CN109512423A - A kind of myocardial ischemia Risk Stratification Methods based on determining study and deep learning - Google Patents
A kind of myocardial ischemia Risk Stratification Methods based on determining study and deep learning Download PDFInfo
- Publication number
- CN109512423A CN109512423A CN201811487253.0A CN201811487253A CN109512423A CN 109512423 A CN109512423 A CN 109512423A CN 201811487253 A CN201811487253 A CN 201811487253A CN 109512423 A CN109512423 A CN 109512423A
- Authority
- CN
- China
- Prior art keywords
- layer
- myocardial ischemia
- neural networks
- convolutional neural
- follows
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
- A61B5/316—Modalities, i.e. specific diagnostic methods
- A61B5/318—Heart-related electrical modalities, e.g. electrocardiography [ECG]
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
- A61B5/316—Modalities, i.e. specific diagnostic methods
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
- A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
- A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7271—Specific aspects of physiological measurement analysis
- A61B5/7275—Determining trends in physiological measurement data; Predicting development of a medical condition based on physiological measurements, e.g. determining a risk factor
Abstract
The invention discloses a kind of based on the myocardial ischemia Risk Stratification Methods for determining study with deep learning.The present invention acquires conventional 12 lead electrocardiogram (ECG) signals, based on the inherent electrocardio dynamic characteristic progress neural net model establishing, identification for determining that the theories of learning contain shallow-layer electrocardiosignal, it obtains realizing the risk stratification to myocardial ischemia with the convolutional neural networks under deep learning frame in behavioral characteristics in electrocardiosignal.Determining learning dynamics modeling method and deep learning classification method has been used in combination in the method for the present invention for the first time, and is applied to the early stage risk stratification of myocardial ischemia, based on conventional 12 lead electrocardiosignals, is not required to add new detection device, simple and convenient, easy to operate.Determine that learning method extracts the behavioral characteristics more sensitive to ischemic state, and deep neural network can autonomous learning data characteristics, portrayed without carrying out further data, reduce the complexity of system.
Description
Technical field
The invention belongs to mode identification technologies, and in particular to a kind of to be lacked based on determining study and the cardiac muscle of deep learning
Blood Risk Stratification Methods.
Background technique
Myocardial ischemia is common cardiovascular disease, seriously threatens the life and health of the mankind.The risk stratification of myocardial ischemia
It is significant to the diagnosis, treatment and prognosis of the state of an illness, it is closely related with human health.Nearly half a century carrys out Medical Imaging skill
Art graduallys mature, become myocardial ischemia, coronary heart disease Clinics and Practices main flow direction.And in clinical practice, part is suffered from
Person's clinical manifestation is asymptomatic myocardial ischemia, although there is apparent coronary artery to change and the objective basis of myocardial ischemia, not companion
Have angina pectoris, be highly susceptible to the ignorance of patient itself He doctor, also become the significant risk of myocardial infarction and sudden death because
One of element.Image check, cannot be in general population screening extensively also due to the reason of itself price and the factor of patients ' psychological
It uses, diagnostic can not be played to greatest extent sometimes only according to image check.In contrast, non-invasive electrocardiology detection technique,
12 especially conventional lead electrocardiogram technologies, simple and practical, audient is wider, as the important supplement part of Medical Imaging,
It plays an important role on diagnosis of myocardial ischemia and risk stratification.Rationally, the noninvasive electrocardiology inspection of selection orderly can provide
Different from the ecg information of image information, the early detection of myocardial ischemia is realized, the mutual supplement with each other's advantages with imageological examination, further
Improve the accuracy rate and recall rate of clinical diagnosis.
Above-mentioned background and there are aiming at the problem that, in Chinese invention patent application: it is a kind of based on determine the theories of learning the heart
(application number: the cardiac muscle that a determining learning dynamics theory is proposed in 201310496628.0) lacks myocardial ischemia aided detection method
The new method of blood early diagnosis, so that original indiscoverable small morbid information is sufficiently indicated.However, this method is only given
The qualitative subjective results by figure whether at random are gone out, standard planning can not have been carried out to it using quantization means.In middle promulgated by the State Council
Bright patent application: a kind of electrocardio dynamics data quantitative analysis method (application number: proposes one kind in 201710587538.0)
Electrocardio dynamics data figure quantitative analysis method, to the above results carry out quantizating index extraction, but index extraction link still
There are sizable subjectivity, cannot provide exact index extraction quantity and index quantity and last diagnostic accuracy it
Between association.
