CN117198463A - Method for identifying needle electrode electromyography motor neuron damage by crossing data fields - Google Patents
Method for identifying needle electrode electromyography motor neuron damage by crossing data fields Download PDFInfo
- Publication number
- CN117198463A CN117198463A CN202311464049.8A CN202311464049A CN117198463A CN 117198463 A CN117198463 A CN 117198463A CN 202311464049 A CN202311464049 A CN 202311464049A CN 117198463 A CN117198463 A CN 117198463A
- Authority
- CN
- China
- Prior art keywords
- network
- data
- training
- motor neuron
- neuron damage
- 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
- 238000000034 method Methods 0.000 title claims abstract description 51
- 210000002161 motor neuron Anatomy 0.000 title claims abstract description 45
- 230000004693 neuron damage Effects 0.000 title claims abstract description 37
- 238000002567 electromyography Methods 0.000 title abstract description 53
- 238000012549 training Methods 0.000 claims abstract description 94
- 238000012360 testing method Methods 0.000 claims abstract description 51
- 210000003205 muscle Anatomy 0.000 claims abstract description 27
- 238000000605 extraction Methods 0.000 claims abstract description 22
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 14
- 230000006870 function Effects 0.000 claims description 29
- 238000011478 gradient descent method Methods 0.000 claims description 6
- 238000012886 linear function Methods 0.000 claims description 6
- 238000011176 pooling Methods 0.000 claims description 6
- 201000010099 disease Diseases 0.000 claims description 4
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 claims description 4
- 238000006073 displacement reaction Methods 0.000 claims description 3
- 230000008447 perception Effects 0.000 claims description 3
- 230000003902 lesion Effects 0.000 claims 2
- 238000013135 deep learning Methods 0.000 abstract description 5
- 238000007781 pre-processing Methods 0.000 abstract 1
- 206010002026 amyotrophic lateral sclerosis Diseases 0.000 description 21
- 238000002474 experimental method Methods 0.000 description 10
- 230000006378 damage Effects 0.000 description 6
- 208000027418 Wounds and injury Diseases 0.000 description 5
- 238000004458 analytical method Methods 0.000 description 5
- 238000004195 computer-aided diagnosis Methods 0.000 description 5
- 238000001514 detection method Methods 0.000 description 5
- 208000014674 injury Diseases 0.000 description 5
- 208000005264 motor neuron disease Diseases 0.000 description 5
- 208000018360 neuromuscular disease Diseases 0.000 description 4
- 230000035945 sensitivity Effects 0.000 description 4
- 206010041591 Spinal osteoarthritis Diseases 0.000 description 3
- 208000036319 cervical spondylosis Diseases 0.000 description 3
- 238000005070 sampling Methods 0.000 description 3
- 208000005801 spondylosis Diseases 0.000 description 3
- 206010028570 Myelopathy Diseases 0.000 description 2
- 238000002679 ablation Methods 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000003745 diagnosis Methods 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 208000034869 Cervical myelopathy Diseases 0.000 description 1
- 208000026072 Motor neurone disease Diseases 0.000 description 1
- 208000029033 Spinal Cord disease Diseases 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000003759 clinical diagnosis Methods 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000003748 differential diagnosis Methods 0.000 description 1
- 238000013399 early diagnosis Methods 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000004445 quantitative analysis Methods 0.000 description 1
- 238000000638 solvent extraction Methods 0.000 description 1
Landscapes
- Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
Abstract
The invention relates to a method for identifying needle electrode electromyogram motor neuron damage by crossing data fields, which comprises the following steps: 1. dividing a training set and a testing set according to individuals for electromyography data of different muscle parts of the acquired motor neuron damage and the acquired non-motor neuron damage; 2. preprocessing such as sliding window division, baseline drift removal, standardization and the like is carried out on each individual data; 3. constructing a cross-data-domain pre-training network comprising a feature extraction network and a projection head, randomly selecting a plurality of preprocessed sample data for each individual data, and inputting the preprocessed sample data into the network for learning; 4. in the identification stage, a single-branch wavelet convolutional neural network is adopted as a classifier, the probability of suffering from motor neuron damage is output, and a final identification result is obtained. The invention realizes comprehensive deep learning and common feature extraction of electromyography data features of different individuals, different parts, different diseased states and different groups of motor neuron damage and non-motor neuron damage.
