A rapid depression detection method based on weighted degree transition network

Abstract

The application provides a rapid depression detection method based on a weighted degree transition network, comprising the following steps: collecting an electroencephalogram signal of a subject and pre-processing to obtain a one-dimensional electroencephalogram time sequence; mapping the one-dimensional electroencephalogram time sequence into a complex network by using a weighted level visual graph algorithm; extracting a degree sequence and an intensity sequence; taking a degree value set after deduplication as a node of a new network; tracking a degree value transition path of adjacent time points; calculating a product of a degree value difference and an intensity value difference of adjacent time points as an edge weight of the degree transition path; constructing a weighted degree transition network; normalizing the network to obtain a transition probability matrix; calculating a Shannon entropy and an MPR statistical complexity of the transition probability matrix as a joint feature vector; and using a machine learning classifier to realize depression state recognition. The application does not need to preset parameters such as embedding dimension, and can quickly and accurately capture abnormal nonlinear dynamic patterns in the electroencephalogram signal of a depression patient by fusing dynamic difference features of degrees and intensities.

Description

Technical Field

[0001] This invention belongs to the field of electroencephalogram (EEG) signal processing technology and relates to a rapid depression detection method based on a weighted degree transfer network. Background Technology

[0002] Depression, one of the most common mood disorders, is characterized by persistent low mood, loss of interest, poor concentration, slowed thinking, and decreased motivation. These clinical manifestations are closely related to brain function, resulting in significant abnormalities in neural rhythm structure, spectral power distribution, and functional connectivity between brain regions in patients with depression. Numerous studies have demonstrated that depression has specific biomarkers at the neurodynamic level. Therefore, objective identification methods based on electroencephalography (EEG) are of great significance in auxiliary diagnosis. EEG signals are generated by the synchronous firing of a large number of neurons and can reflect the dynamic activity patterns of the cerebral cortex in real time. Compared with other neuroimaging methods, EEG has advantages such as high temporal resolution, non-invasiveness, and low cost, and has been widely used in mood disorder research and clinical auxiliary detection.

[0003] Patients with depression typically exhibit significant differences across multiple characteristic dimensions of EEG. The brain functional changes caused by depression are highly dynamic and cross-scale, making the generation process of EEG signals inherently highly nonlinear. The coupling and fluctuation of neural activity across different spatial and temporal scales make it difficult for traditional methods based on linear assumptions to fully characterize the underlying neural mechanisms of depression.

[0004] To capture these complex changes, nonlinear dynamic analysis methods are often used in research, such as the Lyapunov index to assess the chaos and stability of the system. However, despite the theoretical advantages of this method, its results are highly sensitive to small perturbations, especially in real EEG signals, where noise, artifacts, and physiological fluctuations significantly reduce its stability and reliability. This leads to the output index being prone to bias and lacking robustness, thus limiting its application in large-scale data processing and real-time depression detection tasks. Therefore, how to robustly extract depression-related nonlinear features in complex noisy environments and effectively improve the model's sensitivity and stability to differences in depression representations is a significant challenge in current research. Developing more reliable analysis methods for the dynamics, non-stationarity, and cross-scale features of EEG is of great value for achieving high-precision auxiliary diagnosis of depression.

[0005] Furthermore, degree distribution entropy based on horizontal visualization can also be used to describe the structural information of EEG signals. Degree distribution entropy reflects the heterogeneity and information complexity of the network through the probability distribution entropy of node degrees in a statistical graph, and can capture changes in connectivity characteristics of different brain regions in EEG signals. However, degree distribution entropy is based solely on the statistical distribution of node degrees, ignoring the specific connection patterns and dynamic evolution processes between nodes, making it difficult to fully reveal the temporal nonlinear characteristics of EEG signals. Degree distribution entropy is also sensitive to noise; artifacts and measurement errors in EEG data can lead to deviations in the degree distribution, affecting the accurate calculation of entropy values. In addition, this indicator typically reflects overall network characteristics and struggles to capture subtle differences in local brain regions, limiting its application in fine-grained mental state detection.

[0006] Therefore, existing technologies suffer from problems such as sensitivity to noise, reliance on preset parameters, and inability to fully capture the spatiotemporal dynamic nonlinear characteristics of EEG signals, resulting in insufficient accuracy, robustness, and interpretability in depression detection.

