CN111261289A

CN111261289A - Heart disease detection method based on artificial intelligence model

Info

Publication number: CN111261289A
Application number: CN201811448564.6A
Authority: CN
Inventors: 王云霞; 何毅钒
Original assignee: Shanghai Turing Medical Technology Co ltd
Current assignee: Shanghai Turing Medical Technology Co ltd
Priority date: 2018-11-30
Filing date: 2018-11-30
Publication date: 2020-06-09

Abstract

The invention provides a heart disease detection method based on an artificial intelligence model, which comprises the steps of constructing a deep learning identification model of an electrocardiogram vector diagram and an electrocardiogram nonlinear system dynamic diagram, constructing a machine learning classification model of quantitative data, biochemical data, human physiological information data and clinical information data of heart dynamic pathological characteristics and the like, assigning corresponding weight values to heart disease results obtained by identifying and classifying different models, and obtaining a comprehensive judgment result of heart disease detection. The detection method of the heart disease provided by the invention solves the technical problems of a model processing method of dynamic signals with continuous heart electrical activity, the modeling analysis of multi-pathology characteristic quantitative data, the fusion judgment of the same disease under different models and the like. The heart disease detection method provided by the invention improves the accuracy and detection efficiency of heart disease detection, and the diagnosis effect is continuously improved along with the increase of index data of different types of pathological features expanded into a database.

Description

Heart disease detection method based on artificial intelligence model

Technical Field

The invention relates to the field of detection of heart diseases, in particular to a heart disease detection method based on an artificial intelligence model.

Background

The heart disease is a common circulatory system disease, comprises rheumatic heart disease, hypertensive heart disease, myocarditis, coronary heart disease and the like, is a disease with high mortality rate, and brings huge economic burden and life disaster to families of patients. The world health organization states: among the ten diseases most likely to cause human death, heart disease is the first. However, if the patient with heart disease can be diagnosed early, different patients can be treated as soon as possible by effective and necessary treatment means, so that the catastrophic consequences caused by the sudden heart disease can be avoided; therefore, the acquisition of index data of multiple pathological characteristics of the person to be detected is carried out in a targeted manner, the early intervention and treatment of the heart disease are effectively carried out, and the method has important significance for the detection and treatment of the current heart disease.

In order to acquire index data of multiple pathological characteristics of a person to be tested, the prior art also has the following problems to be solved: (1) in the aspect of mining and utilizing index data information of multiple pathological characteristics of a person to be tested, continuous heart dynamic signals are still difficult to process and analyze based on an artificial intelligence model; (2) in the analysis process of the dynamic pathological features of the heart electrical activity of the person to be tested, the mined dynamic pathological information and static pathological information are too limited to comprehensively and objectively reflect the complex heart electrical activity process, and are not beneficial to early intervention or further accurate treatment of the heart disease of the person to be tested in the later stage of a doctor.

Therefore, there is a need to provide an improved technical solution to overcome the technical problems in the prior art.

Disclosure of Invention

In order to achieve the aforementioned object, an aspect of the present invention provides a method for detecting a heart disease based on an artificial intelligence model, comprising the following steps:

s1, acquiring index data of multiple pathological features of a specific heart disease, wherein the index data of the multiple pathological features of the specific heart disease comprises one or more of quantitative data of dynamic pathological features of the heart, index data of image features of an electrocardiograph vector image and dynamic data of an electrocardiograph nonlinear system;

and S2, performing machine learning on the index data of the multiple pathological features of the specific heart disease acquired in the step S1 to acquire a machine learning judgment model of the specific heart disease.

In at least one embodiment, in the method for detecting a heart disease based on an artificial intelligence model as described above, the step S1 includes: one or more of quantitative index data of geometric characteristics, quantitative index data of nonlinear dynamic characteristics, quantitative index data of model characteristics, quantitative index data of time domain characteristics and quantitative index data of frequency domain characteristics; preferably, the index data of multiple pathological features of the specific heart disease in step S1 further includes one or more of quantitative index data of auxiliary pathological features, physiological information data of human body, and clinical information data; more preferably, the quantitative index data of the auxiliary pathological features comprise quantitative data of ECG morphological indexes, and/or index data of electrocardiogram image features, and/or biochemical data.

In at least one embodiment, in the method for detecting heart diseases based on artificial intelligence model as described above, the machine learning decision model for specific heart diseases in step S2 includes at least one of support vector machine, convolutional neural network, recurrent neural network, bayesian classifier, K nearest neighbor algorithm, K-means algorithm, linear regression, logistic regression, multiple nonlinear regression fitting method, Adaboost algorithm, hidden markov model, extreme learning machine, random forest algorithm, decision tree algorithm, clustering algorithm, generative countermeasure network, stacked autoencoder, full connection network, unsupervised pre-training network, deep belief network, deep bolman machine, and neural tensor network;

in at least one embodiment, in the method for detecting a heart disease based on an artificial intelligence model as described above, the machine learning determination model for a specific heart disease in step S2 includes a machine learning classification model and a deep learning identification model; the data input into the machine learning classification model comprises quantitative data of the dynamic pathological features of the heart, and the data input into the deep learning identification model comprises index data of the image features of the vector cardiogram and dynamic data of a nonlinear electrocardiogram system; preferably, the data input into the machine learning classification model further includes one or more of quantitative index data of auxiliary pathological features, human physiological information data and clinical information data; more preferably, the machine learning classification model is selected from at least one of a support vector machine, a bayesian classifier, a K-nearest neighbor algorithm, a K-means algorithm, a linear regression, a logistic regression, a multivariate nonlinear regression fitting method, an Adaboost algorithm, a hidden markov model, an extreme learning machine, a random forest algorithm, a decision tree algorithm, and a clustering algorithm, and the deep learning recognition model is selected from at least one of a convolutional neural network, a recurrent neural network, a generative confrontation network, a stacked autoencoder, a fully-connected network, an unsupervised pre-training network, a deep belief network, a deep boltzmann machine, and a neural tensor network.

In at least one embodiment, in the method for detecting heart disease based on artificial intelligence model as described above, the method further includes step S3: acquiring index data of multiple pathological features of a person to be detected, and inputting the index data into the machine learning judgment model of the specific heart disease in the step S2 to obtain a detection result of the specific heart disease; the index data of the multi-pathological characteristics of the person to be tested comprises one or more of quantitative data of dynamic pathological characteristics of the heart of the person to be tested, index data of image characteristics of an electrocardiograph vector diagram and dynamic data of an electrocardiograph nonlinear system.

In at least one embodiment, in the method for detecting a cardiac disease based on an artificial intelligence model as described above, the index data of the multiple pathological characteristics of the person to be detected in step S3 further includes one or more of quantitative index data of auxiliary pathological characteristics of the person to be detected, physiological information data of a human body, and clinical information data; preferably, the quantitative index data of the auxiliary pathological features of the person to be tested in step S3 includes quantitative data of ECG morphology indexes and/or biochemical data; preferably, the detection result of the specific heart disease in step S3 further includes using the index data of the multiple pathological features of the person to be tested to respectively perform specific heart disease determination, so as to obtain an index determination result of the heart disease of the person to be tested; more preferably, the method for detecting a heart disease further includes assigning a weight value to the output result of the machine learning determination model for the specific heart disease and the index determination result of the heart disease of the person to be detected in step S2 to perform heart disease detection, so as to obtain a comprehensive determination result of the heart disease of the person to be detected; the index judgment result of the heart disease of the person to be detected comprises one or more of judgment information of the specific heart disease of the quantitative data of the dynamic pathological features of the heart of the person to be detected, judgment information of the specific heart disease of the quantitative index data of the auxiliary pathological features of the person to be detected, judgment information of the specific heart disease of the human physiological information data of the person to be detected and judgment information of the specific heart disease of the clinical information data of the person to be detected; more preferably, the output result of the machine learning judgment model for the specific heart disease in step S2 further includes an output result of the machine learning classification model and an output result of the deep learning identification model, and the output result of the machine learning classification model and the output result of the deep learning identification model are assigned with weight values.

In at least one embodiment, in the method for detecting a heart disease based on an artificial intelligence model as described above, the step S3 includes: the method comprises the following steps of obtaining quantitative index data of one or more of dynamic geometric characteristics, quantitative index data of nonlinear dynamic characteristics, quantitative index data of model characteristics, quantitative index data of time domain characteristics and quantitative index data of frequency domain characteristics of an electrocardio nonlinear system of a person to be tested.

The invention also provides a heart disease detection method based on the artificial intelligence model, which comprises the following steps:

step 1, acquiring index data of multiple pathological characteristics of a person to be detected, wherein the index data of the multiple pathological characteristics of the person to be detected comprises quantitative data of heart dynamic pathological characteristics of the person to be detected, and/or index data of image characteristics of an electrocardiograph of the person to be detected, and/or dynamic data of an electrocardiograph nonlinear system of the person to be detected, and/or quantitative index data of auxiliary pathological characteristics of the person to be detected, and/or human body physiological information data of the person to be detected, and/or clinical information data of the person to be detected;

the quantitative data of the heart dynamic pathological features of the person to be tested comprises: one or more of quantitative index data of dynamic geometric characteristics of an electrocardio nonlinear system of a person to be tested, quantitative index data of nonlinear dynamic characteristics, quantitative index data of model characteristics, quantitative index data of time domain characteristics and quantitative index data of frequency domain characteristics;

step 2, inputting the index data of the multiple pathological features of the person to be tested, which is acquired in the step 1, into a machine learning judgment model of the specific heart disease;

and 3, outputting a specific heart disease detection result, wherein the specific heart disease detection result comprises an output result of the machine learning judgment model of the specific heart disease in the step 2.

In at least one embodiment, in the method for detecting a cardiac disease based on an artificial intelligence model as described above, the detection result of the specific cardiac disease in step 3 further includes threshold determination information of the specific cardiac disease of the quantitative data of the dynamic pathological features of the heart of the person to be detected, and/or threshold determination information of the specific cardiac disease of the quantitative index data of the auxiliary pathological features of the person to be detected, and/or determination information of the specific cardiac disease of the clinical information data of the person to be detected; preferably, the detection result of the specific heart disease in step 3 further includes weighted value that is assigned to the output result information of the machine learning judgment model of the specific heart disease in step 3, and/or the judgment information of the specific heart disease of the quantitative data of the dynamic pathological features of the heart of the person to be measured, and/or the judgment information of the specific heart disease of the quantitative index data of the auxiliary pathological features of the person to be measured, and/or the judgment information of the specific heart disease of the human physiological information data of the person to be measured, and/or the judgment information of the specific heart disease of the clinical information data of the person to be measured.

In still another aspect, the present invention provides a heart disease detection product, which uses the artificial intelligence model-based heart disease detection method as described above.

In a further aspect, the invention provides a use of a product for detecting heart diseases as described above for detecting heart diseases.