Description
Technical Field
The invention relates to the technical field of computer-aided medical treatment, in particular to a method for identifying needle electrode electromyogram motor neuron damage across data fields.
Background
Electromyography (EMG) is a common detection tool for clinical diagnosis of neuromuscular diseases, and provides clinical reference for diagnosis and treatment schemes of motor neuron diseases through acquisition and analysis of EMG data of different muscle parts of a subject. Amyotrophic lateral sclerosis (Amyotrophic Lateral Sclerosis, ALS), also known as motor neuron disease, has similar clinical manifestations and electromyographic features as cervical myelopathy (Cervical Spondylotic Myelopathy, CSM), and presents difficulties in clinical differential diagnosis and treatment. The sensitivity of different muscle sites to motor neuron damage is different, and EMG data of different diseased states and different muscle sites can be considered to come from different data fields. The difficulty of EMG data acquisition of different muscle parts or the difference of tolerance of a subject, how to select the muscle parts with high sensitivity to the motor neuron damage and better tolerance of the subject as EMG detection parts, and the common characteristics of the motor neuron damage of different data fields are extracted, which are important problems to be solved by a computer-aided diagnosis method. The invention mainly relates to a deep learning method for identifying needle electrode electromyogram motor neuron damage by crossing data fields.
Compared with the traditional manual detection method, the computer-aided diagnosis method for the neuromuscular diseases based on the EMG data can greatly improve the efficiency and accuracy of quantitative analysis of the EMG data, provide auxiliary reference for doctors, and achieve the purposes of early discovery, early diagnosis and early treatment of motor neuron diseases. At present, in the computer-aided diagnosis of motor neuron diseases based on EMG data, the data set division is mainly to divide sample data of different individuals into a training set and a test set, and the division mode according to the individuals is rarely considered. Meanwhile, in the existing EMG analysis, the sensitivity of different muscle parts in diagnosis of neuromuscular diseases is rarely analyzed, and the EMG data crossing the data domain is not analyzed.
Therefore, the method has important significance for improving the computer-aided diagnosis performance of EMG data motor neuron diseases, extracting EMG common characteristics and searching EMG acquisition sensitive muscle parts based on EMG data learning of crossing individuals and crossing data domains.
Disclosure of Invention
The invention aims at a cross-data-domain learning method for extracting common features, constructs a cross-data-domain learning network, realizes the extraction of the common features of cross-data-domain EMG data, and improves the classification performance of EMG data neuromuscular disease computer-aided diagnosis.
