CN108113647A - A kind of electrocardiosignal sorter and method - Google Patents
A kind of electrocardiosignal sorter and method Download PDFInfo
- Publication number
- CN108113647A CN108113647A CN201611067279.0A CN201611067279A CN108113647A CN 108113647 A CN108113647 A CN 108113647A CN 201611067279 A CN201611067279 A CN 201611067279A CN 108113647 A CN108113647 A CN 108113647A
- Authority
- CN
- China
- Prior art keywords
- layer
- electrocardiosignal
- grader
- training
- sparse
- 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
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
-
- 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]
- A61B5/346—Analysis of electrocardiograms
- A61B5/349—Detecting specific parameters of the electrocardiograph cycle
- A61B5/352—Detecting R peaks, e.g. for synchronising diagnostic apparatus; Estimating R-R interval
-
- 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
Landscapes
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Engineering & Computer Science (AREA)
- Surgery (AREA)
- Animal Behavior & Ethology (AREA)
- Cardiology (AREA)
- Veterinary Medicine (AREA)
- Physics & Mathematics (AREA)
- Public Health (AREA)
- Biophysics (AREA)
- Pathology (AREA)
- Biomedical Technology (AREA)
- Heart & Thoracic Surgery (AREA)
- Medical Informatics (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Physiology (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Signal Processing (AREA)
- Psychiatry (AREA)
- Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
Abstract
The present invention relates to electrocardiosignal sorting technique field, more particularly to a kind of electrocardiosignal sorter and method.The electrocardiosignal sorter includes：Data input module：For the normal data of the clinical data of acquisition and MIT databases to be input in parameter training module；Parameter training module：For extracting the electrocardiosignal feature of the clinical data, the sparse own coding algorithm parameter of training multilayer, and classifier parameters are returned with the normal data training of the MIT databases；Categorised decision module：For according to the sparse own coding algorithm parameter of the multilayer and recurrence classifier parameters structure electrocardiosignal grader, passing through the electrocardiosignal grader and carrying out electrocardiosignal classification.The present invention is with clinical data when training sample makes classification results have more authenticity for clinical diagnosis closer to the situation of actual clinical patient；By the deep learning for realizing mass data the nicety of grading of grader grader is trained to greatly improve.
Description
Technical field
The present invention relates to electrocardiosignal sorting technique field, more particularly to a kind of electrocardiosignal sorter and method.
Background technology
As the ratio of heart class sickness influence crowd is higher and higher, base of the electrocardiosignal classification as Diagnosing Cardiac disease
This step is come into being with corresponding demand.Present clinically most common electrocardiosignal sorting technique is all based on greatly time-domain
The signature analysis of frequency domain, the signature analysis (such as R -- R interval, P ripples lack) of waveform shape, heart variability it is short-term
The methods of analyzing (HRV) is come rhythm abnormality signal of classifying.And with the development of science and technology, for example, deep learning, intensified learning,
Data mining and big data etc. it is burning hot, on the basis of conventional method combining above-mentioned technology perhaps can largely promote
Into the development of electrocardiosignal sorting technique.
Nearly all method is all followed by first feature selecting and feature extraction in all electrocardiosignal sorting techniques, so
Classification decision-making afterwards is realized.Such as have plenty of the shape feature based on waveform, have plenty of the frequency domain character based on waveform, also have base
It is based on neutral net and based on support vector machines etc. in wavelet analysis.
Relevant references on electrocardiosignal classification include：
[1]Thanapatay D,Suwansaroj C,Thanawattano C.ECG beat classification
method for ECG printout with Principle Components Analysis and Support Vector
Machines[C].Electronics and Information Engineering(ICEIE),2010International
Conference On.IEEE,2010,1:V1-72-V1-75.
[2]Kallas M,Francis C,Kanaan L,et al.Multi-class SVM classification
combined with kernel PCA feature extraction of ECG signals[C]
.Telecommunications(ICT),2012 19th International Conference on.IEEE,2012:1-5.
[3]Vaidotas Marozas；Leif Sornmo；Arunas Lukosevicius.An Echo State
Neural Network for QRST Cancellation during Atrial Fibrillation[J].IEEE
Transactions on Biomedical Engineering,2012,59(10):2950–2957.
[4]Jambukia S H,Dabhi V K,Prajapati H B.Classification of ECG signals
using machine learning techniques:A survey[C].Computer Engineering and
Applications(ICACEA),2015International Conference on Advances in.IEEE,2015:
714-721.
Author is used with reference to principal component analysis (PCA) and support vector machines (SVM) structure classification in article [1] and [2]
In preprocessing part with PCA, a support vector machine classifier reality is built to different classes of come electrocardiosignal of classifying respectively for device
Now more classification finally achieve good classification results, but the sample analysis for having mark can only be utilized with support vector machines, right
There is no good handling result in the data of clinical patient.The classification side of block-based neutral net has been used in article [3]
Method realizes the Classification and Identification of ECG (electrocardiogram, electrocardiogram) signal, and author is by building with four in this method
A neuron is the nerve network system of a block, by the weight coefficient of propagated forward and gradient descent method learning network, is learned
Neuron input/output structure during habit in block is variable, the structure to electrocardiosignal classification is finally trained, although method
Relatively new but also can only be to there is labeled data to be trained, and nicety of grading is not high enough, particularly specificity be not high enough.
