A training method for multi-concept discrimination network for EEG signal analysis

Abstract

The present invention is applicable to the field of deep learning technology, and provides a training method for a multi-concept discrimination network for electroencephalogram signal analysis, the method comprising: constructing a multi-concept discrimination network for analyzing electroencephalogram signals, wherein the multi-concept discrimination network comprises a first module constructed based on an encoder-decoder and a second module constructed based on a generative adversarial network and comprising a plurality of discriminators, and performing semi-supervised learning training on the multi-concept discrimination network according to an electroencephalogram signal training set until a target loss function of the multi-concept discrimination network converges, thereby not only reducing the network's dependence on a large amount of labeled data, but also improving the network's generalization ability and robustness in different patient groups, and at the same time, achieving the heterogeneity and complexity of electroencephalogram signals by focusing on multiple discriminative perspectives, improving the accuracy and reliability of electroencephalogram signal analysis, and facilitating its clinical application in diverse environments.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep learning, and in particular relates to a training method for a multi-concept discrimination network for electroencephalogram signal analysis. Background Art

[0002] Major Depressive Disorder (MDD) is a widespread, chronic mental illness that manifests in people of all ages, especially in adolescents, where the incidence of MDD is particularly significant, affecting approximately 13% of adolescents. Among those diagnosed with depression, nearly 50% are reported to engage in (Non-Suicidal Self-Injury, NSSI) behavior. NSSI behavior has attracted widespread attention. The increasing incidence of NSSI among adolescents with depression highlights the urgent need for effective tools and methods to understand, diagnose, and intervene in these behaviors. Early detection of NSSI behavior is essential for effective treatment of the disease and prevention of further serious harm. Automated detection methods not only help manage potential psychological problems, but also significantly reduce the risk of individuals engaging in more dangerous or life-threatening behaviors in the future.

[0003] As a method that directly reflects the neural dynamics of the central nervous system, electroencephalogram (EEG) has shown great value in detecting abnormal brain activity. Due to the non-invasiveness, high temporal resolution, and sensitivity to brain dynamics of EEG, it has been widely used in a variety of disease fields. For example, it has played an important role in epilepsy research, Alzheimer's Disease (AD) research, diagnosis and monitoring of depression, analysis of autism spectrum disorder (ASD), and schizophrenia research. These applications fully demonstrate the diversity and importance of EEG in understanding and managing nervous system and psychiatric diseases. In addition, extracting spatiotemporal features from EEG signals is crucial to revealing disease-related brain activity patterns. For example, in epilepsy detection, EEG WieNet significantly improved detection accuracy by extracting spatiotemporal features by developing a multi-scale convolutional neural network (CNN). Liu et al. proposed a three-dimensional convolutional attention neural network (3DCANN) for EEG emotion recognition, which effectively captures the relationship between dynamic channels by integrating spatiotemporal feature extraction with EEG channel attention modules. In this regard, Zhang et al. introduced two deep learning-based frameworks that can preserve the spatiotemporal feature representation of raw EEG signals to accurately identify human intentions. These frameworks combine convolutional neural networks and recurrent neural networks, both in cascaded and parallel configurations, to effectively explore spatiotemporal information. In addition, Gao et al. developed a new EEG-based spatiotemporal convolutional neural network (ESTCNN) for detecting driver fatigue through the spatiotemporal structure of multi-channel EEG signals. These studies show that spatiotemporal feature extraction plays an important role in improving the interpretability and performance in EEG-related disease research.

[0004] However, in existing EEG-based disease research, the individual variability of EEG signals and the high reliance on label information pose considerable challenges. This individual variability may stem from differences in brain anatomy, age, gender, and even disease status, often leading to inconsistencies in data interpretation, which in turn significantly affects the generalization ability of intelligent models. The need for accurate and extensive labeled data usually requires a lot of time and cost. In addition, labels are often subject to human errors and biases, which may introduce noise into the training process, causing the model to perform well on a specific dataset but fail to generalize effectively to new, unseen data. Summary of the invention

[0005] The purpose of the present invention is to provide a training method, device, equipment and storage medium for a multi-concept identification network for electroencephalogram signal analysis, aiming to solve the problem that the existing technology relies on a large amount of labeled data, resulting in weak generalization ability and robustness of the trained network and inaccurate results of EEG analysis.

[0006] In one aspect, the present invention provides a method for training a multi-concept discrimination network for electroencephalogram signal analysis, the method comprising the following steps:

[0007] Constructing a multi-concept discrimination network for analyzing electroencephalogram signals, wherein the multi-concept discrimination network includes a first module constructed based on an encoder-decoder and a second module constructed based on a generative adversarial network and including a plurality of discriminators;

[0008] The multi-concept identification network is trained by semi-supervised learning according to the electroencephalogram signal training set until the target loss function of the multi-concept identification network converges.

[0009] Preferably, the step of performing semi-supervised learning training on the multi-concept identification network according to the EEG signal training set until the target loss function of the multi-concept identification network converges comprises:

[0010] Extracting features of training samples in the electroencephalogram signal training set through the first module to obtain sample spatiotemporal features of corresponding training samples;

[0011] Based on the training samples and the sample spatiotemporal features, a target loss value of the target loss function is calculated to update the network parameters of the multi-concept identification network according to the target loss value.

[0012] Preferably, the first module includes a first encoder, a decoder and a second encoder, a decoder, and the step of extracting features from the training samples in the EEG signal training set by the first module includes:

[0013] Extracting spatial features from the training sample using the first encoder to obtain a first feature;

[0014] Performing time feature extraction on the first feature by the second encoder to obtain a second feature;

[0015] Performing a nonlinear transformation on the second feature by using a multilayer perceptron in the first module to obtain a third feature;

[0016] Based on the third feature, the sample spatiotemporal feature of the training sample is obtained.

