CN108647614A - The recognition methods of electrocardiogram beat classification and system - Google Patents
The recognition methods of electrocardiogram beat classification and system Download PDFInfo
- Publication number
- CN108647614A CN108647614A CN201810404828.1A CN201810404828A CN108647614A CN 108647614 A CN108647614 A CN 108647614A CN 201810404828 A CN201810404828 A CN 201810404828A CN 108647614 A CN108647614 A CN 108647614A
- Authority
- CN
- China
- Prior art keywords
- heart
- wavelet packet
- ecg signal
- electrocardiogram
- neural networks
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
-
- 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
- A61B5/72—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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B2576/00—Medical imaging apparatus involving image processing or analysis
- A61B2576/02—Medical imaging apparatus involving image processing or analysis specially adapted for a particular organ or body part
- A61B2576/023—Medical imaging apparatus involving image processing or analysis specially adapted for a particular organ or body part for the heart
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
Abstract
The invention discloses a kind of electrocardiogram beat classification recognition methods and system, this method to include:Step S1:ECG signal in abnormal electrocardiogram database is pre-processed;Step S2:It is chosen from abnormal electrocardiogram database and passes through pretreated ECG signal, then extracted the heart from the pretreated ECG signal of the process of selection and clap, obtain the heart and clap sample set;Step S3:It is clapped from the heart and randomly selects a part of heart bat in sample set as training sample set, the heart of remainder, which is clapped, is used as test sample collection;Step S4:Training sample set is inputted in convolutional neural networks model and is trained, realizes the extraction of ECG signal feature, which includes spatial pyramid pond layer;Step S5:In characteristic information and test sample collection the input convolutional neural networks model that extraction is obtained, and classified to the output result of convolutional neural networks model using grader.The present invention can improve the accuracy rate of beat classification identification.
Description
Technical field
The present invention relates to field of medical technology, and in particular to a kind of electrocardiogram beat classification recognition methods and system.
Background technology
Electrocardiogram is the figure for the potential change for drawing diversified forms from body surface by electrocardiograph, and it is emerging to reflect heart
The electrical activity process put forth energy.Electrocardiogram has important reference value in terms of heart basic function and its pathological study.Utilize meter
Electrocardiosignal is identified calculation machine automatic diagnostics and accuracy rate of diagnosis can be improved in classification, and doctor is made to know from cumbersome figure
It frees in not working, and electrocardiogram develops into the important foundation of home health care equipment in the future.
Artificial intelligence and machine learning have been widely used in the identification and classification of heart bat, such as SVM points of existing method
Class device, LS-SVM graders, PSO-SVM graders, PSO-RBF graders, neural network etc..Electrocardiosignal identity recognizing technology
By limitations such as de-noising, feature extractions, the recognition effect of electrocardiosignal is caused to be difficult to be promoted.In order to realize satisfied classification
Can, how to choose suitable feature becomes particularly important.Artificial intelligence and machine learning at present is widely used in this field, good
Classification results be unable to do without the characteristics extraction appropriate clapped electrocardio center of fiqure, feature extraction is exactly from initial electrocardiosignal spy
Most representational character subset is selected in sign, these character subsets have better generalization ability, can improve electrocardiogram
The accuracy of beat classification.
The especially neural network algorithm of machine learning in recent years, it is particularly aobvious in language identification and image procossing etc. effect
It writes.Feature can be automatically extracted by building neural network model, in the case of some complexity classification, the feature that automatically extracts
Value has better effect in Classification and Identification.Depth convolutional network this include numerous hidden layers network structure, have pass
The incomparable ability to express of machine application of uniting and feature learning are horizontal.Therefore its be applied to deep learning algorithm train with
Come, good effect is obtained in the Study of recognition of many large sizes, but existing convolutional neural networks (CNN) require input
The size of data is consistent, and this requirement is artificial, mainly since grader (SVM/Softmax) or full linking layer need admittedly
The vector of measured length.This artificial operation can lead to the loss of image useful information, influence nicety of grading.
Invention content
The purpose of the present invention is to provide a kind of electrocardiogram beat classification recognition methods and systems, can improve beat classification
The accuracy rate of identification.
To achieve the above object, technical scheme of the present invention provides a kind of electrocardiogram beat classification recognition methods, including:
Step S1:ECG signal in abnormal electrocardiogram database is pre-processed, the pretreatment includes removal base
Line drift processing and denoising;
Step S2:From the abnormal electrocardiogram database choose pass through the pretreated ECG signal, then from
The heart is extracted in the process pretreated ECG signal of the selection to clap, and is obtained the heart and is clapped sample set;
Step S3:A part of heart is randomly selected from heart bat sample set to clap as training sample set, remainder
The heart, which is clapped, is used as test sample collection;
Step S4:The training sample set is inputted in convolutional neural networks model and is trained, realizes ECG signal
The extraction of feature, wherein the convolutional neural networks model is five-layer structure, respectively the first convolutional layer, maximum pond layer, the
Two convolutional layers, spatial pyramid pond layer, full articulamentum, wherein spatial pyramid pond layer is used for the volume Two
It is unified that the data of lamination output carry out size;
Step S5:Obtained characteristic information and the test sample collection input convolutional neural networks are extracted by described
In model, and classified to the output result of the convolutional neural networks model using grader, is realized to the variety classes heart
The Classification and Identification of bat.
Further, in the step S1, using wavelet packet analysis method to the electrocardio in the abnormal electrocardiogram database
Figure signal carries out dry processing.
