Multi-lead ECG data estimation methods, computer equipment, and storage media

Abstract

This application discloses a method, computer device, and storage medium for multi-lead ECG data extrapolation. The method includes: acquiring raw ECG data; extracting features from the raw ECG data to obtain corresponding feature parameters to be matched; determining a target extrapolation model that matches the feature parameters to be matched from multiple preset extrapolation models; and using the target extrapolation model to extrapolate leads from the raw ECG data to generate target ECG data with a greater number of leads than the raw ECG data. Through this method, new leads can be extrapolated from the raw ECG data to expand the number of leads, thereby obtaining target multi-lead ECG data and improving the accuracy of the extrapolated multi-lead results.

Description

Technical Field

[0001] This application relates to the field of medical device technology, and in particular to methods for calculating multi-lead electrocardiogram data, computer equipment, and storage media. Background Technology

[0002] In the medical field, the emergence of various medical testing equipment and technologies has provided strong support and protection for people's healthy lives. Among them, electrocardiography (ECG) is a technique that uses an electrocardiograph to record the waveforms of electrical activity changes generated by the heart during each cardiac cycle from the body surface. Through ECG, medical personnel can determine the health status of the person being tested.

[0003] An electrocardiogram (ECG) may include several leads. Generally, the more leads, the more complete the information obtained. However, in some cases, certain leads may be difficult to measure, or lead signals may be lost due to electrode detachment or poor contact, or excessive noise may affect the analysis. Insufficient lead signals make it difficult to assess the health status of the person being tested. Summary of the Invention

[0004] The main technical problem addressed by this application is to provide a method, computer equipment, and storage medium for calculating multi-lead electrocardiogram (ECG) data, which can calculate new leads based on original ECG data to expand the number of leads, thereby obtaining target multi-lead ECG data and improving the accuracy of the multi-lead calculation results.

[0005] To address the aforementioned technical problems, the first technical solution adopted in this application is: providing a method for extrapolating multi-lead electrocardiogram (ECG) data. This method includes: acquiring raw ECG data; extracting features from the raw ECG data to obtain corresponding feature parameters to be matched; determining a target extrapolation model that matches the feature parameters to be matched among multiple preset extrapolation models; and using the target extrapolation model to extrapolate leads from the raw ECG data to generate target ECG data with a lead count greater than the number of leads in the raw ECG data.

[0006] To address the aforementioned technical problems, the second technical solution adopted in this application is to provide a computer device, which includes a processor, a memory, and a communication circuit. The communication circuit and the memory are respectively coupled to the processor. The memory is used to store computer programs, and the processor is used to read and execute the computer programs to implement the multi-lead electrocardiogram data estimation method provided in the first technical solution.

[0007] To solve the above-mentioned technical problems, the third technical solution adopted in this application is: to provide a computer-readable storage medium that stores a computer program that can be read and executed by a processor to realize the multi-lead electrocardiogram data estimation method provided by the first technical solution.

[0008] The beneficial effects of this application are as follows: Unlike existing technologies, this method acquires raw electrocardiogram (ECG) data, extracts features from the raw ECG data to obtain corresponding matching feature parameters, determines a target estimation model that matches the matching feature parameters from multiple preset estimation models, and uses the target estimation model to perform lead estimation on the raw ECG data to generate target ECG data with a greater number of leads than the original ECG data. By extracting the matching feature parameters corresponding to the raw ECG data and determining a target estimation model that matches the matching feature parameters from multiple different estimation models, and then using the target estimation model to perform lead estimation on the raw ECG data, this method can... New leads are extrapolated from the original ECG data to expand the number of leads, thus obtaining target ECG data with multiple leads. Since the target extrapolation model used to extrapolate the number of leads matches the matching feature parameters corresponding to the original ECG data, on the one hand, it can take into account the individual differences of the test subjects and reduce the impact of individual differences on the accuracy of the extrapolation results, which is conducive to improving the accuracy of the extrapolation results of multiple leads. On the other hand, the extrapolation process involves less computation, thus requiring less computing power, and can be completed using a single-chip microcomputer. In addition, the ECG data used does not include the test subjects' gender, age, height and other private information, which is conducive to protecting the privacy of the test subjects. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating an embodiment of the multi-lead electrocardiogram data estimation method of this application;

[0010] Figure 2 A flowchart illustrating the process of determining feature parameters or a pre-defined classification model;

[0011] Figure 3 A flowchart illustrating the process of determining the sub-training database and sub-inference model. Detailed Implementation

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0013] Through long-term research, the inventors have discovered that electrocardiography (ECG) is a technique that uses an electrocardiograph to record the waveforms of electrical activity changes generated by the heart during each cardiac cycle from the body surface. Through ECG, medical personnel can assess the health status of the person being tested. An ECG can include several leads, i.e., multi-lead ECG data. Multi-lead ECG data refers to the data recorded and analyzed from multiple leads used in ECG testing. Each lead represents an electrophysiological signal from a different part of the heart, typically including 12 standard leads. These leads cover various aspects of the heart, such as the front chest, back, left and right sides. Multi-lead ECG can provide more comprehensive cardiac information, which is helpful in analyzing various cardiovascular diseases. Generally speaking, the more leads there are, the more complete the information can be detected. However, in some cases, certain lead signals are difficult to measure, or electrode detachment or poor contact may lead to the loss of corresponding lead signals or excessive noise affecting analysis. Insufficient lead signals hinder the assessment of the person's health status. To address this technical problem, this application provides the following embodiments.

