CN-122004760-A - Single-channel EEG sleep stage method based on feature map neural network

CN122004760ACN 122004760 ACN122004760 ACN 122004760ACN-122004760-A

Abstract

The invention provides a single-channel EEG sleep stage method based on a feature map neural network, and relates to the technical field of sleep stage. The method comprises the steps of preprocessing and segmenting a single-channel EEG signal, extracting multiple types of features from each time window, constructing different features as nodes in a graph, constructing a weighted adjacent matrix based on statistical correlation among the features, and modeling the feature graph by using a graph convolution network to realize high-precision classification of the EEG signal. The graph convolution network updates the characteristics through neighbor node weighting aggregation, learns global structural characteristics without depending on massive samples, and adapts to the characteristics of single-channel EEG data. The method realizes high-precision sleep stage based on single-channel EEG without complex multi-channel acquisition equipment, and meets the requirements of equipment cost control and convenient acquisition in clinical scenes.

Inventors

ZHOU HUI
YU WENCHANG
ZHU RONGJIE
WANG YONGQI
LOU LI

Assignees

无锡特文思达健康科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260129

Claims (10)

1. A single-channel EEG sleep stage method based on a feature map neural network is characterized by comprising the following steps: sequentially carrying out band-pass filtering, artifact removal processing and signal segmentation on an original single-channel EEG signal to obtain an EEG sample; respectively extracting time domain features, frequency domain features and nonlinear dynamics features from each EEG sample; Constructing the extracted time domain features, frequency domain features and nonlinear dynamics features into feature graphs; Modeling the feature map by adopting a map convolution network, and outputting a sleep stage result through map convolution layer feature updating, global pooling and classification calculation.
2. The feature map neural network based single channel EEG sleep staging method according to claim 1, characterized in that the de-artifacting process uses an amplitude threshold based strategy, for each EEG time window x, if its absolute amplitude maximum exceeds a preset threshold Judging the time window as a signal segment containing artifacts and eliminating, wherein the judgment rule is as follows: In the formula, The absolute amplitude maximum of the EEG time window x is represented.
3. The single channel EEG sleep staging method based on feature map neural network of claim 1, wherein the signal segments divide continuous EEG signals into non-overlapping time windows X of length T seconds, Wherein M is the total number of time windows, , Is the signal sampling rate.
4. The single channel EEG sleep staging method based on feature map neural network of claim 1, wherein the time domain features include mean, standard deviation, skewness, kurtosis, hjorth parameters, zero crossing rate; the frequency domain features comprise power spectrum density, frequency band relative energy, spectrum centroid and spectrum entropy; The nonlinear dynamics features comprise sample entropy, approximate entropy, hurst index and fractal dimension.
5. The single channel EEG sleep staging method based on feature map neural network of claim 1, wherein the feature map is a weighted undirected map Wherein V is a node set, each node corresponds to an extracted feature, A is a weighted adjacency matrix, and is constructed based on statistical correlation among features; Wherein the node set , A feature vector representing the i-th feature node, , The original value of the ith feature in the current time window; the mean value of the ith feature in the training set; standard deviation of the ith feature in the training set; The weighted adjacency matrix A is based on the Pearson correlation coefficient among the features Constructing, generating weighted adjacency matrix by thresholding strategy : In the formula, Is a correlation threshold.
6. The single-channel EEG sleep staging method based on feature map neural network of claim 1, wherein the map convolution network adopts a three-layer map convolution network, and the single-layer update formula is: In the formula, Representing a first layer node characteristic matrix; A learnable weight matrix representing a first layer; adding a self-loop to the adjacency matrix A ; A degree matrix; is a nonlinear activation function; The input and output dimensions of the three-layer graph rolling network are respectively a first layer of input Kx1 and output Kx64, a second layer of input Kx64 and output Kx128, and a third layer of input Kx128 and output Kx64.
7. The single channel EEG sleep staging method based on feature map neural network according to claim 1, wherein the global pooling is combined with mean pooling and maximum pooling as follows: In the formula, The global average information is reserved for calculating the node characteristic average value; reserving local significant information for calculating the node characteristic maximum value; The average value and the maximum value are spliced to form a fixed length vector, and the fixed length vector is input into a classifier.
8. The single channel EEG sleep staging method based on feature map neural network according to claim 7, wherein said classification calculation specifically comprises: The graph vectors are input into the Dropout layer and the full connection layer, and the category probabilities are output through a softmax function: In the formula, Is a learnable weight; Is biased; a predicted class distribution for the samples.
9. The single channel EEG sleep staging method based on feature map neural network of claim 1, wherein the feature map is modeled using a map convolution network, model training using cross entropy loss in combination with L2 regularization as a loss function: wherein M is the number of samples in batch, C is the number of categories; one-hot encoding for real tags; Is a predictive probability; a set of all learnable parameters; Regularizing the coefficients for L2.
10. The single channel EEG sleep staging method based on feature map neural network according to claim 1, wherein the sleep stage results comprise W-, N1-, N2-, N3-and REM-stages.

