Depression screening system based on sleep physiological signals

Abstract

The invention discloses a depression screening system based on sleep physiological signals, and the system comprises a data collection module which is used for collecting sleep electroencephalogram data of a user in real time through a brain-computer interface device; the data processing module is used for preprocessing the acquired electroencephalogram data; the method has the advantages that the sleep staging technology is combined, the CNN-BiLSTM mixed architecture algorithm is adopted to process and analyze sleep electroencephalogram data, the algorithm integrates the advantages of CNN and BiLSTM, and high-precision staging of sleep stages can be achieved; a DepNet2D model is introduced, by constructing a two-dimensional network for spectrum information and deeply analyzing association and influence among different frequency bands, characteristic electroencephalogram spectrum information of a depression patient is accurately extracted, the depression diagnosis method based on sleep staging not only improves the diagnosis accuracy, but also enhances the objectivity and persuasion, and the diagnosis accuracy is improved. Powerful support is provided for precise treatment of depression.

Description

Technical Field

[0001] The invention belongs to the field of electronic information technology, and in particular relates to a depression screening system based on sleep physiological signals. Background Art

[0002] In recent years, with the rapid development of deep learning technology, its application in the medical field has become increasingly widespread, especially in the sleep staging and depression diagnosis based on EEG signals, deep learning technology has shown great potential; in terms of the current research status of depression classification based on EEG signals, researchers have made significant progress in the field of depression recognition by carefully designing deep learning models; Acharya et al. used the CNN model to classify the EEG data of patients with depression and normal people, and successfully verified the effectiveness of deep learning in depression diagnosis. However, the study was limited by a small sample size and the generalization ability of the model is still questionable; Li et al. went a step further on this basis. They combined CNN with transfer learning technology and deeply explored the important role of different levels of information in EEG signals in depression recognition, providing a reference for EEG classification. It provides a more solid theoretical basis for identifying depression, but the classification performance of the model is not very good; Kwon et al. proposed an innovative deep learning model based on asymmetric images of prefrontal EEG to achieve preliminary screening of patients with depression, but due to the small amount of data, the study was insufficient in evaluating the generalization ability of the model and can only provide a preliminary reference; the combined model with EEG signal emotion recognition function developed by Lu Guanming et al. (2021) showed advantages in feature extraction and processing of dynamic time series information, with an accuracy rate of 86.08%, but their research pointed out that a larger sample size is needed to support model training; Li Mei et al. (2021) achieved a high accuracy rate of 98.00% by combining an algorithm-based convolutional neural network (CNN) with wireless sensors, but the disadvantage is that its cost is too expensive; Joao Carreira et al. (2018) adopted a 3D convolutional network with effective connections in major depressive disorder (MDD). Although they proposed new features such as PDC and DMN connections, the accuracy of their implementation lacked more evaluation indicators, with an accuracy of 80.90%; Qayyum et al. (2022) demonstrated a method combining convolutional neural networks with GRU, which showed superiority over LSTM in short sequence learning, with an accuracy of 90.45%, but this method still had some shortcomings in processing the length of time window segments; Narazaki et al. (2019)'s research was based on a 22-layer ResNet convolutional neural network, which was convenient for developing classifiers and using algorithms to assign weights, but its accuracy was only 78.94%; In summary, although the above studies have made progress in the application of deep learning in depression diagnosis, they generally face problems such as insufficient sample size, high cost, and the need to improve model accuracy and generalization ability.

[0003] In order to further accurately diagnose depression, many scholars have analyzed the relationship between many EEG signals and depression. Among them, sleep signals, as a stable EEG signal, have attracted much attention. Researchers have deeply explored the relationship between sleep EEG signals and depression and found that they are closely related. The study of SteigerA et al. showed that long-term use of antidepressants will have a significant impact on the characteristics of sleep EEG signals in patients with depression. This finding provides a solid theoretical basis for the diagnosis of depression based on sleep EEG signals. Armitage R et al. focused on young female patients with depression and found that their Delta wave sleep EEG signals were significantly different from those of normal people. This finding provides valuable clues for exploring the sleep EEG specificity of depression. Leistedt S et al. calculated the Scaling parameters of different sleep stages and revealed that the parameter values ​​of patients with depression in specific sleep stages were abnormal, which strongly confirmed that sleep is closely related to depression and can be analyzed numerically. Pillai V et al. (2011) explored the genetic characteristics of depression by analyzing the sleep EEG information of patients with depression. The experimental results showed that patients with severe depression showed more intensive rapid eye movement (REM) sleep and shorter slow wave sleep (SWS). This combination of characteristics is also considered to be the genetic characteristics of major depression.

