A classification model construction method and system based on brain network normative modeling

Abstract

This invention belongs to the field of brain network technology and relates to a method and system for constructing a classification model based on standardized brain network modeling. The construction method includes: acquiring electroencephalogram (EEG) signal data of a subject; determining the partial orientation coherence adjacency matrix corresponding to the EEG signal data; processing the partial orientation coherence adjacency matrix according to a first formula to determine the individualized causal directed brain network partial orientation coherence matrix corresponding to the EEG signal data to be tested, wherein the feature path length and network density in the individualized causal directed brain network partial orientation coherence matrix are in an optimal balance state; predicting the individualized directed brain network partial orientation coherence matrix to be tested using a standardized benchmark model, and determining the deviation between the predicted value of the standardized benchmark model and the true value of the individualized directed brain network partial orientation coherence matrix; the standardized benchmark model is a model established based on the brain network adjacency matrix of healthy subjects; and constructing a classification model using the deviation as a classification feature.

Description

Technical Field

[0001] This invention relates to the field of brain network technology, specifically to a method and system for constructing a classification model based on brain network normative modeling. Background Technology

[0002] Major depressive disorder (MDD) is one of the most widespread mental illnesses, characterized by significant and persistent low mood, loss of interest, and lack of energy.

[0003] Currently, the diagnosis of depression mainly relies on interviews and scale assessments, but the results are inconsistent, with high rates of misdiagnosis and missed diagnosis, and a lack of objective diagnostic methods. Brain network methods can quantitatively characterize the information interaction characteristics in the brain networks of patients with malignant depression (MDD), potentially enabling objective identification of MDD. Existing research typically constructs brain network models for MDD patients and healthy subjects, then compares the means between different groups using intergroup statistical tests. Significantly different brain network features are then used for classification to construct a classification model, achieving objective detection and identification of MDD.

[0004] However, this method faces some limitations. On the one hand, when comparing brain networks among different groups, a sparsity threshold-based approach is often used, defining the sparsity range through small-world properties. This uniform threshold method, to some extent, ignores the heterogeneity between individuals, leading to inaccurate results. On the other hand, the method of comparing inter-group means relies excessively on the assumption of intra-group homogeneity, often ignoring the significant heterogeneity in individual pathological manifestations under the same diagnosis. This results in unreliable category labels output to classify individuals into disease groups.

[0005] Therefore, there is an urgent need for a classification model construction method based on brain network normative modeling, which extracts the most discriminative features to build a classification model and achieves accurate detection and identification of MDD. Summary of the Invention

[0006] To address the aforementioned technical issues, this application provides a method and system for constructing a classification model based on brain network normative modeling. By extracting the most discriminative features to construct a classification model, accurate identification of MDD can be achieved.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] In a first aspect, the present invention provides a method for constructing a classification model based on brain network canonical modeling, the method comprising the following steps:

[0009] Acquire the subject's electroencephalogram (EEG) signal data;

[0010] Determine the partial orientation coherence adjacency matrix corresponding to the EEG signal data;

[0011] The partial orientation coherence adjacency matrix is ​​processed according to the first formula to determine the individualized causal directed brain network partial orientation coherence matrix corresponding to the EEG signal data to be tested. The feature path length and network density in the individualized causal directed brain network partial orientation coherence matrix are in the optimal balance state.

[0012] The standard baseline model was used to predict the partial orientation coherence matrix of the individualized directed brain network to be tested, and the deviation between the predicted value of the standard baseline model and the true value of the partial orientation coherence matrix of the individualized directed brain network was determined. The standard baseline model is a model based on the adjacency matrix of the brain network of healthy subjects.

[0013] Construct a classification model that uses the deviation as a classification feature.

[0014] As one possible implementation, a classification model is constructed that uses the bias as a classification feature, including:

[0015] After applying a threshold to the deviation, extreme positive and extreme negative deviations are determined.

[0016] Based on extreme positive and extreme negative biases, a causal directed connection bias graph is generated.

[0017] Based on the number of edges with extreme positive and extreme negative deviations in the causal directed connection deviation graph, determine the proportion of extreme positive and extreme negative deviations in each edge.

[0018] Classification features are determined based on the proportions of extreme positive and extreme negative biases.

