CN116186544A - Single-channel electroencephalogram sleep stage-dividing method based on deep learning - Google Patents

Single-channel electroencephalogram sleep stage-dividing method based on deep learning Download PDF

Info

Publication number
CN116186544A
CN116186544A CN202310285377.5A CN202310285377A CN116186544A CN 116186544 A CN116186544 A CN 116186544A CN 202310285377 A CN202310285377 A CN 202310285377A CN 116186544 A CN116186544 A CN 116186544A
Authority
CN
China
Prior art keywords
sequence
capsule
sleep
layer
duration
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
Application number
CN202310285377.5A
Other languages
Chinese (zh)
Inventor
彭虎
陈瑾
韩志会
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hefei University of Technology
Original Assignee
Hefei University of Technology
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Hefei University of Technology filed Critical Hefei University of Technology
Priority to CN202310285377.5A priority Critical patent/CN116186544A/en
Publication of CN116186544A publication Critical patent/CN116186544A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • 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
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • 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
    • A61B5/369—Electroencephalography [EEG]
    • A61B5/372—Analysis of electroencephalograms
    • 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
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D30/00—Reducing energy consumption in communication networks
    • Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks

Landscapes

  • Health & Medical Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Psychiatry (AREA)
  • Public Health (AREA)
  • Surgery (AREA)
  • Medical Informatics (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Pathology (AREA)
  • Mathematical Physics (AREA)
  • Veterinary Medicine (AREA)
  • Evolutionary Computation (AREA)
  • Animal Behavior & Ethology (AREA)
  • Theoretical Computer Science (AREA)
  • Fuzzy Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physiology (AREA)
  • Signal Processing (AREA)
  • Data Mining & Analysis (AREA)
  • Psychology (AREA)
  • Software Systems (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)

Abstract

The invention discloses a single-channel electroencephalogram sleep stage method based on deep learning, which comprises the following steps: 1. constructing a training set from the polysomnography and the sleep label; 2. building an electroencephalogram sleep stage network based on deep learning; 3. constructing an MFE loss function; 4. training a deep learning model; 5. and carrying out sleep staging on the original single-channel electroencephalogram signal by using the trained single-channel sleep staging network, so that the automatic sleep staging of the electroencephalogram signal can be realized.