The heart disease detection method of the invention has the following technical effects:

1) by adopting the machine learning technology, the automatic extraction and intelligent diagnosis of the features can be completed, the accuracy is higher, and the diagnosis effect can be continuously improved along with the increase of index data of multiple pathological features of the heart in the database.

2) The dynamic pathological information features of the electrocardio nonlinear system are more abundant, the performances of the detection accuracy, the detection efficiency and the like are obviously improved, and the continuous dynamic signals can be processed, so that early intervention and accurate treatment on the heart disease of the person to be detected in the later period of the doctor are facilitated.

3) The diagnosis process does not need the intervention of specialized doctors, and common users can simply and conveniently complete self-diagnosis and self-monitoring, and can also be used as a computer-aided diagnosis system to be deployed in places such as community hospitals or village and town hospitals and the like which lack specialized doctors.

Drawings

The disclosure of the present invention is illustrated with reference to the accompanying drawings. It is to be understood that these drawings are solely for purposes of illustration and are not intended as a definition of the limits of the invention.

FIG. 1 shows a flow chart of the present invention for constructing a machine learning decision model;

FIG. 2 shows a flow diagram of the aided detection of the machine learning decision model of the present invention.

Detailed Description

The experimental methods of the following examples, which are not specified under specific conditions, are generally determined according to national standards. If there is no corresponding national standard, it is carried out according to the usual international standards, to the conventional conditions or to the conditions recommended by the manufacturer.

In the present invention, all embodiments and preferred embodiments mentioned herein may be combined with each other to form a new technical solution, if not specifically stated.

In the present invention, all the technical features mentioned herein and preferred features may be combined with each other to form a new technical solution, if not specifically stated.

In the present invention, unless otherwise specified, the index data of the electrocardiographic image feature mentioned herein includes, but is not limited to, an electrocardiogram, a preprocessed electrocardiogram, a truncated electrocardiogram, a quantified data of ECG morphology index, and electrocardiographic data; pre-processed electrocardiograms include, but are not limited to, compressed electrocardiograms, contrasted electrocardiograms, or amplified electrocardiograms, etc.; the truncated electrocardiogram includes, but is not limited to, a portion of any size and any shape that is present on the electrocardiogram. Electrocardiographic data refers to data that will include, but is not limited to, single lead electrocardiographic data, 12 lead electrocardiographic data, or multi-lead electrocardiographic data.

In the present invention, if not specifically stated, the index data of the image features of the vector cardiogram mentioned herein includes, but is not limited to, the vector cardiogram, the preprocessed vector cardiogram, the intercepted vector cardiogram and the vector cardiogram data; the preprocessed vector cardiograms include, but are not limited to, compressed vector cardiograms, contrast-changed vector cardiograms, amplified vector cardiograms and the like; the truncated vector cardiogram includes, but is not limited to, a portion of any size and any shape that is present on the vector cardiogram. The electrocardiograph vector data refers to a multi-dimensional electrocardiograph data set obtained by converting electrocardiograph data, wherein the conversion includes but is not limited to converting 12-lead or multi-lead electrocardiograph data into three-dimensional data or converting 12-lead or multi-lead electrocardiograph data into multi-dimensional data.

In the present invention, unless otherwise specified, the electrocardiographic nonlinear system dynamic data mentioned herein includes, but is not limited to, an electrocardiographic nonlinear system dynamic graph, an electrocardiographic kinetic graph, a preprocessed electrocardiographic kinetic graph, a truncated electrocardiographic kinetic graph, and electrocardiographic kinetic data; pre-processed electrocardiographic images include, but are not limited to, compressed electrocardiographic images, contrast-modified electrocardiographic images, or magnified electrocardiographic images, among others; the truncated electrocardiographic image includes, but is not limited to, a portion of any size and any shape that is present on the electrocardiographic image. The electrocardiogram dynamic data refers to multidimensional data obtained by intercepting electrocardiogram vector data through waves, bands or intervals and then by a modeling method.

In the present invention, the testees mentioned herein include, but are not limited to, hospital testees, physical examination personnel, heart disease patients, and the like, unless otherwise specified.

In the present invention, unless otherwise specified, the quantitative index data of geometric features, the quantitative index data of nonlinear dynamical features, the quantitative index data of model features, the quantitative index data of time domain features, and the quantitative index data of frequency domain features mentioned herein are all used for reflecting the dynamic pathological features of the heart. Geometric characteristics include, but are not limited to, one or more of trend, slope, direction, shape, circularity, sphericity, uniformity, eccentricity, variability, and angle; the nonlinear dynamics characteristics comprise one or more of entropy, complexity, relevance dimension, Lyapunov index spectrum and maximum Lyapunov index spectrum; the entropy is preferably information entropy, wavelet entropy or approximate entropy; the complexity is preferably C₀Complexity, Kolmogorov complexity, or LZ complexity; model features include, but are not limited to, AR model coefficients, or TVAR model features; the time-frequency features include, but are not limited to, short-time fourier transform features, wavelet transform features, or a combination of both; the frequency domain features include, but are not limited to, fast fourier transform features.

In the present invention, unless otherwise specified, the threshold value determination information of the specific cardiac disease of the quantitative data of the dynamic pathological features of the heart mentioned herein includes one or more of threshold value determination of the specific cardiac disease of the quantitative index data of the geometric features of the electrocardiographic nonlinear system dynamics, threshold value determination of the specific cardiac disease of the quantitative index data of the nonlinear dynamical features, threshold value determination of the specific cardiac disease of the quantitative index data of the model features, threshold value determination of the specific cardiac disease of the quantitative index data of the time-domain features, and threshold value determination of the specific cardiac disease of the quantitative index data of the time-domain features.

In the invention, the machine learning method for constructing the machine learning judgment model adopts one or more methods including but not limited to a support vector machine, a convolutional neural network, a cyclic neural network, a Bayesian classifier, a K neighbor algorithm, a K mean algorithm, a linear regression, a logistic regression, a multivariate nonlinear regression fitting method, an Adaboost algorithm, a hidden Markov model, an extreme learning machine, a random forest algorithm, a decision tree algorithm, a clustering algorithm, a generative countermeasure network, a stacked automatic encoder, a full-connection network, an unsupervised pre-training network, a deep belief network, a deep Boltzmann machine, a neural tensor network and the like.

In the present invention, the method for detecting a heart disease based on an artificial intelligence model mentioned herein can be used in an artificial intelligence device for disease detection, such as a wearable device for heart disease detection, if not specifically mentioned.

The present invention will be further described with reference to the following examples. It should be understood that the following preferred examples are illustrative only and are not intended to limit the scope of the invention.

Example 1 sample set construction and Pre-processing of sample data

1. Creation of sample sets

The method comprises the following steps of constructing sample set data for a machine learning judgment model, wherein the specific construction method comprises the following steps:

(1) composition of the sample set: n clinically known heart healthy individuals (n >500) and m clinically known heart disease individuals (m >1000) were included as a sample population; relevant pathology data of a sample population associated with a particular cardiac disorder is collected as sample set data.

(2) Setting of sample label: index data of gold standard index of heart disease and expert consensus are adopted as the label of sample data.

2. Preprocessing of sample data

After obtaining the aforementioned sample set data, the sample set data is preprocessed: carrying out preprocessing such as filtering or batch normalization on the obtained sample data, and carrying out data processing according to different data requirements to obtain sample data meeting the requirements; in this case, sample data for myocardial ischemia detection is required to be acquired as 10s electrocardiographic data, and time processing of 10s electrocardiographic data is required to satisfy a standard requirement common to electrocardiographic data. The preprocessing of the electrocardiographic data refers to conventional filtering operation, so that the subsequent operation is sufficiently stable and reliable, and the effectiveness of the method is not influenced by the filtering method.

Example 2 acquisition of quantitative index data of multiple pathological characteristics of attributes associated with specific cardiac diseases

After obtaining the sample set data preprocessed in example 1, acquisition of quantitative index data of various pathological features of the relevant attributes of a specific cardiac disease is performed. The quantitative index data of multiple pathological characteristics of the related attributes of the heart diseases comprise quantitative data of dynamic pathological characteristics of the heart, and/or index data of image characteristics of an electrocardiograph, and/or dynamic data of an electrocardiograph nonlinear system, and/or quantitative index data of auxiliary pathological characteristics, and/or human physiological information data, and/or clinical information data and the like; the specific operation process is carried out according to the following steps:

acquisition of quantitative index data of main pathological features and threshold judgment standard of corresponding heart diseases

The quantitative index data of the main pathological features refers to quantitative data of dynamic pathological features of the heart, and/or index data of image features of an electrocardiogram, and/or dynamic data of an electrocardio nonlinear system, and the like.

In an embodiment, the quantitative data of the dynamic pathological features of the heart includes, but is not limited to, quantitative index data of geometric features, and/or quantitative index data of nonlinear dynamic features, and/or quantitative index data of model features, and/or quantitative index data of time domain features, and/or quantitative index data of frequency domain features.

The following is a method for acquiring quantified data of a specific dynamic pathological feature of a heart, which is only used for illustration and is not limited to the method for acquiring or quantifying the data.

One of the quantitative indicators of the geometric features is recorded as

Wherein k is the cycle number of the electrocardio data, i is the ith cycle of the electrocardio data, and F_iIs the form factor of the ith period. One of the quantitative indicators of the nonlinear dynamics is recorded as

f (t) is a dynamic data sequence of the electrocardio nonlinear system,

the Fourier transform sequence of the dynamic data sequence of the electrocardio nonlinear system. One of the quantitative indicators of the model features is recorded as

Wherein, c_iThe time dispersion characteristic of each bit of the dynamic data of the electrocardio nonlinear system is obtained. One of the quantization indexes of the time domain features is recorded as

Wherein n is the number of data in each period, j is the sequence number of specific data in each period, var_jIs the jth data x of the ith cycle_ijThe variance of (c). One of the quantization indexes of the frequency domain features is recorded as

Wherein, w_iTo make it possible to

Minimum, f_i(n) is the frequency of the dynamic data of the electrocardio nonlinear systemDomain data. Finally, the sample set data preprocessed in embodiment 1 is imported to obtain the quantization index data of the geometric features, the quantization index data of the nonlinear dynamics features, the quantization index data of the model features, the quantization index data of the time domain features, and the quantization index data of the frequency domain features of each sample. The threshold value judgment method of the heart diseases of the quantitative index data of the dynamic pathological features of the heart obtains the critical diagnosis value of the specific heart diseases of the quantitative index data of the dynamic pathological features of the heart through a LOGISTIC regression method, and judges the specific heart diseases. In some embodiments, the quantitative data of the dynamic pathological features of the heart can be further selected from corresponding quantitative data acquired by other known technologies.

