Depression diagnosis and analysis method based on decision tree structure contrast learning

Abstract

The invention discloses a depression diagnosis and analysis method based on decision tree structure contrast learning. The depression diagnosis and analysis method comprises the following steps: step 1, preprocessing owned data; 2, selecting the characteristics of one modal to describe the coding process of the model; step 3, inputting the selected features into an encoder; 4, inputting yi obtained by encoding into a projection encoder to carry out feature dimension reduction; step 5, calculating classification loss; step 6, constructing a contrast learning loss according to the characterization; 7, repeating the steps 2-6 to obtain respective encoders and representations y and z of the four modal data; and 8, according to the gradient direction of the Lall, carrying out iterative updating on the adjustable weights in the encoders according to the gradient descending direction, and completing the learning process of the model. According to the method, MDD is diagnosed by using data of various modes, and the risk genes and brain regions related to the MDD are analyzed by combining the decision tree structure with the white-box characteristic, so that the diagnosis precision of the MDD is effectively improved.

Description

Technical Field

[0001] The invention relates to a depression diagnosis and analysis method based on decision tree structure contrast learning, and belongs to the technical field of imaging genomics. Background Art

[0002] Depression is a common affective mental disorder. Due to the lack of disease awareness among depression patients and the lack of early screening methods, most patients have already developed severe symptoms when diagnosed. Using patient information to build a depression diagnosis model can greatly help practitioners diagnose depression.

[0003] In the field of neuroimaging, magnetic resonance imaging (MRI) is widely used in the prediction and diagnosis of mental illness and neurological diseases. For example, MRI information is used to train a variety of machine learning and deep learning schemes (XGBoost, SVM, ViT, Transformer, etc.). They extract characteristic information of the sample brain from MRI to train the MDD diagnosis model, and continuously improve the model to make it more accurate and interpretable in diagnosis. However, these methods only use imaging data to train the model. Related studies have shown that using a single biological indicator cannot fully explain MDD, and its causes include genetic, environmental, psychological and biological factors. In order to enable the diagnostic model to learn more categories of sample features, many recent studies have begun to use deep learning models for disease diagnosis. They use multi-modal imaging information to train related models, such as multi-connected representation learning networks, multi-modal cross-Transformers, etc., and have achieved good results in diagnostic efficiency and accuracy. However, due to the black box characteristics of deep learning, it is difficult for researchers to perform attribution analysis on these models, and other data that can be obtained in the clinical process, such as genetic information, age, family history, etc., cannot be used.

[0004] At present, the diagnostic models of neurological diseases have made great progress in the fields of machine learning and deep learning. Machine learning methods are inferior to deep learning models in diagnostic accuracy, and the interpretability of deep learning models is not perfect, and it is impossible to determine the brain areas or functional connections related to the disease. In addition, due to the high dimensionality and complex associations of genetic data, existing models for the diagnosis of mental illness rarely pay attention to multimodal data of imaging and genes. Summary of the invention

[0005] The purpose of the present invention is to address the defects and shortcomings of the above-mentioned prior art and propose a depression diagnosis and analysis method based on decision tree structure contrast learning. The classification network is trained by utilizing brain magnetic resonance imaging (MRI) and single nucleotide polymorphisms (SNP) information commonly used in MDD diagnosis. Disease-related features are analyzed while achieving disease diagnosis. MMSC simultaneously utilizes data from four modalities, namely, two MRI (structural MRI and resting-state functional MRI), SNP, and clinical evaluation indicators, to perform multimodal data fusion to mine sample features, and the learned features are used in the diagnosis of MDD. The model effectively improves the diagnostic accuracy of MDD.

[0006] The technical solution adopted by the present invention to solve the technical problem is: a depression diagnosis and analysis method based on decision tree structure contrast learning, the method comprising the following steps:

[0007] Step 1: First, preprocess the data you have;

[0008] Step 2: Select the features of one of the modes to describe the encoding process of the model;

[0009] Step 3: Input the selected features into the encoder;

[0010] Step 4: Input the encoded yi into the projection encoder for feature dimensionality reduction;

[0011] Step 5: Calculate the classification loss;

[0012] Step 6: Construct contrastive learning loss based on representation;

[0013] Step 7: Repeat steps 2 to 6 to obtain the encoders and representations y and z of the four modal data respectively;

[0014] Step 8: According to L all The weights that can be adjusted in the above encoders are iteratively updated in the direction of gradient descent to complete the learning process of the model.

[0015] Furthermore, step 1 includes: for MRI data, using the resting state functional data processing assistant (DPARSF 2.3) MRI toolkit for preprocessing and assigning it to the standard MNI (montreal neurological institute) space, obtaining a 116*116 functional connection matrix through the AAL (automated anatomical labeling) template, and then calculating its clustering coefficient (CC) and mean gray matter volume (GMV) through the matrix as the features of rs-fMRI and sMRI respectively, for SNP information, using the number of the smallest allele to perform additive coding of 0, 1, and 2, for other clinical information, the discrete data are one-hot encoded, and all data are standardized, that is, the mean is 0 and the standard deviation is 1.