Summary of the invention
It is of the existing technology the purpose of the present invention is overcoming the problems, such as, it provides a kind of based on determining study and deep learning
Myocardial ischemia Risk Stratification Methods, it is a kind of more it is simple it is accurate automatically, be suitble to clinical use based on determining study and depth
The myocardial ischemia Risk Stratification Methods of study.
The specific technical solution of the present invention is achieved by the steps of:
Step 1 obtains electrocardiosignal dynamics behavioral characteristics: being determined to the 12 lead electrocardiogram (ECG) datas collected
Dynamic modeling is practised, to obtain electrocardio dynamic characteristic data.
Step 2, data prediction: the electrocardio dynamic characteristic data of acquisition are normalized.
Step 3, building convolutional neural networks model: the structure of the convolutional neural networks of building is input layer-convolutional layer-pond
Change the full articulamentum-output layer of layer-convolutional layer-pond layer-.
Step 4, training convolutional neural networks: the data in the good training set of step 2 normalized are input to step 3
It is trained in the convolutional neural networks of building, trained process includes preceding to trained and Reverse optimization.
Step 5, Classification and Identification: the electrocardio dynamic characteristic data in the good test set of step 2 normalized are inputted
Into trained convolutional neural networks model, the high-risk individuals and low danger individual of test data set center myocardial ischemia are predicted.
In the above method, it is determined Learning Motive modeling described in step 1 and refers to by set generally acknowledged effective
It is a kind of special based on determining the myocardial ischemia aided detection methods of the theories of learning original electrocardiogram (ECG) data and being converted to electrocardio dynamics
Levy data.In electrocardiosignal dynamic modeling link, by determining the theories of learning, realized using RBF neural non-thread to electrocardio
Property the accurately identification of the dynamic part of system, and the knowledge learnt is saved in the form of constant value RBF neural weight
Come.The electrocardio dynamic mode of time-varying can be approached by the accurate neural network in the dynamic part of system to indicate.
The step of being determined Learning Motive modeling to electrocardiosignal is as follows: (1) by 12 lead electrocardiogram ECG numerical value numbers
According to conventional filtering is carried out, three-dimensional electrocardial vector diagram data is converted to by set generally acknowledged effective transformation law, is indicated are as follows:
V (t)=[vx(t),vy(t),vz(t)]T, t=1 in formula, m is sampling instant, then intercepts the ST- in three-dimensional data
T segment data, to extract ST-T ring.(2) dynamic RBF neural network identifier is used, to the built-in system dynamic benefit of ST-T ring
With determining that learning algorithm carries out the accurate RBF neural in part and approach, acquisition about in ECG ST-T segment signal power
Learn characteristic:
Wherein FST(V (t)) indicates built-in system dynamic characteristic,It is constant value neural network weight vector, S (V
It (t)) is Gaussian radial basis function.As shown in above formula, constant value weight matrix that obtained system dynamic knowledge preservesIt is exactly electrocardio dynamic characteristic data.
It is min- to the method that electrocardio dynamic characteristic data are normalized described in step 2 in the above method
Max method.
In the above method, the training of convolutional neural networks described in step 4 includes preceding to training and Reverse optimization two steps
Suddenly.
Detailed process is as follows for forward direction training:
1. convolution layer operation: formula expression are as follows:Wherein, l is the convolution number of plies,For l
Convolution kernel of j-th of the characteristic pattern of layer between l-1 layers of ith feature figure connection, MjFor the set of input data, b is each
The biasing of characteristic pattern is exported, " * " indicates that convolution algorithm, f () indicate excitation function,Indicate l j-th of characteristic pattern of layer.
2. pond layer operation: formula expression are as follows:It is adopted under down () expression in formula
Sample function,Indicate the corresponding coefficient of l j-th of characteristic pattern of layer,For its corresponding biasing term coefficient.
Reverse optimization is to carry out tuning using the mode of learning for having supervision, by making convolutional neural networks mould after tuning
The network weight of each layer of hidden layer in type and biasing can be optimal value.The specific calculating process of Reverse optimization is as follows:
(1) the gradient of convolutional layer calculates: formula expression are as follows:Wherein, u be convolution it
The characteristic pattern generated afterwards, up () indicate that up-sampling function, " * " indicate point multiplication operation.It, can for the characteristic pattern that convolutional layer gives
The gradient of bias term and the gradient of corresponding convolution kernel, calculation formula are corresponded in the hope of this feature figure are as follows: Wherein, J is cost function,Xiang ShiIn in convolution withBy the one of element multiplication
Block region, x, y indicate the coordinate in characteristic pattern.
