A heart disease prediction system based on ECG signals and large models

Abstract

The present invention discloses a heart disease prediction system based on electrocardiogram signals and a large model, and relates to the field of electrocardiogram signal processing. The system comprises: using a heart suit to collect the electrocardiogram signals of the user, and using a data processing center to perform baseline offset suppression and noise processing operations on the electrocardiogram signals; extracting a variability sequence from the processed electrocardiogram signals, drawing an electrocardiogram scatter plot according to the variability, and partitioning and numbering the scatter points in the scatter plot to obtain the sequence information entropy of the heart rate variability sequence; segmenting the electrocardiogram signals based on the sequence information entropy and the peaks, and obtaining the waveform segments of a single heart beat according to the segmentation results, inputting them into a medical prediction large model to output the characteristic changes of the electrocardiogram signals; comparing the characteristic changes with the pre-stored disease characteristic standards to output the disease prediction results. The present invention combines the complex characteristics of the electrocardiogram signals with the prediction ability of the large model, and can significantly improve the diagnostic accuracy and early detection rate of heart diseases.

Description

Technical Field

[0001] The present invention relates to the field of electrocardiogram signal processing, and in particular to a heart disease prediction system based on electrocardiogram signals and a large model. Background Art

[0002] Heart disease is a type of disease that affects the structure and function of the heart, usually affecting blood circulation, the heart's ability to pump blood, or the heart's own electrical activity. There are many types of heart disease, including coronary heart disease, heart failure, arrhythmia, valvular disease, etc. At the same time, in order to avoid the sudden occurrence of heart disease, most of the existing technical solutions are based on doctors' on-site diagnosis and analysis reports based on electrocardiogram data, which have serious problems with timeliness and human interference factors.

[0003] Heart disease prediction is a multidisciplinary task that usually combines traditional clinical evaluation, laboratory tests, imaging examinations, and emerging artificial intelligence technologies and genomics research. By comprehensively considering personal lifestyle, genetic background, clinical history and other information, doctors can develop more personalized prediction and prevention plans for patients. Early prediction and intervention are crucial to reducing the risk of heart attacks, especially for people at high risk. Timely detection of problems can greatly improve treatment outcomes and quality of life.

[0004] In the existing technology, deep learning technology is usually used to process ECG signals to predict heart disease. However, the prediction process lacks baseline offset suppression and noise processing for ECG signals, so that the noisy ECG signals will reduce the clarity of the ECG and make the signal waveform unstable, which is not conducive to detecting the real changes in the heart. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a heart disease prediction system based on electrocardiogram signals and a large model, so as to achieve the purpose of improving the diagnostic accuracy and early detection rate of heart diseases.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A heart disease prediction system based on electrocardiogram signals and a large model, the heart disease prediction system comprising:

[0008] The ECG signal acquisition and processing unit is used to collect the ECG signals of the user using the cardiac clothing, and to perform baseline offset suppression and noise processing operations on the ECG signals using the data processing center;

[0009] A signal sequence information entropy acquisition unit is used to extract a variability sequence from the processed ECG signal, draw an ECG scatter plot according to the variability, and perform partition number encoding on the scattered points in the scatter plot to obtain the sequence information entropy of the heart rate variability sequence;

[0010] A signal characteristic change acquisition unit is used to segment the ECG signal based on sequence information entropy and wave peaks, obtain a waveform segment of a single heart beat according to the segmentation result, and input the waveform segment into a large medical prediction model to output characteristic changes of the ECG signal;

[0011] The disease prediction result output unit is used to compare the feature changes with the pre-stored disease feature standards, and determine the state source of the waveform segment according to the comparison results, and output the disease prediction result based on the source result.

[0012] Preferably, the ECG signal acquisition processing unit includes:

[0013] The ECG signal acquisition and transmission module is used to collect the ECG signals of the user using the ECG sensor built into the cardiac clothing, and transmit the ECG signals to the data processing center using wireless transmission technology;

[0014] The signal singular point elimination processing module is used to screen the singular points in the ECG signal by using the median filtering technology, and eliminate the ECG signal whose singular point is less than the set information threshold;

[0015] A signal trend term elimination processing module is used to construct a multi-order polynomial based on discrete points, determine the coefficients in the polynomial using the least squares principle to obtain a polynomial fitting model, and fit the ECG signal according to the polynomial fitting model to eliminate the signal trend term;

[0016] The filtering baseline shift suppression processing module is used to set a morphological filter based on mathematical morphological filtering technology to perform noise reduction and expansion operations on the ECG signal, and obtain the ECG signal after baseline shift suppression and noise removal according to the operation results.

[0017] Preferably, the signal sequence information entropy acquisition unit includes:

[0018] The variability sequence extraction module is used to perform minimum entropy deconvolution processing on the processed ECG signal and construct a variability sequence feature space based on the processing results and sequence extraction technology;

[0019] The ECG scatter plot drawing module is used to generate scatter points based on the variability sequence feature space, obtain the horizontal and vertical coordinates of the scatter points on the plane, and draw the scatter points on the plane according to the horizontal and vertical coordinates to obtain the ECG scatter plot;

[0020] The scatter number coding processing module is used to use the data point distribution matrix to screen the partition number feature index to construct a high-dimensional feature space, and combine the maximum variance expansion technology to obtain the sequence real space;

[0021] The sequence information entropy calculation acquisition module is used to input the sequence real space into the hypersphere classification vector machine for hypersphere construction and classification recognition, and obtain the probability of the partition number appearing based on the recognition result to determine the sequence information entropy.

[0022] Preferably, performing minimum entropy deconvolution processing on the processed ECG signal, and constructing a variability sequence feature space based on the processing result and sequence extraction technology includes:

[0023] Initialize the filter coefficients, set the filter's objective function according to the filter coefficients and the minimum entropy principle, and analyze the cross-correlation function between the processed ECG signal and the objective function;

[0024] Analyze the autocorrelation function of the processed ECG signal, determine the matrix representation of the ECG signal based on the autocorrelation function and the cross-correlation function, and obtain the ECG signal after deconvolution processing;

[0025] Based on the ECG signal after deconvolution processing, a multi-information entropy feature space is constructed, the number of data points in the ECG signal is analyzed, and the probability of occurrence of each data point in the multi-information entropy feature space is determined;

[0026] According to the probability results, a data point distribution matrix is ​​constructed to measure the representation ability of multiple information entropy features in the same data point category, and based on the size of the representation ability, effective data points are selected as vectors to construct the variability sequence feature space.

[0027] Preferably, using the data point distribution matrix to screen the partition number feature index to construct a high-dimensional feature space, and combining the maximum variance expansion technology to obtain the sequence real space includes:

[0028] The form of the global metric matrix is ​​set based on the data point distribution matrix, and the global metric matrix is ​​expanded to obtain a low-dimensional manifold embedding. The inner product matrix of the low-dimensional output is introduced as a high-dimensional feature space according to the embedding result.