[0017] Preferably, the first module includes the first encoder and decoder constructed using a convolutional neural network and the second encoder and decoder constructed using a bidirectional gated recurrent unit.

[0018] Preferably, the EEG signal training set includes first, second and third labels, the second module includes first, second, third and fourth discriminators, and the step of calculating the target loss value of the target loss function based on the training samples and the sample spatiotemporal features includes:

[0019] Based on the training sample, calculating a first loss value of a first loss function of a first discriminator;

[0020] Calculating a second loss value of a second loss function of a second discriminator based on the sample spatiotemporal features and the first label;

[0021] Based on the sample spatiotemporal features and the second label, calculating a third loss value of a third loss function of a third discriminator;

[0022] Calculating a fourth loss value of a fourth loss function of a fourth discriminator based on the sample spatiotemporal features and the third label;

[0023] Based on the first, second, third and fourth loss values, the target loss value is determined.

[0024] In another aspect, the present invention provides a training device for a multi-concept discrimination network for electroencephalogram signal analysis, the device comprising:

[0025] A network construction unit, used to construct a multi-concept identification network for analyzing electroencephalogram signals, wherein the multi-concept identification network includes a first module constructed based on an encoder-decoder and a second module constructed based on a generative adversarial network and including a plurality of discriminators;

[0026] The network training unit is used to perform semi-supervised learning training on the multi-concept identification network according to the electroencephalogram signal training set until the target loss function of the multi-concept identification network converges.

[0027] Preferably, the network training unit comprises:

[0028] A feature extraction unit, configured to extract features from training samples in the electroencephalogram signal training set through the first module to obtain sample spatiotemporal features corresponding to the training samples;

[0029] A loss value calculation unit is used to calculate a target loss value of the target loss function based on the training samples and the sample spatiotemporal features, so as to update the network parameters of the multi-concept identification network according to the target loss value.

[0030] Preferably, the first module includes a first encoder, a decoder and a second encoder, a decoder, and the feature extraction unit includes:

[0031] A first feature extraction unit, configured to extract spatial features from the training sample using the first encoder to obtain a first feature;

[0032] A second feature extraction unit, configured to perform time feature extraction on the first feature through the second encoder to obtain a second feature;

[0033] A third feature obtaining unit, configured to perform a nonlinear transformation on the second feature through a multilayer perceptron in the first module to obtain a third feature;

[0034] The spatiotemporal feature obtaining unit is used to obtain the sample spatiotemporal feature of the training sample based on the third feature.

[0035] On the other hand, the present invention also provides a computing device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps described in the above-mentioned method for training a multi-concept discrimination network for electroencephalogram signal analysis are implemented.

[0036] On the other hand, the present invention also provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, it implements the steps described in the above-mentioned training method for a multi-concept discrimination network for electroencephalogram signal analysis.

[0037] The present invention constructs a multi-concept identification network for analyzing electroencephalogram (EEG) signals, wherein the multi-concept identification network includes a first module constructed based on an encoder-decoder and a second module constructed based on a generative adversarial network and including a plurality of discriminators. Semi-supervised learning training is performed on the multi-concept identification network according to an EEG signal training set until the target loss function of the multi-concept identification network converges, thereby not only reducing the network's dependence on a large amount of labeled data, but also improving the generalization ability and robustness of the network in different patient groups. At the same time, it is possible to cope with the heterogeneity and complexity of EEG signals by focusing on multiple discriminative perspectives, thereby improving the accuracy and reliability of EEG signal analysis, and facilitating its clinical application in diverse environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a flowchart of the implementation of the training method of the multi-concept discrimination network for electroencephalogram signal analysis provided in the first embodiment of the present invention;

[0039] Figure 2 This is a sample sampling schematic diagram provided in the first embodiment of the present invention;

[0040] Figure 3 is a flowchart for implementing feature extraction of training samples provided in the first embodiment of the present invention;

[0041] Figure 4 Schematic diagram of various experimental results of a performance verification experiment on a multi-concept identification network provided in the first embodiment of the present invention;

[0042] Figure 5 is a schematic diagram of the structure of a training device for a multi-concept discrimination network for electroencephalogram signal analysis provided in Embodiment 2 of the present invention;

[0043] Figure 6 It is a schematic diagram of the structure of a computing device provided in Embodiment 3 of the present invention. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0045] The terms "first", "second", "third", etc. in the present invention are used to distinguish similar or similar objects or entities, and do not necessarily mean to limit a specific order or sequence, unless otherwise noted. It should be understood that the terms used in this way can be interchangeable under appropriate circumstances, for example, they can be implemented in an order other than those given in the diagrams or descriptions of the embodiments of the present disclosure.

[0046] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments:

[0047] Embodiment 1:

[0048] Figure 1 The implementation process of the training method of the multi-concept identification network for electroencephalogram signal analysis provided in the first embodiment of the present invention is shown. For the convenience of explanation, only the part related to the embodiment of the present invention is shown, which is described in detail as follows:

[0049] In step S101, a multi-concept identification network for analyzing electroencephalogram signals is constructed, wherein the multi-concept identification network includes a first module constructed based on an encoder-decoder and a second module constructed based on a generative adversarial network and including a plurality of discriminators.