[0016] Furthermore, step 2 includes: assuming that the i-th modal feature in the model input is x i , which feeds this feature into the first encoder in the model To perform feature selection, where P i Represents the feature selection matrix, which consists of 0 and 1, and each row has only one value of 1. In this way, the input features are screened, and while ensuring the sparseness of the features, the important features are selected as the input of the subsequent model.

[0017] Further, the step 3 comprises: first To enable the decision tree to be iteratively updated, the sigmoid function and the ReLU activation function are used to construct the formula:

[0018] f(x;Θ)=W2(ReLU(W1x+b1))+b2

[0019] Among them, x represents a variable that needs to be judged by the decision tree, W and b represent the weights that can be iteratively updated in the decision tree nodes, and Θ represents the weight values ​​of W and b input into the formula.

[0020] The following differentiable decision tree function is obtained:

[0021] t(x i )=f(σ(x i1 -s1),σ(x i2 -s2),...,σ(x in -s n );Θ)

[0022] Among them, s is used to constrain the output of the decision tree node to prevent the size of the x value from affecting the function.

[0023] The output of the encoding layer is:

[0024]

[0025] Furthermore, the step 4 includes: the projection encoder is a common MLP network, that is:

[0026] z i =σ(W2·g(W1·y i +b1)+b2).

[0027] Furthermore, the step 5 includes: using the sample label l j and the sample representation y obtained in step 3 i Predicting classification results This process is also implemented using MLP, and the cross entropy loss is calculated with the true label of the sample:

[0028]

[0029] Furthermore, step 6 comprises: transforming the representation z obtained in step 4 i As input to the contrastive learning loss:

[0030]

[0031] Since the feature encoding process of different modalities is parallel, it is assumed here that the features of other modalities have been encoded, where N represents the number of samples, M represents the number of feature modes, A(j) represents the N samples after removing the jth sample, P(j) represents other samples with the same classification label as sample j, and supervised contrastive learning loss uses an exponential function to calculate the correlation between the two representations, where the numerator represents feature z ij The correlation between the features of positive samples of the same category as other categories. The denominator represents the feature z ij The correlation degree between the features and the rest. The · in the numerator and denominator represents the inner product of the matrix.

[0032] Furthermore, step 7 includes: by calculating the classification loss of each representation y and the contrastive learning loss of the representation z, the final loss of the model can be obtained:

[0033]

[0034] Furthermore, step 8 includes: in the iterative process, L all The value of is close to the minimum value, which reduces the Euclidean distance between the positive feature groups and increases the Euclidean distance between the negative feature groups.

[0035] Beneficial effects:

[0036] 1. The present invention proposes a new MDD diagnostic method, called multimodal supervised contrastive learning (MMSC), which aims to diagnose MDD using data from multiple modalities, and analyzes MDD-related risk genes and brain regions in combination with a decision tree structure with white-box characteristics. MMSC simultaneously utilizes data from four modalities, namely, two MRIs (structural magnetic resonance imaging and resting-state functional magnetic resonance imaging), SNPs, and clinical evaluation indicators, to perform multimodal data fusion to mine sample features, and uses the learned features in the diagnosis of MDD. The model effectively improves the diagnostic accuracy of MDD.

[0037] 2. The method of the present invention utilizes brain magnetic resonance imaging (MRI) and single nucleotide polymorphisms (SNP) information commonly used in MDD diagnosis to train a classification network, and analyzes disease-related features while achieving disease diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is the MMSC network structure diagram of the present invention.

[0039] Figure 2 This is a structural diagram of the differentiable decision tree of the coding layer of the present invention.

[0040] Figure 3 This is a schematic diagram of the MDD-related risk brain areas obtained through the model analysis of the present invention. DETAILED DESCRIPTION

[0041] The invention will be further described in detail below in conjunction with the accompanying drawings.

[0042] like Figures 1 to 3 As shown, the present invention proposes a depression diagnosis and analysis method based on decision tree structure contrast learning, which specifically includes the following steps:

[0043] Step 1: First, preprocess the available data. For MRI data, use the resting state functional data processing assistant (DPARSF 2.3) MRI toolkit for preprocessing and assign it to the standard MNI (Montreal Neurological Institute) space. Use the AAL (automated anatomical labeling) template to obtain a 116*116 functional connectivity matrix. Then use the matrix to calculate the clustering coefficient (CC) and mean gray matter volume (GMV) as the features of rs-fMRI and sMRI, respectively. For SNP information, the number of the smallest allele is used for additive coding of 0, 1, and 2. For other clinical information, the discrete data is one-hot encoded. All data are standardized, that is, the mean is 0 and the standard deviation is 1.