Description

Single-channel electroencephalogram sleep stage-dividing method based on deep learning
Technical Field
The invention belongs to the field of electroencephalogram sleep stage, and particularly relates to a single-channel electroencephalogram sleep stage method based on deep learning.
Background
Sleep occupies one-third of the life time of a person, and during the sleep, electroencephalogram changes variously, and the changes are different along with the depth of sleep, and the sleep is divided into two states according to different characteristics of the electroencephalogram: non-eye fast movement sleep (also called normal phase sleep, slow wave sleep, synchronous sleep, quiet sleep, NREM sleep) and eye fast movement sleep, which are distinguished by whether there is eye paroxysmal fast movement and different brain wave characteristics, are necessary for evaluation measurement of sleep quality, some sleep-specific simulation methods perform electric signals recorded in a sleep-staged Polysomnogram (PSG) recorded by sensors attached to the brain epidermis, the PSG includes electroencephalogram (EEG), electroencephalogram (EOG), electromyogram (EMG) and Electrocardiograph (ECG), regarding comfort of a subject during sleep monitoring, single-channel electroencephalogram signals are increasingly used for sleep staging, the PSG is divided into periods of 30s, sleep specialists manually divide each period into different stages according to decision rules formulated by rechchanffen and Kales (R & K) and american sleep medical society (AASM), in clinical practice, artificial sleep staging is very energy and cost intensive and automatic research, and thus, various methods for classifying sleep are available for two kinds of sleep stage of critical study: traditional machine-learning-based and deep learning-based approaches typically involve manually extracting features and sleep stage classification. Electroencephalographic signals typically use artificial feature extraction to extract time-domain or frequency-domain features, and then conventional machine learning algorithms such as Support Vector Machines (SVMs), random Forests (RFs), decision trees, and Hidden Markov Models (HMMs) will be used to train sleep stage classification models based on hand-made features, which, while achieving acceptable performance, require prior knowledge of human engineering and expertise in the relevant arts, and require human feature extraction, which takes a long time. In addition, the deep learning method is also applied to single-channel electroencephalogram signal sleep stage. These deep learning-based methods can be divided into two categories depending on the composition of the network: convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). CNN is most often used to extract waveform features related to sleep, some researchers have successfully applied CNN to sleep stage tasks, e.g., tsinalis et al use one-dimensional convolution to extract features and use the largest aggregate layer to remove redundant information, sors et al propose a 14-layer deep CNN for sleep stage, perslev et al propose a CNN of encoder-decoder structure for sleep stage, fiorillo et al propose a multi-scale CNN for extracting features of different scales for sleep stage, there is some time correlation between brain electrical signals, CNN cannot learn the time dependence of these signals, but RNN can learn, e.g., michaeli et al propose a cascaded RNN structure for sleep stage, some methods combine CNN and RNN, extract features with RNN to learn time dependence. For example, supra et al propose a CNN and RNN based model using CNN to obtain a 30 second duration signal characterization and long-short-time memory (LSTM) to learn transition rules between duration phases. In addition, there are methods that combine CNN and attention mechanisms, feature extraction with CNN, and learn time correlation between duration phases with attention mechanisms. Eldele et al combine CNN with channel attention and use an attention mechanism-based module to quantify inter-dependencies between features, qu et al study transition rules between durations by using CNN to extract features and attention mechanisms. Current methods use CNNs to extract features, RNNs or attention mechanisms to learn the time dependence of the duration signal. However, none of these methods can sufficiently extract the time dependence of the history signal, RNN is insufficient in learning ability for long time series time dependence, the attention mechanism needs to rely on a longer history sequence for event-dependent learning of the signal on sleep brain electricity, and learning ability for transition relations between durations is weaker than RNN.
Disclosure of Invention
The invention aims to solve the defects in the prior art, and provides a single-channel electroencephalogram sleep stage method based on deep learning, so that the time correlation of time duration signals can be more fully learned, the generalization capability of a single-channel sleep stage model can be improved, and the accuracy of single-channel sleep stage can be improved.
In order to achieve the aim of the invention, the invention adopts the following technical scheme:
the invention discloses a single-channel electroencephalogram sleep stage method based on deep learning, which is characterized by comprising the following steps of:
step 1, acquiring a polysomnography of a subject and a sleep label set corresponding to the polysomnography, selecting an electroencephalogram sleep signal of one channel in the polysomnography, and dividing the electroencephalogram sleep signal under one channel into N non-overlapping duration signal sequences X= { X 1 ，x 2 ，...，x i ，...，x N Building a tag sequence set y= { Y according to sleep tags corresponding to each duration 1 ，y 2 ，...，y i ，...，y N Simultaneously counting the number { N } of the tags of each sleep category in the sleep tags corresponding to all durations f |f=1, 2, …, D }; wherein x is i Represents the i-th duration signal sequence, and x i ＝{e i,1 ，e i,2 ，...，e i，j ，...，e i，l }，y i Representing the ith duration signal sequence x i Corresponding sleep tag sequences, and y i ＝{y i，1 ，y i,2 ，...，y i,j ，...，y i,l }，e i，j ∈R 1×m ，e i，j Representing the jth duration signal, y, in the ith duration signal sequence i,j Tag indicating the jth duration signal in the ith duration signal sequence, N f Tag number representing the f-th sleep category, f e {1,2,..once., D }, y i,j E {1,2,., D }, D is the total category of sleep sessions, j ε {1,2 … l }, l is x i The number of duration signals present in a sample point is represented by mNumber, and m=θ×δ, θ represents the sampling rate, δ represents the sampling time;
step 2, setting up an electroencephalogram sleep stage network based on deep learning, comprising the following steps: the duration feature extraction module, the context encoder and the capsule network are formed;
step 2.1, the duration feature extraction module extracts the j-th duration signal e i，j After processing, the j-th duration signal e is obtained i，j Feature map a of (2) i，j Thereby obtaining the ith duration signal sequence x i Feature map sequence a of (2) i ＝{a i,1 ，a i,2 ，...，a i，j ，...，a i，l }；
Step 2.2, the context encoder is used for the feature map sequence a i Time correlation learning is carried out to obtain a characteristic diagram sequence H i ＝{H i，1 ，H i，2 ，…，H i，j ，...H i，l }；
Step 2.3, constructing a capsule network, which consists of a primary capsule layer and a digital capsule layer;
step 2.3.1, the primary capsule layer vs. the feature map sequence H i Jth sub-feature map H i，j After treatment, a primary capsule is obtained
Figure BDA0004139619540000031
Thereby obtaining the sequence x with the ith duration i Is the jth duration signal e i，j Corresponding activated primary capsule sequence v i，j ＝{v 1 i，j ，v 2 i，j ，...，v k i，j， ...，v n i，j }；
Step 2.3.2 Primary Capsule sequence v to be activated i，j Inputting into the digital capsule layer for processing to obtain j-th duration signals e related to all D sleep categories i，j Corresponding capsule sequences
Figure BDA0004139619540000032
Step 2.3.3, sequence of capsules U i，j Medium f high-grade characteristic capsule
Figure BDA0004139619540000033
Taking l 2 Norms, thereby obtaining the ith duration sequence x i J-th duration signal e i，j Probability of presence of the f-th sleep class +.>
Figure BDA0004139619540000034
And then obtain the ith duration sequence x i J-th duration signal e i，j Predicting probability vectors of the existence of all sleep categories output via the sleep staging network
Figure BDA0004139619540000035
