A machine learning-based auxiliary prediction method for myocardial infarction

Abstract

The present invention provides a method for auxiliary prediction of myocardial infarction based on machine learning, which belongs to the field of machine learning technology. Based on the electrocardiogram recording and analysis system, the II lead electrocardiogram of each target patient in different body positions is collected, and the II lead electrocardiogram is preprocessed to obtain the current electrocardiogram; all current electrocardiograms of all target patients are classified by body position, and the HRV features under each body position classification are extracted to construct body position feature items; all body position feature items are divided into a training set and a prediction set, and the machine learning model is learned based on the training set and the machine learning model after learning is optimized based on the prediction set to obtain a prediction model; the II lead electrocardiogram of a new patient is collected and input into the prediction model for auxiliary prediction and output of myocardial infarction. The prediction model of myocardial infarction patients is established by using the machine learning method to improve the auxiliary prediction ability for myocardial infarction.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to a method for auxiliary prediction of myocardial infarction based on machine learning. Background Art

[0002] At present, the prediction and diagnosis of myocardial infarction (MI) mainly rely on traditional medical examination methods such as electrocardiogram (ECG), blood biochemical indicators, and medical history. In the diagnosis of acute MI, ECG changes are key indicators, combined with blood indicators such as myocardial markers such as troponin, which can help doctors make judgments. In addition, clinically, a comprehensive assessment is conducted based on risk factors such as the patient's age, gender, family history, and lifestyle habits to determine the individual's risk of MI. In recent years, there have also been some studies that use data-driven machine learning models to predict MI based on these clinical data. These models usually improve the accuracy of predictions through data preprocessing, feature extraction, and machine learning algorithm modeling.

[0003] However, there are the following problems and defects in the prior art:

[0004] Insufficient early warning capabilities: Traditional methods based on electrocardiograms and blood biochemical indicators mainly rely on physiological changes when the disease is already obvious, which leads to limited early warning capabilities for myocardial infarction and inability to identify potential risks in a timely manner.

[0005] Lack of consideration of different body positions and dynamic physiological changes: Traditional myocardial infarction diagnosis and prediction methods are usually based on data in static body positions, while ignoring the effects of different body positions and respiratory states on cardiac physiological characteristics, which may lead to an incomplete assessment of individual myocardial infarction risk.

[0006] Limitations of machine learning models: Although machine learning-based prediction models have improved prediction capabilities, most models rely on structured clinical data, which are susceptible to noise and incomplete information during collection and processing. In addition, most existing models are based on static data and lack effective use of temporal dynamic data (such as changes in HRV in different body positions).

[0007] Hidden and sudden physiological changes: Myocardial infarction is often accompanied by hidden and sudden occurrence, and early physiological changes are difficult to capture through conventional examination methods. Traditional examination methods lack attention to dynamic physiological characteristics (such as changes in heart rate in different body positions), making it difficult to provide effective early warning information before myocardial infarction occurs.

[0008] Lack of systematic research on multi-position data: Traditional methods usually collect data in a static position, ignoring the effects of different positions such as standing and lying on the physiological state of the heart. These changes in body position will have a significant impact on the autonomic nervous system, thereby changing heart rate variability. Failure to consider these factors makes risk assessment incomplete.

[0009] Complexity of temporal dynamic data modeling: Physiological indicators such as ECG signals and HRV have dynamic characteristics, and the changes under different body positions and respiratory states are complex. Existing machine learning models are difficult to effectively capture these complex temporal dynamic characteristics. In addition, obtaining high-quality multi-body position data requires strict control of the experimental design and data collection process, which increases the difficulty of research and the complexity of technical implementation.

[0010] Therefore, this application proposes a method for auxiliary prediction of myocardial infarction based on machine learning. Summary of the invention

[0011] The present invention provides an auxiliary prediction method for myocardial infarction based on machine learning, which is used to analyze the short-term HRV characteristics under multiple body posture changes, and then use machine learning methods to establish a prediction model for myocardial infarction patients, thereby improving the auxiliary prediction ability for myocardial infarction, so as to intervene as early as possible and reduce the risk of myocardial infarction.

[0012] The present invention provides a method for auxiliary prediction of myocardial infarction based on machine learning, comprising:

[0013] Step 1: Based on the ECG recording and analysis system, the lead II ECG of each target patient in different body positions is collected, and the lead II ECG is preprocessed to obtain the current ECG;

[0014] Step 2: Classify the body positions of all current ECGs of all target patients, extract the HRV features under each body position classification, and construct the body position-feature entries;

[0015] Step 3: Divide all body position-feature items into a training set and a prediction set, learn the machine learning model based on the training set, and optimize the learned machine learning model based on the prediction set to obtain a prediction model;

[0016] Step 4: Collect the II lead electrocardiogram of the new patient and input it into the prediction model for auxiliary prediction and output of myocardial infarction.

