A method and system for assessing the severity of depression based on causal representation learning

Abstract

The present invention discloses a method and system for assessing the severity of depression based on causal representation learning, which relates to the field of artificial intelligence technology. An interview text between a subject and a psychologist and the score of each depression symptom are input into a depression assessment model that has been trained, and an estimated score of depression is output to assess the severity of the subject's depression. The training process of the depression assessment model is as follows: a training data set is constructed, and feature extraction is performed on interview samples in the training data set to obtain high-order semantic features of each interview sample; a causal decoupling representation module is used to decouple the high-order semantic features to obtain causal representations between depression symptoms, and a downstream depression estimation module is used to causally process the causal representations and then output a predicted total score of depression for the interview sample. The method and system for assessing the severity of depression achieve accurate assessment of the severity of depression and also provide certain insights into psychological mechanisms.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for assessing the severity of depression based on causal representation learning. Background Art

[0002] Depression is a mental illness that brings some distress to people. Therefore, there is an urgent need to understand the psychopathological mechanism of depression and develop an automated method to predict its severity, which will help to develop a more convenient, faster and more reliable method for assessing the severity of depression.

[0003] Existing studies have proposed deep learning models to assess the severity of depression. Although these deep learning models achieve high performance by capturing complex and unobservable features, they cannot provide an explanation for the psychopathological mechanisms of depression, resulting in low accuracy in assessing the severity of depression. Summary of the invention

[0004] Based on the technical problems existing in the background technology, the present invention proposes a method and system for assessing the severity of depression based on causal representation learning, which achieves accurate assessment of the severity of depression and provides certain insights into psychological mechanisms.

[0005] The present invention proposes a method for assessing the severity of depression based on causal representation learning, which inputs the interview text between the subject and the psychologist and the score of each depression symptom into the trained depression assessment model, and outputs the estimated score of depression to assess the severity of the subject's depression;

[0006] The training process of the depression assessment model is as follows:

[0007] Step 1: Obtain M semi-structured interview samples that contain both depression assessment sub-item scores and total scores to form a training data set, and perform feature extraction on the interview samples in the training data set to obtain high-order semantic features of each interview sample;

[0008] Step 2: Send the high-order semantic features of each interview text to the causal decoupling representation module in turn, encode the high-order semantic features based on the encoder, learn the decoupled potential symptom representation of depression through the supervision signal, establish the causal relationship between the potential symptom representations based on the causal discovery algorithm Notears, send the potential symptom representation to the directed acyclic graph and combine the causal relationship between the potential symptom representations to obtain the causal representation between the depression symptoms, and construct the first loss function to train the causal decoupling representation module;

[0009] Step three, freeze the causal representations between depression symptoms and send them to the downstream depression estimation module, calculate the total causal effect between the nodes in the causal graph based on the causal representation and normalize them to obtain the normalized causal effect coefficient between the nodes, calculate each node representation based on the normalized causal effect coefficient, concatenate each node representation and send it to the ordinal regression module to output the total score prediction value of depression of the interview sample, so as to evaluate the severity of depression and construct a second loss function to train the downstream depression estimation module.

[0010] Furthermore, in step three, the representation of each node is calculated based on the normalized causal effect coefficient, specifically:

[0011] The normalized causal effect coefficient is used as a linear combination to calculate the representation of each node. At this time, the node representation output by each node is as follows:

[0012] ;

[0013] in, represents the node representation, Representation Node The neighbor nodes of Indicates Layer Causal Effect Node To Node The causal effect coefficient, Indicates The weight matrix of the layer causal effects, Representation Node The corresponding causal representation, is the activation function, Represents the total number of layers of causal effect mechanisms.

[0014] When the prediction layer adopts a multi-head causal effect mechanism, the average strategy is used to return the node representation of each node:

[0015] ;

[0016] in, Representation Node The neighbor nodes of for The node index in .

[0017] Furthermore, in step 3, each node representation is concatenated and sent to the ordinal regression module. Specifically, the regression layer of the ordinal regression module is composed of (N-1) binary classifiers, each binary classifier outputs a corresponding depression prediction value, and all depression prediction values ​​are added together to obtain the total score prediction value of depression, thereby obtaining the depression assessment result;

[0018] ;

[0019] ;

[0020] ;

[0021] in, represents the total score prediction value of depression, It represents the concatenated representation after concatenating all node representations. Representation of splicing Through the The depression prediction value generated by the binary classifier, Indicates Concatenated representations in binary classifiers The probability of being predicted as class 1, Representation of splicing In the The corresponding category labels in the binary classifiers are is the activation function, Representation of splicing The linear transformation of represents the weight matrix, Indicates The bias term of a binary classifier.

[0022] Furthermore, in step 2, the first loss function The construction is as follows:

[0023] ;

[0024] in, is the reconstruction loss of the causal disentangled representation, represents the KL loss of the posterior distribution and the prior distribution, represents the KL loss between the approximate posterior of the latent variable and the conditional prior distribution, Represents data distribution Take expectations, Represents the latent variable Distribution Take expectations, Represents a given latent variable After that, reconstruct the data The logarithmic function of represents the KL divergence, Random noise representing the encoder parameterization The posterior distribution of Represents random noise The prior distribution of The latent variables representing the encoder parameterization The posterior distribution of Indicates a given condition Latent variables The prior distribution of and are the weight parameters used to adjust the KL divergence term and reconstruct the data Corresponding to the high-order semantic features of the interview text.

[0025] Furthermore, in step three, the second loss function The construction is as follows:

[0026] ;

[0027] in, represents the weight matrix, Table offset term, represents the total number of interview samples, Indicates The weights of the binary classifiers, Indicates splicing representation, Representation of splicing The linear transformation of Representation of splicing In the The corresponding category labels in the binary classifier.

