Brain disease diagnosis model based on multilevel causal intervention network

Abstract

The invention discloses a brain disease diagnosis model based on a multilevel causal intervention network, and relates to the field of brain function network analysis and medical auxiliary research. The method specifically comprises the following steps: a brain network pre-construction stage: performing data preprocessing, extracting an fMRI time sequence, and constructing input features required by a model by using tangent Pearson embedding; feature extraction of a brain network is realized by using a three-layer Transform network of a self-attention mechanism; fusing multiple layers of attention by using an HCI module and correspondingly dividing brain network features into causal and non-causal parts; applying DCI to apply a first-order causal constraint to the model to reduce interference of non-causal factors, introducing ICAC into a causal part, and applying a second-order causal constraint to the model to further optimize the model; and realizing classification of brain diseases by using an orthogonal clustering (OCR) readout function. According to the method, non-causal interference factors can be effectively filtered, and high-order function connection information is extracted, so that the accuracy and robustness of brain disease analysis are improved.

Description

Technical Field

[0001] The present invention relates to the field of brain function network analysis and medical assistance research, and in particular to a brain disease diagnosis model based on a multi-level causal intervention network with hierarchical causal perception capability. Background Art

[0002] In recent years, functional connectivity networks (FCNs) based on resting-state cerebral functional magnetic resonance imaging (rs-fMRI) have played a vital role in the detection and analysis of neurological diseases. It measures the dependence of functional activities between any two brain regions. Abnormal changes in functional connectivity are used as a basis for diagnosing neurological diseases, which helps to understand the basic mechanisms of neurological diseases and improve diagnostic and treatment strategies. Therefore, functional connectivity derived from rs-fMRI data has become a powerful method to quantify and map brain function. On this basis, the introduction of advanced methods in machine learning and computational neuroscience has opened up new avenues for analyzing complex neuroimaging data and developing predictive models for ASD detection.

[0003] The analysis shows that functional connectivity networks can be used as effective biomarkers to identify patients with brain diseases. However, traditional methods usually only capture direct pairwise relationships between brain regions, thereby ignoring a large amount of high-order information. In addition, fMRI data contains signals directly related to the subject's physiological information, but is also interfered by various external factors that are unrelated to the subject's physiological state, resulting in a large number of non-causal factors mixed in, thereby introducing noise and potentially misleading research results. In addition, these non-causal factors are often heterogeneous between different sites, which is the main cause of domain shift in multi-site datasets.

[0004] In order to solve these problems, a brain disease diagnosis method model based on a multi-level causal intervention network is proposed. This method model integrates high-order FC and brain converters to continuously capture more comprehensive high-order features in the data. At the same time, by using a hierarchical causal-aware integrator as an observation lens, a series of causal interventions are implemented to guide the model to learn truly invariant causal features, with excellent performance and excellent inter-site versatility; a large number of experiments on real data sets of autism spectrum disorder ASD and major depressive disorder MDD show that this method is superior to the current state-of-the-art methods and demonstrates its important clinical value. Summary of the invention

[0005] The purpose of the present invention is to propose a brain disease diagnosis model based on a multi-level causal intervention network, which is used to solve the problem that the existing technology can only capture the direct pairwise relationship between brain regions in the functional connection network analysis, ignores the high-order information, and leads to insufficient accuracy in the identification of brain diseases. At the same time, the fMRI data contains external factors unrelated to the physiological state of the subject, mixed with non-causal interference signals, generates noise and may mislead the research results.

[0006] To achieve the above objectives, the present invention proposes a brain disease diagnosis model based on a multi-level causal intervention network, the steps of which are as follows:

[0007] S1, the brain network model pre-construction stage, firstly, data pre-processing and fMRI time series extraction are performed, and the input features required for the model are constructed using Tangent Pearson Embedding (TPE);

[0008] S2, feature extraction stage, using a 3-layer Transformer network with self-attention mechanism to extract features of the brain network;

[0009] S3, causal brain connection division stage, using the HCI module to fuse multiple layers of attention and divide brain network features into causal and non-causal parts accordingly;

[0010] S4, causal constraint stage, DCI is used to impose first-order causal constraints on the model, and ICAC is used to impose second-order causal constraints on the model;

[0011] S5, classification and recognition stage, using orthogonal clustering OCR readout function to realize the classification of brain diseases.