[0017] Preferably, preprocessing the lead II electrocardiogram to obtain the current electrocardiogram includes:

[0018] Inputting the II lead ECG into the noise analysis model, several different types of noise calibration bands are obtained;

[0019] Determine the initial proportion of each type of noise calibration band based on the lead II electrocardiogram;

[0020] Determine the acquisition components involved in the acquisition process of the ECG recording and analysis system, and retrieve the possible noise types of the relevant acquisition components from the component-noise comparison table;

[0021] Performing component testing on each acquisition component in the electrocardiogram recording and analysis system to obtain a lateral vibration discrete spectrum and a longitudinal vibration discrete spectrum of the corresponding acquisition component;

[0022] Analyze a first correlation coefficient between the lateral vibration discrete spectrum and a standard lateral spectrum and a second correlation coefficient between the longitudinal vibration discrete spectrum and a standard longitudinal spectrum;

[0023] Determine a noise compensation coefficient of a corresponding acquisition component based on the first correlation coefficient and the second correlation coefficient;

[0024]

[0025] Wherein, d1 is the corresponding first correlation coefficient; d2 is the corresponding second correlation coefficient; is the signal variance based on the discrete spectrum of transverse vibration and the standard transverse spectrum; is the signal variance based on the longitudinal vibration discrete spectrum and the standard longitudinal spectrum; max is the maximum value symbol; is the number of discrete signals existing in the discrete spectrum of lateral vibration; is the number of discrete signals in the longitudinal vibration discrete spectrum; e is the exponential function symbol; is the corresponding noise compensation coefficient;

[0026] Assigning noise weights to the noise types of the noise calibration band based on the component weights of each acquisition component in the electrocardiogram recording and analysis system, possible noise types, and noise compensation coefficients;

[0027]

[0028] in, represents the noise weight assigned to the noise type of the qth noise calibration band; Indicates the number of acquisition components that are consistent with the noise type of the qth noise calibration band; represents the component weight of the i1th consistent acquisition component; represents a function of a variable; Indicates the noise compensation coefficient of the i1th acquisition component that has a consistent existence; Indicates all The variance of Indicates the noise setting weight for the i1th acquisition component with the same noise type as the qth noise calibration band; It indicates the probability of noise generation with the same noise type as the qth noise calibration band when there is a consistent acquisition component in the i1th; Indicates all The variance of ; a1 represents the set probability; It represents the sum of the noise compensation coefficients involved in all acquisition components in the ECG recording and analysis system;

[0029] According to the noise weight corresponding to each noise calibration band and the initial proportion, the elimination coefficient of the corresponding noise calibration band is obtained, and each elimination coefficient is compared with the preset coefficient respectively;

[0030] If all the elimination coefficients are smaller than the preset coefficients, all the elimination coefficients are sorted in order of size, and the noise calibration band corresponding to the elimination coefficient ranked first is locked and regarded as the first calibration band;

[0031] Otherwise, the noise calibration band that is greater than or equal to the preset coefficient is regarded as the first calibration band;

[0032] A noise elimination method consistent with the noise type of the first calibration band is selected from the noise elimination database, and noise elimination is performed on the lead II electrocardiogram to obtain a current electrocardiogram.

[0033] Preferably, obtaining the elimination coefficient corresponding to the noise calibration band includes:

[0034]

[0035] in, represents the elimination coefficient of the qth noise calibration band; represents the initial proportion of the qth noise calibration band; represents the energy proportion of the qth noise calibration band; Indicates the total number of noise calibration bands; Indicates the horizontal width value of the qth noise calibration band; Indicates the horizontal width value of the lead II electrocardiogram; represents the energy value of the qth noise calibration band; Represents the total energy value of the lead II electrocardiogram.

[0036] Preferably, all current electrocardiograms of all target patients are classified by body position, including:

[0037] The body position descriptions corresponding to the single body position and the combined body position are obtained, and the electrocardiograms that are consistent with the corresponding body position descriptions are selected from all current electrocardiograms as the body position classification results.

[0038] Preferably, HRV features under each body position classification are extracted to construct body position-feature entries, including:

[0039] Extract the time domain features, frequency domain features and nonlinear features of each electrocardiogram under the same individual position classification to obtain the initial feature set;

[0040] Performing a t-test on the initial feature set to obtain a significant feature set;

[0041] Based on each significant feature in the significant feature set and in combination with the result of mapping the patient attributes of the target patient onto the corresponding body position, a body position-feature entry is obtained.

[0042] Preferably, all body position-feature items are divided into a training set and a prediction set, including:

[0043] Based on the questionnaire survey results of each target patient before the single-position acquisition and the motion capture video during the single-position acquisition, the behavior fluctuation coefficient of the corresponding target patient under the corresponding single-position acquisition is determined;

[0044] All behavioral fluctuation coefficients involved are classified according to the body position classification, and the division ratio under the corresponding body position classification is obtained in combination with the set ratio to divide the body position-feature items under the corresponding body position classification to obtain the training set and the prediction set.

[0045] Preferably, determining the behavior fluctuation coefficient of the corresponding target patient under the corresponding single body position acquisition includes:

[0046] Collecting a behavior action set in the motion capture video, wherein the behavior action set includes a plurality of independent behavior actions and an action time point of each independent behavior action;

[0047] Compare and analyze the maximum amplitude posture of each independent action with the heart rate affecting posture set to determine whether it affects the heart rate;

[0048] If an impact occurs, the first impact coefficient is obtained from the attitude-impact comparison table;

[0049] Otherwise, the generation time point of the maximum amplitude action posture is locked, and the remaining action postures at the same time point as the generation time point are searched from the behavior action set;

[0050] If the number of other action postures is 0, the first influence coefficient is determined to be 0;

[0051] Otherwise, based on the coherence state of each remaining action posture and the maximum amplitude action posture, the corresponding remaining action postures and the maximum amplitude action posture are subjected to action coherence drawing to obtain a coherent action posture;

[0052] Match the continuous action posture in the whole body posture database to obtain the second influence coefficient;

[0053] Based on the influence coefficient of each independent action, the temporal connection relationship between different independent actions, and the action integration connection relationship, the initial fluctuation coefficient is obtained;

[0054] The behavioral influencing factors before collection are extracted from the questionnaire survey results, and the initial fluctuation coefficient is optimized to obtain the behavioral fluctuation coefficient.

