A precise intervention and prediction method for rTMS in depression and anxiety disorders

Abstract

The present invention discloses an rTMS precise intervention and prediction method for depression and anxiety disorders, including: using a cognitive reappraisal task to induce the active emotion regulation process of patients with major depressive disorder and generalized anxiety disorder; during the process of inducing the active emotion regulation process of patients by the cognitive reappraisal task, collecting fMRI and EEG data, and obtaining fused fMRI-EEG data after processing; using a dynamic causal model DCM to find the point with the strongest functional connection between the DLPFC and the amygdala of each patient as the rTMS intervention target; performing individualized rTMS intervention on patients with unipolar and bipolar depression for 2 weeks and evaluating the curative effect, and collecting EEG data during the intervention process to monitor brain activities in real time; extracting imaging brain network indexes before and after the intervention to establish a curative effect prediction model. The present invention improves the curative effect of the rTMS intervention method for patients with depression and anxiety disorders by precisely locating the rTMS intervention target and monitoring brain activities in real time during the intervention process; establishing a curative effect prediction model to realize the precise diversion of clinical treatment of depression and anxiety patients.

Description

Technical Field

[0001] The present invention relates to the technical field of neuromodulation, and particularly relates to a method for precise intervention and prediction of rTMS for depression and anxiety disorders. Background Art

[0002] Major Depressive Disorder (MDD) is a depressive disorder characterized by typical symptoms such as depressed mood, loss of interest or pleasure, and recurrent thoughts of death (Marx et al., 2023). The quality of life of MDD patients is significantly lower than that of normal people, which is not only caused by the disease itself and the resulting impairments in social, occupational, and other important functions (American Psychiatric Association, 2013), but also the result of the frequent comorbidity of MDD with multiple other mental illnesses (Marx et al., 2023). 43% of MDD patients will comorbid with Generalized Anxiety Disorder (GAD) (Kalin, 2020). GAD is an anxiety disorder mainly characterized by excessive anxiety and worry about many events or activities (such as work or school performance), accompanied by physical symptoms such as muscle tension, fatigue, reduced ability to concentrate, and sleep disorders (American Psychiatric Association, 2013). Comorbidity not only leads to more severe symptoms in patients than those of a single disease (Zhou et al., 2017), but also makes treatment difficult. 53% of MDD patients with anxiety have significantly higher treatment difficulty than ordinary patients and greater treatment side effects (Fava et al., 2008). The comorbidity of MDD and GAD is not only difficult to distinguish in terms of symptom manifestations (Kalin, 2020), but the neural mechanisms they rely on also have similarities. Studies have shown that the two mental illnesses have similar lesion patterns in the salience network (McTeague et al., 2020). Repeated Transcranial Magnetic Stimulation (rTMS) is a non-invasive, safe, and effective intervention for treating mental and neurological diseases (Lin et al., 2022). Using rTMS to intervene in different targets of the left dorsolateral prefrontal cortex (DLPFC) can regulate different downstream networks, and the regulation of these different downstream networks has been proven to be closely related to the alleviation of different depressive and anxiety symptoms (Kalin, 2020). The development of brain imaging information processing technology has provided an opportunity for the adjuvant treatment of MDD and GAD. Neuroregulation techniques based on neuroimaging and neuroelectrophysiology to lock key circuits and construct individualized brain maps have been proven to be effective in the treatment of mental illnesses such as MDD (Alhelali et al., 2022; Goldwaser et al., 2021) and GAD (Parikh et al., 2022).In addition, the continuous progress of brain-computer interface technology has opened up new avenues for the adjuvant treatment of depressive disorders and anxiety disorders.

[0003] Currently, the determination of rTMS intervention targets mainly relies on two technologies: methods based on craniofacial surface measurement and methods based on brain imaging. For methods based on craniofacial surface measurement, since TMS itself cannot obtain the anatomical structure information of the cerebral cortex, it is challenging to accurately locate specific cortical targets. For existing methods based on brain imaging, most rTMS interventions for MDD and GAD often locate intervention targets based on calculating the brain functional connection network with rs-fMRI. However, the instability characteristics of resting-state brain functional activities and the lack of attention to key brain regions for emotion regulation in the calculation of global brain functional activities may lead to poor precision and individualized locking effects on neural functional circuits.

[0004] During the rTMS intervention process, existing practices often use neuro-navigation technology as a means to ensure the effectiveness of intervention targets, and monitor and adjust the position of the stimulation coil through an infrared camera (Neacsiu et al., 2018). However, most current neuro-navigation systems display intervention targets by combining the structural MRI data of patients, lacking the monitoring of brain function during the rTMS intervention and the dynamic visualization of intervention targets, resulting in the difficulty of implementing an intervention paradigm that evaluates and adjusts rTMS intervention parameters based on real-time monitoring of patients' brain function, and unable to ensure the effect of rTMS intervention on improving brain function.

[0005] In terms of constructing a treatment efficacy prediction model, although there have been attempts to construct machine learning (Hasanzadeh et al., 2019; Cash et al., 2019) and deep learning (Lin et al., 2018; Mehltretter et al., 2019; Chang et al., 2019) prediction models for rTMS intervention efficacy, most existing results use machine learning algorithms, and the models for predicting rTMS intervention efficacy based on single-modal MRI or EEG signals lack the fusion analysis of fMRI and EEG data, and cannot ensure the accuracy of model prediction results from both time and space scales. Summary of the Invention

[0006] The object of the present invention is to provide an rTMS precise intervention and prediction method for depressive and anxiety disorders to solve the above problems.

[0007] To achieve the above object, the technical solution adopted by the present invention includes the following steps:

[0008] S1: Using a cognitive reappraisal task, the pictures presented to the patients were from the International Affective Picture System, including 64 negative pictures and 32 neutral pictures. The patients were required to perform "viewing" or "cognitive reappraisal" processing tasks on the pictures, and the IDS-30 and HAMA were used to measure the severity of depressive and anxiety symptoms respectively.