[0014] like Figure 1 As shown in the embodiment of the multi-lead ECG data estimation method of this application, the multi-lead ECG data estimation method includes: S100: acquiring raw ECG data. S200: extracting features from the raw ECG data to obtain corresponding feature parameters to be matched. S300: determining a target estimation model that matches the feature parameters to be matched among multiple preset estimation models. S400: using the target estimation model to perform lead estimation on the raw ECG data to generate target ECG data with a greater number of leads than the original ECG data.

[0015] The process of constructing a multi-lead (N leads, M < N) electrocardiogram (ECG) system by extrapolating signals from known few leads (M leads) of ECG data is called the few-lead extrapolation multi-lead ECG technique. Based on the ECG data of known actual measurement sites in the original ECG data, this technique can be used to extrapolate ECG signals from ECG sites that were not actually measured, thereby obtaining target ECG data with a larger number of leads.

[0016] For example, the raw ECG data is obtained from a 12-lead detection, while the target ECG data is 18 leads. An 18-lead ECG adds six additional leads (V3R, V4R, V5R, V7, V8, and V9) compared to a 12-lead ECG, effectively detecting myocardial ischemia in the left ventricular posterior wall and right ventricle, which is undetectable by a conventional 12-lead ECG. This provides a significant advantage for comprehensive assessment of blood supply to multiple parts of the myocardium. However, measuring an 18-lead ECG requires attaching twelve electrodes to the subject's chest, while a 12-lead ECG only requires six. The increased number of electrodes introduces more interference and increases the installation difficulty. Therefore, during static ECG monitoring, an 18-lead ECG can be obtained based on a synchronous 12-lead ECG signal. In other embodiments, the raw and target ECG data can be 10 and 12 leads, or 12 and 15 leads, respectively.

[0017] For example, in Holter monitoring, electrode detachment and poor contact often lead to signal loss in certain leads or excessive noise affecting analysis. For children and pediatric patients, the smaller chest area necessitates the selection of fewer chest leads for acquisition, inevitably resulting in some information loss and negatively impacting clinical diagnosis. Therefore, Holter monitoring can utilize fewer leads (<12 leads) to obtain a 12-lead ECG.

[0018] By extracting the feature parameters to be matched from the original ECG data and determining the target estimation model that matches the feature parameters among multiple different estimation models, the target estimation model is then used to estimate the leads from the original ECG data. This allows for the estimation of new leads based on the original ECG data, thus expanding the number of leads and obtaining target ECG data with multiple leads. Since the target estimation model used to estimate the number of leads matches the feature parameters to be matched from the original ECG data, it can take into account the individual differences of the tested subjects, reducing the impact of individual differences on the accuracy of the estimation results and improving the accuracy of the multi-lead estimation results. On the other hand, the estimation process involves less computation, thus requiring less computing power, and can be completed using a single-chip microcomputer.

[0019] Compared to optimizing lead estimation algorithms with specific computational models (e.g., employing various deep learning network models), the method described in this application simplifies model complexity and significantly reduces computational load. Compared to constructing specific computational models for each individual or specific population, the method described in this application is easier to implement and takes into account individual variations and the diversity of sources of individual differences, thus improving the accuracy of multi-lead estimation results. Furthermore, the ECG data used in the method described in this application does not include privacy information such as the gender, age, and height of the tested subjects, which helps protect the privacy of the tested subjects.

[0020] It should be noted that, in some embodiments, the raw ECG data obtained in step S100 can be a set of data obtained from detecting the subject in the same time period, in which case the raw ECG data used in steps S200 and S400 are both this set of data. In other embodiments, the raw ECG data obtained in step S100 can be two sets of data obtained from detecting the subject in two close time periods, in which case the raw ECG data used in steps S200 and S400 can be one of the two sets of data.

[0021] The following is a detailed description of the embodiments of the multi-lead electrocardiogram data estimation method of this application.

[0022] S100: Acquire raw electrocardiogram data.

[0023] Specifically, raw ECG data can be obtained from the subject using electrodes with a limited number of leads. Subsequent calculations involving multiple leads can then yield more ECG data than the raw data. For example, if the subject's raw ECG data is 12 leads, calculations from 12 to 18 leads are needed.

[0024] S200: Extract features from the raw electrocardiogram data to obtain the corresponding feature parameters to be matched.

[0025] Raw electrocardiogram (ECG) data can include a large amount of data, and can be presented as an ECG chart. Specific characteristics of raw ECG data include: continuous variables, meaning variables that can take any value within a certain range, such as the QRS axis (A). QRS P wave, electrocardiographic axis (A) P ), T-wave electrical axis (A) T ), QRS amplitude in lead V1 (AP) V1 ), QRS amplitude in lead V2 (AP) V2 ), QRS amplitude in lead V3 (AP) V3 QRS duration, PR interval, AP V1 With AP V2 The ratios, etc.; qualitative variables, such as whether there is left ventricular high voltage (LVHV), whether there is left atrial hypertrophy (LAE), whether there is ST segment elevation, etc. The feature parameters to be matched can be at least one variable included in the original ECG data, which can be used to guide the finding of the target extrapolation model corresponding to the original ECG data. For example, the feature parameters to be matched can be the QRS axis (A... QRS ) and P wave electrocardiographic axis (A P According to the QRS axis (A) QRS ) and P wave electrocardiographic axis (A P A target estimation model corresponding to the original electrocardiogram data can be found.

