CN108256452A - A kind of method of the ECG signal classification of feature based fusion - Google Patents
A kind of method of the ECG signal classification of feature based fusion Download PDFInfo
- Publication number
- CN108256452A CN108256452A CN201810012811.1A CN201810012811A CN108256452A CN 108256452 A CN108256452 A CN 108256452A CN 201810012811 A CN201810012811 A CN 201810012811A CN 108256452 A CN108256452 A CN 108256452A
- Authority
- CN
- China
- Prior art keywords
- feature
- sample
- features
- dimensional
- extraction
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
-
- 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
- G06F2218/04—Denoising
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/15—Biometric patterns based on physiological signals, e.g. heartbeat, blood flow
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computational Linguistics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
Abstract
The invention discloses a kind of methods of the ECG signal classification of feature based fusion, include the following steps：Electrocardiogram (ECG) data in the original continuous time is pre-processed to obtain the discrete values sample of holocentric jump, and to these one-dimensional signal sample extraction 1D CNN convolution features and PQRST numerical characteristics；Mixing operation is carried out to two kinds of features of extraction, makes different type, the feature integration one of different dimensions, the representative feature set jumped as holocentric；Each sample of test set and training set passes through Fusion Features, can obtain the feature vector of one 100 dimension；Classify to the sample after Fusion Features, obtain the nicety of grading of each electrocardiosignal.The present invention extracts one-dimensional convolution feature and PQRST features and the method for blending the two, so as to improve the accuracy rate of electrocardiosignal computer-aided diagnosis system.
Description
Technical field
The present invention relates to machine learning techniques and biomedical engineering technology fields more particularly to a kind of feature based to merge
ECG (electrocardiogram) Modulation recognition method.
Background technology
Angiocardiopathy caused by cardiac arrhythmia is the big health problem that the whole world faces, it can cause sufferer temporary
Shock is even died suddenly.Currently, Precise Diagnosis is to cope with the maximally efficient measure of angiocardiopathy with treatment in time.ECG is current
The means of most important detect and diagnose heart disease.However, a large amount of image informations generated in the inspection of disease easily make doctor
Raw fatigue, and diagnostic accuracy is influenced by subjective factors such as vocational ability, the experiences of doctor.In this context, pass through machine learning
Method judges whether heart is problematic or concrete type of heart disease obtains the research hotspot of extensive concern as one.
It is specific belonging to detected person's heartbeat in the computer-aided diagnosis technology of heart disease especially arrhythmia cordis
Type needs to realize by sorting technique.To obtain good accuracy rate of diagnosis, need to find can be with the accurate description arrhythmia cordis heart
Jump the characteristic quantity of sample.The feature of common description heartbeat includes morphological feature, textural characteristics and wavelet character etc..Patent " one
Kind electrocardiosignal classifying identification method " (CN 107184198A) carries out electrocardiogram rhythm and pace of moving things information to original electrocardiographicdigital figure Wave data
With the extraction of PQRST (five basic waves of electrocardiogram) waveform, the digitlization number of electrocardiogram rhythm and pace of moving things information and PQRST waveforms is obtained
According to so as to complete the Classification and Identification of electrocardiosignal, having played the effect of morphological feature well.In general, more feature dimensions
Number can make grader modeling more accurately with stabilization, so as to improve classification accuracy.But in fact, when intrinsic dimensionality is higher,
Wherein there may be interdepend or with the incoherent redundancy feature of class object, the presence of these features can so that operation is complicated
Degree raising causes to analyze feature and the time lengthening needed for training pattern.Meanwhile the disaggregated model of structure also can be more complicated,
Its generalization ability is caused to decline, dimension disaster occurs.By selecting the feature extracted from heartbeat signal, it is uncorrelated or
The feature of redundancy can be removed, and so as to reach reduction Characteristic Number, improved model accuracy and reduced the purpose of run time.Cause
This, in recent years in the relevant multinomial research of heartbeat signal, feature selecting is widely applied.Wherein, patent " feature based
The arrhythmia classification method of selection " (CN106377247A) calculates each spy using the method for Relif (feature weight algorithm)
Levy weight, initialization of population instructed according to feature weight, and according to individual adaptation degree quality according to select probability, crossover probability and
Mutation probability selected respectively, intersect and mutation operation obtains the next generation, so as to achieve the purpose that feature selecting.
