Depression detection method based on EEG channel selection and multi-dimensional feature fusion
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-21
- Publication Date
- 2026-04-10
AI Technical Summary
Existing EEG-based methods for detecting depression rely on doctors' subjective experience, have low accuracy, and the computational complexity of brain functional networks constructed by high-density EEG is high, making it difficult to effectively mine spatial information of abnormal changes in the brain topology of patients with depression.
An improved k-means algorithm is used to select key EEG channels. Multidimensional features are extracted by constructing brain functional networks and super-brain functional networks. A multidimensional feature fusion strategy based on standard deviation is combined to reduce computational complexity and improve detection accuracy.
It significantly improves the accuracy of depression detection, reduces computational complexity, and fully mines depression-related information in EEG signals through multi-dimensional feature complementarity.
Smart Images
Figure CN121817885A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of depression detection technology, and more specifically to a depression detection method based on EEG channel selection and multidimensional feature fusion. Background Technology
[0002] Depression (DP) is a common mental illness, characterized by persistent low mood and loss of interest. According to the World Health Organization, approximately 5% of the global population suffers from depression, and more than 700,000 DP patients die by suicide each year. The internationally recognized most effective treatment for DP is early identification and intervention to prevent irreversible damage to brain function over time. Traditional DP diagnostic methods based on scales and face-to-face interviews rely on the physician's subjective experience and the veracity of information provided by the patient, resulting in a clinical diagnostic accuracy rate often below 50%. Therefore, exploring objective, accurate, and efficient DP testing methods is of significant value and importance.
[0003] With the rapid development of neuroimaging technology over the past few decades, EEG has been widely used in the detection of mental disorders such as DP due to its advantages of being non-invasive, having high temporal resolution, and being easy to operate. However, EEG-based DP detection research often treats the corresponding electrodes of the channels as isolated nodes, ignoring the correlation between them, making it difficult to discover the spatial information changes caused by abnormal changes in the brain topology of DP patients.
[0004] In recent years, with the increasing attention paid to graph-based brain function networks (BFNs), EEG-based BFNs have yielded continuous results in depressive disorder (DP) research. However, BFNs constructed from high-density EEG often have a large number of redundant channels, leading to increased computational complexity. Channel selection technology, as one of the effective methods to reduce information redundancy and computational complexity, has therefore attracted the attention of researchers. Existing research (BT Zhang, et al., “Brain functional networks based on resting-state EEG data for major depressive disorder analysis and classification”, IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2021, 29(1):215-229.) has pointed out that DP is closely related to abnormal changes in neural signals. Single-dimensional features are insufficient to fully extract the valuable information contained in the EEG signals of DP patients. Therefore, multi-dimensional feature fusion is an effective way to solve this problem.
[0005] To address the aforementioned problems, this invention provides a method for suppressing DP detection based on EEG channel selection and multi-dimensional feature fusion. By using EEG channel selection technology, the computational complexity of the DP detection process is reduced, and by using multi-feature fusion technology, complementary information from different dimensions is achieved, thereby improving the accuracy of DP detection. Summary of the Invention
[0006] The purpose of this invention is to provide a depression detection method based on EEG channel selection and multidimensional feature fusion, which can effectively improve the accuracy of depression detection while reducing computational complexity.
[0007] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:
[0008] Step 1: Input the raw EEG channel data X = {X1, X2, ..., X} n};
[0009] Step 2: Improve k-means for selecting key EEG channels from high-density EEG channels;
[0010] Step 3: Extract low-dimensional time-domain (LDT) features from key EEG channels that are closely related to depression;
[0011] Step 4: Construct a brain functional network (BFN) using key EEG channels. The electrodes corresponding to the channels serve as nodes in the BFN, the synchronicity between nodes is represented by edges, and the phase lag index (PLI) is used as the carrier for quantifying synchronicity. The calculation process is as follows:
[0012]
[0013] In the formula, It is an EEG signal x i The instantaneous phase of (t), It is x i The Hilbert transform of (t) is used, where sign is the sign function; after the BFN is constructed, its low-dimensional spatial domain (LDS) features closely related to depression are extracted;
[0014] Step 5: Construct a superbrain functional network (HBFN) using key EEG channels and extract its high-dimensional spatial (HDS) features closely related to depression;
[0015] Step 6: Design a multi-dimensional feature fusion strategy based on standard deviation to fuse low-dimensional time-domain features, low-dimensional spatial features, and high-dimensional spatial features to achieve high-accuracy DP detection.