[0026] Optionally, feature extraction is performed on the raw electrocardiogram data to obtain corresponding feature parameters to be matched, including:

[0027] S210: Identify valid ECG data from the raw ECG data.

[0028] The raw ECG data is large in volume, and some data segments may be interfered with and have unclear features. Therefore, it is necessary to select data segments that are free from interference and have clear features as valid ECG data.

[0029] S220: Extract features from valid electrocardiogram data to obtain the corresponding feature parameters to be matched.

[0030] Thus, guided by the matching feature parameters obtained from effective electrocardiogram data, the target estimation model can be found more accurately, which helps to improve the accuracy of the results of multi-lead estimation.

[0031] S300: Determine the target inference model that matches the feature parameters to be matched among multiple different preset inference models.

[0032] Different preset estimation models can be used to estimate from few leads to multiple leads, but the degree of matching between different preset estimation models and the original ECG data varies. Preset estimation models with a higher degree of matching with the original ECG data have higher accuracy in estimating multiple leads. Through step S300, the target estimation model with a high or highest degree of matching with the feature parameters to be matched can be determined, which helps to improve the accuracy of the multi-lead estimation results.

[0033] In some embodiments, the feature parameter to be matched can be matched one by one with different preset inference models. When a preset inference model that matches the feature parameter meets a preset requirement, the preset inference model is determined as the target inference model, which helps to improve efficiency. In some embodiments, the feature parameter to be matched can be matched with all preset inference models. After matching, the preset inference model with the highest degree of matching with the feature parameter to be matched is selected as the target inference model, which helps to improve the accuracy of multi-lead inference.

[0034] Optionally, the multiple preset inference models include a total inference model and multiple sub-inference models. The total inference model corresponds to the total training database, and the multiple sub-inference models each correspond to a different sub-training database. There is no overlap between the different sub-training databases, and the total training database is a collection of all the sub-training databases.

[0035] Among several different preset inference models, a target inference model that matches the feature parameters to be matched is determined, including:

[0036] S310: Determine whether there is a sub-inference model among multiple sub-inference models that matches the feature parameters to be matched.

[0037] S320: If it exists, the sub-inference model that matches the feature parameters to be matched will be used as the target inference model.

[0038] Specifically, each sub-training database is a different subset of the total training database. Each sub-training database can be obtained by partitioning the total training database. The total training database includes different multi-lead ECG data, where multi-lead refers to a number of leads greater than the original ECG data of the subject. For example, if the original ECG data of the subject is 12-lead ECG data, and the calculation needs to be performed from 12 to 18 leads, then the total training database includes different 18-lead ECG data. Corresponding to the total training database, each sub-training database also includes different multi-lead ECG data. The overall calculation model can be obtained by training on the total training database, and each sub-calculation model can be obtained by training on each sub-training database separately.

[0039] Different sub-estimation models and the overall estimation model can be used to estimate from few leads to multiple leads. However, different sub-estimation models have varying degrees of matching with the original ECG data. Sub-estimation models with a higher degree of matching with the original ECG data have higher accuracy in estimating multiple leads. Through steps S310 and S320, the target estimation model with a high or highest degree of matching with the feature parameters to be matched can be determined, which helps to improve the accuracy of the multi-lead estimation results.

[0040] Optionally, the degree of matching between multiple sub-inference models and the feature parameters to be matched can be determined by methods such as Euclidean distance, regression analysis, SVM support vector machine, and neural network classification model.

[0041] Optionally, determining the target inference model that matches the feature parameters to be matched among multiple preset inference models further includes:

[0042] S330: If there is no sub-inference model among the multiple sub-inference models that matches the feature parameter to be matched, then the overall inference model shall be used as the target inference model that matches the feature parameter to be matched.

[0043] When there is no sub-estimation model among the multiple sub-estimation models that matches the feature parameters to be matched, the total estimation model can be used as the target estimation model that matches the feature parameters to be matched because the total estimation model has a larger data volume and higher compatibility with the original ECG data.

[0044] Optionally, the sub-inference model that matches the feature parameters to be matched is used as the target inference model, including:

[0045] S321: Calculate the similarity between the feature parameters to be matched and the feature parameters corresponding to each sub-inference model.

[0046] S322: Select the sub-inference model with the highest similarity from multiple sub-inference models and use it as the target inference model.