[0044] Step 2: Select one of the modal features to describe the encoding process of the model. Assume that the i-th modal feature in the model input is x i , which feeds this feature into the first encoder in the model To perform feature selection, where P i Represents the feature selection matrix, which consists of 0 and 1, and each row has only one value of 1. In this way, the input features are screened, and while ensuring the sparseness of the features, the important features are selected as the input of the subsequent model;

[0045] Step 3: Input the selected features into the encoder, first To enable the decision tree to be iteratively updated, the sigmoid function and the ReLU activation function are used to construct the formula:

[0046] f(x;Θ)=W2(ReLU(W1x+b1))+b2

[0047] Among them, x represents a variable that needs to be judged by the decision tree, W and b represent the discriminant weights that can be iteratively updated in the decision tree nodes, and Θ represents the W and b weight values ​​input into the formula.

[0048] The following differentiable decision tree function is obtained:

[0049] t(x i )=f(σ(x i1 -s1),σ(x i2 -s2),...,σ(x in -s n );Θ)

[0050] Among them, s is used to constrain the output of the decision tree node to prevent the size of the x value from affecting the function.

[0051] The output of the encoding layer is:

[0052]

[0053] Step 4: Input the encoded yi into the projection encoder for feature dimensionality reduction. The projection encoder is a common MLP network, namely:

[0054] z i =σ(W2·g(W1·y i +b1)+b2);

[0055] Step 5: Calculate the classification loss using the sample label l j The sample representation y obtained in step 3 above i Predicting classification results This process is also implemented using MLP, and the cross entropy loss is calculated with the true label of the sample:

[0056]

[0057] Step 6: Construct contrastive learning loss based on the representation, and transform the representation z obtained in step 4 above i As input to the contrastive learning loss:

[0058]

[0059] Since the feature encoding process of different modalities is parallel, it is assumed here that the features of other modalities have been encoded, where N represents the number of samples, M represents the number of feature modes, A(j) represents the N samples after removing the jth sample, P(j) represents other samples with the same classification label as sample j, and supervised contrastive learning loss uses an exponential function to calculate the correlation between the two representations, where the numerator represents feature z ij The correlation between the features of positive samples of the same category as other categories. The denominator represents the feature z ij The correlation degree with the rest of the features. The · in the numerator and denominator represents the inner product of the matrix.

[0060] Step 7: Repeat steps 2 to 6 to obtain the encoders and representations y and z of the four modal data. By calculating the classification loss of each representation y and the comparative learning loss of the representation z, the final loss of the model can be obtained:

[0061]

[0062] Step 8: According to L allThe weights that can be adjusted in the above encoders are iteratively updated in the direction of gradient descent to complete the learning process of the model. all The value of is close to the minimum value, which reduces the Euclidean distance between the positive feature groups and increases the Euclidean distance between the negative feature groups.

[0063] The simulation experiment and data of the present invention are as follows:

[0064] Common mental illness diagnostic models cannot effectively utilize multimodal data. MMSC effectively improves the diagnostic accuracy of MDD while solving the above problems. The present invention makes full use of multimodal data (imaging, genetics, and clinical) related to mental illness to construct an MDD diagnostic model. The model innovatively utilizes a contrastive learning structure to fuse multimodal data to eliminate heterogeneity between modalities, thereby further improving the accuracy and stability of diagnosis.

[0065] Due to the unsupervised nature of contrastive learning, high-dimensional MDD data cannot be effectively utilized. This paper proposes a new method based on the classic SimCLR contrastive learning model loss. We innovated and improved the model by using multimodal features as the positive example group of input features under supervised conditions. We greatly expanded the number of positive example groups in the numerator of the loss function from a single positive example to P(j)M positive examples, i.e. z' in the numerator, which strengthened the model's contrastive learning ability and rewrote its loss function as follows:

[0066]

[0067] By utilizing the supervised training method in classification problems, the number of positive sample groups in the contrastive learning process is expanded, thereby strengthening the model's contrast and learning capabilities and making the resulting representation more representative.

[0068] At the same time, the model incorporates a highly interpretable decision tree structure into the coding. While performing classification, it can analyze the model's decision-making process to obtain risk characteristics that are highly correlated with MDD, which is of great significance for the analysis and understanding of MDD pathology.