[0019] As one possible implementation, classification features are determined based on the proportions of extreme positive and extreme negative biases, including:

[0020] The extreme positive and extreme negative deviation ratios of each edge are compared with the corresponding extreme positive and extreme negative deviation ratios of the same edge in the healthy control group using a permutation test. When p is less than a preset threshold, the edge is determined as a classification feature.

[0021] As one possible implementation, after acquiring the subject's EEG signal data but before determining the partial orientation coherence adjacency matrix corresponding to the EEG signal data, the method further includes:

[0022] Preprocessing operations are performed on the EEG signal data. The preprocessing operations include at least one or more of the following: notch filtering, bandpass filtering, and artifact removal.

[0023] As one possible implementation, the partially oriented coherent adjacency matrix is ​​processed according to the first optimization formula, including:

[0024] Based on the constraints of the first formula, the sparsity threshold in the partial orientation coherent adjacency matrix is ​​determined.

[0025] Based on the sparsity threshold, the feature path length and network density under the sparsity threshold are determined.

[0026] As one possible implementation, the method also includes:

[0027] Obtain physiological covariates from healthy subjects; the physiological covariates should include at least one or more of age, sex, and years of education.

[0028] The standard baseline model is trained based on physiological covariates.

[0029] As one possible implementation, the deviation between the prediction bias of the canonical baseline model and the true value of the biased orientation coherence matrix of the individualized directed brain network is determined, including:

[0030] The bias is determined based on the second formula. In the second formula, the bias is negatively correlated with the predicted value of the canonical baseline model, positively correlated with the true value of the partial orientation coherence matrix of the individualized directed brain network, and negatively correlated with the prediction uncertainty and canonical variance of the canonical baseline model.

[0031] As one possible implementation, the individualized causal directed brain network partial orientation coherence matrix corresponding to the EEG signal data to be tested is determined, including:

[0032] The partial orientation coherence matrix of the individualized causal directed brain network is weighted, and the weight values ​​are the same as the partial orientation coherence values ​​in the partial orientation coherence adjacency matrix.

[0033] One possible approach is to acquire the subject's electroencephalogram (EEG) signal data, including:

[0034] Sound stimulation is applied to the test subject, and the test subject's electroencephalogram (EEG) signals are collected.

[0035] The EEG signal was windowed using the sliding window method to obtain the EEG signal data.

[0036] Secondly, this invention provides a classification model construction system based on brain network canonical modeling, the system comprising:

[0037] The acquisition module is used to acquire the subject's electroencephalogram (EEG) signal data;

[0038] The first determining module is used to determine the partial orientation coherence adjacency matrix corresponding to the EEG signal data;

[0039] The second determining module is used to process the partial orientation coherence adjacency matrix according to the first formula to determine the individualized causal directed brain network partial orientation coherence matrix corresponding to the EEG signal data to be tested, wherein the feature path length and network density in the individualized causal directed brain network partial orientation coherence matrix are in a balanced state.

[0040] The third determination module is used to predict the partial orientation coherence matrix of the individualized directed brain network to be tested using the standardized benchmark model, and to determine the deviation between the prediction bias value of the standardized benchmark model and the true value of the partial orientation coherence matrix of the individualized directed brain network; the standardized benchmark model is a model established based on the adjacency matrix of the brain network of healthy subjects.

[0041] The building module is used to construct classification models that use the deviation as a classification feature.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] 1. The classification model construction method and system based on brain network standard modeling proposed in this invention can address the problem that existing brain network threshold processing methods ignore individual heterogeneity. Based on the first optimization formula, optimization is performed to achieve an optimal balance between feature path length and network density. Here, feature network length can represent brain network efficiency, and network density can represent wiring cost. The individualized optimal balance causal directed brain network model constructed based on efficiency and cost solves the problem of no standard for threshold selection to a certain extent.

[0044] 2. The classification model construction method based on brain network normative modeling proposed in this invention can overcome the limitation of traditional inter-group mean comparison relying on intra-group homogeneity, fully consider the heterogeneity between individuals, characterize the degree of individualized variation deviation, and achieve objective detection and identification of MDD.