[0012] Preferably, in S1, the preprocessed data is divided into 200 brain regions of interest using a CC200 template to construct a brain network matrix.

[0013] Preferably, the connection patterns between different brain regions are different. Tangent Pearson Embedding (TPE) is used to extract the characteristics of the brain network matrix. Suppose the fMRI time series input of the subject is Where t represents the time point, the calculation formula of the tangent Pearson embedding TPE is as follows:

[0014]

[0015] Among them, x i , x j , x k is an fMRI time series element, i, j, k are integer indexes, r represents the number of brain region ROIs, Cov(x i ,x k ) is x i and x kThe covariance between them, σ represents the standard deviation, P ij represents the covariance between the i-th row and j-th column of the Pearson correlation matrix, M is the number of subjects, log m is the matrix logarithm, P is the Pearson correlation matrix, P i is the element of the i-th row of the Pearson correlation matrix, and TPE is the tangent Pearson embedding matrix.

[0016] Preferably, in S2, a Transformer network is constructed for feature extraction. The data fit must be re-evaluated and the attention mechanism adjusted to correspond to the unique features of the brain network data, using the connection profile tp i As the initial node features to establish the core network, where tp i is the corresponding row of node i in the tangent Pearson embedding matrix TPE.

[0017] Preferably, the self-attention mechanism is an important module of Transformer. The traditional self-attention module is defined as:

[0018]

[0019] Among them, <.,.> represents the inner product, and the input Z of the first layer 0 This is the TPE mentioned above, l represents the number of layers, Z l and Z l-1 Represent the network characteristics of layer l and layer l-1 respectively, is a learnable parameter, represent The size of the first dimension;

[0020] Use a multi-head attention mechanism with H heads to work simultaneously, and combine their outputs through connections to form the overall output of the multi-head attention module, where H is an integer;

[0021] Z l-1 The i-th and j-th rows of and attention For projection and The linear kernel between RBF and RBF nonlinear kernel The linear kernel is improved, and the formula is as follows:

[0022]

[0023] Preferably, in S3, in the causal brain connection division stage, the necessity of attention fusion lies in that the attention matrices at different levels capture multi-scale features from local to global respectively, and the global attention representation is obtained by the HCI method, and the formula is as follows:

[0024]

[0025] A f =α1A1+α2A2+α3A3;

[0026] Among them, ⊙ represents the element product, α1, α2, and α3 are the fusion weights of A1, A2, and A3 respectively. is the attention matrix of the first layer of the i-th head, is the attention matrix of the second layer of the i-th head, is the attention matrix of the third layer of the i-th head, A f is the fused attention matrix.

[0027] Preferably, in order to separate the causal and non-causal parts, keep A f The first γ percent of the largest value in the γ-th percentile is set to zero, which is expressed as The remaining areas are non-causal focal points, represented by γ represents the proportion of causal features. At this time, Focusing only on non-causal regions, Focus on the causal area;

[0028] Execute the forward propagation process of Transformer, respectively and Substitute the attention matrix part in the forward propagation to obtain the candidate causal part Z C and the non-causal part Z N .

[0029] Preferably, in S4, DCI is applied to impose a first-order causal constraint on the model, and feature Z is input into the readout function to obtain the output feature make and Represent the candidate causal prediction of the mth sample and the non-causal prediction of the nth sample respectively. Use different non-causal predictions as interventions and combine them with the causal prediction to form a composite prediction. The formula is as follows:

[0030]

[0031] Where sigm represents the Sigmoid function, For composite forecast;

[0032] Then, by intervening in the loss function The first-order causal constraints are imposed on the model, and the formula is as follows:

[0033]

[0034] in, is the cross entropy loss function, E n [.] indicates the expectation of n, Var n [.] means taking the variance of n, y m is the category label of sample m;

[0035] DCI can effectively alleviate the spurious correlations caused by sampling bias. In the case that the non-causal subgraph does not contain spurious correlations, the combination is unbiased, thereby increasing the possibility of learning pure causal correlations.