In conclusion existing electrocardiosignal sorting technique has a drawback in that：
(1) precision is not high in actual use for the sorting technique of traditional time-domain and frequency-domain, it is impossible to meet the need of practical application
It asks, it is difficult to meet the process demand in the case of mass data；
(2) existing supervised learning method cannot be used for grader instruction for a large amount of no label electrocardiogram (ECG) datas of clinical acquisitions
Practice, it is limited using traditional supervised learning training data, it is impossible to the individual difference in effective adaptation application；
(3) the existing method for electrocardiosignal classification is mostly higher to the discrimination of normal type, and for quantity
Less electrocardio beat classification of type precision is very poor.
The content of the invention
The present invention provides a kind of electrocardiosignal sorter and methods, it is intended to solve existing skill at least to a certain extent
One of above-mentioned technical problem in art.
To solve the above-mentioned problems, the present invention provides following technical solutions：
A kind of electrocardiosignal sorter, including：
Data input module：For the normal data of the clinical data of acquisition and MIT databases to be input to parameter training
In module；
Parameter training module：For extracting the electrocardiosignal feature of the clinical data, the sparse own coding of training multilayer is calculated
Method parameter, and return classifier parameters with the normal data training of the MIT databases；
Categorised decision module：For according to the sparse own coding algorithm parameter of the multilayer and the recurrence classifier parameters structure heart
Electric signal grader carries out electrocardiosignal classification by the electrocardiosignal grader.
The technical solution that the embodiment of the present invention is taken further includes：The data input module is additionally operable to：By the clinic of acquisition
The normal data of data and MIT databases carries out denoising, goes baseline, normalization pretreatment, and extracts using R ripples as basic point each
The waveform in R cycles.
The technical solution that the embodiment of the present invention is taken further includes：The sparse own coding of multilayer includes an input layer and three layers hidden
Layer is hidden, the parameter training module training sparse own coding algorithm parameter of multilayer specifically includes：At the beginning of each parameter of random initializtion
Initial value is successively trained by the algorithm of sparse own coding, and waveform is intercepted 340 points as ripple to be trained using centered on R ripples
Section selects the number of nodes of input layer and three layers of hidden layer, cost function is constantly minimized by gradient descent method respectively, determines
The weight coefficient of input layer and first layer hidden layer, using the output characteristic of first layer hidden layer as the input of second layer hidden layer
Feature, the weight coefficient between training first layer hidden layer and second layer hidden layer, and with the output characteristic of second layer hidden layer
Input feature vector as third layer hidden layer trains the weight coefficient between second layer hidden layer and third layer hidden layer, completes institute
State the training of the sparse own coding parameter of multilayer.
The technical solution that the embodiment of the present invention is taken further includes：The categorised decision module includes softmax and returns classification
Device, the softmax return the instruction that grader returns grader by the use of the normal data of a part of MIT databases as softmax
Practice data, grader is returned using the output characteristic of the third layer hidden layer of the sparse own coding algorithm of the multilayer as softmax
Input layer, according to normal data mark training softmax return grader input layer and output layer between parameter, together
When add in phase, S-T segment, R-wave amplitude between the wave character RR of extraction, obtain trained softmax and return grader.
The technical solution that the embodiment of the present invention is taken further includes：The categorised decision module further includes fine-adjusting unit, described
The fine tuning algorithm of fine-adjusting unit backpropagation, the entire softmax of fine tuning successively return grader, obtain final
Softmax returns grader, and exports electrocardiosignal classification results.
Another technical solution that the embodiment of the present invention is taken is：A kind of electrocardiosignal sorting technique, comprises the following steps：
Step a：The normal data of the clinical data of acquisition and MIT databases is input in parameter training module；
Step b：The electrocardiosignal feature of the clinical data, the sparse own coding algorithm parameter of training multilayer are extracted, and is transported
Classifier parameters are returned with the normal data training of the MIT databases；
Step c：According to the sparse own coding algorithm parameter of the multilayer and return classifier parameters structure electrocardiosignal classification
Device carries out electrocardiosignal classification by the electrocardiosignal grader.
The technical solution that the embodiment of the present invention is taken further includes：The step a is further included：By the clinical data of acquisition and
The normal data of MIT databases carries out denoising, goes baseline, normalization pretreatment, and extracts each R cycles by basic point of R ripples
Waveform.
The technical solution that the embodiment of the present invention is taken further includes：In the step b, the sparse own coding of multilayer includes
One input layer and three layers of hidden layer, the trained sparse own coding algorithm parameter of multilayer specifically include：Random initializtion is respectively joined
Several initial values is successively trained by the algorithm of sparse own coding, and waveform is intercepted using centered on R ripples to 340 points as waiting to instruct
Experienced wave band selects the number of nodes of input layer and three layers of hidden layer, cost letter is constantly minimized by gradient descent method respectively
Number, is determined the weight coefficient of input layer and first layer hidden layer, is hidden using the output characteristic of first layer hidden layer as the second layer
The input feature vector of layer, the weight coefficient between training first layer hidden layer and second layer hidden layer, and with second layer hidden layer
Output characteristic trains the weight system between second layer hidden layer and third layer hidden layer as the input feature vector of third layer hidden layer
Number, completes the training of the sparse own coding parameter of the multilayer.