For index data of image features of the cardiac vector graph, acquiring tagged cardiac electrical signal data e (T), T being 1,2 and … T, preprocessing the acquired cardiac electrical signal data, and then converting to obtain a cardiac vector graph (VCG) (see KorsJ.A. and the like, published in 1990 in the thesis of European Heart Journal 11(12): 1083), and obtaining corresponding cardiac vector data X (T), Y (T) and Z (T); and 4 grades of quantification of the vector cardiogram according to the severity of the specific heart disease are divided into no, mild, moderate and severe according to the label, and the values are respectively assigned as 0,1, 2 and 3.

For the dynamic data of the electrocardio nonlinear system, collecting labeled cardiac electric signal data e (T), wherein T is 1,2 and … T; carrying out preprocessing such as filtering and baseline drift on the collected electrocardio-electric signal data, then obtaining a neural network model of the electrocardio-electric signal data or the electrocardio-vector data by adopting a mathematical operation method on the electrocardio-electric signal data or the electrocardio-vector data, and calculating to obtain the dynamic data of the electrocardio-nonlinear system:

wherein S (X (t)), S (Y (t)), and S (Z (t)) are Gaussian radial basis functions;

and

is a weight vector through a constant neural network; wherein, the mathematical operation method includes but is not limited to an adaptive system identification method; the adaptive system identification method employs methods including, but not limited to, modeling methods including, but not limited to, methods of gaussian radial basis function neural network modeling. Finally, visually displaying the obtained dynamic data of the electrocardio nonlinear system to obtain a dynamic graph of the electrocardio nonlinear system; then, 4 grades of quantification is carried out on the electrocardio nonlinear system dynamic graph of the category according to the severity of the specific heart disease, and the 4 grades are respectively assigned with 0,1, 2 and 3.

(II) acquisition of quantitative index data of auxiliary pathological features and threshold judgment standard of corresponding heart diseases

In the present invention, the quantitative index data for assisting pathological features refers to a quantitative value for assisting pathological feature attributes of heart diseases.

Quantitative index data of auxiliary pathological features, including but not limited to index data of electrocardiogram image features, and/or biochemical data; the indicator data of the electrocardiogram image features comprises quantized data of ECG morphology indicators, including but not limited to P-wave amplitude values, ST-elevated values, ST-depressed values and U-wave amplitude values. The following is a specific extraction method for auxiliary pathological feature quantization index data:

quantitative data of ECG morphology index and threshold judgment standard corresponding to heart disease

Electrocardiographic data can be obtained by a common electrocardiograph, and information is determined according to 2009 electrocardiographic standardization and interpretation guidelines (Wagner GS, Macfarland P, Wellens H, et al AHA/ACCF/HRS criteria for the standardization and interpretation of the electrocardiograph: part VI, elementary/information: a scientific description from the American healthcare association electronics and Arrhytmia Committee, ncil on clinical Cardiagnostics, the American College of clinical information, and the Heart resistive type J Am color 2009,53:1003), and different index parameters corresponding to different diseases. For example, the diagnosis criteria of myocardial infarction with complete left bundle branch block and myocardial infarction with complete left bundle branch block are as follows: (a) the QRS wave main wave is raised to the upper lead ST section by more than or equal to 0.1 mV; (b) the pressure of the ST section of the V1-V3 lead is more than or equal to 0.1mV, so that the S wave is obvious; the upper two are consistency changes of the ST segment. (c) The elevation of the QRS wave main wave to the lower lead ST segment is more than or equal to 0.5mV, which is called uncoordinated ST segment change.

2. Biochemical data and threshold decision criteria for corresponding heart disease

In cardiovascular diseases, biochemical indexes adopted by detection include but are not limited to high-sensitivity C-reactive protein, myocardial enzyme and the like, and corresponding experimental data can be obtained through routine hospital physical examination.

a. High sensitive C-reactive protein: not less than 2mg/L has a greater risk of cardiovascular disease.

b. A cardiac enzyme: in acute myocardial infarction, myocardial cell necrosis releases various enzymes in myocardium, including aspartate Aminotransferase (AST), lactate dehydrogenase (LD or LDH), Creatine Kinase (CK), isozyme, a-hydroxybutyrate dehydrogenase (a-HBD), and the like.

c. Other biochemical indexes are as follows: including but not limited to myoglobin, creatine kinase isozyme, C-reactive protein, triglyceride, high density lipoprotein, low density lipoprotein, lactate dehydrogenase, and the like, may also be used for disease detection.

3. Gold standard index for specific heart diseases and threshold judgment standard for corresponding heart diseases

The gold standard index of the specific heart disease refers to biochemical indexes, imaging indexes, metabolic markers and the like which are generally accepted for the specific heart disease. As in the present invention, the FFR index for the detection of myocardial ischemia is used; detection indexes of myocardial infarction, such as troponin, creatine kinase isoenzyme, imaging indexes and the like; coronary artery angiography index for coronary heart disease detection, and the like.

(III) human body physiological information data and judgment standard corresponding to heart disease

The human physiological information data includes but is not limited to sex, age, weight, smoking history, drinking history, exercise status, nationality and/or region, etc.; the human physiological information data also includes, but is not limited to, pathological data related to the heart activity state: dynamic electrocardiographic detection duration, fastest (slow) heart rate, average heart rate, total heart beat number, premature ventricular count, P-wave dispersion, P-wave (Q-wave) time limit, Q-wave depth, QRS time limit, QTc interval, QT/RR slope, QT interval variability, T-peak end time, J-wave height, J-wave dispersion, atrioventricular conduction block status, cardiomyopathy history, coronary heart disease history, myocardial infarction history, sudden death family history, cardiac function NYHA classification, hypertension history, valvular disease history, congenital heart disease history, diabetes history, cerebrovascular disease history, cardiovascular family history, genetic disease history, systolic pressure, diastolic pressure, body mass index, whether or not a pacemaker is installed, coronary bypass status, coronary stent status, ICD treatment status, lead ablation status, B-receptor blocker usage status, calcium channel antagonist usage status, ACEI/ARB usage status, QRS usage, Diuretic use, history of antiarrhythmic drug use, history of digitalis drug use, history of lipid lowering drug use, non-persistent ventricular rate, Lowns grade, SDNN, SDANN, ASDNN, rmsd, pNN50, pNN50a, TD, heart rate deceleration force, fQRS, RMS40, LAS, microvolt T-wave electrical alternation, homocysteine, LDL, HDL, brain natriuretic peptide, NT-proBNP, creatinine, left atrial diameter, left ventricular end diastolic diameter, left ventricular posterior wall thickness, left ventricular septum thickness, left ventricular ejection fraction, FS, mitral regurgitation, tricuspid regurgitation, aortic regurgitation, phasic ventricular wall motion abnormalities, LM, LAD, LCX, RCA, TIMI grade, and intima-media thickness, etc.; the decision criteria for the corresponding heart disease refer to hospital diagnostic records and published medical diagnostic rules.

(IV) clinical information data and judgment standards corresponding to heart diseases:

in the present embodiment, the clinical information data includes, but is not limited to, the condition information of the patient, such as bradycardia, blood pressure decrease, palpitation, sore throat and burning sensation, throat tightness, toothache, lassitude and hypodynamia, vertigo, chest stuffiness and pain, and/or shortness of breath, and the data of medical record or EMR (electronic medical record), examination, image, diagnosis, prescription, treatment, evaluation table of the patient. In this embodiment, for clinical information data, 4 levels of quantification of none, mild, moderate and severe will be assigned with values of 0,1, 2 and 3; specifically, two types of discrimination can be performed such as sleepiness and stickiness, and 1 and 0 are assigned according to the presence or absence.

(V) preprocessing quantitative index data of multiple pathological features of heart disease related attributes

After obtaining the quantitative index data of various pathological features of the relevant attributes of the specific heart diseases, preprocessing operations including but not limited to missing values and normalization are carried out. The quantitative index data of various pathological features of the relevant attributes of the specific heart disease is the input vector x for subsequent processing. For the above input vector x, x ═ x (x)¹,…xⁿ) (ii) a Wherein x is_i∈RⁿAnd n is the number of the feature vectors, and quantitative index data corresponding to main pathological features, quantitative index data corresponding to auxiliary pathological features, human physiological information data, clinical information data and the like. The normalization process is performed on the input vector x as described above with reference to the following equation:

x*＝(x_i－x_min)/(x_max－x_min)

x_maxis the minimum value, x, in the sample data_minIs the minimum value in the sample data; x is the number of_iIs the input vector of the ith sample.

Embodiment 3 construction and optimization of machine learning model for preprocessed quantitative index data of pathological features of heart diseases

The embodiment is to perform further research on the basis of embodiment 2, and the embodiment is mainly to construct a machine learning model of a specific heart disease, input preprocessed quantitative index data of pathological features of the heart disease obtained in embodiment 2, and train and optimize the constructed machine learning model of the specific heart disease.

(I) constructing machine learning judgment model of specific heart diseases

As shown in fig. 1, the quantitative index data of multiple pathological features of the attributes related to the specific cardiac disease obtained in embodiment 2 is used as input data to perform machine learning, construct a machine learning model adapted to the specific cardiac disease, and implement a one-to-one correspondence between the quantitative index data of each pathological feature and the attributes related to the specific cardiac disease; the method specifically comprises the following steps: the machine learning judgment model of multiple pathological characteristic quantitative indexes such as electrocardio nonlinear system dynamic data, and/or quantitative data of dynamic pathological characteristics of the heart, and/or index data of image characteristics of an electrocardiograph, and/or quantitative data of ECG morphological indexes, and/or biochemical data, and/or human physiological information data, and/or clinical information data is constructed.

The machine learning judgment model is divided into a machine learning classification model and a deep learning identification model. The machine learning classification model is selected from at least one of a support vector machine, a Bayesian classifier, a K-nearest neighbor algorithm, a K-means algorithm, a linear regression, a logistic regression, a multivariate nonlinear regression fitting method, an Adaboost algorithm, a hidden Markov model, an extreme learning machine, a random forest algorithm, a decision tree algorithm, and a clustering algorithm, and the deep learning identification model is selected from at least one of a convolutional neural network, a recurrent neural network, a generative confrontation network, a stacked autoencoder, a fully connected network, an unsupervised pre-training network, a deep belief network, a deep boltzmann machine, and a neural tensor network. Input data for the machine learning classification model include, but are not limited to: dynamic data of an electrocardio nonlinear system and/or quantitative data of dynamic pathological characteristics of a heart, and/or index data of image characteristics of an electrocardiograph vector graph, and/or quantitative data of ECG morphological indexes, and/or biochemical data, and/or human physiological information data, and/or clinical information data and other various pathological characteristic quantitative indexes; the input data of the deep learning identification model comprises but is not limited to electrocardio nonlinear system dynamic data, and/or quantitative data of dynamic pathological features of the heart, and/or index data of image features of vector cardiogram, and/or quantitative data of ECG morphology indexes.