[0036] Preferably, in S4, ICAC is used to impose second-order causal constraints on the model, and Z C Expand it into a column vector. Suppose the column vector has K elements, each element represents a neuron, and the kth neuron is z k , neuron z k Z C The causal attribution of is defined as the value z', and the average causal effect of z' is The calculation formula is as follows:

[0037] E[Z c |do(z -k = z′)] = ∫ Z C ·p(Z c |do(z -k =z'))dZ c

[0038] ξ=∫(∫Z C ·p(Z C |do(z=z′))dZ C )dz′

[0039]

[0040] Among them, z represents the entire column vector, do(.) is the intervention method, p(.) represents the process of deep network calculation probability, z -k represents z divided by z k For all values ​​other than , ξ is the baseline value and E[.|.] is the conditional expectation;

[0041] The intervention variable z′ satisfies a uniform distribution within the range of z, that is, for the i-th element z of z′ ′i , ξ is the mean distribution of z, representing its baseline value; the do(.) intervention method only retains z each time k The value of z k The distribution of all values ​​other than k , repeat the above method to quantify the true causal value of each variable to the model;

[0042] The causal influence of all features on the i-th sample label is quantified by the invariant causal attribution ICA vector, which is the feature neuron z k The causal attribution of provides a metric, and the ICA vector of the k-th feature pair label y is defined as follows:

[0043]

[0044] By analyzing all K dimensions, the ICA vector f of the i-th sample is generated i , the formula is as follows:

[0045] f i =[(f i 1 ) T ,(f i 2 ) T ,…,(f i k ) T ,…,(f i K ) T ] T ;

[0046] In order to evaluate the difference between the ICA vectors of the input features on the output y across domains, a contrast loss function is proposed In order to match the input ICA vector with the output of all instances of the same class across domains, the loss is optimized by minimizing the distance between inputs of the same class and maximizing the distance between inputs of different classes. The inputs of the same class are regarded as positive matches of the i-th sample, and the inputs of different classes are regarded as negative samples. The contrast loss formula is as follows:

[0047]

[0048] Among them, f i 、f j and f k are the ICA vectors of samples i, j, and k, respectively, y is the label, and τ is the temperature parameter for scaling cosine similarity.

[0049] This loss function encourages the embeddings of samples with the same label to be close, while pushing the embeddings of samples with different labels apart; by carefully designing this contrastive loss and combining it with the main loss function, we ensure that the model learns to effectively maintain inter-class similarities, thereby enhancing its ability to capture potential causal relationships.

[0050] Preferably, in S5, after intervention and treatment, Z CFor the causal features most relevant to the disease, the orthogonal clustering readout function OCR is used to extract these features for the final classification;

[0051] Set O cluster centers, each with V dimensions, and use the Softmax projection operator to calculate the probability P of assigning node i to cluster O io , the formula is as follows:

[0052]

[0053] in, Represents Z C The i-th row, u o is an orthogonal basis obtained by uniformly initializing O cluster centers using Xavier and using the Gram-Schmidt method;

[0054] The final embedding Z G is obtained through aggregate learning, composed of the causal feature Z C The calculated distribution ρ is calculated as follows:

[0055] Z G =ρ T Z C ;

[0056] Z G Bring it into MLP to obtain the final classification result and use the cross entropy function Constraints are performed; the readout module of the non-causal branch also uses the OCR function and passes the cross entropy loss function Conduct individual training;

[0057] Model training involves 4 losses, and the total loss for disease classification is expressed as:

[0058]

[0059] Among them, λ1, λ2 and λ3 are hyperparameters of balance contribution, and ε represents the current number of training times.

