is converted into a heat map (linear interpolation between rows) to obtain an electric activation map (EAD). Linear interpolation is an interpolation method for one-dimensional data, which estimates the value according to the left and right adjacent data points of the point to be interpolated in the one-dimensional data sequence. Usually, the average value is used for interpolation between leads.

[0109] EAD technology can accurately reveal the delay or block problem in cardiac electrical conduction and indicate where these problems occur near the lead. This graph not only provides qualitative information on the degree of myocardial ischemia and cardiac dyssynchrony, but also allows quantitative and localization analysis, thereby assisting doctors in assessing whether patients are suitable for cardiac resynchronization therapy. By calculating the delay time of each lead in EAD and the delay difference between them, doctors can determine the degree of electrical dyssynchrony in a specific ventricular region. These calculation results constitute the dyssynchronization parameter system (DPS), which provides a standardized index for evaluating and monitoring treatment effect. If some parameters in DPS are reduced, it indicates that the electrical dyssynchrony condition of a specific ventricular region (such as between the interventricular septum and the left ventricle, between the left and right ventricles, or within the left ventricle) has improved, and vice versa.

[0110] 4. Pattern recognition:

[0111] Classify and locate atrial / ventricular dyssynchrony or myocardial ischemia using machine learning algorithms, including atrial dyssynchrony, ventricular dyssynchrony, and atrial and ventricular dyssynchrony. The electric activation map can be used as an image classification vector input into the VGG19 model of transfer learning for deep learning to achieve image classification, outputting preliminary auxiliary diagnostic results of detecting atrial / ventricular dyssynchrony or myocardial ischemia or no detection. This is because the intermediate variable of the VGG19 network is a replaceable feature vector, facilitating transfer learning and maintaining high accuracy. Build an intelligent diagnostic system to improve the accuracy and automation level of diagnosis using machine learning and deep learning techniques, achieving auxiliary diagnosis by inputting electric activation maps and outputting classification results.

[0112] 5. Result visualization:

[0113] Design an intuitive user interface to display high-frequency electrocardiogram signals, ventricular conduction time differences, electric activation sequence maps (electric activation maps), dyssynchrony time differences (DYS), and operation options, etc.

[0114] As Figure 4 , Figure 5 shown, the electrical activation map of the atrial / ventricular dyssynchrony patient has a significant difference in V4, V6 leads and the remaining leads. The V6 lead of the dyssynchrony patient has a more obvious delay so that the activation point red band of V5 lead is broken. According to the algorithm, the main delay leads are V5 and V6 leads, and the maximum lead delay time difference is 24 ms.

[0115] As Figure 6 shown, a composition of an ultrahigh frequency electrocardiogram mapping analysis system is shown:

[0116] The preprocessing module:

[0117] The Butterworth IIR filter is used for preprocessing to filter out low-frequency baseline drift; the 50 Hz power frequency interference is filtered out using a notch filter. The Pan-Tompkins algorithm is used for R-wave-based heartbeat division, and the waveform of 120 ms before and after the R-wave peak is cut.

[0118] The feature extraction module:

[0119] The quantitative indicators of atrial / ventricular dyssynchrony are realized, including QRS complex width, interventricular conduction time difference, electrical activation sequence map (electrical activation map), dyssynchrony time difference (DYS), etc.

[0120] First, the amplitude envelope is calculated in the frequency range of 100-150 Hz, 150-250 Hz, 200-300 Hz, 250-350 Hz, 300-400 Hz, 350-450 Hz, 400-500 Hz, …, 850-950 Hz, 900-1000 Hz through Hilbert transform; the signal averaging technique is applied to improve the signal-to-noise ratio, the waveform of 120 ms before and after each QRS wave divided in the time period is averaged, and the signal is low-pass filtered (passband 0-40 Hz), forming the waveform corresponding to each lead. Arranging it from top to bottom according to V1-V8 leads, the 2D matrix formed represents the distribution of electrical activity in time (relative to R wave time) and space (V lead).