The technical solution that the embodiment of the present invention is taken further includes：In the step c, the recurrence grader is
Softmax returns grader, and the structure electrocardiosignal grader specifically includes：With the normal data of a part of MIT databases
The training data of grader is returned as softmax, with the output of the third layer hidden layer of the sparse own coding algorithm of the multilayer
Feature returns the input layer of grader as softmax, and the defeated of grader is returned according to the mark training softmax of normal data
Enter the parameter between layer and output layer, while add in phase, S-T segment, R-wave amplitude between the wave character RR of extraction, trained
Softmax return grader.
The technical solution that the embodiment of the present invention is taken further includes：The step c is further included：It is calculated with the fine tuning of backpropagation
Method, the entire softmax of fine tuning successively return grader, obtain final softmax and return grader, and export electrocardiosignal
Classification results.
Compared with the prior art, the advantageous effect that the embodiment of the present invention generates is：The electrocardiosignal of the embodiment of the present invention
Sorter and method determine the parameter of sparse own coding algorithm by the data of training clinical patient, in combination with sparse self-editing
Code algorithm and softmax return grader and build optimal electrocardiosignal grader, and add in classify for electrocardio it is most important
Surface realizes that softmax returns the classification results of grader together with the depth characteristic trained；The present invention is with clinical
Data make classification results have more authenticity for clinical diagnosis closer to the situation of actual clinical patient when training sample；It is logical
It crosses and realizes that the deep learning of mass data trains grader so that the nicety of grading of grader greatly improves；With sparse own coding
The mode that algorithm extraction signal characteristic and surface combine helps to improve the nicety of grading of negligible amounts type signal, solves
The problem of feature extraction difficulty present in existing electrocardiosignal classification and the low signal type nicety of grading of negligible amounts.
Description of the drawings
Fig. 1 is the structure diagram of the electrocardiosignal sorter of the embodiment of the present invention；
Fig. 2 is the structure diagram of the sparse own coding of the embodiment of the present invention；
Fig. 3 is that the sparse own coding combination softmax of the embodiment of the present invention returns the structure chart that grader combines；
Fig. 4 is the training flow chart of the sparse own coding algorithm of the embodiment of the present invention；
Fig. 5 is the flow chart of the electrocardiosignal sorting technique of the embodiment of the present invention.
Specific embodiment
In order to make the purpose , technical scheme and advantage of the present invention be clearer, with reference to the accompanying drawings and embodiments, it is right
The present invention is further elaborated.It should be appreciated that specific embodiment described herein is only to explain the present invention, not
For limiting the present invention.
Referring to Fig. 1, it is the structure diagram of the electrocardiosignal sorter of the embodiment of the present invention.The embodiment of the present invention
Electrocardiosignal sorter includes data input module, parameter training module, categorised decision module and control module.
Data input module is used to go the normal data for having mark in the clinical data of acquisition and MIT databases
It makes an uproar, go the pretreatments such as baseline, normalization, and the waveform in each R cycles is extracted using R ripples as basic point, it will be pre- by input function
Normal data in treated clinical data and MIT databases is input in parameter training module；Specifically, data input mould
Block from different sources will there are the data of different leads and sample rate to input and by bandpass filtering removal noise, pass through
The sample rate of all data is unified for 360Hz by resampling function, and data normalization being shaken for [0,1] with normalized function
Width is come out the single beat periodicity extraction of electrocardiogram (ECG) data using the method for detecting R ripples.The embodiment of the present invention select using R ripples as
Center intercepts a segment of 340 points as training sample, and superposition intercepts N number of cycle, i.e. training sample is 340 row N row
Vector composition；Due to the huge parallel computer that multinuclear is used in the case where hardware condition allows of data volume, selection is suitable
The computer of check figure can realize the input processing of data within a short period of time.
Parameter training module is used to extract the electrocardiosignal feature of clinical data, training by the sparse own coding algorithm of multilayer
The sparse own coding algorithm parameter of multilayer, and the parameter of the normal data training softmax recurrence graders with MIT databases.
Wherein, since sparse own coding algorithm is not required training data that can utilize existing substantial amounts of clinical data with label, lead to
Cross electrocardiosignal feature extraction to clinical data more can slice-of-life data characteristics, eliminate due to can only be led with normal data
The defects of areal variation or racial difference of cause；And clinical data amount bigger, more rare features can be extracted.Specifically
Ground, the embodiment of the present invention extract electrocardio with the sparse own coding of one four layers (including an input layer and three layers of hidden layer)
The feature of signal, it is specific as shown in Fig. 2, be the structure diagram of the sparse own coding of the embodiment of the present invention, sparse own coding by
One encoder and a decoder composition, it is special that the feature drawn by constantly adjusting weight coefficient decoder is equal to input
Sign.
