阅读说明：本技术 基于多导联脑电最优导联选择的抑郁检测系统 (Depression detection system based on optimal lead selection of multi-lead electroencephalogram ) 是由 胡斌 张晓炜 沈健 李娜 潘静 李俊磊 吴漫兮 于 2019-11-22 设计创作，主要内容包括：本发明提供的一种基于多导联脑电最优导联选择的抑郁检测系统,有效去除多导联脑电信号引起的信息冗余,降低计算复杂度,提高抑郁检测的识别率。包括：多导联脑电信号获取模块、脑电信号预处理模块、特征提取模块、导联选择模块、抑郁检测模块；多导联脑电信号获取模块用于获取静息状态采集的多导联脑电信号,脑电信号预处理模块用于对获取的多导联脑电信号进行预处理,特征提取模块用于提取进行最优导联选择的有效特征,所述导联选择模块用于采用修改版本的核排列作为目标函数,利用新的二元粒子群优化法来对目标函数进行导联选择,求取最优化导联子集,抑郁检测模块用于对选择出来的最优化导联子集进行识别,构建最优化抑郁检测分类模型,来进行抑郁检测。(The depression detection system based on the optimal lead selection of the multi-lead electroencephalogram effectively removes information redundancy caused by the multi-lead electroencephalogram signals, reduces the calculation complexity and improves the identification rate of depression detection. The method comprises the following steps: the device comprises a multi-lead electroencephalogram signal acquisition module, an electroencephalogram signal preprocessing module, a feature extraction module, a lead selection module and a depression detection module; the device comprises a multi-lead electroencephalogram signal acquisition module, an electroencephalogram signal preprocessing module, a characteristic extraction module, a lead selection module, a new binary particle swarm optimization method and a depression detection module, wherein the multi-lead electroencephalogram signal acquisition module is used for acquiring multi-lead electroencephalogram signals acquired in a resting state, the electroencephalogram signal preprocessing module is used for preprocessing the acquired multi-lead electroencephalogram signals, the characteristic extraction module is used for extracting effective characteristics for optimal lead selection, the lead selection module is used for adopting modified version nuclear arrangement as a target function, conducting lead selection on the target function by utilizing the new binary particle swarm optimization method, obtaining an optimized lead subset, and the depression detection module is used for identifying the selected optimized lead subset, constructing an optimized depression.)

1. A depression detection system based on multi-lead electroencephalogram optimal lead selection is characterized by comprising a multi-lead electroencephalogram signal acquisition module, an electroencephalogram signal preprocessing module, a feature extraction module, a lead selection module and a depression detection module; the multi-lead electroencephalogram signal acquisition module is used for acquiring multi-lead electroencephalogram signals acquired in a resting state, the electroencephalogram signal preprocessing module is used for preprocessing the acquired multi-lead electroencephalogram signals, the feature extraction module is used for extracting effective features for optimal lead selection, the lead selection module is used for adopting modified version kernel arrangement as a target function, conducting lead selection on the target function by using a new binary particle swarm optimization method, obtaining an optimal lead subset, and the depression detection module is used for identifying the selected optimal lead subset, constructing an optimal depression detection classification model and conducting depression detection.

2. The system for detecting depression based on multi-lead electroencephalogram optimal lead selection according to claim 1, wherein 64-lead electroencephalogram signals acquired by the multi-lead electroencephalogram signal acquisition module are acquired on a Brain Products platform, 64-lead 10-20 international standard electrode caps are used as acquisition equipment, the scalp impedance of the sensor is lower than 20 kilo-ohms, the sampling rate is set to 1000Hz, and the reference electrode and the ground electrode are respectively set to FCz and AFz.

3. The system for detecting depression based on optimal lead selection of multi-lead brain electricity according to claim 2, characterized in that the multi-lead brain electricity signal obtaining module obtains 64 lead brain electricity signals of resting state of 5 minutes under resting state of quiet closed eyes of healthy subject and depression patient under the same condition.

4. The system for detecting depression based on optimal lead selection of multi-lead electroencephalogram according to claim 3, characterized in that the electroencephalogram signal preprocessing module performs data preprocessing on the collected 64-lead electroencephalogram signals, and comprises four steps of changing reference electrodes, down-sampling, band-pass filtering and artifact removal:

1) changing the reference electrode, namely resetting the original reference electrode position FCz to lead potentials of two mastoid positions;

2) the down-sampling is to reduce the sampling rate of the electroencephalogram signals from 1000Hz to 256 Hz;

3) the band-pass filtering is to carry out the band-pass filtering of 1Hz-40Hz to remove the direct current interference and the high-frequency noise signal;

4) the artifact removal is to perform visual inspection on the original signals, select the minimum 70-second resting state electroencephalogram signals containing artifact signals, bad electrodes and the like, and further remove the artifact signals from the selected 70-second resting state electroencephalogram signals by using an independent component analysis method.

