SG-Lasso-based depression automatic detection method

Abstract

The invention discloses an SG-Lasso-based depression (DP) automatic detection method, which is characterized in that a super-brain function network (HBFN) is constructed and optimized through electroencephalogram (EEG) signals, and efficient DP detection is realized. The method comprises the following steps: inputting a multi-channel EEG signal, constructing an HBFN by using SG-Lasso, establishing a mathematical model based on sparsity and stability, optimizing the HBFN at different frequency bands, extracting different types of clustering coefficients in the HBFN, and checking and evaluating differences among clustering coefficient groups through nonparametric arrangement at different frequency bands. And selecting clustering coefficients with significant differences among the groups as potential markers for DP detection, and finally verifying the effectiveness of the potential markers on DP automatic detection by adopting a classifier. According to the method, the HBFN is optimized in combination with sparsity and stability, the accuracy of DP detection is remarkably improved, and a new technical support is provided for clinical DP auxiliary diagnosis.

Description

Technical Field

[0001] This invention relates to the field of mental illness detection technology, and in particular to an automated method for detecting depression based on SG-Lasso. Background Technology

[0002] Difficulty developing (DP) is a common mental disorder, affecting approximately 350 million people worldwide, with about one million suicides annually due to DP. Given the high prevalence and disabling nature of DP, exploring early identification and intervention is crucial to prevent irreversible brain damage over time.

[0003] Traditional face-to-face interview-based DP detection methods are time-consuming and labor-intensive, heavily reliant on the doctor's experience and the authenticity of the information provided by the patient. With the development of biomedical engineering, neuroimaging, and machine learning technologies, DP detection methods based on physiological data have gained increasing attention. Compared to other physiological signals, EEG is widely used in BFN construction and automated DP detection due to its advantages such as high temporal resolution, non-masking properties, and non-invasiveness. However, most BFNs capture the connection information between paired EEG nodes, reflecting only second-order relationships between nodes and neglecting higher-order relationships. The human brain, as one of the most complex systems in the universe, is essentially an interaction between multiple neuronal nodes. Therefore, exploring the relationships between multiple nodes in the brain through HBFN has significant application value.

[0004] Typical HBFN construction methods based on minimum absolute shrinkage and the selection operator (Lasso) only consider the sparsity within ROIs, while group Lasso (G-Lasso) only considers the sparsity between ROIs. Therefore, this invention provides an HBFN construction method based on sparse G-Lasso (SG-Lasso). Fully connected HBFNs typically increase the computational cost of subsequent machine learning algorithms (DP detection). To address this, this invention provides a mathematical model based on the relationship between sparsity and stability to optimize HBFNs, thereby reducing computational complexity and ensuring that detection accuracy is improved while effectively reducing computational complexity. Summary of the Invention

[0005] The purpose of this invention is to provide an automatic DP detection method based on SG-Lasso, which can effectively improve the accuracy of DP detection while reducing the computational complexity of the algorithm, and thus be used in clinical DP auxiliary diagnosis.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an automatic detection method for depression based on SG-Lasso, the steps of which are as follows:

[0007] Step 1: Input EEG signal X = [x1, x2, ..., x N ];

[0008] Step 2: Construct the Superbrain Functional Network (HBFN) based on SG-Lasso;

[0009] Step 3: Use one-way ANOVA to explore the frequency bands where there are significant differences between the DP group and the normal control (NC) group;

[0010] Step 4: Establish a mathematical model based on the relationship between sparsity and stability to optimize HBFN in different frequency bands;

[0011] Step 5: Extract the three types of clustering coefficients for different frequency bands in HBFN, and use non-parametric permutation test to evaluate the inter-group differences of the three clustering coefficients to explore potential markers for DP detection;

[0012] Step 6: Use the KNN classifier to evaluate the effectiveness of latent markers for automatic DP detection.