[0047] The feature parameter to be matched and the feature parameter corresponding to each sub-inference model can be the same variable, but the value or state of the variable is different. For example, the feature parameter to be matched is the QRS axis (A). QRS P wave, electrocardiographic axis (A) P ) and T-wave electrical axis (A T Correspondingly, the characteristic parameters of each sub-calculation model are the QRS electrocardiographic axis (A). QRS P wave, electrocardiographic axis (A) P ) and T-wave electrical axis (A T For example, the feature parameters to be matched include the QRS axis (A). QRS P wave, electrocardiographic axis (A) P ) and T-wave electrical axis (A T The feature parameters to be matched consist of multiple variables, including A, and can be expressed as y = A. QRS A P A T …, the characteristic parameters corresponding to the sub-inference model can be expressed as… Here, i corresponds to different sub-calculation models and sub-databases. For example, when i = 1,

[0048] The higher the similarity between the feature parameters corresponding to the sub-estimation model and the feature parameters to be matched, the higher the accuracy of multi-lead estimation using the sub-estimation model. Through steps S321 and S322, the sub-estimation model with the highest similarity to the feature parameters to be matched can be determined as the target estimation model, which helps to improve the accuracy of multi-lead estimation results.

[0049] In some embodiments, the similarity between the feature parameters to be matched and the feature parameters corresponding to each sub-inference model can be calculated using the Euclidean distance method. Specifically, the similarity between y and... The Euclidean distance is used to determine the sub-inference model corresponding to the minimum Euclidean distance, which is then used as the target inference model.

[0050] Optionally, both the total training database and the sub-training databases include preset ECG data. Before determining the target inference model that matches the feature parameters to be matched among several different preset inference models, the process includes:

[0051] A341: Feature extraction is performed on the preset ECG data in each sub-training database to obtain the feature parameters corresponding to each sub-inference model. In this way, the feature parameters can reflect the characteristics of the corresponding sub-inference model, and the feature parameters to be matched can be matched with the sub-inference model through the feature parameters.

[0052] The preset ECG data includes different multi-lead ECG data. For example, if it is necessary to use the sub-calculation model to calculate from 12 leads to 18 leads, then the preset ECG data includes different 18-lead ECG data.

[0053] like Figure 2 As shown, the features extracted from the preset ECG data in each sub-training database can specifically be ECG variables. ECG variables can include: continuous variables, that is, variables that can take any value within a certain range, such as the QRS axis (A). QRS P wave, electrocardiographic axis (A) P ), T-wave electrical axis (A) T ), QRS amplitude in lead V1 (AP) V1 ), QRS amplitude in lead V2 (AP) V2 ), QRS amplitude in lead V3 (AP) V3 QRS duration, PR interval, AP V1 With AP V2 The ratios, etc.; qualitative variables, such as whether there is left ventricular high voltage (LVHV), whether there is left atrial hypertrophy (LAE), whether there is ST segment elevation, etc. The characteristic parameters corresponding to each sub-calculation model can be one ECG variable, or can include two or more ECG variables, or can include the calculation results obtained by calculating the above variables according to a certain formula.

[0054] Optionally, feature extraction is performed on the preset ECG data in each sub-training database to obtain the feature parameters corresponding to each sub-inference model, including:

[0055] A3411: Use statistical analysis methods to determine the effective features that influence the partitioning of the sub-training database from the extracted features.

[0056] A3412: Use the effective features of each preset ECG data in the sub-training database to obtain the feature parameters corresponding to the sub-inference model.

[0057] For example, effective features influencing the partitioning of the sub-training database can be determined from the extracted features using methods such as Pearson chi-square test, T-test, Fisher discriminant method, and principal component analysis.

[0058] Each sub-training database can be obtained by dividing the total training database. The differences between the sub-training databases can be reflected in the different values ​​or states of the effective features. The effective features implicitly contain the basis for dividing the sub-training databases; that is, the effective features corresponding to each sub-training database can reflect the characteristics of that sub-training database. For example, for the extracted feature A... QRS When the ANOVA test shows P < α (α is a preset parameter, which can be 0.01, 0.05, etc.), it is considered that A... QRS For effective features that influence the partitioning of the sub-training database, otherwise A is considered... QRS This is an invalid feature.

[0059] Through steps A3411 and A3412, effective features affecting the partitioning of the sub-training database can be identified among all extracted features. By utilizing these effective features to obtain the feature parameters corresponding to the sub-inference model, the amount of data computation can be reduced, and the effectiveness of the feature parameters in reflecting the characteristics of the sub-inference model can be improved.

[0060] Optionally, the feature parameters corresponding to the sub-inference model are obtained by utilizing the effective features of each preset ECG data in the sub-training database, including:

[0061] A3413: Mean the effective features in the sub-training database to obtain the corresponding effective feature mean.

[0062] A3414: The mean values ​​of the effective features corresponding to the effective features are used to form a parameter vector to obtain the feature parameters corresponding to the sub-inference model.

[0063] This reduces the amount of data computation and improves the effectiveness of feature parameters in reflecting the characteristics of the sub-inference model. For example, effective features include the QRS axis (A... QRS P wave, electrocardiographic axis (A) P ) and T-wave electrical axis (A T The mean of the effective features corresponding to the effective features is (including multiple variables, the effective features are...). The feature parameters corresponding to the sub-inference model (i.e., the parameter vector composed of the effective feature mean) can be expressed as: Here, i corresponds to different sub-calculation models and sub-databases. For example, when i = 1,

[0064] Optionally, the sub-inference model that matches the feature parameters to be matched is used as the target inference model, including:

[0065] S323: Input the feature parameters to be matched into the preset classification model.