5. The system for depression detection based on optimal lead selection for multi-lead brain electrical waves according to claim 4, wherein said feature extraction module comprises the steps of linear feature extraction: the linear feature extraction is to extract three linear features including power spectrum maximum power, power spectrum average power and power spectrum center frequency from the preprocessed electroencephalogram signal; firstly, solving a power spectrum of each section of electroencephalogram signal by using a Welch method, and solving a maximum value of the power spectrum signal by using a corresponding section of power spectrum signal to obtain the maximum power of the power spectrum; calculating the average value of the power spectrum to obtain the average power of the power spectrum; and solving the center frequency to obtain the center frequency of the power spectrum.

6. The system for depression detection based on optimal lead selection for multi-lead brain electrical waves according to claim 4, wherein said feature extraction module comprises the steps of nonlinear feature extraction: the nonlinear feature extraction is to extract two nonlinear features including Kolmogorov entropy and LZ complexity from the preprocessed electroencephalogram signal;

kolmogorov entropy characterizes the degree of chaos of the system by the associated integral of several growing embedding dimensions:

LZ complexity is based on the coarse granularity of the measurements to find the signal complexity, which proves to be bounded by:

7. The system for depression detection based on optimal lead selection for multi-lead brain electrical waves according to claim 1, characterized in that said lead selection module comprises the step of optimizing lead weight selection: firstly, dividing a feature set containing all brain electricity into a training set and a testing set, and implementing the following lead weight selection steps on the training set: firstly, organizing feature data in a training set into a feature matrix f according to leads and using class labels y^TCalculating a target matrix L; generating P groups of lead selection weights w by using an NBPSO method from the initialized P groups of w, and calculating a P groups of lead subset S_iAnd computing a lead selection kernel matrix K for the P sets of lead subsets_SMeasuring the advantages and disadvantages of the lead subsets obtained by selection by using the mKTA as a target function, and calculating P groups of mKTA kernel arrangement values; then, according to the P groups of mKTA values, updating the weight P _ best of all leads corresponding to the P groups of individual optimal mKTA values and the weight g _ best of all leads corresponding to the global optimal mKTA values, and updating the P groups of lead weights and the lead subsets by using NBPSO; updating P sets of lead subsets S using the updated P sets of lead weights w_iAnd computing a lead selection kernel matrix K_SThis is iterated until the optimized lead weights w are output.

8. The system for depression detection based on optimal lead selection for multi-lead brain electrical waves according to claim 7, wherein said depression detection module comprises the steps of constructing an optimized depression detection classification model: first, the optimized lead subset training set S corresponding to the training set and the test set is calculated according to the calculated optimized lead weight_trAnd optimizing lead subset test set S_te(ii) a And (3) constructing an optimized depression detection classification model by optimizing depression detection parameters by using a ten-fold cross-validation method: will be provided withDerived optimized lead subset training set S_trAnd training class labels

9. The system for depression detection based on optimal lead selection for multi-lead brain electrical waves according to claim 8, wherein said depression detection module comprises the step of testing an optimized depression detection classification model: test set S of optimized lead subset to be found_teInputting into the constructed optimized depression detection classification model: f (S)_te)＝W_modelk(S_teAnd S) + b, obtaining a depression test result through calculation, comparing the depression test result output by the test label and the optimized depression detection model, and calculating corresponding model performance evaluation indexes such as accuracy, sensitivity, specificity and the like.

10. The system for depression detection based on optimal lead selection for multi-lead brain electrical waves according to claim 9, wherein the depression detection module comprises the steps of depression detection aided diagnosis: firstly, preprocessing the acquired electroencephalogram signal data, extracting effective characteristics to obtain a characteristic matrix f_newComputing a diagnosis-aided optimized lead feature subset S using the optimized lead weights_newAnd then inputting the obtained auxiliary diagnosis optimized lead feature subset into an optimized depression detection classification model, and calculating a depression detection result: f (S)_new)＝W_modelk(S_new,S)+b。

Technical Field

The invention relates to a depression detection system based on multi-lead electroencephalogram optimal lead selection, and belongs to the technical field of depression auxiliary medical treatment.