[0060] In the experiment, it was found that since the model cannot distinguish non-causal features well in the early stage of training, if the constraints are too strong, the feature extraction of the model will become more chaotic; on the contrary, in the later stage of training, the constraints should be strengthened to obtain better results. Based on this observation, and The weights are configured separately and This weight gradually increases as the training progresses, and the three loss functions are combined according to the designed weights to achieve better overall performance.

[0061] Therefore, the present invention proposes a brain disease diagnosis model based on a multi-level causal intervention network, which has the following beneficial effects:

[0062] (1) The present invention can effectively filter out non-causal interference factors and extract high-order functional connectivity information to improve the accuracy and robustness of brain disease analysis.

[0063] (2) The present invention integrates high-order FC and brain transformers to continuously capture more comprehensive high-order features in the data, and guides the model to learn truly invariant causal features by utilizing a hierarchical causal-aware integrator as an observation lens.

[0064] (3) The present invention has both excellent performance and outstanding inter-site versatility, outperforming the current most advanced method models and demonstrating its important clinical value.

[0065] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 is a flow chart of a brain disease diagnosis model based on a multi-level causal intervention network of the present invention;

[0067] Figure 2 It is the result of biomarker visualization of the brain disease diagnosis model based on the multi-level causal intervention network of the present invention; wherein, (a) represents the 5 most important brain connections and their influence weights, (b) represents the connection strength of the 5 most important brain connections of ASD patients, and (c) represents the connection strength of the 5 most important brain connections of normal people. DETAILED DESCRIPTION

[0068] In order to make the technical solutions, advantages and purposes of the present invention clearer, the technical solutions of the embodiments of the present invention are clearly and completely described below. The described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the protection scope of this application.

[0069] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0070] like Figure 1As shown in the figure, functional connectivity networks can be used as effective biomarkers for identifying brain diseases, but traditional methods ignore high-order information, and fMRI data are susceptible to noise and inter-site domain offset. To this end, the present invention proposes a brain disease diagnosis model based on a multi-level causal intervention network, which integrates high-order FC and brain converters to capture more comprehensive features, and performs causal intervention through hierarchical causal-aware integrators to help the model learn truly invariant causal features. The experimental results on autism spectrum disorder (ASD) and major depressive disorder (MDD) datasets are better than existing method models and have potential for clinical application. The specific processing process includes the following steps in total:

[0071] (1) In the brain network pre-construction stage, data preprocessing is performed to extract fMRI time series, and Tangent Pearson Embedding (TPE) is used to construct the input features required for the model.

[0072] (2) In the feature extraction stage, a three-layer Transformer network with self-attention mechanism is used to extract features of the brain network.

[0073] (3) In the causal brain connection division stage, the HCI module is used to fuse multiple layers of attention and divide the brain network features into causal and non-causal parts accordingly.

[0074] (4) In the causal constraint stage, DCI is used to impose first-order causal constraints on the model to reduce the interference of non-causal factors. At the same time, ICAC is introduced into the causal part to impose second-order causal constraints on the model to further optimize the model.

[0075] (5) In the classification and recognition stage, the orthogonal clustering (OCR) readout function is used to classify brain diseases.

[0076] Among them, steps (2) to (5) are an end-to-end deep network. The deep network model needs to be trained in advance, and then the brain network matrix extracted in step (1) is input into it to complete the entire process of steps (2) to (5). Therefore, the actual operation process mainly requires the following three steps:

[0077] 1. Brain network pre-construction: extract fMRI time series through data preprocessing, and use Tangent Pearson Embedding (TPE) to construct the input features required by the model.

[0078] 2. Model training: Use the training data set to train subsequent models until the expected standards are met.

[0079] 3. Classification and recognition: input the data to be diagnosed into the trained model to achieve classification and recognition of the corresponding brain diseases.

[0080] The following is a specific example using the ABIDE data set to illustrate the three practical operation steps of the present invention:

[0081] 1) Brain network pre-construction; The ABIDE-I database includes 573 NC and 539 ASD. The preprocessed rs-fMRI data are publicly accessible through the Preprocessed Connectome Project (PCP) and preprocessed using the Configurable Connectome Analysis Pipeline (CPAC). After removing invalid samples with missing data, 1035 subjects were selected, including 530 NC and 505 ASD. The brain area was divided into 200 regions of interest (ROIs) using the Consensus Clustering 200 (CC200) template. After the above processing, the data of each sample was processed into a 200×200 matrix. After that, all the data were used using the TPE algorithm to obtain the required brain network matrix, where the dimension of the brain network matrix of each sample was still 200×200.