[0029] Analyze the distance matrix between the internal data points and the preset distance points in the high-dimensional feature space, assign weights to the distance matrix as partition numbers, and perform proportional fixing operations on the partition numbers according to the dimensionality reduction method;

[0030] The position of the data point in the partition is determined according to the partition number, and the corresponding area number is assigned to the data point to obtain a number sequence. The maximum variance expansion technology is used to perform feature decomposition on the number sequence, and the real space of the sequence is obtained based on the decomposition result.

[0031] Preferably, the signal characteristic change acquisition unit includes:

[0032] ECG signal peak search module, used to implement adaptive bandpass filtering on ECG signals according to the maximum spectral kurtosis principle, and combine the filtering results, sequence information entropy and optimal harmonic energy and algorithm to search for ECG signal peak data;

[0033] An ECG signal segmentation processing module is used to segment a single cycle of the ECG signal according to the peak data, extract a single heart beat signal within the cycle, and generate a waveform segment based on the single heart beat signal;

[0034] The medical prediction big model building module is used to extract ECG features based on historical ECG data, generate encoders and decoders based on the deep learning big model, and use ECG features as input to train the encoders and decoders to obtain the medical prediction big model;

[0035] The ECG signal characteristic change acquisition module is used to input the waveform segment into the medical prediction model and obtain the characteristic changes of the ECG signal based on the model output results.

[0036] Preferably, the electrocardiogram signal is subjected to adaptive bandpass filtering according to the maximum spectral kurtosis principle, and the peak data of the electrocardiogram signal is searched by combining the filtering result, sequence information entropy and optimal harmonic energy and algorithm, including:

[0037] A tree search structure is constructed according to the maximum spectral kurtosis principle, the frequency band corresponding to each data point in the ECG signal is determined based on the tree search results, and the filter parameters are determined by the frequency band to perform bandpass filtering on the ECG signal;

[0038] Based on the filtering processing result, a filtering signal corresponding to each data point is obtained, envelope demodulation is performed on the filtering signal, an envelope signal corresponding to each data point is obtained according to the demodulation result, and the kurtosis of the envelope signal is analyzed;

[0039] The frequency band corresponding to the data point with the maximum kurtosis is selected as the optimal filter to perform bandpass filtering on the ECG signal, and the envelope signal of the ECG signal after filtering is obtained based on the filtering result and Hilbert transform;

[0040] Performing Fourier transform operation on the envelope signal based on the optimal harmonic energy and algorithm, obtaining the spectrum distribution of the envelope signal according to the transformation result, and performing Fourier transform operation on the spectrum distribution to generate an inverse envelope spectrum;

[0041] The harmonic energy and maximum value of the ECG signal are searched according to the inverse envelope spectrum and sequence information entropy, and the peak data of the ECG signal is output based on the harmonic energy and maximum value.

[0042] Preferably, the envelope signal expression of the filtered ECG signal is:

[0043] ;

[0044] In the formula, represents the envelope signal of the ECG signal after filtering, represents the ECG signal after filtering, t represents the time period, dτ represents the infinitesimal element for integrating the ECG signal after filtering, and τ represents the integral variable.

[0045] Preferably, segmenting a single cycle of the electrocardiogram signal according to the peak data, extracting a single heart beat signal within the cycle, and generating a waveform segment based on the single heart beat signal includes:

[0046] A structural element is constructed according to the number of peak data and ECG signal data points, a start segmentation point and an end segmentation point are defined by the structural element, and a single cycle of the ECG signal is segmented based on the start segmentation point and the end segmentation point;

[0047] Analyze the peak position of the largest segmentation impulse in the segmentation result, determine the start and end time of the main impulse within the segmentation range based on the position result, and extract a single heart beat signal within a single cycle according to the main impulse information;

[0048] The waveform segment of a single heart beat is extracted based on a single heart beat signal, and the amplitude of the waveform segment is normalized to a preset interval and converted into a standard normal distribution form to obtain the final waveform segment.

[0049] Preferably, extracting ECG features based on historical ECG data, generating an encoder and a decoder based on a deep learning big model, and using the ECG features as input to train the encoder and the decoder to obtain a medical prediction big model includes:

[0050] Obtain historical ECG data based on offline and online collection methods, and extract ECG signal features after data cleaning and preprocessing of the historical ECG data;

[0051] Generate a pre-trained large model including an encoder and a decoder based on the deep learning large model, and set the structural information of the encoder and the decoder according to the attention mechanism and the feedforward network;

[0052] The encoder and decoder are trained after the structural information is set using ECG features as input, and the errors between the encoder and decoder are determined using the training results. The result information of the encoder and decoder is optimized based on the error results, and a large medical prediction model is obtained based on the optimization results.

[0053] The present invention has the following technical effects or advantages:

[0054] 1. The present invention first implements baseline offset suppression and noise processing on the ECG signal, which can enhance the clarity of the ECG and make the waveform of the signal more stable, which is conducive to detecting the real changes of the heart. At the same time, the heart rate variability is analyzed, and by extracting the variability sequence and drawing a scatter plot, the changing law of the heart rhythm can be intuitively analyzed. Finally, the sequence information entropy is calculated to measure the complexity and uncertainty of the ECG signal. The sequence information entropy is combined with the peak to segment the ECG signal, which can ensure that each heartbeat cycle is accurately extracted, which is convenient for subsequent periodic analysis and feature extraction. Through multi-level and multi-step analysis, combined with the complex characteristics of the ECG signal and the predictive ability of the large model, the diagnostic accuracy and early detection rate of heart disease can be significantly improved.

[0055] 2. The present invention can effectively enhance the real changes in the signal by performing minimum entropy deconvolution processing on the processed ECG signal, so as to extract more accurate variability features from the ECG signal. Through minimum entropy deconvolution, high-order features in the signal can be better extracted, laying the foundation for the later identification of potential heart problems. At the same time, according to the processing results and sequence extraction technology, the variability sequence feature space is constructed to clarify the changes in heart rate variability in time or space dimensions, thereby providing an accurate basis for the early detection of the disease, and converting the variability sequence into a scatter plot, so as to achieve the purpose of intuitively displaying the characteristics of the ECG signal and revealing the relationship between different signal features.

[0056] 3. The present invention segments the ECG signal based on sequence information entropy and peaks, and inputs the waveform segments into the medical prediction large model for feature change output. In the process, adaptive bandpass filtering is performed through the spectral kurtosis maximization principle, which can accurately extract effective signals, making the segmentation accuracy of the filtered signal higher, making the prediction of heart disease based on the large model more accurate, providing a reliable foundation for subsequent signal segmentation, feature extraction and disease prediction, and ultimately improving the diagnostic accuracy and efficiency of heart disease. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0058] Figure 1 It is a principle block diagram of a heart disease prediction system based on electrocardiogram signals and a large model according to an embodiment of the present invention.