[0045] 3. The classification model construction method and system based on brain network normative modeling proposed in this invention have discovered potential diagnostic biomarkers for MDD and have important scientific research and clinical value. Attached Figure Description

[0046] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0047] Figure 1 A flowchart illustrating a classification model construction method based on brain network canonical modeling provided in an embodiment of the present invention;

[0048] Figure 2 A structural block diagram of a classification model construction system based on brain network canonical modeling provided in an embodiment of the present invention;

[0049] Figure 3 The graph shows the change of network density as a function of λ after the gamma-ASSR network topology provided in the embodiment of the present invention reaches equilibrium.

[0050] Figure 4 This is a brain network bias connectivity diagram of MDD provided in an embodiment of the present invention. Detailed Implementation

[0051] To facilitate a clear description of the technical solutions in the embodiments of the present invention, the terms "first" and "second" are used to distinguish identical or similar items with essentially the same function and effect. For example, the first threshold and the second threshold are merely used to distinguish different thresholds and do not limit their order. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that the terms "first" and "second" are not necessarily different.

[0052] It should be noted that in this invention, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0053] In this invention, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one" or similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, "at least one of a, b, or c" can represent: a, b, c, a combination of a and b, a combination of a and c, a combination of b and c, or a, b, and c, where a, b, and c can be single or multiple.

[0054] This invention aims to provide a method and system for constructing a classification model based on canonical brain network modeling. By optimizing the partial orientation coherence adjacency matrix corresponding to EEG signal data, an optimal balance between feature path length and network density is achieved. A canonical baseline model is used to predict the partial orientation coherence matrix of individualized directed brain networks, determining the bias. Furthermore, the classification features of the classification model are determined based on the bias. Since the positions of edges with significantly different extreme bias overlap ratios can be located during feature extraction, accurate identification of partial orientation disorder (MDD) can be achieved.

[0055] In a first aspect, embodiments of the present invention provide a method for constructing a classification model based on brain network canonical modeling, see [link to relevant documentation]. Figure 1 The method may include the following steps:

[0056] S110: Acquire the subject's electroencephalogram (EEG) signal data.

[0057] For example, EEG signal data can be acquired through acoustic stimulation. For instance, EEG signal data of subjects under 40 Hz chirp acoustic stimulation can be collected. The subjects include healthy and diseased groups.

[0058] As an optional implementation, step S110 may include:

[0059] Aural stimulation was applied to the test subject, and the electroencephalogram (EEG) signals of the test subject were collected.

[0060] The EEG signal is windowed using a sliding window method to obtain the EEG signal data.

[0061] In this embodiment of the invention, 40 Hz chirp sound stimulation can be used when collecting EEG signal data. The entire experiment is conducted in a quiet environment shielded from electromagnetic interference. Subjects are asked to sit in a comfortable chair. After being informed of the experimental procedure, subjects are reminded to minimize distracting movements such as head shaking, teeth clenching, and leg shaking during the experiment. Subjects are instructed to keep their eyes open. Subjects receive auditory steady-state stimulation through Sony WH-1000XM3 active noise-canceling headphones.

[0062] The high-frequency auditory task experiment in this embodiment of the invention uses a 40 Hz short-term chirp signal paradigm with a signal duration of 1 millisecond and a sound pressure level of 45 dB, which conforms to the cochlear traveling wave delay compensation principle. To further ensure the accuracy of the salt input, the experimenter tests the headphone sound before the experiment to ensure that the sound stimulus can be heard. After the subject presses the start button, the system emits a "beep" sound, and then the stimulation begins, consisting of 28 stimulation trials. The duration of each stimulus is 3 seconds, and the isolation time between stimuli is precisely controlled to 1.5 seconds.

[0063] After acoustic stimulation, electrode caps can be used to acquire data. For example, the Neuroscan SYNAMPS2 amplifier can be used to acquire EEG data, with 62-lead electrode caps placed according to the international standard 10-20 system. Using the left mastoid process (M1) as a reference, the sampling rate is set to 1000Hz. All electrode impedances are controlled below 10kΩ using EEG gel to ensure signal quality before experimental data acquisition.

[0064] To further ensure the purity of the data and eliminate signal interference, the EEG data in step S110 can be preprocessed.