[0121] After constructing the electrical activity conduction matrix, the time difference between the earliest activation lead and the activation time of all leads is taken as the dyssynchrony time difference DYS, and the group with the maximum time difference between adjacent leads is taken as the major delay lead group (Major Delayed Leads, MDL), and the time difference is the major delay time difference (Major Delayed Duration, denoted as δDYS). In subsequent exploration, the effect of this set of index system on the diagnosis and evaluation of atrial / ventricular dyssynchrony or atrial / ventricular is evaluated, and compared with the dyssynchrony parameters and other indicators of echocardiogram.

[0122] Pattern recognition module:

[0123] The electrical activation map can be used as an image classification vector for deep learning using the VGG19 model. An intelligent diagnostic system is constructed to improve the accuracy and automation level of diagnosis using machine learning and deep learning techniques. Deep learning based on VGG19 is performed on the presented electrical activation map to achieve automated and intelligent diagnosis of atrial / ventricular dyssynchrony or myocardial ischemia.

[0124] Visualization module:

[0125] After obtaining the two-dimensional matrix, since the amplitudes of V1-V8 leads are inconsistent in terms of positive and negative, a normalization algorithm is used to normalize each V lead so that its amplitude is between 0 and 1. Based on the normalized matrix, an electrical activation map (heat map, maximum amplitude displayed as red, minimum amplitude displayed as blue) is created, and the envelope signal is arranged in order as V1 to V8 rows. Since there are only 8 leads, i.e. 8 rows of data, linear interpolation is performed between the rows to graphically describe the time-space distribution of electrical activity (depolarization process). The presented time region is ±120ms around the R wave. In the electrical activation map, color represents activation time and order. V3 lead appears red first, i.e. activates earliest, followed by V1, V2, etc.

[0126] User interface:

[0127] After obtaining the electrical activation map, an intuitive user interface is designed to display high-frequency electrocardiogram signals, ventricular conduction time differences, electrical activation sequence maps (electrical activation maps), dyssynchrony time differences (DYS), and operation options.

[0128] Specifically, by modeling simulation and clinical experiment, the changes of each DPS under different atrial / ventricular dyssynchrony or myocardial ischemia conditions can be determined to determine the cardiac electrical activity of patients with different types of dyssynchrony or myocardial ischemia. For a specific patient, a set of suitable diagnostic and mapping DPS can be determined to verify the index system.

[0129] Finally, it should be noted that the above description is only to illustrate the technical solutions of the present application and is not limiting. Although the present application has been described in detail with reference to the preferred arrangement, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application.

Claims

1. A method for ultra-high frequency electrocardiogram mapping and analysis, characterized in that: S1: Acquire high-frequency ECG data: Use a high-frequency ECG acquisition device to acquire high-frequency ECG signal data; the high-frequency ECG signal data refers to ECG waveform data acquired by a high-frequency ECG signal acquisition device with a frequency response range of 0 to 1000 Hz, an ADC accuracy greater than or equal to 24 bits, and 6-18 leads; the leads include chest leads, limb leads, and pressurized limb leads; S2: Data preprocessing: The high-frequency ECG signal data collected by S1 is preprocessed using Butterworth IIR filter and notch filter; S3: R-wave-based heartbeat segmentation: Use the Pan-Tompkins algorithm to find the R wave in the high-frequency ECG signal data collected by the limb leads or pressurized limb leads after S2 preprocessing; and record the time nodes of the waveform at the specified time length before and after the R wave peak; cut the high-frequency ECG signal collected by the chest leads according to the recorded time nodes to obtain the high-frequency ECG signal of one heartbeat, namely the heartbeat; S4: Feature extraction: Calculate the amplitude envelope of the QRS wave in the high-frequency electrocardiogram signal of each frequency band in a continuous frequency band using the Hilbert transform algorithm for the heartbeat obtained in S3; the continuous frequency band refers to a series of continuous frequency ranges, one frequency range is a frequency band, and the difference between the upper limit and the lower limit of each frequency range is the same; S5: Signal averaging: The amplitude envelope of the QRS wave obtained in S4 is averaged using the signal averaging technique; the averaged amplitude envelope is low-pass filtered to form a characteristic vector corresponding to the amplitude envelope of each lead; S6: Draw an electrical activation map: Use the eigenvectors of all chest leads in S1 to construct a feature matrix, namely the electrical activity conduction matrix; calculate the time difference between lead excitations, namely the asynchronous time difference, through the electrical activity conduction matrix; use a normalization algorithm to normalize the electrical activity conduction matrix so that its amplitude is between 0 and 1; use a colormap algorithm to plot the normalized electrical activity conduction matrix into a heat map to obtain the electrical activation map; S7: Pattern recognition: The S6 electrical activation map is input into the VGG19 model based on transfer learning, and the output is a preliminary auxiliary diagnosis result of whether atrial / ventricular dyssynchrony or myocardial ischemia is detected or not; S8: Result visualization: The high-frequency ECG signals collected by 6-18 leads obtained by S1, the electrical activation map and asynchronous time difference obtained by S6, and the preliminary auxiliary diagnosis results obtained by S7 are displayed on the user interface.