[0050] The embodiment of the present invention is applicable to computing devices, such as personal computers, servers, etc. In the embodiment of the present invention, a multi-concept discrimination network for analyzing EEG signals is constructed, wherein the multi-concept discrimination network includes a first module constructed based on an encoder-decoder structure and a second module constructed based on a generative adversarial network (GAN) and including a plurality of discriminators, wherein the first module is used to extract features of the EEG signal, and each discriminator in the second module is used to distinguish different types of EEG signals.

[0051] In a feasible embodiment, the first module includes two sets of encoders and decoders, namely a first encoder, a decoder and a second encoder, a decoder, wherein the first encoder and decoder are CNN encoders and decoders constructed using Convolutional Neural Networks (CNN), which are used to extract spatial information, and the second encoder and decoder are Bidirectional Gated Recurrent Unit (BiGRU) encoders and decoders constructed using Bidirectional Gated Recurrent Unit (BiGRU), which are used to extract temporal information. Thus, by combining CNN and BiGRU, it is possible to capture the interactions between different brain regions and the changes at different time points, thereby enhancing the representation of complex EEG dynamics.

[0052] In another feasible embodiment, in order to enhance the robustness and accuracy of EEG signal classification, especially when dealing with the complexity of non-stationary time series data, a multi-concept discriminator (i.e., the second module) is introduced into the hybrid CNN-BiGRU encoder-decoder structure. The second module aims to improve the quality and generalization ability of the multi-concept discrimination network by processing multiple aspects of the data. Here, the second module includes first, second, third, and fourth discriminators to cover signal quality, gender differences, domain differences, and specific disease characteristics. Specifically, the first discriminator is a specific signal discriminator for distinguishing between real and generated EEG signals to enhance the effectiveness of the features; the second discriminator is a specific gender discriminator for distinguishing EEG patterns of different genders to extract more generalized features; the third discriminator is a specific domain discriminator for aligning features from labeled source domains, unlabeled source domains, and unknown target domains to reduce differences between domains; the fourth discriminator is a specific disease discriminator for identifying EEG features of depressed adolescents to distinguish between NSSI behavior patients and non-NSSI behavior patients among depressed adolescents.

[0053] In step S102, semi-supervised learning training is performed on the multi-concept identification network according to the EEG signal training set until the target loss function of the multi-concept identification network converges.

[0054] In the embodiment of the present invention, each training sample in the electroencephalogram signal training set is an EEG signal, and the electroencephalogram signal training set includes three labels, namely, first, second, and third labels, for grouping the training samples. Specifically, the first label is a gender label, for distinguishing whether the training sample comes from a male or female patient, and the second label is a domain label, for distinguishing whether the training sample comes from a labeled source domain (denoted as ), unlabeled source domain (denoted as ) or unknown target domain (denoted as ), the third label is the disease label, which is used to distinguish whether the training samples come from the depression with NSSI (DN+) group or the depression without NSSI (DN-) group. Here, the multi-concept discrimination network is trained by semi-supervised learning according to the labeled EEG signal training set until the target loss function of the multi-concept discrimination network converges, so as to use semi-supervised learning to make the multi-concept discrimination network achieve a stable performance even with a small amount of labeled data.

[0055] As an example, in a specific embodiment, the collected EEG signal data is divided into two areas: source area and target domain In the semi-supervised learning framework, the source domain is divided into Further subdivided into a smaller labeled source domain (τ%) and a large unlabeled source domain (Proportion (1-τ%)), considering the data imbalance (DN+ group: 77 patients, including 65 females and 12 males; DN- group: 37 patients, including 18 females and 19 males) and gender distribution, , and The field implements data sampling strategies such as Figure 2 As shown, the sampling process for the DN+ and DN- groups is explained: for the DN+ group, 18 females (represented by the red shaded graph) were randomly selected from the 65 female candidates, and all 12 males were included to ensure a balanced gender ratio within the DN+ group. For the DN- group, all 18 females (represented by the solid blue graph) were included, while 12 were randomly selected from the 19 available male candidates to ensure a gender distribution comparable to that of the DN+ group. This sampling process resulted in a balanced number of patients in the DN+ and DN- groups, 30 NSSI patients and 30 non-NSSI patients, respectively, and were balanced in terms of condition and gender.

[0056] In a feasible embodiment, the training of the multi-concept discrimination network is achieved through the following steps:

[0057] (1) The first module extracts features from the training samples in the EEG signal training set to obtain the sample spatiotemporal features of the corresponding training samples;

[0058] In the embodiment of the present invention, after the EEG signal training set is input into the multi-concept identification network, firstly, the first module performs feature extraction on the training samples in the EEG signal training set to obtain the sample spatiotemporal features of the corresponding training samples.

[0059] Preferably, if Figure 3 As shown in the figure, feature extraction of training samples is achieved through the following steps:

[0060] In step S301, a first encoder is used to extract spatial features from a training sample to obtain a first feature;

[0061] In this embodiment of the present invention, the CNN encoder processes the input training samples. Perform spatial feature extraction, that is, capture the features of the training samples in the spatial dimension. Specifically, the CNN encoder uses convolution operations to map the input training samples into a more expressive feature space to provide a compact and comprehensive EEG spatial dynamic representation. The obtained spatial feature is called the first feature, which is expressed as ,in, is the number of feature maps in the first feature, is the feature dimension, is the number of EEG channels, Represents the number of sampling points per second, is the set of real numbers.