[0082] 2) Model training: The data were divided into training and test sets using two strategies: 10-fold cross validation and leave-one-site-out. Training was performed in Python under the PyTorch framework. The learning rate was set to 0.00001, and the four hyperparameters were set to γ ​​= 0.8, λ1 = 0.1, λ2 = 0.1, and λ3 = 1.4. The training process lasted 100 times, and the trained model weights were saved for later use.

[0083] 3) Classification and recognition: Input the data to be diagnosed into the trained model to classify and recognize the corresponding brain diseases. The evaluation of the model effect can be carried out in the following two aspects:

[0084] 1. Accuracy Evaluation

[0085] The following indicators are used to evaluate the performance of binary classification of patients with or without abnormalities: accuracy (Acc), area under the receiver operating characteristic curve (AUC), recall (Rec), specificity (Spec), F1 score (F1) and precision (Pre). The results of the model's ten-fold cross validation and site-leave-one-out method in the ABIDE dataset and the comparison with the existing optimal methods are shown in Tables 1 and 2, respectively. It can be seen from the results that this model has excellent disease diagnosis performance.

[0086] Table 1 Comparison results of ten-fold cross validation and existing optimal methods

[0087] ABIDE 10% off acc auc recall spec F1 prec RF 63.665 63.461 54.848 72.074 63.212 65.184 SVM 66.668 66.593 63.373 69.811 66.562 66.918 MVS 67.534 67.494 66.117 68.866 67.238 67.304 BolT 68.35 74.94 63.75 72.68 70.19 68.14 FBNETGEN 67.201 67.024 60.989 73.059 66.598 64.588 bnt 71.457 71.115 66.442 75.786 70.887 73.397 TP-MIDA 70.5237 77.21968 67.707 73.206 70.441 70.714 MBT 74.781 74.051 70.9 77.202 74.107 75.707 Proposed 75.449 74.49 71.599 77.382 74.644 76.572

[0088] Table 2 Comparison results between the leave-one-site method and the existing optimal method

[0089] ABIDE leave one acc auc recall spec F1 prec RF 60.659 61.128 54.723 67.604 59.158 64.9905 SVM 64.477 65.027 62.368 67.689 63.2845 66.2725 MVS 70.0085 69.61 66.206 73.013 68.9335 70.9995 BolT 67.8655 74.7955 59.165 75.46 69.616 65.37 FBNETGEN 70.21684 69.48368 68.39895 70.56684 68.8279 69.0895 bnt 71.6758 70.8295 70.5475 71.11 70.4163 71.463 TP-MIDA 69.2754 74.445 65.282 74.4545 69.401 70.846 MBT 74.806 74.663 69.3835 79.9425 73.819 78.879 Proposed 75.955 75.399 71.6125 79.184 75.2845 77.8785

[0090] 2. Biomarker Visualization

[0091] The causal multi-level fusion attention matrix calculated by the HCI module in the model reflects the importance of each brain region connection and can be used to extract the most discriminative brain region connections for disease diagnosis. These brain region connections can be used as a powerful tool for detecting ASD diseases. The average of all samples was taken, and the five largest connections were extracted as the five most important brain connections for ASD classification. The BrainNetViewer toolbox was used to visualize them, as shown in Figure 2. Figure 2 If you want to further compare the differences in brain connectivity between normal people and patients, you can map the extracted important brain connections to the TPE brain connectivity matrix of normal people and the TPE brain connectivity matrix of ASD patients, take the average of the two types of patients, and use the BrainNet Viewer toolbox to visualize it as shown below: Figure 2 (b) and Figure 2 As shown in (c) in the figure, by comparison, it can be found that normal people and ASD patients show obvious differences in these five most discriminative brain connections.