[0055] Preferably, the machine learning model after learning is optimized based on the prediction set to obtain a prediction model, including:

[0056] Collecting each prediction sample in the prediction set and inputting it into the output result of the learned machine learning model, and constructing an error vector based on the prediction set;

[0057] Locking the first sample corresponding to the error absolute value less than the preset error value from the error vector, and treating the remaining samples as the second sample;

[0058] Analyze the similarity coefficients of the corresponding second sample and each first sample to construct a similarity sequence;

[0059] Classifying the similar sequences into similarity levels, and retaining the first sample with a high level as the third sample;

[0060] At the same time, obtaining a fourth sample corresponding to the highest similarity from the similar sequence;

[0061] Extracting difference features between the fourth sample and the corresponding second sample, and performing feature expansion and inconsistent feature replacement on the corresponding second sample according to the difference features to obtain two corresponding new samples;

[0062] If the third sample is one, then the fourth sample is the third sample;

[0063] If there are multiple third samples, then all the third samples are analyzed with the same sub-feature description to obtain uniform features, and a new sample is constructed based on all the uniform features;

[0064] A new sample is obtained based on each second sample to continue learning the machine learning model.

[0065] Compared with the prior art, the present invention has the following beneficial effects:

[0066] By analyzing the short-term HRV characteristics under changes in body position, and then using machine learning methods to establish a prediction model for patients with myocardial infarction, the auxiliary prediction ability for myocardial infarction is improved, so as to intervene as early as possible and reduce the risk of myocardial infarction.

[0067] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0068] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0070] Figure 1 This is a flow chart of a method for auxiliary prediction of myocardial infarction based on machine learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0071] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0072] The present invention provides a method for assisting prediction of myocardial infarction based on machine learning, such as Figure 1 As shown, including:

[0073] Step 1: Based on the ECG recording and analysis system, the lead II ECG of each target patient in different body positions is collected, and the lead II ECG is preprocessed to obtain the current ECG;

[0074] Step 2: Classify the body positions of all current ECGs of all target patients, extract the HRV features under each body position classification, and construct the body position-feature entries;

[0075] Step 3: Divide all body position-feature items into a training set and a prediction set, learn the machine learning model based on the training set, and optimize the learned machine learning model based on the prediction set to obtain a prediction model;

[0076] Step 4: Collect the II lead electrocardiogram of the new patient and input it into the prediction model for auxiliary prediction and output of myocardial infarction.

[0077] In this embodiment, the working principle of the ECG recording and analysis system is: based on the ECG electrode, weak physiological electrical signals are extracted from the human body, sent to the preamplifier and filtering circuit through the lead input network and the lead selector, the AC interference is suppressed by the high-frequency filtering circuit and the low-noise differential amplifier circuit, and then converted into a digital synchronous ECG signal by A / D and sent to the controller part, and then sent to the recorder and display after digital processing and filtering conversion to obtain the II lead ECG.

[0078] In this embodiment, the preprocessing refers to performing noise processing on the electrocardiogram.

[0079] In this embodiment, the body position refers to the standing position for spontaneous breathing, the lying position, and the supine position for controlling the respiratory rate.

[0080] In this embodiment, the acquisition time of the II lead electrocardiogram is generally 15 minutes.

[0081] In this embodiment, the body position classification can be the classification of a single body position, the classification of any two body positions, or the classification of three body positions, and the comprehensive electrocardiogram of multiple body positions can more comprehensively reflect the patient's heart health status in different states, thereby improving the accuracy of the prediction.

[0082] In this embodiment, HRV features include time domain features, frequency features and nonlinear features. The time domain features of HRV are mainly divided into statistical features, long-term features and short-term features. Minimum, range, median, mean, standard deviation (SDNN), skewness (third momentum) and kurtosis (fourth momentum) are the main statistical features extracted from the RR interval. The standard deviation of the continuous difference between adjacent RR intervals (SDSD) is the main long-term extraction feature of HRV. The short-term features of HRV include the number of RR intervals that differ by more than 50ms (NN50), the number of RR intervals that differ by more than 20ms (NN20), the percentage of continuous RR differences greater than 50ms (pNN50), and the percentage of continuous RR differences greater than 20ms (pNN20). In addition, the root mean square of continuous differences between normal heartbeats (rMSSD) is extracted from the HRV signal. A total of 16 features are extracted from each electrocardiogram segment in the time domain, and a geometric domain feature triangular_index is included.

[0083] HRV has three main frequency bands. The very low frequency (VLF) band, i.e., frequencies between 0.0033 and 0.04 Hz, exhibits temperature regulation mechanisms. The low frequency (LF) band, i.e., frequencies between 0.04 and 0.15 Hz, exhibits mainly sympathetic nerve activity, and the high frequency (HF) band, i.e., frequencies between 0.15 and 0.4 Hz, exhibits parasympathetic nerve activity.

[0084] Total power, VLF power, LF power, HF power, LF to HF ratio, normalized HF power (HFnu), and normalized LF power (LFnu) were extracted. In total, seven features were extracted from each ECG segment in the frequency domain.