1. Construction of support vector machine classification model for specific heart diseases

For the construction of the support vector machine model, the input vector is x ═ x (x)¹,…xⁿ) The decision rule is:

the weight is y₁a₁,…y_Na_N. The input vector x is mapped into a high-dimensional feature space by a predetermined non-linear mapping phi, and then an optimal hyperplane is constructed in the high-dimensional space. According to the collected sample set { (x)_i,y_i) 1,2, …, k, where k is the number of sample data; find an optimal functional relationship y ═ f (x) that reflects the sample data. After the training is completed, a machine learning classification model f (x) sgn (Σ) is obtained_{Support vector}y_ia_iK(x_i,x)+b₀) (ii) a Wherein, a_iIs a lagrange multiplier; b₀Is an offset; x is the number of_i∈RⁿN is the number of feature vectors, x_iQuantitative index data of various pathological features corresponding to the relevant attributes of the heart diseases and the like in example 2; y is_iTo desired output, y_iE { +1, -1}, corresponding to the classification of the subject (i.e. patient with the disease or not); if the subject is a healthy person, y _i1 ═ 1; if the subject is a patient with a particular heart disease, y_i-1; here, we select the radial basis kernel function: k (x)_i,x)＝exp(-‖x_i-x‖²/2σ²) Where σ denotes the kernel mesoscale parameter of the radial basis kernel function, the kernel parameter σ²Reflects the distribution or range characteristics of the training sample data, determines the width of the local domain, and has a larger kernel parameter sigma²Meaning a lower variance.

2. Construction of convolutional neural network and cyclic neural network recognition models for specific cardiac diseases

The artificial intelligent convolution neural network model is a multi-layer network connection structure for simulating a neural network, an input signal sequentially passes through each hidden layer, a series of complex mathematical processing is carried out in the hidden layer, some characteristics of an object to be identified are automatically abstracted layer by layer, then the characteristics are taken as input and transmitted to the hidden layer of a higher level for calculation until the whole signal is reconstructed by the full connection layers of the last layers, and a Softmax function is used for carrying out logistic regression to achieve multi-target classification. The method for constructing the three-dimensional neural network recognition model of the electrocardiogram nonlinear system dynamic graph (or VCG) comprises the following steps:

(1) construction of three-dimensional convolution neural network of electrocardio nonlinear system dynamic graph (or VCG)

The deep convolutional neural network recognition model of the electrocardiogram (or VCG) nonlinear system dynamic graph comprises the following frameworks: (a) inputting a three-dimensional convolution layer: defining an input 28 × 28 ecg nonlinear system dynamic graph, wherein the number of convolution layers is 3, performing convolution on the input image, the sizes of convolution kernels are all 3 × 3, and the pooling layer adopts a maximum pooling method; (b) an active layer: setting an activation function to Relu; (c) maximum three-dimensional pooling layer: reducing feature dimensions; (d) anti-overfitting layer: a percentage of the number of neurons is disconnected at each training to prevent data overfitting; (e) full connection layer: there are 1 full connection layer.

The convolutional neural network model of the electrocardiogram (or VCG) nonlinear system dynamic diagram comprises the following steps: (1) creating a metadata base: storing a three-dimensional electrocardiogram nonlinear system dynamic graph (or VCG) with a health or disease label into a metadata base; (2) and constructing a deep convolution neural network taking a three-dimensional electrocardiogram nonlinear system dynamic graph (or VCG) with a health or disease label as an input. (3) And (3) bringing the electrocardio nonlinear system dynamic graph (or VCG) in the metadata base into a deep convolutional neural network for training, and establishing a three-dimensional electrocardio nonlinear system dynamic graph (or VCG) deep convolutional neural network recognition model. The deep convolutional neural network recognition model takes a three-dimensional matrix image as input, and maximally retains three-dimensional space information of a three-dimensional electrocardio nonlinear system dynamic graph (or VCG): and (4) performing spatial feature extraction on the three-dimensional matrix by using the three-dimensional convolution layer and the three-dimensional convolution kernel at the input part of the network, and then performing the next operation.

(2) Construction of cyclic neural network of electrocardio nonlinear system dynamic graph (or VCG)

The recurrent neural network includes three layers: the system comprises an input layer, a hidden layer and an output layer, wherein the input layer receives a convolution neural network output result of an electrocardio nonlinear system dynamic graph (or VCG) as input data, updates the node state of the hidden layer of the network through an activation function, and predicts the heart disease information of a detection sample through the output layer; the state of the hidden layer stores heart disease information of a large number of samples, the relation between the heart disease information of the large number of samples and the current time information can be mined, and the judgment method for the heart disease at the t moment of the recurrent neural network RNN is as follows:

calculating y according to the corresponding information input by the recurrent neural network structure by using a forward propagation algorithm_tRealizes the judgment of heart diseases, has the time step length of 6 and is x with continuous time_t-5……x_tGenerating a primary sequence prediction y_t(ii) a Output of cardiac disease state at time t y_tThe calculation is carried out from the time t-5 to the forward transfer one by one, and the calculation relationship is as follows: h is_t＝σ(Ux_t+Wx_t-1+ b); wherein h is_tIs the hidden state of the RNN model at the time t, 230 dimensions; x is the number of_tIs a heart disease state vector at time t, 192 dimensions; σ is the tanh activation function, b is the 230-dimensional linear relationship bias vector, and U, W, V is the parameter of RNN, U size 192 × 230, W size 230 × 230, V size 230 × 1, shared per stage. Model output at time t O_tComprises the following steps: o is_t＝Vh_t+ c, c is a one-dimensional offset; predicted output y at time t_tComprises the following steps: y is_t＝θ(O_t) And θ is the softmax activation function.

The first use needs to train the parameters of the recurrent neural network, and the training process of the recurrent neural network is as follows: and obtaining the parameter U, W, V, the linear relation offset vector b and the one-dimensional offset c through a back propagation algorithm by one round of iteration of a gradient descent method.

The back propagation algorithm defines a loss function

The time step τ is 6, and then the formula of the back propagation V, c gradient is obtained:

wherein: y is_tIs the predicted output at time t and,

is the actual output at time t.

The gradient defining the hidden state at time t of the sequence is:

therefore, the method comprises the following steps:

gradient calculation expression of backpropagation W, U, b:

wherein: x is the number of_tIs the cardiac disease state vector at time t, h_tIs the hidden state of the RNN model at time t, the gradient of the hidden state at time t of the sequence being

The training process of the recurrent neural network RNN is as follows: (a) initializing the value of each U, W, V, b and c to be a random value, and taking the value range of [0,1]](ii) a (b) for iter to 1to training iteration steps (200); (c) the for start is 1to the data acquisition quantity-5; (d) computing y using a forward propagation algorithm_t(ii) a (e) Calculating a loss function L; (f) back propagation algorithmCalculating the partial derivative values of all hidden layer nodes by using the output layer node values, and updating U, W, V, b and c; and finally, outputting a judgment result.

(II) optimization of machine learning decision models for specific cardiac disorders

(1) Training and optimizing the model: and continuously training a machine learning judgment model of the quantitative index data of the multiple pathological characteristics of the specific heart disease through a large amount of sample data with known labels, and performing a large amount of weight experiments to obtain the optimal weight distribution of the quantitative indexes of the various pathological characteristics of the relevant attributes of the specific heart disease.

(2) Testing of the model: and inputting a batch of sample data with known labels in the training optimized machine learning model to test the training optimized model.

Example 4 evaluation of machine learning decision models for specific cardiac disorders

In order to comprehensively evaluate the performance of the machine learning model, the training optimized model obtained in embodiment 3 is measured by using the accuracy, sensitivity and specificity indexes of heart disease detection, a large amount of sample data with known labels is input into the training optimized machine learning model, and a related judgment result is output. Detecting the accuracy, sensitivity and specificity of the machine learning judgment model according to the judgment result output by the machine learning model; and comprehensively evaluating and optimizing the application performance and the deficiency of the machine learning judgment model of the specific heart disease to complete the establishment of the machine learning judgment model of the specific heart disease.

The accuracy, sensitivity and specificity of the detection result of the machine learning judgment model are specifically defined as follows:

accuracy ═ true positive sample number + true negative sample number)/(true positive sample number + false positive sample number + true negative sample number + false negative sample number);

sensitivity ═ true positive sample number/(true positive sample number + false negative sample number);

the specificity is the number of true negative samples/(number of true negative samples + number of false positive samples);

wherein the number of true positive samples represents the number of samples actually suffering from a heart disease and detected as suffering from a heart disease; the number of false positive samples represents the number of samples that were actually not heart disease but were detected as heart disease; the number of true negative samples represents the number of samples that actually did not suffer from a heart disease and were detected as not suffering from a heart disease; the number of false negative samples represents the number of samples that actually suffered a heart disease and were detected as not suffering a heart disease; for a heart disease detection model, the higher the three indexes are, the better the model can be for the heart disease, and the better the model effect is.

Through the above manner, optimized machine learning judgment models are obtained, specifically, a support vector machine classification model of a specific heart disease, a convolutional neural network recognition model of a vector electrocardiogram (VCG), a convolutional neural network recognition model of a dynamic diagram of a nonlinear electrocardiogram system, a convolutional neural network and cyclic neural network recognition model of a VCG, and a convolutional neural network and cyclic neural network recognition model of a dynamic diagram of a nonlinear electrocardiogram system.

The inventor selects 1714 sample data, evaluates the constructed convolutional neural network and cyclic neural network identification model, and obtains experimental data evaluated by the convolutional neural network and cyclic neural network identification model of the electrocardiogram nonlinear system dynamic diagram, which is specifically shown in table 1; the experimental result shows that the accuracy of the convolutional neural network and the cyclic neural network recognition model of the electrocardio nonlinear system dynamic graph is 89.0 percent, the sensitivity is 91.7 percent, and the specificity is 81.5 percent; is the optimal machine learning judgment model in the several models; experimental data for evaluation of other models, such as the convolutional neural network and cyclic neural network recognition models of VCG, are not listed here.