[0009] S2: During the process of inducing the patients' active emotion regulation by the cognitive reappraisal task, fMRI and EEG data were collected synchronously. SPM 12 was used to preprocess the fMRI data, and EEGLAB was used to preprocess the EEG.

[0010] S3: Using SPM 12, the fMRI data and EEG data were registered to the MNI standard brain uniformly to ensure that the fMRI and EEG data were in the same coordinate system. The ROI_MNI_v4 version of the AAL atlas was used to map the fMRI and EEG data to the same brain regions and spatial coordinate system, and the fMRI and EEG data were merged to obtain the fused fMRI-EEG data.

[0011] S4: Define a generalized linear model, and select the left DLPFC, right DLPFC, left amygdala, and right amygdala as VOIs to construct a full DCM model for DCM analysis to determine the rTMS intervention target.

[0012] S5: Conduct individualized rTMS intervention on the patients for 2 weeks.

[0013] When evaluating the efficacy of the patients after rTMS intervention, the reduction rate of the total scores of the IDS-30 and HAMA scales before and after the intervention was used to determine whether the rTMS intervention was effective. A reduction rate of ≥50% was considered effective, and <50% was considered ineffective. When evaluating the intervention efficacy of MDD patients, the IDS-30 was mainly referred to, and when evaluating the intervention efficacy of GAD patients, the HAMA was mainly referred to. For comorbid patients, both the IDS-30 and HAMA were referred to. The cognitive reappraisal task was used to judge the improvement of the patients' functions before and after rTMS intervention, and fMRI and EEG data were collected synchronously.

[0014] SPM 12 was used to preprocess the post-test fMRI data, and EEGLAB was used to preprocess the post-test EEG.

[0015] S6: Extract the imaging brain network indexes before and after the intervention, including the degree of nodes, betweenness centrality, global efficiency, etc., and perform feature selection to establish a prediction model, and output the brain network index with the largest change after rTMS intervention and the prediction of whether the patient is effective after receiving rTMS intervention.

[0016] Among them, a composite deep learning model including a 3D CNN branch and an RNN branch is established. The 3D CNN is used to process fMRI data, and the RNN is used to process EEG data. Finally, a joint branch is used to fuse the features of the two types of data. The matching Python IDE editor is Pycharm. The specific steps are as follows:

[0017] Step1: Construct a 3D CNN branch to process fMRI data, including the following layers:

[0018] Input layer: Input fMRI data, which is a three-dimensional tensor;

[0019] Two CNN layers: Input the output signal of the input layer into two convolutional layers. The number of convolutional kernels is 8 and 16 respectively, and the shape is (3, 3). Use tanh as the activation function. Each CNN layer uses L2 regularization to prevent overfitting;

[0020] Two max-pooling layers: The pooling window size is 2×2×2, and the stride is 2. Each max-pooling layer uses L2 regularization to prevent overfitting;

[0021] Dropout layer: The dropout probability is 0.25;

[0022] Two fully connected layers: The two fully connected layers use softmax as the activation function, and the number of nodes is 64 and 2 respectively. The first fully connected layer performs batch normalization on the data; the second fully connected layer outputs the possibility that the patient belongs to the severe depressive disorder group or the generalized anxiety group;

[0023] Dropout layer: The dropout probability is 0.25;

[0024] Use the cross-entropy function as the loss function, and adopt the following method (Lin et al., 2022): For the sampled data I i (x, y, z) and its label M is the number of samples, and the CNN output result is W, b are model parameters, h W,b (·) represents the model function, and the loss calculation formula is:

[0025]

[0026] Step2: Construct an RNN branch to process EEG data, including the following steps:

[0027] Construct an input layer to input EEG data, with a size of 32×32;

[0028] Build 9 CNN layers with a convolutional kernel shape of (3,3), a stride of 1, use the ReLU function as the activation function, and perform zero-padding of one pixel in the middle convolutional layers;

[0029] Add a pooling layer after the 4th, 8th, and 9th convolutional layers. Select max pooling as the pooling type, with a pooling window size of 2×2 and a stride of 2, and use the ReLU function as the activation function;

[0030] Build a flatten layer to flatten the data into a one-dimensional tensor;

[0031] Build two fully connected layers with 512 and 128 neurons respectively, and use ReLU as the activation function;

[0032] Build an LSTM layer with 128 neurons;

[0033] Build an output layer to output the binary classification prediction results of the model, and use the softmax function as the activation function;

[0034] Step 3: Build a feature fusion layer to fuse the features of fMRI and EEG data, including the following steps:

[0035] Build a convolutional layer and a pooling layer to extract the shared features of the two modalities. Among them, the convolutional layer has 32 convolutional kernels, with a convolutional kernel shape of (3,3), and uses ReLU as the activation function; select max pooling as the pooling type, and the convolutional kernel shape of the pooling layer is (3,3);

[0036] Design two branches. Branch 1: Use a fully connected layer with the ReLU activation function to output the brain network metrics that change the most after rTMS intervention;

[0037] Branch 2: Use a fully connected layer with two nodes and the softmax as the activation function. These two nodes represent the effective and ineffective categories respectively, and output the prediction of whether the patient is effective after receiving rTMS intervention.

[0038] Furthermore, when collecting fMRI data and EEG data, use the 10-20 international standard lead system to set 64 sampling electrodes to obtain the EEG data of the cognitive reappraisal task; the fMRI parameters are as follows: [TR]=2000ms, [TE]=30ms, voxel size is 3×3×3mm 3 , matrix shape is 90×90, number of layers is 60, a total of 180 time points are collected, and the scanning duration is 6 minutes.

[0039] Furthermore, the preprocessing of the EEG data includes steps such as electrode positioning, deleting useless electrodes, band-pass filtering, downsampling rate, data segmentation and baseline correction, removing bad segments, performing ICA, and artifact removal. The preprocessing of the fMRI data includes excluding motion artifacts, fluctuations, etc. using MRIQC 23.2.0; performing temporal slice correction using SPM 12; performing head motion correction using SPM 12, with rigid motion correction to correct the motion of images at all time points; mapping the fMRI data to the MNI standard brain; using spatial smoothing to improve the signal-to-noise ratio; using a 0.01 Hz high-pass filter to remove trends and low-frequency drifts in the data; regressing the fMRI data using the 6 rigid parameters of motion correction to remove motion-related signals; removing cerebrospinal fluid and white matter signals.