[0004] In summary, the diagnosis method of depression based on EEG signals has made certain research progress; however, there are still many problems in this field, such as low accuracy and weak interpretability of the method. Summary of the invention

[0005] The purpose of the present invention is to provide a depression screening system based on sleep physiological signals to improve the accuracy and objectivity of diagnosis.

[0006] To achieve the above-mentioned purpose, the present invention provides the following technical solution: a depression screening system based on sleep physiological signals, comprising:

[0007] A data acquisition module, wherein the data acquisition module is used to collect the user's sleep EEG data in real time through a brain-computer interface device;

[0008] A data processing module, which pre-processes the collected EEG data, including applying a zero-phase filter to limit the frequency range and using independent component analysis to separate independent components in the mixed signal to reduce noise and artifacts;

[0009] A sleep staging module, which includes a CNN-BiLSTM-based model. The model first converts the preprocessed EEG data into time-frequency signals through short-time Fourier transform and constructs a feature matrix. Subsequently, CNN is used to extract features in the time-frequency domain, and BiLSTM is used to explore the dependencies and associations between EEG signals from the time domain dimension to achieve automatic staging of sleep stages.

[0010] A depression screening module, which is based on the DepNet2D model, processes REM sleep EEG data, extracts spectrum features through a two-dimensional convolutional neural network, then learns high-level features through a fully connected layer, and uses a Softmax activation function for binary classification to screen whether the user suffers from depression;

[0011] A user interaction module is used to receive user instructions and display a sleep report.

[0012] As a preferred technical solution of the present invention, user instructions are received to display a sleep report, including sleep staging results and a depression screening report. The depression screening report includes a radar chart of the degree of depression symptoms based on the Becker-Rafanson Mania Scale and a pie chart of the tendency of depressive symptoms to occur.

[0013] As a preferred technical solution of the present invention, in the CNN-BiLSTM model, the CNN part performs feature extraction through convolution operations, batch normalization, activation functions, maximum pooling and dropout layers, and the BiLSTM part processes time series data through LSTM networks in two directions to capture long-term time dependencies and correlations.

[0014] As a preferred technical solution of the present invention, the DepNet2D model includes three convolutional layers, each of which is followed by a ReLU activation function and a maximum pooling layer for extracting and retaining significant features, and performing high-level feature learning and classification judgment through a fully connected layer.

[0015] As a preferred technical solution of the present invention, the following steps are taken when performing data preprocessing:

[0016] Divide data samples: take frames as the basic unit of judgment, and divide 30 seconds of continuous sleep EEG data into one sleep data sample;

[0017] Set sleep stage labels: According to the AASM sleep stage standard, the sleep cycle is divided into: awake period, N1 stage, N2 stage, N3 stage and REM stage. Set the corresponding sleep stage labels to "0, 1, 1, 1, 2", corresponding to the three stages of awake period, deep sleep period and REM period;

[0018] Filtering of sleep EEG data: Use Butterworth bandpass filters with cutoff frequencies of 0.1 Hz and 50 Hz to filter the EEG signals of two channels, and form a group for every 5 segments of filtered data samples and discriminant labels to construct a feature matrix.

[0019] As a preferred technical solution of the present invention, it also includes a CNN-BiLSTM network structure, which consists of four layers: a convolution layer, a pooling layer, a LSTM-based BiLSTM hidden layer and a multi-task softmax layer. This architecture has four main features: the convolution layer simultaneously accommodates 3 convolution kernels of different sizes; the pooling strategy of the pooling layer is more suitable for capturing the shift invariance of time signals; LSTM units have been proven to extract long-term time dependencies used in many sequence processing applications. In the BiLSTM module, the accuracy of sleep staging can be improved by LSTM units, and the extracted generalized sequential representation of the EEG frame pattern obtained from the first two layers is processed together with the bidirectional LSTM network; the multi-task softmax layer used is adapted to joint classification and prediction, so that the feature extraction of sleep EEG data and the training of the model can have more reliable results, and the probability of various sleep staging results is finally output through the Softmax layer classifier.

[0020] As a preferred technical solution of the present invention, the DepNet2D model fully utilizes the spectrum information of REM sleep EEG data to construct a two-dimensional network model of the spectrum information and analyze the correlation and mutual influence between different frequency bands.

[0021] As a preferred technical solution of the present invention, it also includes constructing a DepNet2D feature extraction model, and the corresponding method is as follows:

[0022] The spectrum information of the original sleep EEG of the REM period is extracted, and two channels are taken. The spectrum information is input into the model, and multiple convolutional layers are stacked for feature extraction. After passing through the convolutional layer, the ReLU activation function and the maximum pooling layer are respectively passed through, and the size of the feature map is reduced by downsampling operation, and the most significant features are retained; then batch normalization is performed; the Dropout layer is connected to each pooling layer, and then the Flatten layer is connected to flatten the features extracted by the convolutional layer and the pooling layer into a one-dimensional vector, and finally passed through two fully connected layers.