[0085] The nonlinear features are mainly used to model and analyze the ECG patterns before, during, and after sleep apnea. The cardiovascular index (CVI), cardiac sympathetic index (CSI), and modified CVI signals are extracted from the ECG. The approximate entropy of the complexity of the HRV time series, the sample entropy, are also calculated.

[12] . A total of eight nonlinear HRV features were extracted. A total of 32 HRV features were extracted.

[0086] In this embodiment, the body position-feature entry refers to the body positions existing under different body position categories and the HRV features matching the existing body positions.

[0087] In this embodiment, the training set and the prediction set are for learning and predicting optimization of the learning model, and the training set and the prediction set are obtained by dividing all body position-feature items, and the divided set contains the II lead electrocardiogram corresponding to the body position-feature item.

[0088] The beneficial effect of the above technical solution is: by analyzing the short-term HRV characteristics under changes in body position, and then using machine learning methods to establish a prediction model for myocardial infarction patients, the auxiliary prediction ability for myocardial infarction is improved, and then the application of single-lead electrocardiogram in out-of-hospital diagnosis is promoted.

[0089] The present invention provides a method for assisting prediction of myocardial infarction based on machine learning, which preprocesses the II lead electrocardiogram to obtain the current electrocardiogram, including:

[0090] Inputting the II lead ECG into the noise analysis model, several different types of noise calibration bands are obtained;

[0091] Determine the initial proportion of each type of noise calibration band based on lead II ECG;

[0092] Determine the acquisition components involved in the acquisition process of the ECG recording and analysis system, and retrieve the possible noise types of the relevant acquisition components from the component-noise comparison table;

[0093] Performing component testing on each acquisition component in the electrocardiogram recording and analysis system to obtain a lateral vibration discrete spectrum and a longitudinal vibration discrete spectrum of the corresponding acquisition component;

[0094] Analyze a first correlation coefficient between the lateral vibration discrete spectrum and a standard lateral spectrum and a second correlation coefficient between the longitudinal vibration discrete spectrum and a standard longitudinal spectrum;

[0095] Determine a noise compensation coefficient of a corresponding acquisition component based on the first correlation coefficient and the second correlation coefficient;

[0096]

[0097] Wherein, d1 is the corresponding first correlation coefficient; d2 is the corresponding second correlation coefficient; is the signal variance based on the lateral vibration discrete spectrum and the standard lateral spectrum; is the signal variance based on the longitudinal vibration discrete spectrum and the standard longitudinal spectrum; is the maximum value symbol; is the number of discrete signals existing in the discrete spectrum of lateral vibration; is the number of discrete signals in the longitudinal vibration discrete spectrum; e is the exponential function symbol; is the corresponding noise compensation coefficient;

[0098] Assigning noise weights to the noise types of the noise calibration band based on the component weights of each acquisition component in the electrocardiogram recording and analysis system, possible noise types, and noise compensation coefficients;

[0099]

[0100] in, represents the noise weight assigned to the noise type of the qth noise calibration band; Indicates the number of acquisition components that are consistent with the noise type of the qth noise calibration band; Indicates the component weight of the i1th collection component that has a consistent represents a function of a variable; Indicates the noise compensation coefficient of the i1th acquisition component that has a consistent existence; Indicates all The variance of Indicates the noise setting weight for the i1th acquisition component with the same noise type as the qth noise calibration band; It indicates the probability of noise generation with the same noise type as the qth noise calibration band when there is a consistent acquisition component in the i1th; Indicates all The variance of; a1 represents the set probability; It represents the sum of the noise compensation coefficients involved in all acquisition components in the ECG recording and analysis system;

[0101] According to the noise weight corresponding to each noise calibration band and the initial proportion, the elimination coefficient of the corresponding noise calibration band is obtained, and each elimination coefficient is compared with the preset coefficient respectively;

[0102] If all the elimination coefficients are smaller than the preset coefficients, all the elimination coefficients are sorted in order of size, and the noise calibration band corresponding to the elimination coefficient ranked first is locked and regarded as the first calibration band;

[0103] Otherwise, the noise calibration band that is greater than or equal to the preset coefficient is regarded as the first calibration band;

[0104] A noise elimination method consistent with the noise type of the first calibration band is selected from the noise elimination database, and noise elimination is performed on the lead II electrocardiogram to obtain a current electrocardiogram.

[0105] Preferably, obtaining the elimination coefficient corresponding to the noise calibration band includes:

[0106]

[0107] in, represents the elimination coefficient of the qth noise calibration band; represents the initial proportion of the qth noise calibration band; represents the energy proportion of the qth noise calibration band; Indicates the total number of noise calibration bands; Indicates the horizontal width value of the qth noise calibration band; Indicates the horizontal width value of the lead II electrocardiogram; represents the energy value of the qth noise calibration band; Represents the total energy value of the lead II electrocardiogram.

[0108] In this embodiment, the noise analysis model is pre-trained, and the neural network model is trained based on the results of sample analysis of different electrocardiograms and the noise existing in the electrocardiogram by electrocardiogram experts. Therefore, after analyzing the lead II electrocardiogram, the locations where different noise types exist can be directly locked, and the electrocardiogram covered by the same noise type can be regarded as the noise calibration band of the noise type.

[0109] In this embodiment, the initial proportion = the horizontal length of the corresponding noise calibration band / the horizontal length of the II lead electrocardiogram.

[0110] In this embodiment, the acquisition components are the building blocks present in the system, such as ECG electrodes, preamplifiers, filter circuits, high-frequency filter circuits, low-noise differential amplifier circuits, A / D converters, controllers, and the like.