TABLE 1 evaluation of convolutional neural network and cyclic neural network recognition models for electrocardiograph nonlinear system dynamic graphs

Example 5 weight optimization of decision outcome for cardiac diseases

Inputting a large amount of sample data with known labels into the support vector machine decision model, the convolutional neural network of the electrocardiograph nonlinear system dynamic graph (and/or VCG) and the recurrent neural network recognition model optimally screened in embodiment 4, obtaining the output results of the corresponding support vector machine classification model, the convolutional neural network of the electrocardiograph nonlinear system dynamic graph (and/or VCG) and the recurrent neural network recognition model, combining the results with the quantitative data of the dynamic pathological features of the heart, the quantitative index data of the auxiliary pathological features, the human physiological information data or the clinical information data extracted in embodiment 2, assigning a weight value suitable for a specific heart disease, and then performing a weight assignment experiment with statistical significance to obtain different output results (the output results of the support vector machine classification model, the convolutional neural network of the electrocardiograph nonlinear system dynamic graph (and/or VCG) and the recurrent neural network recognition model output Output results, threshold value judgment of heart diseases of quantitative data of dynamic pathological characteristics of the heart, threshold value judgment of heart diseases of quantitative information of auxiliary pathological characteristics, human physiological information data and clinical information data), and storing the optimal weights of different output results aiming at the specific heart diseases to obtain a comprehensive judgment system with optimized weights aiming at the specific heart diseases, namely a support vector machine classification model output result with weight, a convolution neural network and circulation neural network identification model output result of an electrocardiogram nonlinear system dynamic graph (and/or VCG) with weight, a threshold value judgment of heart diseases of quantitative data of heart dynamic pathological characteristics with weight, a threshold value judgment of heart diseases of auxiliary pathological characteristics with weight, and a comprehensive judgment system of human physiological information data with weight and clinical information data with weight.

Example 6 auxiliary detection of cardiac diseases by machine learning decision models

In this embodiment, on the basis of embodiments 1to 5, the comprehensive determination system with optimized weights for specific cardiac diseases constructed in embodiment 5 is used to perform detection of specific cardiac diseases. In this embodiment, quantitative index data of various pathological features of a person to be tested is mainly collected and input into the comprehensive judgment system with optimized weights for the specific cardiac diseases, which is constructed in embodiment 5, so as to perform accurate detection and rapid identification of different cardiac diseases, as shown in fig. 2. The specific method for detecting the heart disease by the machine learning judgment model comprises the following steps:

step one, collecting index data of various pathological characteristics of a person to be detected and preprocessing

Acquiring index data of various pathological characteristics of heart disease related attributes of a person to be tested, and correspondingly preprocessing the acquired index data of various pathological characteristics of the person to be tested, wherein the preprocessing method is recorded according to the embodiment 1 and the embodiment 2, so that the acquired index data becomes standard data suitable for a machine learning judgment model; the selection of index data of various pathological characteristics of heart disease related attributes needs to be screened and determined according to specific heart diseases.

Step two, obtaining quantitative index data of multiple pathological characteristics of specific heart disease related attributes

By using the method described in embodiment 2, from the preprocessed normative data of multiple pathological features obtained in the first step, quantitative data of dynamic pathological features of the heart of the person to be tested, index data of image features of an vectogram, dynamic data of an electrocardiograph nonlinear system, quantitative index data of auxiliary pathological features, physiological information data of a human body, and clinical information data are obtained.

Step three, outputting judgment results of specific heart diseases

And (3) inputting the quantitative index data of various pathological characteristics related to the specific heart disease acquired in the step two into the comprehensive judgment system with the optimized weight value of the specific heart disease constructed in the embodiment 5, and outputting the final comprehensive judgment result of the corresponding heart disease to obtain the detection information of the heart health of the person to be detected.

Example 7 comparison of the technical Effect of the method of the present invention with that of the conventional method

This embodiment mainly verifies the actual application of the detection method described in embodiment 6. The inventor adopts the diagnosis of 30 general cardiologists in the hospital and the detection method described in the embodiment 6 to respectively detect the myocardial ischemia of three groups of testees (#1Trial group, #2Trial group and #3Trial group), and the number of the testees in each group is 120; the three groups of persons to be tested all adopt quantitative data of heart dynamic pathological characteristics, quantitative index data of image characteristics of the vector electrocardiogram, an electrocardiogram nonlinear system dynamic diagram, quantitative data of ECG morphological indexes, biochemical data, human body physiological information data and clinical information data of the persons to be tested. The quantitative data of the dynamic pathological features of the heart are quantitative data of Lyapunov index spectra; the quantitative data of the ECG morphology indexes are acquired by performing corresponding processing and operation described in embodiment 2 through ST elevation values; biochemical data, selecting a high-sensitivity C-reactive protein detection value to perform corresponding processing and operation as described in embodiment 2; human body physiological information data, age, sex, smoking history and drinking history are selected to carry out corresponding processing and operation as described in embodiment 2; clinical information data, selected from the conditions of palpitation, chest stuffiness and pain and vertigo, were processed and calculated as described in example 2.

In the #1Trial group, 60 persons over the age of 40 were present, 30 men (15 who smoked and drunk, and 15 who did not smoke and drink), and 30 women (15 who smoked and drunk, and 15 who did not smoke and drink); the number of people under age 40 is 60, male is 30 (15 who smoke and drink, and 15 who do not smoke and drink), female is 30 (15 who smoke and drink, and 15 who do not smoke and drink); similarly, the selection requirements of the #2Trial group and the person to be tested in the #3Trial group are the same as the selection requirements of the #1Trial group. Meanwhile, the output results of the support vector machine classification model, the output results of the VCG convolutional neural network and cyclic neural network recognition model, the output results of the electrocardio nonlinear system dynamic graph convolutional neural network and cyclic neural network recognition model, the threshold judgment of the heart diseases of the quantitative data of the dynamic pathological features of the heart, the threshold judgment of the heart diseases of the quantitative information of the auxiliary pathological features, and the weight values of the human physiological information data and the clinical information data are respectively 0.15, 0.2, 0.3, 0.15, 0.1, 0.05 and 0.05; the quantitative data and biochemical data of ECG morphological indexes are selected as the auxiliary pathological feature quantitative information, and the corresponding weight values are 0.3 and 0.7 respectively.

As shown in tables 2 and 3, in the #1Trial group experiment, 120 persons to be tested were tested, the number of myocardial ischemia individuals was determined to be 90 by both the coronary angiography and the fractional flow reserve, the average number of myocardial ischemia individuals diagnosed by each doctor in the general cardiac department of the third hospital was 46, and the average total consumed time of each doctor was 270 min; the average number of subjects with myocardial ischemia detected using the method described in example 6 was 84, which took a total of 3.2 min. In the #2Trial group experiment, 120 persons to be tested are tested, the number of myocardial ischemia individuals is determined to be 80 through two technical means of coronary angiography and fractional flow reserve, the number of myocardial ischemia individuals diagnosed by 30 ordinary heart departments in the trimethyl hospital is 40 on average, and the total consumed time of each doctor is 265min on average; the average number of subjects with myocardial ischemia detected using the method described in example 6 was 74, which took a total of 3.1 min. In a #3Trial group experiment, 120 persons to be tested are tested, the number of myocardial ischemia individuals is determined to be 70 through two technical means of coronary angiography and fractional flow reserve, the number of myocardial ischemia individuals diagnosed by 30 ordinary heart departments in the trimethyl hospital is 34 on average, and the total consumed time of each doctor is 260min on average; the average number of subjects with myocardial ischemia detected using the method described in example 6 was 63, which took a total of 3.1 min. The experimental results show that the method has remarkably improved and obvious improvement on the performance of the method in the aspects of detection accuracy, detection time consumption, continuous processing of dynamic information and the like when the myocardial ischemia condition is detected.

TABLE 2 comparison of the test results of the method of the present invention with those of the conventional method

TABLE 3 comparison of the test results of the method of the present invention with those of the conventional method

It can be seen that the method of the present invention has several advantages over the prior art: the dynamic pathological characteristic information of the heart disease is more abundant, the detection accuracy, the detection efficiency and the like are obviously improved, and continuous dynamic signals can be processed.

The inventor finds that, in the actual operation process, the quantitative index data of the multiple pathological characteristics described in embodiment 2 can be partially or completely selected, and the method described in embodiment 2 and/or embodiment 3 is selected and/or combined to detect the heart disease, so as to achieve the ideal detection effect of the heart disease.

Embodiment 8 wearable device for cardiac disease detection

The embodiment describes wearable equipment for detecting heart diseases, which comprises wearable clothes, a heart signal sensor, a monitoring chip connected with the heart signal sensor, a conductive electrode and an electric connecting lead, wherein the wearable clothes are of a double-layer structure, and the conductive electrode is embedded in an interlayer of the wearable clothes and is connected with the heart signal sensor through the electric connecting lead; the conductive electrodes include ten electrodes, four of which are limb electrodes and six of which are chest electrodes. The monitoring chip acquires the cardiac shock signal and the electrocardiosignal which are monitored by the cardiac signal sensor in real time. The monitoring chip totally or partially adopts the method described in the embodiment 1to the embodiment 6 to analyze and process the electrocardiosignals. The wearable device provided by the embodiment can be used for detecting diseases or symptoms, including but not limited to acute coronary syndrome, myocardial ischemia, myocardial infarction, sudden cardiac death, myocarditis, arrhythmia, myocardial infarction, angina pectoris, coronary heart disease, coronary artery disease, heart failure, and the like.

In some improved schemes, the monitoring chip can be connected with the heart signal sensor through WIFI or Bluetooth to acquire signals monitored by the sensor, and the Bluetooth can adopt HC-05 serial port Bluetooth;

in some refinements, the monitoring chip comprises: the heart disease prediction method comprises a signal monitoring module (used for acquiring the cardiac shock signal and the electrocardio signal of the person to be detected monitored by the heart signal sensor in real time), a model processing module (used for analyzing and processing the methods in the embodiments 1to 6 according to the electrocardio signal) and a heart disease recognition module (used for predicting the heart disease of the person to be detected according to the result of the model processing module).

In some improvements, the signal monitoring module adopts an AD7061 analog-to-digital conversion chip and an AD8232 chip; the model processing module adopts an AD7061 analog-to-digital conversion chip, a MAX4472 integrated operational amplifier, a Samsung KMR7X0001M-B511FLASH memory chip and a Samsung S3C2410ARM9 embedded processor chip.

Embodiment 9. myocardial ischemia detection method based on convolutional neural network model

The embodiment mainly relates to a detection method of myocardial ischemia based on a convolutional neural network model, which comprises the following steps:

s1, acquiring an electrocardiographic vector diagram of a myocardial ischemia sample population; a vector cardiogram is obtained by the method of the embodiment 2. The myocardial ischemia sample population is the myocardial ischemia patients which can not be diagnosed by electrocardiogram.

And S2, constructing a convolutional neural network model of the vector cardiogram detected by myocardial ischemia, wherein the setting parameters of the convolutional neural network model of the vector cardiogram detected by myocardial ischemia are the same as those in the embodiment 3.

By collecting the vectorcardiogram of the person to be tested described in embodiment 7 and performing an experiment, it is found that the myocardial ischemia detection accuracy of the convolutional neural network model of the vectorcardiogram of this embodiment is improved by at least 10% as compared with the diagnosis accuracy of 30 general cardiologists in the hospital.