[0023] As a preferred technical solution of the present invention, multiple convolutional layers are stacked for feature extraction, and the corresponding contents are as follows: three two-dimensional convolutions are used, the number of convolution kernels of the first convolution layer is 16, the number of convolution kernels of the second convolution layer is 36, and the number of convolution kernels of the third convolution layer is 48, and the convolution kernel sizes of the three convolution layers are all (3,3).

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] The present invention combines the sleep staging technology and uses the CNN-BiLSTM hybrid architecture algorithm to process and analyze sleep EEG data. The algorithm combines the advantages of CNN and BiLSTM and can achieve high-precision division of sleep stages. The DepNet2D model is introduced to build a two-dimensional network for spectrum information, deeply analyze the correlation and influence between different frequency bands, and accurately extract the characteristic EEG spectrum information of patients with depression. The depression diagnosis method based on sleep staging not only improves the diagnostic accuracy, but also enhances its objectivity and persuasiveness, providing strong support for the precise treatment of depression.

[0026] The present invention adopts a CNN-BiLSTM hybrid architecture algorithm, which can achieve high-precision division of sleep stages, and focuses on the electroencephalogram (EEG) signals of the rapid eye movement (REM) sleep period, and fully studies the pathophysiological mechanism of depression; depression is diagnosed by analyzing sleep EEG data; sleep EEG data is objective and is not affected by the patient's subjective consciousness, so it can provide a more objective basis for diagnosis; at the same time, the present invention introduces the DepNet2D model, constructs a two-dimensional network for spectrum information, further analyzes the correlation and influence between different frequency bands, learns and identifies patterns from a large amount of data, thereby improving the accuracy of diagnosis; further improves the accuracy and objectivity of diagnosis, by using sleep EEG data for diagnosis, combined with the CNN-BiLSTM network and the DepNet2D model, the present invention can greatly reduce the subjective factors in the diagnosis process, and provide a more objective and accurate basis for diagnosis; in this way, the influence of doctors' subjective judgment and patients' self-report can be avoided or reduced, thereby improving the accuracy and objectivity of depression diagnosis;

[0027] By utilizing sleep EEG data, the present invention combines the CNN-BiLSTM network and the DepNet2D model to provide a more objective and accurate basis for diagnosis, reduce the influence of subjective factors, and make more accurate results for depression symptoms at an earlier stage, thereby achieving early intervention; early intervention can help patients better understand and cope with their emotional problems, take effective treatment measures, and avoid further aggravation of symptoms or other complications; the depression diagnosis system of the present invention utilizes sleep EEG data for early screening and diagnosis, and has high objectivity and accuracy; by discovering the depression symptoms of patients at an early stage and conducting intervention treatment, the effect and prognosis of treatment can be effectively improved, the pain and burden of patients can be alleviated, and their quality of life can be improved;

[0028] By analyzing the patient's sleep EEG data, we can learn about the patient's sleep quality, sleep structure, characteristics of brain activity during sleep, and other information, so as to better understand the individual differences of patients; the severity of depression varies from person to person, some patients may only have mild depression, while others may have reached the level of severe depression; through the analysis of sleep EEG data, we can objectively evaluate the severity of the patient's depression and provide an important reference for formulating personalized treatment plans;

[0029] The present invention is based on the analysis results of the patient's REM sleep EEG data and the spectrum information of the sleep EEG. It has the advantages of higher accuracy and screening rate, and is more scientific and referenceable. Doctors can formulate more personalized and targeted treatment plans; through the treatment plan formulated according to the patient's specific situation, the patient's physiological and psychological state can be better adjusted, the symptoms of depression can be alleviated, and the treatment effect and prognosis can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a technical flow chart of the present invention;

[0031] Figure 2 This is a schematic diagram of cutting sleep EEG data through a sliding time window of the present invention;

[0032] Figure 3 Constructing a graph for the model of the present invention;

[0033] Figure 4 This is a system module diagram of the depression screening system of the present invention. DETAILED DESCRIPTION

[0034] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0035] Example 1

[0036] See also Figure 1-Figure 4 , which is the first embodiment of the present invention, provides a depression screening system based on sleep physiological signals, which has the characteristics of high precision, high efficiency and high user-friendliness. Facing the diagnosis difficulties of 322 million depression patients worldwide, the system can not only realize the identification of depression by analyzing sleep EEG signals, but also monitor and optimize the sleep quality of patients in real time, aiming to improve the diagnostic accuracy to realize the screening function of depression;