The invention is realized by the following technical scheme, which comprises the following steps:
first, EMG data of different muscle parts of the acquired motor neuron damage and non-motor neuron damage are obtained, and EMG data of each person are individualEMG data from different muscle parts in different disease states are called EMG individual data from different data fields, and EMG individual data from two different data fields respectively form a training setAnd test set；
Second step, EMG data for the ith individualThe sliding window division with the window length of L and the overlapping rate of 50% is adopted to obtainThe EMG sample data are subjected to baseline drift removal by a polynomial fitting method to obtain preprocessed sample data；
Third, in the pre-training stage, for training setAnd test setEMG data of each individual in (3)Randomly selecting a plurality of preprocessed sample dataThe information is input into a pre-training network for learning, and in the process, the training set needs the label information, and the test set does not need the label information. The method comprises the following steps:
(1) Each of the training set and the test setIs a plurality of sample data of (1)As input data of the three branch networks of the pre-training network, normal samples in the training set, diseased samples in the training set and test set samples are included; wherein, the diseased sample refers to a sample with motor neuron damage, and the normal sample refers to a sample without motor neuron damage, which can be a normal human sample or a cervical spondylosis sample without motor neuron damage;
(2) The three branch networks of the pre-training network comprise two online networks and an item mark network, and the three networks comprise a feature extraction network and a projection head; the feature extraction network uses a multi-branch wavelet convolutional neural network and comprises 3 wavelet convolutional layers, 2 pooling layers and 2 batch standardization layers, the kernel function of each wavelet convolutional layer is constrained by a selected wavelet basis function, and parameters such as scale factors, displacement factors, amplitude weight coefficients and the like of the wavelet convolutional layer can be learned and updated in training through a gradient descent method; the projection head is a multi-layer perception network and is formed by connecting a linear function, a batch standardization layer, a linear rectification function and a linear function;
(3) The input of the two online networks is a normal sample and a diseased sample in the training set, and the input of the target network is a sample in the test set. Calculating the inter-class loss between two online networks, calculating the loss between the online network and the target network across the data domains, and weighting the two losses to form the total loss of the network, specifically:
intra-class loss for cross data domainsThe cosine similarity is used to measure the distance between the data fields, and the distance between the two data fields is shortened by using the cosine similarity, and the function is as follows:
wherein,representing samples from the training set and the test set, respectively, S and T representing the total number of samples in the training set and the test set,anda feature extraction network in the network and the target network respectively,andprojection heads in a network and a target network respectively, wherein F is a cosine similarity function, and E is a mean function;
for inter-class loss, a leaky intra-class loss function is employedThe function is as follows:
wherein,representing two samples from the training set, diseased and normal, respectively, C being a constant, LR being a leaky linear rectification function;
the total loss of the network isWhereinThe weighting factor is a value from 0 to 1.
(4) And updating parameters of the online network by adopting a gradient descent method, weighting the target network parameters by the current network parameters and the parameters of the online network, and finally obtaining the trained pre-training network.
And fourthly, in the identification stage, a single-branch wavelet convolutional neural network is adopted as a classifier, a high-order embedded feature is extracted based on a feature extraction network in a pre-training network, and then the high-order embedded feature is input into the classifier for training and testing, so that the probability of motor neuron damage is output, and the identification and classification of electromyographic signals are realized. The method comprises the following steps:
(1) Respectively extracting high-order embedded features from the training set data and the test set data based on a feature extraction network in the pre-training network;
(2) The method comprises the steps of adopting a single-branch wavelet convolutional neural network as a classifier, inputting high-order embedded features of a training set into the classifier for training, and testing a test set based on the trained classifier, wherein the wavelet convolutional neural network comprises 1 wavelet convolutional layer, 2 1-dimensional common convolutional layers, 2 pooling layers and 2 batch standardization layers;
(3) Outputting the probability of motor neuron damage, and realizing the identification and classification of the electromyographic signals.
Compared with the prior art, the invention has the beneficial effects that:
(1) The invention provides a deep learning framework for extracting common characteristics, which can extract common characteristics of motor neuron injury based on EMG data, and screen out muscle parts with high sensitivity to motor neuron injury and better tolerance of a subject as EMG detection muscle parts.
(2) The invention can effectively improve the identification performance of ALS and CSM electromyography and the identification capability of the damage of the motor neurons of the needle electrode electromyography.
(3) The pre-training neural network can reduce the difference between two different data domains on one hand, and can pre-classify the data according to the labels of the training set on the other hand, so as to realize the extraction of the common characteristics of the different data domains.
(4) The designed classifier is based on a single-branch wavelet convolutional neural network, less training parameters are required, more hidden features can be extracted, and the classifying efficiency and accuracy are improved.
Drawings
FIG. 1 is a flowchart of a pre-training phase algorithm of the present invention;
FIG. 2 is a flowchart of an identification phase algorithm according to the present invention;
fig. 3 is a schematic diagram of a feature extraction network according to the present invention.