Further, dry place is carried out to the ECG signal in the abnormal electrocardiogram database using wavelet packet analysis method
Reason includes:
WAVELET PACKET DECOMPOSITION is carried out to ECG signal, obtains corresponding WAVELET PACKET DECOMPOSITION tree;
Best tree is obtained from the WAVELET PACKET DECOMPOSITION tree using principle of minimum cost;
High frequency is obtained from the Best tree using the adaptive threshold selection method based on Stein unbiased possibility predication principles
The wavelet packet analysis coefficient of sequence uses and obtains low frequency sequence from the Best tree by the method for floating threshold value of foundation of signal energy
The wavelet packet analysis coefficient of row;
It is carried out according to the wavelet packet analysis coefficient of the high frequency series and the wavelet packet analysis coefficient of the low frequency sequence
The reconstruct of signal.
Further, the grader is Softmax function category devices.
To achieve the above object, technical scheme of the present invention additionally provides a kind of electrocardiogram beat classification identifying system, packet
It includes:
Preprocessing module, for being pre-processed to the ECG signal in abnormal electrocardiogram database, the pretreatment packet
Include removal baseline drift processing and denoising;
The heart claps sample acquisition module, passes through the pretreated electrocardio for being chosen from the abnormal electrocardiogram database
Then figure signal extracts the heart from the process of the selection pretreated ECG signal and claps, obtains the heart and clap sample set;
Module is chosen, a part of heart bat is randomly selected in sample set as training sample set, residue for being clapped from the heart
The partial heart, which is clapped, is used as test sample collection;
First processing module is trained for inputting the training sample set in convolutional neural networks model, is realized
The extraction of ECG signal feature, wherein the convolutional neural networks model is five-layer structure, respectively the first convolutional layer, most
Great Chiization layer, the second convolutional layer, spatial pyramid pond layer, full articulamentum, wherein spatial pyramid pond layer for pair
It is unified that the data of the second convolutional layer output carry out size;
Second processing module, for extracting obtained characteristic information and the test sample collection input volume by described
In product neural network model, and classified to the output result of the convolutional neural networks model using grader, realization pair
The Classification and Identification that the variety classes heart is clapped.
Further, the preprocessing module uses wavelet packet analysis method to the electrocardiogram in the abnormal electrocardiogram database
Signal carries out dry processing.
Further, the preprocessing module includes:
Resolving cell obtains corresponding WAVELET PACKET DECOMPOSITION tree for carrying out WAVELET PACKET DECOMPOSITION to ECG signal;
Best tree extraction unit, for using principle of minimum cost to obtain Best tree from the WAVELET PACKET DECOMPOSITION tree;
Threshold value quantizing unit, for using the adaptive threshold selection method based on Stein unbiased possibility predication principles from institute
The wavelet packet analysis coefficient that high frequency series are obtained in Best tree is stated, it is the method for floating threshold value of foundation from described to use using signal energy
The wavelet packet analysis coefficient of low frequency sequence is obtained in Best tree;
Reconfiguration unit, for according to the wavelet packet analysis coefficient of the high frequency series and the wavelet packet of the low frequency sequence
Coefficient of analysis carries out the reconstruct of signal.
Further, the grader is Softmax function category devices.
Electrocardiogram beat classification recognition methods provided by the invention, by adding space gold in convolutional neural networks model
Word tower basin layer realizes that the size of feature is unified using spatial pyramid pond layer, the spy clapped changeable ruler feeling may be implemented
Sign extraction is not required to carry out size unification to the data for inputting convolutional neural networks, is conducive to the primitive character for retaining more oversensitive bat,
And then the accuracy rate of Classification and Identification can be improved.
Description of the drawings
Fig. 1 is a kind of flow chart for electrocardiogram beat classification recognition methods that embodiment of the present invention provides;
Fig. 2 is a kind of schematic diagram for WAVELET PACKET DECOMPOSITION tree that embodiment of the present invention provides;
Fig. 3 is a kind of schematic diagram for Best tree that embodiment of the present invention provides;
Fig. 4 is a kind of schematic diagram for convolutional neural networks model that embodiment of the present invention provides.
Specific implementation mode
With reference to the accompanying drawings and examples, the specific implementation mode of the present invention is described in further detail.Implement below
Example is not limited to the scope of the present invention for illustrating the present invention.
It is a kind of flow chart for electrocardiogram beat classification recognition methods that embodiment of the present invention provides referring to Fig. 1, Fig. 1,
The electrocardiogram beat classification recognition methods includes:
Step S1:ECG signal in abnormal electrocardiogram database is pre-processed, the pretreatment includes removal base
Line drift processing and denoising;
Step S2:From the abnormal electrocardiogram database choose pass through the pretreated ECG signal, then from
The heart is extracted in the process pretreated ECG signal of the selection to clap, and is obtained the heart and is clapped sample set;
Step S3:A part of heart is randomly selected from heart bat sample set to clap as training sample set, remainder
The heart, which is clapped, is used as test sample collection;
Step S4:The training sample set is inputted in convolutional neural networks model and is trained, realizes ECG signal
The extraction of feature, wherein the convolutional neural networks model is five-layer structure, respectively the first convolutional layer, maximum pond layer, the
Two convolutional layers, spatial pyramid pond layer, full articulamentum, wherein spatial pyramid pond layer is used for the volume Two
It is unified that the data of lamination output carry out size;
Step S5:Obtained characteristic information and the test sample collection input convolutional neural networks are extracted by described
In model, and classified to the output result of the convolutional neural networks model using grader, is realized to the variety classes heart
The Classification and Identification of bat.
