An Automated Depression Detection Method Based on MBFN

Abstract

The present invention discloses an automatic depression detection method based on MBFN, which ensures the effectiveness of the constructed MBFN, reconstructs the original EEG signal sequence after multi-scaling based on a variable sliding window, so as to fully explore the depression information implicit in the EEG; generates a sample entropy matrix using sample entropies of different scales for each channel, calculates the mean of each row in the sample entropy matrix, so as to select the optimal EEG channel; uses the optimal EEG channel to realize MBFN construction with the mean of bidirectional transfer entropy as a carrier; through the inter-group comparison between the depression group and the normal control group, discovers the significantly changed intra-layer network and inter-layer network in the MBFN, evaluates the correlation between the MBFN network metric index in the significantly changed network layer and the PHQ-9 score, thereby exploring the potential markers for depression detection, and realizing high-precision automatic depression detection.

Description

Technical Field

[0001] The present invention relates to the technical field of electroencephalogram (EEG) signal processing and brain function network construction, and in particular to an automatic depression detection method based on MBFN. Background Art

[0002] Depression is a common mental illness, with persistent low mood and loss of interest in life as its typical symptoms. According to the World Health Organization (WHO) report, about 5% of adults worldwide suffer from depression, about 25% of patients with depression have suicidal tendencies, and more than 700,000 patients with depression die from suicide each year. It is internationally recognized that the most effective way to treat depression is early identification and early intervention to avoid irreversible damage to brain function over time. The traditional depression diagnosis method based on PHQ-9 face-to-face interviews relies on the subjective experience of doctors and the authenticity of information provided by patients, so that the clinical diagnosis accuracy rate is often maintained at around 47%. Therefore, it is of great value and significance to explore objective, accurate and efficient automatic identification methods for depression.

[0003] As an imaging technology that reads electrical activity from brain tissue, electroencephalography (EEG) reflects the degree of synchronization of neurons. The release and binding of neurotransmitters can change the excitability or inhibition of neurons. Neurotransmitters are important chemicals in the brain. Clinical studies have shown that neurotransmitters in patients with depression are in an unbalanced state. EEG-based depression research has achieved fruitful results in the past few decades. Most studies focus on independent EEG electrode data or the fusion of EEG electrode data with other data, ignoring the correlation between different EEG electrode data. Therefore, it is difficult to reveal the physiological mechanism of abnormal changes in brain function hidden behind depression.

[0004] The human brain is a super complex neural network that transmits information through neurons. With the deepening of complex network research based on graph theory, more and more evidence shows that neurological diseases such as depression are essentially caused by abnormal brain functional connectivity rather than local brain damage. Existing studies (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.) have shown that the BFN of patients with depression showed a randomized trend compared with normal controls through a single-layer brain functional network (BFN) analysis based on multi-channel resting-state EEG. The study also used the single-layer BFN related features to obtain a 93.31% depression detection accuracy rate.

[0005] Single-layer BFN based on EEG can realize the analysis and automatic detection of depression to a certain extent. At the same time, the current BFN research is undergoing an expansion from the classic single layer to the multi-layer, and the multi-layer BFN (MBFN) research based on EEG is gradually attracting attention. Unfortunately, there are few MBFN results related to depression. To this end, the present invention provides an automatic detection method for depression based on MBFN by exploring the reconstruction of multi-channel EEG signals and channel selection. Summary of the invention

[0006] The purpose of the present invention is to solve the above problems and provide an automatic depression detection method based on multi-channel EEG signal reconstruction and selection combined with MBFN.

[0007] To achieve the above purpose, the technical solution adopted by the present invention comprises the following specific steps:

[0008] Step 1: Input 128-channel raw EEG data X = {x(1), x(2), …, x(n)};

[0009] Step 2: Multi-scale the EEG signal based on the variable sliding window (VSW). When the scale factor s = 1, VSW obtains the same signal sequence as the original EEG signal sequence. When the scale factor s ≠ 1, VSW reconstructs the original EEG signal sequence after multi-scale.