Fig. 3 is that the sparse own coding combination softmax of the embodiment of the present invention returns the structure chart that grader combines.Start with
Machine initializes the initial value of each parameter, is successively trained by the algorithm of sparse own coding, waveform is intercepted 340 centered on R ripples
A point directly inputs former data, first layer hides layer choosing as wave band to be trained for the input layer of sparse own coding
Appropriate number of node is selected, chooses sigmoid functions as transmission function, definition reconstructs feature and is originally inputted between feature
Cost function constantly minimizes cost function by gradient descent method, determines the weight system of input layer and first layer hidden layer
Number；The number of nodes of second layer hidden layer is selected, it is special as the input of second layer hidden layer using the output characteristic of first layer hidden layer
Sign trains the weight coefficient between first layer hidden layer and second layer hidden layer in the same fashion；The similary training second layer is hidden
The weight coefficient between layer and third layer hidden layer is hidden, third layer hidden layer is the output of sparse own coding therefore, is finally completed
The training of sparse own coding parameter.In sparse own coding algorithm, it is first determined the number of plies of hidden layer, it is then determined that every layer of section
Points, the transmission function between input layer and hidden layer and hidden layer and hidden layer use sigmoid functions herein, i.e.,
The weight coefficient of the encoder of own coding and the weight coefficient of decoder transposition each other, i.e.,
Then the parameter of self-encoding encoder is：{W、bh、bf}。
Assuming that reconstructed error function is L (X, Y), rule of thumb since this paper transmission functions are sigmoid functions, then reconstruct
Error function takes cross entropy： One is proposed according to reconstructed error function
Whole loss function：
JAE(θ)=∑X∈sL (X, f (g (X))) (3)
The parameter that just can obtain network by constantly carrying out minimization to loss function.
It is the training flow chart of the sparse own coding algorithm of the embodiment of the present invention also referring to Fig. 4.Sparse own coding is calculated
Method specifically includes：
(1) the initial data of ECG signal, is obtained, passes through R point recognition detections each heart rate cycle of waveform and minute window.If it obtains
The initial data taken is X (x1,x2,…xn,)。
(2) network is initialized, gives the input layer number n, first layer hidden layer number of nodes m and second of autoencoder network
Layer hidden layer h gives the sparse value of network, learning rate, average activation value, maximum iteration etc.；It is initial according to number of network node
Change parameter θ1, θ2。
(3) transmission function h (), f () are defined, cost function J is defined according to reconstruction of function L (X, Y)AE(θ)。
(4) input data is inputted to the cost function for calculating the first layer network, cost letter is minimized by back-propagation algorithm
Number adjusts weight coefficient
It is whether optimal using gradient detection parameters, into next step if optimal, if not optimal correction network parameter θ1,
Continue to minimize until maximum iteration.
(5) first layer parameter θ is obtained1.Obtained parameter is substituted into transmission function：
The output characteristic value of first layer hidden layer is obtained, using the output characteristic value of first layer hidden layer as the defeated of the second layer
Enter, carry out similar training and obtain the parameter value θ between first layer hidden layer and second layer hidden layer2。
(6) the parameter θ second layer obtained2Substitute into second transmission function：
Obtain the output characteristic value of second layer hidden layer.
(7) the rest may be inferred, using the output characteristic value of second layer hidden layer as the input of third layer, carries out similar training
The parameter value between second layer hidden layer and third layer hidden layer is obtained, and obtains the output characteristic value of third layer hidden layer, most
Sparse own coding algorithm is completed in training eventually.
Categorised decision module is used to return classifier parameters structure according to the sparse own coding algorithm parameter of multilayer and softmax
Electrocardiosignal grader carries out electrocardiosignal classification by the electrocardiosignal grader.Specifically, categorised decision module includes
One softmax returns grader and a fine-adjusting unit, and softmax returns mark of the grader in a part of MIT databases
Quasi- data return the training data of grader as softmax, and data first pass through one four layers of parameter and train to come for front
Sparse own coding algorithm encoded after feature, using the output characteristic value of sparse own coding algorithm third layer hidden layer as
Softmax returns the input layer of grader, i.e., the feature after sparse own coding for softmax recurrence graders input, according to
The mark of normal data is returned the ginseng between the input layer and output layer of grader by the method training softmax of supervised learning
Number, while three phase, S-T segment, R-wave amplitude surfaces between the wave character RR of extraction are added in, obtain trained softmax
Return grader；The fine tuning algorithm of fine-adjusting unit backpropagation, the entire softmax of fine tuning successively return grader, finally
It obtains the high and applied widely softmax of a nicety of grading and returns grader, and export electrocardiosignal classification results.
Control module include computer and control unit, for control respectively data input module, parameter training module and
The work of categorised decision module.