But above research is all based on the method for manually extracting ECG signal feature, and feature extraction and selection scheme are being tested
In selected with being combined, this not only needs largely to be verified, also test people analysis and combination ability.It and will
Feature extraction is separately carried out with feature selecting, the influence of the more and easy external environment of step.And the depth to grow up in recent years
Habit technology compensates for this defect just, it is to integrate feature extraction and selection, and the original sample of input can be carried automatically
Very high level conceptual, most representative feature are taken, although deep network takes larger, the feature redundancy of extraction is small and operates
Simply, the raising of nicety of grading is acted on notable.Patent " a kind of removable electrocardiogram monitoring system and monitoring method "
(CN106344005A) characteristic features are extracted using CNN (convolutional neural networks) network for cardiac electrical numeric type signal,
The speed of service is fast and can be used for the transplanting of wearable device.But the feature of this method extraction is mainly for local feature, to key
Point position feature, is not to pay close attention to very much if the numerical value of PQRST positions, and the numerical value of this key point position at five plays heartbeat type
Decisive role is arrived, the location point information value of each heartbeat is different.Thus the type feature to the effect of heartbeat compared with
It, can not be ignored to be important.
Invention content
The present invention provides a kind of method of the ECG signal classification of feature based fusion, the present invention extracts one-dimensional convolution
Feature and PQRST features and the method for blending the two, so as to improve the accurate of electrocardiosignal computer-aided diagnosis system
Rate, it is described below：
A kind of method of the ECG signal classification of feature based fusion, the described method comprises the following steps：
Electrocardiogram (ECG) data in the original continuous time is pre-processed to obtain the discrete values sample of holocentric jump, and one-dimensional to these
Sample of signal extracts 1D-CNN convolution feature and PQRST numerical characteristics；
Mixing operation is carried out to two kinds of features of extraction, makes different type, the feature integration one of different dimensions, as list
The representative feature set of heartbeat；Each sample of test set and training set pass through Fusion Features, can obtain one 100 dimension feature to
Amount；
Classify to the sample after Fusion Features, obtain the nicety of grading of each electrocardiosignal.
The electrocardiogram (ECG) data in the original continuous time pre-processes to obtain the discrete values sample of holocentric jump, and to these
One-dimensional signal sample extraction 1D-CNN convolution feature and PQRST numerical characteristics are specially：
Original heartbeat data is acted on the junction filter of low frequency and high frequency, it is therefore an objective to remove data noise and baseline drift
It moves；
The heartbeat data collection of single people in continuous time after processing is split, obtains the data sample that a large amount of holocentrics are jumped
This；
Single heartbeat data after segmentation, the one-dimensional convolution of height of the extraction abstract are acted on using one-dimensional convolutional neural networks
Feature；Single heartbeat data of the same purpose after segmentation obtains what holocentric was jumped using different frequency and the frequency window of initial position
PQRST features.
The 1D-CNN convolution feature is specially：
1) convolutional layer：
First convolutional layer sets 5 convolutional channels, and the convolution kernel of each channel is disposed as 21*1；Second convolutional layer
There is provided 16 convolutional channels, the convolution kernel size per channel is 13*1；Third convolutional layer is then provided with 20 convolutional channels,
Each channel convolution kernel is 9*1；The convolution kernel of one layer of single channel 1*1 size, Jiang Getong are added after third convolutional layer
Road feature integration is one-dimensional characteristic vector；
2) full articulamentum：
The one-dimensional characteristic vector of integration abstract characteristics and classifies to sample in last layer again；Using two layers
Artificial nerve network model is set as every layer of neuron 75,5, that is, the convolution feature after integrating shares 75 dimensions, reaches and is divided into
The purpose of 5 class diseases.
It is described extraction PQRST features be specially：
Using the time window of 60Hz, 10 QRS waves are extracted from R-50ms to R+100ms, wherein R is the position of ECG R wave
It puts；For T wave characteristics, the time window of 20Hz is fixed on the position from R+150ms to R+500ms, obtains 8 characteristic points；
The characteristic point of 7 P waves is extracted, window is from R-200ms to R-100ms, sample rate 60Hz；
Each samples pictures generate 25 dimensional features, and 25 dimensional feature is by the full auxiliary 1D-CNN features that are used to electrocardiosignal
Classify.
Two kinds of features of described pair of extraction carry out mixing operation：
The feature for the dense layers that neuron in 1D-CNN networks is 75 is chosen, then by this 75 Victoria C NN feature and 25 dimensions
PQRST features are merged, and obtain 100 dimensional feature vectors of each sample.