[0092] Therefore, the brain disease diagnosis model based on a multi-level causal intervention network proposed in the present invention can effectively filter out non-causal interference factors and extract high-order functional connection information to improve the accuracy and robustness of brain disease analysis; at the same time, by using a hierarchical causal-aware integrator as an observation lens, it guides the model to learn truly invariant causal features, demonstrating its important clinical value.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A brain disease diagnosis model based on a multi-level causal intervention network, including the following steps: S1, the brain network model pre-construction stage, firstly, data pre-processing and fMRI time series extraction are performed, and the input features required for the model are constructed using Tangent Pearson Embedding (TPE); S2, feature extraction stage, using a 3-layer Transformer network with self-attention mechanism to extract features of the brain network; S3, causal brain connection division stage, using the HCI module to fuse multiple layers of attention and divide brain network features into causal and non-causal parts accordingly; S4, causal constraint stage, DCI is used to impose first-order causal constraints on the model, and ICAC is used to impose second-order causal constraints on the model; S5, classification and recognition stage, using orthogonal clustering OCR readout function to realize the classification of brain diseases.

2. According to the brain disease diagnosis model based on the multi-level causal intervention network of claim 1, in S1, the preprocessed data is divided into 200 brain regions of interest using the CC200 template to construct a brain network matrix.

3. According to the brain disease diagnosis model based on the multi-level causal intervention network of claim 2, the connection patterns between different brain regions are different, and the characteristics of the brain network matrix are extracted by using the tangent Pearson embedding TPE. The calculation formula of the tangent Pearson embedding TPE is as follows: in, x i , x j , x k is an fMRI time series element, i, j, k are integer indexes, r represents the number of brain region ROIs, Cov(x i , x k ) is x i and x k The covariance between them, σ represents the standard deviation, P ij represents the covariance between the i-th row and j-th column of the Pearson correlation matrix, M is the number of subjects, log m is the matrix logarithm, P is the Pearson correlation matrix, P i is the element of the i-th row of the Pearson correlation matrix, and TPE is the tangent Pearson embedding matrix.

4. According to claim 1, the brain disease diagnosis model based on the multi-level causal intervention network, in S2, the Transformer network is constructed for feature extraction, the data fit must be re-evaluated and the attention mechanism must be adjusted to correspond to the unique features of the brain network data, using the connection profile tp i As the initial node features to establish the core network, where tp i is the corresponding row of node i in the tangent Pearson embedding matrix TPE.

5. According to the brain disease diagnosis model based on the multi-level causal intervention network of claim 4, the self-attention mechanism is an important module of Transformer. The traditional self-attention module is defined as: in, <.,.> represents the inner product, l represents the number of layers, Z l and Z l-1 Represent the network characteristics of layer l and layer l-1 respectively, is a learnable parameter, represent The size of the first dimension; Use a multi-head attention mechanism with H heads to work simultaneously, and combine their outputs through connections to form the overall output of the multi-head attention module, where H is an integer; Z l-1 The i-th and j-th rows of and attention For projection and The linear kernel between RBF and RBF nonlinear kernel The linear kernel is improved, and the formula is as follows:

6. According to the brain disease diagnosis model based on the multi-level causal intervention network of claim 1, in S3, in the causal brain connection division stage, the necessity of attention fusion lies in that the attention matrices at different levels capture multi-scale features from local to global respectively, and the global attention representation is obtained by the HCI method, and the formula is as follows: A f =α1A1+α2A2+α3A3; in, ⊙ represents the element-wise product, α1, α2, and α3 are the fusion weights of A1, A2, and A3 respectively. is the attention matrix of the first layer of the i-th head, is the attention matrix of the second layer of the i-th head, is the attention matrix of the third layer of the i-th head, A f is the fused attention matrix.

7. According to the brain disease diagnosis model based on the multi-level causal intervention network of claim 6, in order to separate the causal and non-causal parts, retain A f The first γ percent of the largest value in the γ-th percentile is set to zero, which is expressed as The remaining areas are non-causal focal points, represented by γ represents the proportion of causal features. At this time, Focusing only on non-causal regions, Focus on the causal area; Execute the forward propagation process of Transformer, respectively and Substitute the attention matrix part in the forward propagation to obtain the candidate causal part Z C and the non-causal part Z N .