[0111] In this embodiment, the component-noise comparison table includes different collection components and the noise types that match the components. That is, the system will be tested before leaving the factory for use, and the collection component 1 may generate noise 1 and noise 2, and the collection component 2 may generate noise 2, and so on. The possible noise type corresponding to each collection component can be directly determined.

[0112] In this embodiment, during the process of testing the component, since the component itself generates vibration information, vibrations in two directions are captured, one in the lateral direction and the other in the longitudinal direction.

[0113] In this embodiment, the vibration discrete spectrum may be the lateral vibration amplitude and the longitudinal vibration amplitude collected at an interval of 0.5 seconds, thereby forming a lateral vibration discrete spectrum and a longitudinal vibration discrete spectrum.

[0114] In this embodiment, the standard transverse spectrum and the standard longitudinal spectrum are pre-set for comparative analysis, and the number of collection points is consistent with the collected transverse vibration discrete spectrum and longitudinal vibration discrete spectrum, and alignment analysis is sufficient.

[0115] In this embodiment, the correlation coefficient=1-the average value of the absolute values ​​of the vibration amplitude differences of all the acquisition points / the average value of the amplitude values ​​of the corresponding standard spectrum.

[0116] In this embodiment, the signal variance is the variance of the absolute values ​​of the amplitude differences at all acquisition points.

[0117] In this embodiment, the sum of the weights of all acquisition components in the electrocardiogram recording and analysis system is 1.

[0118] In this embodiment, the value of the preset coefficient is 0.6.

[0119] In this embodiment, the noise elimination database includes different noise elimination methods, which belong to the existing elimination technology.

[0120] The beneficial effect of the above technical solution is: by analyzing the electrocardiogram through the noise analysis model, it is convenient to obtain different types of noise calibration bands, and by retrieving the possible noise types through the reference table and combining the noise compensation coefficient calculated by the initial proportion and the correlation coefficient, the elimination coefficient is effectively obtained, which facilitates noise elimination, that is, it can ensure the elimination of the noise that needs to be eliminated the most, and ensure the accuracy of the captured electrocardiogram as much as possible, providing reliable samples for subsequent model learning.

[0121] The present invention provides a method for assisting prediction of myocardial infarction based on machine learning, which classifies the body positions of all current electrocardiograms of all target patients, including:

[0122] The body position descriptions corresponding to the single body position and the combined body position are obtained, and the electrocardiograms that are consistent with the corresponding body position descriptions are selected from all current electrocardiograms as the body position classification results.

[0123] In this embodiment, the body position representation is: standing position with spontaneous breathing, lying position and supine position with controlled respiratory rate, mainly for classifying the electrocardiogram to obtain the features under different body position combinations.

[0124] The beneficial effect of the above technical solution is: by performing single and combined posture classification, it effectively ensures that samples under different combinations of situations participate in model learning, thereby ensuring the comprehensiveness of model learning.

[0125] The present invention provides a myocardial infarction auxiliary prediction method based on machine learning, extracting HRV features under each body position classification, and constructing body position-feature entries, including:

[0126] Extract the time domain features, frequency domain features and nonlinear features of each electrocardiogram under the same individual position classification to obtain the initial feature set;

[0127] Performing a t-test on the initial feature set to obtain a significant feature set;

[0128] Based on each significant feature in the significant feature set and in combination with the result of mapping the patient attributes of the target patient onto the corresponding body position, a body position-feature entry is obtained.

[0129] In this embodiment, a t-test was performed on 32 HRV features of myocardial infarction patients and healthy subjects, and 15 statistically significant HRV features were screened out. These significant features can provide important references for subsequent myocardial infarction detection and diagnosis.

[0130] In this embodiment, the initial feature set includes 32 HRV features, the significant feature set includes 15 HRV features, and the patient attributes refer to health attributes and myocardial infarction attributes.

[0131] The beneficial effects of the above technical solution are: extracting the initial feature set of the electrocardiogram, and combining it with the t-test to effectively obtain the significant feature set, making it convenient to obtain the body position-feature items, and ensuring the accuracy and effectiveness of sample construction.

[0132] The present invention provides a myocardial infarction auxiliary prediction method based on machine learning, which divides all body position-feature items into a training set and a prediction set, including:

[0133] Based on the questionnaire survey results of each target patient before the single-position acquisition and the motion capture video during the single-position acquisition, the behavior fluctuation coefficient of the corresponding target patient under the corresponding single-position acquisition is determined;

[0134] All behavioral fluctuation coefficients involved are classified according to the body position classification, and the division ratio under the corresponding body position classification is obtained in combination with the set ratio to divide the body position-feature items under the corresponding body position classification to obtain the training set and the prediction set.

[0135] In this embodiment, the questionnaire survey results are based on questionnaires sent to patients, including questionnaires on fever, asthma, infection, abnormal thyroid function test, electrolyte imbalance, shock, hypertensive emergency, tumor, stroke, mental illness, taking specific medications and foods, no excessive activity in the 24 hours before data collection, mood swings, smoking, and drinking caffeinated, alcohol or tea.

[0136] In this embodiment, the motion capture video refers to a video of the patient's behavior and posture captured during the process of collecting data from the patient.

[0137] In this embodiment, the behavior fluctuation coefficient is determined in order to adjust the set ratio, and the set ratio is 8:2.

[0138] In this embodiment, the division ratio .