The advantageous effect of technical solution provided by the invention is：
1st, the present invention is directed to heart disease computer-aided diagnosis, while the classification capacity and physical significance of considering feature,
It proposes to one-dimensional signal extraction 1D-CNN (one-dimensional convolutional neural networks) features and PQRST features, and two kinds of features is melted
Conjunction obtains one-dimensional characteristic vector；Traditional classifier is acted on the fusion feature can obtain the classification situation of holocentric jump collection；
2nd, compare through experiment, when applied to classification diagnosis, the present invention has a clear superiority in multiple evaluation indexes.
Description of the drawings
Fig. 1 is one-dimensional convolution frame diagram；
Fig. 2 is the PQRST feature extraction locations drawing；
Fig. 3 is the execution flow chart of steps of the present invention.
Specific embodiment
To make the object, technical solutions and advantages of the present invention clearer, embodiment of the present invention is made below further
It is described in detail on ground.
According to above-mentioned analysis, the embodiment of the present invention considers to obtain single heartbeat, extraction to the pretreatment of continuous heartbeat signal
The PQRST location point numerical characteristics that holocentric is jumped, the representativeness that one-dimensional convolutional neural networks (1D-CNN) extraction holocentric is recycled to jump
Two characteristic use Feature Fusions are carried out mixing operation by feature later, and last application class device acts on the feature of extraction
Complete the classification jumped to holocentric.
Embodiment 1
The purpose of the embodiment of the present invention is based on one-dimensional electrocardiosignal, extracts one-dimensional convolution feature and PQRST features,
And the algorithm for blending the two, so as to improve the accuracy rate of electrocardiosignal computer-aided diagnosis system.The embodiment of the present invention
Technical solution it is as follows：
101：Electrocardiogram (ECG) data in the original continuous time is pre-processed to obtain the discrete values sample of holocentric jump, and to these
One-dimensional signal sample extraction 1D-CNN convolution feature and PQRST numerical characteristics；
Specifically comprise the following steps：
(1) Signal Pretreatment：Original heartbeat data is acted on the junction filter of low frequency and high frequency, it is therefore an objective to remove data
Noise and baseline drift；
(2) signal is divided：According to the beat pulse situation of normal person, to the beats of single people in continuous time after processing
It is split according to collection, so as to obtain the data sample that a large amount of holocentrics are jumped.A variety of different types of diseases may be included in the sample,
Multiple types are integrated into five major class diseases of NSVFQ for judging by the present invention according to internationalization criteria for classification (AAMI standard).
(3) signal characteristic abstraction：Single heartbeat data after segmentation, height of the extraction are acted on using one-dimensional convolutional neural networks
The one-dimensional convolution of abstract is special；Single heartbeat data of the same purpose after segmentation, utilizes different frequency and the frequency of initial position
Window obtains the PQRST features that holocentric is jumped.
102：Mixing operation is carried out to two kinds of features of extraction, makes different type, the feature integration one of different dimensions, makees
The representative feature set jumped for holocentric；
Fusion Features of each sample of test set and training set Jing Guo the embodiment of the present invention, available one 100 dimension
Feature vector, the vector characterize the attribute of sample well, and correlation with each other is very low between feature and whole redundancy is small.
103：Classify to the sample after Fusion Features, obtain the nicety of grading of each electrocardiosignal, according to more classification
The judgment criteria of signal understands classification situation of all categories and whole, so as to verify the feasibility of this method than accurate.
In conclusion after processing of the ECG raw data sets Jing Guo above step 101- steps 103, holocentric jump can obtain
The representative feature and its classification accuracy of collection, obtained classification measurement is than the more objective superiority for giving expression to this method.
Embodiment 2
The 1D-CNN features and PQRST features that the embodiment of the present invention extracts one-dimensional electrocardiosignal using Feature Fusion
It is merged, and is classified with more classical classification to fusion feature, compared to the classification situation that one-dimensional characteristic is applied alone, the present invention
The feature of embodiment extraction has higher nicety of grading.Below in conjunction with the accompanying drawings 1 and 2, the embodiment of the present invention 1 is made further in detail
Thin explanation：
201：Data set is pre-processed and is divided, obtains one-dimensional single heartbeat data collection containing 288 data points；
The embodiment of the present invention uses public database MIT-BIH (known to those skilled in the art, the embodiment of the present invention
This is not repeated) in arrhythmia cordis data set, wherein having the monitoring data of 48 subject's each half an hours, more than totally 10 ten thousand
A holocentric is jumped.