[0065] As an optional implementation, after step S110, the classification model construction method based on brain network canonical modeling provided in this embodiment of the invention may further include:

[0066] Preprocessing operations are performed on the EEG signal data. The preprocessing operations include at least one or more of the following: notch filtering, bandpass filtering, and artifact removal.

[0067] Because some EEG signal data contains 100 Hz and 150 Hz harmonic power frequency interference, notch filtering at 100 Hz and 150 Hz can be applied to the EEG signal data.

[0068] Because some EEG signal data contain interference signals that do not originate from the activity of brain nerve cells, such as: electrooculography artifacts caused by strong electrical signals generated by blinking and eye movement; electromyography artifacts caused by facial and neck muscle contractions when frowning, clenching teeth, and swallowing; electrocardiogram artifacts caused by weak electrical signals from the heartbeat; and channel artifacts caused by slight background noise generated by the electronic device itself. In practical applications, principal component analysis (PCA) can be performed on the EEG signal data first. Then, the ICLabel plugin in the EEGLAB toolkit can be used to remove artifacts from the data. The ICLabel exclusion criteria are: components labeled "EEG", "EMG", "ECG", "linear noise", and "channel noise" with a probability greater than 70% are removed. Finally, using the stimulus start label as a baseline, a data segment is extracted from 1 second before the stimulus start to 3 seconds after the stimulus start. The baseline removal time range is set between -1 and 0 seconds, resulting in a 60*3000*28 EEG data matrix for each subject, where 60 represents the number of electrodes, 3000 represents the EEG data points 3 seconds after the stimulus start label, and 28 represents 28 trials. The frequency range of interest is set to the gamma band (30-50 Hz).

[0069] Bandpass filtering refers to the selective allowing or blocking of signals within a specific frequency range using electronic or mathematical methods. In practical applications, the `pop_eegfiltnew` function can be used to perform bandpass filtering on data from 0.1 to 70 Hz.

[0070] After preprocessing the brain signals, the acquired EEG signals can be windowed using a sliding window method. In practical applications, the partial directed coherence (PDC) adjacency matrix has 1000 frequency points, representing a frequency range from 0 to 100 Hz. The number of time windows is fixed at 25 (3 seconds per trial, window length 0.6 s, step size 0.1 s). The frequency band of interest is set to 30-50 Hz, and the frequency points representing 30-50 Hz are selected and averaged.

[0071] S130: Determine the partial orientation coherence adjacency matrix corresponding to the EEG signal data.

[0072] Partially Oriented Coherence Adjacency Matrix (PDC) can be used to construct causal directed brain networks. For a specific time window and a specific brainwave frequency, the PDC value is a number between 0 and 1. The PDC value indicates the extent to which the activity information of the "source" brain region (sender) has a direct, directional causal influence on the subsequent activity information of the "target" brain region (receiver), after excluding the influence of activity in all other brain regions. The larger the value, the stronger the direct driving effect of the "source" on the "target" at that moment and rhythm.

[0073] S150: Process the partial orientation coherence adjacency matrix according to the first formula to determine the individualized causal directed brain network partial orientation coherence matrix corresponding to the EEG signal data to be tested. The feature path length and network density in the individualized causal directed brain network partial orientation coherence matrix are in the optimal balance state.

[0074] As an optional implementation, the partially oriented coherent adjacency matrix is ​​processed according to the first optimization formula, including:

[0075] Based on the constraints of the first formula, the sparsity threshold in the partial orientation coherent adjacency matrix is ​​determined.

[0076] Based on the sparsity threshold, the feature path length and network density under the sparsity threshold are determined.

[0077] The operational mode of brain networks is to minimize wiring costs while maximizing network efficiency. This cost-efficiency competitive optimization model is quantified in the connectivity landscape model. By establishing a competitive relationship between modularity and integration, the network achieves co-optimization through mutual trade-offs. To express this competitive relationship, this embodiment of the invention introduces a first formula to process the partially oriented coherent adjacency matrix.