[0059] In the figure:

[0060] 1. ECG signal acquisition and processing unit; 2. Signal sequence information entropy acquisition unit; 3. Signal feature change acquisition unit; 4. Disease prediction result output unit. DETAILED DESCRIPTION

[0061] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0062] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0063] It should be noted that the terms used herein are only for describing specific embodiments, and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that the terms "include" and "have" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0064] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0065] See also Figure 1 The present invention provides a heart disease prediction system based on electrocardiogram signals and a large model, the heart disease prediction system comprising:

[0066] The ECG signal acquisition and processing unit 1 is used to collect the ECG signals of the user using the cardiac clothing, and to perform baseline shift suppression and noise processing operations on the ECG signals using the data processing center.

[0067] In this embodiment, the ECG signal acquisition processing unit 1 includes:

[0068] The ECG signal acquisition and transmission module is used to collect the user's ECG signals using the ECG sensor built into the cardiac clothing, and transmit the ECG signals to the data processing center using wireless transmission technology.

[0069] Specifically, in the process of acquiring ECG signals, an ergonomic heart garment can be pre-designed, and the heart garment has multiple built-in ECG sensors to ensure that the patient's ECG signals can be collected comprehensively and accurately. The heart garment uses wireless transmission technology to transmit the collected ECG signals to the data processing center in real time.

[0070] The signal singular point elimination processing module is used to screen the singular points in the ECG signal using the median filtering technology, and eliminate the ECG signals whose singular points are smaller than the set information threshold.

[0071] Specifically, when using the median filtering technology to screen the singular points in the ECG signal and remove the ECG signals with singular points less than the set information threshold, a suitable median filtering window size can be pre-selected (usually selected according to the signal sampling rate and the target noise characteristics), the data points in each sliding window are sorted, and the middle value is taken as the output value of the current window, and the singular points in the signal that deviate greatly from the median value in the window are replaced with the median value to smooth the signal, and a deviation threshold (such as the rate of change of the signal amplitude or the difference from the median value) is set to identify abnormal points. According to the important features of the ECG signal (such as the R wave peak value, RR interval, etc.), an information threshold is set to judge the quality of the signal and evaluate the proportion of valid ECG waveforms or the number of singular points in the signal. If the number of singular points in the signal exceeds the threshold, or the proportion of valid waveforms is lower than the threshold, it is marked as an invalid signal, and the invalid signal segment is removed, or the small missing signal segment is completed by interpolation and other methods, and the ECG signal with higher quality after the singular points are removed is saved for subsequent analysis.

[0072] The signal trend item elimination processing module is used to construct a multi-order polynomial based on discrete points, determine the coefficients in the polynomial using the least squares principle to obtain a polynomial fitting model, and fit the ECG signal according to the polynomial fitting model to eliminate the signal trend item.

[0073] Specifically, a multi-order polynomial is constructed based on discrete points, the coefficients in the polynomial are determined using the least squares principle to obtain a polynomial fitting model, and the ECG signal is fitted according to the polynomial fitting model. When eliminating the signal trend term, a suitable polynomial order n can be selected. The selection of the polynomial order usually depends on the complexity of the signal. If the order is too low, the trend term may not be accurately fitted, and if the order is too high, overfitting may occur. The polynomial model is constructed, assuming that the signal trend term T(x) can be expressed as an n-order polynomial: T(x)=a0+a1x+a2x 2 +…+a n x n , where a0, a1, a2, …, a n is the coefficient to be determined. The polynomial coefficient is solved by minimizing the sum of squares of the fitting error. Specifically, the least squares method can be used to calculate the fitting values ​​corresponding to all time points using the obtained polynomial model T(x) and obtain the trend term. This process is to use the polynomial function at each discrete point x i Prediction is performed on T(x i ) is the fitting result of the ECG signal based on the polynomial model, which represents the trend component of the signal. i ) to get the detrending signal y detrended (x i )=y i -T(x i), this process effectively removes the trend term (low-frequency changes) in the ECG signal and retains the high-frequency part or important dynamic features in the ECG signal.

[0074] The filtering baseline shift suppression processing module is used to set a morphological filter based on mathematical morphological filtering technology to perform noise reduction and expansion operations on the ECG signal, and obtain the ECG signal after baseline shift suppression and noise removal according to the operation results.

[0075] Specifically, when a morphological filter is set based on mathematical morphological filtering technology to perform noise reduction and dilation operations on the ECG signal, and the ECG signal after baseline offset suppression and noise removal is obtained according to the operation results, a low-pass filter or a band-pass filter (such as 0.5Hz-40Hz) can be used to remove power frequency noise and high-frequency interference to provide a smoother signal for subsequent morphological processing. Based on set theory, it mainly relies on two basic operations, dilation and erosion, and their combination operations: opening and closing. A suitable structural element B, that is, a predefined sliding window, is selected for morphological operations. At the same time, the opening operation is used to extract the local minimum curve in the ECG signal, which is used as the fitting baseline to calculate the baseline drift component of the ECG signal, and the baseline component is subtracted from the original ECG signal to eliminate the low-frequency fluctuation caused by the baseline drift, and the high-frequency noise is further removed by the morphological filter.

[0076] The signal sequence information entropy acquisition unit 2 is used to extract the variability sequence from the processed ECG signal, draw an ECG scatter plot according to the variability, and perform partition number encoding on the scatter points in the scatter plot to obtain the sequence information entropy of the heart rate variability sequence.

[0077] The signal sequence information entropy acquisition unit 2 includes:

[0078] The variability sequence extraction module is used to perform minimum entropy deconvolution processing on the processed ECG signal and construct the variability sequence feature space based on the processing results and sequence extraction technology.

[0079] Specifically, when performing minimum entropy deconvolution processing on the processed ECG signal and constructing the variability sequence feature space according to the processing results and the sequence extraction technology, the filter coefficients can be initialized, the objective function of the filter can be set according to the filter coefficients and the minimum entropy principle, and the cross-correlation function between the processed ECG signal and the objective function can be analyzed; the autocorrelation function of the processed ECG signal is analyzed, and the matrix representation of the ECG signal is determined based on the autocorrelation function and the cross-correlation function to obtain the ECG signal after deconvolution processing; a multi-information entropy feature space is constructed based on the ECG signal after deconvolution processing, the number of data points in the ECG signal is analyzed, and the probability of occurrence of each data point in the multi-information entropy feature space is determined; a data point distribution matrix is ​​constructed according to the probability results, the representation ability of the multi-information entropy features in the same data point category is measured, and effective data points are screened as vectors based on the size of the representation ability to construct the variability sequence feature space.