The inventor carries out further demonstration experiments on the basis of the above experiments, and finds that the myocardial ischemia detection accuracy of the convolutional neural network model of the electrocardio nonlinear system dynamic graph is improved by at least 12% compared with the diagnosis accuracy of 30 common cardiologists in trimethyl hospitals by acquiring the electrocardio nonlinear system dynamic graph of the person to be tested in embodiment 7 and inputting the electrocardio nonlinear system dynamic graph into the convolutional neural network model of the electrocardio nonlinear system dynamic graph in embodiment 3; then, quantitative index data of the nonlinear dynamics characteristics of the person to be tested in embodiment 7 are collected and input into the machine learning judgment model in embodiment 3, and the detection accuracy of the machine learning judgment model of the quantitative indexes of the nonlinear dynamics characteristics is found to be improved by at least 7% compared with the diagnosis accuracy of 30 common cardiac department doctors in trimethyl hospital; then, the output result of a machine learning judgment model of the quantitative index of the nonlinear dynamics characteristics and the output result of the convolutional neural network model of the vector cardiogram are respectively assigned with weighted values of 0.5 and 0.5, and the accuracy of myocardial ischemia detection is improved by at least 25 percent compared with the diagnosis accuracy of 30 common cardiologists in the hospital of trimethyl; then, the judgment information of the heart diseases considering the quantitative index data of the nonlinear dynamics characteristics of the person to be detected is added, the output result of the machine learning judgment model of the quantitative indexes of the nonlinear dynamics characteristics, the output result of the convolutional neural network model of the cardiac vector graph in the embodiment and the weight values of the judgment information of the heart diseases of the quantitative index data of the nonlinear dynamics characteristics of the person to be detected are distributed to be 0.1, 0.6 and 0.3, and the accuracy of myocardial ischemia detection is improved by at least 32 percent compared with the diagnosis accuracy of 30 common cardiologists in trimethyl hospitals. Finally, the output result of the machine learning judgment model of the quantitative index of the nonlinear dynamics characteristics, the output result of the convolutional neural network model of the vectorcardiogram described in this embodiment, the output result of the dynamic graph convolutional neural network model of the electrocardiograph nonlinear system, and the weight values of the judgment information of the heart diseases of the quantitative index data of the nonlinear dynamics characteristics of the person to be detected are assigned to 0.1, 0.4, 0.3, and 0.2, and the accuracy of myocardial ischemia detection is improved by at least 39% compared with the diagnosis accuracy of 30 common cardiologists in trimethyl hospitals.

In step S2, the machine learning decision model for the specific heart disease may further be at least one of Adaboost algorithm, hidden markov model, extreme learning machine, random forest algorithm, decision tree algorithm, generative opponent network, stacked automatic encoder, deep belief network, deep boltzmann machine, and neural tensor network in machine learning.

Example 10 detection method of myocardial infarction with complete left bundle branch block based on circulatory neural network

The embodiment mainly relates to a method for detecting myocardial infarction accompanied by complete left bundle branch block based on a circulatory neural network, which comprises the following steps:

s1, acquiring index data of multiple pathological characteristics of the complete left bundle branch block of the myocardial infarction, wherein the index data of the multiple pathological characteristics of the complete left bundle branch block of the myocardial infarction is an electrocardiogram nonlinear system dynamic diagram;

and S2, performing machine learning on the index data of the multiple pathological characteristics of the myocardial infarction complicated by the complete left bundle branch block obtained in the step S1 to construct a circulatory nerve network model of the myocardial infarction complicated by the complete left bundle branch block, wherein the specific parameters are the same as those of the embodiment 3.

By collecting the dynamic graph of the electrocardiographic nonlinear system of the person to be tested in embodiment 7 to perform experiments, it is found that the detection accuracy of the complete left bundle branch block due to myocardial infarction of the recurrent neural network model of the dynamic graph of the electrocardiographic nonlinear system is improved by at least 15% compared with the diagnosis accuracy of 30 common cardiologists in the hospital.

The inventor carries out further demonstration experiments on the basis of the above experiments, acquires quantitative index data of the geometric characteristics of the person to be tested in embodiment 7, inputs the quantitative index data into the machine learning judgment model in embodiment 3, and finds that the detection accuracy of the machine learning judgment model of the quantitative indexes of the geometric characteristics is improved by at least 9% compared with the diagnosis accuracy of 30 common cardiologists in trimethyl hospital; when 120 persons to be detected are detected, the detection time consumption of a machine learning judgment model of quantitative indexes of geometric features is 84 times of the average diagnosis time consumption of 30 ordinary cardiologists in the trimethyl hospital; compared with the diagnosis accuracy of 30 common cardiologists in the third hospital, the determination accuracy of the myocardial infarction accompanied by the complete left bundle branch block of the quantitative index of the geometric characteristics has no obvious difference.

The inventor carries out further experiments on the basis of the above experiments, and finds that when the detection method of the embodiment is fused with the quantitative index data of the auxiliary pathological features of the person to be detected, the human physiological information data or the judgment information of the myocardial infarction accompanied with complete left bundle branch block of the clinical information data, the method is favorable for obtaining richer dynamic pathological feature information of the heart disease of the person to be detected, and is favorable for later-stage precise treatment and drug intervention.

In step S2, the recurrent neural network model may be replaced with a neural tensor network, and different vectors are directly computed interactively through tensors by using a neural tensor network layer (NTN), so as to implement matching between the feature representation and the similarity; the specific experimental data of the inventor are omitted; the index data of the pathological features of the complete left bundle branch block of the myocardial infarction can also be quantitative index data of other pathological features of the related attributes of the specific heart disease obtained from the embodiment 2.

Example 11 detection method of myocardial infarction with complete left bundle branch block based on artificial intelligence model

The embodiment mainly relates to a method for detecting myocardial infarction accompanied by complete left bundle branch block based on an artificial intelligence model, which comprises the following steps:

s1, acquiring index data of multiple pathological features of complete left bundle branch conduction block of myocardial infarction, wherein the index data of the multiple pathological features of the complete left bundle branch conduction block of myocardial infarction is index data of image features of electrocardiogram;

and S2, performing machine learning on the index data of the multiple pathological characteristics of the myocardial infarction complicated with the complete left bundle branch block obtained in the step S1 to construct a machine learning judgment model of the myocardial infarction complicated with the complete left bundle branch block.

Step S2, the machine learning decision model for the specific heart disease is at least one of a bayesian classifier, a K-nearest neighbor algorithm, a K-means algorithm, a linear regression, a logistic regression, a multivariate nonlinear regression fitting method, an Adaboost algorithm, a hidden markov model, an extreme learning machine, a random forest algorithm, a decision tree algorithm, a clustering algorithm, a generative countermeasure network, a stacked autocoder, a fully connected network, an unsupervised pre-training network, a deep belief network, a deep boltzmann machine, and a neural tensor network in machine learning.

The machine learning method of the logistic regression is obtained by further research on the basis of the machine learning method of the linear regression and is used for solving the problem of complex classification which is difficult to solve by the machine learning method of the linear regression. The machine learning method of the logistic regression is carried out according to the following steps: n clinically known arrhythmia individuals (n-120) and m clinically known contrast-blocking individuals (m-120) were included as a test sample population, quantitative index data of various pathological characteristics of heart disease-related attributes of the test sample population were collected according to the method described in example 2, and training was performed based on a logistic regression model using a Gridsearch function under a skearn module. Similarly, an electrocardiogram of the person to be diagnosed is acquired, processed in the same manner as described above, and input to the trained logistic regression model to obtain a classification result. After the inventor constructs the logistic regression model by using the method, the experimental verification proves that the training accuracy of the logistic regression model is more than 65%. Similarly, the machine learning method herein may also employ a multivariate non-linear regression fitting method or a linear regression method.

A machine learning method for a bayesian classifier, comprising: (1) obtaining symptom information of a patient, wherein the symptom information of the patient is obtained from symptom results extracted from the sample sets of example 1 and example 2; (2) and inputting the symptom information into a constructed multi-label Bayes classifier to obtain a prediction set of the disease of the patient, wherein the multi-label Bayes classifier is trained by using symptom results extracted from the sample sets of the embodiment 1 and the embodiment 2. The disease type is N, a Bayesian classifier with N labels is constructed, and the classification result of each classifier is as follows: { di, -di }, i.e., the samples belonging to the disease di and not belonging to the disease di, the input of the samples used for training is the symptom result S extracted from the sample set of example 1, and the symptom input of the patient is subjected to the multi-label Bayesian classifier during calculation, so as to obtain the disease set L which the patient may suffer from. (3) Determining a recommended combination of medical exams from a medical knowledge-graph based on the prediction set, wherein the medical knowledge-graph includes symptom entities, disease entities, single exam entities, and exam group entities, the exam group entities being determined by computing a frequent itemset of the single exam entities. The inventor establishes the Bayesian classifier by using the method and then conducts experimental verification, and finds that the accuracy of training by using the Bayesian classifier is more than 80%.

The machine learning method for the K-nearest neighbor algorithm comprises the following steps: acquiring 240 sample data sets; wherein 120 samples are arrhythmic, 120 samples are contrast blocking, and each sample comprises symptom results acquired in example 1 and example 2; secondly, after preprocessing the data set in the first step is carried out to meet the data format supported by the KNN model, screening the preprocessed data, carrying out data normalization processing on the preprocessed data, and constructing a training data set and a verification data set; (III) KNN model construction: the specific model construction method comprises the following steps of (1) calculating the Euclidean distance; (2) sorting according to the increasing relation of the distances; (3) selecting K points with the minimum distance, wherein K is 5-10; (4) determining the occurrence frequency of the category where the first K points are located; (5) and returning the category with the highest frequency of occurrence in the former K points as the prediction classification of the test data. After the KNN model is constructed by the method, the accuracy of the training by the KNN algorithm is more than 81.7 percent after experimental verification.

In this embodiment, the index data of the pathological features of the complete left bundle branch block of myocardial infarction can also be quantitative index data of other pathological features of which the relevant attributes of a specific heart disease are obtained from the example 2.

Similarly, the present embodiment may also use a K-means algorithm to perform the clustering algorithm analysis.

Embodiment 12 cardiac disease classification method based on AdaBoost classifier

On the basis of the research, the inventor further develops a heart disease detection method based on dynamic data of an electrocardio nonlinear system of an AdaBoost classifier, and the method comprises the following steps:

the number of samples was 240: 120 samples are the dynamic data of the electrocardio nonlinear system of the arrhythmia patient, and 120 samples are the dynamic data of the electrocardio nonlinear system of the contrast blocking patient. Each sample of the data is trained and given a weight, which constitutes a weight vector D, the dimensionality being equal to the number of samples of the data set. Initially, these weights are all equal, first a weak classifier is trained on a training data set and the error rate of the classifier is calculated, then the weak classifier is trained again on the same data set, but during the second training, the weights of the samples in the data set are adjusted according to the error rate of the classifier, the weight of the correctly classified sample is reduced, and the weight of the incorrectly classified sample is increased, but the sum of these weights remains not 1.