Step 3, constructing a loss function l of the electroencephalogram sleep stage network by using the formula (7) MFE ：
Figure BDA0004139619540000036
In the formula (7), the amino acid sequence of the compound,
Figure BDA0004139619540000037
is the j-th duration signal e i，j Is the true label y of (2) i，j The existence probability of the f-th class in the single-hot coding of (a);
and 4, training the sleep stage network by using a gradient descent algorithm, and minimizing the loss function to optimize network parameters until the loss function converges to obtain a sleep stage model for realizing sleep stage of the single-channel electroencephalogram signal.
The single-channel electroencephalogram sleep stage-dividing method based on deep learning is also characterized in that the duration feature extraction module in the step 2.1 sequentially comprises a one-dimensional step length of S 1 The convolution kernel is K 1 A first convolutional layer, a first BN normalization layer, a first Relu nonlinear active layer, a pooling size k m1 Step length s m1 Is the first maximum pooling of (2)A layer; step length S 2 The convolution kernel is K 2 A second BN normalization layer, a second Relu nonlinear activation layer; step length S 2 The convolution kernel is K 2 A third BN normalization layer, a third Relu nonlinear activation layer; step length S 2 The convolution kernel is K 2 A fourth convolution layer, a fourth BN normalization layer, a fourth Relu nonlinear activation layer, a pooling size of k m2 Step length s m2 Is stacked with the second largest pooling layer.
The context encoder in step 2.2 comprises: a forward LSTM cell and a backward LSTM cell;
the characteristic diagram sequence a i Is the j-th feature map a of (2) i，j Obtaining the forward hidden state at output j time after processing by the forward LSTM unit
Figure BDA0004139619540000041
Jth feature map a i，j Obtaining the backward hidden state of the output j moment after the processing of the backward LSTM unit>
Figure BDA0004139619540000042
Will->
Figure BDA0004139619540000043
And->
Figure BDA0004139619540000044
After splicing, a characteristic diagram sequence a is obtained i A j-th sub-feature map output by the context encoder, < >>
Figure BDA0004139619540000045
Thereby obtaining the characteristic diagram sequence a i Feature map sequence H output through context encoder i ＝{H i，1 ，H i，2 ，…，H i，j ，…H i，l }。
The primary capsule layer in the step 2.3.1 is formed by a one-dimensional convolution kernel with the size of K p X 1, number of convolution kernels N p Is a convolution layer of (2)A length of l d Is composed of a Reshape operation layer;
the characteristic diagram sequence H i Jth sub-feature map H i，j The primary characteristic map P is obtained after the convolution layer treatment of the primary capsule layer i，j After the Reshape operation layer is processed, outputting a j-th sub-feature diagram H i，j Primary capsule sequence s of (2) i，j And s i，j ＝{s 1 i，j ，s 2 i，j ，...，s k i，j ，...，s n i，j }，
Figure BDA0004139619540000046
Representing the i-th duration sequence x i J-th duration signal e i，j Corresponding primary capsule sequence s i，j N is the primary capsule sequence s i，j The number of the capsules, and n=l P /l d ；l d Represents the kth capsule->
Figure BDA0004139619540000047
Length of l P Representing P i，j K e {1,2,3 …, n };
using a squaring activation function on the primary capsule sequence s i，j The kth capsule of (a)
Figure BDA0004139619540000048
After treatment, an activated primary capsule is obtained>
Figure BDA0004139619540000049
Thereby obtaining the sequence x with the ith duration i Is the jth duration signal e i，j Corresponding activated primary capsule sequence v i，j ＝{v 1 i，j ，v 2 i，j ，...，v k i，j ，...，v n i，j }。/>
The digital capsule layer in the step 2.3.2 calculates a primary capsule sequence v by using the formula (1) i，j Mapped f-th subspace prediction capsule orderColumn of
Figure BDA00041396195400000410
The kth predictive capsule of (a)>
Figure BDA00041396195400000411
Thereby calculating the primary capsule sequence v using formula (2) i，j Mapped f-th subspace predictive capsule sequence +.>
Figure BDA00041396195400000412
Figure BDA00041396195400000413
Figure BDA00041396195400000414
In the formula (1) and the formula (2), W f Representing the f-th matrix of the learner,
Figure BDA00041396195400000415
representation matrix W f Is the kth sub-matrix of (c); f.epsilon. {1,2,3 …, D };
the digital capsule layer calculates the f-th subspace predictive capsule sequence by utilizing the method (3)
Figure BDA0004139619540000051
Similarity matrix between each prediction capsule of (a)>
Figure BDA0004139619540000052
Figure BDA0004139619540000053
In the formula (3), the amino acid sequence of the compound,
Figure BDA0004139619540000054
representing predicted capsule sequence->
Figure BDA0004139619540000055
Is a transpose of (2);
the digital capsule layer calculates the f-th subspace predictive capsule sequence by utilizing the method (4)
Figure BDA0004139619540000056
The kth sub-capsule in (a)
Figure BDA0004139619540000057
Weight occupied ∈>
Figure BDA0004139619540000058
Figure BDA0004139619540000059
In the formula (4), the amino acid sequence of the compound,
Figure BDA00041396195400000510
represents->
Figure BDA00041396195400000511
Is->
Figure BDA00041396195400000512
Similarity matrix representing dimensions n x n>
Figure BDA00041396195400000513
The p-th row and the o-th column of the elements;
the digital capsule layer calculates the f-th subspace predictive capsule sequence by utilizing the method (5)
Figure BDA00041396195400000514
The length of the output is l d+1 Advanced feature capsule->
Figure BDA00041396195400000515
Thereby obtaining the jth associated with all D sleep classesDuration signal e i,j Corresponding capsule sequence->
Figure BDA00041396195400000516
Figure BDA00041396195400000517
The electronic equipment comprises a memory and a processor, wherein the memory is used for storing a program for supporting the processor to execute any single-channel electroencephalogram sleep stage method, and the processor is configured for executing the program stored in the memory.
The invention relates to a computer readable storage medium, wherein a computer program is stored on the computer readable storage medium, and is characterized in that the computer program is executed by a processor to execute any step of the single-channel electroencephalogram sleep stage method.
Compared with the prior art, the invention has the beneficial effects that:
1. according to the invention, the scalar network combined by the traditional LSTM and the full-connection layer is converted into the vector network combined by the LSTM and the capsule network by adopting the mode of combining the LSTM by the capsule network, so that the defect of insufficient generalization of the mode of combining the traditional LSTM and the full-connection layer is overcome, and the learning capacity of the LSTM on the internal time correlation of sleep signals can be fully improved.
2. The invention completes the route process from the primary route to the advanced route by adopting the route mechanism based on the self-attention mechanism, overcomes the defect of large training consumption caused by the fact that the traditional dynamic route mechanism cannot be trained in parallel, reduces the complexity of capsule network training, and ensures that the capsule network is successfully applied to the electroencephalogram sleep stage task.
Drawings
FIG. 1 is a block diagram of a single channel sleep stage neural network designed in the present invention;
FIG. 2 is a diagram of a CNN module for feature extraction according to the present invention;
FIG. 3 is a schematic diagram of a long and short term memory model unit designed for use in the present invention;
FIG. 4 is a block diagram of a capsule network according to the present invention;
FIG. 5 is a schematic diagram of the self-attention mechanism of a capsule network contemplated for use with the present invention;
FIG. 6 is a diagram of a reconstructed network for regularization of the design model of the present invention;
fig. 7 is a histogram of ablation experimental evaluation parameters of a single-channel staged sleep model designed in the present invention.
Detailed Description
In this embodiment, a single-channel electroencephalogram sleep stage method based on deep learning includes the following steps:
step 1, acquiring a polysomnography of a subject and a sleep tag set corresponding to the polysomnography, selecting an electroencephalogram sleep signal of one channel in the polysomnography, and dividing the electroencephalogram sleep signal under the one channel into N non-overlapping duration signal sequences X= { X 1 ，x 2 ，...，x i ，...，x N And constructing a tag sequence set Y= { Y according to the sleep tags corresponding to each duration 1 ，y 2 ，...，y i ，...，y N Simultaneously counting the number { N } of the tags of each sleep category in the sleep tags corresponding to all durations f |f=1, 2, …, D }; wherein x is i Represents the i-th duration signal sequence, and x i ＝{e i,1 ，e i,2 ，...，e i，j ，...，e i，l }，y i Representing the ith duration signal sequence x i Corresponding sleep tag sequences, and y i ＝{y i，1 ，y i,2 ，...，y i,j ，...，y i，l }，e i，j ∈R 1×m ，e i，j Representing the jth duration signal, y, in the ith duration signal sequence i,j Tag indicating the jth duration signal in the ith duration signal sequence, N f Tag number representing the f-th sleep category, f e {1,2,..once., D }, y i,j E {1,2,., D }, D is the total category of sleep sessions, j ε {1,2 … l }, l is x i Duration signal present in (a)The number, m, represents the number of sampling points, and m=θ×δ, θ represents the sampling rate, δ represents the sampling time;
step 2, building an electroencephalogram sleep stage network based on deep learning, as shown in fig. 1, including: the duration feature extraction module, the context encoder and the capsule network are formed;
step 2.1, the different sleep phases show significant waveform differences, such as Low Amplitude Mixed Frequency (LAMF) waves occurring in the first phase (N1) and K-complexes occurring mainly in the second phase (N2) of NREM sleep, together with the main axis of sleep, are two significant features of this phase, while sawtooth waves are the only typical electroencephalographic feature of REM sleep. The convolutional neural network is widely applied to the aspect of feature extraction, so the convolutional neural network is designed to serve as a feature extraction module. As shown in fig. 2, the duration feature extraction module sequentially uses a one-dimensional step length of S 1 The convolution kernel is K 1 A first convolutional layer, a first BN normalization layer, a first Relu nonlinear active layer, a pooling size k m1 Step length s m1 Is included in the first maximum pooling layer; step length S 2 The convolution kernel is K 2 A second BN normalization layer, a second Relu nonlinear activation layer; step length S 2 The convolution kernel is K 2 A third BN normalization layer, a third Relu nonlinear activation layer; step length S 2 The convolution kernel is K 2 A fourth convolution layer, a fourth BN normalization layer, a fourth Relu nonlinear activation layer, a pooling size of k m2 Step length s m2 Is stacked, the stacked convolution layers can merge the fields of view and can also extract deeper features. And a droupout layer is arranged between the convolution layers, the discarding rate is set to 0.5 for reducing overfitting, and the designed optimal parameters of the feature extraction module formed by the convolution neural network are the results obtained by continuous adjustment in experiments.
Ith duration signal sequence x i The j-th duration signal e in (a) i，j After the processing of the duration feature extraction module, the j-th duration signal e is obtained i，j Feature map a of (2) i,j Thereby obtaining the ith duration signal sequence x i Is a sequence of feature maps of (a)a i ＝{a i,1 ，a i,2 ，...，a i,j ，...，a i，l }；
Step 2.2, constructing a context encoder, comprising: forward LSTM unit and backward LSTM unit for feature map sequence a i And (3) performing time correlation learning:
according to fig. 3, the LSTM network is designed to learn the time dependence of time series, consists of gates with different functions, and different operations, the forget gate determines which information can be discarded or retained, expressed as follows:
f j ＝σ(W f ⊙[h j-1 ，a i，j ]+b f ) (9)
the purpose of updating the input gate and the output gate is to determine whether to replace the memory cell with a candidate value and to generate an active part of the current time step, which can also be expressed as:
i j ＝σ(W i ⊙[h j-1 ，a i，j ]+b i ) (10)
o j ＝σ(W o ⊙[h j-1 ，a i，j ]+b o ) (11)
the following equation governs the behavior of the LSTM cell:
Figure BDA0004139619540000071
Figure BDA0004139619540000072
h j ＝o j *tanh(C j ) (14)
wherein sigma is a sigmoid activation function, tanh is a tanh activation function, W represents a weight matrix, and a i，j To be the eigenvector of the input sequence at time j, h j The hidden state at the j-th moment is represented by b, and b is represented by bias;
sequence of feature maps a i Is the j-th feature map a of (2) i,j Via forward LSTM cellAfter processing to obtain the forward hidden state at output j
Figure BDA0004139619540000073
Jth feature map a i,j Obtaining the backward hidden state of the output j moment after the processing of the backward LSTM unit>
Figure BDA0004139619540000074
Will->
Figure BDA0004139619540000075
And->
Figure BDA0004139619540000076
After splicing, a characteristic diagram sequence a is obtained i The jth sub-feature map of the context encoder output,
Figure BDA0004139619540000081
a two-way long-short-period memory unit is selected, so that the time transition relation of the duration signal can be learned from two directions; thereby obtaining a characteristic diagram sequence a i Feature map sequence H output through context encoder i ＝{H i,1 ,H i,2 ,...,H i,j ,...H i,l }；
Step 2.3, constructing a capsule network, wherein the capsule network is composed of a primary capsule layer and a digital capsule layer as shown in fig. 4;
step 2.3.1, the primary capsule layer is formed by a one-dimensional convolution kernel with size K p X 1, number of convolution kernels N p Is a convolution layer of length l d Is composed of a Reshape operation layer;
sequence of feature maps H i Jth sub-feature map H i,j The primary characteristic map P is obtained after the convolution layer treatment of the primary capsule layer i,j This convolution layer is a one-dimensional convolution layer of 1*1, so the convolution kernel size is set to 1 because the 1*1 convolution layer can better fuse features between different channels, thereby converting to a primary capsule better, increasing nonlinearity, and keeping the number of convolution kernels consistent with the number of channels of the input features. Then pass throughAfter the Reshape operation layer is processed, outputting a j-th sub-feature diagram H i,j Primary capsule sequence s of (2) i,j And s i,j ＝{s 1 i,j ,s 2 i,j ,...,s k i,j ,...,s n i,j }，
Figure BDA0004139619540000082
Representing the i-th duration sequence x i J-th duration signal e i,j Corresponding primary capsule sequence s i,j N is the primary capsule sequence s i,j The number of the capsules, and n=l P /l d ；l d Represents the kth capsule->
Figure BDA0004139619540000083
Length of l P Representing P i,j K e {1,2,3., n };
step 2.3.2 Using the squaring activation function on the Primary Capsule sequence s i,j The kth capsule of (a)
Figure BDA0004139619540000084
After treatment, an activated primary capsule is obtained>
Figure BDA0004139619540000085
The function of the square activation has two functions, namely, the function of normalizing the capsules and the function of improving the nonlinear learning ability of a capsule network, thereby obtaining the sequence x with the ith duration i Is the jth duration signal e i,j Corresponding activated primary capsule sequence v i,j ＝{v 1 i,j ,v 2 i,j ,...,v k i,j ,...,v n i,j }；
Step 2.3.3, as shown in FIG. 5, the primary capsule sequence v to be activated i,j Is input into a digital capsule layer, and a primary capsule sequence v is calculated by using the formula (1) i,j Mapped f-th subspace predictive capsule sequence
Figure BDA0004139619540000086
The kth predictive capsule of (a)>
Figure BDA0004139619540000087
Thereby calculating the primary capsule sequence v using formula (2) i,j Mapped f-th subspace predictive capsule sequence
Figure BDA0004139619540000088
Figure BDA0004139619540000089
Figure BDA00041396195400000810
In the formula (1) and the formula (2), W f Representing the f-th matrix of the learner,
Figure BDA00041396195400000811
representation matrix W f Is the kth sub-matrix of (c); f e {1,2,3., D };
step 2.3.4, calculating the f-th subspace predictive capsule sequence by the digital capsule layer through the method (2)
Figure BDA0004139619540000091
Similarity matrix between each prediction capsule of (a)>
Figure BDA0004139619540000092
Figure BDA0004139619540000093
In the formula (2), the amino acid sequence of the compound,
Figure BDA0004139619540000094
representing predicted capsule sequence->