(2) the gradient of pond layer calculates: the parameter involved in the forward process of the down-sampling of pond layer is each characteristic pattern
An a corresponding weight parameter w and bias term b, if acquiring the residual plot of this layer, the gradient of the two parameters is just easy to
It acquires, which can be expressed with formula are as follows:Rot180 in formula
() function representation rotates 180 degree counterclockwise, and parameter full indicates to carry out complete convolution operation.After obtaining residual plot, w and b
Calculating meet formula:Wherein,
In the above method, Classification and Identification described in step 5 refers to by the trained convolutional neural networks of step 4 to test
The electrocardio dynamic characteristic data of concentration carry out classification prediction, identify myocardial ischemia high-risk individuals and low danger individual.
Compared with the prior art, the invention has the following advantages and beneficial effects:
1, determining learning dynamics modeling and deep learning classification method has been used in combination in the method for the present invention for the first time, and is applied to
The early stage risk stratification of myocardial ischemia.It is special to the dynamic that myocardial ischemia is more sensitive to determine that learning dynamics modeling can be extracted
Sign, deep learning can autonomous learning data characteristics, portrayed without carrying out further data, reduce the complexity of system,
And do not need to increase new detection device, it is simple and convenient, it is easy to operate.
2, the present invention carries out the identification of myocardial ischemia risk stratification using deep learning method, deep compared to traditional method
Degree study learning ability is strong, can learn to better feature, there is better ability to express.This method is with amount of training data
Increase, better recognition effect can be obtained.
3, the present invention carries out myocardial ischemia risk stratification identification by building convolutional neural networks, can pass through training sample
This electrocardio dynamic characteristic data carry out autonomous learning, while the feature for having weight shared, reduce the complexity of model,
The operation of pond layer enhances the robustness of system.The present invention provides one several layers of convolutional neural networks mould in embodiment
Type has reached preferable recognition effect.
Detailed description of the invention
Fig. 1 is the flow chart of center of embodiment of the present invention myocardial ischemia Risk Stratification Methods.
Fig. 2 is the electrocardio dynamic characteristic data display figure of a certain myocardial ischemia high-risk individuals in the embodiment of the present invention.
Fig. 3 is the electrocardio dynamic characteristic data display figure of the low individual of endangering of a certain myocardial ischemia in the embodiment of the present invention.
Fig. 4 is the structural schematic diagram of convolutional neural networks model in the embodiment of the present invention.
Fig. 5 is that convolutional neural networks design parameter configures in the embodiment of the present invention.
Fig. 6 is the parameter setting of the full articulamentum of neural network in the embodiment of the present invention.
Fig. 7 is the classification recognition result of center of embodiment of the present invention electrodynamics characteristic.
Specific embodiment
A specific embodiment of the invention is described in further detail below with reference to examples and drawings, but the present invention
Implementation and protection scope it is without being limited thereto.
The method of the present invention is proposed based on the myocardial ischemia Risk Stratification Methods for determining study with deep learning, is combined for the first time
Determining learning dynamics modeling technique and current most fiery deep learning nerual network technique, first by determining the theories of learning pair
In nonlinear system dynamic implement neural net model establishing in electrocardiosignal, convolutional neural networks algorithm is then utilized, it is dynamic to electrocardio
Mechanical characteristics data carry out identification classification, identify myocardial ischemia high-risk individuals and low danger individual.The method of the present invention is suitable for the heart
The automatic no tagsort identification of electrodynamics signal, the risk stratification applied to myocardial ischemia.Especially in general population,
High-risk result can cause the further concern of clinician and individual itself, and potential high-risk patient is guided to be further examined.
And low danger result can then exclude to reduce unnecessary medical resource wave because of the fear of patients ' psychological caused by non-cardiogenic symptom
Take.
Embodiment
Specific embodiments of the present invention choose the electrocardiogram (ECG) data of hospital clinical acquisition.
As shown in Figure 1, the flow chart of the myocardial ischemia Risk Stratification Methods for the embodiment of the present invention, comprising the following steps:
It obtains electrocardiosignal dynamics behavioral characteristics: learning dynamics is determined to the 12 lead electrocardiogram (ECG) datas collected
Modeling, to obtain electrocardio dynamic characteristic data.The operation of dynamic modeling is as described in illustrating step.Electrocardio dynamic characteristic number
Neural net model establishing and holographic characteristic are carried out in nonlinear system dynamic in electrocardiogram (ECG) data by the determining theories of learning according to referring to
It extracts, and the constant value weight matrix that the system dynamic knowledge acquired is preserved.If by electrocardio dynamic characteristic data into
Row three-dimensional visualization is shown, is the electrocardio dynamic characteristic data display figure of myocardial ischemia high-risk individuals, such as Fig. 3 as shown in Figure 2
Shown is the electrocardio dynamic characteristic data display figure of the low danger individual of myocardial ischemia.
