CN113116361A - Sleep staging method based on single-lead electroencephalogram - Google Patents

Abstract

The invention discloses a single-lead brain electricity based staging method and device. The invention adopts a single-lead EEG signal feature fusion extraction algorithm based on triangular filtering and Convolution (CNN) network, which is characterized in that: for 30-s EEG, extracting the cepstrum characteristic of a signal by using triangular filtering and deep information extracted by two CNN networks with different convolution kernels, and combining the cepstrum characteristic and the deep information in a defined dimension to serve as a characteristic value sent to a classification network; the classification network adopts a Bi _ Lstm + ATTETION network, fully considers the correlation between the front and the back of the information, adds ATTENTION (ATTENTION mechanism), and distributes corresponding weight to the importance degree of the classification result according to the characteristic information so as to improve the accuracy of classification. The algorithm flow of the method is verified on the SLEEP EDF-2013 data set, and the result proves that the method provided by the invention is superior to the traditional method and other classification network algorithms and has better classification effect.

Description

Sleep staging method based on single-lead electroencephalogram

Technical Field

The invention belongs to the field of electroencephalogram signal processing, and particularly relates to a sleep staging method based on single-lead electroencephalogram.

Background

Sleep is an important physiological phenomenon, and sleep stage is one of important evaluation and diagnosis methods for various clinical neurologic diseases. At present, sleep staging based on physiological signals, especially on electroencephalogram signals, has become an important research direction.

EEG (Electroencephalogram) has the characteristics of simple recording, easy operation, no damage, high repeatability and the like. The electroencephalogram signals contain rich brain physiological information, have different waveforms and signal intensities in different sleep periods, have better difference, are weak, and are easily influenced by external factors under the condition of strong background noise, and common interferences include electrooculogram, sharp pulse, white noise and the like. How to effectively remove external noise and interference from the original EEG signal and extract effective characteristic information becomes the key point of the related research work of sleep stage based on the EEG signal.

The traditional signal feature extraction method mainly comprises four aspects, namely a time domain feature analysis method, a frequency domain feature analysis method, a time-frequency domain feature analysis method and a nonlinear dynamics analysis method. The time domain analysis method mainly comprises the following steps: extracting parameters such as mean value, variance, root mean square, peak factor, kurtosis coefficient, form factor, pulse factor and the like from an EEG original signal as characteristics of the signal, and analyzing and researching the signal; the frequency domain analysis method comprises the following steps: extracting characteristic parameters such as frequency spectrum, power spectrum, characteristic frequency, mean square frequency, center of gravity frequency, frequency variance and the like of the signal by methods such as short-time Fourier transform and the like to analyze the signal; the time-frequency domain analysis method comprises the following steps: wavelet transform, hilbert yellow transform, and the like; the nonlinear dynamics analysis method comprises the following steps: entropy, spectral entropy, permutation entropy, wavelet entropy, Empirical Mode Decomposition (EMD) entropy, envelope spectral entropy, etc. of the analysis signal. Although the feature extraction algorithm based on the classical priori knowledge has good stability, some important signal features can be missed due to limited or missing a priori knowledge. Moreover, the traditional feature extraction method needs to perform denoising processing on data independently, the process is complicated, and real-time and rapid classification is difficult to realize in practical application.

The feature extraction method based on deep learning automatically selects signal features with distinctiveness by means of strong learning ability of a deep network. At present, the improvement of calculation power enables people to avoid worries about hardware resources, and the automatic digging of analysis data through a neural network becomes a research hotspot. The electroencephalogram signal also has extremely strong front-back correlation characteristics, and the front-back correlation information in sequence of time sequence can be fully utilized by using the deep learning network, so that a better experimental effect can be achieved. A Recurrent Neural network (Recurrent Neural Networks) can capture context and associated information, and is also a deep network for earliest application timing sequence feature extraction. Long Short Term Memory (LSTM) networks, which are improved on the basis of RNNs, have made important progress in speech signal recognition and are in accordance with life practices. Convolutional Networks (CNN) are a type of Neural Networks with a feedforward function, have characterization learning capacity, can perform invariant classification translation on input information according to the structure of the input information, and obtain better results in sleep stage research which is increasingly applied in recent two years. However, the network based on deep learning is too sensitive to the experimental parameter values, and the result is sporadic.