[0013] Furthermore, the specific process of constructing HBFN in step two is as follows:

[0014] SG-Lasso is used to represent the connectivity between nodes, i.e., the optimization objective function is:

[0015]

[0016] In the formula, x n This is the EEG data from the nth channel, X n It is a matrix composed of all EEG channel data, where the value of the element in the nth column is 0. n This is a coefficient vector representing the weight vector of the hyperedge generated centered on the nth channel, reflecting the degree to which it is influenced by the EEG signals of other channels. ||*||1 is L1 regularization, ||*||2 is L2 regularization, and λ is a regularization parameter. λ1 is used to adjust the sparsity within the region of interest (ROI), and λ2 is used to adjust the sparsity between ROIs. G i It is the i-th ROI.

[0017] Super-edge 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 HBFN correlation matrix H:

[0018] H = [α1, α2, ..., α] N (2)

[0019] The elements in the correlation matrix H are represented as follows:

[0020]

[0021] 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.

[0022] Furthermore, in step four, a mathematical model based on the relationship between sparsity and stability is established. The process of optimizing HBFN in different frequency bands is as follows:

[0023] (1) Determine the optimal network density range and establish a mathematical model for the average node degree K, number of nodes N, and number of edges M of HBFN:

[0024] 2M=K×N (4)

[0025] Establish a mathematical model for the network density ρ in relation to the number of nodes N and the number of edges M:

[0026]

[0027] The mathematical relationship between the average node degree K and the network density ρ is established using equations (4) and (5):

[0028]

[0029] Studies have shown that the optimal network is constructed when the average node degree is greater than the natural logarithm of the number of nodes and the network density is less than 50%.

[0030] K > ln(N) (7)

[0031] From equations (6) and (7), we can see that:

[0032] K=ρ×(N-1)>ln (N) (8)

[0033] That is, the lower limit of HBFN network density is ln(N) / (N-1), and the HBFN network density range is [ln(N) / (N-1), 50%].

[0034] (2) Establishing the optimal balance between network sparsity and stability. Existing research shows that the human brain is a stable small-world network. The Small-World Index (SWI) is used as an indicator to quantify small-world networks, and its calculation formula is as follows:

[0035]

[0036] In the formula, CC and CC random These are the clustering coefficients of the original HBFN and the corresponding random network, CPL and CPL, respectively. randomThese are the feature path lengths of the original HBFN and the corresponding random network, respectively.

[0037] According to the definition of SWI, when SWI is greater than 1, the network exhibits small-world characteristics, and the larger the SWI, the stronger the network stability. Therefore, with a step size of 0.01, we search for the network density ρ corresponding to the maximum SWI within the range of network densities [ln(N) / (N-1), 50%] of the original HBFN in different frequency bands, and use this network density as a threshold to optimize the original HBFN in different frequency bands.

[0038] Furthermore, in step five, the three clustering coefficients extracted from HBFN refer to HCC1, HCC2, and HCC3. Their calculation process and meaning are as follows:

[0039]

[0040] In the formula, N(v) represents the neighboring nodes of node v. When the superedge only contains u and t, but not v, otherwise, HCC1 measures the number of connections that node v does not facilitate between its neighboring nodes.

[0041]

[0042] In the formula, when the hyperedge contains u, t, and v simultaneously, I(u,t,v) = 1; otherwise, I(u,t,v) = 0. HCC2 measures the number of connections that node v facilitates between its neighboring nodes.

[0043]

[0044] In the formula, E(v) represents the set containing the hyperedges of node v, and HCC3 measures the number of hyperedges passing through node v.

[0045] Compared with the prior art, the beneficial effects of the present invention using the above technical solution are as follows:

[0046] (1) BFN can only reflect the second-order connection relationship between paired nodes, while the present invention can capture the higher-order interaction between multiple nodes in the brain by constructing HBFN, which can more comprehensively reflect the complexity of brain activity and thus improve the accuracy of DP detection.