[0010]

[0011] In the formula, is the jth data of the EEG signal of the lth variable sliding window, and all EEG channels are multi-scaled and reconstructed in turn;

[0012] Step 3: Compare the sample entropy of EEG signals reconstructed at different scales of each channel, select the top k channels in the order of sample entropy values ​​in the EEG signals reconstructed at most scales, and use bandpass filters to divide the EEG signals of the top k channels into five sub-bands: delta (0.5-4Hz), theta (4-8Hz), alpha (8-13Hz), beta (13-30Hz) and gamma (30-48Hz);

[0013] Step 4: Use the electrodes corresponding to the first k EEG channels as the shared nodes of MBFN. The functional connections of EEG node pairs within the sub-band generate the intra-layer adjacency matrix. The functional connections of EEG node pairs between different sub-bands generate the inter-layer adjacency matrix. The intra-layer adjacency matrix and the inter-layer adjacency matrix together form the super-adjacency matrix, and then generate MBFN.

[0014] Among them, the calculation of the functional connection of the intra-layer and inter-layer adjacency matrix uses transfer entropy as a carrier. The transfer entropy between the EEG node pairs in two directions is added and averaged, which is the functional connection strength. The calculation process is as follows:

[0015]

[0016] r i and r j Represents two EEG nodes v j and v j Corresponding time series, τ is the time delay between channels, τ = 1ms, p(·) is the probability distribution;

[0017] Step 5: Analyze the changes in intra-layer and inter-layer functional connectivity of MBFN between the depression group and the normal control group, and analyze the total intra-layer coupling strength of MBFN in the significantly changed network layer ( <m>), interlayer coupling strength ( <h>), multi-layer network participation coefficient (MPC), multi-way clustering coefficient (MCC), local efficiency (LocE), global efficiency (GloE) and other metrics were analyzed by Pearson correlation with the clinical examination standard PHQ-9 score to explore potential markers for depression detection;

[0018] Step 6: Use KNN classifier to evaluate the effectiveness of potential markers for depression detection and achieve depression detection.

[0019] Furthermore, the VSW multi-scale EEG signal reconstruction process in step 2 is as follows:

[0020] (1) When the scale factor s = 2, the EEG signal is reconstructed as Right now

[0021] (2) When the scale factor s = 3, the EEG signal is reconstructed as and Among them, the first EEG signal reconstruction sequence is The second EEG signal reconstruction sequence is

[0022] (3) And so on, until s ≥ N / 4 or s = 30, where N is the number of original EEG channels.

[0023] Furthermore, the process of selecting the first k channels in step 3 is as follows:

[0024] (1) The reconstructed EEG signals are sequentially organized into a set of vector sequences Y with a dimension of m. m (1),…,Y m (N-m+1), where Y m (i)={y(i),y(i+1),…,y(i+m-1)},1≤i≤N-m+1;

[0025] (2) Vector Y m (i) and Y m The distance calculation formula between (j) is:

[0026] d[Y m (i),Y m (j)]=max k＝0，...，m-1 (|y(i+k)-y(i+k)|) (5);

[0027] (3) For a given Y m (i), statistics d[Y m (i),Y m The number of j with [j]≤r i , where 1≤j≤nm, then the vector Y m (i) Existence:

[0028]

[0029] For all vector sequences, then B m (r) is defined as:

[0030]

[0031] (4) Increase the dimension to m+1, and count d[Y m+1 (i),Y m+1 The number of j A for which (j)]≤r i , then A i m (r) is defined as:

[0032]

[0033] A m (r) is defined as:

[0034]

[0035] (5) Calculate the sample entropy of the EEG signal as follows:

[0036]

[0037] In the formula, the parameters are set as follows: m = 2, r = 0.15*SD, SD is the standard deviation of the EEG signal of the corresponding scale;

[0038] (6) Based on the sample entropy of different scales of each channel, the following sample entropy matrix is ​​generated:

[0039]

[0040] Where N is the number of EEG channels, s is the multi-scale coefficient;

[0041] (7) Calculate the mean entropy of each row of the sample entropy matrix, and select the top k channels with the highest mean to construct the MBFN.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. The present invention provides a method for reconstructing multi-scale EEG signals based on a variable sliding window (VSW). By reconstructing the multi-scaled EEG signals, the valuable depression information hidden in the EEG can be fully mined from different scales, thereby ensuring the effectiveness of the subsequently constructed MBFN.