Experiment will be training set (SD1) and test set (SD2) according to the different demarcation of patient.Wherein, SD1 is number
101,106,108,109,112,114,115,116,118,119,122,124,201,203,205,207,208,209,215,
220,223 and 230 subject.SD2 is number 100,103,105,111,113,117,121,123,200,202,210,
212,213,214,219,221,222,228,231,232,233 and 234 subject.
The embodiment of the present invention removes the baseline drift of above-mentioned data using median filter, and work is removed using low-pass filter
Frequency interference and high-frequency noise.According to the palmic rate of normal person, i.e., the heartbeat signal after 75 (beat/min) segmentation filterings, at R peaks
Value both sides respectively take 90 and 198 data points, and one-dimensional single heartbeat data collection containing 288 data points is obtained with this.
202：Obtain test and training data；
The model split data set that database used in the embodiment of the present invention is told about according to above-mentioned steps 201.Wherein, training set
Including 5020 single heartbeat datas, test set includes 4947 single heartbeat datas, and two data set sample proportions are close to 1:1, each other
Between without intersect data.Sample distribution between data set is substantially the same, correct in training pattern to later stage test set sample
Classification provides guarantee.
203：1D-CNN models are built to sample：
Experiment one-dimensional convolutional neural networks used mainly handle one-dimensional continuous or discrete signal, and the feature of extraction is height
Abstract separated data, process can be represented with equation below：
Y (k)=h (k) * u (k)=∑ h (k-i) u (i), i ∈ (0, N)
In formula, h (k) is convolution kernel sequence, and u (k) is to be convolved signal sequence, and K values are mutually overlapped for two sequences at least 1
The maximum length of conjunction.
Two sequence convolution processes are to be added to obtain after respective items are multiplied.Parameter increases after convolution, in order to reduce memory
Expense, each convolutional layer are used for down-sampling followed by a pond layer.The effect of full articulamentum is the parameter before integrating this layer
Last classification results are exported for one-dimensional vector and according to artificial neural network principle.The arrangement of specific layer is as follows：
1) convolutional layer：
First convolutional layer sets 5 convolutional channels, and the convolution kernel of each channel is disposed as 21*1；Second convolutional layer
There is provided 16 convolutional channels, the convolution kernel size per channel is 13*1；Third convolutional layer is then provided with 20 convolutional channels,
Each channel convolution kernel is 9*1；
The convolution algorithm effect of three first layers is continuous abstract data characteristics, and the mechanism taken is that port number gradually increases,
And the size of convolution kernel is gradually reduced, which can complete feature the effect of extraction, but until third convolutional layer, feature
The form of expression is still multichannel parallel fashion, and such a feature cannot be directly used to sort operation.Therefore, it is added after third layer
The convolution kernel of one layer of single channel 1*1 size, it is therefore an objective to which each channel characteristics are integrated into one-dimensional characteristic vector, the fusion after facilitating
Operation.
2) full articulamentum：
For this layer similar to artificial neural network, main purpose is the one-dimensional characteristic vector by integration according to multilayer neural network
Characteristic, abstract characteristics and classify again in last layer to sample.This part uses two layers of artificial neural network mould
Type is set as 75,5 to every layer of neuron.Convolution feature after integrating shares 75 dimensions, finally reaches the disease that falls into 5 types
Purpose.
3) to sample extraction PQRST features：
To the samples pictures of input, with three different sample frequencys and the time window intercepting message point of initial position.Such as Fig. 2
As display, using the time window of 60Hz, 10 QRS waves are extracted from R-50ms to R+100ms, wherein R is ECG R wave
Position.For T wave characteristics, the time window of 20Hz is fixed on the position from R+150ms to R+500ms by the embodiment of the present invention, is obtained
To 8 characteristic points.
Finally, the embodiment of the present invention is extracted the characteristic point of 7 P waves, and window is from R-200ms to R-100ms, sample rate
For 60Hz.On the whole, 25 dimensional features can be generated by each samples pictures of aforesaid operations.25 dimensional feature will be used to auxiliary entirely
1D-CNN features is helped to classify electrocardiosignal.