The final classifier can distribute different decision coefficients alpha based on the classification error rate of the trained weak classifiers, and the classifier with low error rate obtains higher decision coefficient, thereby playing a key role in predicting data. alpha is calculated based on the error rate:

alpha＝0.5*ln(1-ε/max(ε,1e-16))；

where ε is the number of samples correctly classified/total number of samples, and max (ε,1e-16) is to prevent the error rate from causing the denominator to be 0.

After alpha is calculated, the weight vector can be updated so that the samples that are misclassified get higher weights and the samples that are correctly classified get lower weights. The calculation formula of D is as follows: if a sample is correctly classified, the weights are updated as: d (m +1, i) ═ D (m, i) × exp (-alpha)/sum (D); if a sample is misclassified, the weights are updated as: d (m +1, i) ═ D (m, i) × exp (alpha)/sum (D); where m is the number of iterations, i.e. the mth classifier trained, i is the ith component of the weight vector, i <, the number of dataset samples. After we update the weights of the samples, the next iteration training can be performed. The AdaBoost algorithm can continuously repeat training and weight adjustment until the iteration times are reached or the training error rate is 0;

through experimental verification using the above method, the inventors found that the accuracy of training using the AdaBoost algorithm was 100%.

In this embodiment, the electrocardiographic nonlinear system dynamic data can also be quantitative index data of other multiple pathological features of the attributes related to the specific heart disease obtained from embodiment 2.

Example 13 detection of myocardial infarction with complete left bundle branch block based on generative antagonistic network

In this embodiment, the inventors have developed a method for detecting myocardial infarction with complete left bundle branch block based on a generative antagonistic network, comprising the steps of:

step 1, acquiring pathological feature data of myocardial infarction accompanied by complete left bundle branch block, wherein the pathological feature data of myocardial infarction accompanied by complete left bundle branch block is dynamic data of an electrocardio nonlinear system;

and 2, clustering the dynamic data of the electrocardio nonlinear system into a plurality of clusters by adopting a k-means algorithm, wherein each cluster comprises 100 pieces of dynamic data of the electrocardio nonlinear system, and normalizing all the dynamic data of the electrocardio nonlinear system to a [0,1] interval.

Step 3, constructing a dynamic data generation model G of the electrocardio nonlinear system, which comprises the following specific steps:

the convolution neural network of the dynamic data of the electrocardio nonlinear system comprises an input layer, a hidden layer and an output layer, and is used for generating the dynamic data of the virtual target electrocardio nonlinear system, and the process is as follows: firstly, converting preprocessed electrocardio nonlinear system dynamic data into a data matrix of a target dimension; then, defining convolution operation, namely obtaining an output matrix through sliding and calculation of a convolution kernel on an original input matrix, obtaining an output matrix with smaller dimensionality and realizing characteristic extraction of data; the convolution kernel is an n x n matrix (generally 3 x 3) with a small dimension, which is also called a weight matrix, the values of the matrix elements can be preset, the sliding step length can be set (generally 1), and each element value in the output matrix is the product of the convolution kernel and the original input matrix covered currently; then, defining deconvolution operation, wherein the form of the deconvolution operation is similar to that of convolution operation, namely, a transposed matrix obtained by a convolution kernel is multiplied by an input matrix, and an output matrix is obtained through conversion, and a matrix with larger dimensionality can be obtained under general conditions, so that data expansion is realized; and finally, performing convolution processing of 3 layers of different structures on the dynamic data matrix of the target dimension electrocardio nonlinear system, wherein the convolution processing comprises convolution and deconvolution of different step lengths to obtain n x m dimensional data, performing sigmoid processing on the n x m dimensional data, and outputting an n x m dimensional output matrix, wherein the n x m dimensional output matrix forms a virtual target electrocardio nonlinear system dynamic data set.

Step 4, constructing a dynamic data discrimination model D of the electrocardio nonlinear system, which comprises the following specific steps:

and the electrocardio nonlinear system dynamic data discrimination model is used for judging the authenticity of the input virtual target electrocardio nonlinear system dynamic data and the target electrocardio nonlinear system dynamic data.

The method comprises the steps of taking virtual target electrocardio nonlinear system dynamic data and target electrocardio nonlinear system dynamic data as input of an electrocardio nonlinear system dynamic data discrimination model, defining a maximum pooling layer and a convolution layer, utilizing the convolution layer to realize dimension expansion feature extraction of the input data, enabling a convolution kernel of the convolution layer to be a matrix of 5 x 5, enabling a step length to be 1 or 2, enabling the maximum pooling layer to realize dimension reduction feature extraction of the input data, namely performing numerical value extraction and dimension reduction on an input matrix with a larger dimension through a window to obtain an output matrix with a smaller dimension, wherein the window is 3 x 3, and the numerical value extraction is to select the maximum numerical value in each corresponding area of the window on an original data matrix to be used as an element of the output matrix.

Convolution and pooling operations of different structures are carried out on the target electrocardio nonlinear system dynamic data and the virtual target electrocardio nonlinear system dynamic data with the dimensionality of n x m, and a probability value is obtained through full connection processing calculation, wherein the range of the probability value is [0,1], namely when the data is judged to be the target electrocardio nonlinear system dynamic data, the probability is 1, and when the data is judged to be the virtual target electrocardio nonlinear system dynamic data, the probability is 0.

And the countermeasure training is a generation countermeasure network formed by the electrocardio nonlinear system dynamic generation model and the electrocardio nonlinear system dynamic discrimination model. The countermeasure and optimization of the generative countermeasure network are realized through the training strategy of alternate training. In the training process, the aim of training the electrocardio nonlinear system dynamic discrimination model D is to maximize the discrimination accuracy of the electrocardio nonlinear system dynamic discrimination model D as much as possible, namely when the probability is 1, the data is discriminated to be from the target electrocardio nonlinear system dynamic data and is marked with 1, and when the probability is 0, the data is discriminated to be from the virtual target electrocardio nonlinear system dynamic data, namely the data is generated by the electrocardio nonlinear system dynamic generation model G and is marked with 0. The training target of the electrocardio nonlinear system dynamic generation model G is to minimize the discrimination accuracy of the electrocardio nonlinear system dynamic discrimination model D.

Example 14 detection of myocardial infarction with complete left bundle branch block based on stacked AutoCoder

In this embodiment, the inventor has developed a method for detecting myocardial infarction with complete left bundle branch block based on a stacked automatic encoder, comprising at least the following steps:

s1, jointly inputting unlabeled electrocardiogram vector data of a training sample and electrocardiogram vector data of an actual label of the same training sample, obtaining the same output after multi-layer coding and decoding, and then taking the electrocardiogram vector data with the actual label in the output as a target detection result;

s2, the noise reduction stacking automatic encoder network comprises a plurality of layers, wherein the first layer is used as an input end and an output end, the noise reduction function is not realized through simple encoding and decoding, the relation among different dimensions is found through a plurality of times of encoding and decoding in the middle layer, the function of restoring the electrocardiogram vector data of the actual label from the unlabeled electrocardiogram vector data is learned from a sample, and the actual label of the unlabeled electrocardiogram vector data is obtained; in specific implementation, S2 further includes the steps of:

a1, generating a first-layer automatic encoder, encoding and decoding input information to obtain output information same as the original input, taking unlabeled electrocardiogram vector data and electrocardiogram vector data of an actual label of the same training sample as a common input F1, and encoding O1 ═ s₁(W₁F₁+b₁) To become an intermediate layer O₁Then decoded to reconstruct F₁’＝s₁(W₂O₁+b₂) The parameters of the model should approximate the reconstructed data to the original vector as much as possible;

a2, taking the output of the first layer encoder as the input of the second layer noise reduction automatic encoder, and also minimizing the reconstruction error of the second layer noise reduction automatic encoder, so that the reconstructed output of the second layer after encoding and decoding is the same as the second layer input;

a3, generating a plurality of layers of noise reduction automatic encoders in the middle;

a4, stacking each layer of noise reduction automatic encoders, inputting the first layer encoding and the second layer encoding … layer encoding in sequence, then performing the nth layer decoding … layer decoding and the first layer encoding in sequence, and outputting the same information as the input;

a5, when in use, using unlabelled electrocardiogram vector data of a training sample and electrocardiogram vector data of an actual label of the same training sample as common input, using the electrocardiogram vector data of the actual label of the same training sample as lost information or shielded information under noise interference, recovering the lost information from the unlabelled electrocardiogram vector data of the training sample through a multi-layer noise reduction automatic encoder, obtaining the unlabelled electrocardiogram vector data of the training sample and the electrocardiogram vector data of the actual label of the same training sample at the last layer, but only using the electrocardiogram vector data of the actual label of the same training sample as output;

and S3, extracting the characteristics layer by the denoising stacking automatic encoder network and recovering the lost information. By extracting features layer by layer and recovering lost information, the detection accuracy of the network employing the noise reduction stacking automatic encoder can be improved.

Through the design of the embodiment, the invention can realize that the noise reduction stacking automatic encoder network comprises a plurality of layers, extracts the characteristics layer by layer and recovers the lost information, and can improve the detection precision. In this embodiment, the electrocardiographic vector data may also be quantitative index data of other multiple pathological features of the attributes related to the specific heart disease obtained in embodiment 2.

Example 15 detection method of myocardial infarction with complete left bundle branch block based on deep belief network

In this embodiment, the inventors have developed a method for detecting myocardial infarction with complete left bundle branch block based on a deep belief network, the method comprising:

step 1, obtaining quantitative data of dynamic pathological characteristics of a heart to be detected, index data of image characteristics of an electrocardiograph vector diagram or dynamic data of an electrocardiograph nonlinear system; acquiring quantitative data of the dynamic pathological characteristics of the heart, index data of the image characteristics of the vector cardiogram or dynamic data of the nonlinear electrocardiograph system corresponding to the quantitative data of the dynamic pathological characteristics of the heart to be detected, the index data of the image characteristics of the vector electrocardiograph or the dynamic data of the nonlinear electrocardiograph system; calculating a similarity value between the data to be detected and the verification data according to a content verification consistency identification algorithm;

step 2, comparing the similarity value with a preset threshold value, and determining the dynamic data of the electrocardio nonlinear system to be detected with the similarity value larger than the threshold value as initial credible dynamic data of the electrocardio nonlinear system;

step 3, extracting multi-dimensional credible characteristics of the dynamic data of the preliminary credible electrocardio nonlinear system, and constructing multi-dimensional credible characteristic vectors;

and 4, inputting the multi-dimensional credible feature vector into a Deep Belief Network (DBN) discrimination model, and outputting a discrimination result representing the credibility of the dynamic data of the preliminary credible electrocardio nonlinear system. For a Deep Belief Network (DBN) discrimination model, establishing a Restricted Boltzmann Machine (RBM) with a two-layer network structure and a DBN with a layer of BP neural network, respectively configuring a learning rate and a weight attenuation in the DBN as a learning rate with a weight updating amount as a self-adjusting range of 5-1 times of a weight, and a weight attenuation of a penalty function comprising a gradient term and 1/2 multiplied by a regularization coefficient, wherein the weight attenuation is obtained by multiplying a weight updating amount as a weight, determining a relation between an error rate and iteration times, and training the DBN by using text data for training in combination with the learning rate, the weight attenuation, the error rate and the iteration times. The limit boltzmann machine (RBM) can be further expanded to a deep boltzmann machine (RBM) to be applied to the embodiment.