Data prediction: the numerical value difference in order to eliminate electrocardio dynamic characteristic data, it is special to the electrocardio dynamics of acquisition
Sign data are normalized, and used method is min-max method.Its calculation method is as follows: assuming that for sequence x1,
x2,···,xnIt is converted:Then new sequences y1,y2,···,yn∈[0,1]。
Construct convolutional neural networks model: the special networks structure that convolutional neural networks are shared with its local weight is in mode
There is unique superiority in identification field.As shown in figure 4, the convolutional neural networks model structure of building is mainly inputted by one
Layer, several hidden layers and an output layer composition.Wherein, hidden layer includes convolutional layer, pond layer and full articulamentum.In convolution mind
In single convolutional layer through network, several characteristic planes are generally included, same characteristic plane shares a set of weight filter.?
In one pond layer of convolutional neural networks, usually there are mean value pond (Mean Pooling) and maximum value pond (Max
Pooling) two kinds of forms.In embodiments of the present invention, the pond layer of convolutional neural networks all uses mean value pondization to operate.Convolution
The convolution sum pondization operation of neural network simplifies model complexity, reduces model parameter number.
For the detection classification of myocardial ischemia, the inputs of convolutional neural networks is electrocardio dynamics sample data X, in network
I-th layer of characteristic pattern is Xi.If i-th layer is convolutional layer, XiIt can indicate are as follows: Vi=f (Xi-1*Wi+bi)；Wherein, WiIt is i-th
The weight vector of layer, * is convolution operation, biFor i-th layer of bias vector, f () is excitation function.In embodiments of the present invention
The activation primitive used is Relu function, expression formula are as follows: f (x)=max (0, x).If i-th layer is pond layer, XiIt can be with table
It is shown as: Vi=mean [Xi-1(n × s+r)], wherein r is the size of pond layer window, and s is the step-length of pond layer, and n is characteristic pattern
Element index, function mean () be used for calculates merging pond window average value.
In order to which using all features extracted, the output of pond layer is passed in flatten layers, the input of multidimensional
One-dimensional, the feature vector of one-dimensional may be expressed as: X=[X1,X2,···,Xi], wherein X is global characteristics vector, and is made
For the input of full articulamentum.Matrix multiplication is executed in full articulamentum, C element vector of final output, wherein C is the number of classification
Amount.In addition, in the final prediction probability for calculating each class using softmax classifier.The operation of full articulamentum may be expressed as:Wherein WcIt is the weight vectors being connect with c-th of neuron of output layer,It is that final output vector is (pre-
Mark label).It is exported in above formulaValue is the positive number from 0 to 1, is expressed as the prediction probability of target category.
In embodiments of the present invention, the convolutional neural networks of building include that 2 convolutional layers and 2 pond layers and one are complete
Articulamentum, as shown in figure 5, for the design parameter setting of convolutional neural networks in embodiment, as shown in fig. 6, to be complete in embodiment
The design parameter of articulamentum is arranged.
Training convolutional neural networks: it is necessary to being trained to network structure after the structure of convolutional neural networks is put up.
Trained purpose is that optimizing network parameter obtains optimal weight and bias, followed by these trained weights and partially
It sets value to configure neural network, the neural network being finally configured is able to carry out the knowledge of the myocardial ischemia with certain accuracy rate
Do not classify.
The training of convolutional neural networks includes preceding to trained and two steps of Reverse optimization.The process of forward direction training is as follows:
(a) convolution layer operation: formula expression are as follows:Wherein, l is the convolution number of plies,It is j-th of l layer
Convolution kernel of the characteristic pattern between l-1 layers of ith feature figure connection, MjFor the set of input data, b is each output feature
The biasing of figure, " * " indicate that convolution algorithm, f () indicate excitation function,Indicate l j-th of characteristic pattern of layer.(b) pond layer is grasped
Make: formula expression are as follows:Down () indicates down-sampling function in formula.Reverse optimization is to adopt
Tuning is carried out with the mode of learning for having supervision, by the net for making each layer of hidden layer in convolutional neural networks model after tuning
Network weight and biasing can be optimal value.The specific calculating process of Reverse optimization is as follows: (a) gradient of convolutional layer calculates:
Formula expression are as follows:Wherein, u is the characteristic pattern generated after convolution, in up () expression
Sampling function, " * " indicate point multiplication operation.For the characteristic pattern that convolutional layer gives, bias term can be corresponded in the hope of this feature figure
The gradient of gradient and corresponding convolution kernel, calculation formula are as follows:Wherein, J is generation
Valence function,Xiang ShiIn in convolution withBy one piece of region of element multiplication, x, y indicate the coordinate in characteristic pattern.