[0040] Furthermore, the DCM analysis includes selecting the VOI time series of the left DLPFC, right DLPFC, left amygdala, and right amygdala extracted from all patients, specifying the A matrix, B matrix, and C matrix, performing model estimation, and selecting the brain region with the strongest connection to the amygdala as the rTMS intervention target.

[0041] Furthermore, when performing individualized rTMS intervention, an online EEG parsing algorithm is combined to make up for the deficiency of the fMRI time resolution. The online EEG parsing algorithm includes time-domain filtering, CSP, online PCA, and online ICA, and a source localization analysis model is used to perform weighted minimum norm estimation inversion and solve the source analysis for the EEG data. Through the online EEG parsing algorithm, it is monitored in real time to ensure that the target brain region activated during the rTMS stimulation is consistent with the brain region with the strongest atlas connection determined by the fMRI, so as to ensure the high spatio-temporal resolution of the stimulation target area. The rTMS stimulation intensity is 100% of the resting motor threshold (rMT), the stimulation frequency is 10 Hz, the stimulation duration is 1 s, the interval is 10 s, the intervention lasts for 2 weeks, and it is continuously intervened for 5 days per week, about 20 minutes per day.

[0042] Furthermore, the feature selection in S6 includes constructing a feature matrix of the fused data of fMRI data and EEG data and selecting a linear kernel as the kernel function, finding the similarity between the two modalities, learning the kernel weights of each modality, and obtaining the fused feature matrix.

[0043] Furthermore, the composite deep learning prediction model in S6 includes the following parts: a 3D CNN branch for processing fMRI data, an RNN branch for processing EEG data, and a feature fusion layer.

[0044] Furthermore, it includes the degree of nodes, betweenness centrality, global efficiency, etc. The variance threshold method is used to perform feature selection on these indicators. The specific steps are as follows:

[0045] Step 1: Represent the modal feature matrices from the two modalities, fMRI and EEG, as X1 and X2 respectively, and select the linear kernel as the kernel function for the data of both modalities;

[0046] Step 2: For each modality i, calculate the kernel matrix K using the selected kernel function i , and the element K ij represents the similarity between modality i and modality j;

[0047] Step 3: Learn the kernel weights for each modality through an optimization problem to determine the contribution of each modality to the final prediction. This process can be expressed by the following formula:

[0048]

[0049] where Y is the label, α is the kernel weight vector, λ is the regularization parameter, and K i is the kernel matrix of modality i.

[0050] Step 4: Weight and fuse the feature matrices of each modality according to the learned kernel weights to obtain the fused feature matrix X fusion , and the calculation formula for this matrix is:

[0051]

[0052] Step 5: Use the fused feature matrix X fusion for subsequent model training and prediction.

[0053] Build a composite deep learning model that includes 3D CNN and RNN branches, where 3D CNN is used to process fMRI data and RNN is used to process EEG data. Finally, use the joint branch to fuse the features of the two types of data. The matching Python IDE editor is Pycharm,

[0054] Furthermore, Step 5 specifically includes the following parts:

[0055] 3D CNN branch:

[0056] Step 1: Input layer: Input fMRI data, which is a three-dimensional tensor;

[0057] Step 2: Two CNN layers: Input the output signal of the input layer into two convolutional layers. The number of convolutional kernels is 8 and 16 respectively, with a shape of (3, 3), and use tanh as the activation function. Each CNN layer uses L2 regularization to prevent overfitting;

[0058] Step 3: Two max - pooling layers: The pooling window size is 2×2×2, and the stride is 2. Each max - pooling layer uses L2 regularization to prevent overfitting;

[0059] Step 4: Dropout layer: The dropout probability is 0.25;

[0060] Step 5: Two fully - connected layers: The two fully - connected layers use softmax as the activation function, and the number of nodes is 64 and 2 respectively. The first fully - connected layer performs batch normalization on the data; the second fully - connected layer outputs the probability that the patient belongs to the MDD group or the GAD group;

[0061] Step 6: Dropout layer: The dropout probability is 0.25;

[0062] Step 7: Use the cross - entropy function as the loss function, and adopt the following method (Lin et al., 2022): For the sampled data I i (x, y, z) and its label M is the number of samples, and the output result of the CNN is W, b are the model parameters, h W,b (·) represents the model function, and the loss calculation formula is:

[0063]

[0064] RNN branch:

[0065] Step 1: Input layer: Input EEG data;

[0066] Step 2: CNN layer: There are 9 CNN layers in total. The shape of the convolutional kernel is (3, 3), the stride is 1, and the ReLU function is used as the activation function. The middle convolutional layers perform zero padding of one pixel to restore the spatial resolution;

[0067] Step 3: Pooling layer: Add a pooling layer after the 4th, 8th, and 9th convolutional layers. The pooling type is max - pooling, the pooling window size is 2×2, the stride is 2, and the ReLU function is used as the activation function;

[0068] Step 4: Flatten layer, flatten the data into a one - dimensional tensor;

[0069] Step 5: Fully - connected layer: Set two fully - connected layers, with the number of neurons being 512 and 128 respectively, and use ReLU as the activation function;

[0070] Step 6: LSTM layer: The number of neurons is 128;

[0071] Step 7: Output layer: Output the binary classification prediction result of the model, and use the softmax function as the activation function;

[0072] Feature fusion layer:

[0073] It includes a convolutional layer and a pooling layer to extract the shared features of the two modalities.

[0074] Among them, the number of convolutional kernels in the convolutional layer is 32, the shape of the convolutional kernel is (3, 3), and ReLU is used as the activation function. Each convolutional kernel performs a convolution operation on the input data to generate a new channel, and the number of new channels is equal to the number of convolutional kernels. The output signal of the convolutional layer is input into the pooling layer, and the pooling type is selected as max pooling, and the shape of the convolutional kernel in the pooling layer is (3, 3);

[0075] Output layer:

[0076] Design two branches. Branch 1: Use a fully connected layer with the ReLU activation function to output the brain network index with the largest change after rTMS intervention. Branch 2: Use a fully connected layer with the softmax as the activation function and two nodes. These two nodes represent the effective and ineffective categories respectively, and output the prediction of whether the patient is effective after receiving rTMS intervention.