[0037] Due to the non-invasiveness and high temporal resolution of sleep EEG signals, we use the DepNet2D (Depression Net 2Dimension) model for sleep staging, focusing on the feature extraction and classification of REM EEG data; DepNet2D learns the spatiotemporal dependence of sleep EEG through a two-dimensional convolutional neural network, effectively captures the characteristic patterns of brain activity, and improves the accuracy of identifying the spectral characteristics of depression; the model has a strong generalization ability due to the reduction in the number of parameters, and the accuracy rate in the depression screening task reached 88.82%, showing higher accuracy and objectivity compared with traditional methods.

[0038] The present invention will use different public data sets to test the generalization performance of the model, continuously adjust parameters and optimize, and integrate the front-end and back-end codes; finally, the EEG acquisition device will be connected to the computer to build an integrated depression diagnosis system that integrates signal acquisition, processing and classification, result output, and music regulation;

[0039] Sleep staging based on CNN-BiLSTM

[0040] The present invention uses sleep EEG signals as the basis for screening patients with depression. Sleep disorders are typical symptoms of many patients with depression. Among them, the shortened latency, increased duration and increased REM density of REM (rapid eye movement) sleep are considered to be biological markers of depression. It is worth noting that changes in REM sleep may appear before the clinical manifestation of depression. Therefore, the analysis of REM sleep data is helpful for early identification and screening of patients with depression. Therefore, this application selects REM sleep EEG data as an indicator for depression screening.

[0041] Preprocessing of sleep EEG data

[0042] The sleep EEG data used in this application comes from the Sleep-EDF public dataset, which is a well-recognized resource and is widely used in the field of sleep research. It contains data from two different groups: one group records the sleep EEG signals of normal people, and the other group records the sleep EEG signals of patients with depression. These data are obtained by placing electrodes during sleep to record brain activity. When performing data preprocessing, this application takes the following steps:

[0043] Divide data samples: take frame (epoch) as the basic unit of judgment, and divide 30 seconds of continuous sleep EEG data into one sleep data sample;

[0044] Set sleep stage labels: According to the AASM sleep stage standard proposed by the American Academy of Sleep Medicine (AASM), the sleep cycle is divided into: awake period, N1 period, N2 period, N3 period and REM period. Set the corresponding sleep stage labels to "0, 1, 1, 1, 2", corresponding to the three stages of awake period, deep sleep period and REM period;

[0045] Filtering of sleep EEG data: To reduce the impact of artifacts, the EEG signals of two channels were filtered using a Butterworth bandpass filter with a cutoff frequency of 0.1 Hz and 50 Hz, and every five filtered data samples and discriminant labels formed a group to construct a feature matrix;

[0046] Sleep staging based on CNN-BiLSTM network model

[0047] The CNN-BiLSTM network model is an architecture that combines a convolutional neural network (CNN) and a bidirectional long short-term memory neural network (BiLSTM) for automatic sleep staging. This combination can effectively extract the features of sleep EEG data and learn the timing information therein, thereby achieving accurate classification of sleep stages. CNN can be well applied to feature extraction of EEG signals through its effective processing capabilities for non-stationary and nonlinear data. In order to better explore the temporal changes between different EEG modalities, the features from different modalities are jointly optimized with the BiLSTM layer to better learn the laws of sleep stage transitions and output accurate sleep staging results. This combination of CNN-BiLSTM network models can not only effectively extract the features of sleep data, but also learn the timing information therein, thereby achieving more accurate sleep staging classification.

[0048] The CNN-BiLSTM network structure used in the present invention consists of four layers: a convolutional layer, a pooling layer, a BiLSTM hidden layer based on LSTM and a multi-task softmax layer. This architecture has four main features: first, its convolutional layer simultaneously accommodates three convolution kernels of different sizes, so it can learn the features of sleep EEG data at different resolutions and reduce the complexity of reconstructing data in the data feature extraction process, which is conducive to building an automatic sleep staging model; second, compared with the common subsampling pooling, the pooling strategy of the pooling layer is more suitable for capturing the shift invariance of time signals, because the specific features of sleep EEG signals may occur at any time position; then, considering the correlation and connection before and after the sleep cycle, the LSTM unit has been proven to extract the long-term time dependence used in many sequence processing applications. To this end, in the BiLSTM module, the present application can improve the accuracy of sleep staging through LSTM units, and the extracted generalized sequential representation of the EEG frame patterns obtained from the first two layers is processed together with the bidirectional LSTM network, and two such bidirectional LSTM layers are stacked together, and the generalized variational features of different EEG patterns are further processed by these LSTM layers, which extract higher-level effective features while considering long-term inter-modal temporal dependencies, thereby improving the accuracy of sleep staging; finally, in contrast to the standard softmax, the multi-task softmax layer used in the present application is adapted to joint classification and prediction, so that the feature extraction of sleep EEG data and the training of the model can have more reliable results, and the probability of each type of sleep staging result is finally output through the Softmax layer classifier.