[0047] (2) Existing Lasso or G-Lasso-based methods only focus on the sparsity within or between ROIs of HBFN, while this invention adopts the SG-Lasso HBFN method, which takes into account both the sparsity within and between ROIs, and establishes a mathematical model of sparsity and stability. By optimizing HBFN, this invention effectively reduces the computational burden brought by fully connected networks, and effectively reduces computational complexity while ensuring detection performance.

[0048] (3) This invention extracts three clustering coefficients (HCC1, HCC2, and HCC3) of HBFN in different frequency bands, and evaluates the differences between groups through nonparametric permutation tests, screening out clustering coefficients with significant differences as potential markers for DP detection. This multi-band, multi-index analysis method enhances the distinguishing ability of features, significantly improves the accuracy of DP detection, and provides strong technical support for clinical diagnosis. Attached Figure Description

[0049] Figure 1 This is a flowchart of the method of the present invention;

[0050] Figure 2 The frequency bands representing significant differences between groups in this embodiment of the invention;

[0051] Figure 3 This illustrates the relationship between the small-world index and network density in an embodiment of the present invention.

[0052] Figure 4 This is a comparison of the differences in the three types of clustering coefficients between two groups in an embodiment of the present invention. Detailed Implementation

[0053] 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.

[0054] Please see Figure 1 As shown, this invention provides a technical solution: an automatic detection method for depression based on SG-Lasso, the steps of which are as follows:

[0055] Step 1: Input 128-channel EEG signal X = [x1, x2, ..., x N Please refer to [link / reference]. Figure 1 ;

[0056] Step 2: Construct a Superbrain Functional Network (HBFN) based on SG-Lasso. EEG channels represent network nodes, and the connections between nodes are represented using SG-Lasso, i.e., optimizing the objective function:

[0057]

[0058] In the formula, x n This is the EEG data from the nth channel, X n It is a matrix composed of all EEG channel data, where the value of the element in the nth column is 0. nThis is a coefficient vector representing the weight vector of the hyperedge generated centered on the nth channel, reflecting the degree to which it is influenced by the EEG signals of other channels. ||*||1 is L1 regularization, ||*||2 is L2 regularization, and λ is a regularization parameter. λ1 is used to adjust the sparsity within the region of interest (ROI), and λ2 is used to adjust the sparsity between ROIs. G i It is the i-th ROI.

[0059] Super-edge 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 HBFN correlation matrix H:

[0060] H = [α1, α2, ..., α] N (2)

[0061] The elements in the correlation matrix H are represented as follows:

[0062]

[0063] 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.

[0064] Step 3: One-way ANOVA was used to explore the frequency bands with significant differences between the DP group and the normal control (NC) group. Studies have shown (X. Li et al., “A Resting-State Brain Functional Network Study in MDD Based on Minimum Spanning Tree Analysis and the Hierarchical Clustering”, Complexity, 2017, 9514369.) that low-frequency EEG signals are closely related to DP. This invention compares the low-frequency absolute brain power between the DP group and the NC group, and uses one-way ANOVA to discover the frequency bands with significant differences between the two groups. The results are as follows: Figure 2 As shown, Figure 2 (a) provides the absolute power between the two groups. Figure 2 (b) The results of the one-way ANOVA comparison of the absolute power between the two groups are provided. ** indicates that there is a difference, ** indicates that the difference is significant, and *** indicates that the difference is extremely significant. As can be seen from the figure, the significant difference between the two groups exists in the delta, theta, alpha, and beta1 frequency bands of HBFN, while there is no difference in other frequency bands.