[0077] Compared with the prior art, the present invention has the following beneficial effects:

[0078] The present invention proposes an individualized rTMS intervention target determination scheme for precisely guiding key neural circuits, accurately locates the rTMS intervention target based on the dual-modal fusion of EEG and task-fMRI to calculate the brain effective connection network, ensures the high spatio-temporal resolution of the rTMS intervention targets for MDD and GAD neuromodulation, and thus obtains better curative effects of individualized rTMS intervention methods for depression and anxiety disorders.

[0079] Combined with the brain-computer interface technology, it can real-time monitor the neuroelectrophysiological activities during the rTMS intervention process and establish a corresponding visualization model, dynamically and real-time monitor the brain activities and timely adjust the coil position to ensure the accuracy of the intervention target, and further ensure the effect of rTMS intervention on improving brain function.

[0080] Based on the fMRI-EEG fused data, it can detect the neural mechanisms of brain network changes before and after precise intervention of individualized rTMS targets for MDD and GAD patients and establish a deep learning prediction model for curative effects, which can screen in advance the depressed and anxious patients suitable for precise target intervention with rTMS and achieve precise patient stratification for clinical treatment. Description of the Drawings

[0081] Figure 1 It is a flowchart of an rTMS precise intervention and prediction method for depression and anxiety disorders provided by the patent of this application;

[0082] Figure 2 Task flow chart of the cognitive reappraisal task provided for this patent;

[0083] Figure 3 64 sampling electrode diagram set according to the 10 - 20 international standard lead system provided for this patent;

[0084] Figure 4 Pre - experiment rTMS intervention efficacy diagram provided for this patent;

[0085] Figure 5 Prediction model architecture diagram provided for this patent. Detailed implementation manners

[0086] The following further explains the detailed implementation manners of the present invention with reference to the accompanying drawings. It should be noted here that the description of these implementation manners is used to help understand the present invention, but does not limit the present invention. In addition, the technical features involved in the various implementation manners of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0087] First, this patent provides a precise rTMS intervention method for depression and anxiety disorders, including:

[0088] As Figure 1 shown, the intervention method includes steps such as collecting EEG and fMRI data from MDD and GAD patients, using DCM analysis to determine individualized rTMS intervention targets, and synchronously collecting EEG data during rTMS intervention.

[0089] As Figure 2 shown, MDD and GAD patients perform a cognitive reappraisal task, that is, a pre - test. 64 negative pictures and 32 neutral pictures from the International Affective Picture System are presented to the patients, and the patients are required to perform processing tasks of "viewing" or "cognitive reappraisal" on the pictures. The task is divided into 5 blocks. At the end of each block, the patients are required to subjectively rate the negative emotion experience on a scale of 1 - 5 (1 point = not at all, 5 points = extremely).

[0090] The severity of depression and anxiety symptoms is measured using IDS - 30 and HAMA respectively.

[0091] As Figure 3 shown, 64 sampling electrodes are set using the 10 - 20 international standard lead system to obtain EEG data of the cognitive reappraisal task, and the sampling frequency is 1000Hz. fMRI data of the cognitive reappraisal task is synchronously collected, and the parameters are as follows: [TR] = 2000ms, [TE] = 30ms, and the voxel size is 3×3×3mm 3, The matrix shape is 90×90, the number of layers is 60, a total of 180 time points are collected, and the scanning duration is 6 min.

[0092] EEGLAB was used to preprocess the pretest EEG data, including: electrode localization, deleting useless electrodes, band-pass filtering from 0.5 - 30 Hz, downsampling to a rate of 250 Hz, segmenting the data and performing baseline correction, excluding bad segments, performing ICA, removing artifacts such as electrooculogram, electrocardiogram, and channel noise.

[0093] SPM12 was used to preprocess the pretest fMRI data, specifically including:

[0094] Step1: MRIQC 23.2.0 was used to check the data and exclude motion artifacts, fluctuations, etc.

[0095] Step2: SPM 12 was used for time slice correction.

[0096] Step3: SPM 12 was used for head motion correction. Rigid motion correction was adopted to correct the motion of images at all time points. For the image I(t) at time point t, I ref is the reference image, the translation parameters are (Δx, Δy, Δz), the rotation parameters are (θx, θy, θz), and the corrected image I′(t) can be obtained by the following formula:

[0097]

[0098] Step4: The fMRI data was mapped onto the MNI standard brain.

[0099] Step4: Spatial smoothing was used to improve the signal-to-noise ratio, and the size of the smoothing kernel was 6 - 8 mm; a 0.01 Hz high-pass filter was used to remove trends and low-frequency drifts in the data.

[0100] Step5: The 6 rigid parameters of motion correction were used to regress the fMRI data to remove motion-related signals. This process can be expressed by the following formula:

[0101] Y(t) = X(t)β + ∈(t)

[0102] where: Y(t) is the fMRI data at time point t. X(t) is the design matrix containing motion parameters. β is the coefficient of motion parameters. ∈(t) is the residual.

[0103] Step6: Cerebrospinal fluid and white matter signals were removed.

[0104] SPM 12 was used to register the fMRI data and EEG data onto the MNI standard brain to ensure that the fMRI and EEG data are in the same coordinate system.

[0105] The fMRI and EEG data were mapped to the same brain regions and spatial coordinate system using the AAL atlas ROI_MNI_v4 version, and the fMRI and EEG data were combined.

[0106] Define a generalized linear model, and use binary-encoded emotional picture stimuli (0: neutral, 1: negative) as independent variables (i.e., psychological variables) to establish a design matrix. Taking the physiological variables of the amygdala during the execution of the cognitive reappraisal task as the dependent variable, estimate the degree to which the independent variable affects the dependent variable by coefficients. The least squares estimation method is used, and the calculation formula is as follows:

[0107] Y = Xβ + ∈

[0108] By adjusting the value of the coefficient vector β, the linear combination of the dependent variable Y and the design matrix X is best fitted, that is, the error term ∈ is minimized.