4) fusion 1D-CNN features and PQRST features：
Traditional Fusion Features mode is that multiple feature vectors are integrated into one-dimensional characteristic vector for classifying.Therefore, originally
Two kinds of multidimensional characteristic also using which, with nonterminal character integration technology, is integrated into one-dimensional spy by inventive embodiments
Sign.Choose the feature of dense (full connection) layer that neuron in 1D-CNN networks is 75, then by this 75 Victoria C NN feature with
25 dimension PQRST features are merged, and obtain 100 dimensional feature vectors of each sample.That is, training sample and test sample
Feature have 100 dimensions.
5) classifier design：
Because there are the quantity of serious energy imbalance, i.e. multiclass sample and few class sample between electrocardiogram (ECG) data sample class
Ratio nearly 100:1.Influence of the unbalanced sample to classification is more serious, and especially general grader only balances sample
Data Classifying Quality it is good, it is and poor to unbalanced sample classification.Fortunately random forest grader (RF) is to sample imbalance
Data sensitive it is relatively low, based on this, the embodiment of the present invention classifies to the sample after Fusion Features using RF graders.
The algorithm is the set of many weak decision trees.Wherein, total training set according to the model split put back into certain number
The training subset of amount, each subset build decision tree by the way of not beta pruning.In the node process of training tree, the spy that uses
Sign takes the mode for having no put back in proportion to be extracted from the feature set of training subset, to ensure the integrality of feature with not weighing
It is multiple.When sample to be tested enters the grader, classification will be carried out with all decision trees and will be compared, each tree gives sample one mark
Label count label quantity of all categories, the most as final classification of the sample of poll.As a result, RF graders can by samples selection,
Decision tree generation, model combination and model verify that four parts are formed, and detailed process is as follows：
1st, samples selection：Assuming that original training set has n d dimensional feature sample, n sample of the extraction put back to every time is as root
Set, there may be identical samples for the set, and the set so as to choose every time is not whole samples, avoids over-fitting.
2nd, achievement process：Nothing selects D dimensional vectors (D with putting back to from the d dimensional feature vectors of a certain training subset at random<D),
The branch of tree is built by the way of fully nonlinear water wave according to this D dimensional feature, until reaching end condition or all samples of the node
It is similar.
3rd, model combines：It repeats 1,2 step k times, k different decision tree can be obtained.Each tree occupies equal decision status,
RF graders are the set of the k decision tree, and each tree will also participate in the decision process of RF graders.
4th, model is verified：Verification set will be become by having neither part nor lot in the sample of decision tree training, as verification classifier performance
New samples.
When new samples enter the grader, all k decision trees judge it and give label, final to use
The most label of poll is assigned to sample to be tested by the mode of ballot, i.e., one kind of probability summation maximum in all labels.
In conclusion the embodiment of the present invention extracts one-dimensional convolution feature and PQRST features and the side for blending the two
Method, so as to improve the accuracy rate of electrocardiosignal computer-aided diagnosis system.
Embodiment 3
With reference to table 1, specific calculation formula, feasibility verification is carried out to the scheme in Examples 1 and 2, is referred to down
Text description：
Table 1 gives the sorting technique of patent " the arrhythmia classification method of feature based selection " (CN106377247A)
The comparing result of the sorting technique proposed with the embodiment of the present invention.Two methods carry out real in identical test and training set
It tests, experiment evaluates classification results using following index：
Wherein, whole accuracy rate is used to show the situation of all sample classifications, and sensitivity and specificity are every for weighing
The classification situation of class sample, i.e. five major class cardiac arrhythmias are represented respectively with sensitivity and specificity.
1 tagsort performance of table compares
This method has apparent excellent on indices compared to patent CN106377247A it can be seen from 1 data of table
Gesture, therefore the feature set obtained through method has better classification capacity.
It will be appreciated by those skilled in the art that attached drawing is the schematic diagram of a preferred embodiment, the embodiments of the present invention
Serial number is for illustration only, does not represent the quality of embodiment.
The foregoing is merely presently preferred embodiments of the present invention, is not intended to limit the invention, it is all the present invention spirit and
Within principle, any modification, equivalent replacement, improvement and so on should all be included in the protection scope of the present invention.