[0049]

[0050] As shown in formula (1), for the input vector Z = {Z1, Z2, ..., Z K}, where z i represents the activation value at the i-th moment; the probability of each type of sleep stage result is output through the Softmax classifier;

[0051] As shown in formula (1), the probability of each sleep stage result is output by the Softmax classifier; where Z is the input vector, which contains the activation value of each category; i represents the activation value of the i-th element in the input vector Z; e is the base of the natural logarithm; k is the length of the input vector Z, that is, the total number of categories, Represents all elements z in the input vector Z i The exponential sum of ;

[0052]

[0053] At the same time, in order to solve the problem of imbalanced sleep data categories, this model uses the sum of the cross entropy errors of each subtask as the classification and prediction error of training, as shown in formula (2), where θ and denote the network parameters and probability distribution of CNN output respectively; the network is trained to minimize the multi-task cross entropy error;

[0054]

[0055] As shown in formula (3), log(y i ) represents the logarithmic probability of the network’s prediction result for sample i, It is category c i By adjusting the weight value, we can pay different attention to samples of different categories to solve the problem of category imbalance;

[0056] As shown in formula (3), by adjusting the weight value, different degrees of attention can be paid to samples of different categories, thereby solving the problem of category imbalance; where L represents the value of the loss function; N represents the total number of samples; Indicates category c i The weight of log(y i ) represents the logarithmic probability of the network’s prediction result for sample i; y i represents the predicted probability of sample i.

[0057] Depression screening based on DepNet2D

[0058] Clinically, EEG spectrum analysis can be used to diagnose and monitor various diseases, which is conducive to objective analysis of various diseases and improves interpretability and accuracy. Patients with depression may show specific spectral characteristics in EEG signals, and may show abnormal EEG activity patterns in specific brain areas or specific frequency bands. In this regard, the present application proposes a DepNet2D model to realize the function of depression screening. The DepNet2D model makes full use of the spectral information of REM sleep EEG data, constructs a two-dimensional network model of spectral information, analyzes the correlation and mutual influence between different frequency bands, and compares the EEG spectrum information of normal people. It can effectively extract the EEG spectrum characteristics of patients with depression to realize the screening function for depression.

[0059] Preprocessing of sleep EEG data

[0060] First, we took out the original sleep EEG data of the REM period for data analysis and data conversion, and then used a zero-phase filter (FIR) to limit the frequency to 0.5Hz to 45Hz, which can help reduce unnecessary noise and will not cause phase distortion. After that, we used independent component analysis (ICA) to separate the mixed signals into independent components, effectively extracting potential sources of EEG activity, and reduce the impact of artifacts. It can improve the quality and accuracy of the signal, making the neural activity patterns of different brain regions during sleep more significant, and then cropped the filtered REM-SLEEP-EEG with a 30s time window and plotted it as a time domain diagram, and then further extracted the spectral information of the EEG signal.

[0061] Building a DepNet2D feature extraction model

[0062] First, extract the original REM sleep EEG spectrum information, take two channels (EEG Fpz-Cz and EEG Pz-Oz), and input the spectrum information into the model. The size of the input layer is (150,150,3). We use stacked multiple convolutional layers for feature extraction, so that we can build a model with more powerful representation capabilities. This application uses three two-dimensional convolutions (Conv2D). The number of convolution kernels in the first convolution layer is 16, the number of convolution kernels in the second convolution layer is 36, and the number of convolution kernels in the third convolution layer is 48. The convolution kernel sizes of the three convolution layers are all (3,3); after passing through the convolution layer, the ReLU activation function and the maximum pooling layer (MaxPooling2D) of size (2,2) are used respectively to reduce the size of the feature map through downsampling operations and retain the most significant features; then batch normalization is performed; in order to reduce the overfitting of the neural network and improve the performance of the model, Dr. The opout layer (ratio = 0.1) is connected to each pooling layer; then, the Flatten layer is connected to flatten the features extracted by the convolution layer and the pooling layer into a one-dimensional vector; finally, after two fully connected layers (Dense), through the combination of multiple fully connected layers, the model can learn higher-level feature representations and map these feature representations to the probability distribution of the output category; the first fully connected layer has 64 neurons and uses the ReLU activation function to learn the nonlinear feature representation of the input data, which is more expressive when learning feature representation; the second fully connected layer (Dense (2, activation = tf.nn.softmax)) has 2 neurons and uses the softmax activation function to convert the output of the neural network into a distribution representing the probability of each category;