Disclosure of Invention

In order to accelerate the classification speed and improve the classification accuracy so as to better realize clinical application, the invention provides a sleep staging method based on the combination of triangular filtering and a Convolutional Neural Network (CNN), and the method can utilize more and more comprehensive signal characteristic values. In addition, the invention uses a bidirectional recurrent neural (Bi _ Lstm) classification network which is more suitable for time sequence signal classification, and adds an ATTENTION mechanism (ATTENTION), thereby further improving the efficiency and the accuracy of classification.

The technical scheme adopted by the invention is as follows:

a sleep staging method based on single-lead electroencephalogram fully utilizes the advantages that analysis of triangular filtering on signal cepstrum features and a Convolutional Neural Network (CNN) can extract potential deep-level information of signals, and features of electroencephalogram signals are extracted, classified, extracted, fused and obtained to obtain signal features which are most beneficial to signal separation; and classifying by using a recurrent neural network Bi _ Lstm which has high discrimination of the time-series signals and is more matched with the signal characteristics, and the specific operation steps are as follows:

firstly, data preprocessing, namely extracting an original EEG (electroencephalogram) signal EEG (electroencephalogram) in a sleep multi-lead map data set, then performing data cleaning, and deleting data without labels; selecting a required signal frequency band, and simply filtering the signal by using a Butterworth filter to reserve a required signal frequency component; then, carrying out normalization processing on the filtered signals to obtain data corresponding to the sleep states one by one;

inputting the experimental data obtained in the step (I) into a triangular filtering cepstrum analysis program, firstly performing Fast Fourier Transform (FFT), then selecting a proper triangular filter to refine the nonlinear description of the signal cepstrum characteristic, obtaining a two-dimensional cepstrum characteristic coefficient matrix related to the EEG signal after triangular filtering, taking the coefficient matrix as the cepstrum characteristic value of the EEG signal, and spreading the two-dimensional matrix into a one-dimensional data form;

thirdly, sending the data in the step one into two one-dimensional convolutional neural networks CNN with different convolutional kernel sizes, extracting potential features, combining the potential features in a specified dimension, and extracting deep level information of the signals in different scales to obtain required final signal features;

and (IV) merging the frequency-reversed data obtained by the processing in the step (II) and the step (III) and deep characteristic data extracted by the CNN in a third dimension, carrying out characteristic fusion, taking the fused characteristic vector as the characteristic data of a signal, carrying out dimension conversion before the data is sent into a classification network, changing the data into a data form corresponding to the labels one by one, carrying out normalization processing on the data, and finally sending the data into a bidirectional recursive neural network for classification to obtain a classification result.

The method fully utilizes the analysis of triangular filtering on the signal cepstrum characteristics and the advantage that a CNN network can extract potential deep information of signals, exerts the stability of classical time-frequency domain characteristic analysis and the high-performance advantage of a neural network, extracts, fuses and acquires the signal characteristics which are most distinguished and beneficial to the sleep state; and a bidirectional recursive cyclic neural network which has high discrimination of the time-series signals and is more matched with the signal characteristics is used for classification, and an attention mechanism is added, so that the method has better classification performance and timeliness.

Drawings

FIG. 1 is a network flow chart of a single-lead brain based staging method of the present invention.

Detailed Description

The invention is described in detail below with reference to the figures and examples.

First, the present invention provides a method for extracting cepstrum features from an EEG signal based on triangular filtering.