[0109] Select the left DLPFC, right DLPFC, left amygdala, and right amygdala as the VOIs. The four brain regions are located at the MNI coordinates (-36, 27, 29), (36, 27, 29), (-24, -6, -16), and (22, -2, -15) respectively.

[0110] Use SPM12 to construct a full DCM model and perform DCM analysis, including the following steps:

[0111] Step1: Select the SPM.mat files generated by constructing the GLM model for all patients.

[0112] Step2: Select the VOI time series of the left DLPFC, right DLPFC, left amygdala, and right amygdala extracted from all patients.

[0113] Step3: Specify the A matrix. This patent presupposes that each VOI brain region and the connections between brain regions have their own balance mechanisms.

[0114] Step4: Specify the B matrix. This patent presupposes that the connections between all VOI brain regions are modulated by the experimental tasks.

[0115] Step5: Specify the C matrix. This patent presupposes that the balance mechanisms of all VOI brain regions themselves are modulated by the experimental tasks.

[0116] Step6: Perform model estimation, and select the brain region with the strongest connection to the amygdala as the rTMS intervention target.

[0117] The individualized rTMS intervention lasted for 2 weeks, with a stimulation intensity of 100% of the resting motor threshold (rMT), a stimulation frequency of 10 Hz, a stimulation duration of 1 s, and an interval of 10 s. The intervention continued for 2 weeks, with 5 consecutive days of intervention per week and approximately 20 minutes per day.

[0118] During the rTMS intervention, an online EEG analysis algorithm was combined to perform online denoising on the collected EEG data, including time-domain filtering, CSP, online PCA, and online ICA.

[0119] For time-domain filtering, the high-pass filter was set to 0.1 Hz and the low-pass filter was set to 30 Hz.

[0120] When using CSP filtering, different covariance matrices were created for different categories of EEG signals. The matrices are as follows:

[0121]

[0122] Among them, represents the variance of this factor and is calculated using the following formula:

[0123]

[0124] Among them, X i is the EEG signal of different categories, is the mean vector of the data.

[0125] Cov represents the covariance of two factors and is calculated using the following formula:

[0126]

[0127] Among them, is the EEG signal of different categories, is the mean vector of the data.

[0128] Generalized eigenvalue decomposition was performed on these covariance matrices. The basic idea is as follows: For a group of matrices A1, A2... A k and B, where A i is an n×n-dimensional matrix and B is an n×n-dimensional positive definite matrix. The form of the generalized eigenvalue problem is:

[0129] A i x = λBx)

[0130] Among them, x is a non-zero eigenvector and λ is the generalized eigenvalue. The generalized eigenvalue decomposition in this general form can be expressed by the following formula:

[0131] Ax = ΛBx

[0132] where, A = [A1, A2... A k is an n×(kn)-dimensional matrix, x is an n×(kn)-dimensional matrix, whose columns are combinations of eigenvectors x. Λ is a diagonal matrix, and the elements on its diagonal are the generalized eigenvalues λ. The final generalized eigenvalue decomposition will generate a set of eigenvalues and corresponding eigenvectors. Select the eigenvectors with the largest or smallest eigenvalues to construct filters. Select the eigenvectors with the largest eigenvalues to enhance the brain activities of interest, and select the eigenvectors with the smallest eigenvalues to suppress other brain activities. Finally, a set of CSP filters is generated, each filter corresponding to a different brain activity pattern, some of which are used for enhancement and some for suppression.

[0133] Online PCA analysis is performed on the synchronously collected EEG data, and the basic idea is as follows: When collecting EEG signals, there are n samplings, and each sampling has m electrodes (in this patent, the number of electrodes is 64). The EEG data is represented as an n×m matrix X, and the calculation formula for the covariance matrix C is as follows:

[0134]

[0135] where, is the mean vector of the data. The eigenvalue decomposition of the covariance matrix is performed using the following formula:

[0136] C = PDP -1

[0137] where, P is the eigenvector matrix, D is the diagonal matrix, and the elements on the diagonal are the eigenvalues. Select the eigenvectors corresponding to the first k eigenvalues to form the projection matrix W, and project the data into the principal component space:

[0138] Y = XW

[0139] where, Y is the data matrix after dimensionality reduction.

[0140] The basic steps of online ICA are shown as follows: First, construct an EEG data matrix X containing m electrodes, with a size of n×m, where n is the number of sampling points. ICA is represented by the following formula in mathematics:

[0141] X = AS

[0142] where, X is the observed signal matrix, A is the unknown mixing matrix, S is the source signal matrix, and the purpose of ICA is to find the source signal matrix S, which specifically includes the following steps:

[0143] Step1: For the data matrix X, first perform centering using the following formula:

[0144]

[0145] where, is the mean value of each column feature of X in the original data matrix.

[0146] Step2: Whiten the data using the following formula:

[0147]

[0148] Perform eigenvalue decomposition on the covariance matrix C using the following formula:

[0149] C = EDE T

[0150] where E is the eigenvector matrix and D is the diagonal eigenvalue matrix. The whitening transformation can be expressed as the following formula:

[0151] X white = D -1 / 2 E T X centered

[0152] Step3: Based on maximizing non-Gaussianity, select negentropy as the measure of non-Gaussianity, use the non-linear function g(·) to approximate the maximization of negentropy, and find the projection weights that maximize negentropy. Its iterative update rule can be expressed by the following formula:

[0153]

[0154] After that, normalize the new weight vector w + using the following formula:

[0155]

[0156] Step4: Orthogonalization and convergence, using the Gram-Schmidt process. For the i-th component, the update rule will consider the previous i - 1 components:

[0157]

[0158] After normalizing w i check for convergence.