2. An ultra-high frequency electrocardiogram mapping and analysis method according to claim 1, characterized in that: The 6-18 leads include: 6, 9, 12, 14 and 18 leads; wherein: The 6 leads are six chest leads from V1 to V6; The 9-lead system is based on the 6-lead system with the addition of three limb leads, I, II, and III. The 12 leads are based on the 9 leads and have three pressurized limb leads added, i.e., the standard 12 leads. The 14-lead system is based on the 12-lead system and adds two chest leads, V7 and V8; The 18-lead system includes four more leads, namely V3R, V4R, V5R and V9, on the basis of the 14-lead system. The V3R lead is located on the right chest symmetrical to V3; the V4R lead is located on the right chest symmetrical to V4; the V5R lead is located on the right chest symmetrical to V5; and V9 is the left posterior chest wall lead.

3. An ultra-high frequency electrocardiogram mapping and analysis method according to claim 1, characterized in that: The data preprocessing method described in S2 is: using a Butterworth IIR filter to filter out the low-frequency baseline drift of the high-frequency ECG signal data; using a notch filter to filter out the power frequency interference of a specified frequency in the high-frequency ECG signal data.

4. An ultra-high frequency electrocardiogram mapping and analysis method according to claim 1, characterized in that: The Hilbert transform algorithm described in S4 converts the divided high-frequency ECG signal into an amplitude envelope signal. The specific steps are as follows: S41: Calculate Fourier transform: Set the Fourier transform of the high-frequency ECG signal x(t) after heart beat division to X(jw), and according to the Fourier transform formula, we get: In formula (1), x * (t) is the conjugate of x(t), j is the imaginary unit; if the phase is changed by 90°, that is, ω is -ω, then x can be calculated * The Fourier transform of (t) is X * (-jω); S42: Calculate the Hilbert transform: The Fourier transform of a real signal contains negative frequencies. If you want to remove the negative frequencies but keep the total power unchanged, multiply X(jw) by 2u(t), where u(t) is a step function, to obtain the high-frequency ECG signal after the Hilbert transform: The frequency response characteristics are: It can be seen from formula (4) that the Hilbert transform shifts the phase of all positive frequency components by 90 degrees and shifts the phase of all negative frequency components by 90 degrees; S43: Construct analytical signal: Convert the real signal into a complex signal, take the high-frequency ECG signal x(t) after the original heartbeat is divided as the real part, and the high-frequency ECG signal after Hilbert transform As the imaginary part, we get: The high-frequency ECG signal x(t) after heart beat division has both amplitude information and phase information. Let x(t) be: x(t)=A(t)cos(ωt+θ(t)) (6) Among them, A(t) represents the amplitude information of the signal, and θ(t) represents the phase information; Substituting formula (6) into formula (5), we get: S44: Calculate and obtain the amplitude envelope signal: Take the absolute value to get the amplitude envelope signal:

5. An ultra-high frequency electrocardiogram mapping and analysis method according to claim 1, characterized in that: The specific steps of signal averaging described in S5 are: S51: average the amplitude envelopes of the QRS waves corresponding to all S4 beats in the same lead and frequency band to obtain the simplified amplitude envelope of the QRS waves; S52: Calculate the median value of each time point in the simplified QRS wave amplitude envelope in all frequency bands, that is, the eigenvector of each lead; the median value is a one-dimensional eigenvector of 1×division time length.

6. An ultra-high frequency electrocardiogram mapping and analysis method according to claim 1, characterized in that: The asynchronous time difference, DYS, described in S6 is obtained by calculating the time difference between the lead that starts to be excited earliest and the time difference between all leads in the electrical activity conduction matrix; the group with the largest time difference between adjacent leads is taken as the main delay lead group, namely MDL, and this time difference is the main delay time difference, namely δDYS; the value of the asynchronous time difference represents the degree of atrial / ventricular asynchrony, and the larger the value, the stronger the degree of atrial / ventricular asynchrony.

7. An ultra-high frequency electrocardiogram mapping and analysis method according to claim 1, characterized in that: The specific method for obtaining the VGG19 model based on transfer learning described in step S7 is: S61: Pre-training the VGG19 model to obtain the pre-training model weight; the intermediate variable of the VGG19 model is a replaceable feature variable, which is the key carrier for transmitting and processing information between the layers of the VGG19 model; S62: Establishing a pattern recognition model, namely a VGG19 model; utilizing a transfer learning algorithm, migrating the weights of the S61 pre-trained model to the VGG19 model, and fine-tuning the VGG19 model using the electrical activation map dataset to obtain a VGG19 model for detecting atrial / ventricular dyssynchrony or myocardial ischemia; the fine-tuning uses the electrical activation map as input to the model, and outputs a preliminary auxiliary diagnosis result of whether atrial / ventricular dyssynchrony or myocardial ischemia is detected or not; S63: Verify the fine-tuned model. The performance indicators of the model include accuracy, F1 score, precision and recall.

8. An ultra-high frequency electrocardiogram mapping and analysis system formed based on the method according to any one of claims 1 to 7, characterized in that: include: Data acquisition module: uses high-frequency ECG acquisition equipment to collect high-frequency ECG signal data; Data preprocessing module: The Butterworth IIR filter unit is used to filter out the low-frequency baseline drift in the collected high-frequency ECG signal data; the notch filter is then used to filter out the power frequency interference of the specified frequency in the collected ECG signal data; Feature extraction module: uses the Hilbert transform unit to calculate the amplitude envelope of the ECG signal data processed by the data preprocessing module within the specified frequency range; The amplitude envelope obtained by the Hilbert transform unit is averaged by the signal averaging technology unit; the amplitude envelope averaged by the signal averaging technology unit is filtered by the low-pass filtering unit to form a waveform corresponding to the amplitude envelope corresponding to each lead; Visualization module: The eigenvector of the specified lead is constructed into a feature matrix, i.e., an electrical activity conduction matrix, through a matrix construction unit; the electrical activity conduction matrix is ​​input into a calculation unit for calculating asynchronous time difference to calculate the time difference between lead activations; the electrical activity conduction matrix is ​​plotted into a heat map using a colormap algorithm unit to obtain an electrical activation map; Pattern recognition module: The electrical activation map of the visualization module is input into the VGG19 model unit to obtain a preliminary auxiliary diagnosis result of whether atrial / ventricular dyssynchrony or myocardial ischemia is detected or not; User interface: displays the high-frequency ECG signals of the data acquisition module, the electrical activation map and asynchronous time difference of the visualization module, and the preliminary auxiliary diagnosis results and operation options of the pattern recognition module.

Images