Referring to Fig. 5, it is the flow chart of the electrocardiosignal sorting technique of the embodiment of the present invention.The electrocardio of the embodiment of the present invention
Modulation recognition method comprises the following steps：
Step 10：By the normal data for having mark in the clinical data of acquisition and MIT databases carry out denoising, go baseline,
The pretreatments such as normalization, the waveform in each R cycles is extracted using R ripples as basic point, and passes through input function and faces pretreated
Normal data in bed data and MIT databases is input in parameter training module；
In step 10, data input module is defeated by the data with different leads and sample rate from different sources
Enter and noise is removed by bandpass filtering, the sample rate of all data is unified for by 360Hz by resampling function, with normalization
Data normalization is the amplitude of [0,1] by function, is gone out the single beat periodicity extraction of electrocardiogram (ECG) data using the method for detecting R ripples
Come, the embodiment of the present invention selects to intercept a segment of 340 points using centered on R ripples as training sample, and superposition intercepts N number of week
Phase, i.e. training sample form for the vector of 340 row N row；It is used since data volume is huge in the case where hardware condition allows more
The parallel computer of core selects the computer of suitable check figure that can realize the input processing of data within a short period of time.
Step 20：The electrocardiosignal feature of clinical data is extracted by the sparse own coding algorithm of multilayer, training multilayer is sparse
Own coding algorithm parameter, and the parameter of the normal data training softmax recurrence graders with MIT databases；
In step 20, since sparse own coding algorithm is not required training data that can be utilized existing big with label
The clinical data of amount, by the electrocardiosignal feature extraction to clinical data more can slice-of-life data characteristics, eliminate due to
The defects of areal variation caused by normal data or racial difference can only be used；And clinical data amount bigger, it can extract more
More rare features.Specifically, the embodiment of the present invention is with the dilute of one four layers (including an input layer and three layers of hidden layer)
Thin own coding extracts the feature of electrocardiosignal, specific as shown in Fig. 2, being that the structure of the sparse own coding of the embodiment of the present invention is shown
It is intended to, sparse own coding is made of an encoder and a decoder, is obtained by constantly adjusting weight coefficient decoder
The feature gone out is equal to input feature vector.
Fig. 3 is that the sparse own coding combination softmax of the embodiment of the present invention returns the structure chart that grader combines.Start with
Machine initializes the initial value of each parameter, is successively trained by the algorithm of sparse own coding, waveform is intercepted 340 centered on R ripples
A point directly inputs former data, first layer hides layer choosing as wave band to be trained for the input layer of sparse own coding
Appropriate number of node is selected, chooses sigmoid functions as transmission function, definition reconstructs feature and is originally inputted between feature
Cost function constantly minimizes cost function by gradient descent method, determines the weight system of input layer and first layer hidden layer
Number；The number of nodes of second layer hidden layer is selected, it is special as the input of second layer hidden layer using the output characteristic of first layer hidden layer
Sign trains the weight coefficient between first layer hidden layer and second layer hidden layer in the same fashion；The similary training second layer is hidden
The weight coefficient between layer and third layer hidden layer is hidden, third layer hidden layer is the output of sparse own coding therefore, is finally completed
The training of sparse own coding parameter.In sparse own coding algorithm, it is first determined the number of plies of hidden layer, it is then determined that every layer of section
Points, the transmission function between input layer and hidden layer and hidden layer and hidden layer use sigmoid functions herein, i.e.,
The weight coefficient of the encoder of own coding and the weight coefficient of decoder transposition each other, i.e.,
Then the parameter of self-encoding encoder is：{W、bh、bf}。
Assuming that reconstructed error function is L (X, Y), rule of thumb since this paper transmission functions are sigmoid functions, then reconstruct
Error function takes cross entropy： One is proposed according to reconstructed error function
Whole loss function：
JAE(θ)=∑X∈sL (X, f (g (X))) (3)
The parameter that just can obtain network by constantly carrying out minimization to loss function.
It is the training flow chart of the sparse own coding algorithm of the embodiment of the present invention also referring to Fig. 4.Sparse own coding is calculated
The idiographic flow of method is as follows：
Step 21：The initial data of ECG signal is obtained, passes through R point recognition detections each heart rate cycle of waveform and minute window.
If the initial data obtained is X (x1,x2,…xn,)。
Step 22：Initialize network, give autoencoder network input layer number n, first layer hidden layer number of nodes m and
Second layer hidden layer h gives the sparse value of network, learning rate, average activation value, maximum iteration etc.；According to number of network node
Initiation parameter θ1, θ2。
Step 23：Transmission function h (), f () are defined, cost function J is defined according to reconstruction of function L (X, Y)AE(θ)。
Step 24：Input data is inputted to the cost function for calculating the first layer network, is minimized by back-propagation algorithm
Cost function adjusts weight coefficient
It is whether optimal using gradient detection parameters, into next step if optimal, if not optimal correction network parameter θ1,
Continue to minimize until maximum iteration.
Step 25：Obtain first layer parameter θ1.Obtained parameter is substituted into transmission function：
The output characteristic value of first layer hidden layer is obtained, using the output characteristic value of first layer hidden layer as the defeated of the second layer
Enter, carry out similar training and obtain the parameter value θ between first layer hidden layer and second layer hidden layer2。
Step 26：The parameter θ that the second layer is obtained2Substitute into second transmission function：
Obtain the output characteristic value of second layer hidden layer.
Step 27：The rest may be inferred, using the output characteristic value of second layer hidden layer as the input of third layer, carries out similar
Training obtains the parameter value between second layer hidden layer and third layer hidden layer, and obtains the output characteristic of third layer hidden layer
Sparse own coding algorithm is completed in value, final training.