[0078] For example, the first formula can be:

[0079]

[0080] in, . It is control and The parameters of a linear combination of two features. The normalized feature path length represents network efficiency, i.e., integration, and is defined as follows:

[0081]

[0082] In the above formula, This represents the characteristic path length of the causal directed brain network. The maximum feature path length achievable by a connected network is given by the following formula:

[0083]

[0084] and This represents network density, which indicates cabling costs, i.e., the degree of modularity, and is defined as follows:

[0085]

[0086] In the above formula, Let be the number of edges in the network. Represents the number of nodes.

[0087] The optimization process of network topology aims to minimize This means that while reducing the length of the feature path and increasing network efficiency, the network density is reduced and the wiring cost is decreased, thus reflecting the competitive trade-off between efficiency and cost.

[0088] In one optional implementation, the partially oriented coherent adjacency matrix is ​​processed according to a first optimization formula, including:

[0089] Based on the constraints of the first formula, the sparsity threshold in the partial orientation coherent adjacency matrix is ​​determined.

[0090] Based on the sparsity threshold, the feature path length and network density under the sparsity threshold are determined.

[0091] It should be noted that the constraints of the first formula can be given parameters. A sparsity thresholding method is used to retain the top T% of edges based on their connection strength. Initial threshold. =100, and the sparsity threshold T is decreased by a step size of 0.1. The sparsity is calculated successively for each threshold T. , and Minimize The value of T at that time is used to obtain the equilibrium state. , And the corresponding adjacency matrix.

[0092] As an optional implementation, after determining the individualized causal directed brain network partial orientation coherence matrix corresponding to the EEG signal data to be tested, the method further includes:

[0093] The partial orientation coherence matrix of the individualized causal directed brain network is weighted, and the weight values ​​are the same as the partial orientation coherence values ​​in the partial orientation coherence adjacency matrix.

[0094] It should be noted that the final adjacency matrix is ​​a binary adjacency matrix, which is then converted into a weighted adjacency matrix. The weights are consistent with the PDC values ​​at the corresponding positions in the fully connected matrix. An optimized adjacency matrix is ​​generated for each subject in each time window, which is the individual-specific causal directed brain network. The adjacency matrix of this individual-specific causal directed brain network is used for the subsequent canonical modeling.

[0095] S170: The standard baseline model is used to predict the partial orientation coherence matrix of the individualized directed brain network to be tested, and the deviation between the predicted value of the standard baseline model and the true value of the partial orientation coherence matrix of the individualized directed brain network is determined; the standard baseline model is a model based on the adjacency matrix of the brain network of healthy subjects.

[0096] The canonical baseline model uses the adjacency matrix of the brain networks of healthy subjects as the training set and employs a wrapped Bayesian linear regression algorithm for model estimation. This embodiment of the invention uses the open-source Predictive Clinical Neuroscience Toolkit (PCN toolkit) for canonical modeling. The canonical baseline model is trained using the adjacency matrix of the brain networks of 71 healthy subjects as the training set (HCtrain).

[0097] The goodness of fit of the standard baseline model is evaluated by explained variance (EV), mean standardized log loss (MSLL), standardized meansquared error (SMSE), skewness, and kurtosis.

[0098] As an optional implementation, when modeling the baseline model, the influence of covariates on the model should also be considered. The method further includes:

[0099] Obtain physiological covariates from healthy subjects; the physiological covariates include at least one or more of age, sex, and years of education.

[0100] Due to HC train The sample's age range is 23-54 years old. This embodiment of the invention introduces a nonlinear transformation function. Warp Sin Arcsinh and basis functions B-spline It models the nonlinear relationship between covariates and each PDC value in a causal directed brain network, with a modeling range of -5 to 110. It has the ability to estimate trend continuity in the covariate space, and this modeling approach allows for reasonable estimation of individuals outside the training data range.

[0101] After standardizing the modeling, the deviation between the model's predicted results and the actual causal directed brain network PDC values ​​in the test set is estimated using test set data.

[0102] S190: Construct a classification model that uses the deviation as a classification feature.

[0103] The classification model in this embodiment of the invention can be constructed using three common machine learning classification algorithms: linear support vector machine (SVM), random forest (RF), and linear discriminator (LDA).

[0104] The classification model construction method based on brain network normative modeling provided in this embodiment of the invention can better consider the heterogeneity between individuals and characterize the degree of deviation of individualized variation by using the deviation between samples of different groups of subjects as classification features, so as to achieve objective detection and identification of MDD.