[0016] Furthermore, the process of improving k-means in step two and selecting key EEG channels from high-density EEG channels is as follows:
[0017] (1) The initial number of clusters was determined using the Kalinsky-Harabas criterion (CH):
[0018]
[0019] In the formula, SS b It is the inter-cluster separation, SS w It represents the intra-cluster separation, where N is the number of EEG channels and k is the number of clusters.
[0020]
[0021] (2) Determine the initial cluster centers of the k clusters:
[0022]
[0023] In the formula, i∈[1,k], d is the damping coefficient, B(i) represents the set of channels related to EEG channel i, sim(j,i) represents the correlation between channel j and channel i, PR(j) represents the PageRank value of channel j, and L(j) represents the number of channels connected to channel j.
[0024] (3) Using the k initial cluster centers determined in the previous step, for X = {X1, X2, ..., X...} n Clustering is performed to obtain k clusters C. i :
[0025]
[0026] Where i = 1, 2, ..., k, m ≤ n, ∪ C i =X, calculate the mean of the EEG channels within each cluster:
[0027]
[0028] Calculate the distance between each channel and the mean of the corresponding cluster.
[0029]
[0030] calculate Mean and standard deviation:
[0031]
[0032] Establish an adaptive threshold model:
[0033] θ i_adapt =|μ i -σ i | (10)
[0034] An adaptive thresholding model is used to select EEG channels within each cluster whose distance is less than the adaptive threshold as key channels:
[0035]
[0036] In the formula, S i S represents the key EEG channels selected within each cluster, while S represents all key EEG channels.
[0037] Furthermore, the HBFN construction process in step five is as follows:
[0038] We use lasso to solve a sparse linear regression model to construct the HBFN, which represents the connection relationships between nodes. The optimization objective function is:
[0039]
[0040] In the formula, X m This is the EEG data from the m-th channel, A m It is a matrix composed of all EEG channel data, where the value of the element in the m-th column is 0. m is a coefficient vector, representing the weight vector of other channels acting on the m-th channel when generating a hyperedge centered on the node corresponding to the m-th channel. ||*||2 is L2 regularization, and ||*||1 is L1 regularization. λ is a regularization parameter used to control the sparsity of HBFN; based on the introduction of a specific regularization parameter λ in (14), the hyperedge e i Given the i-th channel and coefficient vector a i The channels with corresponding positive elements are composed of coefficient vectors that form the correlation matrix H of HBFN:
[0041] H = [α1, α2, ..., α] N (14)
[0042] The elements in the correlation matrix H are represented as follows:
[0043]
[0044] In the formula, v i It is a node in HBFN, e j It is a superedge in HBFN, when v i ∈e j At that time, H ij =1, otherwise H ij =0.
[0045] Furthermore, the implementation process of the multi-dimensional feature fusion strategy based on standard deviation in step six is as follows:
[0046] (1) Eliminate the dimensionality differences between different features and normalize LDT features, LDS features, and HDS features:
[0047]
[0048] In the formula, x' is the normalized eigenvalue, x is the eigenvalue before normalization, and x' is the normalized eigenvalue. min It is the smallest eigenvalue before normalization, x max The largest eigenvalue before normalization;
[0049] (2) Calculate the normalized standard deviation of each feature:
[0050]
[0051] In the formula, This is the normalized mean of each feature. The weight coefficients for each feature are calculated as follows:
[0052]
[0053] (3) Multidimensional feature fusion based on the following formula:
[0054] F fus ={w1f LDT_1 w2f LDT_2 ,…,w i f LDS_1 ,w i+1 f LDS_2 ,…,w n f HDS_3} (19)
[0055] Compared with the prior art, the beneficial effects of the present invention are:
[0056] 1. Traditional k-means clustering is sensitive to the initial cluster centers and the number of clusters, and is prone to getting trapped in local optima. This invention determines the optimal number of clusters using the Kalinsky-Hallabus criterion (CH) and selects the initial cluster centers based on the PageRank value, significantly improving the stability and accuracy of the clustering results, thereby more effectively screening out key channels from high-density EEG channels.
[0057] 2. By calculating the mean and standard deviation of channel distances within a cluster, an adaptive threshold model is designed to dynamically select channels with distances less than a threshold as key EEG channels. This method avoids the limitations of fixed thresholds, enabling more accurate identification of key channels related to dynamic programming (DP), while reducing computational complexity.