The process of extracting the cepstrum feature of the EEG signal based on the triangular filtering in the invention is shown as follows:

a basic frequency domain transformation stage: and performing framing operation on the preprocessed EEG data, wherein the original sampling frequency rate of the signal is fs and is generally set to be 100Hz, so that a 30s signal is selected as one frame. The triangular filter is mainly used for processing speech signals originally, and when the triangular filter is used for processing electroencephalogram signals, a proper windowing length and a window moving scale are selected according to characteristics of the electroencephalogram signals, and the signals are subjected to Fast Fourier Transform (FFT) after being framed.

(II) entering a stage of introducing a 24-order triangular filter bank:

according to the formula

The common frequency scale is converted into a Mel frequency scale by a triangular filtering method, wherein f represents the original frequency of the signal and has the unit of Hz, Mel (f) represents the Mel scale, and 2595 is an adjusting coefficient for coordinate conversion. In the experiment, a 24-order triangular filter group is selected and multiplied by the frequency spectrum obtained in the step one, then basic logarithm operation is carried out, Discrete Cosine Transform (DCT) is carried out to obtain a two-dimensional cepstrum coefficient matrix and the two-dimensional cepstrum coefficient matrix is flattened so as to adapt to the data form of feature fusion in the following process, and extraction of the cepstrum characteristic value of the EEG signal by the triangular filter is completed.

Second, ATTENTION (ATTENTION) mechanisms are added to the classification network.

The ATTENTION mechanism is a solution proposed by imitating human ATTENTION, and simply, high-value information is quickly screened from a large amount of information. The method is mainly used for solving the problem that the final reasonable vector representation is difficult to obtain when the input sequence of the LSTM/RNN model is long, and the method is characterized in that the intermediate result of the LSTM is reserved, the LSTM is learned by a new model, and the LSTM is associated with the output, so that the purpose of information screening is achieved.

Thirdly, the invention is a method for sleep staging research based on triangle filtering, Convolution (CNN) network and recurrent neural network classification network.

As shown in fig. 1, the specific steps of sleep staging based on triangular filtering and Convolution (CNN) network and recursive cyclic neural classification network proposed by the present invention are as follows:

the method comprises the steps that (I) data are input into a CNN network in a dotted frame, multi-level and multi-scale feature information is needed, so as shown in the figure, two one-dimensional CNN networks with different basic convolution kernels are arranged, one convolution kernel is half of sampling frequency, namely, the convolution kernel is set to be fs/2, the other convolution kernel is set to be 4 times of the sampling frequency, namely, set to be fs x 4, then after three times of basic convolution operation, pooling operation is respectively carried out, the two CNN networks are combined in the third dimension of the data, in order to further reduce parameters of the networks and complex co-adaptation relations among neurons, a dropout layer with the random loss rate of 0.5 is added, and finally needed potential deep-level feature information is obtained.

And (3) after the cepstrum feature is extracted from the other path of EEG data through triangular filtering, combining the extracted feature value with the feature value extracted from the CNN network in the step (I), and performing feature screening to obtain the signal feature value finally sent to the classification network. As shown in part B of the figure: one path of the feature value after feature fusion passes through a full-connection (full-connect) network, and simple normalization processing is carried out to be used as a part of classification features before softmax operation; and the other path of the data is used for further extracting hidden information through two layers of Bi-lstm networks, and utilizing the characteristic of good memorability of the Bi-lstm networks to obtain front and back related information of feature data, after the classification features are extracted through the Bi-lstm networks, the forward network feature values and the backward network feature values are combined to obtain 1024 classification feature values, and then the 1024 classification feature values are added with the classification features obtained by the front fully-connected networks according to the graph and input into an ATTENTION network part. And automatically giving different weights to different characteristic values according to the influence on the classification result, and then performing softmax operation to obtain five different classification results.

Fourthly, the invention provides an overall flow chart of the sleep stage dividing method of the single-lead brain based on the triangular filtering and the convolution network.