[0007] A search revealed a Chinese invention patent application with a similar technical approach to this invention, entitled "Method and System for Detecting Mental State Based on EEG Signals with Entropy Complexity Plane," application number 202311810547.3. This application is used as a prior art document to briefly describe the technical features that distinguish this invention from the prior art document:

[0008] ① Different network construction methods: The comparison file is based on time series ontology, obtains the delayed embedding matrix through phase space reconstruction, and then converts each row into ordinal mode. Its overall processing only symbolizes the time series and does not involve any form of network structure construction, nor does it capture the explicit transition relationship between different states. In contrast, this invention first constructs a first-order weighted horizontal visualization of EEG signals, explicitly establishes the visual connection relationship between time points, and constructs a second-order weighted degree transition network by combining node degree sequence and node strength sequence, thereby mapping EEG signals into a dynamic network structure with time evolution. This network not only retains local connection information, but also represents the transition law of node degree over time.

[0009] ② Different definitions of entropy index: The comparison document calculates the standard permutation entropy based on the probability distribution of ordinal patterns, reflecting the uncertainty of time series at the symbolic pattern level; while this invention proposes a new information metric, namely weighted degree transfer entropy, for the transfer probability matrix obtained based on weighted degree transfer network. Its measurement object changes from symbolic pattern to the transfer process of network node degree, which can better indicate the dynamic change characteristics of EEG signal. Summary of the Invention

[0010] The purpose of this invention is to address the problems existing in the prior art by providing a rapid depression detection method based on a weighted degree transfer network, which solves the problems of insufficient accuracy, robustness and interpretability in the detection of depression in the prior art.

[0011] To achieve the above objectives, this invention provides a rapid depression detection method based on a weighted degree transfer network, comprising:

[0012] The subjects' electroencephalogram (EEG) signals were collected and preprocessed to obtain a clean EEG time sequence;

[0013] A weighted horizontal visualization algorithm is introduced to map the one-dimensional EEG time series into a first-order weighted undirected graph, and the degree value and node strength of each node in the network are calculated to generate a degree sequence and a strength sequence.

[0014] A weighted degree transfer network is constructed based on the degree sequence and intensity sequence;

[0015] The weighted adjacency matrix of the weighted degree transition network is normalized into a transition probability matrix. The weighted degree transition entropy and MPR statistical complexity are calculated based on the probability matrix to form a two-dimensional feature vector.

[0016] The machine learning classifier, pre-trained based on the two-dimensional feature vector input, outputs a detection result as to whether the subject suffers from depression.

[0017] Specifically:

[0018] The subjects' electroencephalogram (EEG) signals were preprocessed to obtain a clean EEG time sequence, specifically including the following steps:

[0019] A unified and appropriate reference standard is re-established for all electrodes of the EEG signal to reduce the bias introduced by the reference electrodes and improve signal quality and comparability.

[0020] The EEG signal is input into a preset low-pass filter and a preset high-pass filter to filter out signals with frequencies higher than a first threshold and signals with frequencies lower than a second threshold in the EEG signal.

[0021] Independent component analysis was performed to decompose the electroencephalogram (EEG) signal into multiple independent components.

[0022] Eliminate signal artifacts in the electroencephalogram (EEG) signals, including electrooculogram (EOG) artifacts and electromyogram (EMG) artifacts;

[0023] The EEG signals are sampled to obtain the one-dimensional EEG time series.

[0024] A weighted horizontal visualization algorithm is introduced to map the one-dimensional EEG time series into a first-order weighted undirected graph, generating the degree sequence and intensity sequence. This specifically includes the following steps:

[0025] The length is time series Mapped to a weighted undirected graph In the formula, For a one-dimensional EEG time series in the 1st The amplitude corresponding to each sampling time. For a set of nodes, For edge set, For weight set;

[0026] For any two nodes and An edge is established if and only if the values ​​of all data points between them are less than the values ​​of the two data points.

[0027] For each connection, its weight is calculated based on the magnitude difference and time interval of the data points. ;

[0028] Calculate the degree value of each node. That is, with nodes The number of directly connected edges forms a degree sequence. ;

[0029] Calculate the node strength of each node. That is, with nodes The sum of the weights of all connected edges constitutes the intensity sequence. .