The electrocardiogram beat classification recognition methods that embodiment of the present invention provides, by increasing in convolutional neural networks model
If spatial pyramid pond layer, realizes that the size of feature is unified using spatial pyramid pond layer, may be implemented to variable size
The feature extraction that the heart is clapped is not required to carry out size unification to the data for inputting convolutional neural networks, is conducive to retain more oversensitive bat
Primitive character, and then the accuracy rate of Classification and Identification can be improved.
Preferably, in order to retain more local features and realize that the denoising of signal may be used in above-mentioned steps S1
Wavelet packet analysis method carries out dry processing to the ECG signal in the abnormal electrocardiogram database, and wavelet packet analysis method is not only right
Low frequency signal is decomposed, and is also decomposed to high-frequency signal, by using wavelet packet analysis method, not only realizes denoising, and
And by being decomposed to high and low frequency, more minutias can be extracted, are as follows:
Step A:WAVELET PACKET DECOMPOSITION is carried out to ECG signal, obtains corresponding WAVELET PACKET DECOMPOSITION tree;
First, suitable wavelet basis function and Decomposition order is selected to carry out wavelet packet point to ECG signal (ECG signal)
Solution, obtains low frequency coefficient and high frequency coefficient.Enough characteristic mass are reconstructed since the difference of Decomposition order directly affects, it is possible to
The optimal Decomposition number of plies is determined with " noise margin method ", for example, it is 4 layers to obtain Decomposition order, selects sym4 as wavelet basis function,
It is as shown in Figure 2 that it obtains corresponding WAVELET PACKET DECOMPOSITION tree;
Step B:Best tree is obtained from the WAVELET PACKET DECOMPOSITION tree using principle of minimum cost;
I.e. according to principle of minimum cost, determine that Optimum Wavelet Packet calculates Best tree, for a given entropy standard
(using minimum Shannon entropys), calculating Optimum Wavelet Packet, such as obtained Best tree are as shown in Figure 3;
Step C:It is obtained from the Best tree using the adaptive threshold selection method based on Stein unbiased possibility predication principles
The wavelet packet analysis coefficient for taking high frequency series is used and is obtained from the Best tree using signal energy as the method for floating threshold value of foundation
The wavelet packet analysis coefficient of low frequency sequence;
The threshold value quantizing for carrying out wavelet packet analysis coefficient, for the wavelet packet analysis coefficient of high frequency series, using based on
The adaptive threshold selection of Stein unbiased possibility predication principles, for the wavelet packet analysis coefficient of low frequency sequence, may be used with
Signal energy is that the method for floating threshold value of foundation obtains;
Step D:According to the wavelet packet analysis coefficient of the obtained high frequency series of step C and the wavelet packet of the low frequency sequence
Coefficient of analysis carries out the reconstruct of signal;
The reconstruct of signal is carried out according to the decomposition coefficient of the best wavelet packet basis after quantization, the signal for reconstructing gained is
By the signal that best wavelet packet basis is handled, the signal after reconstruct is as the size of original signal;
ECG signal after reconstruct does detection to QRS wave, P waves, T waves and the heart is clapped mark and realigned, then divides to obtain
The different hearts claps (P waves starting point to T waves terminal), and normalized is done to heart bat.
The ECG signal in MIT-BIH abnormal electrocardiogram databases is pre-processed through the above way, then therefrom
It chooses and passes through above-mentioned pretreated ECG signal, extract the heart from the ECG signal of selection later and clap, obtain the heart and clap sample
This collection, for example, the heart bat for randomly selecting 70% in sample set can be clapped from the heart as training sample set, the heart of residue 30%
It claps and is used as test sample collection;
Wherein, the convolutional neural networks model in embodiment of the present invention is as shown in figure 4, the output of the first convolutional layer connects
The input of maximum pond layer, the input of output the second convolutional layer of connection of maximum pond layer, the output connection of the second convolutional layer are empty
Between pyramid pond layer input, the output of spatial pyramid pond layer connects the input of full articulamentum;
Wherein, in convolutional neural networks model in embodiments of the present invention, convolutional layer (the first i.e. above-mentioned convolution
Layer, the second convolutional layer) fuzzy filter is can be regarded as, so that original signal feature is enhanced and reduces noise, in convolutional layer, on
One layer of feature vector and the convolution kernel of current layer carry out convolution, and the result of convolution algorithm forms this after activation primitive
The Feature Mapping of layer, convolutional layer output can be indicated with following formula:
Wherein,Indicate the corresponding feature vector of j-th of convolution kernel of l layers of convolutional layer, MjIndicate the receiving of Current neural member
Domain,Indicate i-th of weighting coefficient of l j-th of convolution kernel of layer,Indicate the corresponding biasing system of l j-th of convolution kernel of layer
Number, and f (z) is activation primitive, calculation formula is:
Maximum pond layer can be described as sub-sampling layer, and sub-sampling is considered as a kind of special convolution process, maximum pond layer
Sub-sample is carried out to data using the principle of local correlations, retains useful information while reducing data dimension, using pond
Change operation and keep feature, makes feature that there is displacement, scaling invariance, sub-sampling layer to have the function of Further Feature Extraction, count
Calculating formula is:
Wherein, down () indicates sub-sampling operating method, and Max-pooling operating methods may be used in the present invention,Indicate weighting coefficient,Indicate biasing coefficient.