Background

Depression is a major mental disorder that can last for weeks, months or even years, and it can cause impairment in the ability of an individual to live. According to the world health organization predictions, depression will become the second largest disease threatening human life in 2020. Depression can significantly affect the thinking, behavior, sensation and well-being of humans. Furthermore, according to the world health organization forecast, 3.5 million people worldwide are affected by depression, and nearly 100 million depressed patients suicide annually. Depression has become a serious threat to human life. Although depression can be treated by effective clinical means, a large number of patients suffering from depression worldwide are suffering from depression due to neglect, untimely diagnosis and misdiagnosis. Therefore, the early detection of depression is crucial to cure depression, thereby improving the life quality of depression patients, ensuring the physical and mental health of the depression patients, and simultaneously, the social and economic pressures brought by depression can be directly reduced.

Currently, the main clinical means of diagnosing and detecting depression are patient self-rating scales and clinician interviews, but patient self-rating scales are often subjective and clinician interviews are often limited by their clinical skills and methods of diagnosis. Furthermore, there is no objective assessment criteria in clinical practice for depression diagnosis and detection. With the increase in the number of patients with depression, effective follow-up treatment of patients with depression through early detection and evaluation has become increasingly difficult. In order to obtain a more objective detection result, researchers at home and abroad detect depression by behavior signals such as expressions, voices and postures, and a plurality of machine learning and deep learning methods have achieved some achievements, but because the behavior signals are indirect signals and have controllability and disguise, the signals may be interfered by subjective factors of a tested person, and the potential and real mental states of the tested person may not be reflected sometimes. The electroencephalogram signal can directly reflect the working state in the brain, is spontaneous and not controlled by human, and has high time resolution. Moreover, the electroencephalogram has the characteristics of safe acquisition, no wound, low cost, simplicity and convenience in acquisition and the like. Therefore, the electroencephalogram signals are feasible and promising to provide objective depression detection, and the electroencephalogram-based depression detection research provides a more convenient and new way for clinical diagnosis and treatment of depression. In recent years, the research of depression detection based on electroencephalogram has also received wide attention from experts and scholars.

At present, the identification rate of the depression detection research based on the electroencephalogram signals is still to be improved. Moreover, because clinically acquired electroencephalogram signals often have the characteristics of multiple leads, high sampling rate, high density and the like, the information redundancy and the calculation complexity of the clinical electroencephalogram signals are greatly improved, and great challenges are brought to machine learning and deep learning technologies. The large number of leads may cause the machine learning and deep learning models to over-fit, and even, unrelated leads may cause a loss of depression detection recognition rate. In addition, when electroencephalogram signals are clinically collected, a large amount of time is consumed by setting a large number of leads, and meanwhile, discomfort of a tested person can be caused, so that the collected electroencephalogram signals of the tested person cannot accurately reflect the mental state of the tested person.

Disclosure of Invention

Aiming at the problems in the prior art, the invention provides a depression detection system based on multi-lead electroencephalogram optimal lead selection, which aims at the problems of low recognition rate of electroencephalogram signals in clinical depression detection and high information redundancy and calculation complexity of the multi-lead electroencephalogram signals, effectively removes the information redundancy caused by the multi-lead electroencephalogram signals, reduces the calculation complexity and improves the recognition rate of depression detection.

The technical scheme of the invention is as follows:

1. a depression detection system based on multi-lead electroencephalogram optimal lead selection is characterized by comprising a multi-lead electroencephalogram signal acquisition module, an electroencephalogram signal preprocessing module, a feature extraction module, a lead selection module and a depression detection module; the multi-lead electroencephalogram signal acquisition module is used for acquiring multi-lead electroencephalogram signals acquired in a resting state, the electroencephalogram signal preprocessing module is used for preprocessing the acquired multi-lead electroencephalogram signals, the feature extraction module is used for extracting effective features for optimal lead selection, the lead selection module is used for selecting leads of the target functions by adopting modified version of nuclear array mKTA as the target functions and utilizing a New Binary Particle Swarm Optimization (NBPSO) method, and the depression detection module is used for identifying the selected lead optimized subsets and constructing an optimized depression detection classification model for depression detection.

2. The 64-lead electroencephalogram signals acquired by the multi-lead electroencephalogram signal acquisition module are acquired on a Brain Products platform, a 64-lead 10-20 international standard electrode cap is used as acquisition equipment, the scalp impedance of the sensor is lower than 20 kilo-ohms, the sampling rate is set to be 1000Hz, and the reference electrode and the grounding electrode are respectively set to be FCz and AFz.