[0159] Step5: Independent component analysis, find all the weight vectors:

[0160] W = [w1, w2,..., w n

[0161] Estimate the independent components in the following way:

[0162] S = W T X white

[0163] ​Using the source localization analysis model, perform weighted minimum norm estimation inversion to solve the source analysis for EEG data, including the following steps:

[0164] Step1: Establish a head model, including the following steps: Check the quality of the 3DT1-weighted sequence images of the MRI to ensure that the images have no obvious motion artifacts, artifacts, or other artifacts; Register the MRI images with the spatial information of the electrode positions; Segment the MRI images using FreeSurfer; Use the three-dimensional reconstruction software BrainSuite to convert the segmentation results of the MRI images into a three-dimensional surface model.

[0165] Step2: Use the finite element method or the analytical method to calculate the transfer matrix of the electrical signal on the cerebral cortex. The formula is as follows:

[0166]

[0167] σ is the electrical conductivity of the tissue, r i and r j are the spatial positions of source i and electrode j respectively, Ω represents the surface of the cerebral cortex, and n(r) is the surface normal vector.

[0168] Step3: Perform weighted minimum norm estimation using the following formula:

[0169]

[0170] where, is the estimated source signal, L is the transfer matrix, d is the observed EEG data, W is the weighting matrix, and λ is the regularization parameter.

[0171] Step4: Use the transfer matrix to convert the estimated source signal into the source activity distribution on the cerebral cortex. The formula is as follows:

[0172]

[0173] where is the source signal, is the estimated value of the electrode signal, (L T L) -1 L T is the pseudo-inverse matrix used to convert the electrode signal back to the source space.

[0174] Step5: Project the source activity onto the cerebral cortex surface of the head model, use Brainstorm to visualize the source activity on the cerebral cortex, overlay the pre-determined rTMS target positions on the cerebral cortex, and compare them with the activated brain regions.

[0175] Manually adjust the coil to ensure that the activated brain region and the intervention target are consistent.

[0176] When evaluating the efficacy of rTMS intervention in patients, the reduction rate of the total scores of the IDS-30 and HAMA scales before and after the intervention is used to determine whether the rTMS intervention is effective. A reduction rate of ≥50% is considered effective, and <50% is considered ineffective (Corlier et al., 2020; Diefenbach et al., 2016). The IDS-30 is mainly used to evaluate the intervention efficacy in MDD patients, the HAMA is mainly used to evaluate the intervention efficacy in GAD patients, and both the IDS-30 and HAMA are referred to for patients with comorbidities; the cognitive reappraisal task is used to judge the improvement of patients' functions before and after rTMS intervention, and fMRI and EEG data are collected synchronously, that is, the post-test. The efficacy of rTMS intervention is as Figure 4 shown.

[0177] As Figure 1 shown, the prediction method includes establishing an efficacy prediction step according to the improvement of symptoms.

[0178] The collected post-test EEG data of the cognitive reappraisal task is preprocessed using EEGLAB, and the post-test fMRI data is preprocessed using SPM12. The preprocessing steps are the same as those of the pre-test.

[0179] According to whether the reduction rate of IDS-30 and HAMA before and after rTMS intervention in patients is greater than or equal to 50%, the patients are divided into a treatment-effective group and a treatment-ineffective group.

[0180] Extract the imaging brain network indexes before and after the intervention, including the degree of nodes, betweenness centrality, global efficiency, etc. The variance threshold method is used to perform feature selection on these indexes. The specific steps are as follows:

[0181] Step1: The modal feature matrices from the two modalities of fMRI and EEG are respectively denoted as X1 and X2, and a linear kernel is selected as the kernel function for the data of the two modalities.

[0182] Step2: For each modality i, the kernel matrix K is calculated using the selected kernel function i , and the element K ij of the kernel matrix represents the similarity between modality i and modality j.

[0183] Step3: By optimizing the problem, learn the kernel weights of each modality to determine the contribution of each modality to the final prediction. This process can be expressed by the following formula:

[0184]

[0185] where Y is the label, α is the kernel weight vector, λ is the regularization parameter, and K i is the kernel matrix of modality i.

[0186] Step 4: Weight and fuse the feature matrices of each modality according to the learned kernel weights to obtain the fused feature matrix X fusion , and the calculation formula of this matrix is:

[0187]

[0188] Step 5: Use the fused feature matrix X fusion for subsequent model training and prediction.

[0189] As Figure 5 shown, a composite deep learning model containing 3D CNN and RNN branches is established, where 3D CNN is used to process fMRI data, and RNN is used to process EEG data. Finally, the joint branch is used to fuse the features of the two types of data. The matching Python IDE editor is Pycharm, and the specific steps are as follows:

[0190] 3D CNN branch:

[0191] Step 1: Input layer: Input fMRI data, which is a three-dimensional tensor.

[0192] Step 2: Two CNN layers: Input the output signal of the input layer into two convolutional layers. The numbers of convolutional kernels are 8 and 16 respectively, and the shape is (3, 3). Tanh is used as the activation function. Each CNN layer uses L2 regularization to prevent overfitting.

[0193] Step 3: Two max pooling layers: The pooling window size is 2×2×2, and the stride is 2. Each max pooling layer uses L2 regularization to prevent overfitting.

[0194] Step 4: Dropout layer: The dropout probability is 0.25.

[0195] Step 5: Two fully connected layers: The two fully connected layers use softmax as the activation function, and the numbers of nodes are 64 and 2 respectively. The first fully connected layer performs batch normalization on the data. The second fully connected layer outputs the probabilities that the patient belongs to the MDD group or the GAD group.

[0196] Step 6: Dropout layer: The dropout probability is 0.25.

[0197] Step 7: Use the cross-entropy function as the loss function and adopt the following method (Lin et al., 2022): For the sampled data I i (x, y, z) and its label M is the number of samples, and the CNN output result is W, b are the model parameters, h W,b(·) represents the model function, and the loss calculation formula is as follows:

[0198]

[0199] RNN branch:

[0200] Step1: Input layer: Input EEG data.

[0201] Step2: CNN layer: There are 9 CNN layers in total. The convolution kernel shape is (3,3), the stride is 1, and the ReLU function is used as the activation function. Zero-padding of one pixel is performed in the middle convolutional layer to restore the spatial resolution.