Claims (5)
- A kind of 1. method of the ECG signal classification of feature based fusion, which is characterized in that the described method comprises the following steps：Electrocardiogram (ECG) data in the original continuous time is pre-processed to obtain the discrete values sample of holocentric jump, and to these one-dimensional signals Sample extraction 1D-CNN convolution feature and PQRST numerical characteristics；Mixing operation is carried out to two kinds of features of extraction, makes different type, the feature integration one of different dimensions, is jumped as holocentric Representative feature set；Each sample of test set and training set passes through Fusion Features, can obtain the feature vector of one 100 dimension；Classify to the sample after Fusion Features, obtain the nicety of grading of each electrocardiosignal.
- 2. the method for the ECG signal classification of a kind of feature based fusion according to claim 1, which is characterized in that described Electrocardiogram (ECG) data in the original continuous time is pre-processed to obtain the discrete values sample of holocentric jump, and to these one-dimensional signal samples Extraction 1D-CNN convolution feature and PQRST numerical characteristics are specially：Original heartbeat data is acted on the junction filter of low frequency and high frequency, it is therefore an objective to remove data noise and baseline drift；The heartbeat data collection of single people in continuous time after processing is split, obtains the data sample that a large amount of holocentrics are jumped；Single heartbeat data after segmentation is acted on using one-dimensional convolutional neural networks, the one-dimensional convolution of height of the extraction abstract is special Sign；Single heartbeat data of the same purpose after segmentation obtains what holocentric was jumped using different frequency and the frequency window of initial position PQRST features.
- 3. the method for the ECG signal classification of a kind of feature based fusion according to claim 1, which is characterized in that described 1D-CNN convolution features are specially：1) convolutional layer：First convolutional layer sets 5 convolutional channels, and the convolution kernel of each channel is disposed as 21*1；Second convolutional layer setting 16 convolutional channels, the convolution kernel size per channel are 13*1；Third convolutional layer is then provided with 20 convolutional channels, each Channel convolution kernel is 9*1；The convolution kernel of one layer of single channel 1*1 size is added after third convolutional layer, by each channel spy Sign is integrated into one-dimensional characteristic vector；2) full articulamentum：The one-dimensional characteristic vector of integration abstract characteristics and classifies to sample in last layer again；It is artificial using two layers Neural network model is set as every layer of neuron 75,5, that is, the convolution feature after integrating shares 75 dimensions, reaches 5 classes of classification The purpose of disease.
- 4. the method for the ECG signal classification of a kind of feature based fusion according to claim 1, which is characterized in that described Extracting PQRST features is specially：Using the time window of 60Hz, 10 QRS waves are extracted from R-50ms to R+100ms, wherein R is the position of ECG R wave；It is right In T wave characteristics, the time window of 20Hz is fixed on the position from R+150ms to R+500ms, obtains 8 characteristic points；The characteristic point of 7 P waves is extracted, window is from R-200ms to R-100ms, sample rate 60Hz；Each samples pictures generate 25 dimensional features, which carries out electrocardiosignal the full auxiliary 1D-CNN features that are used to Classification.
- 5. the method for the ECG signal classification of a kind of feature based fusion according to claim 1, which is characterized in that described Carrying out mixing operation to two kinds of features of extraction is specially：The feature for the dense layers that neuron in 1D-CNN networks is 75 is chosen, then by this 75 Victoria C NN feature and 25 dimension PQRST Feature is merged, and obtains 100 dimensional feature vectors of each sample.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810012811.1A CN108256452A (en) | 2018-01-06 | 2018-01-06 | A kind of method of the ECG signal classification of feature based fusion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810012811.1A CN108256452A (en) | 2018-01-06 | 2018-01-06 | A kind of method of the ECG signal classification of feature based fusion |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108256452A true CN108256452A (en) | 2018-07-06 |