The system specifically comprises data preprocessing, triangular filtering and Convolution (CNN) network characteristic value extraction, characteristic fusion and shaping, classification by using a recurrent neural network, obtaining and storing classification results, constructing a confusion matrix (cvmax), and obtaining key evaluation indexes.

1. Raw EEG data was acquired, and the 0.3-50Hz portion of the EEG signal was screened out using a 5 th order butterworth filter, and then normalized for subsequent calculations.

2. Feature extraction: and respectively sending the processed data into a set triangular filtering program and a CNN (CNN) feature extraction network to respectively extract features.

3. Feature fusion and data shaping: and (3) performing basic feature fusion on the features extracted in the step (2), using a basic tf.concat () function, selecting the dimension of the feature value, merging, and then performing data shaping (reshape) operation to adapt to the data format required by the classified network.

4. Inputting the processed data in the step 3 into a recurrent neural network with an ATTENTION structure for classification, selecting a Focal _ loss cost function in order to further solve the problem of unbalanced data distribution and improve the accuracy of classification, better paying ATTENTION to the classification samples with less data volume and improving the network performance and convergence rate. In order to avoid the contingency of the experiment, 20-fold cross validation is used, so that the reliability of the experimental result of the user is improved.

5. After the data passes through a classification network, a confusion matrix is constructed together with the original data, and the required key parameters are calculated: overall accuracy (Overall-accuracy), cohn's kappa score, and Overall harmonic mean index (Macro-F1 accuracy). In order to show the good performance of the network system, a comparison test is carried out by adopting a traditional method and other methods, and 20-fold cross validation is carried out by adopting the same data set SLEEP-EDF2013, and the experimental result also shows that the accuracy and the performance of the method are superior to those of other models, as shown in the table 1:

TABLE 1

Claims (1)

1. A sleep staging method based on single-lead electroencephalogram fully utilizes the advantages that analysis of triangular filtering on signal cepstrum features and a Convolutional Neural Network (CNN) can extract potential deep-level information of signals, and features of electroencephalogram signals are extracted, classified, extracted, fused and obtained to obtain signal features which are most beneficial to signal separation; and classifying by using a Bi _ Lstm network which has high discrimination on the time-series signals and is more matched with the signal characteristics, and specifically operating steps are as follows:

firstly, data preprocessing, namely extracting an original EEG (electroencephalogram) signal EEG (electroencephalogram) in a sleep multi-lead map data set, then performing data cleaning, and deleting data without labels; selecting a required signal frequency band, and simply filtering the signal by using a Butterworth filter to reserve a required signal frequency component; then, carrying out normalization processing on the filtered signals to obtain data corresponding to the sleep states one by one;

inputting the experimental data obtained in the step (I) into a triangular filtering cepstrum analysis program, firstly performing Fast Fourier Transform (FFT), then selecting a proper triangular filter to refine the nonlinear description of the signal cepstrum characteristic, obtaining a two-dimensional cepstrum characteristic coefficient matrix related to the EEG signal after triangular filtering, taking the coefficient matrix as the cepstrum characteristic value of the EEG signal, and spreading the two-dimensional matrix into a one-dimensional data form;

thirdly, sending the data in the step one into two one-dimensional convolutional neural networks CNN with different convolutional kernel sizes, extracting potential features, combining the potential features in a specified dimension, and extracting deep level information of the signals in different scales to obtain required final signal features;

and (IV) merging the frequency-reversed data obtained by the processing in the step (II) and the step (III) and deep characteristic data extracted by the CNN in a third dimension, carrying out characteristic fusion, taking the fused characteristic vector as the characteristic data of a signal, carrying out dimension conversion before the data is sent into a classification network, changing the data into a data form corresponding to the labels one by one, carrying out normalization processing on the data, and finally sending the data into a bidirectional recursive neural network for classification to obtain a classification result.

Images