[0080] The ECG scatter plot drawing module is used to generate scatter points based on the variability sequence feature space, obtain the horizontal and vertical coordinates of the scatter points on the plane, and draw the scatter points on the plane according to the horizontal and vertical coordinates to obtain the ECG scatter plot.

[0081] Specifically, when generating scatter points based on the variability sequence feature space, obtaining the horizontal and vertical coordinates of the scatter points on the plane, and drawing the scatter points on the plane according to the horizontal and vertical coordinates to obtain an ECG scatter plot, feature points can be extracted based on the variability sequence feature space, such as the R wave peak, the start and end points of the P wave and the T wave, etc., and the variability sequence is calculated based on the feature points. The variability sequence can be the time interval (RR interval) between adjacent R wave peaks, the voltage difference between adjacent R wave peaks, etc., two variability sequences are selected as the horizontal and vertical coordinates, and scatter data are generated according to the selected horizontal and vertical coordinates. Each scatter point represents the variability characteristics of a time point. A drawing tool (such as Matplotlib, Plotly, etc.) is used to draw a scatter plot, and the generated scatter data is drawn on the plane, and the horizontal and vertical coordinates correspond to the values ​​of the variability sequence, respectively.

[0082] It needs to be explained that the electrocardiogram is a graphic record obtained by measuring the electrical activity of the heart at different time points, which can reflect the electrophysiological state of the heart and the functional abnormalities of various parts of the heart. The following is a brief introduction to the P wave, R wave, Q wave, S wave, T wave, RR interval, and QRS interval:

[0083] P wave: represents the potential change of the depolarization of the two atria (left atrium and right atrium). The electrical activity of the heart starts from the sinoatrial node, indicating the process of atrial depolarization, that is, after the sinoatrial node sends out an electrical signal, the electrical activity is transmitted to the atrium, which produces the P wave. The P wave is an important wave group on the electrocardiogram. Its morphology, amplitude, and duration have important clinical significance. The normal P wave duration generally does not exceed 0.12 seconds, and the amplitude in the limb leads generally does not exceed 0.25mV, and in the chest leads generally does not exceed 0.2mV. Abnormalities in the P wave can indicate a variety of heart diseases.

[0084] R wave: represents the potential change of right ventricular depolarization and is part of the QRS complex. The appearance of the R wave marks the beginning of ventricular depolarization. On the electrocardiogram, the QRS complex reflects the entire process of ventricular depolarization, and the R wave is the first upward peak in this process. The duration and amplitude of the R wave are important parameters for assessing heart health. Under normal circumstances, the amplitude of the R wave is within a certain range and the duration is greater than 0.1 seconds. Changes in the R wave can indicate heart disease.

[0085] Q wave: The first downward wave in the QRS complex is called the Q wave. Based on the size of the Q wave and the location of the lead where the Q wave appears, it can be used to preliminarily determine whether a person has suffered a myocardial infarction. It can also be used to preliminarily determine whether the myocardial infarction is acute or old. The specific clinical significance of the appearance of the Q wave in the ECG signal requires a comprehensive analysis based on symptoms and other information.

[0086] S wave: The S wave represents part of the ventricular depolarization process, especially related to the depolarization of the left ventricle. On the electrocardiogram, the S wave usually follows the R wave and is deflected downward. The normal value range of the S wave is generally between 0 and 1. Exceeding this range may indicate the possibility of left ventricular hypertrophy.

[0087] T wave: represents the change in potential during the repolarization process of the two ventricles. The specific shape of the T wave is related to the changes in the myocardial repolarization process. Under normal circumstances, the T wave should be an upward peak, representing good heart function and a normal repolarization process. Changes in the shape, amplitude and duration of the T wave can reflect certain diseases or abnormalities of the heart.

[0088] RR interval: refers to the distance between two R waves on the electrocardiogram, that is, the time interval between two ventricular beats. The normal value of the RR interval is 0.6 to 1.0 seconds, which can reflect the condition of the ventricular rhythm. The prolongation or shortening of the RR interval represents bradycardia or tachycardia, respectively. In addition, irregular RR interval may indicate arrhythmia, such as atrial fibrillation.

[0089] QRS interval: refers to the time course from the starting point of the QRS complex (i.e. the starting point of the Q wave, if there is no Q wave, the starting point of the R wave shall prevail) to the end point of the QRS complex (i.e. the end point of the S wave or if the R wave is the last peak, the end point of the R wave shall prevail). The QRS interval represents the entire process of ventricular depolarization, and its duration is usually between 0.06 and 0.10 seconds. A prolongation of the QRS interval may indicate abnormal conditions such as ventricular hypertrophy, bundle branch block or myocardial infarction.

[0090] In summary, the P wave, R wave, Q wave, S wave, T wave, RR interval and QRS interval of the ECG signal are all important indicators for assessing heart health, each of which has unique clinical significance and diagnostic value.

[0091] The scatter number encoding processing module is used to use the data point distribution matrix to screen the partition number feature indicators to construct a high-dimensional feature space, and combine the maximum variance expansion technology to obtain the real space of the sequence.

[0092] Specifically, when using the data point distribution matrix to screen the partition number feature indicators to construct a high-dimensional feature space, and combining the maximum variance expansion technology to obtain the sequence real space, the form of the global metric matrix can be set based on the data point distribution matrix, and the global metric matrix can be expanded to obtain a low-dimensional manifold embedding, and the inner product matrix of the low-dimensional output is introduced as the high-dimensional feature space according to the embedding result; the distance matrix between the data points inside the high-dimensional feature space and the preset distance points is analyzed, and the distance matrix is ​​weighted as the partition number, and the partition number is fixed in proportion according to the dimensionality reduction method; the position of the data point in the partition is determined according to the partition number, and the corresponding area number is assigned to the data point to obtain a numbering sequence, and the maximum variance expansion technology is used to perform feature decomposition on the numbering sequence, and the sequence real space is obtained based on the decomposition result.

[0093] The sequence information entropy calculation acquisition module is used to input the sequence real space into the hypersphere classification vector machine for hypersphere construction and classification recognition, and obtain the probability of the partition number appearing based on the recognition result to determine the sequence information entropy.

[0094] Specifically, when the real space of the sequence is input into the hypersphere classification vector machine for hypersphere construction and classification recognition, the probability of the occurrence of the partition number is obtained based on the recognition result to determine the sequence information entropy, and the training data set can be used to train the HSVM model (hypersphere classification vector machine). HSVM classifies data of different categories by constructing a hypersphere. The center and radius of the hypersphere are determined by optimizing the objective function. For each test data point, the probability of belonging to each category is calculated by calculating the distance between the data point and the center of each hypersphere. The closer the distance, the higher the probability of belonging to the category. The calculated probability is normalized so that the sum of the probabilities of all categories is 1. The probability of the occurrence of each partition number is obtained from the classification recognition result to calculate the information entropy.