Step 30：Classifier parameters structure electrocardiosignal is returned according to the sparse own coding algorithm parameter of multilayer and softmax
Grader carries out electrocardiosignal classification by electrocardiosignal grader, and exports electrocardiosignal classification results.
In step 30, the method for building electric signal grader specifically includes：With the criterion numeral in a part of MIT databases
According to the training data that grader is returned as softmax, obtain trained softmax and return grader；Data first pass through one
A four layers of parameter trains the feature after the sparse own coding algorithm come is encoded for front, with sparse own coding algorithm the
The output characteristic value of three layers of hidden layer returns the input layer of grader as softmax, i.e., the feature after sparse own coding is
Softmax returns the input of grader, according to the mark of normal data, returns and divides by the method training softmax of supervised learning
Parameter between the input layer and output layer of class device, while add in phase, S-T segment, R-wave amplitude three between the wave character RR of extraction
Surface obtains trained softmax and returns grader.The fine tuning algorithm of backpropagation is used again, and fine tuning successively is entire
Softmax returns grader, finally obtains the high and applied widely softmax of a nicety of grading and returns grader.
The electrocardiosignal sorter of the embodiment of the present invention and method by the data of training clinical patient determine it is sparse from
The parameter of encryption algorithm returns grader in combination with sparse own coding algorithm and softmax and builds optimal electrocardiosignal point
Class device, and add in and realize that softmax is returned together with the depth characteristic trained for the most important surface of electrocardio classification
The classification results of grader；The present invention is with clinical data when training sample makes classification results closer to the feelings of actual clinical patient
Condition has more authenticity for clinical diagnosis；Grader grader is trained by the deep learning for realizing mass data
Nicety of grading greatly improves；It is helped to improve in a manner that sparse own coding algorithm extracts signal characteristic and surface combines
The nicety of grading of negligible amounts type signal solves feature extraction difficulty and quantity present in existing electrocardiosignal classification
The problem of less signal type nicety of grading is low.
The foregoing description of the disclosed embodiments enables professional and technical personnel in the field to realize or use the present invention.
A variety of modifications of these embodiments will be apparent for those skilled in the art, it is as defined herein
General Principle can be realized in other embodiments without departing from the spirit or scope of the present invention.Therefore, it is of the invention
The embodiments shown herein is not intended to be limited to, and is to fit to and the principles and novel features disclosed herein phase one
The most wide scope caused.
Claims (10)
1. a kind of electrocardiosignal sorter, which is characterized in that including：
Data input module：For the normal data of the clinical data of acquisition and MIT databases to be input to parameter training module
In；
Parameter training module：For extracting the electrocardiosignal feature of the clinical data, the sparse own coding algorithm ginseng of training multilayer
Number, and return classifier parameters with the normal data training of the MIT databases；
Categorised decision module：For according to the sparse own coding algorithm parameter of the multilayer and recurrence classifier parameters structure electrocardio letter
Number grader passes through the electrocardiosignal grader and carries out electrocardiosignal classification.
2. electrocardiosignal sorter according to claim 1, which is characterized in that the data input module is additionally operable to：
The normal data of the clinical data of acquisition and MIT databases is subjected to denoising, goes baseline, normalization pretreatment, and using R ripples as base
Point extracts the waveform in each R cycles.
3. electrocardiosignal sorter according to claim 2, which is characterized in that the sparse own coding of multilayer is defeated including one
Enter layer and three layers of hidden layer, the parameter training module training sparse own coding algorithm parameter of multilayer specifically includes：It is random initial
Change the initial value of each parameter, successively trained by the algorithm of sparse own coding, waveform is intercepted 340 points centered on R ripples makees
For wave band to be trained, the number of nodes of input layer and three layers of hidden layer is selected respectively, is constantly minimized by gradient descent method
Cost function determines the weight coefficient of input layer and first layer hidden layer, using the output characteristic of first layer hidden layer as second
The input feature vector of layer hidden layer, the weight coefficient between training first layer hidden layer and second layer hidden layer, and it is hidden with the second layer
The output characteristic for hiding layer is trained as the input feature vector of third layer hidden layer between second layer hidden layer and third layer hidden layer
Weight coefficient completes the training of the sparse own coding parameter of the multilayer.
4. electrocardiosignal sorter according to claim 3, which is characterized in that the categorised decision module includes
Softmax returns grader, and the softmax returns grader by the use of the normal data of a part of MIT databases as softmax
Return grader training data, using the output characteristic of the third layer hidden layer of the sparse own coding algorithm of the multilayer as
Softmax return grader input layer, according to normal data mark training softmax return grader input layer with it is defeated
Go out the parameter between layer, while add in phase, S-T segment, R-wave amplitude between the wave character RR of extraction, obtain trained softmax
Return grader.
5. electrocardiosignal sorter according to claim 4, which is characterized in that the categorised decision module further includes micro-
Unit, the fine tuning algorithm of the fine-adjusting unit backpropagation are adjusted, the entire softmax of fine tuning successively returns grader, obtains
Final softmax returns grader, and exports electrocardiosignal classification results.