Figure BDA0004139619540000095
Is a transpose of (2);
step 2.3.5, calculating the f-th subspace predictive capsule sequence by the digital capsule layer through the method (4)
Figure BDA0004139619540000096
The kth sub-capsule of->
Figure BDA0004139619540000097
Weight occupied ∈>
Figure BDA0004139619540000098
Figure BDA0004139619540000099
In the formula (4), the amino acid sequence of the compound,
Figure BDA00041396195400000910
represents->
Figure BDA00041396195400000911
Is->
Figure BDA00041396195400000912
Similarity matrix representing dimensions n x n>
Figure BDA00041396195400000913
The p-th row and the o-th column of the elements;
step 2.3.5, calculating the f-th subspace predictive capsule sequence by the digital capsule layer through the method (5)
Figure BDA00041396195400000914
The length of the output is l d+1 Advanced feature capsule->
Figure BDA00041396195400000915
Thereby obtaining the j-th duration signal e related to all the D sleep categories i,j Corresponding capsule sequence->
Figure BDA00041396195400000916
The attention-based routing mechanism comprises two parts, the first part being a matrix W f It is intended to generate predictive capsules. The second part is the process of assigning weights and generating new capsules from the attention mechanism. The two interact to route for self-attention mechanism. />
Figure BDA00041396195400000917
Step 2.3.6 for Capsule sequence U i,j Medium f high-grade characteristic capsule
Figure BDA00041396195400000918
Taking l 2 Norms, thereby obtaining the ith duration sequence x i J-th duration signal e i,j Probability of presence of the f-th sleep class +.>
Figure BDA00041396195400000919
And then obtain the ith duration sequence x i J-th duration signal e i,j Predicting probability vectors of the presence of all sleep categories output via a sleep staging network
Figure BDA00041396195400000920
Step 3, constructing a loss function l of the electroencephalogram sleep stage network by utilizing the step 7 MFE ：
Figure BDA00041396195400000921
In the formula (7), the amino acid sequence of the compound,
Figure BDA0004139619540000101
is the j-th duration signal e i,j Is the true label y of (2) i,j Class f existence probability in the single thermal coding of (1), since sleep stage task is an unbalanced distribution classification task, using balanced distribution loss function will increase networkThe probability of the signal being identified as the maximum number of classes in the training set, which causes errors, can be well solved by using an MFE weighted loss function which forces the network to prioritize the classes with small sample size, thereby alleviating the problem of unbalanced distribution caused by the sample size, and in addition, the loss function has priori property due to the fact that the distribution of the verification set cannot be accurately determined, without adjusting the super-parameters
To improve network performance, reconstruction regularization is used, as shown in fig. 6, by placing a full connection layer at the end of the network to reconstruct the original input signal, reducing the network's overfitting, the euclidean distance between the reconstructed signal and the original signal is used to calculate the reconstruction loss, by multiplying the regularized loss by a coefficient γ and adding it to the loss function, the total loss being:
l total ＝l MFE +γ*l Recon (15)
and 4, training the sleep stage network by using a gradient descent algorithm, and minimizing a loss function to optimize network parameters until the loss function converges to obtain a sleep stage model for realizing sleep stage of the single-channel electroencephalogram signal.
In this embodiment, an electronic device includes a memory for storing a program supporting the processor to execute the above method, and a processor configured to execute the program stored in the memory.
In this embodiment, a computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the method described above.
Examples:
experimental data set: the model was evaluated by a Sleep-EDFx dataset in which 153 PSG recordings were used, from studies of age effects on healthy Subjects (SC), collected by 78 participants, for comparison with existing methods by dividing Sleep-EDFx into two datasets, sleep-EDF-39 and Sleep-EDF-153, sleep-EDF-39 included the first 20 subjects of SC, excluding 13 subjects with only one night of data, and other 19 persons with two night of data; there were 78 subjects in the Sleep-EDF-153. To demonstrate the broad applicability of the model to different channels, the method of the present invention was tested on the Fpz-Cz channel, the Pz-Oz channel and the Eog channel, respectively:
as shown in table 1, the model was evaluated using a k-fold cross-validation scheme, with k equal to 20 for Sleep-EDF-39, which means that 19 subjects 'data were used as training sets, 1 subject's data were used as test sets, and 10% of the remaining training set was used for validation, with mini-batch for training set 20.
Table 1: sample distribution and number of subjects for sleep stage class for each dataset
Dataset N s W N1 N2 N3 REM Total
Sleep-EDF-39 20 10197 2804 17799 5703 7717 44220
Sleep-EDF-153 78 69824 21522 69132 13039 25835 199352
Experimental parameters: the length l of each sequence is 15, and parameters lr, beta of Adam optimizer are used for updating the weight of the model 1 And beta 2 Set to 10-4, 0.90 and 0.999 respectively, for regularization, the L2 regularization term is added to the loss function with a weight of 10-3, the reconstruction loss in the capsule network is also applied to the loss function with a weight γ set to 10-4. The epochs of the maximum number of networks trained is set to 200 and if the loss of the validation set does not decrease within 20 iterations, the network will stop training and save the weight with the lowest validation loss as the optimization weight. Training was accelerated by using Tensorflow as the framework of the experiment and using a NVIDIA GeForce GTX 1070 video card with 8GB video memory.
Evaluation parameters: evaluating the performance of the model by using precision per class (PR), recall per class (RE), F1 score per class (F1), macroscopic average F1 score (MF 1) and overall precision (ACC), and Kappa coefficient (k) of Cohen, PR reflecting the proportion of true positive samples in positive cases determined by the classifier, RE reflecting the proportion of true positive cases in all samples determined to be positive by the classifier, F1 being a comprehensive evaluation based on PR and RE, F1 being a more efficient test method when F1 is high, MF1 being the average of all classes F1 reflecting the overall performance of the model, kappa coefficient (k) of Cohen's being used to measure the consistency of two different sleep stage results, corresponding to the model results and labels formulated by the sleep expert, the calculation formula being:
Figure BDA0004139619540000111
Figure BDA0004139619540000112
Figure BDA0004139619540000113
Figure BDA0004139619540000114
Figure BDA0004139619540000115
Figure BDA0004139619540000116
wherein TP represents true positive, TP f Represents true positive of class f, TN represents true negative, TN f Represents true negative of class f, FP refers to false positive, FP f Representing a false positive of class f, FN is a false negative, FN f Represents a false negative of class f, P e Is the probability of hypothetical contingency, a hyper-parameter.
Experimental evaluation: classification performance references table 2, table 3 and table 4 list the confusion matrices applied to proposed models of Fpz-Cz, pz-Oz and Eog channels, respectively, the diagonal elements of the confusion matrices representing the consistency of the labels with expert classification, while the right side of the table represents the evaluation parameters for each category, the higher the evaluation parameters, the better the performance of Fpz-Cz channel is for the channel used, while the performance of Eog channels is the worst, the best performance being at W and N2 phases and the worst performance being at N1 phase, which is easily confused with other sleep phases, mainly due to the small amount of data at N1 phase, and the similarity of N1 phase with other sleep phases, which makes N1 phase more difficult to distinguish from other sleep phases.
Table 2: the output confusion matrix for the model of the Fpz-Cz channel signal in the Sleep-EDF-39 dataset is used.
Figure BDA0004139619540000121
Table 3: the output confusion matrix for the model of the Pz-Oz channel signal in the Sleep-EDF-39 dataset is used.
Figure BDA0004139619540000122