[0058] 3. This invention not only extracts low-dimensional temporal features (LDT) from key EEG channels, but also extracts low-dimensional spatial features (LDS) and high-dimensional spatial features (HDS) by constructing a brain functional network (BFN) and a super-brain functional network (HBFN), respectively. Through a multi-dimensional feature fusion strategy based on standard deviation, the complementary nature of features from different dimensions is achieved, fully mining depression-related information in EEG signals and significantly improving the accuracy of depression detection (DP). Attached Figure Description
[0059] Figure 1 This is a flowchart of the method of the present invention;
[0060] Figure 2 These are the CH values corresponding to different cluster numbers in the embodiments of the present invention;
[0061] Figure 3 These are the cluster centers obtained based on the PageRank values corresponding to each EEG channel in this embodiment of the invention;
[0062] Figure 4 This invention improves the k-means clustering and EEG channel selection results in embodiments of the invention.
[0063] Figure 5 This refers to the HBFN and its corresponding correlation matrix in an embodiment of the present invention.
[0064] Figure 6 A comparison of DP detection performance using different features in embodiments of the present invention;
[0065] Figure 7 This invention provides a comparison of DP detection performance with different numbers of channels in embodiments of the invention.
[0066] Figure 8 This is a comparison of the time complexity before and after channel selection in an embodiment of the present invention. Detailed Implementation
[0067] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.
[0068] Please see Figure 1 As shown, this invention provides a technical solution: a DP detection method based on EEG channel selection and multidimensional feature fusion, the method comprising the following steps:
[0069] Step 1: Input the raw 128-channel EEG data X = {X1, X2, ..., X} n},like Figure 1 As shown in (a).
[0070] Step 2: Improve k-means for selecting key EEG channels from a 128-channel high-density EEG channel (see...) Figure 1 (a)), the specific process is as follows:
[0071] (1) The initial number of clusters was determined using the Kalinsky-Harabas criterion (CH):
[0072]
[0073] In the formula, SS b It is the inter-cluster separation, SS w This represents the intra-cluster separation, where N is the number of EEG channels (N = 128), and k is the number of clusters.
[0074]
[0075] Figure 2 CH values corresponding to different numbers of clusters are provided. Therefore, the number of clusters selected in this embodiment is k=4.
[0076] (2) Determine the initial cluster centers for the four clusters:
[0077]
[0078] In the formula, i∈[1,k], d is the damping coefficient, B(i) represents the set of channels related to EEG channel i, sim(j,i) represents the correlation between channel j and channel i, PR(j) represents the PageRank value of channel j, and L(j) represents the number of channels connected to channel j. Figure 3 PageRank values for each EEG channel are provided. The first four EEG channels are selected as the initial cluster centers, i.e., EEG channels numbered 22, 37, 65 and 115 are selected.
[0079] (3) Using four initial cluster centers of 22, 37, 65, and 115, for X = {X1, X2, ..., X...} n Clustering was performed, resulting in 4 clusters C. i ,like Figure 4 As shown in (a).
[0080]
[0081] Where i = 1, 2, ..., k, m ≤ n, ∪ C i =X, calculate the mean of the EEG channels within each cluster:
[0082]
[0083] Calculate the distance between each channel and the mean of the corresponding cluster.
[0084]
[0085] calculate Mean and standard deviation:
[0086]
[0087] Establish an adaptive threshold model:
[0088] θ i_adapt =|μ i -σ i | (10)
[0089] An adaptive thresholding model is used to select EEG channels within each cluster whose distance is less than the adaptive threshold as key channels:
[0090]
[0091] In the formula, S i These are the key EEG channels selected within each cluster, and S represents all 58 key EEG channels, such as... Figure 4 As shown in (b).
[0092] Step 3: Extract low-dimensional time-domain (LDT) features closely related to depression from key EEG channels, such as differential entropy (DE) and power spectral density (PSD). Figure 1 As shown in (b).
[0093] Step 4: Construct a brain functional network (BFN) using key EEG channels. The electrodes corresponding to the channels serve as nodes in the BFN, the synchronicity between nodes is represented by edges, and the phase lag index (PLI) is used as the carrier for quantifying synchronicity. The calculation process is as follows:
[0094]
[0095] In the formula, It is an EEG signal x i The instantaneous phase of (t), It is x i The Hilbert transform of (t), where sign is the sign function; after the BFN is constructed, its low-dimensional spatial domain (LDS) features closely related to depression are extracted, such as the feature path length (CPL) and clustering coefficient (CC), as follows. Figure 1 As shown in (b).