In the present invention, convolutional neural networks model can automatically generate high-level characteristic (weights and threshold value) by training, first
First training sample is sent into convolutional neural networks model and is trained, input vector is obtained, is counted compared with given target vector
Loss function is calculated, calculation formula is as follows:
Wherein, L is loss function (secondary variance), ykFor output vector, dkFor target vector;
Weights are obtained after calculating update according to L and threshold value obtains, and calculation is as follows:
Wherein, α represents learning rate, and j represents the neural unit of hidden layer, and k represents output layer unit, and M represents output nerve
The number of first unit, hjHidden layer output vector is represented, W is the weights of adjustment, and δ is the threshold value for needing to adjust.
Table 1 lists the structure of the convolutional neural networks model in embodiment of the present invention and the information of each layer, has altogether
It is of five storeys, wherein distributed M in the first convolutional layercon_1=6 length are the convolution kernel of 5 sampled points, and the second convolutional layer contains
Mcon_2=12 length are the convolution kernel of 5 sampled points, and the heart due to inputting the network model is clapped length and differed, and intentionally claps
It is sent into convolutional neural networks model with single channel, for example, by using NsingleTo indicate the length of heart bat, NsingleBy first
The feature vector length N obtained after convolutional layer convolutioncon_1It indicates, the feature vector obtained after the second convolutional layer convolution is long
Use Ncon_2It indicates, Ncon_2It is 7 by the length of feature vector after the layer operation of spatial pyramid pond, realizes decentraction bat
Scale it is unified, full articulamentum is finally sent into, by Mcon_2The characteristic value that a length is 7 pull into feature that a length is 84 to
Amount.
Table 1
The present invention is placed on second by adding spatial pyramid pond layer (SPP layers) in convolutional neural networks model
Between convolutional layer and full articulamentum, ensure that fixed feature vector exports by using the pondization operation of multiple and different sizes, from
And realize the input of any scale, for example, the method that Max pooling are operated as pondization can be selected, maximum pond is exactly
Select the maximum value of feature vector as the value of Chi Huahou;
Specifically, can installation space container be μ ∈ [1,2,4], the feature vector exported after last layer convolution operation
Number be Mcon_2, each feature vector Ncon_2It indicates, spatial pyramid pond layer processes 1 feature vector
Journey is as follows:It will be fed into the SPP layers of a feature vector N firstcon_2Three parts are copied into, respectively with size Ncon_2/ μ, step-length
Ncon_2/ μ does maximum pondization operation and obtains characteristic value, finally sorts to obtain length to be N to all characteristic valuessppFeature vector.It is right
Each feature vector that second convolutional layer is sent does operation as above, the M that will finally obtaincon_2×NsppA characteristic value feeding connects entirely
Layer is connect, algorithm is as follows:
The heart that training sample is concentrated is clapped to be sent into convolutional neural networks model and is trained, due to being sent into the heart of the network
It claps size to differ, so every batch of sample number is 1, automatically generates high-level characteristic (the i.e. updated weights of backpropagation and threshold
Value), character subset is obtained, characteristic value therein is exactly continuous study, updates obtained threshold value and weights, realizes ECG signal
The extraction of feature;
Later, characteristic information (high-level characteristic) and test sample collection input convolutional neural networks mould said extracted arrived
In type, and classified to the output result of convolutional neural networks model using grader, realization divides variety classes heart bat
Class identifies, for example, Softmax functions can be selected as grader, (N), atrial premature beats (A), ventricular premature beat are clapped to the normal heart
(V), the pace-making heart claps (/), right bundle branch block (R), left bundle branch block (L) totally 6 kinds of hearts bat progress Classification and Identifications;
Specifically, in the present invention, using convolutional neural networks model to the feature extraction of ECG signal the step of is as follows:
Step 101:Arrange parameter initializes convolutional neural networks model, random numbers of the setting weights W between [0,1], threshold
Value δ is 0, and learning rate α is 0.1, setting frequency of training epochs=60;
Step 102:The heart that training sample is concentrated is clapped feeding convolutional neural networks model to be trained, due to being sent into the net
The heart of network is clapped size and is differed, so the sample number that batch is sent into every wheel training is 1, while the target of given corresponding sample is defeated
Go out vector dk(label of correct classification);
Step 103:Reality output vector is calculated using formula (1), formula (2), formula (3) and algorithm 1 (such as Fig.2)
yk, by reality output vector ykWith target output vector dkIt brings formula (4) into, calculates penalty values L;
Step 104:According to L and formula (5), formula (6), updated weights W and threshold value δ is obtained.
It repeats step 102- steps 104epochs times, obtains W, the high level that δ is automatically extracted as convolutional neural networks model
Feature.
Later, high-level characteristic and test sample collection are sent into convolutional neural networks model to test, execute a step
102- steps 104 finally classify test result feeding grader.
Wherein, in the present invention, training convolutional neural networks model needs to rely on a large amount of data, because depth network
There are many parameter, and parameter is more, and the search space of model is bigger, it is necessary to have enough data that could preferably depict model and exist
Distribution spatially is made for example, can clap the whole hearts being partitioned into from 46 records in MIT-BIH abnormal electrocardiogram databases
For data set, totally 100300 hearts are clapped, however, different classifications is concentrated with different ratios in data, wherein what the normal heart was clapped
Quantity is 73542, accounts for the 73.3% of data set, out of proportion in order to solve the problems, such as, can therefrom randomly select a part just
Normal heart umber of beats, and collectively form the heart with the abnormal heart bat in data set and clap sample set, the heart obtained after processing claps sample set such as table 2
Shown, 70% heart bat for randomly selecting training sample concentration is sent into network as training set, and remaining 30%, for testing, makees
For test sample collection;
Table 2
Type | N | / | A | V | L | R | Sum |
Heart umber of beats | 6000 | 3616 | 2480 | 6676 | 8069 | 5916 | 32757 |
The electrocardiogram beat classification recognition methods that embodiment of the present invention provides in convolutional neural networks model by drawing
Enter spatial pyramid pond layer, realizes that the data length exported to last layer (the second convolutional layer) is unified, and use Softmax letters
Number makees grader, and process training twice can be quickly obtained 83% classification accuracy, and loss function (selecting cross entropy) is received
The speed held back is fast, and 92% or more accuracy rate can be reached after training 20 times.