Table 4: the output confusion matrix for the model of the EOG channel signals in the Sleep-EDF-39 dataset is used.
Figure BDA0004139619540000131
By using ablation experiments to represent the function of each component of the model, by comparing the performance of the model with and without the part-capsule network, the model for comparison consists of the following four basic models:
(1) the method comprises the following steps Cnn+lstm, comprising one CNN and one unidirectional LSTM;
(2) the method comprises the following steps CNN+Bi-LSTM comprising a CNN and a Bi-directional LSTM;
(3) the method comprises the following steps CNN+Bi-LSTM+CapsNet, comprising a CNN and a Bi-directional LSTM and capsule network, without reconstruction loss;
(4) the method comprises the following steps The proposed model, sleepcapsule net, contains a CNN and a reconstruction penalty for a bi-directional LSTM and capsule network.
By evaluating the performance of different structures with macroscopic averaged F1 fraction (MF 1), overall Accuracy (ACC), and Kappa coefficients for Cohen's, FIG. 7 depicts a histogram of performance of various models under different evaluation parameters, bi-directional LSTM performs better than uni-directional LSTM models, mainly because uni-directional LSTM models can capture only one directional time dependence, while Bi-directional LSTM models can capture both directional time dependence, which is beneficial for learning transition rules in sleep stages, and CNN+Bi-LSTM+CapsNet models perform better than CNN+Bi-LSTM models, which is higher in all three evaluation indexes, indicating that adding a capsule network can improve the performance of a Bi-LSTM network, and by comparing the impact of reconstruction losses on network performance, reconstruction losses can improve the performance of the network, thereby reducing overfitting, and achieving optimal performance by minimizing reconstruction regularization losses.
By comparing the model with the most advanced methods, the present invention evaluates parameters such as macroscopic average F1 score (MF 1), overall Accuracy (ACC), and Cohen's Kappa coefficient.
Table 5 shows the results of the comparison with the current most advanced methods, the model comprising DeepSleepNet, sleepEEGNet, tinySleepNet, CCRRSleepNet, attnNet, EOGNET, the model being superior to the other models in terms of accuracy, up to 85.84% on the Sleep-EDF-39 and up to 83.4% on the Sleep-EDF-153, in particular the model being superior to the other models in terms of the following.
In terms of W-phase recognition accuracy and generalization performance over large datasets, using EOG channel signals for staging, the model achieves superior results compared to EOGNet, exceeding EOGNet in all respects, the model is a sequence-to-sequence model, using multiple consecutive durations as inputs, outputting sleep stages corresponding to each duration, while deepsenet employs a two-step training method, first learning a representation of each epoch, then training RNN to learn transition relationships between epochs, but the model does not require two-step learning, the first half of the model is comparable to the model of tinysenet, except if features extracted from CNN are flattened and fed LSTM, tinySleepNet require small volume data enhancement during training, attnNet is a network that uses multiple attentions to capture time dependence, unlike the network of the present invention, it can achieve optimal results by using three epochs as inputs, and outputting intermediate epoch tags. Furthermore, its loss function is a hyper-parametric loss function, which lacks reasonable interpretation, requiring hyper-parametric adjustments for different data sets, whereas the network of the present invention is not.
Table 5 evaluation parameter graphs comparing the methods herein with the most advanced methods
Figure BDA0004139619540000151
In terms of W-phase recognition accuracy and generalization performance over large datasets, using EOG channel signals for staging, the model achieves superior results compared to EOGNet, exceeding EOGNet in all respects, the model is a sequence-to-sequence model, using multiple consecutive durations as inputs, outputting sleep stages corresponding to each duration, while deepsenet employs a two-step training method, first learning a representation of each epoch, then training RNN to learn transition relationships between epochs, but the model does not require two-step learning, the first half of the model is comparable to the model of tinysenet, except if features extracted from CNN are flattened and fed LSTM, tinySleepNet require small volume data enhancement during training, attnNet is a network that uses multiple attentions to capture time dependence, unlike the network of the present invention, it can achieve optimal results by using three epochs as inputs, and outputting intermediate epoch tags. Furthermore, its loss function is a class-aware loss function, lacking reasonable interpretation, requiring hyper-parameter adjustment for different data sets, whereas the network of the present invention is not.
The model in the invention can be applied to different EEG channels (Fpz-Cz and Pz-Oz) and EOG channels without modifying the structure and training algorithm of the model, and from the experimental results carried out on each channel, the use of the Fpz-Cz channel for Sleep staging obtains better performance than other channels, which indicates that the Fpz-Cz channel has substantially more Sleep related information, the best performance is in the W stage and the worst performance is in the N1 stage, the identification of the N1 stage is always the difficulty of Sleep staging task, on one hand, the data of the N1 stage is relatively limited, on the other hand, the staging itself is difficult and is easy to confuse with other stages, the model proposed in the invention can be used for Sleep staging from untreated original single brain electrical signals, this is friendly for subjects and persons lacking expertise in the Sleep area, as it reduces the discomfort of the measurement of the subject by using single channel EEG signals, which however also increase the risk of Sleep stage inaccuracy, do not always appear to be Sleep related signatures, even containing different degrees and types of noise signals, e.g. in Sleep-EDF-39, the subject 11 verification accuracy on the Fpz-Cz channel is only 60%, which is the worst of 20-fold cross-verification results, mainly due to the presence of high intensity noise signals in the verification set, leading to other classes of duration being identified as W-or R-cycles, the model in the present invention captures this inherent spatial relationship by introducing a network of capsules to obtain a more accurate, compared with the LSTM, the Bi-LSTM can consider time reference information from two directions, so that Sleep stage is more accurate, the self-attention routing mechanism used in the invention can fully consider the inherent similarity among capsules, and parameters are fewer, a capsule network can be widely applied to Sleep stage tasks based on EEG signals through the high-efficiency self-attention mechanism, in clinical practice, sleep experts need to refer to single-channel EEG signals and other channels, the characteristics related to Sleep, such as Sleep spindle waves, are not simultaneously present in all EEG signals, eyeball motion and electromyographic signals are important to judge Sleep stage, in addition, the scoring styles of the experts are different, and different expert scoring styles are different, so that different scoring results are caused, and even so, the overall kappa value of a model in the invention on the Sleep-EDF-39 exceeds 0.8, therefore, the invention has the more practical value compared with the method of learning the Sleep stage data in the invention by the most advanced method, such as the advanced Sleep stage, and the method has the more time-consuming value. The model of the invention is used for sleep stage, the model can be trained on data accumulated in hospitals, the sleep stage can be completed in a short time without manually extracting characteristics, which is beneficial to a monitored person to quickly make psychological or physiological adjustment according to the result, in particular, the invention can reduce the uncomfortable feeling of the infant in infant sleep monitoring, which is beneficial to monitoring the development of the infant, and the invention can be more convenient for individual sleep monitoring by combining the model of the invention with a physical electroencephalogram signal acquisition tool.