[0063] The binary cross-entropy loss function is used as shown in formula (4). For binary classification, this loss function can measure the difference between the probability distribution of the model output and the true label, thereby guiding the optimization process of the model parameters. By using the Adam optimizer, we can dynamically adjust the learning rate according to the first-order moment estimate and second-order moment estimate of the gradient of each parameter. Compared with the traditional gradient descent algorithm, Adam has a faster convergence speed and better performance. Accuracy is used as an evaluation indicator, which can intuitively reflect the classification performance of the model.

[0064]

[0065] For N data, where t i is the true label with a value of 0 or 1, p i is the Softmax probability of the i-th category;

[0066] For N data, L represents the value of the loss function; t i is the true label of the i-th sample, which takes the value of 0 or 1; p i is the predicted probability that the i-th sample belongs to the positive class; log(p i ) is the logarithm of the predicted probability that the i-th sample belongs to the positive class; log(1-p i ) is the logarithm of the predicted probability that the i-th sample belongs to the negative class; N is the total number of samples; It means summing all N samples.

[0067] Example 2

[0068] See also Figure 1-Figure 4 , which is the second embodiment of the present invention, is based on the previous embodiment, except that:

[0069] Due to the noise problem of non-invasive EEG signals, in order to ensure the reliability of experimental research, the present invention needs to rely on the high-precision equipment of South China Brain Control Intelligent Technology Co., Ltd. The present invention uses the LT3 EEG cap;

[0070] The present invention uses the carefully selected SleepEDF public data set with practical credibility and the data collected in cooperation with the brain-computer interaction and hybrid intelligence team of South China Normal University. These sample data have practical credibility, which can ensure the reliability of the experimental equipment and the scientific nature of the collection process, and can obtain a more accurate depression screening model under the training of the CNN-BiLSTM network model and the DepNet model; in addition, the present invention improves the quality and reliability of subsequent processing by strictly screening the data, can achieve the diversity of the collected data and the rigor of the screening process, and ensure the generalization ability of the model and the credibility of the experimental results; the depression screening system of the present invention has completed the sleep staging module and the depression screening module;

[0071] Sleep staging module based on CNN-BiLSTM

[0072] Forming the feature matrix

[0073] For the EEG data obtained after preprocessing, this application uses the short-time Fourier transform method in time-frequency analysis to analyze the signal whose amplitude changes with time into a signal whose time changes with frequency; through the sliding window mechanism, the window size and step size are set, the window is allowed to slide on the time domain signal, and the Fourier transform of each window is calculated respectively, and the frequency signals corresponding to the formed different time windows are spliced ​​to form a time-frequency signal; then the corresponding Mel spectrum is obtained by calculation, and every 5 consecutive 30s EEG signal data are grouped into one group to form a feature matrix;

[0074] Feature extraction

[0075] ①Construction of CNN-BiLSTM model

[0076] This application uses a CNN-BiLSTM network model to extract features from the matrix, that is, the network model is composed of a convolutional neural network (CNN) and a bidirectional long short-term memory neural network (BiLSTM). Considering that EEG multimodal PSG (including EEG, EOG, EMG, etc.) sleep signal data has the characteristics of non-stationary and nonlinearity, traditional methods are difficult to capture long-term changes in sleep stages; and the CNN model is suitable for processing network structure data, and its core ideas are: local perception field, weight sharing and downsampling pooling layer.

[0077] Considering that sleep data contains a large amount of temporal information, the CNN network can only capture short-term spatiotemporal features, and is limited in capturing long-term temporal features. Therefore, this system inputs the vector obtained based on the CNN model into the bidirectional long short-term memory network BiLSTM based on bidirectional temporal structure modeling, and processes sleep information in the front and back directions through two completely independent long short-term memory networks LSTM, respectively extracting information features in different dimensions, so as to better explore the dependence and correlation of EEG signals before and after each time point.