Detailed Description
The following describes embodiments of the present invention in detail with reference to the accompanying drawings, and the embodiments and processes of the present invention are given by implementing the embodiments of the present invention on the premise of the technical solution of the present invention, but the scope of protection of the present invention is not limited to the following embodiments.
The invention takes Electromyography (EMG) data as an application object, and learns and classifies EMG data of amyotrophic lateral sclerosis (Amyotrophic Lateral Sclerosis, ALS) and EMG data of normal EMG data and cervical spondylosis spinal cord disease (Cervical Spondylotic Myelopathy, CSM) respectively. As shown in table 1, the data used in the experiment were derived from the public dataset EMGLAB and the third hospital of beijing university, and contained two datasets in total:
data set 1 is an ALS, normal and CSM data set acquired by a Keypoint system and provided by a third hospital of Beijing university, and the sampling rate is 24kHz. Are collected from tibialis Anterior (ATM), right first interosseous (rFIM) and Sternocleidomastoid (STER), respectively.
Data set 2 (BBM data set) was downloaded from the published data set of EMGLABs, which is an ALS and normal data set sampled from the Biceps Brachii (BBM) using the Keypoint system, at a sampling rate of 23437Hz. BBM datasets are limited in number of people, recorded as individuals.
In the following experiments, data set 1 was used to evaluate the performance of the proposed method in ALS and CSM identification, and data set 1 and data set 2 were used to demonstrate the effectiveness of the proposed method in cross-site and cross-population experiments.
TABLE 1 composition of datasets
As shown in fig. 1, the embodiment flow of the inter-individual and inter-data-domain learning method based on contrast learning of the invention is as follows:
first, EMG data of different muscle parts of the acquired motor neuron damage and non-motor neuron damage are obtained, and EMG data of each person are individualEMG data from different muscle parts in different disease states are called EMG individual data from different data fields, and EMG individual data from two different data fields respectively form a training setAnd test set。
During the pre-training phase, 1/5 of all samples of the training set and test set samples are used. In the classifier training stage, all samples and labels of the training set are used for training the classifier; in the test phase, all samples of the test set are used for testing. To avoid possible bias in the sample partitioning process, five-fold cross-validation was used to reduce the impact of different individuals, demonstrating the performance of the proposed method.
Second step, EMG data for the ith individualThe sliding window division with the window length of L and the overlapping rate of 50% is adopted to obtainThe EMG sample data are subjected to baseline drift removal by a polynomial fitting method to obtain preprocessed sample data. The method comprises the following steps:
first, for the original EMG dataSliding window division is performed by using window length 6000, and 50% overlap ratio exists between the windows, and EMG data in each analysis window is regarded as one sampleThe method comprises the steps of carrying out a first treatment on the surface of the In this example, taking the BBM dataset as an example, 5028 ALS samples and 5344 normal control samples can be obtained;
then, for each EMG sample data, a polynomial fitting method is used to remove baseline drift. In this embodiment, an original baseline is obtained by fitting a signal polynomial, and the original baseline is subtracted from the original EMG signal to obtain baseline-shifted EMG data.
Third, in the pre-training stage, for training setAnd test setEMG data of each individual in (3)Randomly selecting a plurality of preprocessed sample dataThe information is input into a pre-training network for learning, and in the process, the training set needs label information and the test set does not need labels. The specific operation is as follows:
(1) Each of the training set and the test setIs of 12 samplesAs input data of the three branch networks of the pre-training network, normal samples in the training set, diseased samples in the training set and test set samples are included; wherein, the diseased sample refers to a sample with motor neuron damage, and the normal sample refers to a sample without motor neuron damage, which can be a normal human sample or a cervical spondylosis sample without motor neuron damage;
(2) The three branch networks of the pre-training network comprise two online networks and an item mark network, and the three networks comprise a feature extraction network and a projection head; the data is embedded after passing through the feature extraction network, and then projection is obtained after the embedding. The characteristic extraction network is shown in figure 3, and the characteristic extraction network uses a multi-branch wavelet convolutional neural network, and comprises 3 wavelet convolutional layers, 2 pooling layers and 2 batch standardization layers, wherein the kernel function of each wavelet convolutional layer is constrained by the selected wavelet basis function, and parameters such as scale factors, displacement factors, amplitude weight coefficients and the like can be learned and updated in training by a gradient descent method; the projection head is a simple multi-layer perception network and is formed by connecting a linear function, a batch standardization layer, a linear rectification function and a linear function.