3. The multi-lead electroencephalogram signal acquisition module acquires 64 lead electroencephalogram signals of healthy testees and depression patients in a resting state of resting for 5 minutes under the same condition in a quiet eye-closing resting state.

4. The electroencephalogram signal preprocessing module carries out data preprocessing on the collected 64 electroencephalogram signals, and the data preprocessing comprises four steps of changing a reference electrode, down-sampling, band-pass filtering and artifact removal:

1) changing the reference electrode, namely resetting the original reference electrode position FCz to lead potentials of two mastoid positions;

2) the down-sampling is to reduce the sampling rate of the electroencephalogram signals from 1000Hz to 256 Hz;

3) the band-pass filtering is to carry out the band-pass filtering of 1Hz-40Hz to remove the direct current interference and the high-frequency noise signal;

4) the artifact removal is to perform visual inspection on the original signals, select the minimum 70-second resting state electroencephalogram signals containing artifact signals, bad electrodes and the like, and further remove the artifact signals from the selected 70-second resting state electroencephalogram signals by using an independent component analysis method.

5. The feature extraction module comprises the steps of linear feature extraction: the linear feature extraction is to extract three linear features including power spectrum maximum power, power spectrum average power and power spectrum center frequency from the preprocessed electroencephalogram signal; firstly, solving a power spectrum of each section of electroencephalogram signal by using a Welch method, and solving a maximum value of the power spectrum signal by using a corresponding section of power spectrum signal to obtain the maximum power of the power spectrum; calculating the average value of the power spectrum to obtain the average power of the power spectrum; and solving the center frequency to obtain the center frequency of the power spectrum.

6. The feature extraction module comprises the steps of nonlinear feature extraction: the nonlinear feature extraction is to extract two nonlinear features including Kolmogorov entropy and LZ complexity from the preprocessed electroencephalogram signal;

kolmogorov entropy characterizes the degree of chaos of the system by the associated integral of several growing embedding dimensions:

where m is the embedding dimension, C_m(r) is the associated integral of the embedding dimension;

LZ complexity is based on the coarse granularity of the measurements to find the signal complexity, which proves to be bounded by:

and c (n) can be normalized by b (n) to:

where n is the signal length, c (n) is the complexity of the signal, and b (n) is the complexity of the signal after binarization.

7. The lead selection module comprises the steps of optimizing lead weight selection: dividing a feature set containing all brain electricity into a training set and a testing set, and implementing the following lead weight selection steps on the training set: firstly, organizing feature data in a training set into a feature matrix f according to leads and using class labels y^TComputing a targetA matrix L; generating P groups of lead selection weights w by using an NBPSO method from the initialized P groups of w, and calculating a P groups of lead subset S_iAnd computing a lead selection kernel matrix K for the P sets of lead subsets_SMeasuring the advantages and disadvantages of the lead subsets obtained by selection by using the mKTA as a target function, and calculating P groups of mKTA kernel arrangement values; then, according to the P groups of mKTA values, updating the weight P _ best of all leads corresponding to the P groups of individual optimal mKTA values and the weight g _ best of all leads corresponding to the global optimal mKTA values, and updating the P groups of lead weights and the lead subsets by using NBPSO; updating P sets of lead subsets S using the updated P sets of lead weights w_iAnd computing a lead selection kernel matrix K_SThis is iterated until the optimized lead weights w are output.

8. The depression detection module comprises the steps of constructing an optimized depression detection classification model: first, the optimized lead subset training set S corresponding to the training set and the test set is calculated according to the calculated optimized lead weight_trAnd optimizing lead subset test set S_te(ii) a And (3) constructing an optimized depression detection classification model by optimizing depression detection parameters by using a ten-fold cross-validation method: training set S of optimized lead subset to be obtained_trAnd training class labels

Dividing the model into 10 equal parts randomly, taking 9 parts of the model as training models of a training set in turn, taking the remaining 1 part of the model as a verification set verification model, and searching for optimized parameters of the depression detection model; then, an optimized depression detection classification model f is constructed on the obtained optimized lead subset training set by using the found optimization parameters (S)_tr)，f(S_tr)＝W_modelk(S_tr,S_tr) + b, wherein W_modelIs the weight coefficient, k is the kernel function, b is the bias term.