[0030] The edge weight The calculation formula is:

[0031]

[0032] In the formula, For a one-dimensional EEG time series in the 1st The amplitude corresponding to each sampling time; For a one-dimensional EEG time series in the 1st The amplitude corresponding to each sampling time.

[0033] The construction of the weighted degree transfer network specifically includes the following steps:

[0034] Traverse the degree sequence Extract all unique degree values ​​that appear in the set. , will set The elements in the definition are nodes of the weighted degree transfer network. ,in The total number of uniqueness values;

[0035] Traverse the degree sequence in chronological order, if time... degree value equal And at all times degree value equal Then at node With nodes Establish a directed edge between them;

[0036] Calculation time arrive Instantaneous transfer intensity Its calculation formula combines the difference in degree and the difference in strength:

[0037]

[0038] In the formula, For a moment The degree value; For a moment The degree value; For a moment Node strength; For a moment The node strength.

[0039] By summing the instantaneous transition intensities corresponding to all identical state transitions, a weighted adjacency matrix for the weighted degree transition network is generated. .

[0040] The weighted adjacency matrix elements in The calculation formula is:

[0041]

[0042] In the formula, This is the Kronecker function, which has a value of 1 if and only if the two parameters inside the parentheses are equal, otherwise it has a value of 0; The length of the one-dimensional EEG time series.

[0043] The feature quantization specifically includes the following steps:

[0044] Weighted adjacency matrix Normalization yields the weighted degree transition probability matrix. ,in:

[0045]

[0046] Based on probability matrix Calculate the weighted degree transition entropy :

[0047]

[0048] In the formula, Probability matrix The elements in.

[0049] Based on probability matrix Calculating the statistical complexity of MPR Specifically, it includes the following steps:

[0050] Calculate the probability matrix With uniform distribution The Jansen-Shannon divergence between :

[0051]

[0052] Calculate normalized divergence :

[0053]

[0054] Multiplying the normalized divergence by the weighted degree transfer entropy gives the MPR complexity:

[0055]

[0056] In the formula, For probability distribution Shannon entropy, For all elements to be The uniform distribution matrix; sharp distribution matrix With uniform distribution matrix The Jansen-Shannon divergence between them; A sharp distribution matrix is ​​a matrix where a single element is 1 and all other elements are 0. It is used to obtain... The maximum value, to achieve Scale normalization.

[0057] Based on the transition probability matrix, the method for detecting whether a subject suffers from depression includes the following steps:

[0058] The Shannon entropy and MPR complexity are concatenated to obtain the two-dimensional feature vector. ;

[0059] The two-dimensional feature vector The dynamic patterns are input into a machine learning classifier to distinguish between depression and healthy control groups.

[0060] The method further includes:

[0061] Acquire electroencephalogram (EEG) signal sample data;

[0062] The EEG signal sample data is divided into multiple sample data using a preset data unit;

[0063] The sample data is divided into a training set and a test set according to a preset ratio;

[0064] Build a classifier;

[0065] The classifier is trained based on the training set, and the trained classifier is validated based on the test set to obtain the depression detection model.

[0066] The beneficial effects of this invention are as follows:

[0067] This invention constructs a weighted degree transfer network, modeling the dynamic evolution of degree sequences as a weighted directed graph. This eliminates the need for pre-setting parameters such as embedding dimension, avoiding subjectivity in parameter selection. Furthermore, it calculates transfer weights by fusing changes in node degree values ​​and node strength, simultaneously capturing the dynamic differences in network topology and node connection strength, thus more accurately characterizing the nonlinear dynamic patterns of EEG signals. In addition, this invention combines weighted degree transfer entropy with MPR statistical complexity to form a joint feature vector, which can describe EEG signals from both information content and structural complexity dimensions, thereby significantly improving the accuracy and robustness of depression detection. Specifically:

[0068] 1. Enhanced feature representation capability: This invention constructs a joint feature vector based on a weighted degree transfer network, effectively integrating the weighted degree transfer entropy with MPR statistical complexity. This enables simultaneous characterization of EEG signals from both multidimensional structural characteristics and probability distribution characteristics, resulting in more comprehensive and richer feature representation.