[0095] In order to facilitate understanding of the above technical solution of this embodiment, the method for obtaining sequence information entropy in the actual process of this embodiment is described in detail below.

[0096] Step 1: Extraction of variant sequences;

[0097] Assume that there is ECG signal data with a length of 1000 sampling points. The ECG signal is filtered, denoised and baseline drift corrected. The filter coefficient is [0.5, 0.3, 0.2]. According to the minimum entropy principle, the objective function J is set as:

[0098] ;

[0099] Among them, m a represents the processed ECG signal, h b represents the filter coefficient, y a represents the ECG signal after deconvolution processing, N represents the number of sampling points of the ECG signal, and M represents the order of the filter, that is, the number of filter coefficients;

[0100] The cross-correlation function between the processed ECG signal and the target function is calculated, and the autocorrelation function of the processed ECG signal is calculated based on the cross-correlation function. Based on the autocorrelation function and the cross-correlation function, the matrix representations Q and P of the ECG signal are determined. The ECG signal y after deconvolution is obtained by solving the matrix equation Q·h=P. The R wave peak, the start and end points of the P wave and the T wave are extracted from the ECG signal after deconvolution, and the time interval (RR interval) between adjacent R wave peaks and the voltage difference between adjacent R wave peaks are calculated. The RR interval is selected as the horizontal coordinate, and the voltage difference between adjacent R wave peaks is selected as the vertical coordinate. Scatter data are generated according to the selected horizontal and vertical coordinates.

[0101] Step 2: Draw the ECG scatter plot;

[0102] Extract feature points: Assume that the extracted R wave peak position is:

[0103] [100, 200, 300, 400, 500, 600, 700, 800, 900];

[0104] Calculate the RR interval of the variability sequence: [100, 100, 100, 100, 100, 100, 100, 100];

[0105] Voltage difference: Assume the voltage difference is [0.1, 0.2, 0.15, 0.18, 0.17, 0.16, 0.14, 0.13];

[0106] Generate scattered data: Each scattered point represents the variability characteristics of a time point, for example:

[0107] Scatter point 1: (100, 0.1); Scatter point 2: (100, 0.2); Scatter point 3: (100, 0.15); Scatter point 4: (100, 0.18); Scatter point 5: (100, 0.17); Scatter point 6: (100, 0.16); Scatter point 7: (100, 0.14); Scatter point 8: (100, 0.13);

[0108] Use Matplotlib to draw a scatter plot and set the chart's title, axis labels, scales, and other properties.

[0109] Step 3: Scatter number coding processing;

[0110] Assume that the data point distribution matrix is ​​D, where each element represents the position of a data point in the feature space. Based on the data point distribution matrix, select effective partition number feature indicators, and combine the maximum variance expansion technology to expand the data point distribution matrix into a low-dimensional manifold embedding to obtain a high-dimensional feature space.

[0111] Assume that the global metric matrix is ​​G, and perform an expansion operation on the global metric matrix to obtain a low-dimensional manifold embedding. The inner product matrix of the low-dimensional output is introduced as the high-dimensional feature space. The distance matrix between the internal data points of the high-dimensional feature space and the preset distance points is analyzed, and the distance matrix is ​​weighted to obtain the partition number. The partition number is fixed in proportion according to the dimensionality reduction method, the position of the data point in the partition is determined according to the partition number, and the corresponding area number is assigned to the data point to obtain a number sequence.

[0112] Step 4: Calculate sequence information entropy;

[0113] Assume that the training data set contains 100 samples, each sample has 10 features, and the training data set is used to train the HSVM model. A hypersphere is constructed for classification and recognition. For each test data point, the probability of belonging to each category is calculated, and the calculated probability is normalized so that the sum of the probabilities of all categories is 1. The probability of each partition number appearing is obtained from the classification and recognition results, and the information entropy is calculated using the information entropy formula;

[0114] Specific numerical examples:

[0115] RR interval: [100, 100, 100, 100, 100, 100, 100, 100];

[0116] Voltage difference: [0.1, 0.2, 0.15, 0.18, 0.17, 0.16, 0.14, 0.13];

[0117] Scattered data:

[0118] Scatter point 1: (100, 0.1); Scatter point 2: (100, 0.2); Scatter point 3: (100, 0.15); Scatter point 4: (100, 0.18); Scatter point 5: (100, 0.17); Scatter point 6: (100, 0.16); Scatter point 7: (100, 0.14); Scatter point 8: (100, 0.13);

[0119] The partition numbers are [1, 2, 1, 2, 1, 2, 1, 2]; the probability of each partition number appearing is [0.5, 0.5]; the information entropy O = -(0.5log0.5 + 0.5log0.5) = 1;

[0120] Through the above steps, the variability sequence can be completely extracted from the processed ECG signal, the ECG scatter plot can be drawn, and the sequence information entropy of the heart rate variability sequence can be calculated. Therefore, the present invention generates a variability sequence based on the ECG signal and converts it into a scatter plot, so as to achieve the purpose of intuitively displaying the characteristics of the ECG signal and revealing the relationship between different signal characteristics.

[0121] The signal characteristic change acquisition unit 3 is used to segment the ECG signal based on sequence information entropy and wave peaks, obtain the waveform segment of a single heart beat according to the segmentation result, and input the waveform segment into the medical prediction model to output the characteristic change of the ECG signal.

[0122] In this embodiment, the signal characteristic change acquisition unit 3 includes:

[0123] The ECG signal peak search module is used to perform adaptive bandpass filtering on the ECG signal according to the principle of maximum spectral kurtosis, and combine the filtering results, sequence information entropy, optimal harmonic energy and algorithm to search for the ECG signal peak data.

[0124] Specifically, when the ECG signal is subjected to adaptive bandpass filtering according to the maximum spectral kurtosis principle, and the filtering result, sequence information entropy, and optimal harmonic energy and algorithm are combined to search for the peak data of the ECG signal, a tree search structure can be constructed according to the maximum spectral kurtosis principle, and the frequency band corresponding to each data point in the ECG signal is determined based on the tree search results, and the ECG signal is subjected to bandpass filtering by using the frequency band to determine the filtering parameters; based on the filtering result, a filtered signal corresponding to each data point is obtained, envelope demodulation is performed on the filtered signal, and the envelope signal corresponding to each data point is obtained according to the demodulation result. signal, and analyze the kurtosis of the envelope signal; select the frequency band corresponding to the data point with the maximum kurtosis as the optimal filter to perform bandpass filtering on the ECG signal, and obtain the envelope signal of the filtered ECG signal based on the filtering result and Hilbert transform; perform Fourier transform operation on the envelope signal based on the optimal harmonic energy and algorithm, obtain the spectrum distribution of the envelope signal according to the transformation result, and perform Fourier transform operation on the spectrum distribution to generate an inverse envelope spectrum; search for the harmonic energy and maximum value of the ECG signal according to the inverse envelope spectrum and sequence information entropy, and output the peak data of the ECG signal based on the harmonic energy and maximum value.