(3) The input of the two online networks is a normal sample and a diseased sample in a training set respectively, and the input of the target network is a sample in a test set; calculating the inter-class loss between two online networks, calculating the loss between the online network and the target network across the data domains, and weighting the two losses to form the total loss of the network, specifically:
intra-class loss for cross data domainsThe cosine similarity is used to measure the distance between the data fields, and the distance between the two data fields is shortened by using the cosine similarity, and the function is as follows:
wherein,representing samples from the training set and the test set, respectively, S and T representing the total number of samples in the training set and the test set,anda feature extraction network in the network and the target network respectively,andprojection heads in a network and a target network respectively, wherein F is a cosine similarity function, and E is a mean function;
for inter-class loss, a leaky intra-class loss function is employedThe function is as follows:
wherein,two samples from the training set are shown diseased and normal, C is a constant, LR is a leaky linear rectification function, respectively.
The total loss of the network isWhereinThe weighting factor is a value from 0 to 1.
(4) And updating parameters of the online network by adopting a gradient descent method, weighting the target network parameters by the current network parameters and the parameters of the online network, and finally obtaining the trained pre-training network.
In the recognition stage, as shown in fig. 2, a single-branch wavelet convolutional neural network is adopted as a classifier, a high-order embedded feature is extracted based on a feature extraction network in a pre-training network, and then the high-order embedded feature is input into the classifier for training and testing, so that the probability of motor neuron damage is output, and the recognition and classification of electromyographic signals are realized. The method comprises the following steps:
(1) Respectively extracting high-order embedded features from the training set data and the test set data based on a feature extraction network in the pre-training network;
(2) The method comprises the steps of adopting a single-branch wavelet convolutional neural network as a classifier, inputting high-order embedded features of a training set into the classifier for training, and testing a testing set based on the trained classifier; the wavelet convolution neural network comprises 1 wavelet convolution layer, 2 1-dimensional common convolution layers, 2 pooling layers and 2 batch standardization layers;
(3) Outputting the probability of motor neuron damage, and realizing the identification and classification of the electromyographic signals.
In this example, all of the following experiments were performed on a Ubuntu system, with GPU NVIDIA Geforce 1080Ti, CUDA10.2. The relevant configuration during pre-training is as follows: 30 epochs, batch size 128, learning rate 0.0003; the relevant configuration during the classifier training phase is as follows: 60 epochs, batch size 16, learning rate 0.0003.
For both dataset 1 and dataset 2, the Keypoint system was used to sample from different muscle sites and different groups at approximately the same sampling rate (i.e., dataset 1 was an asian population sample and dataset 2 was a european population sample). In the experiments of ALS and CSM identification, considering the effectiveness of the sternocleidomastoid muscle on ALS and CSM identification in clinic, rFIM and ATM are respectively adopted as training sets, and STER is adopted as a test set to identify two diseases. In cross-population experiments, rFIM and BBM were used alternately as training and testing sets.