[0069] 2. Significantly improved classification performance; On the MPHC EEG dataset, the joint feature vector proposed in this invention achieves the best performance in terms of accuracy, sensitivity, and specificity, which is significantly improved compared with the single feature method and can more reliably identify depression samples.

[0070] 3. High feature stability; According to the violin diagram and feature space distribution results, the joint feature vector is more concentrated in the distribution of different samples and has higher separation between categories, which has better stability and robustness, and is conducive to dealing with individual differences and noise effects in EEG signals.

[0071] 4. The calculation process is highly interpretable; the features of this invention are all derived from the weighted degree transition matrix and the probability and statistics model, which have clear mathematical and physical meanings. The interpretability is better than that of black-box deep learning features, which is convenient for clinical research and subsequent expansion applications. Attached Figure Description

[0072] Figure 1 This is a schematic diagram of the network architecture according to an embodiment of the present invention;

[0073] Figure 2 This is a schematic diagram of the weighted degree transfer network and its weighted adjacency matrix construction according to an embodiment of the present invention;

[0074] Figure 3 This is a violin diagram illustrating the classification accuracy of the MPHC dataset in an embodiment of the present invention.

[0075] Figure 4 This is a schematic diagram of the classification sensitivity violin of the MPHC dataset in an embodiment of the present invention;

[0076] Figure 5 This is a schematic diagram of the classification specificity of the MPHC dataset in an embodiment of the present invention.

[0077] Figure 6 This is a visualization diagram of the joint feature vector of the MPHC dataset in an embodiment of the present invention;

[0078] Figure 7 This is a bar chart illustrating the classification index of the MPHC dataset in an embodiment of the present invention. Detailed Implementation

[0079] The technical solutions of this embodiment will now be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0080] This embodiment provides a rapid depression detection method based on a weighted degree transfer network. It extracts deep nonlinear features from EEG signals by constructing a two-stage network model and finally uses a machine learning classifier to achieve accurate identification of depression.

[0081] like Figure 1 As shown, the architecture of this embodiment includes a signal acquisition module, a preprocessing module, a feature extraction module, and a depression detection module; wherein:

[0082] The signal acquisition module is used to acquire the electroencephalogram (EEG) signals of the target object.

[0083] The preprocessing module is used to perform preliminary processing on the acquired raw EEG signals to suppress noise interference, improve signal quality, and obtain a purer and more reliable EEG time series, providing a reliable data foundation for subsequent feature extraction. Specifically, the preprocessing module performs operations such as electrode rereference, filtering, independent component analysis, artifact removal, and sample segmentation.

[0084] In the electrode rereference step, a unified reference benchmark needs to be reset for all electrodes. Different reference methods will affect the amplitude and morphology of the EEG signal. Rereference can reduce the deviation introduced by the reference electrode itself and improve the spatial resolution and comparability of the signal. The rereference method used in this embodiment includes average reference or binaural reference.

[0085] In the filtering step, the EEG signal is input into a preset low-pass filter and a high-pass filter. In this embodiment, a 0.1Hz high-pass filter is used to filter out low-frequency drift and baseline changes below 0.1Hz, while a 50Hz low-pass filter is used to filter out high-frequency power supply interference above 50Hz. The purpose of filtering is to retain the effective frequency bands related to brain activity and remove environmental noise and physiological noise.

[0086] In the independent component analysis step, complex electroencephalogram (EEG) signals are decomposed into multiple independent components. Since EEG signals are often interfered with by physiological activities such as eye movements and muscle movements, independent component analysis technology can separate the above-mentioned mixed signal sources. As an existing technology in the field, it helps to identify and remove components that are not related to brain activity.

[0087] The artifact removal step mainly involves eliminating signal artifacts such as electrooculography (EOG) artifacts and electromyography (EMG) artifacts. In this embodiment, by combining the results of independent component analysis, existing technical tools such as ICLabel are used to identify and remove the aforementioned artifact components, ensuring that the final EEG signal is purer.

[0088] In the sample partitioning step, since the amount of EEG signal sample data may be small, directly extracting features for classification can easily lead to model overfitting. Therefore, this embodiment uses a sample partitioning method to expand the dataset. That is, using 2000 sampling points as a data unit, the long-term EEG signal is divided into multiple short samples, thereby effectively increasing sample diversity and improving the model's generalization ability.