(b) gradient of pond layer calculates: the parameter involved in the forward process of the down-sampling of pond layer is that each characteristic pattern is corresponding
An one weight parameter w and bias term b, if acquiring the residual plot of this layer, the gradient of the two parameters is just easy to acquire, should
Process can be expressed with formula are as follows:After obtaining residual plot, w and b
Corresponding gradient calculating meets formula: Wherein,
The back-propagation algorithm used in embodiments of the present invention is stochastic gradient descent method (SGD, Stochastic
Gradient Descent).Training process is completed based on the iteration that forward and backward is transmitted, and defines each of training data concentration
The size of batch, the loss function of SGD is defined as:Wherein, Q is instruction
Practice the batch size concentrated and select each sample size, Y is required output vector (true tag).Above formula loss function also by
Referred to as intersect entropy function, commonly used a kind of effective loss function calculation formula in classification task.θ is allowed to indicate in model
Can training parameter, such as weight and biasing, then the update rule of parameter may be defined as:Wherein, J
(θ) indicates loss cost function, and η is learning rate,It is gradient operator, t indicates training step.In the training process, hyper parameter
(such as learning rate, small lot size, frequency of training) all has a certain impact to training effect, needs repetition test to adjust them
To reach optimal experiment effect.
In embodiments of the present invention, for 20, Study rate parameter is set as the back-propagation algorithm training batch used
0.001,500 training have been carried out in total, so that model obtains better training effect.
Classification and Identification: it in building with after training for completion convolutional neural networks, needs to collected electrocardio dynamics
Characteristic carries out Classification and Identification, verifies the recognition performance of the myocardial ischemia Risk Stratification Methods of the embodiment of the present invention.
In embodiments of the present invention, learning model building is determined to hospital's electrocardiogram (ECG) data of acquisition, obtains 1524 hearts altogether
The characteristic of electrodynamics characteristic wherein Healthy People has 116, and the characteristic with myocardial ischemia has 1408.?
It is used to training convolutional neural networks with 75% (1143) characteristic of total data in embodiment, the 25% (381 of total data
Item) characteristic is used to be tested, wherein including 37 Healthy People characteristics and 344 myocardial ischemia characteristics.?
The confusion matrix for the myocardial ischemia risk stratification tested in the embodiment of the present invention is as shown in Figure 7.
In order to verify the validity of the method for the present invention, carry out assessment models for the heart using sensitivity, specificity and accuracy rate
The performance of myocardial ischemia detection.Sensitivity refers to the percentage for actually having disease correctly classified by model, specificity refer to by
The percentage without disease that model is correctly classified, accuracy rate refer to the percentage correctly classified by model.Sensitivity, specificity
It is defined as follows with accuracy rate:
Wherein, TP, TN, FP, FN respectively represent true positives, true negative, the quantity of false positive and false negative.
In embodiments of the present invention by the experiment to electrocardio dynamic characteristic data, obtained sensitivity, specificity and
Accuracy rate is respectively 90.98%, 75%, 90.83%.Can be seen that the method for the present invention by the experiment of embodiment can be effectively
The discrimination for improving myocardial ischemia detection is capable of providing quick recognition methods especially in the case where mass data collection.It is real
It applies example and demonstrates the validity of the method for the present invention and the reliability of testing result, recognition result of the invention can be used as cardiovascular disease
The important references of disease diagnosis, provide quickly and effectively tool for the diagnosis state of an illness.Above-described embodiment is that the present invention is preferably real
Mode is applied, but embodiment of the present invention are not limited by the above embodiments, it is other any without departing from spirit of the invention
Essence and made changes, modifications, substitutions, combinations, simplifications under principle, should be equivalent substitute mode, are included in this hair
Within bright protection scope.