[0066] S324: Determine the database labels that match the feature parameters to be matched using a preset classification model. Each sub-inference model corresponds to a sub-training database with its own database labels.

[0067] S325: Use the sub-inference model corresponding to the sub-training database pointed to by the database label as the target inference model.

[0068] Specifically, the preset classification model is used to guide the feature parameter to be matched to find a matching database label. The database label corresponds to the sub-training database, and the sub-training database corresponds to the sub-inference model. Therefore, after the feature parameter to be matched finds a matching database label through the preset classification model, the database label can be used to determine which sub-inference model matches the feature parameter to be matched.

[0069] In some embodiments, the preset classification model is a Support Vector Machine (SVM) sub-database classification model. The sub-inference model that matches the feature parameters to be matched can be determined by the SVM method. The specific process is as follows: input the feature parameters to be matched into the SVM sub-database classification model, perform classification prediction, obtain the sub-training database that best matches the feature parameters to be matched, and use the sub-inference model corresponding to the sub-training database as the target inference model.

[0070] Optionally, both the total training database and the sub-training databases include preset ECG data. Before determining the target inference model that matches the feature parameters to be matched among several different preset inference models, the process includes:

[0071] A342: Extract features from the preset ECG data in each sub-training database to obtain a preset classification model.

[0072] In this way, the pre-defined classification model can perceive the characteristics of the sub-training database. The feature parameters to be matched can determine the database labels of the sub-training databases that match them through the pre-defined classification model, and then determine the sub-inference model that matches them.

[0073] Optionally, feature extraction is performed on preset ECG data in each sub-training database to obtain a preset classification model, including:

[0074] A3421: Use statistical analysis methods to determine the effective feature set that influences the partitioning of the sub-training database from the extracted features.

[0075] For example, the sub-training database can be preprocessed using Fisher's discriminant method or principal component analysis to obtain an effective feature set.

[0076] The effective feature set can be a collection of effective features. Each sub-training database can be obtained by dividing the total training database. The differences between the sub-training databases can be reflected in the differences in their effective feature sets. The effective feature set implicitly contains the basis for dividing the sub-training databases. In other words, the effective feature set corresponding to each sub-training database can reflect the characteristics of that sub-training database.

[0077] A3422: Construct a preset classification model using the effective feature set of each preset ECG data and the sub-training database to which the preset ECG data belongs.

[0078] For example, the preset classification model is a Support Vector Machine (SVM) sub-database classification model. The effective feature set of each preset ECG data and the database labels of the sub-training database can be used as input, and an appropriate kernel function and parameters can be selected to build the SVM sub-database classification model.

[0079] Through steps A3421 and A3422, the effective feature set that influences the partitioning of the sub-training database can be determined from all the extracted features. The preset classification model can be obtained by using the effective feature set, which can reduce the amount of data computation and improve the accuracy of the preset classification model in perceiving the characteristics of the sub-training database.

[0080] The following describes in detail the generation process of the overall estimation model and each sub-estimate model.

[0081] Optionally, before determining the target inference model that matches the feature parameters to be matched among multiple pre-defined inference models, the process includes:

[0082] A310: A total training database is established using multiple preset ECG data sets as training data. The number of leads in the preset ECG data sets is the same as the number of leads in the target ECG data sets.

[0083] For example, if the target ECG data is 18 leads, then the preset number of leads for ECG data is also 18 leads, and the total training database includes multiple 18-lead ECG data sets.

[0084] A320: Use the total training database as the training set to train and generate the total inference model.

[0085] For example, the overall inference model is one of the following: a linear regression model such as a small binary regression model or a minimum binary regression model based on principal component analysis, or a nonlinear model such as a support vector machine or a neural network model.

[0086] A330: Divide multiple preset ECG data into multiple sub-training databases in the total training database.

[0087] A340: Each sub-training database is used as a training set to train and generate the corresponding sub-inference model.

[0088] For example, each sub-inference model is one of the following: a linear regression model such as a small binary regression model or a minimum binary regression model based on principal component analysis; or a nonlinear model such as a support vector machine or a neural network model. Furthermore, the model category of each sub-inference model is consistent with that of the overall inference model.

[0089] By using multiple preset electrocardiogram data as training data, the overall prediction model and each sub-prediction model can be trained and generated without involving the privacy information of the test subjects such as gender, age, height, weight, and chest circumference. This eliminates the need to actually measure the height, weight, and chest circumference of the test subjects in advance, avoids the problem of data grouping difficulties and the problem of mismatch of old individual coefficients due to individual changes, and also helps to protect the privacy of the test subjects.

[0090] By dividing the data into multiple sub-training databases and training corresponding sub-inference models, individual differences in the training data sources can be taken into account, reducing the impact of individual variations and differences in the training data sources on the accuracy of the inference results, which is beneficial to improving the accuracy of the inference results for multi-lead inference.

[0091] It should be noted that in some embodiments, steps A330 and A340 can be performed alternately, that is, after obtaining a sub-training database in step A330, the corresponding sub-inference model can be trained and generated in step A340. In other embodiments, steps A330 and A340 can also be performed sequentially, that is, after obtaining all sub-training databases in step A330, the corresponding sub-inference model can be trained and generated in step A340.