The electrocardiogram beat classification recognition methods that embodiment of the present invention provides, can not only solve existing convolutional Neural net
The problem of network will seek unification to input data size, and wavelet packet analysis is used in the pretreatment stage of data, wavelet packet is not
Only low frequency signal is decomposed, high-frequency signal is also decomposed, and then the raising of classification performance may be implemented.
The convolutional neural networks model of the present invention not only allows for input variable sized data, and all data can pass through list
One network training realizes weights and shares, avoids the complex operations of Multi net voting switching.The present invention is in feature extraction better than biography
The convolutional neural networks model of system, have higher nicety of grading, be sent into network model the heart clap can without artificial processing,
The size all clapped by the original heart, truly realizes artificial intelligence.
In addition, embodiment of the present invention additionally provides a kind of electrocardiogram beat classification identifying system, including:
Preprocessing module, for being pre-processed to the ECG signal in abnormal electrocardiogram database, the pretreatment packet
Include removal baseline drift processing and denoising;
The heart claps sample acquisition module, passes through the pretreated electrocardio for being chosen from the abnormal electrocardiogram database
Then figure signal extracts the heart from the process of the selection pretreated ECG signal and claps, obtains the heart and clap sample set;
Module is chosen, a part of heart bat is randomly selected in sample set as training sample set, residue for being clapped from the heart
The partial heart, which is clapped, is used as test sample collection;
First processing module is trained for inputting the training sample set in convolutional neural networks model, is realized
The extraction of ECG signal feature, wherein the convolutional neural networks model is five-layer structure, respectively the first convolutional layer, most
Great Chiization layer, the second convolutional layer, spatial pyramid pond layer, full articulamentum, wherein spatial pyramid pond layer for pair
It is unified that the data of the second convolutional layer output carry out size;
Second processing module, for extracting obtained characteristic information and the test sample collection input volume by described
In product neural network model, and classified to the output result of the convolutional neural networks model using grader, realization pair
The Classification and Identification that the variety classes heart is clapped.
Wherein, in embodiments of the present invention, the preprocessing module uses wavelet packet analysis method to the abnormal electrocardiogram
ECG signal in database carries out dry processing.
Wherein, in embodiments of the present invention, the preprocessing module includes:
Resolving cell obtains corresponding WAVELET PACKET DECOMPOSITION tree for carrying out WAVELET PACKET DECOMPOSITION to ECG signal;
Best tree extraction unit, for using principle of minimum cost to obtain Best tree from the WAVELET PACKET DECOMPOSITION tree;
Threshold value quantizing unit, for using the adaptive threshold selection method based on Stein unbiased possibility predication principles from institute
The wavelet packet analysis coefficient that high frequency series are obtained in Best tree is stated, it is the method for floating threshold value of foundation from described to use using signal energy
The wavelet packet analysis coefficient of low frequency sequence is obtained in Best tree;
Reconfiguration unit, for according to the wavelet packet analysis coefficient of the high frequency series and the wavelet packet of the low frequency sequence
Coefficient of analysis carries out the reconstruct of signal.
Wherein, in embodiments of the present invention, the grader is Softmax function category devices.
Although above having used general explanation and specific embodiment, the present invention is described in detail, at this
On the basis of invention, it can be made some modifications or improvements, this will be apparent to those skilled in the art.Therefore,
These modifications or improvements without departing from theon the basis of the spirit of the present invention belong to the scope of protection of present invention.
Claims (8)
1. a kind of electrocardiogram beat classification recognition methods, which is characterized in that including:
Step S1:ECG signal in abnormal electrocardiogram database is pre-processed, the pretreatment includes removal baseline drift
Move processing and denoising;
Step S2:It is chosen from the abnormal electrocardiogram database and passes through the pretreated ECG signal, then from described
The heart is extracted in the process pretreated ECG signal of selection to clap, and is obtained the heart and is clapped sample set;
Step S3:It is clapped from the heart and randomly selects a part of heart bat in sample set as training sample set, the heart of remainder is clapped
As test sample collection;
Step S4:The training sample set is inputted in convolutional neural networks model and is trained, realizes ECG signal feature
Extraction, wherein the convolutional neural networks model is five-layer structure, respectively the first convolutional layer, maximum pond layer, volume Two
Lamination, spatial pyramid pond layer, full articulamentum, wherein spatial pyramid pond layer is used for second convolutional layer
It is unified that the data of output carry out size;
Step S5:Obtained characteristic information and the test sample collection input convolutional neural networks model are extracted by described
In, and classified to the output result of the convolutional neural networks model using grader, what the variety classes heart was clapped in realization
Classification and Identification.
2. electrocardiogram beat classification recognition methods according to claim 1, which is characterized in that in the step S1, adopt
Dry processing is carried out to the ECG signal in the abnormal electrocardiogram database with wavelet packet analysis method.