[0065] Step 4: Establish a mathematical model based on the relationship between sparsity and stability to optimize HBFN in different frequency bands. The specific implementation process is as follows:

[0066] (1) Determine the optimal network density range and establish a mathematical model for the average node degree K, number of nodes N, and number of edges M of HBFN:

[0067] 2M=K×N (4)

[0068] Establish a mathematical model for the network density ρ in relation to the number of nodes N and the number of edges M:

[0069]

[0070] The mathematical relationship between the average node degree K and the network density ρ is established using equations (4) and (5):

[0071]

[0072] A study (S.Ji et al., “Study on Brain Function Network of Emotional Conflict Response in College Students with Anxiety”, Chinese Journal of Biomedical Engineering, 2020, 39(2):145-151.) shows that the network is optimal when the average node degree is greater than the natural logarithm of the number of nodes and the network density is less than 50%.

[0073] K > ln(N) (7)

[0074] From equations (6) and (7), we can see that:

[0075] K=ρ×(N-1)>ln (N) (8)

[0076] That is, the lower limit of the HBFN network density is ln(N) / (N-1), and the HBFN network density range is [ln(N) / (N-1), 50%]. In this embodiment, the number of nodes N is the number of channels, N=128. Therefore, the HBFN network density range is [3.82%, 50%].

[0077] (2) Establishing the optimal balance between network sparsity and stability. Existing research (DJ Watts., "Collective dynamics of 'small-world' networks", Nature, 393(6684):440-442.) shows that the human brain is a stable small-world network. The Small-World Index (SWI) is used as an indicator to quantify small-world networks, and its calculation formula is as follows:

[0078]

[0079] In the formula, CC and CC random These are the clustering coefficients of the original HBFN and the corresponding random network, CPL and CPL, respectively. random These are the feature path lengths of the original HBFN and the corresponding random network, respectively.

[0080] According to the definition of SWI, when SWI is greater than 1, the network exhibits small-world characteristics; the larger the SWI, the stronger the network stability. Therefore, within the original HBFN network density range of [3.82%, 50%] in different frequency bands, the network density corresponding to the maximum SWI was searched in the four frequency bands delta, theta, alpha, and beta1 with a step size of 0.01. The results are as follows. Figure 3 As shown, the small-world exponent reaches its maximum value when the network densities for the delta, theta, alpha, and beta1 frequency bands are 0.36, 0.46, 0.44, and 0.35, respectively, after which the small-world exponent slowly decreases or fluctuates. Therefore, using network densities of 0.36, 0.46, 0.44, and 0.35 as thresholds, HBFN is optimized for the delta, theta, alpha, and beta1 frequency bands, respectively. Connections greater than the threshold are retained, while connections less than the threshold are deleted.

[0081] Step 5: Extract the HCC1, HCC2, and HCC3 clustering coefficients from the four frequency bands delta, theta, alpha, and beta1 in HBFN:

[0082]

[0083] In the formula, N(v) represents the neighboring nodes of node v. When the superedge only contains u and t, but not v,

[0084] otherwise, HCC1 measures the number of connections that node v does not facilitate between its neighboring nodes.

[0085]

[0086] In the formula, when the hyperedge contains u, t, and v simultaneously, I(u,t,v) = 1; otherwise, I(u,t,v) = 0. HCC2 measures the number of connections that node v facilitates between its neighboring nodes.

[0087]

[0088] In the formula, E(v) represents the set containing the hyperedges of node v, and HCC3 measures the number of hyperedges passing through node v.

[0089] Nonparametric permutation tests were used to assess the inter-group differences in the cluster coefficients of HCC1, HCC2, and HCC3. The results are as follows: Figure 4 As shown, the significance indexes include: HCC1_Delta, HCC1_Alpha, HCC2_Delta, HCC2_Theta, HCC2_Alpha, HCC2_beta1, HCC3_Delta, and HCC3_Alpha. These significance indexes are used as potential markers for DP detection.