[0078] ② Combine CNN convolutional neural network to extract features in time and frequency domain

[0079] After the feature matrix composed of the Mel spectrum is grouped and input into the model, the convolution operation layer performs local correlation and window sliding processing on the input data, and then performs batch normalization processing; then, the activation function is applied and the data is pooled at the maximum, and the dropout layer is connected to each pooling layer to make the data more robust; then the Flatten layer is connected to realize the transition of data from the convolution layer to the fully connected layer. Then, after the Dense one-dimensional fully connected layer, normalization and ReLU activation and the final data block splicing between different channels, the transition from the convolution layer to the long short-term memory neural network is realized;

[0080] ③ Connect the BiLSTM network model to explore the dependence and correlation of EEG signals from the time domain dimension

[0081] In the BiLSTM part, the model uses a bidirectional long short-term memory network to transform the data in ② and obtain a matrix of a completely new size as the input of the classifier;

[0082] This application uses a multi-classification logarithmic loss function, corresponding to the Softmax classifier, and optimizes the loss function through Adam to dynamically adjust the learning rate of each parameter; the data in ③ above is input into the Softmax layer for three classification tasks, and the corresponding sleep states (W, LS, SWS, REM) are obtained respectively;

[0083] Depression screening module based on DepNet2D

[0084] Preprocessing of sleep EEG data

[0085] The application of the DepNet2D model in depression screening first relies on comprehensive preprocessing of REM sleep EEG data; by extracting data from the original EEG signal and applying a zero-phase filter (FIR) to limit the frequency range (0.5Hz to 45Hz), this method effectively reduces noise and avoids phase distortion; then, independent component analysis (ICA) is used to separate the independent components in the mixed signal, which enhances the clarity of the signal and reduces artifacts, thereby more accurately reflecting the neural activity patterns of different brain regions; the preprocessed data is cropped through a 30-second time window, plotted as a time domain diagram, and the spectrum information is extracted, providing cleaned feature data for subsequent model input;

[0086] Feature extraction for depression:

[0087] In the feature extraction stage, the DepNet2D model processes the extracted REM EEG spectrum information by constructing a two-dimensional convolutional neural network (CNN). The size of the spectrum information received by the input layer is (150,150,3). The model contains three convolutional layers with 16, 36 and 48 convolution kernels, and the size of the convolution kernel is (3,3); each convolutional layer is followed by a ReLU activation function and a maximum pooling layer of size (2,2) to extract and retain significant features, and batch normalization and Dropout layers are used to enhance the robustness of the model and prevent overfitting; the features extracted by the convolutional layer and the pooling layer are flattened into a one-dimensional vector, and then passed through two fully connected layers for high-level feature learning; the first fully connected layer contains 64 neurons and uses the ReLU activation function to learn the nonlinear features of the input data; the second fully connected layer uses the Softmax activation function to convert the network output into a distribution representing the probability of each category, which is used for the binary classification task, that is, to determine whether it is depression;

[0088] Model training and evaluation

[0089] During the model training process, the binary cross entropy function is used as the loss function to measure the difference between the probability distribution of the model output and the true label, and the learning rate is dynamically adjusted through the Adam optimizer to achieve faster convergence and better performance; the classification performance of the model is evaluated through accuracy, which directly reflects the effect of DepNet2D in depression screening. This whole process, combined with efficient data processing and deep learning methods, enables the DepNet2D model to effectively extract the EEG spectrum characteristics of patients with depression and provide accurate depression screening functions;

[0090] Depression Screening System

[0091] At present, the sleep staging module and depression screening module have been integrated with the front end to build a depression screening system based on sleep physiological signals. When the user wears the LT3 EEG cap as a brain-computer interface device and connects to the system back end to collect sleep EEG data, real-time monitoring of sleep EEG signals can be achieved.

[0092] When the user clicks the "End Sleeping" button, the system will generate a sleep report; when clicking the sleep report, the system will present the user's specific data on sleep, such as average monthly sleep duration, average monthly deep sleep duration, etc., and set sleep goals for the user; at the same time, the system will summarize the user's sleep records, sleep status this week, and sleep depth distribution in the form of a chart to help users intuitively understand their own sleep conditions.

[0093] In addition, after performing sleep staging through the CNN-BiLSTM network, the system extracts REM sleep EEG signals, and after performing EEG spectrum analysis through DepNet2D, the system can learn the user's sleep data characteristics and perform depression screening and classification. The system will generate a radar chart of depression symptom severity based on the Becker-Rafanson Mania Rating Scale (BRMS) in clinical medicine, and based on the algorithm, deduce the user's tendency to develop depressive symptoms, obtain a pie chart of depression tendency, and generate a depression report for the user.