Description

Single-channel EEG sleep stage method based on feature map neural network Technical Field The invention relates to the technical field of sleep stage, in particular to a single-channel EEG sleep stage method based on a feature map neural network. Background Electroencephalogram (Electroencephalography, EEG) is a non-invasive neural signal acquisition technique that records the weak potential changes produced by simultaneous discharge of a cerebral cortical neuron population by arranging electrodes on the surface of the scalp. Due to the advantages of millisecond time resolution, safe and noninvasive acquisition process, relatively low equipment cost and the like, EEG has become one of the most commonly used signal forms in neuroscience research and clinical neurological disease auxiliary diagnosis. Currently, EEG has been widely used in a number of important areas such as seizure detection, sleep staging, assessment of consciousness disturbance, emotion recognition, brain-computer interfaces, and the like. However, the EEG signal itself has significant complexity. On the one hand, EEG is a typical non-stationary signal, the statistical properties of which change with time, and on the other hand, EEG reflects the superposition of a large number of neuronal population activities, with obvious nonlinearities and randomness. Furthermore, EEG is extremely susceptible to interference from non-brain-derived signals such as myoelectricity, blinking, electrocardiograms, and environmental noise, so that signal-to-noise ratio is low. Together, these factors lead to automated analysis and high-precision classification of EEG signals, still facing significant challenges. Traditional EEG classification methods typically employ a two-stage framework of "feature engineering+machine learning classifiers". Researchers first manually extract time-domain, frequency-domain, or non-linear features from the EEG signal and then model using classifiers such as support vector machines (Support Vector Machine, SVM), random Forest (Random Forest), or k-nearest neighbors. The method achieves certain effect in a specific application scene, but the core assumption is that all the features are mutually independent, and the features are input as a model only by a simple splicing mode, so that potential statistical dependency relationships and high-order interaction structures between the features are difficult to be described, and the expression capacity and generalization performance of the model are limited. In recent years, deep learning methods have received a great deal of attention in the field of EEG signal analysis. Convolutional Neural Networks (CNNs) are capable of automatically learning the local pattern of the timing signal by local convolution kernels, and cyclic neural networks (RNNs) and variants thereof, such as LSTM, have certain advantages in modeling long-term dependencies. However, for single channel EEG data, constructing a depth model directly on the original time series often requires a large number of training samples, while model parameters are large in scale, and over-fitting problems are prone to occur. Furthermore, such models often lack explicit interpretability, and it is difficult to understand the basis of discrimination of the model from a physiological or statistical perspective. Disclosure of Invention The invention aims to provide a single-channel EEG sleep stage method based on a feature map neural network, which comprises the steps of firstly preprocessing and segmenting the single-channel EEG signal, extracting multiple types of features from each time window, then constructing different features as nodes in a map, constructing a weighted adjacent matrix based on statistical correlation among the features, and finally modeling the feature map by using a map convolution network to realize high-precision classification of the EEG signal so as to solve the problems in the prior art. In order to achieve the above purpose, the invention provides a single-channel EEG sleep stage method based on a feature map neural network, which comprises the following steps: sequentially carrying out band-pass filtering, artifact removal processing and signal segmentation on an original single-channel EEG signal to obtain an EEG sample; respectively extracting time domain features, frequency domain features and nonlinear dynamics features from each EEG sample; Constructing the extracted time domain features, frequency domain features and nonlinear dynamics features into feature graphs; Modeling the feature map by adopting a map convolution network, and outputting a sleep stage result through map convolution layer feature updating, global pooling and classification calculation. As a preferred embodiment, the de-artefact process employs an amplitude threshold based strategy, for each EEG time window x, if its absolute amplitude is maximumExceeding a preset thresholdJudging the time window as a signal segment containing arti