[0125] Among them, the envelope signal expression of the ECG signal after filtering is:

[0126] ;

[0127] In the formula, represents the envelope signal of the ECG signal after filtering, represents the ECG signal after filtering, t represents the time period, dτ represents the infinitesimal element for integrating the ECG signal after filtering, and τ represents the integral variable.

[0128] The ECG signal segmentation processing module is used to segment a single cycle of the ECG signal according to the peak data, extract a single heart beat signal within the cycle, and generate a waveform segment based on the single heart beat signal.

[0129] Specifically, when a single cycle of the ECG signal is segmented according to the peak data, and a single heart beat signal within the cycle is extracted, and a waveform segment is generated based on the single heart beat signal, a structural element can be constructed according to the peak data and the number of ECG signal data points, and the starting segmentation point and the ending segmentation point are defined by using the structural element, and the single cycle of the ECG signal is segmented based on the starting segmentation point and the ending segmentation point; the peak position with the largest segmentation impact in the segmentation result is analyzed, and the starting time and the ending time of the main impact within the segmentation range are determined based on the position result, and the single heart beat signal within the single cycle is extracted according to the main impact information; the waveform segment of the single heart beat is extracted based on the single heart beat signal, and the amplitude of the waveform segment is normalized to a preset interval and converted into a standard normal distribution form to obtain the final waveform segment.

[0130] The medical prediction big model building module is used to extract ECG features based on historical ECG data, generate encoders and decoders based on the deep learning big model, and use ECG features as input to train the encoder and decoder to obtain the medical prediction big model.

[0131] Specifically, when extracting ECG features based on historical ECG data, generating encoders and decoders based on the deep learning big model, and using the ECG features as input to train the encoder and decoder to obtain the medical prediction big model, the historical ECG data can be obtained based on offline and online collection methods respectively, and the ECG signal features can be extracted after data cleaning and preprocessing of the historical ECG data; a pre-trained big model including an encoder and a decoder is generated based on the deep learning big model, and the structural information of the encoder and the decoder is set according to the attention mechanism and the feedforward network; the ECG features are used as input to train the encoder and the decoder after the structural information is set, and the training results are used to judge the errors of the encoder and the decoder, and the result information of the encoder and the decoder is optimized according to the error results, and the medical prediction big model is obtained based on the optimization results.

[0132] It should be explained that the medical prediction model is mainly based on the pre-trained model of the Transformer model, which is mainly composed of two parts: the encoder and the decoder. The encoder is composed of multiple identical layers stacked together, and each layer contains two sub-layers:

[0133] A multi-head self-attention mechanism (Multi-HeadAttention) and a feedforward network (FeedForwardNetwork). In addition, each sublayer uses residual connection (ResidualConnection) and layer normalization (LayerNormalization). The input sequence is various waveform characteristics of the disease. Assume that the input sequence is Z={z1,z2,…,z n}, where z n The embedding vector representing the nth element in the sequence.

[0134] Query, Key, Value: In the self-attention mechanism, each input element z n is linearly transformed into three vectors e n , k n 、v n ;

[0135] ;

[0136] Where W E , W K , W VAll are learnable weight matrices;

[0137] AttentionScores: Calculate e n With all k j The dot product of , and then applying the scaling factor , (d k It is k j Dimension), and finally the attention weight is obtained through the softmax function;

[0138] Multi-head attention: transform the input Z into query, key, and value vectors of multiple "heads", perform self-attention calculations on each of them, and then concatenate and linearly transform the results.

[0139] Each encoder layer also includes a feed-forward network, which further processes the output of the self-attention layer. The feed-forward network usually contains two linear transformations and a ReLU activation function to output the final prediction result.

[0140] The ECG signal characteristic change acquisition module is used to input the waveform segment into the medical prediction model and obtain the characteristic changes of the ECG signal based on the model output results.

[0141] The disease prediction result output unit 4 is used to compare the characteristic change with the pre-stored disease characteristic standard, and determine the state source of the waveform segment according to the comparison result, and output the disease prediction result based on the source result.

[0142] In this embodiment, when comparing the feature changes with the pre-stored disease feature standards, and judging the state source of the waveform segment based on the comparison results, and outputting the disease prediction results based on the source results, a suitable similarity calculation method (such as Euclidean distance, cosine similarity, etc.) can be used to calculate the similarity between the extracted feature vector and the pre-stored feature standard, and at the same time set a threshold to judge the state source of the waveform segment based on the similarity or distance. If the similarity is higher than a certain threshold, it is considered that the waveform segment belongs to a certain disease type. Based on the state judgment result, the disease prediction result is output. The prediction result can be a specific disease name or a possible probability of the disease. An explanation of the prediction result is provided, including possible disease types, detailed information on feature matching, etc.

[0143] In order to facilitate understanding of the above technical solution of this embodiment, the heart disease prediction method in the actual process of this embodiment is described in detail below.

[0144] Step 1: Assume that there is a segment of ECG signal data with a length of 1000 sampling points. According to the maximum spectral kurtosis principle, a tree search structure is constructed. Based on the tree search results, the frequency band corresponding to each data point in the ECG signal is determined. Assuming that the frequency band range is 1-100 Hz, with a step size of 1 Hz, the filter parameters are determined according to the frequency band range. The ECG signal is band-pass filtered to obtain the filtered signal corresponding to each data point.

[0145] Assume that the filtered signal is [s1,s2,…,s 1000 ], perform envelope demodulation on the filtered signal to obtain the envelope signal corresponding to each data point. Assume that the envelope signal is [e1, e2, …, e 1000 ], analyze the kurtosis of the envelope signal, select the frequency band corresponding to the data point with the largest kurtosis as the optimal filter, assume that the frequency band corresponding to the data point with the largest kurtosis is 30-50Hz, use the frequency band of 30-50Hz to perform bandpass filtering on the ECG signal, assume that the filtered signal is [y1,y2,…,y 1000 ], perform Hilbert transform on the filtered signal to obtain the envelope signal;

[0146] Perform Fourier transform on the envelope signal to obtain the spectrum distribution, and perform Fourier transform on the spectrum distribution to generate an inverse envelope spectrum. Search for the harmonic energy and maximum value of the ECG signal based on the inverse envelope spectrum and the sequence information entropy. Assume that the maximum value is 150. Output the peak data of the ECG signal based on the harmonic energy and the maximum value. The peak position is [100, 200, 300, 400, 500, 600, 700, 800, 900];

[0147] Step 2: Construct a structural element based on the number of peak data and ECG signal data points. Assume that the structural element is [10, 20, 30, 40, 50, 60, 70, 80, 90], and use the structural element to define the start segmentation point and the end segmentation point:

[0148] Start segmentation point: [90, 190, 290, 390, 490, 590, 690, 790, 890];

[0149] End segmentation point: [110, 210, 310, 410, 510, 610, 710, 810, 910];

[0150] The single cycle of the ECG signal is segmented based on the start segmentation point and the end segmentation point, the peak position with the largest segmentation impact in the segmentation result is analyzed, and the start and end times of the main impact within the segmentation range are determined to determine the main impact position; the single heart beat signal within a single cycle is extracted according to the main impact information, a waveform segment is generated based on the single heart beat signal, and the amplitude of the waveform segment is normalized to a preset interval and converted into a standard normal distribution form.