Family
ID=62725873
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810012811.1A Pending CN108256452A (en) | 2018-01-06 | 2018-01-06 | A kind of method of the ECG signal classification of feature based fusion |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108256452A (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109497986A (en) * | 2018-11-22 | 2019-03-22 | 杭州脉流科技有限公司 | Electrocardiogram intelligent analysis method, device, computer equipment and system based on deep neural network |
| 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 |
| CN109620152A (en) * | 2018-12-16 | 2019-04-16 | 北京工业大学 | A kind of electrocardiosignal classification method based on MutiFacolLoss-Densenet |
| CN110664395A (en) * | 2019-09-29 | 2020-01-10 | 京东方科技集团股份有限公司 | Image processing method, image processing apparatus, and storage medium |
| CN110840402A (en) * | 2019-11-19 | 2020-02-28 | 山东大学 | Atrial fibrillation signal identification method and system based on machine learning |
| CN111557660A (en) * | 2020-06-08 | 2020-08-21 | 东北大学 | Arrhythmia identification method under sub-population deep learning framework |
| CN111626114A (en) * | 2020-04-20 | 2020-09-04 | 哈尔滨工业大学 | Electrocardiosignal arrhythmia classification system based on convolutional neural network |
| CN111919224A (en) * | 2020-06-30 | 2020-11-10 | 北京小米移动软件有限公司 | Biological feature fusion method and device, electronic equipment and storage medium |
| CN112932431A (en) * | 2021-01-26 | 2021-06-11 | 山西三友和智慧信息技术股份有限公司 | Heart rate identification method based on 1DCNN + Inception Net + GRU fusion network |
| CN113133767A (en) * | 2021-03-22 | 2021-07-20 | 浙江工业大学 | Electrocardiosignal classification method combining wave band extraction and multi-feature fusion |
| WO2022052300A1 (en) * | 2020-09-08 | 2022-03-17 | 浙江大学山东工业技术研究院 | Electrocardio parameter calculation method based on deep learning |
| CN115211866A (en) * | 2022-09-07 | 2022-10-21 | 西南民族大学 | Arrhythmia classification method and system and electronic equipment |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20120179059A1 (en) * | 2010-02-03 | 2012-07-12 | Zhijun Gu | Analyzing Electrocardiogram Signals |
| CN106344005A (en) * | 2016-10-28 | 2017-01-25 | 张珈绮 | Mobile ECG (electrocardiogram) monitoring system and monitoring method |
| CN107137072A (en) * | 2017-04-28 | 2017-09-08 | 北京科技大学 | A kind of ventricular ectopic beating detection method based on 1D convolutional neural networks |
| CN107184198A (en) * | 2017-06-01 | 2017-09-22 | 广州城市职业学院 | A kind of electrocardiosignal classifying identification method |
-
2018
- 2018-01-06 CN CN201810012811.1A patent/CN108256452A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20120179059A1 (en) * | 2010-02-03 | 2012-07-12 | Zhijun Gu | Analyzing Electrocardiogram Signals |
| CN106344005A (en) * | 2016-10-28 | 2017-01-25 | 张珈绮 | Mobile ECG (electrocardiogram) monitoring system and monitoring method |
| CN107137072A (en) * | 2017-04-28 | 2017-09-08 | 北京科技大学 | A kind of ventricular ectopic beating detection method based on 1D convolutional neural networks |
| CN107184198A (en) * | 2017-06-01 | 2017-09-22 | 广州城市职业学院 | A kind of electrocardiosignal classifying identification method |
Non-Patent Citations (5)
| Title |
|---|
| DAN LI,ET.AL: ""Classification of ECG Signals Based on 1D Convolution Neural Network"", 《2017 IEEE 19TH INTERNATIONAL CONFERENCE ON E-HEALTH NETWORKING, APPLICATIONS AND SERVICES (HEALTHCOM)》 * |
| MUHAMMAD ZUBAIR ET.AL: ""An Automated ECG Beat Classification System Using Convolutional Neural Networks"", 《2016 IEEE》 * |
| QINGXUE ZHANG ET.AL: ""HeartID: A Multiresolution Convolutional Neural Network for ECG-Based Biometric Human Identification in Smart Health Applications"", 《SPECIAL SECTION ON BODY AREA NETWORKS》 * |
| SERKAN KIRANYAZ ET.AL: ""Convolutional Neural Networks for Patient-Specific ECG Classification"", 《2015 37TH ANNUAL INTERNATIONAL CONFERENCE OF THE IEEE，ENGINEERING IN MEDICINE AND BIOLOGY SOCIETY(EMBC)》 * |
| WEI LU ET AL: ""Feature fusion for imbalanced ECG data analysis"", 《BIOMEDICAL SIGNAL PROCESSING CONTROL》 * |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109497986A (en) * | 2018-11-22 | 2019-03-22 | 杭州脉流科技有限公司 | Electrocardiogram intelligent analysis method, device, computer equipment and system based on deep neural network |
| 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 |