[0096] Step 5: Construct the Superbrain Functional Network (HBFN) using key EEG channels. The HBFN construction process is as follows:
[0097] We use lasso to solve a sparse linear regression model to construct the HBFN, which represents the connection relationships between nodes. The optimization objective function is:
[0098]
[0099] In the formula, X m This is the EEG data from the m-th channel, A m It is a matrix composed of all EEG channel data, where the value of the element in the m-th column is 0. m is a coefficient vector, representing the weight vector of other channels acting on the m-th channel when generating a hyperedge centered on the node corresponding to the m-th channel. ||*||2 is L2 regularization, and ||*||1 is L1 regularization. λ is a regularization parameter used to control the sparsity of HBFN; based on the introduction of a specific regularization parameter λ in (14), the hyperedge e i Given the i-th channel and coefficient vector a i The channels with corresponding positive elements are composed of coefficient vectors that form the correlation matrix H of HBFN:
[0100] H = [α1, α2, ..., α] N (14)
[0101] The elements in the correlation matrix H are represented as follows:
[0102]
[0103] In the formula, v i It is a node in HBFN, e j It is a superedge in HBFN, when v i ∈e j At that time, H ij =1, otherwise H ij =0. Figure 5 The correlation matrix and corresponding HBFN are provided.
[0104] High-dimensional spatial domain (HDS) features closely related to depression were extracted from HBFN, namely the three different types of clustering coefficients of HBFN: HCC1(v), HCC2(v), and HCC2(v), such as... Figure 1 As shown in (b).
[0105] Step Six: Design a multi-dimensional feature fusion strategy based on standard deviation, fusing low-dimensional time-domain features, low-dimensional spatial-domain features, and high-dimensional spatial-domain features to achieve high-accuracy depression detection. The specific process is as follows:
[0106] (1) Eliminate the dimensional differences between different features, normalize LDT features DE and PSD, LDS features CPL and CC, and HDS features HCC1(v), HCC2(v), and HCC2(v), and calculate the formula as follows:
[0107]
[0108] In the formula, x' is the normalized eigenvalue, x is the eigenvalue before normalization, and x' is the normalized eigenvalue. min It is the smallest eigenvalue before normalization, x max The largest eigenvalue before normalization;
[0109] (2) Calculate the normalized standard deviation of each feature:
[0110]
[0111] In the formula, This is the normalized mean of each feature. The weight coefficients for each feature are calculated as follows:
[0112]
[0113] Tables 1 and 2 provide the standard deviations and weighting coefficients for the DP group and the normal control group, respectively.
[0114] Table 1 Standard deviation and weighting coefficients of DP group
[0115]
[0116] Table 2 Standard deviation and weighting coefficient of the normal control group
[0117]
[0118] (3) Multidimensional feature fusion based on the following formula:
[0119] F fus ={w1f LDT_1 w2f LDT_2 ,…,w i f LDS_1 ,w i+1 f LDS_2 ,…,w n f HDS_3} (19)
[0120] The embodiments use the KNN classifier to verify the detection performance of the DP detection method provided by the present invention. Existing research (H.Cai, et al., “A case-based reasoning model for depression based on three-electrode EEG data”, IEEE Transactions on Affective Computing, 2018, 11(3):383-392. and H.Cai, et al., “Feature-level fusion approaches based on multimodal EEG data for depression recognition”, Information Fusion, 2020, 59:127-138.) shows that KNN generally outperforms other classifiers in DP detection. In the embodiments, the KNN parameters are set to n_neighbor = 3, algorithm = 'ball_tree', and metric = 'euclideanDis'.
[0121] Figure 6 The performance comparison of DP detection under different feature usage strategies is provided. LDT-DF represents using only low-dimensional temporal features, LDS-DF represents using only low-dimensional spatial features, HDS-DF represents using only high-dimensional spatial features, UMFF represents direct multi-dimensional feature fusion, and WMFF represents multi-dimensional feature fusion based on standard deviation. It can be seen that the multi-dimensional feature fusion strategy based on standard deviation is superior to other feature usage strategies in terms of accuracy, F1, AUC and other evaluation indicators.
[0122] Figure 7 A comparison of DP detection performance with different numbers of channels is provided. The line graph shows the DP detection accuracy under different numbers of channels, and the channel topography map shows the channel distribution corresponding to the inflection point of detection performance. Blue indicates that the channel is not selected, with a value of 0, and red indicates that the channel is selected, with a value of 1. It can be seen that after the number of channels reaches 33, the detection performance improves rapidly and continuously. In this embodiment, the highest DP detection accuracy of 96.14% is reached when 58 channels are selected. 58 channels is the inflection point of detection performance. Before the number of channels reaches 91, the detection performance shows a downward trend. After that, the decline in detection performance stops and begins to rise slowly until all 128 channels are reached.
[0123] Figure 8 A comparison of time complexity before and after channel selection is provided. It can be seen that, compared with 128 full EEG channels, the 58 EEG channels selected in this invention significantly improve detection accuracy and AUC while effectively reducing runtime.
[0124] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
[0125] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.