In tables 2 and 3, the recognition results of ALS and CSM on the sternocleidomastoid muscle are provided, comparing the recognition performance of the proposed method with other existing methods. The method 1 and the method 2 use a traditional machine learning method. Wherein, the method 1 uses linear discriminant analysis (Linear Discriminant Analysis, LDA) as classifier after extracting features; method 2 uses an integrated classifier for identification classification. Method 3 uses an extended one-dimensional convolutional neural network to identify ALS from the original EMG signal. In table 2, the training set is the rFIM dataset of dataset 1 and the test set is the stem dataset of dataset 2. In table 3, the training set is the ATM dataset of dataset 1 and the test set is the STER dataset of dataset 2. As can be seen from tables 2 and 3, the accuracy of the two conventional machine learning methods is significantly lower than that of the two deep learning methods. In contrast, the deep learning method of method 3 and the inventive method more easily learns the commonality of motor neuron damage. Compared with the method 3, the method has more stable identification performance and better robustness when the test is performed on the sternocleidomastoid muscle under the condition of training different muscle parts.
TABLE 2 ALS and CSM recognition results with rFIM as training set and STER as test set
TABLE 3 ALS and CSM recognition results with ATM as training set and STER as test set
Tables 4 and 5 provide ALS and normal recognition experiments across muscle sites and across groups, using rFIM and BBM as training sets, BBM, rFIM, ATM as test sets, respectively. From tables 4 and 5, the learning model provided by the invention can extract the common characteristics of motor neuron injuries of different muscle parts and different groups, and can accurately identify normal EMG data of motor neuron injury ALS and non-motor neuron injury. The method has higher classification performance in single muscle part and cross-part experiments, and has good robustness to various different muscle part experiments. Under the condition that certain muscle parts are inconvenient to collect, the method can effectively utilize the EMG data of the existing muscle parts, and can obtain good identification performance.
TABLE 4 ALS and Normal recognition results Using rFIM as training set
TABLE 5 ALS and Normal recognition results Using BBM as training set
The function of each algorithm module in the invention is shown in the ablation experiment of table 6, in which BBM is used as a training data set and rFIM is used as a test data set. The results in Table 6 demonstrate that pre-training is critical to the performance of the proposed algorithm, with the pre-training module removed, the algorithm performance is reduced by 12.62%. Leakage intra-class loss functions and new classifiers also represent important contributions to the model performance improvement.
Table 6 ablation experimental results
In the comprehensive view, in the EMG data analysis of cross-individual and cross-data fields, the method can more effectively extract the common characteristics of the motor neuron damage, and improve the accuracy and the robustness of the motor neuron damage detection of cross-individual, cross-part, cross-group and other cross-data fields.
The foregoing description has fully described the embodiments of the invention. It should be noted that any modifications to the specific embodiments of the invention may be made by those skilled in the art without departing from the scope of the invention as defined in the appended claims. Accordingly, the scope of the claims of the present invention is not limited to the foregoing detailed description.
Claims (3)
1. A method for identifying needle electrode electromyographic motor neuron damage across data fields, the method comprising the steps of:
(1) Electromyographic data of different muscle parts of acquired motor neuron damage and non-motor neuron damage, the EMG data of the ith person is an individualEMG data of different muscle parts with different disease states are called from different data fields, and EMG individual data from two different data fields respectively form a training set +.>And test set->；
(2) For each ofEMG data of individualsAdopting sliding window division with window length L and overlapping rate of 50% to obtain +.>The EMG sample data are subjected to baseline drift removal by a polynomial fitting method to obtain pretreatment sample data>；
(3) During the pre-training stage, for training setAnd test set->EMG data of each individual +.>Randomly selecting a number of preprocessed sample data +.>Inputting the information into a pre-training network for learning, wherein in the process, the training set needs label information, and the testing set does not need label information;
(4) In the identification stage, a single-branch wavelet convolutional neural network is adopted as a classifier, a high-order embedded feature is extracted based on a feature extraction network in a pre-training network, then the high-order embedded feature is input into the classifier for training and testing, the probability of motor neuron damage is output, and the identification and classification of electromyographic signals are realized.