[0089] In the feature extraction module, the one-dimensional EEG time series obtained from the preprocessing module is mapped into a first-order weighted horizontal visualization and a second-order weighted degree transfer network, and the weighted degree transfer entropy is calculated through the weighted adjacency matrix of the weighted degree transfer network. MPR statistical complexity measure Thus, a two-dimensional feature vector is formed. It is used to characterize the spatiotemporal dynamics of pure EEG signals. As an indicator of the temporal irreversibility of EEG signals, it provides accurate data input for subsequent depression detection modules and improves detection accuracy.

[0090] The feature extraction module is used to perform first-order weighted level visual image mapping, second-order weighted degree transition network construction, transition probability matrix calculation, and weighted degree transition entropy calculation. MPR statistical complexity measure Calculation and other steps.

[0091] For the first-order weighted horizontal visibility mapping step, the feature extraction module can apply horizontal visibility rules and edge weights to the one-dimensional EEG time series to obtain the first-order weighted horizontal visibility map. The weighted horizontal visibility map is a graph structure generated by mapping the continuous time series of pure EEG signals, which can comprehensively express its geometric shape and dynamic evolution characteristics in the time dimension. The weighted horizontal visibility map can transform the original EEG signal into a graph structure representation with topological structure and dynamic relationships, so as to better reveal the potential nonlinear characteristics and temporal dependencies in the signal. It helps to map the one-dimensional time series into a multi-dimensional graph space, so that the implicit structural patterns and dynamic evolution laws in the EEG signal can be revealed, providing a richer and more interpretable feature basis for subsequent depression detection.

[0092] like Figure 2 As shown, in this embodiment, a first-order weighted horizontal view is constructed, with a length of... EEG time series Mapped to a weighted undirected graph In the formula, For a one-dimensional EEG time series in the 1st The amplitude corresponding to each sampling time. For a set of nodes, For edge set, This is the weight set.

[0093] The first-order weighted level view is constructed according to the following formula:

[0094]

[0095] In the formula, For a one-dimensional EEG time series in the 1st The amplitude corresponding to each sampling time; For a one-dimensional EEG time series in the 1st The amplitude corresponding to each sampling time; For a one-dimensional EEG time series in the 1st The amplitude corresponding to each sampling time; that is, for any two nodes and An edge is established if and only if the values ​​of all data points between them are less than the values ​​of the two data points.

[0096] For each connection, its weight is calculated based on the magnitude difference and time interval of the data points. The definition is as follows:

[0097]

[0098] Calculate the degree value of each node. That is, with nodes The number of directly connected edges forms a degree sequence. ;

[0099] Calculate the node strength of each node. That is, with nodes The sum of the weights of all connected edges constitutes the intensity sequence. .

[0100] For the construction steps of the second-order weighted degree transfer network, the feature extraction module can obtain the degree sequence from the first-order weighted level view. and intensity sequence Construct a weighted degree transition network; where: the degree sequence represents the number of connection edges of each node in the weighted horizontal visibility, and is arranged sequentially according to the order of the nodes in the time series; the intensity sequence in the weighted horizontal visibility represents the sum of the weights of the connection edges of each node in the weighted horizontal visibility, and is arranged sequentially according to the order of the nodes in the time series; based on the degree sequence and intensity sequence, generate the weighted adjacency matrix of the weighted degree transition network by accumulating the instantaneous transition intensities corresponding to all identical state transitions. .

[0101] Traverse the degree sequence Extract all unique degree values ​​that appear in the set. , will set The elements in the definition are nodes of the weighted degree transfer network. ,in, This represents the total number of uniqueness values.

[0102] Traverse the degree sequence in chronological order, if time... degree value equal And at all times degree value equal Then at node With nodes Establish a directed edge between them.

[0103] Calculation time arrive Instantaneous transfer intensity Its calculation formula combines the difference in degree and the difference in strength:

[0104]

[0105] In the formula, They are time points , The degree value; They are time points , The node strength.