[0139] The beneficial effect of the above technical solution is: to comprehensively determine the behavior fluctuation coefficient from the questionnaire survey results and the motion capture video, and then to achieve reasonable adjustment of the set ratio, to ensure the rationality of the division of the training set and the prediction set, and then to ensure the effectiveness of the machine learning model learning.

[0140] The present invention provides a myocardial infarction auxiliary prediction method based on machine learning, which determines the behavior fluctuation coefficient of a corresponding target patient under a corresponding single body position acquisition, including:

[0141] Collecting a behavior action set in the motion capture video, wherein the behavior action set includes a plurality of independent behavior actions and an action time point of each independent behavior action;

[0142] Compare and analyze the maximum amplitude posture of each independent action with the heart rate affecting posture set to determine whether it affects the heart rate;

[0143] If an impact occurs, the first impact coefficient is obtained from the attitude-impact comparison table;

[0144] Otherwise, the generation time point of the maximum amplitude action posture is locked, and the remaining action postures at the same time point as the generation time point are searched from the behavior action set;

[0145] If the number of other action postures is 0, the first influence coefficient is determined to be 0;

[0146] Otherwise, based on the coherence state of each remaining action posture and the maximum amplitude action posture, the corresponding remaining action postures and the maximum amplitude action posture are subjected to action coherence drawing to obtain a coherent action posture;

[0147] Match the continuous action posture in the whole body posture database to obtain the second influence coefficient;

[0148] Based on the influence coefficient of each independent action, the temporal connection relationship between different independent actions, and the action integration connection relationship, the initial fluctuation coefficient is obtained;

[0149] The behavioral influencing factors before collection are extracted from the questionnaire survey results, and the initial fluctuation coefficient is optimized to obtain the behavioral fluctuation coefficient.

[0150] In this embodiment, an independent action refers to a continuous action performed by the user, and the setting standard of the independent action is set in advance. As long as the standard is met, it can be regarded as an independent action. For example, the user's arm from the beginning of stretching to the end of stretching and falling is an independent action.

[0151] In this embodiment, the action time point of an independent behavior action includes several continuous time points. For example, the time point when the arm starts to extend is t1, and the time point when the arm ends to fall is t6. At this time, the continuous time points involved in t1~t6 are the action time points corresponding to the independent behavior action.

[0152] In this embodiment, the maximum amplitude action posture refers to the action profile of the independent behavior action at each time point, and the posture with the largest action profile is selected as the maximum amplitude action posture.

[0153] In this embodiment, the heart rate influencing posture set includes a plurality of pre-set motion postures, and based on the heart rate influencing coefficient of each motion posture, under normal circumstances, the larger the amplitude, the greater the corresponding heart rate impact.

[0154] In this embodiment, the purpose of the comparative analysis is to find the most matching posture to determine whether there is an influence, so as to facilitate the subsequent determination of the influence coefficient.

[0155] In this embodiment, the posture-influence comparison table includes different action postures and heart rate influence coefficients matching the postures, and the value range of the first influence coefficient is 0 to 1.

[0156] In this embodiment, for the independent action of the user's arm from the beginning of stretching to the end of stretching and falling, the time point at which the corresponding maximum amplitude action posture under t1~t6 is generated is, for example, time t3. At this time, the coherent state refers to obtaining the amplitude of change of other parts except the arm at time t3, and then outlining the posture to obtain a coherent action posture.

[0157] In this embodiment, the whole body posture database includes multiple whole body postures of all parts of the body and the influence coefficient of each whole body posture on the heart rate, which are all pre-set. Therefore, by matching the continuous action posture with the whole body posture database, the influence coefficient of the whole body posture with the highest matching degree is extracted and regarded as the second influence coefficient.

[0158] In this embodiment, the initial fluctuation coefficient=the average value of the influence coefficient of each independent behavior action×ln(2+the relation variance of all time connection relationships+the relation variance of all action blending connection relationships).

[0159] In this embodiment, the behavior fluctuation coefficient=initial fluctuation coefficient×(1+behavior influence factor).

[0160] In this embodiment, the relationship variance of all time connection relationships=the number of intersection time points under all time connection relationships / the variance of the number of union time points under the time connection relationships.

[0161] The relation variance of all action blending relationships = the variance of the intersection action contour length under all action blending relationships / the variance of the union action contour length under all action blending relationships.

[0162] In this embodiment, the behavior influence factor is obtained by accumulating all factors retrieved from the questionnaire survey results and multiplying them by the weights of the corresponding factors, and the value range of the behavior influence factor is 0 to 0.3.

[0163] The beneficial effects of the above technical solution are: capturing behavioral actions from the video to obtain the maximum amplitude movement posture and comparing and analyzing it with the heart rate influencing posture set, and then determining the coefficient under the influencing condition and the influence coefficient under the remaining movement posture conditions, to achieve reliable calculation of the initial fluctuation coefficient, and subsequently obtaining the behavioral influencing factor from the questionnaire survey results to further optimize the coefficient to obtain the behavioral fluctuation coefficient, providing a reliable basis for proportion adjustment.