2. A method of identifying needle electrode electromyographic motor neuron lesions across data fields according to claim 1, wherein said pre-training process of step (3) is specifically as follows:
(1) Each of the training set and the test setIs>As input data of the three branch networks of the pre-training network, normal samples in the training set, diseased samples in the training set and test set samples are included; wherein the diseased sample refers to a sample with motor neuron damage and the normal sample refers to a sample without motor neuron damage;
(2) The three branch networks of the pre-training network comprise two online networks and an item mark network, and the three networks comprise a feature extraction network and a projection head; the feature extraction network uses a multi-branch wavelet convolutional neural network and comprises 3 wavelet convolutional layers, 2 pooling layers and 2 batch standardization layers, the kernel function of each wavelet convolutional layer is constrained by a selected wavelet basis function, and the scale factors, the displacement factors and the amplitude weight coefficients of the wavelet convolutional layer can be learned and updated in training by a gradient descent method; the projection head is a multi-layer perception network and is formed by connecting a linear function, a batch standardization layer, a linear rectification function and a linear function;
(3) The input of the two online networks is a normal sample and a diseased sample in a training set respectively, and the input of the target network is a sample in a test set; calculating the inter-class loss between two online networks, calculating the loss between the online network and the target network across the data domains, and weighting the two losses to form the total loss of the network, specifically:
intra-class loss for cross data domainsThe cosine similarity is used to measure the distance between the data fields, and the distance between the two data fields is shortened by using the cosine similarity, and the function is as follows:
，
wherein,representing samples from the training set and the test set, respectively, S and T representing the total number of samples in the training set and the test set, respectively, +.>And->Feature extraction network in the network and the target network, respectively,>and->Projection heads in a network and a target network respectively, wherein F is a cosine similarity function, and E is a mean function;
for inter-class loss, a leaky intra-class loss function is employedThe function is as follows:
，
wherein,representing two samples from the training set, diseased and normal, respectively, C being a constant, LR being a leaky linear rectification function;
the total loss of the network isWherein->Taking the value from 0 to 1 as a weighting factor;
(4) And updating parameters of the online network by adopting a gradient descent method, weighting the target network parameters by the current network parameters and the parameters of the online network, and finally obtaining the trained pre-training network.
3. A method of identifying needle electrode electromyographic motor neuron lesions across data fields according to claim 1, wherein said classification of step (4) is specifically as follows:
(1) Respectively extracting high-order embedded features from the training set data and the test set data based on a feature extraction network in the pre-training network;
(2) The method comprises the steps of adopting a single-branch wavelet convolutional neural network as a classifier, inputting high-order embedded features of a training set into the classifier for training, and testing a test set based on the trained classifier, wherein the wavelet convolutional neural network comprises 1 wavelet convolutional layer, 2 1-dimensional common convolutional layers, 2 pooling layers and 2 batch standardization layers;
(3) Outputting the probability of motor neuron damage, and realizing the identification and classification of the electromyographic signals.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311464049.8A CN117198463A (en) | 2023-11-06 | 2023-11-06 | Method for identifying needle electrode electromyography motor neuron damage by crossing data fields |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311464049.8A CN117198463A (en) | 2023-11-06 | 2023-11-06 | Method for identifying needle electrode electromyography motor neuron damage by crossing data fields |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117198463A true CN117198463A (en) | 2023-12-08 |
Family
ID=89003808
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311464049.8A Pending CN117198463A (en) | 2023-11-06 | 2023-11-06 | Method for identifying needle electrode electromyography motor neuron damage by crossing data fields |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117198463A (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180247224A1 (en) * | 2017-02-28 | 2018-08-30 | Nec Europe Ltd. | System and method for multi-modal graph-based personalization |