[0106] By summing the instantaneous transition intensities corresponding to all identical state transitions, a weighted adjacency matrix for the weighted degree transition network is generated. The calculation formula is as follows:

[0107]

[0108] In the formula, This is the Kronecker function, which has a value of 1 if and only if the two parameters inside the parentheses are equal, otherwise it has a value of 0; The length of the one-dimensional EEG time series.

[0109] For the transition probability matrix calculation step, the feature extraction module can use the weighted adjacency matrix of the weighted degree transition network. Obtain the weighted degree transition probability matrix Weighted adjacency matrix We have already been able to fully characterize the weighted degree transition relationship in the weighted degree transition network. Based on this, by analyzing the matrix... After normalization, the weighted transition probability matrix can be obtained. The above normalization process transforms the original weighted connections into state transition probabilities, which helps to highlight the relative transition preferences between different nodes and provides a more stable probabilistic description for subsequent complexity calculations and feature analysis.

[0110] Weighted transition probability matrix Its definition is as follows:

[0111]

[0112] For weighted transition entropy MPR statistical complexity measure The calculation steps, the feature extraction module can use the weighted transition probability matrix To calculate the weighted degree transfer entropy and the MPR statistical complexity measure.

[0113] Based on probability matrix Calculate the weighted degree transition entropy :

[0114]

[0115] In the formula, Probability matrix The elements in.

[0116] Based on probability matrix Calculating the statistical complexity of MPR It includes the following steps:

[0117] First, calculate the probability matrix. With uniform distribution The Jansen-Shannon divergence between :

[0118]

[0119] Then, calculate the normalized divergence. :

[0120]

[0121] Finally, the MPR complexity is obtained by multiplying the normalized divergence by the weighted degree transition entropy:

[0122]

[0123] In the formula, For probability distribution Shannon entropy; For all elements to be The uniform distribution matrix; sharp distribution matrix With uniform distribution The Jansen-Shannon divergence between them; A sharp distribution matrix is ​​a matrix where a single element is 1 and all other elements are 0. It is used to obtain... The maximum value, to achieve Scale normalization.

[0124] The depression detection module transfers the weighted degree entropy obtained from the feature extraction module. MPR statistical complexity measure Combined into a two-dimensional feature vector The data is then input into a pre-trained machine learning classifier. Specifically, in this embodiment, a Support Vector Machine (SVM) classifier is used, leveraging its superior performance in high-dimensional data to ensure accurate classification of the entropy complexity plane data extracted from each sample. In this embodiment, the parameters of the SVM classifier are set as follows: kernel function parameter kernel=linear, indicating that the kernel function is a linear function, standardization is enabled, and the seed of the random number generator is seed=1.

[0125] After training, the classifier can learn the different distribution patterns of EEG signals in the feature space between patients with depression and healthy controls. Finally, the classifier outputs the detection result to indicate whether the subject has depression.

[0126] The EEG signal sample data used in this embodiment is the MPHC dataset, which is publicly available data. Specifically, the MPHC dataset uses a 21-channel EEG cap to collect 5 minutes of resting-state data with eyes open and closed from the measurement subjects, with a sampling rate of 256Hz. The measurement subjects include 34 users with depression (17 females and 17 males, with an age range of 40.3±12.9 years) and 30 users without depression (9 females and 21 males, with an age range of 38.3±15.6 years).

[0127] When training the depression classifier, in order to increase the diversity of samples and expand the training set, and improve the robustness of subsequent analysis, the preprocessing module can divide the EEG signal sample data. A preset number of EEG data points are selected from the middle part of the EEG signal sample data, and the EEG data is divided into multiple sample data points using a preset data unit. In this embodiment, data with a length of 20,000 points is selected from the middle part of the EEG signal, and the data is divided into units of 2,000 points. The EEG signal of each measurement object can be divided into 10 samples.

[0128] To improve detection accuracy, the preprocessing module can also divide the sample data into a training set and a test set according to a preset ratio, train a classifier based on the training set, and verify the trained classifier based on the test set to obtain a depression classifier.

[0129] To evaluate the robustness and generalization ability of this model, this embodiment employs a ten-fold cross-validation method. Specifically, all EEG signal sample data are divided into training and test sets in a 9:1 ratio. In each round of validation, the features are trained on the training set, and then the trained model is applied to an independent test set for validation. This process is repeated 10 times, each time using a different combination of training and test sets to ensure the accuracy and reliability of the results.