[0062] In step S302, a second encoder is used to extract a time feature from the first feature to obtain a second feature;

[0063] In this embodiment of the present invention, the first feature extracted It is then input into the BiGRU encoder to process the feature vector sequence generated by the CNN encoder. , thereby further revealing the relationship of brain activity in the temporal dimension and encapsulating the key temporal dynamics of the entire EEG signal sequence. Specifically, in the BiGRU encoder, the forward GRU layer generates a forward hidden state sequence, and the backward GRU layer generates a backward hidden state sequence, and these two sequences are then concatenated to form a new, more informative deep feature representation vector , the deep feature representation vector is called the second feature.

[0064] In step S303, the second feature is nonlinearly transformed by the multilayer perceptron in the first module to obtain a third feature;

[0065] In the embodiment of the present invention, the second feature The input is sent to the Multi-Layer Perceptron (MLP) in the first module. The MLP transforms the feature representation into a finer feature space and captures the nonlinear relationship between feature elements. The feature processed by the MLP is called the third feature and is expressed as .

[0066] In step S304, based on the third feature, the sample spatiotemporal feature of the training sample is obtained.

[0067] In the embodiment of the present invention, the third feature Flattening operation is performed to generate a rich vector representation containing more information, which is called sample spatiotemporal feature and expressed as .

[0068] The above steps S301 to S304 combine the spatial features extracted by the CNN layer and the temporal features extracted by the recursive layer to achieve feature extraction of training samples, thereby effectively capturing the spatiotemporal representation of NSSI-related brain activities and providing a comprehensive understanding of the dynamics of EEG signals.

[0069] (2) Based on the training samples and the sample spatiotemporal characteristics, the target loss value of the target loss function is calculated to update the network parameters of the multi-concept identification network according to the target loss value.

[0070] In the embodiment of the present invention, preferably, the calculation of the target loss value is achieved through the following steps:

[0071] ① Based on the training sample, calculate the first loss value of the first loss function of the first discriminator;

[0072] In an embodiment of the present invention, in order to identify the quality of the signal generated by the CNN-BiGRU encoder-decoder architecture, a specific signal discriminator is designed, which aims to enhance the feature representation capability by distinguishing the generated signal from the real signal. By providing adversarial feedback, the specific signal discriminator plays an important role in improving the feature representation capability of the entire network, ensuring that the generated signal is of high quality and effectively presents the required characteristics. Here, based on the training sample, the first loss value of the first loss function of the first discriminator is calculated. Specifically, the first loss function of the first discriminator is: ,in, is the first discriminator, For the generator, is the GAN loss, is the reconstruction loss, For balance and to encourage the generation of high-quality EEG signals.

[0073] In one possible embodiment, the GAN loss Defined as: , reconstruction loss Defined as: ,in, is the training sample, i.e. the original EEG signal, For the generator Reconstructed EEG signal, specifically, the first module is used as a generator, then That is, the first module is training samples The reconstructed samples obtained after encoding and decoding processing.

[0074] ② Based on the sample spatiotemporal features and the first label, calculate the second loss value of the second loss function of the second discriminator;

[0075] In the embodiment of the present invention, considering the gender differences in NSSI patients and the differences in EEG pattern representation between men and women, a specific gender discriminator is designed to capture and utilize gender-specific EEG features. By integrating the specific gender discriminator, the multi-concept discrimination network can not only consider gender differences, but also improve the overall classification accuracy, ensuring that the subtle differences in EEG signals related to gender can be fully represented and used for better prediction results. Here, based on the sample spatiotemporal features and the first label, the second loss value of the second loss function of the second discriminator is calculated. Specifically, the second loss function of the second discriminator is defined as:

[0076] ,in, For the generator Parameters, Represents a gender classifier (i.e. the parameters of the second discriminator), and Represent the number of male and female patients in each batch, For the The temporal and spatial characteristics of the samples of male patients, For the The spatial and temporal characteristics of the samples of female patients, the loss function This aims to ensure that the multi-concept discrimination network effectively captures gender-related EEG features, which plays a key role in processing gender differences.

[0077] ③ Based on the sample spatiotemporal features and the second label, calculate the third loss value of the third loss function of the third discriminator;

[0078] In an embodiment of the present invention, due to individual differences in EEG signals, a specific domain discriminator is designed to align the feature distributions between the labeled source domain, the unlabeled source domain, and the unknown target domain. The specific domain discriminator aims to reduce the differences between these domains by ensuring that the features extracted from different sources remain consistent. By aligning these distributions, the specific domain discriminator helps to minimize the domain gap, thereby enhancing the generalization ability of the multi-concept discrimination network on new and unseen data. Here, based on the sample spatiotemporal features and the second label, the third loss value of the third loss function of the third discriminator is calculated. Specifically, the third loss function of the third discriminator is defined as:

[0079] ,in, Representation domain classifier (i.e. the parameters of the third discriminator), represents the total number of samples of all patients in each batch, For the training samples, It indicates A domain label indicator of whether it is from a labeled source domain, an unlabeled source domain, or an unknown target domain, for Corresponding sample spatiotemporal characteristics. Loss function By minimizing the ability of the domain classifier to distinguish different domains, the first module is encouraged to generate domain-invariant features. Therefore, the multi-concept discrimination network can generate features that are robust in different domains and improve the generalization ability across subjects.

[0080] ④ Based on the sample spatiotemporal features and the third label, calculating a fourth loss value of a fourth loss function of a fourth discriminator;

[0081] In the embodiment of the present invention, in order to effectively capture the differential EEG features between DN+ and DN-, a specific disease discriminator is designed, which plays a key role in distinguishing patients according to EEG patterns, ensuring that the unique features associated with each disease can be accurately identified. Here, based on the sample spatiotemporal features and the third label, the fourth loss value of the fourth loss function of the fourth discriminator is calculated. Specifically, the fourth loss function of the fourth discriminator is defined as:

[0082] ,in, Represents a binary label indicating whether NSSI is present (DN+ or DN-), represents the number of samples in the labeled source domain batch, is the disease classifier (i.e. the fourth discriminator), is the sigmoid activation function. This loss function is designed to drive the multi-concept discrimination network to accurately classify patients based on EEG features and enhance its predictive ability to identify NSSI.