[0164] The present invention provides a method for assisting prediction of myocardial infarction based on machine learning, which optimizes the learned machine learning model based on a prediction set to obtain a prediction model, including:

[0165] Collecting each prediction sample in the prediction set and inputting it into the output result of the learned machine learning model, and constructing an error vector based on the prediction set;

[0166] Locking the first sample corresponding to the error absolute value less than the preset error value from the error vector, and treating the remaining samples as the second sample;

[0167] Analyze the similarity coefficients of the corresponding second sample and each first sample to construct a similarity sequence;

[0168] Classifying the similar sequences into similarity levels, and retaining the first sample with a high level as the third sample;

[0169] At the same time, obtaining a fourth sample corresponding to the highest similarity from the similar sequence;

[0170] Extracting difference features between the fourth sample and the corresponding second sample, and performing feature expansion and inconsistent feature replacement on the corresponding second sample according to the difference features to obtain two corresponding new samples;

[0171] If the third sample is one, then the fourth sample is the third sample;

[0172] If there are multiple third samples, then all the third samples are analyzed with the same sub-feature description to obtain uniform features, and a new sample is constructed based on all the uniform features;

[0173] A new sample is obtained based on each second sample to continue learning the machine learning model.

[0174] In this embodiment, the output result is the output predicted HRV feature, and is represented by a specific numerical value.

[0175] In this embodiment, the error vector={the absolute value of the numerical difference between each actual HRV feature and the corresponding predicted HRV feature}, and the absolute value of the numerical difference is the absolute value of the error.

[0176] In this embodiment, the preset error value is 0.1.

[0177] In this embodiment, for example, the error vector={u1, u2, u3, u4}, wherein the first sample is the sample corresponding to u1 and u2, and the second sample is the sample corresponding to u3 and u4.

[0178] In this embodiment, the similarity coefficient is calculated based on the similarity function.

[0179] In this embodiment, the similarity sequence={the similarity coefficient between the second sample and the sample corresponding to u1 corresponds to the similarity coefficient between the second sample and the sample corresponding to u2}.

[0180] In this embodiment, the similarity level classification refers to classifying similarity coefficients between 0.7 and 1 as high levels, and classifying similarity coefficients between 0 and 0.7 as low levels.

[0181] In this embodiment, the difference feature refers to a feature corresponding to a different value.

[0182] In this embodiment, there are two new samples: the second sample and the difference feature, and the feature after replacement.

[0183] For example, the difference feature is u01, the second sample is y1, and the feature in the second sample that is different from the fourth sample is u02 (an inconsistent feature). In this case, u02 is replaced with u01.

[0184] In this embodiment, the uniform features refer to the extraction of sub-features with the same number of values ​​and the largest number from the same sub-feature description, and the extraction process is regarded as an analysis process.

[0185] In this embodiment, the new sample is a uniform feature extracted based on each sub-feature description.

[0186] In this embodiment, the machine model is implemented using a multi-model fusion method, combining traditional logistic regression, support vector machine (SVM) and deep learning model to give full play to the advantages of different models and improve prediction accuracy and robustness.

[0187] The beneficial effect of the above technical solution is: by acquiring new samples to continue learning the model, thereby ensuring the prediction accuracy of the model.

[0188] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for assisting prediction of myocardial infarction based on machine learning, characterized in that: include: Step 1: Based on the ECG recording and analysis system, the lead II ECG of each target patient in different body positions is collected, and the lead II ECG is preprocessed to obtain the current ECG; Step 2: Classify the body positions of all current ECGs of all target patients, extract the HRV features under each body position classification, and construct the body position-feature entries; Step 3: Divide all body position-feature items into a training set and a prediction set, learn the machine learning model based on the training set, and optimize the learned machine learning model based on the prediction set to obtain a prediction model; Step 4: Collect the II lead electrocardiogram of the new patient and input it into the prediction model for auxiliary prediction and output of myocardial infarction; Among them, the machine learning model after learning is optimized based on the prediction set to obtain a prediction model, including: Collecting each prediction sample in the prediction set and inputting it into the output result of the learned machine learning model, and constructing an error vector based on the prediction set; Locking the first sample corresponding to the error absolute value less than the preset error value from the error vector, and treating the remaining samples as the second sample; Analyze the similarity coefficients of the corresponding second sample and each first sample to construct a similarity sequence; Classifying the similar sequences into similarity levels, and retaining the first sample with a high level as the third sample; At the same time, obtaining a fourth sample corresponding to the highest similarity from the similar sequence; Extracting difference features between the fourth sample and the corresponding second sample, and performing feature expansion and inconsistent feature replacement on the corresponding second sample according to the difference features to obtain two corresponding new samples; If the third sample is one, then the fourth sample is the third sample; If there are multiple third samples, then all the third samples are analyzed with the same sub-feature description to obtain uniform features, and a new sample is constructed based on all the uniform features; A new sample is obtained based on each second sample to continue learning the machine learning model.