[0090] Step Six: Verify the effectiveness of the above significant difference indicators as potential markers for DP identification. Existing research (H.Cai, et al., “Feature-level fusion approaches based on multimodal EEG data for depression recognition”, Information Fusion, 2020, 59:127-138.) shows that the KNN classifier is generally superior to other classifiers in the field of DP detection. Therefore, in this embodiment, the KNN classifier is used as the detection tool to verify the detection performance of the above 8 potential markers. 10-fold cross-validation is used, and the average detection performance results are shown in Table 1. It can be seen that the potential marker HCC3_Alpha achieved the highest accuracy of 91.36%, with corresponding sensitivity and specificity of 89.52% and 92.91%, respectively. In addition, in this embodiment, the identifiers HCC1_Delta, HCC1_Alpha, HCC2_Delta, HCC2_Theta, HCC2_Alpha, HCC2_beta1, HCC3_Delta, and HCC3_Alpha were input into the KNN classifier, and an accuracy of 96.44% was obtained. The results for each fold are shown in Table 2.

[0091] Table 1. DP detection performance of 8 potential markers

[0092]

[0093]

[0094] Table 2. DP detection performance of each fold of the 10-fold cross-validation.

[0095] 1 2 3 4 5 6 7 8 9 10 mean accuracy 95.84％ 96.52％ 97.18％ 95.19％ 96.01％ 97.49％ 97.31％ 95.54％ 96.94％ 96.36％ 96.44％ Sensitivity 96.20％ 95.81％ 99.01％ 94.52％ 96.49％ 97.85％ 95.33％ 96.44％ 97.22％ 96.06％ 96.49％ Specificity 95.20％ 97.52％ 96.12％ 96.22％ 95.33％ 97.26％ 98.05％ 94.56％ 96.75％ 96.56％ 96.36％

[0096] To verify the mathematical model based on sparsity and stability in the method provided in this patent and to optimize the performance of HBFN, the DP detection performance of HCC1_Delta, HCC1_Alpha, HCC2_Delta, HCC2_Theta, HCC2_Alpha, HCC2_beta1, HCC3_Delta, and HCC3_Alpha identifiers jointly input into the KNN classifier was compared before and after HBFN optimization on a computer configured with an Intel Core i7-12700H 2.30GHz processor, 16.00GB RAM, and Windows 10 operating system. The results are shown in Figure 3. It can be seen that the optimized HBFN achieved a detection accuracy of up to 96.44% while effectively reducing runtime and memory resource consumption. This result indicates that the optimized HBFN can effectively reduce redundancy and invalid connections, thereby improving DP detection performance and reducing computational consumption.

[0097] Table 3 Comparison of HBFN detection performance before and after optimization

[0098] Accuracy (%) Runtime (ms) Resource usage (MB) HBFN before optimization 93.21% 92.75 76.45 Optimized HBFN 96.44% 94.72 210.44

[0099] To verify the performance of the HBFN construction method based on SG-Lasso provided in this patent, a comparative experiment was designed to compare the HBFN construction methods based on Lasso, G-Lasso, and SG-Lasso. HBFNs were constructed using all three methods and optimized. The features HCC1_Delta, HCC1_Alpha, HCC2_Delta, HCC2_Theta, HCC2_Alpha, HCC2_beta1, HCC3_Delta, and HCC3_Alpha were extracted and used as latent markers, which were then input into a KNN classifier for DP detection. The results are shown in Table 4. It can be seen that the HBFN construction method based on SG-Lasso achieved the highest detection accuracy of 96.44%.

[0100] Table 4 Comparison of HBFN detection performance before and after optimization

[0101] Lasso G-Lasso SG-Lasso accuracy 87.25％ 89.03％ 96.44％ Sensitivity 89.20％ 92.21％ 96.49％ Specificity 85.74％ 86.13％ 96.36％

[0102] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the above-mentioned inventive concept. For example, technical solutions formed by replacing the above-mentioned features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. An automated method for detecting depression based on SG-Lasso, characterized in that, The steps of this method are as follows: Step 1: Input EEG signal X = [x1, x2, ..., x N ]; Step 2: Construct the Superbrain Functional Network (HBFN) based on SG-Lasso; Step 3: Use one-way ANOVA to explore the frequency bands where there are significant differences between the DP group and the normal control (NC) group; Step 4: Establish a mathematical model based on the relationship between sparsity and stability to optimize HBFN in different frequency bands; Step 5: Extract the three types of clustering coefficients for different frequency bands in HBFN, and use non-parametric permutation test to evaluate the inter-group differences of the three clustering coefficients to explore potential markers for DP detection; Step 6: Use the KNN classifier to evaluate the effectiveness of latent markers for automatic DP detection.