[0083] ⑤ Based on the first, second, third and fourth loss values, determine the target loss value.

[0084] In the embodiment of the present invention, the target loss function of the multi-concept identification network is formed by weighted combination of the first, second, third and fourth loss functions. Specifically, the target loss function is:

[0085] ,in, , and is the weighting coefficient of the corresponding loss function. Through this target loss function, the multi-concept discrimination network can generate robust and multi-dimensional EEG feature representations that are invariant to signal noise and can sensitively capture gender differences, domain adaptability, and specific disease characteristics, especially in the classification of NSSI. Here, based on the calculated first, second, third, and fourth loss values, the target loss function is The target loss value is calculated to obtain the target loss value, and the network parameters of the multi-concept identification network are updated according to the target loss value. The above steps (1) to (2) are repeated to continue training the multi-concept identification network until the relative change of the target loss value is less than a certain threshold (that is, the target loss function converges), and the training is completed.

[0086] By optimizing the objective loss function through the above steps ① to ⑤, the multi-concept discrimination network can better adapt to the complexity associated with EEG signal analysis, thus facilitating the study of depression and NSSI.

[0087] In order to evaluate the performance of the multi-concept identification network, the following experiments were conducted on the trained multi-concept identification network:

[0088] Experimental Dataset: EEG dataset, which includes recordings from 114 adolescent patients diagnosed with depression. Specifically, according to the NSSI criteria, the adolescent patients were divided into two groups: the DN+ group, including 77 subjects (65 females), and the DN- group, including 37 subjects (18 females). Each experimenter completed a total of 35 trials, each trial involved watching natural images related to social pain, and each trial lasted for 5 seconds. During this process, EEG data were simultaneously recorded through a 63-channel BrainAmp system, following The international 10-20 system was used, and the sampling rate of data collection was 500 Hz. A standard preprocessing process was applied to the raw EEG signal to remove artifacts, including eye movements, muscle activity, and environmental noise. After artifact correction, the sampling rate was reduced from 500 Hz to 384 Hz, thereby reducing the dimension of the data and making subsequent processing and analysis more efficient. In order to increase the sample size for modeling, each 5-second trial was further divided into multiple 1-second samples. Therefore, the final data format can be expressed as follows: 114 participants × 35 trials × 5 seconds × 63 channels × 384 sampling points;

[0089] (i) Using the above experimental dataset and a semi-supervised 10-fold cross-validation protocol, the proposed multi-concept identification network (denoted as NSSI-Net) is compared with various machine learning and deep learning models. The experimental results are shown in Figure 2. Figure 4 As shown in (a), Figure 4The methods shown in (a) in the figure for comparison with NSSI-Net can be divided into two parts: traditional machine learning methods (including: Support Vector Machine (SVM), Transfer Component Analysis (TCA), Subspace Alignment (SA), Kernel Correlation Component Analysis (KPCA), Random Forest (RF), Adaptive Boosting (Adaboost), Correlation Alignment (CORAL), Geodesic Flow Kernel (GFK)) and deep learning methods (including: Deep Adaptation Network (DAN), Domain-Adversarial Neural Network (DANN), Deep Correlation Alignment (DCORAL), Deep Domain Confusion (DDO), ... Confusion, DDC), among the traditional machine learning methods, Adaboost has the highest accuracy of 57.85±06.57, followed by CORAL with an accuracy of 57.07±03.45. Other traditional models, such as SA, RF, and SVM, performed relatively poorly, with accuracies ranging from 48.09% to 56.33. In contrast, deep learning methods provide more competitive results, among which the best performing model is DDC with an accuracy of 62.56±2.52, DAN with an accuracy of 61.46±03.86, and DCORAL with 60.91±02.95%. These models are better than most traditional methods, but still inferior to NSSI-Net. Another deep learning method, DANN, performed moderately with an accuracy of 58.20±02.80. Through a comprehensive comparison of the above methods, the NSSI-Net of the present invention stood out as the best performing model with an accuracy of 70.00±13.90%, surpassing both traditional and deep learning methods. It was 7.44% higher than the next best model (DDC), highlighting its effectiveness in capturing the complex patterns associated with adolescent NSSI behaviors and establishing its position as a promising tool for early detection and intervention.