[0044] 2. To avoid the problem of a large number of weak connections in a fully connected network interfering with the core connections, the present invention provides a channel selection method based on bidirectional sample entropy, which is used to select the top k channels closely related to depression to construct an MBFN.

[0045] 3. Through the inter-group difference analysis of MBFN, the present invention found that in the MBFN of patients with depression, the functional connectivity of theta-theta and alpha-alpha within the layer network and the functional connectivity of delta-theta, delta-alpha and delta-beta between the layers were all reduced, and the reduction of the intra-layer functional connectivity was more significant than that of the inter-layer functional connectivity. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flow chart of the method of the present invention;

[0047] Figure 2 The variable sliding window multi-scale EEG signal and reconstruction process of the embodiment of the present invention;

[0048] Figure 3 is the sample entropy of 128 channels on 30 time scales in the embodiment of the present invention;

[0049] Figure 4 is the mean value of sample entropy in each row in the sample entropy matrix of the embodiment of the present invention;

[0050] Figure 5 The results of the intra-layer and inter-layer functional connectivity analysis of MBFN between groups in the embodiment of the present invention are shown;

[0051] Figure 6 The correlation analysis results of the MBFN metric and the PHQ-9 score when p<0.01 in the embodiment of the present invention;

[0052] Figure 7 Depression detection performance of MBFN constructed for different numbers of EEG channels in embodiments of the present invention. DETAILED DESCRIPTION

[0053] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below in conjunction with specific implementation methods.

[0054] See also Figure 1 As shown, the present invention provides a technical solution: an automatic detection method for depression based on multi-channel EEG signal reconstruction and selection and MBFN construction, the method steps are as follows:

[0055] Step 1: Input 128-channel raw EEG data X = {x(1), x(2), ..., x(n)}, such as Figure 1 (a)

[0056] Step 2: Multi-scale the EEG signal based on the variable sliding window (VSW). When the scale factor s = 1, VSW obtains the same signal sequence as the original EEG signal sequence. When the scale factor s ≠ 1, VSW reconstructs the original EEG signal sequence after multi-scale. like Figure 2 As shown,

[0057]

[0058] The specific multi-scale reconstruction process is as follows:

[0059] (1) When the scale factor s = 2, the EEG signal is reconstructed as Right now

[0060] (2) When the scale factor s = 3, the EEG signal is reconstructed as and Among them, the first EEG signal reconstruction sequence is The second EEG signal reconstruction sequence is

[0061] (3) And so on, until s ≥ N / 4 or s = 30, where N is the number of original EEG channels.

[0062] Step 3: Compare the sample entropy of EEG signals reconstructed at different scales for each channel, select the top k channels in the order of sample entropy values ​​in the EEG signals reconstructed at most scales, and use bandpass filters to divide the EEG signals of the top k channels into five sub-bands: delta (0.5-4Hz), theta (4-8Hz), alpha (8-13Hz), beta (13-30Hz) and gamma (30-48Hz). The process is summarized as follows:

[0063] (1) The reconstructed EEG signals are sequentially organized into a set of vector sequences Y with a dimension of m. m (1),…,Y m (N-m+1), where Y m (i)={y(i),y(i+1),…,y(i+m-1)},1≤i≤N-m+1;

[0064] (2) Vector Y m (i) and Y m The distance calculation formula between (j) is:

[0065] d[Y m (i),Y m (j)]=max k＝0，...，m-1 (|y(i+k)-y(i+k)|) (5);

[0066] (3) For a given Y m (i), statistics d[Y m (i),Y m The number of j with [j]≤r i , where 1≤j≤nm, then the vector Y m (i) Existence:

[0067]