[0092] Optionally, the multiple preset ECG data in the total training database are divided into multiple sub-training databases, including:

[0093] A331: The overall calculation model is used to perform lead calculations on each preset ECG data in the overall training database to obtain the calculation results corresponding to the preset ECG data.

[0094] A332: The set of preset ECG data corresponding to the calculation results that pass the preset calculation standard is used as a sub-training database, and the set of preset ECG data corresponding to the calculation results that do not pass the preset calculation standard is used as a transitional dataset.

[0095] A333: Cascade the transition dataset to obtain other sub-training databases.

[0096] This setup allows for a gradual, gradient-based division of the total training database into different sub-training databases. Each sub-training database has different characteristics, and the sub-inference models generated from each sub-training database are applicable to different test subjects. The sub-inference model that matches the test subject is more targeted to the test subject's original ECG data, which helps improve the accuracy of multi-lead inference results.

[0097] Specifically, such as Figure 3 As shown, the process of dividing the multiple preset ECG data in the total training database into multiple sub-training databases is as follows:

[0098] Step A331 includes the following steps A3311. A3311: The total training database is represented as T1. The total training database T1 is used as the training set to train and generate the total prediction model S1. The data of the few leads of the preset ECG data in the total training database T1 are used as the input of the total prediction model S1 to obtain the prediction results of multiple leads.

[0099] Step A332 includes the following steps A3321, A3322, and A3323. A3321: Set a preset estimation standard and determine whether the estimation result obtained in step A3311 passes the preset estimation standard. The preset estimation standard can determine the degree of similarity between the multi-lead estimation result and the preset ECG data of the multi-lead. A3322: Construct the first sub-training database G1 using the preset ECG data corresponding to the estimation results that pass the preset estimation standard. A3323: Use the set of preset ECG data corresponding to the estimation results that do not pass the preset estimation standard as the transition dataset T2. After step A3322, step A340 can be executed, specifically: using the sub-training database G1 as the training set to train and generate the corresponding sub-estimate model M1.

[0100] Optionally, the transition dataset can be cascaded to obtain other sub-training datasets, including:

[0101] A334: Determine if the current transition dataset contains zero data.

[0102] A335: If so, stop cascading processing.

[0103] A336: If not, then use the current transition dataset as the training set to train and generate a transition estimation model. Use the transition estimation model to perform lead estimation on each preset ECG data in the current transition dataset to obtain the estimation results corresponding to each preset ECG data in the current transition dataset.

[0104] A337: The set of preset ECG data corresponding to the calculation results that pass the preset calculation standard is used as a sub-training database, and the set of preset ECG data corresponding to the calculation results that do not pass the preset calculation standard is used as the next transition dataset.

[0105] If the transition dataset has zero data volume, it means that multiple preset ECG data points from the main training database have been divided into sub-training databases. Once the division process is complete, the cascading processing stops, and step A340 is executed to train and generate the sub-inference model corresponding to the sub-training database. If the transition dataset has zero data volume, it means that the process of dividing the sub-training database is incomplete, and the cascading processing continues to obtain a new sub-training database.

[0106] This setup allows for a gradual, gradient-based recursive partitioning of the transition dataset, resulting in different characteristics for each sub-training database. The sub-inference models trained from each sub-training database are applicable to different test subjects. The sub-inference model that matches the test subject is more targeted to the test subject's original ECG data, which helps improve the accuracy of multi-lead inference results from the test subject's original ECG data.

[0107] Optionally, the transitional inference model can be one of the following: a linear regression model such as a small binary regression model or a minimum binary regression model based on principal component analysis; or a nonlinear model such as a support vector machine or a neural network model. Furthermore, each transitional inference model is consistent with the model category of the overall inference model.

[0108] Specifically, the process of cascading the transition dataset to obtain other sub-training datasets is as follows:

[0109] Step A336 includes the following steps A3361 and A3362. A3361: Using the transition dataset Ti (i = 2, 3, 4…) obtained in the previous steps as the training set, train the corresponding transition estimation model Si (i = 2, 3, 4…). A3362: Using the data from the few leads portion of the preset ECG data in the transition dataset Ti (i = 2, 3, 4…) as inputs to the transition estimation model Si (i = 2, 3, 4…), obtain the multi-lead estimation results.

[0110] For example, the process in A3361 is as follows: using the transition dataset T2 from step A3323 as the training set, the corresponding transition estimation model S2 is trained. The process in A3362 is as follows: using the data from the few leads of the preset ECG data in the transition dataset T2 as the input to the transition estimation model S2, the estimation results of multiple leads are obtained.

[0111] For example, the process in A3361 is as follows: using the transition dataset T3 from step A3373 as the training set, the corresponding transition estimation model S3 is trained. The process in A3362 is as follows: using the data from the few leads of the preset ECG data in the transition dataset T3 as the input to the transition estimation model S3, the estimation results of multiple leads are obtained.