[0090] (ii) To further evaluate the performance of the NSSI-Net model in different subgroups and identify any bias or inconsistency in the prediction results, the EEG signal analysis was performed on the confusion matrices of the three groups (including: the entire experimental dataset, the female subgroup, and the male subgroup). Here, DN+ was considered as the positive class and DN- was considered as the negative class. The experimental results are shown in Figure 2. Figure 4 As shown in (b) in Figure 4 As can be seen in (b), on the entire dataset, the overall confusion matrix shows that the true negative rate (TN) is 78.60% and the true positive rate (TP) is 60.70%, indicating that the model is relatively effective in distinguishing DN+ and DN- cases. However, the false positive rate (FP) is 21.40% and the false negative rate (FN) is 39.30%, showing certain limitations, especially in identifying DN+ cases. The higher false negative rate indicates that the model has difficulty in correctly identifying DN+ cases, which may be related to the intrinsic differences in EEG signals of different individuals. In the female subgroup, the confusion matrix shows that the TN rate is 82.86% and the TP rate is 59.35%, indicating that the performance in correctly identifying DN- cases is slightly better than that of DN+ cases. The FP rate is lower at 17.14%, while the FN rate is 40.65%. These results indicate that the model is more accurate in predicting DN- cases in females, which may be due to the fact that NS The higher incidence of SI makes the EEG patterns associated with DN- more prominent, however, the higher false negative rate shows room for improvement in detecting DN+ cases in this subgroup; in the male subgroup, the confusion matrix shows a lower TN rate of 71.92%, while the TP rate is 62.85%, the FP rate is higher, at 28.08%, and the FN rate is lower, at 37.15%. This distribution indicates that the model is more sensitive to DN+ cases, but at the expense of a higher false positive rate. There may be a gender-specific bias, which may be affected by the lower incidence of NSSI in males, making it more difficult for the model to distinguish between DN- and DN+ cases. In summary, the results of the confusion matrix show that although the model performs well overall, there are differences in performance between gender-specific subgroups. The model more accurately identifies DN- cases in females and DN+ cases in males, which may be related to biological and psychological gender differences. However, the higher false negative rate in females still indicates that there is room for improvement in detecting DN+ cases. Future research should focus on these specific differences to further improve the robustness and accuracy of the model.

[0091] ( ) An ablation experiment was conducted on the NSSI-Net model to evaluate the contribution of each discriminator in the model to the performance. Figure 4(c) shows the model results of using semi-supervised cross-subject ten-fold cross validation under different discriminator configurations:

[0092] (1) Contribution of specific signal discriminator: After removing the specific signal discriminator, the accuracy is reduced to 65.47±10.22%, which highlights its key role in extracting the basic features of relevant signals;

[0093] (2) Contribution of specific gender discriminators: After removing the specific gender discriminator, the accuracy slightly decreased to 66.63±09.66%, indicating that gender-specific features contribute to the overall classification performance, although the impact is not as significant as signal-specific features;

[0094] (3) Contribution of domain-specific discriminators: After removing the domain-specific discriminators, the accuracy drops to 64.30±09.88%, showing the importance of domain adaptation in improving model performance;

[0095] (4) Comparison with traditional domain discriminators: The traditional domain adaptation method is based on the DANN framework. The accuracy of the proposed domain discriminator is 67.17±12.99%, which is lower than the performance of the proposed NSSI-Net of 70.00±13.90%. This shows the effectiveness of the proposed multi-concept identification network framework. The results prove the importance of each discriminator in improving the model performance, among which the signal-specific and domain-specific discriminators have the greatest impact;

[0096] In addition to removing individual discriminators to evaluate the contribution of each component, we also explored the impact of different combinations of discriminators (signal+gender, gender+domain, etc.) on the overall prediction accuracy of the model. The performance of the model under different discriminator combinations is shown in Figure 2. Figure 4 As shown in (c):

[0097] (1) Signal + disease specificity: When using only these two components, the model achieved an accuracy of 62.72%, indicating the importance of disease-related information in distinguishing classes;

[0098] (2) Signal + Gender + Disease Specificity: After adding the gender-specific component, the accuracy increased to 64.30%, indicating that gender-specific EEG patterns provide additional predictive power;

[0099] (3) Gender + domain + disease specificity: The accuracy of combining the gender and domain-specific components was 65.47%, indicating that combining domain-related information with gender-specific features further improved the predictive ability, highlighting the importance of domain differences;

[0100] (4) Signal + Gender + Domain + Disease Specificity: When all discriminators were combined, the highest accuracy reached 70.00%, indicating that each discriminator captured unique and complementary aspects of the data and their integration provided the most accurate prediction;

[0101] The above analysis shows that the signal-specific components capture the underlying features, while the gender- and domain-specific components provide additional refinement by accounting for gender- and domain-specific variations. Although the individual components are informative, their combined effect maximizes the predictive performance of the model.

[0102] ( ) In the semi-supervised learning process, the availability of labeled source data is crucial to determine the effectiveness of model training, especially when processing complex EEG signals. Understanding the impact of different proportions of labeled source data helps to evaluate the trade-off between data amount and model performance, providing insights on the amount of labeled data required to achieve the best results, such as Figure 4 As shown in (d) in the figure, we adjusted the proportion of labeled source data from 5% to 85%. The experimental results show that when only 5% of the labeled source domain data is used, the model is difficult to learn effectively and the accuracy is low, only 50.97%. When the proportion increases to 10%, the accuracy is significantly improved and the model reaches an average accuracy of 65.71%. This indicates that the model requires at least 10% of the labeled source data to begin to capture meaningful patterns in the EEG signal. After exceeding the proportion of 10%, the accuracy continues to increase steadily and reaches a plateau when the proportion reaches 75%, with an accuracy of 70.00%. This shows that the model is able to effectively utilize labeled and unlabeled data through its semi-supervised learning framework, and can maintain stable and robust performance even when the labeled source data is relatively limited.

[0103] In an embodiment of the present invention, a multi-concept discrimination network for analyzing electroencephalogram signals is constructed, wherein the multi-concept discrimination network includes a first module constructed based on an encoder-decoder and a second module constructed based on a generative adversarial network and including a plurality of discriminators. Semi-supervised learning training is performed on the multi-concept discrimination network according to an electroencephalogram signal training set until the target loss function of the multi-concept discrimination network converges, thereby not only reducing the network's dependence on a large amount of labeled data, but also improving the generalization ability and robustness of the network in different patient groups. At the same time, it is possible to cope with the heterogeneity and complexity of electroencephalogram signals by focusing on multiple discriminative perspectives, thereby improving the accuracy and reliability of electroencephalogram signal analysis, and facilitating its clinical application in diverse environments.