9. The depression detection module comprises the steps of testing an optimized depression detection classification model: test set S of optimized lead subset to be found_teInputting into the constructed optimized depression detection classification model: f (S)_te)＝W_modelk(S_te,S_tr) And b, obtaining a depression test result through calculation, comparing the depression test result output by the test type label and the optimized depression detection model, and calculating corresponding model performance evaluation indexes such as accuracy, sensitivity, specificity and the like.

10. The depression detection module comprises the steps of depression detection aided diagnosis: firstly, preprocessing the acquired electroencephalogram signal data, extracting effective characteristics to obtain a characteristic matrix f_newComputing a diagnosis-aided optimized lead feature subset S using the optimized lead weights_newAnd then inputting the obtained auxiliary diagnosis optimized lead feature subset into an optimized depression detection classification model, and calculating a depression detection result: f (S)_new)＝W_modelk(S_new,S_tr)+b。

The invention has the technical effects that:

the depression detection system based on the optimal lead selection of the multi-lead electroencephalogram effectively removes information redundancy caused by the multi-lead electroencephalogram, reduces the calculation complexity and improves the identification rate of depression detection aiming at the problems of low identification rate of electroencephalogram signals in clinical depression detection and high information redundancy and calculation complexity of the multi-lead electroencephalogram signals. The invention also solves the problems of poor interpretability, poor recognition effect, insufficient consideration to lead combination, high calculation loss, easy overfitting and the like of the existing lead selection method.

The depression detection method based on multi-lead electroencephalogram optimal lead selection uses modified Kernel Target Alignment (mKTA) as a Target function of lead selection, optimizes the selected lead subset through New Binary Particle Swarm Optimization (NBPSO), and finally solves the optimized lead subset. Because the mKTA value is minimized in the process of obtaining the optimized lead subset, in other words, the class spacing of the depression patient and the normal tested electroencephalogram in the feature space is maximized, the optimal lead subset selected by the lead selection method can be used for depression diagnosis to obtain a relatively ideal depression diagnosis identification rate, and the depression patient and the normal tested electroencephalogram signals can be accurately distinguished. Meanwhile, the objective function mKTA provided by the invention is used for measuring the similarity degree of the kernel function and the objective function, has good interpretability, and similarly, the NBPSO provided by the invention is also a simple and effective function optimization method and also has good interpretability, and the mKTA provided by the invention can also effectively avoid overfitting. The NBPSO provided by the invention selects the lead subsets as a whole, and fully examines the combination relation among different leads to select the optimized lead subset. As is well known, particle swarm optimization is an evolutionary method and can calculate all lead weights of a plurality of groups in parallel, so that the new binary particle swarm optimization provided by the invention can also solve the problem of high calculation loss of the traditional lead selection algorithm.

Drawings

FIG. 1 is a workflow diagram of the depression detection system based on optimal lead selection for multi-lead brain electrical.

Figure 2 is a diagram of the position of a 64 lead 10-20 international standard electrode.

FIG. 3 is a flow chart of lead weight selection by the lead selection module.

Fig. 4 is a flow chart of depression detection by the depression detection module.

Detailed Description

The present invention will be described in further detail with reference to the following drawings and examples.

A depression detection system based on multi-lead electroencephalogram optimal lead selection, comprising: the device comprises a multi-lead electroencephalogram signal acquisition module, an electroencephalogram signal preprocessing module, a feature extraction module, a lead selection module and a depression detection module; the device comprises a multi-lead electroencephalogram signal acquisition module, an electroencephalogram signal preprocessing module, a characteristic extraction module, a lead selection module, a new binary particle swarm optimization method and a depression detection module, wherein the multi-lead electroencephalogram signal acquisition module is used for acquiring multi-lead electroencephalogram signals acquired in a resting state, the electroencephalogram signal preprocessing module is used for preprocessing the acquired multi-lead electroencephalogram signals, the characteristic extraction module is used for extracting effective characteristics for optimal lead selection, the lead selection module is used for adopting modified version nuclear arrangement as a target function, conducting lead selection on the target function by utilizing the new binary particle swarm optimization method, obtaining an optimized lead subset, and the depression detection module is used for identifying the selected optimized lead subset, constructing an optimized.

FIG. 1 is a workflow diagram of the depression detection system based on optimal lead selection for multi-lead brain electrical.

The working process of the depression detection system based on the optimal lead selection of the multi-lead electroencephalogram comprises the following 5 stages:

(1) and in the data acquisition stage of the multi-lead electroencephalogram signal acquisition module, 64-lead electroencephalogram signals in a quiet eye-closing resting state under the same conditions of a healthy subject matched with age, gender and education degree and a depression patient are acquired.