[0112] Step A337 includes the following steps A3371, A3372, and A3373. A3371: Determine whether the estimation result obtained in step A3362 meets the preset estimation criteria. A3372: The preset ECG data corresponding to the estimation results that meet the preset estimation criteria constitute a sub-training database Gi (i = 2, 3, 4…). A3373: The set of preset ECG data corresponding to the estimation results that do not meet the preset estimation criteria is used as a transition dataset Ti+1 (i = 2, 3, 4…). After step A3372, step A340 can be executed, specifically: using the sub-training database Gi (i = 2, 3, 4…) as the training set, train and generate the corresponding sub-estimation model Mi (i = 2, 3, 4…). For example, using the sub-training database G2 as the training set, train and generate the corresponding sub-estimation model M2.

[0113] For example, the process in A3371 is as follows: determine whether the estimation result obtained in step A3362 meets the preset estimation standard. The process in A3372 is as follows: the preset ECG data corresponding to the estimation results that meet the preset estimation standard constitute the sub-training database G2. The process in A3373 is as follows: the set of preset ECG data corresponding to the estimation results that do not meet the preset estimation standard is used as the transition dataset T3. After step A3373, cascaded processing continues, that is, data loop processing is performed, and steps A334, A335, A336 and A337 are executed again to obtain the transition dataset T4.

[0114] This process continues until the calculation results of all preset ECG data pass the preset calculation criteria. The cascading processing is then complete, the sub-training database is partitioned, and a lead calculation model library is constructed through step A340. This library consists of a total calculation model S1 and sub-calculation models Mi (i = 1, 2, 3…). The training set corresponding to the total calculation model S1 is the total training database T1, and the training sets corresponding to the sub-calculation models Mi (i = 1, 2, 3…) are the corresponding sub-training databases Gi (i = 1, 2, 3…). The total calculation model S1 can serve as a backup calculation model. In subsequent lead calculations, if the data to be calculated cannot match the optimal sub-calculation model Mi (i = 1, 2, 3…), the total calculation model S1 is used. Thus, instead of partitioning the sub-training database based on the characteristics of the preset ECG data itself, the partitioning is based on the calculation results of the transitional calculation model, which helps improve the accuracy of multi-lead calculations.

[0115] It should be noted that during the execution of the program from step A3373 to step A3361, the assignment process i = i + 1 needs to be performed first, and then step A3361 is executed.

[0116] The computer device described in the embodiments of this application includes a processor, a memory, and a communication circuit. The communication circuit and the memory are respectively coupled to the processor. The memory is used to store computer programs, and the processor is used to read and execute the computer programs to implement the above-described multi-lead electrocardiogram data estimation method.

[0117] A processor can also be called a CPU (Central Processing Unit). A processor may be an integrated circuit chip with signal processing capabilities. A processor can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor.

[0118] The memory is used to store computer programs and can be RAM, ROM, or other types of storage devices. Specifically, the memory may include one or more computer-readable storage media, which may be non-transitory. The memory may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in the memory is used to store at least one line of program code.

[0119] The processor is used to execute the computer program stored in the memory to implement the multi-lead electrocardiogram data estimation method of this application.

[0120] For a detailed description of the functions and execution processes of each functional module or component in the computer device embodiments of this application, please refer to the description in the embodiments of the multi-lead electrocardiogram data estimation method of this application, which will not be repeated here.

[0121] The computer-readable storage medium described in the embodiments of the storage medium in this application stores a computer program that can be read and executed by a processor to implement the above-described multi-lead electrocardiogram data estimation method.

[0122] For a detailed description of the functions and execution processes of each functional module or component in the storage medium embodiments of this application, please refer to the descriptions in the above embodiments of the multi-lead electrocardiogram data estimation method and the computer device embodiments of this application, which will not be repeated here.

[0123] In summary, this embodiment can extract the matching feature parameters corresponding to the original ECG data, determine the target estimation model that matches the matching feature parameters among multiple different estimation models, and then use the target estimation model to estimate the leads of the original ECG data. This allows for the estimation of new leads based on the original ECG data, expanding the number of leads and obtaining multi-lead target ECG data. Since the target estimation model used to estimate the number of leads matches the matching feature parameters corresponding to the original ECG data, it can take into account the individual differences of the tested subjects, reducing the impact of individual differences on the accuracy of the estimation results, thus improving the accuracy of multi-lead estimation. Furthermore, the estimation process involves less computation, thus requiring less computing power; a single-chip microcomputer can be used to complete the calculations. In addition, the ECG data used in the embodiments of this application does not include the gender, age, height, or other private information of the tested subjects, which helps protect the privacy of the tested subjects.

[0124] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for extrapolating multi-lead electrocardiogram data, characterized in that, include: Obtain raw electrocardiogram data; Feature extraction is performed on the raw electrocardiogram data to obtain the corresponding feature parameters to be matched; Among multiple preset inference models, a target inference model that matches the feature parameters to be matched is determined. The multiple preset inference models include a total inference model and multiple sub-inference models. The total inference model corresponds to a total training database, and the multiple sub-inference models correspond to different sub-training databases. There is no overlap between the different sub-training databases, and the total training database is a set of all the sub-training databases. The target estimation model is used to estimate the leads of the original ECG data to generate target ECG data with a greater number of leads than the original ECG data. The step of determining the target inference model that matches the feature parameters to be matched among multiple preset inference models includes: Determine whether there exists a sub-inference model among the plurality of sub-inference models that matches the feature parameter to be matched; If it exists, the sub-inference model that matches the feature parameters to be matched will be used as the target inference model; If it does not exist, then the overall estimation model is used as the target estimation model that matches the feature parameters to be matched.