[0151] Step 3: Obtain historical ECG data based on offline and online collection methods respectively. Assume that the historical ECG data contains 1000 samples, each sample has 1000 sampling points, perform data cleaning and preprocessing on the historical ECG data, and extract ECG signal features. Assume that the extracted features include RR interval, PR interval, QT interval, QRS interval, etc.

[0152] Generate a pre-trained large model including an encoder and a decoder based on the deep learning large model. Assume that the encoder contains three convolutional layers, and the number of filters in each convolutional layer is 64, 128, and 256 respectively. The decoder contains three deconvolutional layers, and the number of filters in each deconvolutional layer is 256, 128, and 64 respectively. Set the structural information of the encoder and decoder according to the attention mechanism and feedforward network.

[0153] Assuming that the feedforward network contains two fully connected layers, the number of neurons in each fully connected layer is 512 and 256 respectively, the ECG features are used as input to train the encoder and decoder, and the training results are used to judge the error between the encoder and decoder. The result information of the encoder and decoder is optimized according to the error results, and the accuracy of the large medical prediction model is 95%.

[0154] Step 4: Assume that the pre-stored disease characteristic standards include the following disease types:

[0155] A normal electrocardiogram is [100, 120, 300, 80];

[0156] atrial fibrillation: [120, 130, 320, 90];

[0157] Myocardial infarction: [90, 110, 280, 70];

[0158] Use Euclidean distance to calculate the similarity between the extracted feature vector and the pre-stored feature standard: assuming that the feature vector is [100, 120, 300, 80], assuming that the set threshold is 10, if the similarity (distance) is less than 10, the match is considered successful; normal electrocardiogram: the distance is 0, less than the threshold 10, the match is successful; atrial fibrillation: the distance is 31.62, greater than the threshold 10, the match fails; myocardial infarction: the distance is 26.46, greater than the threshold 10, the match fails, and the disease prediction result is output based on the state judgment result.

[0159] To sum up, with the help of the above technical scheme of the present invention, the present invention first implements baseline offset suppression and noise processing on the ECG signal, which can enhance the clarity of the ECG and make the waveform of the signal more stable, which is conducive to detecting the real changes of the heart. At the same time, the heart rate variability is analyzed, and by extracting the variability sequence and drawing a scatter plot, the changing law of the heart rhythm can be intuitively analyzed. Finally, the sequence information entropy is calculated to measure the complexity and uncertainty of the ECG signal. The sequence information entropy is combined with the peak to segment the ECG signal, which can ensure that each heartbeat cycle is accurately extracted, which is convenient for subsequent periodic analysis and feature extraction. Through multi-level and multi-step analysis, combined with the complex characteristics of the ECG signal and the predictive ability of the large model, the diagnostic accuracy and early detection rate of heart disease can be significantly improved. The present invention can effectively enhance the real changes in the signal by performing minimum entropy deconvolution processing on the processed ECG signal, so as to extract more accurate variability features from the ECG signal. Through minimum entropy deconvolution, high-order features in the signal can be better extracted, laying the foundation for the later identification of potential heart problems. At the same time, according to the processing results and sequence extraction technology, the variability sequence feature space is constructed to clarify the changes in heart rate variability in time or space dimensions, thereby providing an accurate basis for the early detection of the disease, and converting the variability sequence into a scatter plot, so as to achieve the purpose of intuitively displaying the characteristics of the ECG signal and revealing the relationship between different signal features.

[0160] The present invention segments the ECG signal based on sequence information entropy and peaks, and inputs the waveform segments into a large medical prediction model for feature change output. In the process, adaptive bandpass filtering is performed through the spectral kurtosis maximization principle, which can accurately extract effective signals, making the segmentation accuracy of the filtered signal higher, making the prediction of heart disease based on the large model more accurate, providing a reliable foundation for subsequent signal segmentation, feature extraction and disease prediction, and ultimately improving the diagnostic accuracy and efficiency of heart disease.

[0161] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0162] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A heart disease prediction system based on electrocardiogram signals and a large model, characterized in that: include: The ECG signal acquisition and processing unit is used to collect the ECG signals of the user using the cardiac clothing, and to perform baseline offset suppression and noise processing operations on the ECG signals using the data processing center; A signal sequence information entropy acquisition unit is used to extract a variability sequence from the processed ECG signal, draw an ECG scatter plot according to the variability, and perform partition number encoding on the scattered points in the scatter plot to obtain the sequence information entropy of the heart rate variability sequence; A signal characteristic change acquisition unit is used to segment the ECG signal based on sequence information entropy and wave peaks, obtain a waveform segment of a single heart beat according to the segmentation result, and input the waveform segment into a large medical prediction model to output characteristic changes of the ECG signal; A disease prediction result output unit, used to compare the characteristic change with the pre-stored disease characteristic standard, and determine the state source of the waveform segment according to the comparison result, and output the disease prediction result based on the source result; The signal sequence information entropy acquisition unit comprises: The variability sequence extraction module is used to perform minimum entropy deconvolution processing on the processed ECG signal and construct a variability sequence feature space based on the processing results and sequence extraction technology; ECG scatter plot drawing module, used to generate ECG scatter plot based on variability sequence feature space; The scatter point number encoding processing module is used to set the form of the global metric matrix based on the data point distribution matrix, expand the global metric matrix to obtain a low-dimensional manifold embedding, and introduce the inner product matrix of the low-dimensional output as the high-dimensional feature space according to the embedding result; analyze the distance matrix between the data points and the preset distance points in the high-dimensional feature space, and assign weights to the distance matrix as the partition number, and perform a proportional fixing operation on the partition number according to the dimensionality reduction method; determine the position of the data point in the partition according to the partition number, and assign corresponding area numbers to the data points to obtain a numbering sequence, use the maximum variance expansion technology to perform feature decomposition on the numbering sequence, and obtain the real space of the sequence based on the decomposition result.