[0105] As an optional implementation, a classification model is constructed using the deviation as a classification feature, including:

[0106] After applying a threshold to the deviation, extreme positive and extreme negative deviations are determined.

[0107] Based on extreme positive and extreme negative biases, a causal directed connection bias graph is generated.

[0108] Based on the number of edges with extreme positive and extreme negative deviations in the causal directed connection deviation graph, determine the proportion of extreme positive and extreme negative deviations in each edge.

[0109] Classification features are determined based on the proportions of extreme positive and extreme negative biases.

[0110] As an optional implementation, determining the deviation between the prediction bias of the standard baseline model and the true value of the partial orientation coherence matrix of the individualized directed brain network includes:

[0111] The bias is determined based on the second formula. In the second formula, the bias is negatively correlated with the predicted value of the canonical baseline model, positively correlated with the true value of the partial orientation coherence matrix of the individualized directed brain network, and negatively correlated with the prediction uncertainty and canonical variance of the canonical baseline model.

[0112] As an example, the bias can be represented by the Z-score. This applies to each subject for each time window sample. i In a causal directed brain network, each edge (i.e., the PDC value between each pair of nodes) j Above, combined with the predicted PDC value The actual PDC value Forecast uncertainty and normalized variance Calculate the Z-score The second formula is as follows, which quantifies the deviation of an individual's PDC value from the predicted value of the normative baseline model:

[0113]

[0114] in, This reflects the degree to which the individual sample's PDC deviates from the model's prediction, while also taking into account the model's uncertainty. Between any pair of nodes, this embodiment of the invention focuses on samples that significantly deviate from the canonical baseline model's prediction, assuming that these deviations reflect the pathophysiological characteristics of the disease.

[0115] To characterize the significant positive or negative bias of individual sample PDC values ​​under the canonical model, this embodiment of the invention employs two complementary methods. The absolute value of the preset threshold can be 1.65, i.e., a Z-score greater than 1.65 or less than -1.65 (corresponding to a one-tailed test in statistics). P <0.05) corresponds to extreme positive and extreme negative deviations, respectively, which is used to filter the deviation mapping.

[0116] Assuming the brain network signal data of the subjects is collected from 60 nodes, after thresholding the Z-score, a causal directed connectivity extreme bias map is generated for each subject. Excluding the 60 diagonal edges, there are a total of 60 × 59 = 3540 edges, corresponding to the extreme bias Z-scores for each edge. The number of edges with extreme positive and extreme negative biases for each subject in each time window is calculated. Next, the proportion of extreme positive and extreme negative biases in each edge is calculated, resulting in extreme positive and extreme negative bias overlap maps. This is used to analyze the heterogeneity among individuals. The extreme bias sample overlap map visually represents the heterogeneity between samples from different groups of subjects. The larger the overlap ratio on a certain edge, the more samples exhibiting extreme bias in that group, thus reflecting the common characteristics of most samples in that group. Conversely, if the overlap ratio on all edges is small, it indicates that samples exhibiting extreme bias within the group are widely distributed across different edges, revealing significant heterogeneity among individuals.

[0117] To further characterize the intergroup differences in the overlap ratio of extreme biases in causal directed brain network connectivity among different disease groups, this embodiment of the invention uses a nonparametric statistical method based on permutation tests.

[0118] As one possible implementation, determining classification features based on the proportions of extreme positive and extreme negative biases can include:

[0119] The extreme positive and extreme negative deviation ratios of each edge are compared with the corresponding extreme positive and extreme negative deviation ratios of the same edge in the healthy control group using a permutation test. When p is less than a preset threshold, the edge is determined as a classification feature.

[0120] Specifically, for each group, the overlap ratio of extreme positive and negative biases for each edge is calculated, which represents the proportion of subject samples exhibiting extreme biases within that group. Then, the overlap ratio of the patient group is subtracted from the corresponding overlap ratio of the healthy control test group to obtain the disease-control difference overlap plot. Next, the group labels (i.e., HCtest and patient group labels) are randomly permuted on a per-subject basis rather than on a per-time-window sample basis, and this process is repeated 10,000 times. An empirical distribution is constructed using the null hypothesis that there is no significant difference between the patient and control groups for hypothesis testing. For each edge, the p-value is defined as the proportion in the permuted distribution that exceeds the observed true inter-group difference. Finally, this invention employs FDR correction, using a corrected two-tailed p-value test. FDR <0.05 is used as the criterion for statistical significance.