| CN114611563A (en) * | 2022-05-16 | 2022-06-10 | 北京大学第三医院（北京大学第三临床医学院） | Method for identifying neurogenic damage of needle pole electromyogram in cross-site manner |
| CN114936583A (en) * | 2022-06-14 | 2022-08-23 | 中国科学技术大学 | Teacher-student model-based two-step field self-adaptive cross-user electromyogram pattern recognition method |
| CN116029333A (en) * | 2023-01-04 | 2023-04-28 | 飞矩科技(上海)有限公司 | Distributed system self-adaption-oriented multi-target myoelectricity mode identification method |
-
2023
- 2023-11-06 CN CN202311464049.8A patent/CN117198463A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180247224A1 (en) * | 2017-02-28 | 2018-08-30 | Nec Europe Ltd. | System and method for multi-modal graph-based personalization |
| CN114611563A (en) * | 2022-05-16 | 2022-06-10 | 北京大学第三医院（北京大学第三临床医学院） | Method for identifying neurogenic damage of needle pole electromyogram in cross-site manner |
| CN114936583A (en) * | 2022-06-14 | 2022-08-23 | 中国科学技术大学 | Teacher-student model-based two-step field self-adaptive cross-user electromyogram pattern recognition method |
| CN116029333A (en) * | 2023-01-04 | 2023-04-28 | 飞矩科技(上海)有限公司 | Distributed system self-adaption-oriented multi-target myoelectricity mode identification method |
Non-Patent Citations (2)
| Title |
|---|
| HAOYU ZHANG .ET: "Domain Contrast Network for cross-muscle ALS disease identification with EMG signal", 《ELSEVIER》, pages 1 - 8 * |
| 唐爱明 等: "《新编内科疾病诊疗学》", 30 September 2009, pages: 449 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Sharma et al. | Dermatologist-level classification of skin cancer using cascaded ensembling of convolutional neural network and handcrafted features based deep neural network | |
| Rahman et al. | Sleep stage classification using single-channel EOG | |
| Zhao et al. | Noise rejection for wearable ECGs using modified frequency slice wavelet transform and convolutional neural networks | |
| Abou Jaoude et al. | Detection of mesial temporal lobe epileptiform discharges on intracranial electrodes using deep learning | |
| CN111046731B (en) | Transfer learning method and recognition method for gesture recognition based on surface electromyographic signals | |
| Li et al. | Robust ECG biometrics using GNMF and sparse representation | |
| CN109497990B (en) | Electrocardiosignal identity recognition method and system based on canonical correlation analysis | |
| CN108256629A (en) | The unsupervised feature learning method of EEG signal based on convolutional network and own coding | |
| Du et al. | A method for detecting high-frequency oscillations using semi-supervised k-means and mean shift clustering | |
| Yuan et al. | Wave2vec: Learning deep representations for biosignals | |
| Yildiz et al. | Classification and analysis of epileptic EEG recordings using convolutional neural network and class activation mapping | |
| Huan et al. | Deep convolutional neural networks for classifying body constitution based on face image | |
| Li et al. | Application of MODWT and log-normal distribution model for automatic epilepsy identification | |
| Prasanth et al. | Deep learning for interictal epileptiform spike detection from scalp EEG frequency sub bands | |
| Singh et al. | Detection of epileptic seizure EEG signal using multiscale entropies and complete ensemble empirical mode decomposition | |
| Soni et al. | Graphical representation learning-based approach for automatic classification of electroencephalogram signals in depression | |
| Li et al. | EEG signal classification method based on feature priority analysis and CNN | |
| Wei et al. | A deep learning framework with multi-perspective fusion for interictal epileptiform discharges detection in scalp electroencephalogram | |
| Li et al. | Early diagnosis of Parkinson's disease using Continuous Convolution Network: Handwriting recognition based on off-line hand drawing without template | |
| CN112651927A (en) | Raman spectrum intelligent identification method based on convolutional neural network and support vector machine | |
| Qazi et al. | Electroencephalogram (EEG) Brain Signals to Detect Alcoholism Based on Deep Learning. | |
| CN113988135A (en) | Electromyographic signal gesture recognition method based on double-branch multi-stream network | |
| CN117198463A (en) | Method for identifying needle electrode electromyography motor neuron damage by crossing data fields | |
| Li et al. | Tfformer: A time frequency information fusion based cnn-transformer model for osa detection with single-lead ecg | |
| Hao et al. | Convolutional neural network based on recurrence plot for EEG recognition |
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 |