CN-121971042-A - Single-channel electroencephalogram sleep analysis method and related equipment

CN121971042ACN 121971042 ACN121971042 ACN 121971042ACN-121971042-A

Abstract

The embodiment of the application provides a single-channel electroencephalogram sleep analysis method and related equipment, and belongs to the technical field of electroencephalogram signal processing and sleep health monitoring. The method comprises the steps of preprocessing and time sequence enhancement of single-channel electroencephalogram signals, extracting local and global features by utilizing a double-branch feature encoder, realizing feature distribution alignment between a source domain and a target domain by combining cross-view time sequence contrast learning and condition contrast alignment through a double-contrast learning module, dynamically adjusting sample weights of the target domain based on information gain through a learnable re-weighting module so as to correct label distribution offset, and finally realizing high-precision sleep stage or OSAHS event detection by utilizing a classifier. The method and the device remarkably improve generalization capability and robustness of the model in a cross-subject scene, can realize accurate analysis by only needing single-channel electroencephalogram signals, reduce hardware cost and use threshold, and are suitable for scenes such as home sleep monitoring, remote medical screening and the like.

Inventors

CHEN JUNLONG
YIN YUXIN
CHEN BIANNA
XU ZIHUA
ZHANG TONG

Assignees

华南理工大学

Dates

Publication Date: 20260505
Application Date: 20260126

Claims (10)

1. A single-channel electroencephalogram sleep analysis method, characterized in that the method comprises the following steps: Acquiring single-channel electroencephalogram signals of a source domain and a target domain, and preprocessing the electroencephalogram signals to obtain a source domain sample and a target domain sample; Respectively carrying out time sequence enhancement on the source domain sample and the target domain sample to generate corresponding enhancement views to form four paths of data streams, wherein the four paths of data streams comprise a source domain original stream, a source domain enhancement stream, a target domain original stream and a target domain enhancement stream; inputting the four paths of data streams into a shared double-branch feature encoder, and extracting local time sequence features and fusion features; Performing cross-view time sequence comparison learning based on the local time sequence features, and performing condition countermeasure alignment based on the fusion features to distribute the features of Ji Yuanyu and the target domain; Dynamically generating sample weights according to the prediction confidence and the information gain of the target domain samples so as to correct label distribution offset between the source domain and the target domain; training a classifier based on the fusion characteristics and the sample weights to obtain a trained sleep analysis model; and performing sleep staging or OSAHS event detection on the target domain electroencephalogram signals by using the trained sleep analysis model.
2. The method of claim 1, wherein the preprocessing comprises: resampling the single-channel electroencephalogram signals to a uniform frequency; performing band-pass filtering to preserve sleep related frequency bands; independent normalization of the signal for each subject; dividing the signal into segments with fixed time length, and labeling according to the task.
3. The method of claim 1, wherein the timing enhancement employs a fourier transform proxy strategy to generate an enhanced view of phase randomization and power spectrum preservation.
4. The method of claim 1, wherein the dual branch feature encoder comprises: A local detail branch for extracting high resolution local features; global context branches for extracting long timing dependent features; and the attention fusion module is used for dynamically fusing the outputs of the local detail branches and the global context branches to generate the fusion characteristics.
5. The method of claim 1, wherein the cross-view timing contrast learning comprises: Generating a context vector based on the autoregressive model; Performing a cross prediction task to predict future features of one enhancement view using the context of another enhancement view; A contrast constraint is applied to the context vector, pulling the distance between different enhancement views of the same sample.
6. The method of claim 1, wherein the conditional challenge alignment comprises: splicing the fusion characteristics and the classifier prediction results into a joint representation; Distinguishing the feature sources through a domain discriminator, and enabling the encoder to generate domain invariant features through countermeasure training; The sample weights are introduced in the countering loss, and the high value target domain samples are aligned in a weighted manner.
7. The method of claim 1, wherein dynamically generating sample weights comprises: Weight adjustment is carried out based on the relative entropy between the pseudo tag distribution of the target domain and the real tag distribution of the source domain; Calculating a prediction entropy value based on a Monte Carlo random inactivation technology, and taking the prediction entropy value as an information gain basis; A confidence regularization term is introduced to prevent over-weighting low confidence samples.
8. An electronic device comprising a memory storing a computer program and a processor implementing the method of any of claims 1 to 7 when the computer program is executed by the processor.
9. A computer readable storage medium storing a computer program, characterized in that the computer program, when executed by a processor, implements the method of any one of claims 1 to 7.
10. A computer program product comprising a computer program, characterized in that the computer program, when executed by a processor, implements the method of any one of claims 1 to 7.