2. The method for assisting prediction of myocardial infarction based on machine learning according to claim 1, characterized in that: Preprocessing the lead II electrocardiogram to obtain the current electrocardiogram includes: Inputting the II lead ECG into the noise analysis model, several different types of noise calibration bands are obtained; Determine the initial proportion of each type of noise calibration band based on the lead II electrocardiogram; Determine the acquisition components involved in the acquisition process of the ECG recording and analysis system, and retrieve the possible noise types of the relevant acquisition components from the component-noise comparison table; Performing component testing on each acquisition component in the electrocardiogram recording and analysis system to obtain a lateral vibration discrete spectrum and a longitudinal vibration discrete spectrum of the corresponding acquisition component; Analyze a first correlation coefficient between the lateral vibration discrete spectrum and a standard lateral spectrum and a second correlation coefficient between the longitudinal vibration discrete spectrum and a standard longitudinal spectrum; Determine a noise compensation coefficient of a corresponding acquisition component based on the first correlation coefficient and the second correlation coefficient; ； Wherein, d1 is the corresponding first correlation coefficient; d2 is the corresponding second correlation coefficient; is the signal variance based on the discrete spectrum of transverse vibration and the standard transverse spectrum; is the signal variance based on the longitudinal vibration discrete spectrum and the standard longitudinal spectrum; max is the maximum value symbol; is the number of discrete signals existing in the discrete spectrum of lateral vibration; is the number of discrete signals in the longitudinal vibration discrete spectrum; e is the exponential function symbol; is the corresponding noise compensation coefficient; Assigning noise weights to the noise types of the noise calibration band based on the component weights of each acquisition component in the electrocardiogram recording and analysis system, possible noise types, and noise compensation coefficients; ； ； in, represents the noise weight assigned to the noise type of the qth noise calibration band; Indicates the number of acquisition components that are consistent with the noise type of the qth noise calibration band; represents the component weight of the i1th consistent acquisition component; represents a function of a variable; Indicates the noise compensation coefficient of the i1th acquisition component that has a consistent existence; Indicates all The variance of Indicates the noise setting weight for the i1th acquisition component with the same noise type as the qth noise calibration band; It indicates the probability of noise generation with the same noise type as the qth noise calibration band when there is a consistent acquisition component in the i1th; Indicates all The variance of; a1 represents the set probability; It represents the sum of the noise compensation coefficients involved in all acquisition components in the ECG recording and analysis system; According to the noise weight corresponding to each noise calibration band and the initial proportion, the elimination coefficient of the corresponding noise calibration band is obtained, and each elimination coefficient is compared with the preset coefficient respectively; If all the elimination coefficients are smaller than the preset coefficients, all the elimination coefficients are sorted in order of size, and the noise calibration band corresponding to the elimination coefficient ranked first is locked and regarded as the first calibration band; Otherwise, the noise calibration band that is greater than or equal to the preset coefficient is regarded as the first calibration band; A noise elimination method consistent with the noise type of the first calibration band is selected from the noise elimination database, and noise elimination is performed on the lead II electrocardiogram to obtain a current electrocardiogram.

3. The method for assisting prediction of myocardial infarction based on machine learning according to claim 2, characterized in that: Obtain the elimination coefficient of the corresponding noise calibration band, including: ； in, represents the elimination coefficient of the qth noise calibration band; represents the initial proportion of the qth noise calibration band; represents the energy proportion of the qth noise calibration band; Indicates the total number of noise calibration bands; Indicates the horizontal width value of the qth noise calibration band; Indicates the horizontal width value of the lead II electrocardiogram; represents the energy value of the qth noise calibration band; Represents the total energy value of the lead II electrocardiogram.

4. The method for assisting prediction of myocardial infarction based on machine learning according to claim 1, characterized in that: Body position classification for all current ECGs of all target patients, including: The body position descriptions corresponding to the single body position and the combined body position are obtained, and the electrocardiograms that are consistent with the corresponding body position descriptions are selected from all current electrocardiograms as the body position classification results.

5. The method for assisting prediction of myocardial infarction based on machine learning according to claim 4, characterized in that: Extract the HRV features under each body position classification and construct the body position-feature entries, including: Extract the time domain features, frequency domain features and nonlinear features of each electrocardiogram under the same individual position classification to obtain the initial feature set; Performing a t-test on the initial feature set to obtain a significant feature set; Based on each significant feature in the significant feature set and in combination with the result of mapping the patient attributes of the target patient onto the corresponding body position, a body position-feature entry is obtained.

6. The method for assisting prediction of myocardial infarction based on machine learning according to claim 1, characterized in that: All body position-feature items are divided into training set and prediction set, including: Based on the questionnaire survey results of each target patient before the single-position acquisition and the motion capture video during the single-position acquisition, the behavior fluctuation coefficient of the corresponding target patient under the corresponding single-position acquisition is determined; All behavioral fluctuation coefficients involved are classified according to the body position classification, and the division ratio under the corresponding body position classification is obtained in combination with the set ratio to divide the body position-feature items under the corresponding body position classification to obtain the training set and the prediction set.

7. The method for assisting prediction of myocardial infarction based on machine learning according to claim 6, characterized in that: Determine the behavioral fluctuation coefficient of the corresponding target patient under the corresponding single body position collection, including: Collecting a behavior action set in the motion capture video, wherein the behavior action set includes a plurality of independent behavior actions and an action time point of each independent behavior action; Compare and analyze the maximum amplitude posture of each independent action with the heart rate affecting posture set to determine whether it affects the heart rate; If an impact occurs, the first impact coefficient is obtained from the attitude-impact comparison table; Otherwise, the generation time point of the maximum amplitude action posture is locked, and the remaining action postures at the same time point as the generation time point are searched from the behavior action set; If the number of other action postures is 0, the first influence coefficient is determined to be 0; Otherwise, based on the coherence state of each remaining action posture and the maximum amplitude action posture, the corresponding remaining action postures and the maximum amplitude action posture are subjected to action coherence drawing to obtain a coherent action posture; Match the continuous action posture in the whole body posture database to obtain the second influence coefficient; Based on the influence coefficient of each independent action, the temporal connection relationship between different independent actions, and the action integration connection relationship, the initial fluctuation coefficient is obtained; The behavioral influencing factors before collection are extracted from the questionnaire survey results, and the initial fluctuation coefficient is optimized to obtain the behavioral fluctuation coefficient.

Images