[0068] For all vector sequences, then B m (r) is defined as:

[0069]

[0070] (4) Increase the dimension to m+1, and count d[Y m+1 (i),Y m+1 (j)]≤r, then Defined as:

[0071]

[0072] A m (r) is defined as:

[0073]

[0074] (5) Calculate the sample entropy of the EEG signal as follows:

[0075]

[0076] In the formula, the parameters are set as follows: m = 2, r = 0.15*SD, SD is the standard deviation of the EEG signal of the corresponding scale;

[0077] (6) Based on the sample entropy of different scales of each channel, the following sample entropy matrix is ​​generated:

[0078]

[0079] Where N is the number of EEG channels, s is the multi-scale coefficient, and according to (3) of the specific multi-scale reconstruction process in step 2, the parameter is set to s=30. Figure 3 The calculation results of sample entropy of 128 channels in 30 time scales for depression group and normal control group (NC) are given. Due to space limitation, the present invention only provides the results in odd time scales.

[0080] (7) Calculate the mean entropy of each row in the sample entropy matrix. The result is as follows Figure 4 As shown. Existing studies (J.Shen et al., "An optimal channel selection for EEG-based depression detection viakernel-target alignment", IEEE Journal of Biomedical and Health Informatics, 2021, 25(7): 2545-2556.) have shown that when the total number of EEG channels is greater than 58, selecting a channel rate in the range of 32.76% to 39.06% is the best balance between computational load and depression detection accuracy. Based on this theory, statistics Figure 4 The mean entropy of the samples in the Figure 5 As shown, the present invention selects the first 47 channels 13, 18, 19, 20, 22, 25, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 42, 49, 50, 51, 53, 54, 56, 58, 64, 65, 66, 67, 68, 69, 70, 71, 73, 74, 97, 100, 101, 109, 113, 114, 116, 116, 119, 120, 127, 128 for MBFN construction.

[0081] (8) Research (M. Fraschini, et al., "The effect of epoch length on estimated EEG functional connectivity and brain network organization", Journal of Neural Engineering, 2016, 13 (3): 036015.) shows that when the EEG signal window length is 6 seconds, the constructed brain functional network tends to be optimal. Therefore, in the present invention, the EEG signal sequence is divided into a 6-second time window, and the EEG signals of the first 47 channels are decomposed into five sub-bands such as delta (0.5-4Hz), theta (4-8Hz), alpha (8-13Hz), beta (13-30Hz) and gamma (30-48Hz) using a bandpass filter. Figure 1 (b) as shown.

[0082] Step 4: Use the EEG electrodes corresponding to the first 47 channels as the shared nodes of MBFN. The functional connections of EEG node pairs within the sub-band generate the intra-layer adjacency matrix, and the functional connections of EEG node pairs between different sub-bands generate the inter-layer adjacency matrix, such as Figure 1 As shown in (c), the intra-layer adjacency matrix and the inter-layer adjacency matrix together form a super adjacency matrix, which then generates MBFN, as shown in Figure 1 (d)

[0083] In the above process, the calculation of the functional connection of the adjacency matrix within and between layers uses transfer entropy as a carrier. The transfer entropy between EEG node pairs in two directions is added and averaged, which is the functional connection strength. The calculation process is as follows:

[0084]

[0085] r i and r j Represents two EEG nodes v j and v j Corresponding time series, τ is the time delay between channels, τ = 1ms, p(·) is the probability distribution;

[0086] Step 5: Analyze the changes in intra- and inter-layer functional connectivity of MBFN between the depression group and the normal control group. Figure 5 As shown, Figure 5 (a) is the intra-layer adjacency matrix between the two groups. It can be seen that compared with the NC group, the depression group has weakened functional connectivity in theta-theta and alpha-alpha networks within the layer. Figure 5 (b) is the inter-layer adjacency matrix between the two groups. The upper triangle is the inter-layer adjacency matrix of the depression group, and the lower triangle is the inter-layer adjacency matrix of the NC group. It can be seen that compared with the NC group, the functional connectivity of delta-theta, delta-alpha and delta-beta in the inter-layer network of the depression group is weakened; Figure 5 (c) and Figure 5 (d) The functional connectivity difference matrix and significant difference diagram of the two groups are given respectively. From the horizontal comparison of the intra-layer network and the inter-layer network, it can be found that the decrease in theta-theta and alpha-alpha functional connectivity of the intra-layer network is greater than the decrease in delta-theta, delta-alpha and delta-beta functional connectivity of the inter-layer network.