6. a kind of electrocardiosignal sorting technique, which is characterized in that comprise the following steps：
Step a：The normal data of the clinical data of acquisition and MIT databases is input in parameter training module；
Step b：The electrocardiosignal feature of the clinical data, the sparse own coding algorithm parameter of training multilayer are extracted, and uses institute
The normal data training for stating MIT databases returns classifier parameters；
Step c：According to the sparse own coding algorithm parameter of the multilayer and classifier parameters structure electrocardiosignal grader is returned, is led to
It crosses the electrocardiosignal grader and carries out electrocardiosignal classification.
7. electrocardiosignal sorting technique according to claim 6, which is characterized in that the step a is further included：By acquisition
The normal data of clinical data and MIT databases carries out denoising, goes baseline, normalization pretreatment, and is extracted using R ripples as basic point
The waveform in each R cycles.
8. electrocardiosignal sorting technique according to claim 7, which is characterized in that in the step b, the multilayer is dilute
Dredging own coding includes an input layer and three layers of hidden layer, and the trained sparse own coding algorithm parameter of multilayer specifically includes：With
Machine initializes the initial value of each parameter, is successively trained by the algorithm of sparse own coding, waveform is intercepted 340 centered on R ripples
A point is as wave band to be trained, and the number of nodes of selection input layer and three layers of hidden layer, continuous by gradient descent method respectively
Cost function is minimized, determines the weight coefficient of input layer and first layer hidden layer, is made with the output characteristic of first layer hidden layer
For the input feature vector of second layer hidden layer, the weight coefficient between training first layer hidden layer and second layer hidden layer, and with the
The output characteristic of two layers of hidden layer is as the input feature vector training second layer hidden layer of third layer hidden layer and third layer hidden layer
Between weight coefficient, complete the training of the sparse own coding parameter of the multilayer.
9. electrocardiosignal sorting technique according to claim 8, which is characterized in that in the step c, described return is divided
Class device returns grader for softmax, and the structure electrocardiosignal grader specifically includes：With the mark of a part of MIT databases
Quasi- data return the training data of grader as softmax, with the third layer hidden layer of the sparse own coding algorithm of the multilayer
Output characteristic as softmax return grader input layer, according to the mark of normal data train softmax return classification
Parameter between the input layer and output layer of device, while phase, S-T segment, R-wave amplitude between the wave character RR of extraction are added in, it obtains
Trained softmax returns grader.
10. electrocardiosignal sorting technique according to claim 9, which is characterized in that the step c is further included：With reversed
The fine tuning algorithm of propagation, the entire softmax of fine tuning successively return grader, obtain final softmax and return grader, and
Export electrocardiosignal classification results.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611067279.0A CN108113647A (en) | 2016-11-28 | 2016-11-28 | A kind of electrocardiosignal sorter and method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611067279.0A CN108113647A (en) | 2016-11-28 | 2016-11-28 | A kind of electrocardiosignal sorter and method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108113647A true CN108113647A (en) | 2018-06-05 |
Family
ID=62224852
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201611067279.0A Pending CN108113647A (en) | 2016-11-28 | 2016-11-28 | A kind of electrocardiosignal sorter and method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108113647A (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108937916A (en) * | 2018-08-03 | 2018-12-07 | 西南大学 | A kind of electrocardiograph signal detection method, device and storage medium |
| CN109171670A (en) * | 2018-06-25 | 2019-01-11 | 天津海仁医疗技术有限公司 | A kind of 3D blood vessel imaging algorithm based on reverse Principal Component Analysis |
| CN109259733A (en) * | 2018-10-25 | 2019-01-25 | 深圳和而泰智能控制股份有限公司 | Apnea detection method, apparatus and detection device in a kind of sleep |
| CN110085108A (en) * | 2019-03-08 | 2019-08-02 | 台州学院 | A kind of electrocardiosignal analogue system |
| CN110403601A (en) * | 2019-08-27 | 2019-11-05 | 安徽心之声医疗科技有限公司 | Electrocardiosignal QRS wave group recognition methods based on deep learning |
| CN111084621A (en) * | 2019-12-30 | 2020-05-01 | 上海数创医疗科技有限公司 | QRS wave group form identification method and device based on depth self-encoder |
| CN111680785A (en) * | 2020-05-29 | 2020-09-18 | 山东省人工智能研究院 | ECG signal processing method based on sparse characteristic and antagonistic neural network combination |
| CN113288156A (en) * | 2021-05-18 | 2021-08-24 | 浙江大学 | Method for generating electrocardiographic data of any lead visual angle |
| CN115426940A (en) * | 2020-02-12 | 2022-12-02 | 意锐瑟科技公司 | Non-invasive cardiac monitor and method of inferring patient physiological characteristics using recorded cardiac data |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104523266A (en) * | 2015-01-07 | 2015-04-22 | 河北大学 | Automatic classification method for electrocardiogram signals |