| CN109620152A (en) * | 2018-12-16 | 2019-04-16 | 北京工业大学 | A kind of electrocardiosignal classification method based on MutiFacolLoss-Densenet |
| CN110664395A (en) * | 2019-09-29 | 2020-01-10 | 京东方科技集团股份有限公司 | Image processing method, image processing apparatus, and storage medium |
| CN110664395B (en) * | 2019-09-29 | 2022-04-12 | 京东方科技集团股份有限公司 | Image processing method, image processing apparatus, and storage medium |
| CN110840402A (en) * | 2019-11-19 | 2020-02-28 | 山东大学 | Atrial fibrillation signal identification method and system based on machine learning |
| CN111626114A (en) * | 2020-04-20 | 2020-09-04 | 哈尔滨工业大学 | Electrocardiosignal arrhythmia classification system based on convolutional neural network |
| CN111626114B (en) * | 2020-04-20 | 2022-11-18 | 哈尔滨工业大学 | Electrocardiosignal arrhythmia classification system based on convolutional neural network |
| CN111557660A (en) * | 2020-06-08 | 2020-08-21 | 东北大学 | Arrhythmia identification method under sub-population deep learning framework |
| CN111919224A (en) * | 2020-06-30 | 2020-11-10 | 北京小米移动软件有限公司 | Biological feature fusion method and device, electronic equipment and storage medium |
| WO2022052300A1 (en) * | 2020-09-08 | 2022-03-17 | 浙江大学山东工业技术研究院 | Electrocardio parameter calculation method based on deep learning |
| CN112932431A (en) * | 2021-01-26 | 2021-06-11 | 山西三友和智慧信息技术股份有限公司 | Heart rate identification method based on 1DCNN + Inception Net + GRU fusion network |
| CN113133767A (en) * | 2021-03-22 | 2021-07-20 | 浙江工业大学 | Electrocardiosignal classification method combining wave band extraction and multi-feature fusion |
| CN115211866A (en) * | 2022-09-07 | 2022-10-21 | 西南民族大学 | Arrhythmia classification method and system and electronic equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108256452A (en) | A kind of method of the ECG signal classification of feature based fusion | |
| CN109886273A (en) | A kind of CMR classification of image segmentation system | |
| CN108256453A (en) | A kind of method based on one-dimensional ECG signal extraction two dimension CNN features | |
| CN107909117A (en) | A kind of sorting technique and device based on brain function network characterization to early late period mild cognitive impairment | |
| KR101812406B1 (en) | The method and system for diagnosing skin disease | |
| CN109117864A (en) | Coronary heart disease risk prediction technique, model and system based on heterogeneous characteristic fusion | |
| CN111000553B (en) | Intelligent classification method for electrocardiogram data based on voting ensemble learning | |
| Exarchos et al. | EEG transient event detection and classification using association rules | |
| CN108932452A (en) | Arrhythmia cordis beat classification method based on multiple dimensioned convolutional neural networks | |
| CN106682616A (en) | Newborn-painful-expression recognition method based on dual-channel-characteristic deep learning | |
| CN109934089A (en) | Multistage epileptic EEG Signal automatic identifying method based on supervision gradient lifter | |
| CN106778685A (en) | Electrocardiogram image-recognizing method, device and service terminal | |
| Songdechakraiwut et al. | Topological learning and its application to multimodal brain network integration | |
| CN108511055A (en) | Ventricular premature beat identifying system and method based on Multiple Classifier Fusion and diagnostic rule | |
| Alqudah et al. | Employing image processing techniques and artificial intelligence for automated eye diagnosis using digital eye fundus images | |
| CN108090509A (en) | A kind of adaptive electrocardiogram sorting technique of data length | |
| CN112598661A (en) | Ankle fracture and ligament injury diagnosis method based on machine learning | |
| CN107301409A (en) | Learn the system and method for processing electrocardiogram based on Wrapper feature selectings Bagging | |
| Zhao et al. | Constructing multi-view high-order functional connectivity networks for diagnosis of autism spectrum disorder | |
| Sanjana et al. | Automated diabetic retinopathy detection using transfer learning models | |
| CN109044339A (en) | A kind of ECG signal diagnostic method of limit convolutional neural networks | |
| CN106778830A (en) | Staging method based on double cluster results and AdaBoost | |
| Hassan et al. | A hybrid data mining approach for knowledge extraction and classification in medical databases | |
| CN112561935B (en) | Intelligent classification method, device and equipment for brain images | |
| CN112686091B (en) | Two-step arrhythmia classification method based on deep 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: 20180706 |