[0130] To prevent potential data leakage during the detection process, this embodiment controls the sample distribution rules, ensuring that samples belonging to the same object cannot appear in both the training and test sets simultaneously. This ensures that the model can effectively classify unseen data in practical applications. The above strategy helps the model better adapt to new data, improving the reliability and generalizability of depression detection.

[0131] To further verify the effectiveness of the method of the present invention in the EEG recognition task of depression, this embodiment conducted a systematic experiment on the classification performance on the MPHC dataset, and used a violin plot to show the distribution of different feature methods on three core indicators.

[0132] Different feature methods include: using only weighted degree transition entropy based on the transition probability matrix, using only MPR statistical complexity based on the transition probability matrix, and a joint feature vector that combines weighted degree transition entropy and MPR statistical complexity.

[0133] Unlike tables that directly list means or extreme values, violin plots can simultaneously display the shape of the data distribution, dispersion, median, extreme values, and fold stability, making them more helpful in observing the overall performance of the model across different experimental folds; by comparing the classification accuracy of the three feature methods (e.g., Figure 3 As shown), sensitivity (such as) Figure 4 (as shown) and specificity (e.g.) Figure 5 The distribution differences shown in the figure can be used to more intuitively evaluate the discriminative ability and stability of each feature method.

[0134] The classification performance results during the testing phase show that, on the MPHC dataset, the joint feature vector proposed in this embodiment achieves the best performance in terms of classification accuracy, sensitivity, and specificity. Compared with single features, the joint feature vector can more comprehensively characterize the multi-scale and nonlinear structural features of EEG signals in depression, thus showing significant advantages in classification effect and stability.

[0135] Figure 6 This is a visualization diagram of the joint feature vector of the MPHC dataset provided in this embodiment. The weighted degree transfer entropy is used as the horizontal axis, and the MPR statistical complexity is used as the vertical axis. Together, they constitute a two-dimensional visualization space for the joint feature vector. Each scatter point corresponds to a feature value of an EEG slice sample of a certain test subject in the MPHC dataset. The first type of sample represents the EEG signal slices of depressed subjects, and the second type represents the EEG signal slices of non-depressed subjects. Figure 6 It can be seen that the two types of samples form distinct distribution regions in the joint feature space, which can clearly distinguish between the results of depression and non-depression.

[0136] The quantitative results of classification performance during the testing phase on the MPHC dataset are shown in the table and bar chart below. Figure 7 As shown: the classification accuracy based solely on weighted degree transfer entropy features is 81.80%; the classification accuracy based solely on MPR statistical complexity features is 79.51%; the classification accuracy of the joint feature vector (weighted degree transfer entropy combined with MPR statistical complexity) method proposed in this application reaches 97.38%, which is significantly better than the single feature method.

[0137] Table 1. Quantitative Comparison of Different Classification Methods

[0138]

[0139] In summary, the joint feature vector method proposed in this embodiment can fully integrate the information of entropy-type features and complexity features, improve the discriminative ability of features, and thus achieve better classification performance in the depression recognition task.

[0140] This embodiment also provides a method for parameter-free EEG feature extraction and depression detection based on a weighted degree transfer network, including:

[0141] The subjects' electroencephalogram (EEG) signals were collected and preprocessed to obtain a one-dimensional EEG time series;

[0142] The weighted horizontal visualization algorithm is used to map a one-dimensional EEG time series into a complex network, and the degree sequence and intensity sequence of each node in the network are extracted.

[0143] Using the deduplicated set of degree values ​​as nodes of the new network, the degree transfer path between adjacent time points is traced, and the product of the degree difference and strength difference between adjacent time points is calculated as the edge weight of the degree transfer path, thereby constructing a weighted degree transfer network.

[0144] The network is normalized to obtain the transition probability matrix, and its Shannon entropy and MPR statistical complexity are calculated as joint feature vectors.

[0145] Using an SVM machine learning classifier to identify depression states.

[0146] The above method does not require pre-setting parameters such as embedding dimension, and through the dynamic difference characteristics of fusion degree and weight, it can quickly and accurately capture abnormal nonlinear dynamic patterns in the EEG signals of depression.