It can be understood that the DBN discrimination model is a neural network structure obtained by training the RBM of the multilayer limited Boltzmann machine, and the model progressively extracts the characteristics of input data from a bottom layer to a high layer by adopting a mode of simulating a multilayer structure of a human brain, so that ideal characteristics suitable for mode classification are finally formed, and the classification accuracy is improved. After the multidimensional credible feature vectors are input into a DBN (deep belief network) discrimination model, the DBN discrimination model can quickly and accurately output discrimination results representing the credibility of the dynamic data of the preliminary credible electrocardio nonlinear system.

Example 16 detection method of myocardial infarction with complete left bundle branch block based on random forest algorithm

In a plurality of machine learning algorithms, a random forest is a method for distinguishing and classifying data by using a plurality of classification trees, and the random forest algorithm can process high-dimensional features, can give importance scores of all variables, evaluates the functions of all variables in classification and is easy to realize parallelization.

The inventor develops a detection method of myocardial infarction with complete left bundle branch block based on a random forest algorithm, which comprises the following steps:

step one, data acquisition: aiming at a target population (120 samples are arrhythmia patients, 120 samples are contrast blocking patients), quantitative data of heart dynamic pathological characteristics of cardiovascular and cerebrovascular disease patients in an observation period window, index data of image characteristics of an electrocardiograph, dynamic data of an electrocardiograph nonlinear system, quantitative data of ECG morphological indexes, biochemical data, human physiological information data and clinical information data are collected from a related information system; the acquired data also comprises quantitative index data of other multiple pathological characteristics of the related attributes of the specific heart diseases acquired in the embodiment 2;

step two, data preprocessing: the data preprocessing is used for carrying out a series of integration, cleaning and missing data processing on a data set, so that the data quality is improved;

step three, constructing a random forest prediction model:

(1) randomly putting back and sampling m samples from an original training set by using a Bootstrap method, and performing n _ tree times of sampling to generate n _ tree training sets;

(2) for n _ tree training sets, respectively training n _ tree decision tree models;

(3) for a single decision tree model, assuming that the number of training sample features is n, selecting the best feature to split according to the information gain/information gain ratio/the kini index during each splitting;

(4) each tree is split until all training examples of the node belong to the same class; pruning is not required during the splitting of the decision tree. And forming a random forest by the generated decision trees. For the classification problem, voting according to a plurality of tree classifiers to determine a final classification result; for the regression problem, the final prediction result is determined by the mean value of the predicted values of the multiple trees.

The inventor uses the method to carry out experimental verification, and finds that the accuracy of the random forest algorithm training is more than 85%; in some embodiments, the random forest classifier may perform similar processing instead of a decision tree algorithm; preferably, an extreme learning machine can be added to combine the hidden Markov models to optimize the classification results.

While several embodiments have been provided in the present disclosure, it should be understood that the disclosed systems and methods may be embodied in many other specific forms without departing from the spirit or scope of the present disclosure. These examples are to be considered as illustrative and not restrictive, and the intention is not to be limited to the details given herein. The various methods or ranges of parameters may be combined or integrated in another system, the various methods or ranges of parameters may be combined or integrated in another method, the various methods or ranges of parameters may be combined or integrated in a product of this or other fields, or certain features may be omitted or not implemented, all within the skill of the art in light of the ease with which these may be extended or implemented in accordance with the present disclosure.

Claims

1. A heart disease detection method based on an artificial intelligence model is characterized by comprising the following steps:

step S1, acquiring index data of multiple pathological characteristics of the specific heart disease, wherein the index data of the multiple pathological characteristics of the specific heart disease comprises one or more of quantitative data of dynamic pathological characteristics of the heart, index data of image characteristics of an electrocardiograph vector image and dynamic data of an electrocardiograph nonlinear system;

and step S2, performing machine learning on the index data of the multiple pathological features of the specific heart disease acquired in the step S1, and acquiring a machine learning judgment model of the specific heart disease.

2. The method for detecting heart diseases as claimed in claim 1, wherein the quantitative data of dynamic pathological features of heart in step S1 includes: one or more of quantitative index data of geometric characteristics, quantitative index data of nonlinear dynamic characteristics, quantitative index data of model characteristics, quantitative index data of time domain characteristics and quantitative index data of frequency domain characteristics;

preferably, the index data of multiple pathological features of the specific heart disease in step S1 further includes one or more of quantitative index data of auxiliary pathological features, physiological information data of human body, and clinical information data;

more preferably, the quantitative index data of the auxiliary pathological features comprise quantitative data of ECG morphological indexes, and/or index data of electrocardiogram image features, and/or biochemical data.

3. The method for detecting heart diseases according to claim 1, wherein the machine learning decision model for specific heart diseases at step S2 includes at least one of a support vector machine, a convolutional neural network, a recurrent neural network, a bayesian classifier, a K-nearest neighbor algorithm, a K-means algorithm, a linear regression, a logistic regression, a multivariate nonlinear regression fitting method, an Adaboost algorithm, a hidden markov model, an extreme learning machine, a random forest algorithm, a decision tree algorithm, a clustering algorithm, a generative pair network, a stacked autocoder, a fully connected network, an unsupervised pre-training network, a deep belief network, a deep boltzmann machine, and a neural tensor network;

preferably, the machine learning decision model for the specific heart disease of step S2 includes a machine learning classification model and a deep learning identification model; the data input into the machine learning classification model comprises quantitative data of the dynamic pathological features of the heart, and the data input into the deep learning identification model comprises index data of the image features of the vector cardiogram and dynamic data of a nonlinear electrocardiogram system; preferably, the data input into the machine learning classification model further includes one or more of quantitative index data of auxiliary pathological features, human physiological information data and clinical information data;

more preferably, the machine learning classification model is selected from at least one of a support vector machine, a bayesian classifier, a K-nearest neighbor algorithm, a K-means algorithm, a linear regression, a logistic regression, a multivariate nonlinear regression fitting method, an Adaboost algorithm, a hidden markov model, an extreme learning machine, a random forest algorithm, a decision tree algorithm, and a clustering algorithm, and the deep learning recognition model is selected from at least one of a convolutional neural network, a recurrent neural network, a generative confrontation network, a stacked autoencoder, a fully-connected network, an unsupervised pre-training network, a deep belief network, a deep boltzmann machine, and a neural tensor network.

4. The method for detecting cardiac diseases according to claim 1 or 3, further comprising step S3: acquiring index data of multiple pathological features of a person to be detected, and inputting the index data into the machine learning judgment model of the specific heart disease in the step S2 to obtain a detection result of the specific heart disease;

the index data of the multi-pathological characteristics of the person to be tested comprises one or more of quantitative data of dynamic pathological characteristics of the heart of the person to be tested, index data of image characteristics of an electrocardiograph vector diagram and dynamic data of an electrocardiograph nonlinear system.

5. The method for detecting heart diseases according to claim 4, wherein the index data of the multiple pathological features of the person to be tested in step S3 further includes one or more of quantitative index data of auxiliary pathological features of the person to be tested, physiological information data of a human body, and clinical information data;

preferably, the quantitative index data of the auxiliary pathological features of the person to be tested in step S3 includes quantitative data of ECG morphology indexes and/or biochemical data;

preferably, the detection result of the specific heart disease in step S3 further includes using the index data of the multiple pathological features of the person to be tested to respectively perform specific heart disease determination, so as to obtain an index determination result of the heart disease of the person to be tested;

more preferably, the method for detecting a heart disease further includes assigning a weight value to the output result of the machine learning determination model for the specific heart disease and the index determination result of the heart disease of the person to be detected in step S2 to perform heart disease detection, so as to obtain a comprehensive determination result of the heart disease of the person to be detected; the index judgment result of the heart disease of the person to be detected comprises one or more of judgment information of the specific heart disease of the quantitative data of the dynamic pathological features of the heart of the person to be detected, judgment information of the specific heart disease of the quantitative index data of the auxiliary pathological features of the person to be detected, judgment information of the specific heart disease of the human physiological information data of the person to be detected and judgment information of the specific heart disease of the clinical information data of the person to be detected;

more preferably, the output result of the machine learning judgment model for the specific heart disease in step S2 further includes an output result of the machine learning classification model and an output result of the deep learning identification model, and the output result of the machine learning classification model and the output result of the deep learning identification model are assigned with weight values.

6. The method for detecting heart diseases according to claim 4, wherein the step S3 is implemented by using quantitative data of dynamic pathological features of the heart of the person under test, including: the method comprises the following steps of obtaining quantitative index data of one or more of dynamic geometric characteristics, quantitative index data of nonlinear dynamic characteristics, quantitative index data of model characteristics, quantitative index data of time domain characteristics and quantitative index data of frequency domain characteristics of an electrocardio nonlinear system of a person to be tested.

7. A heart disease detection method based on an artificial intelligence model is characterized by comprising the following steps:

8. The method for detecting a cardiac disease according to claim 7, wherein the detection result of the specific cardiac disease in step 3 further includes threshold value determination information of the specific cardiac disease of the quantitative data of the dynamic pathological features of the heart of the person to be tested, and/or threshold value determination information of the specific cardiac disease of the quantitative index data of the auxiliary pathological features of the person to be tested, and/or determination information of the specific cardiac disease of the clinical information data of the person to be tested;

preferably, the detection result of the specific heart disease in step 3 further includes weighted value that is assigned to the output result information of the machine learning judgment model of the specific heart disease in step 3, and/or the judgment information of the specific heart disease of the quantitative data of the dynamic pathological features of the heart of the person to be measured, and/or the judgment information of the specific heart disease of the quantitative index data of the auxiliary pathological features of the person to be measured, and/or the judgment information of the specific heart disease of the human physiological information data of the person to be measured, and/or the judgment information of the specific heart disease of the clinical information data of the person to be measured.

9. A test product for heart diseases, which uses the method for testing heart diseases according to any one of claims 1to 8.

10. Use of a cardiac disorder detection product according to claim 9 for cardiac disorder detection.