3. electrocardiogram beat classification recognition methods according to claim 2, which is characterized in that use wavelet packet analysis method pair
ECG signal in the abnormal electrocardiogram database carries out dry processing:
WAVELET PACKET DECOMPOSITION is carried out to ECG signal, obtains corresponding WAVELET PACKET DECOMPOSITION tree;
Best tree is obtained from the WAVELET PACKET DECOMPOSITION tree using principle of minimum cost;
High frequency series are obtained from the Best tree using the adaptive threshold selection method based on Stein unbiased possibility predication principles
Wavelet packet analysis coefficient, use and obtain low frequency sequence from the Best tree by the method for floating threshold value of foundation of signal energy
Wavelet packet analysis coefficient;
Signal is carried out according to the wavelet packet analysis coefficient of the high frequency series and the wavelet packet analysis coefficient of the low frequency sequence
Reconstruct.
4. electrocardiogram beat classification recognition methods according to claim 1, which is characterized in that the grader is
Softmax function category devices.
5. a kind of electrocardiogram beat classification identifying system, which is characterized in that including:
Preprocessing module, for being pre-processed to the ECG signal in abnormal electrocardiogram database, the pretreatment includes going
Except baseline drift processing and denoising;
The heart claps sample acquisition module, for being chosen from the abnormal electrocardiogram database by the pretreated electrocardiogram letter
Number, the heart is then extracted from the process of the selection pretreated ECG signal and is clapped, and is obtained the heart and is clapped sample set;
Module is chosen, a part of heart bat is randomly selected in sample set as training sample set, remainder for being clapped from the heart
The heart clap be used as test sample collection;
First processing module is trained for inputting the training sample set in convolutional neural networks model, realizes electrocardio
The extraction of figure signal characteristic, wherein the convolutional neural networks model is five-layer structure, respectively the first convolutional layer, maximum pond
Change layer, the second convolutional layer, spatial pyramid pond layer, full articulamentum, wherein spatial pyramid pond layer is used for described
It is unified that the data of second convolutional layer output carry out size;
Second processing module, for extracting obtained characteristic information and the test sample collection input convolution god by described
Classify to the output result of the convolutional neural networks model through in network model, and using grader, realizes to difference
The Classification and Identification that the type heart is clapped.
6. electrocardiogram beat classification identifying system according to claim 5, which is characterized in that the preprocessing module uses
Wavelet packet analysis method carries out dry processing to the ECG signal in the abnormal electrocardiogram database.
7. electrocardiogram beat classification identifying system according to claim 6, which is characterized in that the preprocessing module packet
It includes:
Resolving cell obtains corresponding WAVELET PACKET DECOMPOSITION tree for carrying out WAVELET PACKET DECOMPOSITION to ECG signal;
Best tree extraction unit, for using principle of minimum cost to obtain Best tree from the WAVELET PACKET DECOMPOSITION tree;
Threshold value quantizing unit, for using the adaptive threshold selection method based on Stein unbiased possibility predication principles from it is described most
The wavelet packet analysis coefficient that high frequency series are obtained in good tree, it is the method for floating threshold value of foundation from described best to use using signal energy
The wavelet packet analysis coefficient of low frequency sequence is obtained in tree;
Reconfiguration unit, for according to the wavelet packet analysis coefficient of the high frequency series and the wavelet packet analysis of the low frequency sequence
Coefficient carries out the reconstruct of signal.
8. electrocardiogram beat classification identifying system according to claim 5, which is characterized in that the grader is
Softmax function category devices.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810404828.1A CN108647614A (en) | 2018-04-28 | 2018-04-28 | The recognition methods of electrocardiogram beat classification and system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810404828.1A CN108647614A (en) | 2018-04-28 | 2018-04-28 | The recognition methods of electrocardiogram beat classification and system |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108647614A true CN108647614A (en) | 2018-10-12 |
Family
ID=63748262
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810404828.1A Pending CN108647614A (en) | 2018-04-28 | 2018-04-28 | The recognition methods of electrocardiogram beat classification and system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108647614A (en) |
Cited By (20)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109480825A (en) * | 2018-12-13 | 2019-03-19 | 武汉中旗生物医疗电子有限公司 | The processing method and processing device of electrocardiogram (ECG) data |
CN109567789A (en) * | 2018-12-03 | 2019-04-05 | 东软集团股份有限公司 | Division processing method, device and the readable storage medium storing program for executing of ECG data |
CN109620211A (en) * | 2018-11-01 | 2019-04-16 | 吉林大学珠海学院 | A kind of intelligent abnormal electrocardiogram aided diagnosis method based on deep learning |