(2) And in the data preprocessing stage of the electroencephalogram signal preprocessing module, the acquired electroencephalogram signals are processed in four steps of changing a reference electrode, down-sampling, band-pass filtering and artifact removing.

(3) And in the feature extraction stage of the feature extraction module, three linear features including power spectrum maximum power, power spectrum average power and power spectrum center frequency and two nonlinear features including Kolmogorov entropy and LZ complexity are extracted from the preprocessed electroencephalogram signals.

(4) In the lead selection stage of the lead selection module, a modified version of a kernel arrangement method is used as an objective function, the objective function is selected by using a new binary particle swarm optimization algorithm, and an optimized lead subset is solved.

(5) And in the depression detection stage of the depression detection module, after the lead selection is completed, classifiers such as a support vector machine and the like are used for identifying the selected optimized lead subset, and an optimized classification model is constructed for depression detection.

The following respectively illustrates the 5 stages of depression detection based on multi-lead electroencephalogram optimal lead selection according to the present invention:

(1) data acquisition phase

During the data collection phase, we selected 15 depression patients and 20 age, gender, and education matched healthy subjects by psychometric scale and physician diagnosis. All the tested subjects had right handedness, normal vision or vision correction, normal hearing and intelligence, and no history of other neurological diseases. Before data collection, all the tested patients do not take any nerve medicine or mental medicine, and the mental state of the tested patients is good during data collection. All data acquisitions tested were performed under the same environmental conditions.

The data acquisition is completed on a Brain Products platform, the resting state electroencephalogram signals of a tested person in a resting state of quiet eyes and closed eyes are acquired for 5 minutes, a 64-lead 10-20 international standard electrode cap is used as the acquisition equipment, and the electrode positions are shown in figure 2. The scalp impedance of the sensor is below 20 kilo-ohms. The sampling rate was set to 1000 Hz. The reference and ground electrodes were set at FCz and AFz, respectively.

(2) Data preprocessing stage

And carrying out data preprocessing on the collected 64 electroencephalogram signals. The data preprocessing mainly comprises four steps of changing a reference electrode, down-sampling, band-pass filtering and artifact removal.

Since the reference electrode is set to FCz when the electroencephalogram signal is acquired, the signal amplitude of the leads in the prefrontal lobe and parietal lobe areas is low. Therefore, the original reference electrode position is firstly reset to the lead potential of the mastoid positions on two sides, the signal amplitude of the leads in the prefrontal lobe and parietal lobe areas is increased, and the subsequent processing is facilitated. Because the original sampling frequency of the electroencephalogram signal is 1000Hz, too high sampling rate can cause too large data volume and influence subsequent processing and analysis. Therefore, the acquired electroencephalogram signals are subjected to down-sampling, and the sampling rate of the electroencephalogram signals is reduced from 1000Hz to 256 Hz. The invention performs 1Hz-40Hz band-pass filtering to remove DC infection and high frequency noise signals. Although the invention collects the brain electrical signals in a resting state with quiet eyes, the influence caused by the electric eye, the eye movement and the myoelectric signals is still difficult to avoid. Therefore, the artifact of the acquired brain electrical signals needs to be removed. Removing artifacts is to visually inspect the original signal and pick out the signals containing artifacts such as: and (3) carrying out minimum 70-second resting state electroencephalogram signals such as body movement, blink signals, eye movement signals, electromyogram signals and bad electrodes, and further removing artifact signals from the selected 70-second resting state electroencephalogram signals by using an independent component analysis method.

(3) Feature extraction stage

The invention uses three linear characteristics and two nonlinear characteristics which are commonly used in depression diagnosis and proved to be effective to extract characteristics of each preprocessed electroencephalogram signal.

1) Three linear features: the power spectrum maximum power, the power spectrum average power and the power spectrum center frequency are included;

and (3) solving a power spectrum of each section of electroencephalogram signal by using a Welch method, and then solving the maximum value, the average value and the center frequency of the corresponding section of power spectrum signal respectively to obtain the maximum power of the power spectrum, the average power of the power spectrum and the center frequency of the power spectrum.