Claims (8)
1. a kind of based on the myocardial ischemia Risk Stratification Methods for determining study with deep learning, which is characterized in that include following step
It is rapid:
Step 1 obtains electrocardiosignal dynamics behavioral characteristics: it is dynamic to be determined study to the 12 lead electrocardiogram (ECG) datas collected
State modeling, to obtain electrocardio dynamic characteristic data；
Step 2, data prediction: the electrocardio dynamic characteristic data of acquisition are normalized；
Step 3, building convolutional neural networks model: the structure of the convolutional neural networks of building is input layer-convolutional layer-pond
Full articulamentum-the output layer of layer-convolutional layer-pond layer-；
Step 4, training convolutional neural networks: the data in the good training set of step 2 normalized are input to step 3 and are constructed
Convolutional neural networks in be trained, trained process includes preceding to trained and Reverse optimization；
Step 5, Classification and Identification: the electrocardio dynamic characteristic data in the good test set of step 2 normalized are input to instruction
In the convolutional neural networks model perfected, the high-risk individuals and low danger individual of test data set center myocardial ischemia are predicted.
2. according to claim 1 based on the myocardial ischemia Risk Stratification Methods for determining study with deep learning, feature
It is, Learning Motive modeling is determined described in step 1 and is referred to through set generally acknowledged effective one kind based on determining
It practises theoretical myocardial ischemia aided detection method and original electrocardiogram (ECG) data is converted to electrocardio dynamic characteristic data.
3. it is according to claim 1 or 2 based on the myocardial ischemia Risk Stratification Methods for determining study with deep learning, it is special
Sign is, is min-max method to the method that characteristic is normalized described in step 2.
4. according to claim 3 based on the myocardial ischemia Risk Stratification Methods for determining study with deep learning, feature
It is, the training of convolutional neural networks described in step 4 includes preceding to trained and two steps of Reverse optimization, wherein forward direction training
Detailed process is as follows:
1. convolution layer operation: formula expression are as follows:Wherein, l is the convolution number of plies, and k is convolution kernel, Mj
For the set of input data, b is the biasing of each output characteristic pattern, and " * " indicates convolution algorithm；
2. pond layer operation: formula expression are as follows:Down () indicates down-sampling letter in formula
Number.
5. according to claim 4 based on the myocardial ischemia Risk Stratification Methods for determining study with deep learning, feature
It is, the Reverse optimization of convolutional neural networks described in step 4 is to carry out tuning using the mode of learning for having supervision, passes through tuning
The network weight and biasing for making each layer of hidden layer in convolutional neural networks model later can be optimal value, reversed excellent
It is as follows to change specific calculating process:
(1) the gradient of convolutional layer calculates: formula expression are as follows:Wherein, u is production after convolution
Raw characteristic pattern, up () indicate that up-sampling function, " * " indicate point multiplication operation；For the characteristic pattern that convolutional layer gives, Neng Gouqiu
It obtains this feature figure and corresponds to the gradient of bias term and the gradient of corresponding convolution kernel, calculation formula are as follows:Wherein, J is cost function,Xiang ShiIn in convolution withBy
One piece of region of element multiplication, x, y indicate the coordinate in characteristic pattern；
(2) the gradient of pond layer calculates: the parameter involved in the forward process of the down-sampling of pond layer is that each characteristic pattern is corresponding
A weight parameter w and a bias term b, if acquiring the residual plot of this layer, the gradient of weight parameter w and bias term b are just very
It is easy to acquire, which is expressed with formula are as follows:Obtaining residual plot
Afterwards, the corresponding gradient calculating of weight parameter w and bias term b meets formula:Its
In,
6. according to claim 5 based on the myocardial ischemia Risk Stratification Methods for determining study with deep learning, feature
It is, Classification and Identification described in step 5 refers to by the trained convolutional neural networks of step 4 to the electrocardio power in test set
It learns characteristic and carries out classification prediction, identify myocardial ischemia high-risk individuals and low danger individual.