[0087] Research (ZJ Bian, et al., "Relative power and coherence of EEG series are related to amnestic mild cognitive impairment in diabetes", Frontiers In Aging Neuroscience, 2014, 6: 10.) points out that the traditional face-to-face interview based on PHQ-9 is the basis for the diagnosis of depression, and the correlation between biological indicators or brain function network measurement indicators and scale scores is widely used to evaluate whether the corresponding indicators can be used as potential markers for diagnosing related diseases. Therefore, the present invention significantly changes the total coupling strength of MBFN in the theta-theta, alpha-alpha, delta-theta, delta-alpha and delta-beta network layers ( <m>), interlayer coupling strength ( <h>), multi-layer network participation coefficient (MPC), multi-way clustering coefficient (MCC), local efficiency (LocE), global efficiency (GloE) and other metrics were analyzed by Pearson correlation analysis with the clinical examination standard PHQ-9 score, p = 0.05, and the analysis results are shown in Table 1. When p < 0.01, the results are as follows Figure 6 shown.

[0088] Table 1 Correlation analysis results between MBFN metrics and PHQ-9 scores

[0089]

[0090]

[0091] Step 6: Use KNN classifier to evaluate the effectiveness of potential markers for depression detection and achieve depression detection.

[0092] In order to evaluate the effectiveness of the MBFN metric related to PHQ-9 scores as a potential marker for depression detection and to achieve depression detection, the present invention designed the following comparative experiment.

[0093] (1) The 13 MBFN metrics with p<0.05 and the 7 MBFN metrics with p<0.01 in Table 1 were respectively used as KNN classifier inputs for depression detection. The results are shown in Table 2. It can be seen from the table that the 7 MBFN metrics with p<0.01 are better than the 13 MBFN metrics with p<0.05 in depression detection performance. Therefore, the 7 MBFN metrics with p<0.01 are used as potential markers for depression detection in the present invention.

[0094] Table 2 Depression detection performance at different significance levels

[0095]

[0096] (2) To verify whether the 47 channels selected by the present invention to construct MBFN are optimal, a comparative experiment was designed based on the above 7 MBFN metrics with p < 0.01 and the KNN classifier, and MBFN with different numbers of channels was used to detect depression. According to the mean sample entropy result of each row in the sample entropy matrix, the channels with lower average sample entropy were added to the optimal EEG channel subset in turn. The detection results are shown in Figure 2. Figure 7 As shown in the figure, the channel topography shows the distribution of EEG electrodes corresponding to the selected channels. Blue indicates the area where the channels are not selected, and red indicates the area where the selected channels are. When the number of channels is 47, the highest depression detection performance can be obtained, and the accuracy, sensitivity, and specificity of depression detection are 95.78%, 96.27%, and 95.29%, respectively. As can be seen from the line chart, the detection performance continues to improve before the number of channels reaches 47. When the number of channels is 1 to 28, the detection performance improves the fastest, and when the number of channels is 28 to 47, the detection performance improvement rate slows down. When the number of channels is 47, it is the turning point of the detection performance, and it can be observed that the detection performance is declining. When the number of channels is 71, the first performance decline stops, and when the number of channels is 93, the second performance decline stops.