2) Two non-linear characteristics: including kolmogorov entropy, LZ complexity;

kolmogorov entropy can characterize the degree of chaos of a system by the associated integration of several growing embedding dimensions:

where m is the embedding dimension, C_m(r) is the associated integral of the embedding dimension;

LZ complexity is based on the coarse granularity of the measurements to find the signal complexity, which proves to be bounded by:

and c (n) can be normalized by b (n) to:

where n is the signal length, c (n) is the complexity of the signal, and b (n) is the complexity of the signal after binarization.

(4) Lead selection phase

First, the feature set containing all brain electricity is divided into training set and testing set. The following lead weight selection procedure is performed on the training set, which is a flow chart for the lead weight selection by the lead selection module, as shown in FIG. 3. First, centralize the trainingFeature data of (2) organizing a feature matrix f by leads_i＝[c_i1,...,c_iE]，f＝[f₁,...,f_n]^TWherein c is_ieRepresenting the characteristic data of the E-th lead, wherein E represents the lead number of the electroencephalogram signals, E is 64 in the embodiment, n is the number of the electroencephalogram signals in the training set, and n represents the segment number of the electroencephalogram signals; by class label y^TObtaining the target matrix L, L ═ yy^T,y＝[y₁,...,y_n]^T,y_i∈{+1,-1}，y_iA presentation class label; then P groups of lead selection weights w are generated by P groups of lead selection weights w initialized randomly by using an NBPSO method, and P groups of lead subsets S are calculated_i，S_i＝wof_i(o denotes the Hadamard product) and computes a lead selection kernel matrix K for the P sets of lead subsets_S，K_Sij＝k(S_i,S_j)＝k(wof_i,wof_j) And k represents a kernel function.

Using a modified Kernel Target Alignment (mKTA) as an objective function to measure the goodness of the lead subsets obtained by selection, P sets of mKTA values were calculated:

where L is the object matrix, and L is yy^T,y＝[y₁,...,y_n]^T,y_i∈{+1,-1}，y_iRepresenting class label, n representing the number of segments of the electroencephalogram signal, f ═ f₁,...,f_n]^T,f_i＝[c_i1,...,c_iE],w＝[w₁,...,w_E],w_e∈{0,1},S_i＝wof_i，f_iIs the characteristic of all leads of the i-th segment of electroencephalogram signal, c_ieFeatures, w, representing the e-th lead of the i-th electroencephalogram signal_eRepresents the weight of the e-th lead when w_eWhen 1, it is indicated that the e-th lead is selected, and when w_eWhen 0, it is indicated that the e-th lead is not selected, w represents the weight of all leads (thus w can be considered as the selected subset of leads), S_iRepresenting the ith segment of electroencephalogramThe signature is passed through a subset of the selected leads, where the eigenvalues of the unselected leads are set to 0 and the eigenvalues of the selected leads are unchanged. K_SIs a kernel matrix for lead selection, K_Sij＝k(S_i,S_j)＝k(wof_i,wof_j). Using mKTA as the objective function, the objective matrix L and lead selection kernel matrix K can be minimized by optimizing the selected lead subset_SThe difference between them. According to the mKTA definition, the smaller the mKTA value is, the target matrix L and the lead selection kernel matrix K are shown_SThe higher the similarity between the two, it means that the selected lead subset can obtain a good recognition rate in depression diagnosis. Therefore, minimizing the mKTA value is also equivalent to maximizing the inter-class distance in the feature space.

Next, according to the above P sets of mKTA values, updating the weights P _ best of all leads corresponding to the P sets of individual optimal mKTA values and the weights g _ best of all leads corresponding to the global optimal mKTA values using a New Binary Particle Swarm Optimization (NBPSO), and updating the above P sets of lead weights and updating the lead subsets using NBPSO:

S(v_t+1)＝2×|Sig(v_t+1)-0.5|

wherein, w_tWeights (lead subsets) representing all leads at the t-th iteration, v_tWeights w representing all leads at the t-th iteration_tVelocity v of_t+1Weights w representing all leads at the updated t +1 th iteration_tThe speed of the motor vehicle is set to be,