8. According to the brain disease diagnosis model based on the multi-level causal intervention network of claim 1, in S4, DCI is applied to impose a first-order causal constraint on the model, and the feature Z is input into the readout function to obtain the output feature make and Represent the candidate causal prediction of the mth sample and the non-causal prediction of the nth sample respectively. Use different non-causal predictions as interventions and combine them with the causal prediction to form a composite prediction. The formula is as follows: Where sigm represents the Sigmoid function, For composite forecast; Then, by intervening in the loss function The first-order causal constraints are imposed on the model, and the formula is as follows: in, is the cross entropy loss function, E n [.] indicates the expectation of n, Var n [.] means taking the variance of n, y m is the category label of sample m.

9. According to the brain disease diagnosis model based on multi-level causal intervention network in claim 7, in S4, ICAC is used to impose second-order causal constraints on the model, and Z C Expand it into a column vector. Suppose the column vector has K elements, each element represents a neuron, and the kth neuron is z k , neuron z k Z C The causal attribution of is defined as the value z', and the average causal effect of z' is The calculation formula is as follows: E[Z C |to(from -k =z′)]=∫Z C ·p(Z C |to(from -k =z'))dZ C ξ=∫(∫Z C ·p(Z C |to(z=z′))dZ C )dz′ in, z represents the entire column vector, do(.) is the intervention method, p(.) is the process of calculating the probability of the deep network, z -k represents z divided by z k For all values ​​other than , ξ is the baseline value and E[.|.] is the conditional expectation; The intervention variable z′ satisfies a uniform distribution within the range of z, that is, for the i-th element z′ i , z′ i ~u(minz i , max z i ), ξ is the average distribution of z, representing its baseline value; the do(.) intervention method only retains z each time k The value of z k The distribution of all values ​​other than k , repeat the above method to quantify the true causal value of each variable to the model; The causal influence of all features on the i-th sample label is quantified by the invariant causal attribution ICA vector, which is the feature neuron z k The causal attribution of provides a metric, and the ICA vector of the k-th feature pair label y is defined as follows: By analyzing all K dimensions, the ICA vector f of the i-th sample is generated i , the formula is as follows: In order to evaluate the difference between the ICA vectors of the input features on the output y across domains, a contrast loss function is proposed In order to match the input ICA vector with the output of all instances of the same class across domains, the loss is optimized by minimizing the distance between inputs of the same class and maximizing the distance between inputs of different classes. The inputs of the same class are regarded as positive matches of the i-th sample, and the inputs of different classes are regarded as negative samples. The contrast loss formula is as follows: Among them, f i 、f j and f k are the ICA vectors of samples i, j, and k, respectively, y is the label, and τ is the temperature parameter for scaling cosine similarity.

10. The brain disease diagnosis model based on a multi-level causal intervention network according to claim 1, in S5, after intervention and treatment, Z C For the causal features most relevant to the disease, the orthogonal clustering readout function OCR is used to extract these features for the final classification; Set O cluster centers, each with V dimensions, and use the Softmax projection operator to calculate the probability P of assigning node i to cluster O io , the formula is as follows: in, Represents Z C The i-th row, u o is an orthogonal basis obtained by uniformly initializing O cluster centers using Xavier and using the Gram-Schmidt method; The final embedding Z G is obtained through aggregate learning, composed of the causal feature Z C The calculated distribution ρ is calculated as follows: WITH G =ρ T WITH C ; Z G Bring it into MLP to obtain the final classification result and use the cross entropy function Constraints are performed; the readout module of the non-causal branch also uses the OCR function and passes the cross entropy loss function Conduct individual training; Model training involves 4 losses, and the total loss for disease classification is expressed as: Among them, λ1, λ2 and λ3 are hyperparameters of balance contribution, and ε represents the current number of training times.

Images