Claims (7)

1. A single-channel electroencephalogram sleep stage method based on deep learning is characterized by comprising the following steps:
step 1, acquiring a polysomnography of a subject and a sleep label set corresponding to the polysomnography, selecting an electroencephalogram sleep signal of one channel in the polysomnography, and dividing the electroencephalogram sleep signal under one channel into N non-overlapping duration signal sequences X= { X 1 ,x 2 ,…,x i ,…,x N Building a tag sequence set y= { Y according to sleep tags corresponding to each duration 1 ,y 2 ,…,y i ,…,y N Simultaneously counting the number { N } of the tags of each sleep category in the sleep tags corresponding to all durations f |f=1, 2, …, D }; wherein x is i Represents the i-th duration signal sequence, and x i ＝{e i,1 ,e i,2 ,…,e i,j ,…,e i,l }，y i Representing the ith duration signal sequence x i Corresponding sleep tag sequences, and y i ＝{y i,1 ,y i,2 ,…,y i,j ,…,y i,l }，e i,j ∈R 1×m ，e i,j Represents the ithThe j-th duration signal in the sequence of duration signals, y i,j Tag indicating the jth duration signal in the ith duration signal sequence, N f Tag number representing the f-th sleep category, f e {1,2,..once., D }, y i,j E {1,2, …, D }, D is the total class of sleep stages, j E {1,2 … l }, l is x i The number of the duration signals existing in the network, m represents the number of sampling points, m=θ×δ, θ represents the sampling rate, and tiger represents the sampling time;
step 2, setting up an electroencephalogram sleep stage network based on deep learning, comprising the following steps: the duration feature extraction module, the context encoder and the capsule network are formed;
step 2.1, the duration feature extraction module extracts the j-th duration signal e i,j After processing, the j-th duration signal e is obtained i,j Feature map a of (2) i,j Thereby obtaining the ith duration signal sequence x i Feature map sequence a of (2) i ＝{a i,1 ,a i,2 ,…,a i,j ,…,a i,l }；
Step 2.2, the context encoder is used for the feature map sequence a i Time correlation learning is carried out to obtain a characteristic diagram sequence H i ＝{H i,1 ,H i,2 ,...,H i,j ,...H i,l }；
Step 2.3, constructing a capsule network, which consists of a primary capsule layer and a digital capsule layer;
step 2.3.1, the primary capsule layer vs. the feature map sequence H i Jth sub-feature map H i,j After treatment, a primary capsule is obtained
Figure FDA0004139619530000011
Thereby obtaining the sequence x with the ith duration i Is the jth duration signal e i,，j Corresponding activated primary capsule sequence v i,j ＝{v 1 i,j ,v 2 i,j ,...,v k i,j ,...,v n i,j }；
Step 2.3.2 Primary Capsule sequence v to be activated i,j Inputting into the digital capsule layer for processing to obtain and allJ-th duration signal e related to D sleep categories i,j Corresponding capsule sequences
Figure FDA0004139619530000012
Step 2.3.3, sequence of capsules U i,j Medium f high-grade characteristic capsule
Figure FDA0004139619530000013
Taking l 2 Norms, thereby obtaining the ith duration sequence x i J-th duration signal e i,，j Probability of presence of the f-th sleep class +.>
Figure FDA0004139619530000014
And then obtain the ith duration sequence x i J-th duration signal e i,j Predicting probability vectors of the existence of all sleep categories output via the sleep staging network
Figure FDA0004139619530000021
Step 3, constructing a loss function l of the electroencephalogram sleep stage network by using the formula (7) MFE ：
Figure FDA0004139619530000022
In the formula (7), the amino acid sequence of the compound,
Figure FDA0004139619530000023
is the j-th duration signal e i,j Is the true label y of (2) i,j The existence probability of the f-th class in the single-hot coding of (a);
and 4, training the sleep stage network by using a gradient descent algorithm, and minimizing the loss function to optimize network parameters until the loss function converges to obtain a sleep stage model for realizing sleep stage of the single-channel electroencephalogram signal.
2. The method for single-channel electroencephalogram sleep stage separation based on deep learning according to claim 1, wherein the duration feature extraction module in step 2.1 sequentially comprises a one-dimensional step length of S 1 The convolution kernel is K 1 A first convolutional layer, a first BN normalization layer, a first Relu nonlinear active layer, a pooling size k m1 Step length s m1 Is included in the first maximum pooling layer; step length S 2 The convolution kernel is K 2 A second BN normalization layer, a second Relu nonlinear activation layer; step length S 2 The convolution kernel is K 2 A third BN normalization layer, a third Relu nonlinear activation layer; step length S 2 The convolution kernel is K 2 A fourth convolution layer, a fourth BN normalization layer, a fourth Relu nonlinear activation layer, a pooling size of k m2 Step length s m2 Is stacked with the second largest pooling layer.
3. The single channel electroencephalogram sleep staging method according to claim 2, characterized in that said context encoder in step 2.2 comprises: a forward LSTM cell and a backward LSTM cell;
the characteristic diagram sequence a i Is the j-th feature map a of (2) i,j Obtaining the forward hidden state at output j time after processing by the forward LSTM unit
Figure FDA0004139619530000024
Jth feature map a i,j Obtaining the backward hidden state of the output j moment after the processing of the backward LSTM unit>
Figure FDA0004139619530000025
Will->
Figure FDA0004139619530000026
And->
Figure FDA0004139619530000027
After splicing, a characteristic diagram sequence a is obtained i At the jth sub-feature map of the context encoder output,
Figure FDA0004139619530000028
thereby obtaining the characteristic diagram sequence a i Feature map sequence H output through context encoder i ＝{H i,1 ,H i,2 ,...,H i,j ,...H i,l }。
4. The method of deep learning-based single-channel electroencephalogram sleep stage according to claim 3, wherein the primary capsule layer in step 2.3.1 is composed of a one-dimensional convolution kernel with size of K p X 1, number of convolution kernels N p Is a convolution layer of length l d Is composed of a Reshape operation layer;
the characteristic diagram sequence H i Jth sub-feature map H i,j The primary characteristic map P is obtained after the convolution layer treatment of the primary capsule layer i,j After the Reshape operation layer is processed, outputting a j-th sub-feature diagram H i,j Primary capsule sequence s of (2) i,j And (2) and
Figure FDA0004139619530000031
Figure FDA0004139619530000032
representing the i-th duration sequence x i J-th duration signal e i,j Corresponding primary capsule sequence s i,j N is the primary capsule sequence s i,j The number of the capsules, and n=l P /l d ；l d Represents the kth capsule->
Figure FDA0004139619530000033
Length of l P Representing P i,j K e {1,2,3., n };
using a squaring activation function on the primary capsule sequence s i,j The kth capsule of (a)
Figure FDA0004139619530000034
After treatment, an activated primary capsule is obtained>
Figure FDA0004139619530000035
Thereby obtaining the sequence x with the ith duration i Is the jth duration signal e i,j Corresponding activated primary capsule sequence v i,j ＝{v 1 i,j ,v 2 i,j ,...,v k i,j ,...,v n i,j }。
5. The method of single channel electroencephalogram sleep stage based on deep learning according to claim 4, wherein the digital capsule layer in step 2.3.2 calculates the primary capsule sequence v by using formula (1) i,j Mapped f-th subspace predictive capsule sequence
Figure FDA0004139619530000036
The kth predictive capsule of (a)>
Figure FDA0004139619530000037
Thereby calculating the primary capsule sequence v using formula (2) i,j Mapped f-th subspace predictive capsule sequence +.>
Figure FDA0004139619530000038
Figure FDA0004139619530000039
Figure FDA00041396195300000310
In the formula (1) and the formula (2), W f Representing the f-th matrix of the learner,
Figure FDA00041396195300000311
representation matrix W f Is the kth sub-matrix of (c); f e {1,2,3., D }; />
The digital capsule layer calculates the f-th subspace predictive capsule sequence by utilizing the method (3)
Figure FDA00041396195300000312
Similarity matrix between each prediction capsule of (a)>
Figure FDA00041396195300000319
Figure FDA00041396195300000313
In the formula (3), the amino acid sequence of the compound,
Figure FDA00041396195300000314
representing predicted capsule sequence->
Figure FDA00041396195300000315
Is a transpose of (2);
the digital capsule layer calculates the f-th subspace predictive capsule sequence by utilizing the method (4)
Figure FDA00041396195300000316
The kth sub-capsule of->
Figure FDA00041396195300000317
Weight occupied ∈>
Figure FDA00041396195300000318
Figure FDA0004139619530000041
In the formula (4), the amino acid sequence of the compound,
Figure FDA0004139619530000042
represents->
Figure FDA0004139619530000043
Is->
Figure FDA0004139619530000044
Similarity matrix representing dimensions n x n>
Figure FDA0004139619530000045
The p-th row and the o-th column of the elements;
the digital capsule layer calculates the f-th subspace predictive capsule sequence by utilizing the method (5)
Figure FDA0004139619530000046
The length of the output is l d+1 Advanced feature capsule->
Figure FDA0004139619530000047
Thereby obtaining the j-th duration signal e related to all the D sleep categories i,j Corresponding capsule sequences
Figure FDA0004139619530000048
Figure FDA0004139619530000049
6. An electronic device comprising a memory and a processor, wherein the memory is configured to store a program that supports the processor to perform the single channel electroencephalogram sleep staging method of any one of claims 1-5, the processor being configured to execute the program stored in the memory.
7. A computer readable storage medium having stored thereon a computer program, characterized in that the computer program when executed by a processor performs the steps of the single channel electroencephalogram sleep staging method of any one of claims 1-5.
CN202310285377.5A 2023-03-22 2023-03-22 Single-channel electroencephalogram sleep stage-dividing method based on deep learning Pending CN116186544A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310285377.5A CN116186544A (en) 2023-03-22 2023-03-22 Single-channel electroencephalogram sleep stage-dividing method based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310285377.5A CN116186544A (en) 2023-03-22 2023-03-22 Single-channel electroencephalogram sleep stage-dividing method based on deep learning