| CN104523264A (en) * | 2014-12-31 | 2015-04-22 | 深圳职业技术学院 | Electrocardiosignal processing method |
-
2016
- 2016-11-28 CN CN201611067279.0A patent/CN108113647A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104523264A (en) * | 2014-12-31 | 2015-04-22 | 深圳职业技术学院 | Electrocardiosignal processing method |
| CN104523266A (en) * | 2015-01-07 | 2015-04-22 | 河北大学 | Automatic classification method for electrocardiogram signals |
Non-Patent Citations (1)
| Title |
|---|
| JIANLI YANG 等: "A novel method of diagnosing premature ventricular contraction based on sparse autoencoder and softmax regression", 《BIO-MEDICAL MATERIALS AND ENGINEERING》 * |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109171670A (en) * | 2018-06-25 | 2019-01-11 | 天津海仁医疗技术有限公司 | A kind of 3D blood vessel imaging algorithm based on reverse Principal Component Analysis |
| CN108937916A (en) * | 2018-08-03 | 2018-12-07 | 西南大学 | A kind of electrocardiograph signal detection method, device and storage medium |
| CN109259733A (en) * | 2018-10-25 | 2019-01-25 | 深圳和而泰智能控制股份有限公司 | Apnea detection method, apparatus and detection device in a kind of sleep |
| CN110085108B (en) * | 2019-03-08 | 2021-10-29 | 台州学院 | Electrocardiosignal simulation system |
| CN110085108A (en) * | 2019-03-08 | 2019-08-02 | 台州学院 | A kind of electrocardiosignal analogue system |
| CN110403601A (en) * | 2019-08-27 | 2019-11-05 | 安徽心之声医疗科技有限公司 | Electrocardiosignal QRS wave group recognition methods based on deep learning |
| CN111084621A (en) * | 2019-12-30 | 2020-05-01 | 上海数创医疗科技有限公司 | QRS wave group form identification method and device based on depth self-encoder |
| CN111084621B (en) * | 2019-12-30 | 2022-09-06 | 上海数创医疗科技有限公司 | QRS wave group form identification method and device based on depth self-encoder |
| CN115426940A (en) * | 2020-02-12 | 2022-12-02 | 意锐瑟科技公司 | Non-invasive cardiac monitor and method of inferring patient physiological characteristics using recorded cardiac data |
| CN111680785A (en) * | 2020-05-29 | 2020-09-18 | 山东省人工智能研究院 | ECG signal processing method based on sparse characteristic and antagonistic neural network combination |
| CN111680785B (en) * | 2020-05-29 | 2021-09-24 | 山东省人工智能研究院 | ECG signal processing method based on sparse characteristic and antagonistic neural network combination |
| CN113288156A (en) * | 2021-05-18 | 2021-08-24 | 浙江大学 | Method for generating electrocardiographic data of any lead visual angle |
| CN113288156B (en) * | 2021-05-18 | 2022-02-11 | 浙江大学 | Method for generating electrocardiographic data of any lead visual angle |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108113647A (en) | A kind of electrocardiosignal sorter and method | |
| Lynn et al. | A deep bidirectional GRU network model for biometric electrocardiogram classification based on recurrent neural networks | |
| Zhang et al. | HeartID: A multiresolution convolutional neural network for ECG-based biometric human identification in smart health applications | |
| CN109567793B (en) | Arrhythmia classification-oriented ECG signal processing method | |
| Ebrahimzadeh et al. | Prediction of paroxysmal Atrial Fibrillation: A machine learning based approach using combined feature vector and mixture of expert classification on HRV signal | |
| Das et al. | ECG beats classification using mixture of features | |
| CN107811626A (en) | A kind of arrhythmia classification method based on one-dimensional convolutional neural networks and S-transformation | |
| CN107194404B (en) | Underwater target feature extraction method based on convolutional neural network | |
| Jun et al. | Premature ventricular contraction beat detection with deep neural networks | |
| CN109948647A (en) | A kind of electrocardiogram classification method and system based on depth residual error network | |
| Sha et al. | A breast cancer classifier using a neuron model with dendritic nonlinearity | |
| CN110458197A (en) | Personal identification method and its system based on photoplethysmographic | |
| CN106251880A (en) | Identify method and the system of physiological sound | |
| CN112508110A (en) | Deep learning-based electrocardiosignal graph classification method | |
| CN109620152A (en) | A kind of electrocardiosignal classification method based on MutiFacolLoss-Densenet | |
| CN111481192B (en) | Electrocardiosignal R wave detection method based on improved U-Net | |
| Reddy et al. | An efficient multilevel thresholding scheme for heart image segmentation using a hybrid generalized adversarial network | |
| CN116361688A (en) | Multi-mode feature fusion model construction method for automatic classification of electrocardiographic rhythms | |
| Kheradpisheh et al. | Mixture of feature specified experts | |
| Kuila et al. | Feature extraction of electrocardiogram signal using machine learning classification | |
| Doshi et al. | (Retracted) Deep belief network-based image processing for local directional segmentation in brain tumor detection | |
| Foeady et al. | Lung cancer classification based on CT scan image by applying FCM segmentation and neural network technique | |
| Kim et al. | Algorithm for classifying arrhythmia using Extreme Learning Machine and principal component analysis | |
| Yang et al. | The study of electrocardiograph based on radial basis function neural network | |
| Muthuvel et al. | ECG signal feature extraction and classification using harr wavelet transform and neural network |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180605 |