CN109730656A (en) * | 2019-01-09 | 2019-05-10 | 中国科学院苏州纳米技术与纳米仿生研究所 | Nerve network system, computer equipment for pulse wave signal classification |
CN109805924A (en) * | 2019-02-15 | 2019-05-28 | 济南大学 | ECG's data compression method and cardiac arrhythmia detection system based on CNN |
CN109998529A (en) * | 2019-04-15 | 2019-07-12 | 北京工业大学 | P wave, T wave detecting method in a kind of ECG signal based on Gaussian function fitting |
CN110584649A (en) * | 2019-09-09 | 2019-12-20 | 武汉大学 | Method and system for identifying abnormal electrocardiosignals in dynamic single-lead electrocardiogram |
CN110619322A (en) * | 2019-10-18 | 2019-12-27 | 武汉大学 | Multi-lead electrocardio abnormal signal identification method and system based on multi-flow convolution cyclic neural network |
CN110717416A (en) * | 2019-09-24 | 2020-01-21 | 上海数创医疗科技有限公司 | Neural network training method for ST segment classification recognition based on feature selection |
CN111261283A (en) * | 2020-01-21 | 2020-06-09 | 浙江理工大学 | Electrocardiosignal deep neural network modeling method based on pyramid type convolution layer |
CN111345815A (en) * | 2020-02-11 | 2020-06-30 | 广州视源电子科技股份有限公司 | Method, device, equipment and storage medium for detecting QRS wave in electrocardiosignal |
CN111449645A (en) * | 2020-03-07 | 2020-07-28 | 河南大学 | Intelligent classification and identification method for electrocardiogram and heartbeat |
CN111488793A (en) * | 2020-02-21 | 2020-08-04 | 广州视源电子科技股份有限公司 | Electrocardiosignal classification method and device, electronic equipment and storage medium |
CN112043260A (en) * | 2020-09-16 | 2020-12-08 | 杭州师范大学 | Electrocardiogram classification method based on local mode transformation |
CN112329609A (en) * | 2020-11-03 | 2021-02-05 | 山东大学 | Feature fusion transfer learning arrhythmia classification system based on 2D heart beat |
CN112336310A (en) * | 2020-11-04 | 2021-02-09 | 吾征智能技术(北京)有限公司 | Heart disease diagnosis system based on FCBF and SVM fusion |
CN112826513A (en) * | 2021-01-05 | 2021-05-25 | 华中科技大学 | Fetal heart rate detection system based on deep learning and specificity correction on FECG |
CN113080996A (en) * | 2021-04-08 | 2021-07-09 | 大同千烯科技有限公司 | Electrocardiogram analysis method and device based on target detection |
CN113349790A (en) * | 2020-10-16 | 2021-09-07 | 山东大学 | Heart beat classification method and system of 2D convolutional neural network based on S-shaped reconstruction |
CN114098757A (en) * | 2021-11-12 | 2022-03-01 | 南京海量物联科技有限公司 | ECG signal monitoring method based on quantum particle swarm optimization |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20060052717A1 (en) * | 2004-08-31 | 2006-03-09 | Mugler Dale H | Method and system for evaluating cardiac ischemia |
CN104615877A (en) * | 2015-01-28 | 2015-05-13 | 辽宁工程技术大学 | Method for conducting signal denoising based on wavelet packet |
CN105997055A (en) * | 2016-07-11 | 2016-10-12 | 吉林大学 | Automatic classification method, system and device of electrocardiosignal ST band |
CN107122788A (en) * | 2017-03-09 | 2017-09-01 | 吉林大学 | A kind of personal identification method and device based on electrocardiosignal |
CN107944353A (en) * | 2017-11-10 | 2018-04-20 | 西安电子科技大学 | SAR image change detection based on profile ripple BSPP networks |
CN107944493A (en) * | 2017-11-30 | 2018-04-20 | 天津大学 | A kind of the first lead electrocardiogram beat classification method based on convolutional neural networks |
-
2018
- 2018-04-28 CN CN201810404828.1A patent/CN108647614A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20060052717A1 (en) * | 2004-08-31 | 2006-03-09 | Mugler Dale H | Method and system for evaluating cardiac ischemia |
CN104615877A (en) * | 2015-01-28 | 2015-05-13 | 辽宁工程技术大学 | Method for conducting signal denoising based on wavelet packet |
CN105997055A (en) * | 2016-07-11 | 2016-10-12 | 吉林大学 | Automatic classification method, system and device of electrocardiosignal ST band |
CN107122788A (en) * | 2017-03-09 | 2017-09-01 | 吉林大学 | A kind of personal identification method and device based on electrocardiosignal |
CN107944353A (en) * | 2017-11-10 | 2018-04-20 | 西安电子科技大学 | SAR image change detection based on profile ripple BSPP networks |
CN107944493A (en) * | 2017-11-30 | 2018-04-20 | 天津大学 | A kind of the first lead electrocardiogram beat classification method based on convolutional neural networks |
Cited By (28)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109620211A (en) * | 2018-11-01 | 2019-04-16 | 吉林大学珠海学院 | A kind of intelligent abnormal electrocardiogram aided diagnosis method based on deep learning |
CN109567789A (en) * | 2018-12-03 | 2019-04-05 | 东软集团股份有限公司 | Division processing method, device and the readable storage medium storing program for executing of ECG data |
CN109480825B (en) * | 2018-12-13 | 2021-08-06 | 武汉中旗生物医疗电子有限公司 | Electrocardio data processing method and device |
CN109480825A (en) * | 2018-12-13 | 2019-03-19 | 武汉中旗生物医疗电子有限公司 | The processing method and processing device of electrocardiogram (ECG) data |
CN109730656A (en) * | 2019-01-09 | 2019-05-10 | 中国科学院苏州纳米技术与纳米仿生研究所 | Nerve network system, computer equipment for pulse wave signal classification |
CN109805924A (en) * | 2019-02-15 | 2019-05-28 | 济南大学 | ECG's data compression method and cardiac arrhythmia detection system based on CNN |
CN109998529A (en) * | 2019-04-15 | 2019-07-12 | 北京工业大学 | P wave, T wave detecting method in a kind of ECG signal based on Gaussian function fitting |
CN109998529B (en) * | 2019-04-15 | 2021-11-05 | 北京工业大学 | Method for detecting P wave and T wave in ECG signal based on Gaussian function fitting |
CN110584649A (en) * | 2019-09-09 | 2019-12-20 | 武汉大学 | Method and system for identifying abnormal electrocardiosignals in dynamic single-lead electrocardiogram |