is a constant of the coefficient of inertia, c₁、c₂Is a normal number, and Rand () are [0,1 ]]As a random function of intervalP _ best represents the weight of all leads corresponding to the individually optimal mKTA value, g _ best represents the weight of all leads corresponding to the globally optimal mKTA value, Sig (v)_t+1) Representing the velocity v_t+1Sigmod function of S (v)_t+1) Is to the velocity v_t+1The conversion function to convert, exchange (w)_t) Representing the weight w for all leads of the t-th round_tThe value is converted from 1 to 0 and vice versa, w_t+1Is the weight of all leads at the t +1 th iteration after update. Selecting a lead subset by using NBPSO, solving the mKTA value corresponding to the subset to update the individual optimal mKTA value and the global optimal mKTA value, and updating the next round of updating speed v of the currently selected lead subset according to the individual optimal mKTA value and the weight of all leads corresponding to the global optimal mKTA value_t+1And converting the speed, and updating the weights w of all leads in the next round through the speed_t+1Until the optimized lead weights are found. The NBPSO provided by the invention is a parallelization optimization method, can calculate and update a plurality of groups of lead weights in parallel, for example, P groups of lead weights are constructed during initialization, and the P groups of lead weights can be calculated in parallel in updating iteration until the optimized lead weights are solved.

Updating P set lead subsets S using updated P set lead weights_iAnd calculating a lead selection kernel function K_SAnd iterating in the above manner until the optimized lead subset S is obtained, the optimized lead subsets obtained by the embodiment of the present invention are Fp1, C4, P3, O2, F7, CP1, FC5, TP9, F1, F2, C1, FC3, F5, C5, AF7, FT8, TP7, FT9, FT10 and Fpz. The area of these leads coincides with the known depression-related brain region.

(5) Depression detection stage

Fig. 4 is a flow chart of depression detection by the depression detection module.

Obtaining optimized lead subsets according to the optimized lead weights obtained in the stage (4), and respectively calculating the optimized lead subsets S corresponding to the training set and the test set_tr＝wof_trAnd S_te＝wof_teWhere w is the found optimal lead weight, f_trIs a training feature set, S_trIs the optimized lead subset training set, f_teIs a set of test features, S_teIs an optimized lead subset test set. In auxiliary diagnosis, auxiliary diagnosis optimized lead subset S is calculated for auxiliary diagnosis EEG characteristics by using optimized lead weight_new＝wof_newWherein f is_newIs a set of auxiliary diagnostic features, S_newIs a set of auxiliary diagnostic optimized lead subset features. Class label y^TClassification as test class labels

And training class labels

The method comprises the following steps of firstly constructing an optimized depression detection classification model: and (3) optimizing depression detection parameters by using a ten-fold cross-validation method to construct an optimized depression detection classification model. Training set S of optimized lead subset to be obtained_trAnd training class labelsDividing the model into 10 equal parts at random, taking 9 parts of the 10 equal parts as training models of a training set in turn, taking the remaining 1 part of the 1 equal part as a verification set to verify the models, and searching for optimized parameters of the depression detection model, in other words, taking each part of data as the verification set to verify the performance of the training models generated by using other data, and generally selecting the model parameter with the highest average verification precision as the optimized parameter of the depression detection model; then, an optimized depression detection classification model f is constructed on the obtained optimized lead subset training set by using the found optimization parameters (S)_tr) The invention uses a classification model of a support vector machine, and the classification model constructed by the classifier is as follows:

f(S_tr)＝W_modelk(S_tr,S_tr) + b, wherein W_modelIs the weight coefficient, k is the kernel function, b is the bias term.

Secondly, the method comprises the step of detecting the classification model for the optimized depressionThe testing steps are as follows: test set S of optimized lead subset to be found_teInputting into the constructed optimized depression detection classification model, f (S)_te)＝W_modelk(S_teS) + b, obtaining depression test result by calculation, and comparing test labels

And a depression test result output by the optimized depression detection model, and calculating corresponding model performance evaluation indexes such as accuracy, sensitivity, specificity and the like.

Then, the method comprises the step of auxiliary diagnosis of depression detection: firstly, preprocessing the acquired electroencephalogram signal data, extracting effective characteristics to obtain a characteristic matrix f_newComputing a diagnosis-aided optimized lead feature subset S using the optimized lead weights_newAnd then inputting the obtained auxiliary diagnosis optimized lead characteristic subset into an optimized depression detection classification model, and obtaining a depression detection result through calculation: f (S)_new)＝W_modelk(S_new,S)+b。

The present invention uses a leave-one-test cross-validation method during the lead selection phase and the depression detection phase. The present invention has 80% depression detection recognition rate, 80% sensitivity and 80% specificity, and the present invention has depression recognition rate of only 65.71% without using the present invention for lead selection.

It should be noted that the above-mentioned embodiments enable a person skilled in the art to more fully understand the invention, without restricting it in any way. All technical solutions and modifications thereof without departing from the spirit and scope of the present invention are covered by the protection scope of the present invention.