[0202] Step3: Pooling layer: A pooling layer is added after the 4th, 8th, and 9th convolutional layers. The pooling type is selected as max pooling, the pooling window size is 2×2, the stride is 2, and the ReLU function is used as the activation function.

[0203] Step4: Flatten layer, flatten the data into a one-dimensional tensor.

[0204] Step5: Fully connected layer: Two fully connected layers are set, with the number of neurons being 512 and 128 respectively, and ReLU is used as the activation function.

[0205] Step6: LSTM layer: The number of neurons is 128.

[0206] Step7: Output layer: Output the binary classification prediction result of the model, and use the softmax function as the activation function.

[0207] Feature fusion layer:

[0208] It includes a convolutional layer and a pooling layer to extract the shared features of the two modalities. Among them, the number of convolution kernels in the convolutional layer is 32, the convolution kernel shape is (3,3), and ReLU is used as the activation function. Each convolution kernel performs a convolution operation on the input data to generate a new channel, and the number of new channels is equal to the number of convolution kernels. The output signal of the convolutional layer is input into the pooling layer, and the pooling type is selected as max pooling, and the convolution kernel shape of the pooling layer is (3,3).

[0209] Output layer:

[0210] Two branches are designed. Branch one: Use a fully connected layer with the ReLU activation function to output the brain network metrics with the largest changes after rTMS intervention. Branch two: Use a fully connected layer with two nodes with the softmax as the activation function. These two nodes represent the effective and ineffective categories respectively, and output the prediction of whether the patient is effective after receiving rTMS intervention.

[0211] In this patent, the present application also provides a computer device, which includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the rTMS precise intervention and prediction method for depression and anxiety disorders described in any one of the above.

[0212] In this patent, the present application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor is caused to execute the rTMS precise intervention and prediction method for depression and anxiety disorders described in any one of the above.

[0213] The embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, without departing from the principle and spirit of the present invention, various changes, modifications, substitutions, and variations of these embodiments still fall within the protection scope of the present invention.

[0214] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be regarded as limiting the claimed rights.

[0215] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only includes an independent technical solution. This narrative manner of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A precise rTMS intervention method for depression and anxiety disorders, characterized in that, It includes the following steps: S1: Using a cognitive reappraisal task, the pictures presented to the patients are from the International Affective Picture System, including 64 negative pictures and 32 neutral pictures. The patients are required to perform a processing task of "watching" or "cognitive reappraisal" on the pictures, and the IDS-30 and HAMA are respectively used to measure the severity of depressive and anxiety symptoms; S2: During the process of inducing active emotion regulation in the patients by the cognitive reappraisal task, fMRI and EEG data are synchronously collected; SPM12 is used to preprocess the fMRI data, and EEGLAB is used to preprocess the EEG; S3: Using SPM 12, the fMRI data and EEG data are uniformly registered to the MNI standard brain to ensure that the fMRI and EEG data are in the same coordinate system. The ROI_MNI_v4 version of the AAL atlas is used to map the fMRI and EEG data to the same brain regions and spatial coordinate system, and the fMRI and EEG data are merged to obtain the fused fMRI-EEG data; S4: Define a generalized linear model, and select the left DLPFC, right DLPFC, left amygdala, and right amygdala as VOIs to construct a full DCM model for DCM analysis to determine the rTMS intervention target; S5: Perform individualized rTMS intervention on the patients for 2 weeks; When evaluating the efficacy of the patients after rTMS intervention, it is determined whether the rTMS intervention is effective according to the reduction rate of the total scores of the IDS-30 and HAMA scales before and after the intervention. A reduction rate ≥ 50% is considered effective, and < 50% is considered ineffective. When evaluating the efficacy of MDD patients, the IDS-30 is mainly referred to, and when evaluating the efficacy of GAD patients, the HAMA is mainly referred to. For comorbid patients, both the IDS-30 and HAMA are referred to simultaneously; The cognitive reappraisal task is used to judge the improvement of the patients' functions before and after rTMS intervention, and fMRI and EEG data are synchronously collected; Use SPM 12 to preprocess the post-test fMRI data, and use EEGLAB to preprocess the post-test EEG; S6: Extract the imaging brain network indexes before and after the intervention, including the degree of nodes, betweenness centrality, and global efficiency, and perform feature selection to establish a prediction model, and output the brain network index with the largest change after rTMS intervention and the prediction of whether the patient is effective after receiving rTMS intervention; Among them, a composite deep learning model including 3D CNN and RNN branches is included, where 3D CNN is used to process fMRI data, and RNN is used to process EEG data. Finally, the joint branch is used to fuse the features of the two types of data, and the matching Python IDE editor is Pycharm.

2. The precise rTMS intervention method for depression and anxiety disorders according to claim 1, wherein: When collecting fMRI data and EEG data, 64 sampling electrodes were set using the 10-20 international standard lead system to obtain EEG data for the cognitive reappraisal task; the fMRI parameters were as follows: [TR]=2000ms, [TE]=30ms, voxel size was 3×3×3mm 3 , the matrix shape was 90×90, the number of layers was 60, a total of 180 time points were collected, and the scanning duration was 6 minutes.

3. The rTMS precise intervention method for depression and anxiety disorders according to claim 1, characterized in that: The preprocessing of the EEG data includes electrode positioning, removing useless electrodes, band-pass filtering, downsampling the rate, segmenting the data and performing baseline correction, removing bad segments, performing ICA, and artifact removal steps. The preprocessing of the fMRI data includes using MRIQC 23.2.0 to exclude motion artifacts and fluctuations; performing temporal slice correction using SPM 12; performing head motion correction using SPM 12, adopting rigid motion correction, and performing motion correction on the images at all time points; mapping the fMRI data onto the MNI standard brain; using spatial smoothing to improve the signal-to-noise ratio; using a 0.01 Hz high-pass filter to remove trends and low-frequency drifts in the data; using the 6 rigid parameters of motion correction to perform regression on the fMRI data to remove motion-related signals; and removing cerebrospinal fluid and white matter signals.