[0121] After permutation test and FDR correction, edges with significant differences in the overlap ratio of extreme positive bias and extreme negative bias between the disease group and the control group are obtained. In this embodiment of the invention, the positions of these edges with significant differences in the overlap ratio of extreme bias (including the overlap ratio of extreme positive bias and the overlap ratio of extreme negative bias) are further located, and the union of the edges with significant differences in the disease group is taken. The deviation Z score of these edges is used as the classification feature.

[0122] This invention employs the ReliefF algorithm for feature extraction in classification. The ReliefF algorithm evaluates the contribution of each feature to classifying a category by repeatedly comparing the feature values ​​of a random sample with its neighbors of the same and different categories, and assigns weights to these features. Features with higher weights are considered more important. The features selected through ReliefF will be further used to train the classification model to improve classification performance.

[0123] Secondly, embodiments of the present invention provide a classification model construction system based on brain network canonical modeling, referencing... Figure 2 The system includes:

[0124] Acquisition module 210 is used to acquire the subject's electroencephalogram (EEG) signal data;

[0125] The first determining module 230 is used to determine the partial orientation coherent adjacency matrix corresponding to the EEG signal data;

[0126] The second determining module 250 is used to process the partial orientation coherence adjacency matrix according to the first formula to determine the individualized causal directed brain network partial orientation coherence matrix corresponding to the EEG signal data to be tested, wherein the feature path length and network density in the individualized causal directed brain network partial orientation coherence matrix are in an optimal balance state.

[0127] The third determining module 270 is used to predict the partial orientation coherence matrix of the individualized directed brain network to be tested using a standardized benchmark model, and to determine the deviation between the prediction bias value of the standardized benchmark model and the true value of the partial orientation coherence matrix of the individualized directed brain network; the standardized benchmark model is a model established based on the adjacency matrix of the brain network of healthy subjects.

[0128] Module 290 is used to construct a classification model that uses the deviation as a classification feature.

[0129] Next, combine Figure 3 and Figure 4 The embodiments of this application will be further described. Figure 3 This is a graph showing the change in network density as a function of λ after the gamma-ASSR network topology reaches equilibrium. Figure 3It can be seen that the network density remains almost constant when λ is between 0.1 and 0.92, and the network structure tends to be stable.

[0130] Figure 4 This is a brain network bias connectivity diagram for MDD. (Through...) Figure 4 It can be seen that the classification model constructed using the embodiments of this application can effectively identify the significant connection results between the MDD group and the healthy subject group.

[0131] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the description of the drawings, in carrying out the claimed invention. In this specification, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple components. A single processor or other unit can implement several of the functions listed in the specification. While certain measures are described in different embodiments, this does not mean that these measures cannot be combined to produce good results.

[0132] Although the invention has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made therein without departing from the spirit and scope of the invention. Accordingly, this specification and drawings are merely illustrative of the invention and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if such modifications and modifications fall within the scope of the invention and its equivalents, the invention is also intended to include such modifications and modifications.