[0104] Embodiment 2:

[0105] Figure 5The structure of the training device for the multi-concept identification network for electroencephalogram signal analysis provided by the second embodiment of the present invention is shown. For the convenience of description, only the part related to the embodiment of the present invention is shown, including:

[0106] The network construction unit 51 is used to construct a multi-concept identification network for analyzing electroencephalogram signals, wherein the multi-concept identification network includes a first module constructed based on an encoder-decoder and a second module constructed based on a generative adversarial network and including a plurality of discriminators.

[0107] The network training unit 52 is used to perform semi-supervised learning training on the multi-concept identification network according to the EEG signal training set until the target loss function of the multi-concept identification network converges.

[0108] Preferably, the network training unit 52 includes:

[0109] A feature extraction unit, configured to extract features from training samples in the electroencephalogram signal training set through a first module to obtain sample spatiotemporal features corresponding to the training samples, wherein the first module includes a first encoder, a decoder and a second encoder, a decoder;

[0110] The loss value calculation unit is used to calculate the target loss value of the target loss function based on the training samples and the sample spatiotemporal characteristics, so as to update the network parameters of the multi-concept identification network according to the target loss value.

[0111] Preferably, the feature extraction unit comprises:

[0112] A first feature extraction unit, configured to extract spatial features from the training sample using a first encoder to obtain a first feature;

[0113] A second feature extraction unit is used to extract a time feature from the first feature through a second encoder to obtain a second feature;

[0114] A third feature obtaining unit, used for performing nonlinear transformation on the second feature through the multi-layer perceptron in the first module to obtain a third feature;

[0115] The spatiotemporal feature acquisition unit is used to obtain the sample spatiotemporal features of the training sample based on the third feature.

[0116] In the embodiment of the present invention, each unit of the training device for the multi-concept discrimination network for electroencephalogram signal analysis can be implemented by a corresponding hardware or software unit, and each unit can be an independent software or hardware unit, or can be integrated into a software or hardware unit, which is not intended to limit the present invention. Specifically, the implementation of each unit can refer to the description of the aforementioned embodiment 1, which will not be repeated here.

[0117] Embodiment three:

[0118] Figure 6 The structure of a computing device provided in the third embodiment of the present invention is shown. For the convenience of description, only the part related to the embodiment of the present invention is shown.

[0119] The computing device 6 of the embodiment of the present invention includes a processor 60, a memory 61, and a computer program 62 stored in the memory 61 and executable on the processor 60. When the processor 60 executes the computer program 62, the steps in the embodiment of the training method of the multi-concept discrimination network for electroencephalogram signal analysis are implemented, for example, Figure 1 Alternatively, when the processor 60 executes the computer program 62, the functions of each unit in the above-mentioned device embodiments are realized, for example Figure 5 Function of the unit shown.

[0120] In an embodiment of the present invention, a multi-concept discrimination network for analyzing electroencephalogram signals is constructed, wherein the multi-concept discrimination network includes a first module constructed based on an encoder-decoder and a second module constructed based on a generative adversarial network and including a plurality of discriminators. Semi-supervised learning training is performed on the multi-concept discrimination network according to an electroencephalogram signal training set until the target loss function of the multi-concept discrimination network converges, thereby not only reducing the network's dependence on a large amount of labeled data, but also improving the generalization ability and robustness of the network in different patient groups. At the same time, it is possible to cope with the heterogeneity and complexity of electroencephalogram signals by focusing on multiple discriminative perspectives, thereby improving the accuracy and reliability of electroencephalogram signal analysis, and facilitating its clinical application in diverse environments.

[0121] The computing device of the embodiment of the present invention can be a personal computer or a server. The steps implemented when the processor 60 in the computing device 6 executes the computer program 62 to implement a training method for a multi-concept identification network for EEG signal analysis can refer to the description of the aforementioned method embodiment, which will not be repeated here.

[0122] Embodiment 4:

[0123] In an embodiment of the present invention, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, the steps in the training method of the multi-concept discrimination network for electroencephalogram signal analysis are implemented, for example, Figure 1 Alternatively, when the computer program is executed by a processor, the functions of each unit in the above-mentioned device embodiments are realized, for example Figure 5 Function of the unit shown.

[0124] In an embodiment of the present invention, a multi-concept discrimination network for analyzing electroencephalogram signals is constructed, wherein the multi-concept discrimination network includes a first module constructed based on an encoder-decoder and a second module constructed based on a generative adversarial network and including a plurality of discriminators. Semi-supervised learning training is performed on the multi-concept discrimination network according to an electroencephalogram signal training set until the target loss function of the multi-concept discrimination network converges, thereby not only reducing the network's dependence on a large amount of labeled data, but also improving the generalization ability and robustness of the network in different patient groups. At the same time, it is possible to cope with the heterogeneity and complexity of electroencephalogram signals by focusing on multiple discriminative perspectives, thereby improving the accuracy and reliability of electroencephalogram signal analysis, and facilitating its clinical application in diverse environments.

[0125] The computer-readable storage medium of the embodiment of the present invention may include any entity or device or recording medium capable of carrying computer program code, for example, ROM / RAM, magnetic disk, optical disk, flash memory and other memories.