Description

Single-channel electroencephalogram sleep analysis method and related equipment Technical Field The application relates to the technical field of electroencephalogram signal processing and sleep health monitoring, in particular to a single-channel electroencephalogram sleep analysis method and related equipment. Background High quality sleep is the basis for maintaining physiological and psychological health of the human body. Clinically, sleep health assessment mainly relies on Polysomnography (PSG), which enables sleep staging and respiratory event detection through overnight recording and analysis of multimodal physiological signals such as electroencephalograms, electrooculography, electromyography, and the like. However, PSG analysis is highly dependent on manual interpretation by a professional technician, and the process is time consuming, expensive and poorly scalable. In recent years, an automated analysis method based on deep learning is gradually rising. Because electroencephalogram signals are highly sensitive to sleep states and respiratory events, models based on single-channel electroencephalograms are of interest. However, the prior art faces two major core challenges in practical cross-subject applications: 1) The characteristic distribution is offset, namely, the electroencephalogram signals of different individuals have obvious statistical distribution differences due to physiological and pathological differences, so that the model generalization capability is weak. 2) Label distribution bias-the sleep stage ratio or the occurrence frequency of respiratory events (i.e. label distribution) of different individuals are greatly different, and the existing method usually ignores the difference, so that a model generates prediction deviation for target individuals with different clinical characteristics. The existing solution mostly adopts an unsupervised domain self-adaptive method to align feature distribution, but semantic discriminant is difficult to maintain, and deviation of label distribution is not considered generally, so that the problem of negative migration or model bias is caused. In addition, the existing method mostly depends on single-view or complex multi-mode input, and cannot fully mine rich time sequence dynamic information in single-channel electroencephalogram signals. Therefore, a unified framework capable of simultaneously solving the characteristic and label distribution offset and realizing high generalization sleep analysis based on single-channel electroencephalogram is needed. Disclosure of Invention The embodiment of the application mainly aims to provide a single-channel electroencephalogram sleep analysis method, electronic equipment, a storage medium and a program product, so as to realize high-precision, high-generalization and low-cost sleep monitoring. In order to achieve the above object, an aspect of an embodiment of the present application provides a single-channel electroencephalogram sleep analysis method, including: Acquiring single-channel electroencephalogram signals of a source domain and a target domain, and preprocessing the electroencephalogram signals to obtain a source domain sample and a target domain sample; Respectively carrying out time sequence enhancement on the source domain sample and the target domain sample to generate corresponding enhancement views to form four paths of data streams, wherein the four paths of data streams comprise a source domain original stream, a source domain enhancement stream, a target domain original stream and a target domain enhancement stream; inputting the four paths of data streams into a shared double-branch feature encoder, and extracting local time sequence features and fusion features; Performing cross-view time sequence comparison learning based on the local time sequence features, and performing condition countermeasure alignment based on the fusion features to distribute the features of Ji Yuanyu and the target domain; Dynamically generating sample weights according to the prediction confidence and the information gain of the target domain samples so as to correct label distribution offset between the source domain and the target domain; training a classifier based on the fusion characteristics and the sample weights to obtain a trained sleep analysis model; and carrying out sleep stage or OSAHS event detection on the brain electrical signal of the target domain by using the trained sleep analysis model. In some embodiments, the preprocessing comprises: resampling the single-channel electroencephalogram signals to a uniform frequency; performing band-pass filtering to preserve sleep related frequency bands; independent normalization of the signal for each subject; dividing the signal into segments with fixed time length, and labeling according to the task. In some embodiments, the timing enhancement employs a fourier transform proxy strategy, generating an enhanced view of phase randomization and p