[0094] Although the embodiments of the present invention have been shown and described, as detailed above, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A depression screening system based on sleep physiological signals, characterized by: include A data acquisition module, wherein the data acquisition module is used to collect the user's sleep EEG data in real time through a brain-computer interface device; A data processing module, which pre-processes the collected EEG data, including applying a zero-phase filter to limit the frequency range and using independent component analysis to separate independent components in the mixed signal to reduce noise and artifacts; A sleep staging module, wherein the sleep staging module includes a CNN-BiLSTM-based model, which first converts the preprocessed EEG data into a time-frequency signal through a short-time Fourier transform and constructs a feature matrix; Subsequently, CNN was used to extract features in the time-frequency domain, and BiLSTM was used to explore the dependencies and associations between EEG signals in the time domain to achieve automatic classification of sleep stages. A depression screening module, which is based on the DepNet2D model, processes REM sleep EEG data, extracts spectrum features through a two-dimensional convolutional neural network, then learns high-level features through a fully connected layer, and uses a Softmax activation function for binary classification to screen whether the user suffers from depression; A user interaction module is used to receive user instructions and display a sleep report.

2. The depression screening system based on sleep physiological signals according to claim 1, characterized in that: Receive user instructions and display a sleep report, including sleep staging results and a depression screening report. The depression screening report includes a radar chart of depression symptom severity based on the Becker-Rafanson Self-Rating Mania Scale and a pie chart of the tendency of depression symptoms to occur.

3. The depression screening system based on sleep physiological signals according to claim 1, characterized in that: In the CNN-BiLSTM model, the CNN part performs feature extraction through convolution operations, batch normalization, activation functions, maximum pooling and dropout layers, and the BiLSTM part processes time series data through LSTM networks in two directions to capture long-term time dependencies and correlations.

4. The depression screening system based on sleep physiological signals according to claim 1, characterized in that: The DepNet2D model consists of three convolutional layers, each of which is followed by a ReLU activation function and a maximum pooling layer to extract and retain significant features, and perform high-level feature learning and classification judgment through a fully connected layer.

5. The depression screening system based on sleep physiological signals according to claim 1, characterized in that: When performing data preprocessing, the following steps are taken: Divide data samples: take frames as the basic unit of judgment, and divide 30 seconds of continuous sleep EEG data into one sleep data sample; Set sleep stage labels: According to the AASM sleep stage standard, the sleep cycle is divided into: awake period, N1 stage, N2 stage, N3 stage and REM stage. Set the corresponding sleep stage labels to "0,1,1,1,2", corresponding to the three stages of awake period, deep sleep period and REM period; Filtering of sleep EEG data: Use Butterworth bandpass filters with cutoff frequencies of 0.1 Hz and 50 Hz to filter the EEG signals of two channels, and form a group for every 5 segments of filtered data samples and discriminant labels to construct a feature matrix.

6. The depression screening system based on sleep physiological signals according to claim 1, characterized in that: It also includes a CNN-BiLSTM network structure, which consists of four layers: a convolutional layer, a pooling layer, a LSTM-based BiLSTM hidden layer, and a multi-task softmax layer. This architecture has four main features: the convolutional layer simultaneously accommodates three convolution kernels of different sizes; the pooling strategy of the pooling layer is more suitable for capturing the shift invariance of temporal signals; LSTM units have been shown to extract long-term temporal dependencies used in many sequence processing applications. In the BiLSTM module, the accuracy of sleep staging can be improved through LSTM units, and the extracted generalized sequential representation of the EEG frame patterns obtained from the first two layers is processed together with the bidirectional LSTM network; the multi-task softmax layer used is adapted to joint classification and prediction, so that the feature extraction of sleep EEG data and the training of the model can have more reliable results. Through the Softmax layer classifier, the probability of each type of sleep staging result is finally output.

7. The depression screening system based on sleep physiological signals according to claim 1, characterized in that: The DepNet2D model makes full use of the spectrum information of REM sleep EEG data to build a two-dimensional network model of spectrum information and analyze the correlation and mutual influence between different frequency bands.

8. The depression screening system based on sleep physiological signals according to claim 7, characterized in that: It also includes building a DepNet2D feature extraction model. The corresponding method is as follows: Extract the original REM sleep EEG spectrum information, take two channels from it, input the spectrum information into the model, use multiple convolutional layers to extract features, and after the convolutional layer, use the ReLU activation function and the maximum pooling layer respectively, and use the downsampling operation to reduce the size of the feature map and retain the most significant features; Then batch normalization is performed; the Dropout layer is connected to each pooling layer, and then the Flatten layer is connected to flatten the features extracted by the convolution layer and the pooling layer into a one-dimensional vector, and finally passes through two fully connected layers.

9. The depression screening system based on sleep physiological signals according to claim 8, characterized in that: Use stacked multiple convolutional layers for feature extraction. The corresponding content is as follows: use three two-dimensional convolutions, the number of convolution kernels in the first convolution layer is 16, the number of convolution kernels in the second convolution layer is 36, and the number of convolution kernels in the third convolution layer is 48. The convolution kernel sizes of the three convolution layers are all (3,3).

Images