[0147] It should be noted that the data involved in this invention (including but not limited to historical operating data) are all data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0148] The databases involved in the various embodiments provided by this invention may include at least one of relational databases and non-relational databases. Non-relational databases may include distributed databases based on blockchain, etc., and are not limited thereto.

[0149] The processors involved in the various embodiments provided by this invention may be general-purpose processors, central processing units, graphics processors, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence processors, etc., and are not limited thereto.

[0150] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of the above technical features, they should be considered within the scope of the present invention.

[0151] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and all such modifications and improvements fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for processing electroencephalogram (EEG) signals based on a weighted degree transfer network, characterized in that, Includes the following steps: The raw EEG signals were preprocessed to remove noise and artifacts, resulting in a one-dimensional EEG time series. Based on the weighted horizontal visualization algorithm, the one-dimensional EEG time series is mapped into a first-order weighted undirected graph; Calculate the degree and strength of each node in a first-order weighted undirected graph, and generate a degree sequence and strength sequence arranged in chronological order. The specific steps include the following: The length is time series Mapped to a weighted undirected graph In the formula, For a one-dimensional EEG time series in the 1st The amplitude corresponding to each sampling time. For a set of nodes, For edge set, For weight set; For any node in the weighted undirected graph and If located at node and All data point values ​​between are less than and Then at node and Establish edges between them; For each established edge, calculate its edge weight. ; Calculate the degree value of each node. The degree sequence is obtained. The degree value For nodes The number of directly connected edges; Calculate the node strength of each node. The intensity sequence is obtained. The node strength For nodes The sum of the weights of all connected edges; Based on the degree sequence and intensity sequence, a weighted degree transfer network is constructed; The weighted adjacency matrix of the weighted degree transition network is normalized to obtain the transition probability matrix; Based on the aforementioned transition probability matrix, the weighted degree transition entropy and MPR statistical complexity are calculated to form a two-dimensional feature vector, which serves as a joint feature characterizing the nonlinear dynamic properties of EEG signals. The two-dimensional feature vector is input into a pre-trained machine learning classifier to obtain the processing result.

2. The method according to claim 1, characterized in that, The edge weight The calculation formula is: ； In the formula, For a one-dimensional EEG time series in the 1st The amplitude corresponding to each sampling time; For a one-dimensional EEG time series in the 1st The amplitude corresponding to each sampling time.

3. The method according to claim 1, characterized in that, The construction of the weighted degree transfer network includes the following steps: Traverse the degree sequence Extract all uniqueness values ​​that have appeared in the data to form a uniqueness value set. and the set of uniqueness values The elements in the definition are the nodes of the weighted degree transfer network. ,in The total number of the uniqueness values; Traverse the degree sequence in chronological order. If at any time degree value equal And at all times degree value equal Then at node With nodes Establish a directed edge between them; Calculation time At the time Instantaneous transfer intensity ; By summing the instantaneous transition intensities corresponding to all identical state transitions, a weighted adjacency matrix of the weighted degree transition network is generated. , Weighted adjacency matrix The elements in.

4. The method according to claim 3, characterized in that, The instantaneous transfer intensity The calculation formula is: ； In the formula, For a moment The degree value; For a moment The degree value; For a moment Node strength; For a moment The node strength.

5. The method according to claim 3, characterized in that, The weighted adjacency matrix elements in The calculation formula is: ； In the formula, This is the Kronecker function, which has a value of 1 if and only if the two parameters inside the parentheses are equal, otherwise it has a value of 0; The length of the one-dimensional EEG time series.

6. The method according to claim 1, characterized in that, The calculation of the weighted degree transition entropy includes the following steps: The weighted adjacency matrix Normalization yields the weighted degree transition probability matrix. The formula is as follows: ； Based on the weighted transition probability matrix The weighted degree transfer entropy H is calculated using the following formula: 。 7. The method according to claim 6, characterized in that, The statistical complexity of MPR includes the following steps: Calculate the weighted transition probability matrix With uniform distribution matrix The Jansen-Shannon divergence between ; Calculate normalized divergence The MPR complexity is obtained by multiplying it by the weighted degree transition entropy: ； In the formula, For all elements to be The uniform distribution matrix.

8. The method according to claim 1, characterized in that, The machine learning classifier used is a support vector machine.

Images