CN110717416A (en) * | 2019-09-24 | 2020-01-21 | 上海数创医疗科技有限公司 | Neural network training method for ST segment classification recognition based on feature selection |
CN110619322A (en) * | 2019-10-18 | 2019-12-27 | 武汉大学 | Multi-lead electrocardio abnormal signal identification method and system based on multi-flow convolution cyclic neural network |
CN111261283B (en) * | 2020-01-21 | 2023-05-09 | 浙江理工大学 | Electrocardiosignal deep neural network modeling method based on pyramid convolution layer |
CN111261283A (en) * | 2020-01-21 | 2020-06-09 | 浙江理工大学 | Electrocardiosignal deep neural network modeling method based on pyramid type convolution layer |
CN111345815A (en) * | 2020-02-11 | 2020-06-30 | 广州视源电子科技股份有限公司 | Method, device, equipment and storage medium for detecting QRS wave in electrocardiosignal |
CN111345815B (en) * | 2020-02-11 | 2023-05-02 | 广州视源电子科技股份有限公司 | Method, device, equipment and storage medium for detecting QRS wave in electrocardiosignal |
CN111488793A (en) * | 2020-02-21 | 2020-08-04 | 广州视源电子科技股份有限公司 | Electrocardiosignal classification method and device, electronic equipment and storage medium |
CN111449645A (en) * | 2020-03-07 | 2020-07-28 | 河南大学 | Intelligent classification and identification method for electrocardiogram and heartbeat |
CN111449645B (en) * | 2020-03-07 | 2021-07-06 | 河南大学 | Intelligent classification and identification method for electrocardiogram and heartbeat |
CN112043260B (en) * | 2020-09-16 | 2022-11-15 | 杭州师范大学 | Electrocardiogram classification method based on local mode transformation |
CN112043260A (en) * | 2020-09-16 | 2020-12-08 | 杭州师范大学 | Electrocardiogram classification method based on local mode transformation |
CN113349790A (en) * | 2020-10-16 | 2021-09-07 | 山东大学 | Heart beat classification method and system of 2D convolutional neural network based on S-shaped reconstruction |
CN112329609A (en) * | 2020-11-03 | 2021-02-05 | 山东大学 | Feature fusion transfer learning arrhythmia classification system based on 2D heart beat |
CN112336310A (en) * | 2020-11-04 | 2021-02-09 | 吾征智能技术(北京)有限公司 | Heart disease diagnosis system based on FCBF and SVM fusion |
CN112336310B (en) * | 2020-11-04 | 2024-03-08 | 吾征智能技术(北京)有限公司 | FCBF and SVM fusion-based heart disease diagnosis system |
CN112826513A (en) * | 2021-01-05 | 2021-05-25 | 华中科技大学 | Fetal heart rate detection system based on deep learning and specificity correction on FECG |
CN113080996A (en) * | 2021-04-08 | 2021-07-09 | 大同千烯科技有限公司 | Electrocardiogram analysis method and device based on target detection |
CN114098757A (en) * | 2021-11-12 | 2022-03-01 | 南京海量物联科技有限公司 | ECG signal monitoring method based on quantum particle swarm optimization |
CN114098757B (en) * | 2021-11-12 | 2024-02-09 | 南京海量物联科技有限公司 | ECG signal monitoring method based on quantum particle swarm optimization |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108647614A (en) | The recognition methods of electrocardiogram beat classification and system | |
CN107822622B (en) | Electrocardiogram diagnosis method and system based on deep convolutional neural network | |
CN104523266B (en) | A kind of electrocardiosignal automatic classification method | |
CN111062250B (en) | Multi-subject motor imagery electroencephalogram signal identification method based on deep feature learning | |
CN111000553B (en) | Intelligent classification method for electrocardiogram data based on voting ensemble learning | |
CN109948647A (en) | A kind of electrocardiogram classification method and system based on depth residual error network | |
CN109934089A (en) | Multistage epileptic EEG Signal automatic identifying method based on supervision gradient lifter | |
CN101859377A (en) | Electromyographic signal classification method based on multi-kernel support vector machine | |
CN111582396B (en) | Fault diagnosis method based on improved convolutional neural network | |
CN106845528A (en) | A kind of image classification algorithms based on K means Yu deep learning | |
CN108171318A (en) | One kind is based on the convolutional neural networks integrated approach of simulated annealing-Gaussian function | |
CN112022153B (en) | Electroencephalogram signal detection method based on convolutional neural network | |
CN108304887A (en) | Naive Bayesian data processing system and method based on the synthesis of minority class sample | |
CN110313894A (en) | Arrhythmia cordis sorting algorithm based on convolutional neural networks | |
CN107423815A (en) | A kind of computer based low quality classification chart is as data cleaning method | |
CN113011386B (en) | Expression recognition method and system based on equally divided characteristic graphs | |
CN109858518A (en) | A kind of large data clustering method based on MapReduce | |
CN110288028A (en) | ECG detecting method, system, equipment and computer readable storage medium | |
CN107301409A (en) | Learn the system and method for processing electrocardiogram based on Wrapper feature selectings Bagging | |
CN106599903A (en) | Correlation weighed least squares-dictionary learning signal reconstruction method | |
CN108388918A (en) | Data characteristics selection method with structure retention performance | |
CN112215082A (en) | Plant leaf image identification method | |
Hassan et al. | A hybrid data mining approach for knowledge extraction and classification in medical databases | |
CN116421200A (en) | Brain electricity emotion analysis method of multi-task mixed model based on parallel training | |
Valério et al. | Deepmammo: deep transfer learning for lesion classification of mammographic images |
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: 20181012 |
|
RJ01 | Rejection of invention patent application after publication |