Publications (1)

Publication Number Publication Date
CN116186544A true CN116186544A (en) 2023-05-30

Family

ID=86446424

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310285377.5A Pending CN116186544A (en) 2023-03-22 2023-03-22 Single-channel electroencephalogram sleep stage-dividing method based on deep learning

Country Status (1)

Country Link
CN (1) CN116186544A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117747117A (en) * 2024-02-21 2024-03-22 安徽星辰智跃科技有限责任公司 Sound-based sleep respiration evaluation and auxiliary adjustment method, system and device

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117747117A (en) * 2024-02-21 2024-03-22 安徽星辰智跃科技有限责任公司 Sound-based sleep respiration evaluation and auxiliary adjustment method, system and device
CN117747117B (en) * 2024-02-21 2024-05-24 安徽星辰智跃科技有限责任公司 Sound-based sleep respiration evaluation and auxiliary adjustment method, system and device

Similar Documents

Publication Publication Date Title
Sun et al. A hierarchical neural network for sleep stage classification based on comprehensive feature learning and multi-flow sequence learning
Cui et al. Automatic Sleep Stage Classification Based on Convolutional Neural Network and Fine‐Grained Segments
Sun et al. A two-stage neural network for sleep stage classification based on feature learning, sequence learning, and data augmentation
Pourbabaee et al. Feature leaning with deep convolutional neural networks for screening patients with paroxysmal atrial fibrillation
CN114052735B (en) Deep field self-adaption-based electroencephalogram emotion recognition method and system
Manjunath et al. A low-power lstm processor for multi-channel brain eeg artifact detection
Li et al. Patient-specific seizure prediction from electroencephalogram signal via multi-channel feedback capsule network
Yulita et al. Combining deep belief networks and bidirectional long short-term memory: Case study: Sleep stage classification
Jinliang et al. EEG emotion recognition based on granger causality and capsnet neural network
CN113925459A (en) Sleep staging method based on electroencephalogram feature fusion
Zhao et al. A deep learning algorithm based on 1D CNN-LSTM for automatic sleep staging
CN116186544A (en) Single-channel electroencephalogram sleep stage-dividing method based on deep learning
Yuan et al. Multivariate sleep stage classification using hybrid self-attentive deep learning networks
Zhu et al. Predicting ICU mortality by supervised bidirectional LSTM networks.
Guan Application of logistic regression algorithm in the diagnosis of expression disorder in Parkinson's disease
Qin et al. Deep multi-scale feature fusion convolutional neural network for automatic epilepsy detection using EEG signals
CN110522446A (en) A kind of electroencephalogramsignal signal analysis method that accuracy high practicability is strong
Giri et al. Combining generative and discriminative neural networks for sleep stages classification
Lv et al. A multilevel temporal context network for sleep stage classification
Rammy et al. Sequence-to-sequence deep neural network with spatio-spectro and temporal features for motor imagery classification
CN116763324A (en) Single-channel electroencephalogram signal sleep stage method based on multiple scales and multiple attentions
CN115399735A (en) Multi-head attention mechanism sleep staging method based on time-frequency double-current enhancement
Huang et al. A two-phase model for sleep staging using single channel eeg
Wang et al. Sleep staging based on multi scale dual attention network
Huang et al. A deep learning-based method for sleep stage classification using physiological signal

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