Claims

1. A method for constructing a classification model based on brain network canonical modeling, characterized in that, Includes the following steps: Acquire the subject's electroencephalogram (EEG) signal data; Determine the partial orientation coherence adjacency matrix corresponding to the EEG signal data of the test subject; The partial orientation coherence adjacency matrix is ​​processed according to the first formula to determine the individualized causal directed brain network partial orientation coherence matrix corresponding to the EEG signal data of the test subject. The feature path length and network density in the individualized causal directed brain network partial orientation coherence matrix are in an optimal balance state. The partial orientation coherence matrix of the individualized directed brain network to be tested is predicted using a standardized benchmark model, and the deviation between the predicted value of the standardized benchmark model and the true value of the partial orientation coherence matrix of the individualized directed brain network is determined; the standardized benchmark model is a model established based on the adjacency matrix of the brain network of healthy subjects. Construct a classification model using the aforementioned deviation as a classification feature; The construction of a classification model using the deviation as a classification feature includes: After applying a threshold to the deviation, extreme positive and extreme negative deviations are determined. Based on the extreme positive bias and the extreme negative bias, a causal directed connection bias graph is generated; Based on the number of edges with extreme positive and extreme negative deviations in the causal directed connection deviation graph, the proportion of extreme positive deviations and the proportion of extreme negative deviations in each edge are determined. The classification features are determined based on the extreme positive bias ratio and the extreme negative bias ratio. The determination of the classification features based on the extreme positive bias ratio and the extreme negative bias ratio includes: The extreme positive deviation ratio and extreme negative deviation ratio of each edge are compared with the corresponding extreme positive deviation ratio and extreme negative deviation ratio of the same edge in the healthy control group. When p is less than a preset threshold, the edge is determined to be a classification feature.

2. The classification model construction method based on brain network canonical modeling according to claim 1, characterized in that, After acquiring the subject's electroencephalogram (EEG) signal data, and before determining the partial orientation coherence adjacency matrix corresponding to the EEG signal data, the method further includes: The EEG signal data is preprocessed, and the preprocessing operation includes at least one or more of notch filtering, bandpass filtering, and artifact removal.

3. The classification model construction method based on brain network canonical modeling according to claim 1, characterized in that, The partial orientation coherent adjacency matrix is ​​processed according to the first formula, including: Based on the constraints of the first formula, the sparsity threshold in the partial orientation coherent adjacency matrix is ​​determined. Based on the sparsity threshold, the feature path length and network density under the sparsity threshold are determined.

4. The classification model construction method based on brain network canonical modeling according to claim 1, characterized in that, The method further includes: Obtain physiological covariates from the healthy subjects; the physiological covariates include at least one or more of age, sex, and years of education. The standard baseline model is trained based on the physiological covariates.

5. The classification model construction method based on brain network canonical modeling according to claim 1, characterized in that, The determination of the deviation between the prediction bias of the canonical baseline model and the true value of the biased orientation coherence matrix of the individualized directed brain network includes: The deviation is determined based on the second formula, in which the deviation is negatively correlated with the predicted value of the canonical baseline model, positively correlated with the true value of the biased orientation coherence matrix of the individualized directed brain network, and negatively correlated with the prediction uncertainty and canonical variance of the canonical baseline model.

6. The classification model construction method based on brain network canonical modeling according to claim 1, characterized in that, Determining the individualized causal directed brain network partial coherence matrix corresponding to the EEG signal data of the test subject includes: The individualized causal directed brain network partial orientation coherence matrix is ​​weighted, and the weight values ​​are the same as the partial orientation coherence values ​​in the partial orientation coherence adjacency matrix.

7. The classification model construction method based on brain network canonical modeling according to claim 1, characterized in that, The acquisition of the subject's electroencephalogram (EEG) signal data includes: Sound stimulation was applied to the test subject, and the electroencephalogram (EEG) signals of the test subject were collected. The EEG signal is windowed using a sliding window method to obtain the EEG signal data.

8. A classification model construction system based on brain network canonical modeling, characterized in that, The classification model building system is used to execute the classification model building method of any one of claims 1 to 7; The classification model building system includes: The acquisition module is used to acquire the subject's electroencephalogram (EEG) signal data; The first determining module is used to determine the partial orientation coherent adjacency matrix corresponding to the EEG signal data; The second determining module is used to process the partial orientation coherence adjacency matrix according to the first formula to determine the individualized causal directed brain network partial orientation coherence matrix corresponding to the EEG signal data, wherein the feature path length and network density in the individualized causal directed brain network partial orientation coherence matrix are in an optimal balance state. The third determination module is used to predict the partial orientation coherence matrix of the individualized directed brain network to be tested using a standardized benchmark model, and to determine the deviation between the prediction bias value of the standardized benchmark model and the true value of the partial orientation coherence matrix of the individualized directed brain network; the standardized benchmark model is a model established based on the adjacency matrix of the brain network of healthy subjects. A construction module is used to construct a classification model that uses the deviation as a classification feature.

Images