[0097] It will be apparent to those skilled in the art that the 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 the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

[0098] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.< / h> < / m> < / h> < / m>

Claims

1. A method for automatic detection of depression based on MBFN, characterized in that: The specific steps are as follows: Step 1: Input 128-channel raw EEG data X = {x(1), x(2), …, x(n)}; Step 2: Multi-scale EEG signals based on variable sliding window VSW. When the scale factor s = 1, VSW obtains the same signal sequence as the original EEG signal sequence. When the scale factor s ≠ 1, VSW reconstructs the original EEG signal sequence after multi-scale. In the formula, is the jth data of the EEG signal of the lth variable sliding window, and all EEG channels are multi-scaled and reconstructed in turn; Step 3: Compare the sample entropy of EEG signals reconstructed at different scales for each channel, select the top k channels in the order of sample entropy values ​​in the EEG signals reconstructed at most scales, and use bandpass filters to divide the EEG signals of the top k channels into five sub-bands: delta (0.5-4Hz), theta (4-8Hz), alpha (8-13Hz), beta (13-30Hz) and gamma (30-48Hz); The process of selecting the first k channels of sample entropy ranking is as follows: (1) The reconstructed EEG signals are sequentially organized into a set of vector sequences Y with a dimension of m. m (1),…,Y m (N-m+1), where Y m (i)={y(i),y(i+1),…,y(i+m-1)},1≤i≤N-m+1; (2) Vector Y m (i) and Y m The distance calculation formula between (j) is: d[Y m (i),Y m (j)]=max k＝0,…,m-1 (|y(i+k)-y(j+k)|) (5); (3) For a given Y m (i), statistics d[Y m (i),Y m The number of j with [j]≤r i , where 1≤j≤nm, then the vector Y m (i) Existence: For all vector sequences, then B m (r) is defined as: (4) Increase the dimension to m+1, and count d[Y m+1 (i),Y m+1 The number of j A for which (j)]≤r i ,but Defined as: Then A m (r) is defined as: (5) Calculate the sample entropy of the EEG signal as follows: In the formula, the parameters are set as follows: m = 2, r = 0.15*SD, SD is the standard deviation of the EEG signal of the corresponding scale; (6) Based on the sample entropy of different scales of each channel, the following sample entropy matrix is ​​generated: Where N is the number of EEG channels, s is the multi-scale coefficient; (7) Calculate the mean entropy of each row of the sample entropy matrix, and select the top k channels with the highest mean to construct the MBFN; Step 4: Use the electrodes corresponding to the first k EEG channels as the shared nodes of MBFN. The functional connections of EEG node pairs within the sub-band generate the intra-layer adjacency matrix. The functional connections of EEG node pairs between different sub-bands generate the inter-layer adjacency matrix. The intra-layer adjacency matrix and the inter-layer adjacency matrix together form the super-adjacency matrix, and then generate MBFN. Among them, the calculation of the functional connection of the intra-layer and inter-layer adjacency matrix uses transfer entropy as a carrier. The transfer entropy between the EEG node pairs in two directions is added and averaged, which is the functional connection strength. The calculation process is as follows: r i and r j Represents two EEG nodes v j and v j Corresponding time series, τ is the time delay between channels, τ = 1ms, p(·) is the probability distribution; Step 5: Analyze the changes in intra-layer and inter-layer functional connectivity of MBFN between the depression group and the normal control group, and analyze the total intra-layer coupling strength of MBFN in the significantly changed network layer ( <m>), interlayer coupling strength ( <h> Pearson correlation analysis was performed between six metrics including multi-layer network participation coefficient (MPC), multi-way clustering coefficient (MCC), local efficiency (LocE), and global efficiency (GloE) and the clinical examination standard PHQ-9 score to explore potential markers for depression detection;< / h> < / m> Step 6: Use KNN classifier to evaluate the effectiveness of potential markers for depression detection and achieve depression detection.

2. The MBFN-based automatic depression detection method according to claim 1, characterized in that: The process of reconstructing EEG signals based on VSW multi-scale in step 2 is as follows: (1) When the scale factor s = 2, the EEG signal is reconstructed as Right now (2) When the scale factor s = 3, the EEG signal is reconstructed as and Among them, the first EEG signal reconstruction sequence is The second EEG signal reconstruction sequence is (3) And so on, until s ≥ N / 4 or s = 30, where N is the number of original EEG channels.

Images