[0126] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A training method for a multi-concept discrimination network for electroencephalogram signal analysis, characterized in that: The method comprises the following steps: Constructing a multi-concept discrimination network for analyzing electroencephalogram signals, wherein the multi-concept discrimination network includes a first module constructed based on an encoder-decoder and a second module constructed based on a generative adversarial network and including a plurality of discriminators, the first module includes a first encoder and a decoder constructed using a convolutional neural network for extracting spatial information and a second encoder and a decoder constructed using a bidirectional gated recurrent unit for extracting temporal information, the second module includes first, second, third, and fourth discriminators, the first discriminator is a specific signal discriminator for distinguishing between real and generated EEG signals to enhance the effectiveness of features, the second discriminator is a specific gender discriminator for distinguishing EEG patterns of different genders to extract more generalized features, the third discriminator is a specific domain discriminator for aligning features from a labeled source domain, an unlabeled source domain, and an unknown target domain to reduce differences between domains, and the fourth discriminator is a specific disease discriminator for identifying EEG features of depressed adolescents to distinguish between NSSI behavior patients and non-NSSI behavior patients among depressed adolescents; The multi-concept identification network is trained by semi-supervised learning according to the electroencephalogram signal training set until the target loss function of the multi-concept identification network converges, wherein the electroencephalogram signal training set comprises first, second and third labels, the first label is a gender label for distinguishing whether the training samples are from male or female patients, the second label is a domain label for distinguishing whether the training samples are from a labeled source domain, an unlabeled source domain or an unknown target domain, and the third label is a disease label for distinguishing whether the training samples are from a depression with self-injury group or a depression without self-injury group.

2. The method according to claim 1, characterized in that The step of performing semi-supervised learning training on the multi-concept identification network according to the electroencephalogram signal training set until the target loss function of the multi-concept identification network converges comprises: Extracting features of training samples in the electroencephalogram signal training set through the first module to obtain sample spatiotemporal features of corresponding training samples; Based on the training samples and the sample spatiotemporal features, a target loss value of the target loss function is calculated to update the network parameters of the multi-concept identification network according to the target loss value.

3. The method according to claim 2, characterized in that The step of extracting features from training samples in the electroencephalogram signal training set by the first module includes: Extracting spatial features from the training sample using the first encoder to obtain a first feature; Performing time feature extraction on the first feature by the second encoder to obtain a second feature; Performing a nonlinear transformation on the second feature by using a multilayer perceptron in the first module to obtain a third feature; Based on the third feature, the sample spatiotemporal feature of the training sample is obtained.

4. The method according to claim 2, characterized in that The step of calculating the target loss value of the target loss function based on the training sample and the sample spatiotemporal characteristics comprises: Based on the training sample, calculating a first loss value of a first loss function of a first discriminator; Calculating a second loss value of a second loss function of a second discriminator based on the sample spatiotemporal features and the first label; Based on the sample spatiotemporal features and the second label, calculating a third loss value of a third loss function of a third discriminator; Calculating a fourth loss value of a fourth loss function of a fourth discriminator based on the sample spatiotemporal features and the third label; Based on the first, second, third and fourth loss values, the target loss value is determined.

5. A training device for a multi-concept discrimination network for electroencephalogram signal analysis, characterized in that: The device comprises: A network construction unit, used for constructing a multi-concept discrimination network for analyzing electroencephalogram signals, wherein the multi-concept discrimination network includes a first module constructed based on an encoder-decoder and a second module constructed based on a generative adversarial network and including a plurality of discriminators, the first module includes a first encoder and a decoder constructed using a convolutional neural network for extracting spatial information and a second encoder and a decoder constructed using a bidirectional gated recurrent unit for extracting temporal information, the second module includes first, second, third and fourth discriminators, the first discriminator is a specific signal discriminator for distinguishing between real and generated EEG signals to enhance the effectiveness of features, the second discriminator is a specific gender discriminator for distinguishing EEG patterns of different genders to extract more generalized features, the third discriminator is a specific domain discriminator for aligning features from a labeled source domain, an unlabeled source domain and an unknown target domain to reduce differences between domains, and the fourth discriminator is a specific disease discriminator for identifying EEG features of depressed adolescents to distinguish between NSSI behavior patients and non-NSSI behavior patients among depressed adolescents; A network training unit is used to perform semi-supervised learning training on the multi-concept identification network according to an electroencephalogram signal training set until the target loss function of the multi-concept identification network converges, wherein the electroencephalogram signal training set includes first, second, and third labels, the first label is a gender label, used to distinguish whether the training sample comes from a male or female patient, the second label is a domain label, used to distinguish whether the training sample comes from a labeled source domain, an unlabeled source domain, or an unknown target domain, and the third label is a disease label, used to distinguish whether the training sample comes from a depression with self-injury group or a depression without self-injury group.

6. The device according to claim 5, characterized in that The network training unit comprises: A feature extraction unit, configured to extract features from training samples in the electroencephalogram signal training set through the first module to obtain sample spatiotemporal features corresponding to the training samples; A loss value calculation unit is used to calculate a target loss value of the target loss function based on the training samples and the sample spatiotemporal features, so as to update the network parameters of the multi-concept identification network according to the target loss value.

7. The device according to claim 6, characterized in that The feature extraction unit comprises: A first feature extraction unit, configured to extract spatial features from the training sample using the first encoder to obtain a first feature; A second feature extraction unit, configured to perform time feature extraction on the first feature through the second encoder to obtain a second feature; A third feature obtaining unit, configured to perform a nonlinear transformation on the second feature through a multilayer perceptron in the first module to obtain a third feature; The spatiotemporal feature obtaining unit is used to obtain the sample spatiotemporal feature of the training sample based on the third feature.

8. A computing device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 4 are implemented.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.

Images