2. The automatic detection method for depression based on SG-Lasso according to claim 1, characterized in that, The process of constructing HBFN in step two is as follows: SG-Lasso is used to represent the connectivity between nodes, i.e., the optimization objective function is: In the formula, x n This is the EEG data from the nth channel, X n It is a matrix composed of all EEG channel data, where the value of the element in the nth column is 0. n This is a coefficient vector representing the weight vector of the hyperedge generated centered on the nth channel, reflecting the degree to which it is influenced by the EEG signals of other channels. ||*||1 is L1 regularization, ||*||2 is L2 regularization, and λ is a regularization parameter. λ1 is used to adjust the sparsity within the region of interest (ROI), and λ2 is used to adjust the sparsity between ROIs. G i It is the i-th ROI. Super-edge 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 HBFN correlation matrix H: H＝[α1,α2,…,α N ] (2) 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.

3. The automatic detection method for depression based on SG-Lasso according to claim 1, characterized in that, Step four involves establishing a mathematical model based on the relationship between sparsity and stability, and optimizing the HBFN in different frequency bands. The process is as follows: (1) Determine the optimal network density range and establish a mathematical model for the average node degree K, number of nodes N, and number of edges M of HBFN: 2M=K×N (4) Establish a mathematical model for the network density ρ in relation to the number of nodes N and the number of edges M: The mathematical relationship between the average node degree K and the network density ρ is established using equations (4) and (5): Studies have shown that the optimal network is constructed when the average node degree is greater than the natural logarithm of the number of nodes and the network density is less than 50%. K > ln(N) (7) From equations (6) and (7), we can see that: K=ρ×(N-1)>ln (N) (8) That is, the lower limit of HBFN network density is ln(N) / (N-1), and the HBFN network density range is [ln(N) / (N-1), 50%]. (2) Establishing the optimal balance between network sparsity and stability. Existing research shows that the human brain is a stable small-world network. The Small-World Index (SWI) is used as an indicator to quantify small-world networks, and its calculation formula is as follows: In the formula, CC and CC random These are the clustering coefficients of the original HBFN and the corresponding random network, CPL and CPL, respectively. random These are the feature path lengths of the original HBFN and the corresponding random network, respectively. According to the definition of SWI, when SWI is greater than 1, the network exhibits small-world characteristics, and the larger the SWI, the stronger the network stability. Therefore, with a step size of 0.01, we search for the network density ρ corresponding to the maximum SWI within the range of network densities [ln(N) / (N-1), 50%] of the original HBFN in different frequency bands, and use this network density as a threshold to optimize the original HBFN in different frequency bands.

4. The automatic detection method for depression based on SG-Lasso according to claim 1, characterized in that, Step five extracts the three clustering coefficients from HBFN, namely HCC1, HCC2, and HCC3. Their calculation process and meaning are as follows: In the formula, N(v) represents the neighboring nodes of node v. When the superedge only contains u and t, but not v, otherwise, HCC1 measures the number of connections that node v does not facilitate between its neighboring nodes. In the formula, when the hyperedge contains u, t, and v simultaneously, I(u,t,v) = 1; otherwise, I(u,t,v) = 0. HCC2 measures the number of connections that node v facilitates between its neighboring nodes. In the formula, E(v) represents the set containing the hyperedges of node v, and HCC3 measures the number of hyperedges passing through node v.

Images