2. The method according to claim 1, characterized in that, The step of using the sub-inference model that matches the feature parameters to be matched as the target inference model includes: Calculate the similarity between the feature parameters to be matched and the feature parameters corresponding to each of the sub-inference models; The sub-inference model with the highest similarity among the plurality of sub-inference models is determined as the target inference model.

3. The method according to claim 2, characterized in that, Both the total training database and the sub-training database include preset electrocardiogram data; before determining the target inference model that matches the feature parameters to be matched among different preset inference models, the process includes: Feature extraction is performed on the preset electrocardiogram data in each of the sub-training databases to obtain the feature parameters corresponding to each sub-inference model.

4. The method according to claim 3, characterized in that, The step of extracting features from the preset ECG data in each of the sub-training databases to obtain feature parameters corresponding to each sub-inference model includes: Statistical analysis methods are used to determine the effective features that influence the partitioning of the sub-training database from the extracted features; The feature parameters corresponding to the sub-inference model are obtained by utilizing the effective features of each preset electrocardiogram data in the sub-training database.

5. The method according to claim 4, characterized in that, The step of obtaining the feature parameters corresponding to the sub-inference model using the effective features of each preset ECG data in the sub-training database includes: The effective features in the sub-training database are averaged to obtain the corresponding effective feature mean. The mean values ​​of the effective features corresponding to the effective features are used to form a parameter vector to obtain the feature parameters corresponding to the sub-inference model.

6. The method according to claim 1, characterized in that, The step of using the sub-inference model that matches the feature parameters to be matched as the target inference model includes: The feature parameters to be matched are input into a preset classification model; The database labels that match the feature parameters to be matched are determined by the preset classification model; wherein, each sub-inference model corresponds to a sub-training database with a corresponding database label; The sub-inference model corresponding to the sub-training database pointed to by the database label is taken as the target inference model.

7. The method according to claim 6, characterized in that, Both the total training database and the sub-training database include preset electrocardiogram data; before determining the target inference model that matches the feature parameters to be matched among different preset inference models, the process includes: Feature extraction is performed on the preset electrocardiogram data in each of the sub-training databases to obtain the preset classification model.

8. The method according to claim 7, characterized in that, The step of extracting features from the preset ECG data in each of the sub-training databases to obtain the preset classification model includes: Statistical analysis methods are used to determine the effective feature set that influences the partitioning of the sub-training database from the extracted features; The preset classification model is constructed using the effective feature set of each preset ECG data and the sub-training database to which the preset ECG data belongs.

9. The method according to claim 1, characterized in that, Before determining the target inference model that matches the feature parameters to be matched among multiple preset inference models, the process includes: The total training database is established using multiple preset ECG data as training data; wherein the number of leads in the preset ECG data is the same as the number of leads in the target ECG data. The total training database is used as the training set to train and generate the total prediction model; The multiple preset electrocardiogram data in the total training database are divided into multiple sub-training databases; Each of the sub-training databases is used as a training set to train and generate the corresponding sub-inference model.

10. The method according to claim 9, characterized in that, The step of dividing the multiple preset ECG data in the total training database into multiple sub-training databases includes: The overall estimation model is used to perform lead estimation on each of the preset ECG data in the overall training database to obtain the estimation result corresponding to the preset ECG data; The set of preset ECG data corresponding to the calculation results that pass the preset calculation standard is used as a sub-training database, and the set of preset ECG data corresponding to the calculation results that do not pass the preset calculation standard is used as a transitional dataset. The transition dataset is cascaded to obtain the other sub-training databases.

11. The method according to claim 10, characterized in that, The cascading processing of the transition dataset to obtain other sub-training databases includes: Determine whether the current transition dataset contains zero data. If so, then stop the cascading process; If not, the current transition dataset is used as the training set to train and generate a transition estimation model. The transition estimation model is then used to perform lead estimation on each of the preset ECG data in the current transition dataset to obtain the estimation results corresponding to each of the preset ECG data in the current transition dataset. The set of preset ECG data corresponding to the calculation results that pass the preset calculation standard is used as a sub-training database, and the set of preset ECG data corresponding to the calculation results that do not pass the preset calculation standard is used as the next transition dataset.

12. The method according to claim 1, characterized in that, The step of extracting features from the original electrocardiogram data to obtain corresponding feature parameters to be matched includes: Identify the valid ECG data from the raw ECG data; Feature extraction is performed on the valid electrocardiogram data to obtain the corresponding feature parameters to be matched.

13. A computer device, characterized in that, include: A processor, a memory, and a communication circuit; the communication circuit and the memory are respectively coupled to the processor, the memory is used to store a computer program, and the processor is used to read and execute the computer program to implement the method as described in any one of claims 1-12.

14. A computer-readable storage medium, characterized in that, The device contains a computer program that can be read and executed by a processor to implement the method as described in any one of claims 1-12.

Images