[0069] Through ten five-fold crossover experiments, we compared the performance of MMSC with other diagnostic models in MDD diagnosis, as shown in Table 1:

[0070] Table 1 Comparison of the performance of MMSC and other diagnostic models in the diagnosis of MDD

[0071]

[0072] Mean ± SD

[0073] Table 1 shows the accuracy (ACC), sensitivity (SEN), specificity (SPE), F1 score, and AUC of each model in MDD diagnosis. The values ​​are composed of the average and standard deviation of the cross-experiment. MMSC has the best results in accuracy, F1 score, and AUC, and is the most stable in sensitivity and specificity. It does not have the abnormal situation of extremely high sensitivity and extremely low specificity in other models. This abnormal result is caused by the model's tendency to judge the sample as an MDD patient.

[0074] 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 the technical solution of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A depression diagnosis and analysis method based on decision tree structure contrast learning, characterized in that: The method comprises the following steps: Step 1: First, preprocess the data you have; Step 2: Select the features of one of the modes to describe the encoding process of the model; Step 3: Input the selected features into the encoder; Step 4: Input the encoded yi into the projection encoder for feature dimensionality reduction; Step 5: Calculate the classification loss; Step 6: Construct contrastive learning loss based on representation; Step 7: Repeat steps 2 to 6 to obtain the encoders and representations y and z of the four modal data respectively; Step 8: According to L all The weights that can be adjusted in the above encoders are iteratively updated in the direction of gradient descent to complete the learning process of the model.

2. A depression diagnosis and analysis method based on decision tree structure contrast learning according to claim 1, characterized in that: The step 1 includes: for MRI data, using the resting state functional data processing assistant MRI toolkit to preprocess and assign it to the standard MNI space, obtaining a 116*116 functional connection matrix through the AAL template, and then calculating its clustering coefficient and average gray matter volume through the matrix as the characteristics of rs-fMRI and sMRI respectively, for SNP information, using the number of the smallest allele to perform additive encoding of 0, 1, and 2, for other clinical information, the discrete data therein are one-hot encoded, and all data are standardized, that is, the mean is 0 and the standard deviation is 1.

3. A depression diagnosis and analysis method based on decision tree structure contrast learning according to claim 1, characterized in that: The step 2 includes: assuming that the i-th modal feature in the model input is x i , which feeds this feature into the first encoder in the model To perform feature selection, where P i Represents the feature selection matrix, which consists of 0 and 1, and each row has only one value of 1. In this way, the input features are screened, and while ensuring the sparseness of the features, the important features are selected as the input of the subsequent model.

4. A depression diagnosis and analysis method based on decision tree structure contrast learning according to claim 1, characterized in that: The step 3 comprises: first To enable the decision tree to be iteratively updated, the sigmoid function and the ReLU activation function are used to construct the formula: f(x;Θ)=W2(ReLU(W1x+b1))+b2 Among them, x represents a variable that needs to be judged by the decision tree, W and b represent the discriminant weights that can be iteratively updated in the decision tree node, and Θ represents the weight values ​​of W and b input into the formula; The following differentiable decision tree function is obtained: t(x i )=f(σ(x i1 -s1),σ(x i2 -s2),...,σ(x in -s n );I) Among them, s is used to constrain the output of the decision tree node to prevent the size of the x value from affecting the function; The output of the encoding layer is:

5. A depression diagnosis and analysis method based on decision tree structure contrast learning according to claim 1, characterized in that: The step 4 includes: the projection encoder is a common MLP network, that is: z i =σ(W2·g(W1·y i +b1)+b2)。 6. A depression diagnosis and analysis method based on decision tree structure contrast learning according to claim 1, characterized in that: The step 5 includes: using sample label l j and the sample representation y obtained in step 3 i Predicting classification results This process is also implemented using MLP, and the cross entropy loss is calculated with the true label of the sample:

7. A depression diagnosis and analysis method based on decision tree structure contrast learning according to claim 1, characterized in that: The step 6 comprises: transforming the representation z obtained in step 4 i As input to the contrastive learning loss: Since the feature encoding process of different modalities is parallel, it is assumed here that the features of other modalities have been encoded, where N represents the number of samples, M represents the number of feature modes, A(j) represents the N samples after removing the jth sample, P(j) represents other samples with the same classification label as sample j, and supervised contrastive learning loss uses an exponential function to calculate the correlation between the two representations, where the numerator represents feature z ij The correlation between the features of positive samples of the same category as other categories. The denominator represents the feature z ij The correlation degree between the features and the rest. The · in the numerator and denominator represents the inner product of the matrix.

8. The depression diagnosis and analysis method based on decision tree structure contrast learning according to claim 1, characterized in that: The step 7 includes: by calculating the classification loss of each representation y and the contrastive learning loss of the representation z, the final loss of the model can be obtained:

9. A depression diagnosis and analysis method based on decision tree structure contrast learning according to claim 1, characterized in that: The step 8 comprises: making L all The value of is close to the minimum value, which reduces the Euclidean distance between the positive feature groups and increases the Euclidean distance between the negative feature groups.

Images