4. A precise rTMS intervention method for depression and anxiety disorders according to claim 1, characterized in that: The DCM analysis includes selecting the VOI time series of the left DLPFC, right DLPFC, left amygdala, and right amygdala extracted from all patients, specifying the A matrix, B matrix, and C matrix, performing model estimation, and selecting the brain region with the strongest connection to the amygdala as the rTMS intervention target.

5. The precise intervention method for rTMS of depression and anxiety disorders according to claim 1, characterized in that: When performing individualized rTMS intervention, an online EEG parsing algorithm is combined to make up for the deficiency of the fMRI time resolution. The online EEG parsing algorithm includes time-domain filtering, CSP, online PCA, and online ICA, and a source localization analysis model is used to perform weighted minimum norm estimation inversion and solve the source analysis for the EEG data. Through the online EEG parsing algorithm, it is monitored in real time and ensured that the target brain region activated during the rTMS stimulation is consistent with the brain region at the point with the strongest atlas connection determined by the fMRI, so as to ensure the high spatio-temporal resolution of the stimulation target area. The rTMS stimulation intensity is 100% of the resting motor threshold (rMT), the stimulation frequency is 10 Hz, the stimulation duration is 1 s, the interval is 10 s, the intervention lasts for 2 weeks, and it is continuously intervened for 5 days per week, 20 minutes per day.

6. The precise rTMS intervention method for depression and anxiety disorders according to claim 1, wherein: The feature selection in S6 includes constructing a feature matrix of the fused data of the fMRI data and the EEG data and selecting a linear kernel as the kernel function, finding the similarity between the two modalities, learning the kernel weights of each modality, and obtaining the fused feature matrix.

7. The rTMS precise intervention method for depression and anxiety disorders according to claim 1, characterized in that: The composite deep learning prediction model in S6 includes the following parts: a 3D CNN branch for processing the fMRI data, an RNN branch for processing the EEG data, and a feature fusion layer is constructed.

8. A precise rTMS intervention method for depression and anxiety disorders according to claim 1, characterized in that: It includes the degree of nodes, betweenness centrality, and global efficiency. The variance threshold method is used to perform feature selection on these metrics. The specific steps are as follows: Step1: Represent the modal feature matrices from the two modalities of fMRI and EEG as , and select the linear kernel as the kernel function for the data of the two modalities; Step 2: For each modality \(i\), calculate the kernel matrix using the selected kernel function , and the elements of the kernel matrix represent the similarity between modality \(i\) and modality ; Step3: Through the optimization problem, learn the kernel weights of each modality and determine the contribution of each modality to the final prediction. This process is represented by the following formula: Among them, is a label, is the kernel weight vector, is the regularization parameter, is the kernel matrix of modality i; Step4: Weight and fuse the feature matrices of each modality according to the learned kernel weights to obtain the fused feature matrix , and the calculation formula of this matrix is: Step5: Use the fused feature matrix for subsequent model training and prediction; A composite deep learning model including 3D CNN and RNN branches is established, where 3D CNN is used to process the fMRI data and RNN is used to process the EEG data. Finally, the features of the two types of data are fused using a joint branch. The matching Python IDE editor is Pycharm.

9. A precise intervention method for rTMS in depression and anxiety disorders according to claim 8, characterized in that: Step5 specifically includes the following parts: 3D CNN Branch: Step1: Input layer: Input fMRI data, which is a three-dimensional tensor; Step2: Two CNN layers: Input the output signal of the input layer into two convolutional layers. The number of convolutional kernels is 8 and 16 respectively, with a shape of (3, 3), and use tanh as the activation function; Each CNN layer uses L2 regularization to prevent overfitting; Step3: Two max pooling layers: The pooling window size is 2×2×2, and the stride is 2. Each max pooling layer uses L2 regularization to prevent overfitting; Step4: Dropout layer: The dropout probability is 0.25; Step5: Two fully connected layers: The two fully connected layers use softmax as the activation function, and the number of nodes is 64 and 2 respectively; The first fully connected layer performs batch normalization on the data; The second fully connected layer outputs the possibility that the patient belongs to the MDD group or the GAD group; Step6: Dropout layer: The dropout probability is 0.25; Step 7: Use the cross-entropy function as the loss function, and adopt the following method: For the sampled data and its label , is the number of samples, the CNN output result is , is the model parameter, represents the model function, and the loss calculation formula is: ； RNN Branch: Step1: Input layer: Input EEG data; Step2: CNN layers: There are 9 CNN layers in total. The shape of the convolutional kernel is (3, 3), the stride is 1, and ReLU function is used as the activation function. Zero-padding of one pixel is performed on the intermediate convolutional layers to restore the spatial resolution; Step3: Pooling layer: Add a pooling layer after the 4th, 8th, and 9th convolutional layers. The pooling type is max pooling, the pooling window size is 2×2, the stride is 2, and ReLU function is used as the activation function; Step4: Flatten layer, flatten the data into a one-dimensional tensor; Step5: Fully connected layer: Set two fully connected layers, with the number of neurons being 512 and 128 respectively, and use ReLU as the activation function; Step6: LSTM layer: The number of neurons is 128; Step7: Output layer: Output the binary classification prediction result of the model, and use the softmax function as the activation function; Feature Fusion Layer: It includes a convolutional layer and a pooling layer to extract the shared features of the two modalities, where the convolutional layer has 32 convolutional kernels, the shape of the convolutional kernel is (3, 3), and ReLU is used as the activation function. Each convolutional kernel performs a convolution operation on the input data to generate a new channel, and the number of new channels is equal to the number of convolutional kernels. Input the output signal of the convolutional layer into the pooling layer. The pooling type is max pooling, and the shape of the convolutional kernel of the pooling layer is (3, 3); Output Layer: Design two branches. Branch 1: Use a fully connected layer with the ReLU activation function to output the brain network metrics that change the most after rTMS intervention. Branch 2: Use a fully connected layer with two nodes and the softmax as the activation function. These two nodes represent the effective and ineffective categories respectively, and output the prediction of whether the patient is effective after receiving rTMS intervention.

Images