2. A heart disease prediction system based on electrocardiogram signals and a large model according to claim 1, characterized in that: The electrocardiogram signal acquisition processing unit comprises: The ECG signal acquisition and transmission module is used to collect the ECG signals of the user using the ECG sensor built into the cardiac clothing, and transmit the ECG signals to the data processing center using wireless transmission technology; The signal singular point elimination processing module is used to screen the singular points in the ECG signal by using the median filtering technology, and eliminate the ECG signal whose singular point is less than the set information threshold; A signal trend term elimination processing module is used to construct a multi-order polynomial based on discrete points, determine the coefficients in the polynomial using the least squares principle to obtain a polynomial fitting model, and fit the ECG signal according to the polynomial fitting model to eliminate the signal trend term; The filtering baseline shift suppression processing module is used to set a morphological filter based on mathematical morphological filtering technology to perform noise reduction and expansion operations on the ECG signal, and obtain the ECG signal after baseline shift suppression and noise removal according to the operation results.

3. The heart disease prediction system based on electrocardiogram signals and large models according to claim 1, characterized in that: The signal sequence information entropy acquisition unit also includes: The sequence information entropy calculation acquisition module is used to input the sequence real space into the hypersphere classification vector machine for hypersphere construction and classification recognition, and obtain the probability of the partition number appearing based on the recognition result to determine the sequence information entropy.

4. A heart disease prediction system based on electrocardiogram signals and large models according to claim 3, characterized in that: The method of performing minimum entropy deconvolution processing on the processed ECG signal and constructing a variability sequence feature space based on the processing result and the sequence extraction technology includes: Initialize the filter coefficients, set the filter's objective function according to the filter coefficients and the minimum entropy principle, and analyze the cross-correlation function between the processed ECG signal and the objective function; Analyze the autocorrelation function of the processed ECG signal, determine the matrix representation of the ECG signal based on the autocorrelation function and the cross-correlation function, and obtain the ECG signal after deconvolution processing; Based on the ECG signal after deconvolution processing, a multi-information entropy feature space is constructed, the number of data points in the ECG signal is analyzed, and the probability of occurrence of each data point in the multi-information entropy feature space is determined; According to the probability results, a data point distribution matrix is ​​constructed to measure the representation ability of multiple information entropy features in the same data point category, and based on the size of the representation ability, effective data points are selected as vectors to construct the variability sequence feature space.

5. The heart disease prediction system based on electrocardiogram signals and large models according to claim 1, characterized in that: The signal characteristic change acquisition unit comprises: ECG signal peak search module, used to perform adaptive bandpass filtering on ECG signals according to the maximum spectral kurtosis principle, and combine the filtering results, sequence information entropy and optimal harmonic energy and algorithm to search for ECG signal peak data; An ECG signal segmentation processing module is used to segment a single cycle of the ECG signal according to the peak data, extract a single heart beat signal within the cycle, and generate a waveform segment based on the single heart beat signal; The medical prediction big model building module is used to extract ECG features based on historical ECG data, generate encoders and decoders based on the deep learning big model, and use ECG features as input to train the encoders and decoders to obtain the medical prediction big model; The ECG signal characteristic change acquisition module is used to input the waveform segment into the medical prediction model and obtain the characteristic changes of the ECG signal based on the model output results.

6. A heart disease prediction system based on electrocardiogram signals and a large model according to claim 5, characterized in that: The method of implementing adaptive bandpass filtering on the ECG signal according to the maximum spectral kurtosis principle and combining the filtering result, sequence information entropy and optimal harmonic energy and algorithm to search for the peak data of the ECG signal includes: A tree search structure is constructed according to the maximum spectral kurtosis principle, the frequency band corresponding to each data point in the ECG signal is determined based on the tree search results, and the filter parameters are determined by the frequency band to perform bandpass filtering on the ECG signal; Based on the filtering processing result, a filtering signal corresponding to each data point is obtained, envelope demodulation is performed on the filtering signal, an envelope signal corresponding to each data point is obtained according to the demodulation result, and the kurtosis of the envelope signal is analyzed; The frequency band corresponding to the data point with the maximum kurtosis is selected as the optimal filter to perform bandpass filtering on the ECG signal, and the envelope signal of the ECG signal after filtering is obtained based on the filtering result and Hilbert transform; Performing Fourier transform operation on the envelope signal based on the optimal harmonic energy and algorithm, obtaining the spectrum distribution of the envelope signal according to the transformation result, and performing Fourier transform operation on the spectrum distribution to generate an inverse envelope spectrum; The harmonic energy and maximum value of the ECG signal are searched according to the inverse envelope spectrum and sequence information entropy, and the peak data of the ECG signal is output based on the harmonic energy and maximum value.

7. A heart disease prediction system based on electrocardiogram signals and a large model according to claim 6, characterized in that: The envelope signal expression of the filtered ECG signal is: ； In the formula, represents the envelope signal of the ECG signal after filtering, represents the ECG signal after filtering, t represents the time period, dτ represents the infinitesimal element for integrating the ECG signal after filtering, and τ represents the integral variable.

8. The heart disease prediction system based on electrocardiogram signals and large models according to claim 7, characterized in that: The step of segmenting a single cycle of the electrocardiogram signal according to the peak data, extracting a single heart beat signal within the cycle, and generating a waveform segment based on the single heart beat signal includes: A structural element is constructed according to the number of peak data and ECG signal data points, a start segmentation point and an end segmentation point are defined by the structural element, and a single cycle of the ECG signal is segmented based on the start segmentation point and the end segmentation point; Analyze the peak position of the largest segmentation impulse in the segmentation result, determine the start and end time of the main impulse within the segmentation range based on the position result, and extract a single heart beat signal within a single cycle according to the main impulse information; The waveform segment of a single heart beat is extracted based on a single heart beat signal, and the amplitude of the waveform segment is normalized to a preset interval and converted into a standard normal distribution form to obtain the final waveform segment.

9. A heart disease prediction system based on electrocardiogram signals and a large model according to claim 8, characterized in that: The method of extracting ECG features based on historical ECG data, generating an encoder and a decoder based on a deep learning model, and using the ECG features as input to train the encoder and the decoder to obtain a medical prediction model includes: Obtain historical ECG data based on offline and online collection methods, and extract ECG signal features after data cleaning and preprocessing of the historical ECG data; Generate a pre-trained large model including an encoder and a decoder based on the deep learning large model, and set the structural information of the encoder and the decoder according to the attention mechanism and the feedforward network; The encoder and decoder are trained with ECG features as input after the structural information is set, and the training results are used to determine the errors between the encoder and decoder. The result information of the encoder and decoder is optimized according to the error results, and a large medical prediction model is obtained based on the optimization results.

Images