Claims
1. A method for detecting depression based on EEG channel selection and multidimensional feature fusion, characterized in that, The specific steps are as follows: Step 1: Input the raw EEG channel data X = {X1, X2, ..., X} n }; Step 2: Improve k-means for selecting key EEG channels from high-density EEG channels; Step 3: Extract low-dimensional time-domain (LDT) features from key EEG channels that are closely related to depression; Step 4: Construct a brain functional network (BFN) using key EEG channels. The electrodes corresponding to the channels serve as nodes in the BFN, the synchronicity between nodes is represented by edges, and the phase lag index (PLI) is used as the carrier for quantifying synchronicity. The calculation process is as follows: In the formula, It is an EEG signal x i The instantaneous phase of (t), It is x i The Hilbert transform of (t) is used, where sign is the sign function; after the BFN is constructed, its low-dimensional spatial domain (LDS) features closely related to depression are extracted; Step 5: Construct a superbrain functional network (HBFN) using key EEG channels and extract its high-dimensional spatial (HDS) features closely related to depression; Step 6: Design a multidimensional feature fusion strategy based on standard deviation to integrate low-dimensional time-domain features, low-dimensional spatial features, and high-dimensional spatial features to achieve high-accuracy depression detection.
2. The method for detecting depression based on EEG channel selection and multidimensional feature fusion according to claim 1, characterized in that, The process of improving k-means and selecting key EEG channels from high-density EEG channels in step two is as follows: (1) The initial number of clusters was determined using the Kalinsky-Harabas criterion (CH): In the formula, SS b It is the inter-cluster separation, SS w It represents the intra-cluster separation, where N is the number of EEG channels and k is the number of clusters. (2) Determine the initial cluster centers of the k clusters: In the formula, i∈[1,k], d is the damping coefficient, B(i) represents the set of channels related to EEG channel i, sim(j,i) represents the correlation between channel j and channel i, PR(j) represents the PageRank value of channel j, and L(j) represents the number of channels connected to channel j. (3) Using the k initial cluster centers determined in the previous step, for X = {X1, X2, ..., X...} n Clustering is performed to obtain k clusters C. i : Where i = 1, 2, ..., k, m ≤ n, ∪ C i =X, calculate the mean of the EEG channels within each cluster: Calculate the distance between each channel and the mean of the corresponding cluster. calculate Mean and standard deviation: Establish an adaptive threshold model: i i_adapt =|μ i -s i | (10) Using an adaptive thresholding model, channels within each cluster whose distance is less than the adaptive threshold are selected as key EEG channels: In the formula, S i S represents the key EEG channels selected within each cluster, while S represents all key EEG channels.
3. The method for detecting depression based on EEG channel selection and multidimensional feature fusion according to claim 1, characterized in that, The HBFN construction process in step five is as follows: We use lasso to solve a sparse linear regression model to construct the HBFN, which represents the connection relationships between nodes. The optimization objective function is: In the formula, X m This is the EEG data from the m-th channel, A m It is a matrix composed of all EEG channel data, where the value of the element in the m-th column is 0. m is a coefficient vector, representing the weight vector of other channels acting on the m-th channel when generating a hyperedge centered on the node corresponding to the m-th channel. ||*||2 is L2 regularization, and ||*||1 is L1 regularization. λ is a regularization parameter used to control the sparsity of HBFN; based on the introduction of a specific regularization parameter λ in (14), the hyperedge e i Given the i-th channel and coefficient vector a i The channels with corresponding positive elements are composed of coefficient vectors that form the correlation matrix H of HBFN: H＝[α1,α2，…,α N ] (14) The elements in the correlation matrix H are represented as follows: In the formula, v i It is a node in HBFN, e j It is a superedge in HBFN, when v i ∈e j At that time, H ij =1, otherwise H ij =0.
4. The method for detecting depression based on EEG channel selection and multidimensional feature fusion according to claim 1, characterized in that, The implementation process of the multidimensional feature fusion strategy based on standard deviation in step six is as follows: (1) Eliminate the dimensionality differences between different features and normalize LDT features, LDS features, and HDS features: In the formula, x' is the normalized eigenvalue, x is the eigenvalue before normalization, and x' is the normalized eigenvalue. min It is the smallest eigenvalue before normalization, x max The largest eigenvalue before normalization; (2) Calculate the normalized standard deviation of each feature: In the formula, This is the normalized mean of each feature. The weight coefficients for each feature are then calculated. (3) Fusing multidimensional features based on the following formula: F fus ={w1f LDT_1 ,w2f LDT_2 ,…,w i f LDS_1 ,w i+1 f LDS_2 ,…,w n f HDS_3 } (19)