7. according to claim 5 based on the myocardial ischemia Risk Stratification Methods for determining study with deep learning, feature
It is, the specific calculating of used min-max method is as follows: assuming that for sequence x1,x2,…,xnIt is converted:Then new sequences y1,y2,…,yn∈[0,1]。
8. according to claim 7 based on the myocardial ischemia Risk Stratification Methods for determining study with deep learning, feature
It is the input of convolutional neural networks for electrocardio dynamics sample data X, i-th layer of characteristic pattern is X in networki；
If i-th layer is convolutional layer, XiIt indicates are as follows: Vi=f (Xi-1*Wi+bi)；Wherein, WiFor i-th layer of weight vector, * is volume
Product operation, biFor i-th layer of bias vector, f () is activation primitive, the expression formula of activation primitive are as follows: f (x)=max (0, x)；
If i-th layer is pond layer, XiIt indicates are as follows: Vi=mean [Xi-1(n × s+r)], wherein r is the size of pond layer window, s
It is the step-length of pond layer, n is the element index of characteristic pattern, and function mean () is used to calculate the average value for merging pond window,
Pond layer window r is equal to step-length s.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811487253.0A CN109512423A (en) | 2018-12-06 | 2018-12-06 | A kind of myocardial ischemia Risk Stratification Methods based on determining study and deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811487253.0A CN109512423A (en) | 2018-12-06 | 2018-12-06 | A kind of myocardial ischemia Risk Stratification Methods based on determining study and deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109512423A true CN109512423A (en) | 2019-03-26 |
Family
ID=65795052
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811487253.0A Pending CN109512423A (en) | 2018-12-06 | 2018-12-06 | A kind of myocardial ischemia Risk Stratification Methods based on determining study and deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109512423A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110522444A (en) * | 2019-09-03 | 2019-12-03 | 西安邮电大学 | A kind of electrocardiosignal method for identifying and classifying based on Kernel-CNN |
| CN110537907A (en) * | 2019-08-26 | 2019-12-06 | 华南理工大学 | Electrocardiosignal compression and identification method based on singular value decomposition |
| CN110929847A (en) * | 2019-11-15 | 2020-03-27 | 国网浙江省电力有限公司电力科学研究院 | Converter transformer fault diagnosis method based on deep convolutional neural network |
| CN110974214A (en) * | 2019-12-20 | 2020-04-10 | 华中科技大学 | Automatic electrocardiogram classification method, system and equipment based on deep learning |
| CN111128298A (en) * | 2019-12-24 | 2020-05-08 | 大连海事大学 | Method and system for obtaining multi-gene risk scores based on deep learning model |
| CN112067916A (en) * | 2019-09-20 | 2020-12-11 | 武汉理工大学 | Time series data intelligent fault diagnosis method based on deep learning |
| CN112633195A (en) * | 2020-12-28 | 2021-04-09 | 杭州电子科技大学 | Myocardial infarction identification and classification method based on frequency domain features and deep learning |
| CN112932535A (en) * | 2021-02-01 | 2021-06-11 | 杜国庆 | Medical image segmentation and detection method |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20040059203A1 (en) * | 1998-10-05 | 2004-03-25 | Guerrero Juan R. | Method and system for analysis of biological signals such as dynamic electrocardiograms and the like |
| CN103549949A (en) * | 2013-10-21 | 2014-02-05 | 华南理工大学 | Myocardial ischemia auxiliary detecting method based on deterministic learning theory |
| US20150374300A1 (en) * | 2014-06-27 | 2015-12-31 | The Regents Of The University Of Michigan | Early detection of hemodynamic decompensation using taut-string transformation |
| CN107260161A (en) * | 2017-07-18 | 2017-10-20 | 上海图灵医疗科技有限公司 | A kind of electrocardio dynamics data quantitative analysis method |
| CN108427921A (en) * | 2018-02-28 | 2018-08-21 | 辽宁科技大学 | A kind of face identification method based on convolutional neural networks |
| US20180235497A1 (en) * | 2014-06-05 | 2018-08-23 | Guangren CHEN | Accurate ECG I-Point and J-Joint Detection and Measurement |
-
2018
- 2018-12-06 CN CN201811487253.0A patent/CN109512423A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20040059203A1 (en) * | 1998-10-05 | 2004-03-25 | Guerrero Juan R. | Method and system for analysis of biological signals such as dynamic electrocardiograms and the like |
| CN103549949A (en) * | 2013-10-21 | 2014-02-05 | 华南理工大学 | Myocardial ischemia auxiliary detecting method based on deterministic learning theory |
| US20180235497A1 (en) * | 2014-06-05 | 2018-08-23 | Guangren CHEN | Accurate ECG I-Point and J-Joint Detection and Measurement |
| US20150374300A1 (en) * | 2014-06-27 | 2015-12-31 | The Regents Of The University Of Michigan | Early detection of hemodynamic decompensation using taut-string transformation |
| CN107260161A (en) * | 2017-07-18 | 2017-10-20 | 上海图灵医疗科技有限公司 | A kind of electrocardio dynamics data quantitative analysis method |
| CN108427921A (en) * | 2018-02-28 | 2018-08-21 | 辽宁科技大学 | A kind of face identification method based on convolutional neural networks |
Non-Patent Citations (2)
| Title |
|---|
| XUNDE DONG ET AL: "ECG beat classiÞcation via deterministic learning", 《NEUROCOMPUTING》 * |
| 王星等: "时频分析在心率变异性研究中的应用", 《生物医学工程研究》 * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110537907A (en) * | 2019-08-26 | 2019-12-06 | 华南理工大学 | Electrocardiosignal compression and identification method based on singular value decomposition |
| CN110522444A (en) * | 2019-09-03 | 2019-12-03 | 西安邮电大学 | A kind of electrocardiosignal method for identifying and classifying based on Kernel-CNN |
| CN110522444B (en) * | 2019-09-03 | 2022-03-25 | 西安邮电大学 | Electrocardiosignal identification and classification method based on Kernel-CNN |
| CN112067916A (en) * | 2019-09-20 | 2020-12-11 | 武汉理工大学 | Time series data intelligent fault diagnosis method based on deep learning |
| CN110929847A (en) * | 2019-11-15 | 2020-03-27 | 国网浙江省电力有限公司电力科学研究院 | Converter transformer fault diagnosis method based on deep convolutional neural network |
| CN110974214A (en) * | 2019-12-20 | 2020-04-10 | 华中科技大学 | Automatic electrocardiogram classification method, system and equipment based on deep learning |
| CN111128298A (en) * | 2019-12-24 | 2020-05-08 | 大连海事大学 | Method and system for obtaining multi-gene risk scores based on deep learning model |
| CN111128298B (en) * | 2019-12-24 | 2022-12-02 | 大连海事大学 | Method and system for obtaining multi-gene risk scores based on deep learning model |
| CN112633195A (en) * | 2020-12-28 | 2021-04-09 | 杭州电子科技大学 | Myocardial infarction identification and classification method based on frequency domain features and deep learning |
| CN112633195B (en) * | 2020-12-28 | 2024-02-13 | 杭州电子科技大学 | Myocardial infarction recognition and classification method based on frequency domain features and deep learning |
| CN112932535A (en) * | 2021-02-01 | 2021-06-11 | 杜国庆 | Medical image segmentation and detection method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109512423A (en) | A kind of myocardial ischemia Risk Stratification Methods based on determining study and deep learning | |
| JP6522161B2 (en) | Medical data analysis method based on deep learning and intelligent analyzer thereof | |
| Deng et al. | Extracting cardiac dynamics within ECG signal for human identification and cardiovascular diseases classification | |
| Sahu et al. | FINE_DENSEIGANET: Automatic medical image classification in chest CT scan using Hybrid Deep Learning Framework | |
| TWI701681B (en) | Prediction model of atrial fibrillation and prediction system thereof | |
| Xiao et al. | Heart coronary artery segmentation and disease risk warning based on a deep learning algorithm | |
| CN111759345A (en) | Heart valve abnormality analysis method, system and device based on convolutional neural network | |
| CN112906748A (en) | 12-lead ECG arrhythmia detection classification model construction method based on residual error network | |
| CN109978873A (en) | A kind of intelligent physical examination system and method based on Chinese medicine image big data | |
| Herath et al. | Autism spectrum disorder diagnosis support model using Inception V3 | |
| CN111462082A (en) | Focus picture recognition device, method and equipment and readable storage medium | |
| Nagadeepa et al. | Artificial Intelligence based Cervical Cancer Risk Prediction Using M1 Algorithms | |
| Chourasia et al. | ECG heartbeat classification using CNN | |
| CN116864140A (en) | Intracardiac branch of academic or vocational study postoperative care monitoring data processing method and system thereof | |
| Phankokkruad | Evaluation of Deep transfer learning models in glaucoma detection for clinical application | |
| CN111329467A (en) | Heart disease auxiliary detection method based on artificial intelligence | |
| Shaheamlung et al. | A comprehensive review of medical expert systems for diagnosis of chronic liver diseases | |
| Kusuma et al. | A Hybrid Model for Skin Disease Classification using Transfer Learning | |
| Selvathi et al. | Deep convolutional neural network-based diabetic eye disease detection and classification using thermal images | |
| Sinha et al. | iHyptn: Predicting Hypertension using PPG signal for Cardiovascular Disease with Machine Learning Models | |
| Chowdary et al. | Multiple Disease Prediction by Applying Machine Learning and Deep Learning Algorithms | |
| CN108346471A (en) | A kind of analysis method and device of pathological data | |
| Viswapriya et al. | A Systematic Method of Stroke Prediction Model based on Big Data and Machine Learning | |
| Duwairi et al. | A deep learning model and a dataset for diagnosing ophthalmology diseases | |
| Wang et al. | Diagnosis of heart disease via cnns |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190326 |
|
| RJ01 | Rejection of invention patent application after publication |