[0028] Furthermore, in step 1, feature extraction is performed on the interview samples in the training data set to obtain high-level semantic features of each interview sample, specifically:

[0029] Encode the conversation sentences in each interview sample based on the BERT encoder to obtain a text vector;

[0030] Based on the temporal convolutional network, the contextual information between the text vectors of each sentence is captured to obtain the high-order semantic features of each interview sample.

[0031] A depression severity assessment system based on causal representation learning, which inputs the interview text between the subject and the psychologist and the score of each depression symptom into the trained depression assessment model, and outputs the estimated depression score to assess the severity of the subject's depression;

[0032] The depression assessment model includes a feature extraction module, a causal decoupling representation module and a downstream depression estimation module. The training process of the depression assessment model is as follows:

[0033] The feature extraction module is used to extract features from interview samples in a training data set to obtain high-order semantic features of each interview sample, wherein the training data set is M semi-structured interview samples that include both depression assessment sub-item scores and total scores;

[0034] The causal decoupling representation module is used to decouple the high-order semantic features of each interview text sent in, encode the high-order semantic features based on the encoder, learn the decoupled potential symptom representation of depression through the supervision signal, establish the causal relationship between the potential symptom representations based on the causal discovery algorithm Notears, send the potential symptom representation into a directed acyclic graph and obtain the causal representation in combination with the causal relationship between the potential symptom representations, and construct a first loss function to train the causal decoupling representation module;

[0035] The downstream depression estimation module is used to evaluate depression based on the causal representations between the frozen depression symptoms, calculate the total causal effect between the nodes in the causal graph based on the causal representations and normalize them to obtain the normalized causal effect coefficients between the nodes, calculate each node representation based on the normalized causal effect coefficients, concatenate each node representation and send it to the ordinal regression module to output the total score prediction value of depression of the interview sample, thereby evaluating the severity of depression and constructing a second loss function to train the downstream depression estimation module.

[0036] The advantages of the depression severity assessment method and system based on causal representation learning provided by the present invention are: a depression assessment model is proposed for depression severity assessment, and the depression assessment model will also learn the causal relationship and dynamic evolution process between the eight symptoms of depression, which provides a psychological pathogenesis of depression to a certain extent; specifically, the depression assessment model first learns the causal representation of depressive symptoms, and then uses a graph causal effect network to simultaneously model symptom nodes and causal relationships, thereby achieving accurate assessment of depression severity and explaining the causal relationship between symptoms, providing certain insights into psychological mechanisms. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a structural schematic diagram of the present invention;

[0038] Figure 2 Flowchart for training a model for depression evaluation. DETAILED DESCRIPTION

[0039] Below, the technical solution of the present invention is described in detail through specific embodiments. Many specific details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific implementation disclosed below.

[0040] like Figure 1 and 2 As shown, the present invention proposes a method for assessing the severity of depression based on causal representation learning, which inputs the interview text between the subject and the psychologist and the score of each depression symptom into the trained depression assessment model, and outputs the estimated score of depression to assess the severity of the subject's depression;

[0041] The training process of the depression assessment model is as follows:

[0042] Step 1: Obtain M semi-structured interview samples that contain both depression assessment sub-item scores and total scores to form a training data set, and perform feature extraction on the interview samples in the training data set to obtain high-order semantic features of each interview sample;

[0043] The feature extraction module extracts features from the interview samples. The feature extraction module includes a BERT encoder and a one-dimensional temporal convolutional neural network (TCN). First, the BERT encoder is used to encode the conversation sentences in each interview sample to obtain a text vector.

[0044] ;

[0045] in, is the text vector of the conversation sentence, represents the turn of a semi-structured conversation, represents the dimension of the text vector, Indicates the encoding separator.

[0046] In order to capture the contextual information between each utterance, the temporal convolutional network (TCN) is then used to capture the contextual information between the text vectors of each utterance to obtain high-order semantic features of each interview sample.

[0047] ;

[0048] in, As high-order semantic features of each interview text in the training dataset.

[0049] Step 2: Send the high-order semantic features of each interview text to the causal decoupling representation module in turn, encode the high-order semantic features based on the encoder, learn the decoupled potential symptom representation of depression through the supervision signal, establish the causal relationship between the potential symptom representations based on the causal discovery algorithm Notears, send the potential symptom representation to the directed acyclic graph and combine the causal relationship between the potential symptom representations to obtain the causal representation, and construct the first loss function to train the causal decoupling representation module;

[0050] The causal decoupled representation module includes an encoder, potential symptom decoupled representation, causal discovery of symptoms, causal graph of symptoms and a decoder. The encoder is composed of a 3-layer fully connected MLP, and the decoder is also composed of the same. The decoupled depression symptom representation is learned through supervision signals, and then the causal relationship between symptoms is established by using the causal discovery algorithm Notears. The decoupled symptom representation is sent to a directed acyclic graph to obtain a causal representation.

[0051] The results of causal representation learning can usually be expressed as a causal graph, which is an intuitive way to show the causal relationship between variables, while the causal representation is a specific mathematical description of this causal relationship. This implementation first uses the causal discovery algorithm Notears to establish the causal relationship between potential symptom representations, which is equivalent to pre-training, and then uses the causal decoupling representation module to establish the causal representation between depression symptoms from high-dimensional data / text.

[0052] Notears is a method for learning causal structures through continuous optimization, which consists of the keywords non-combinatorial optimization, trace of matrix exponential, augmented Lagrangian and structural learning.

[0053] Specifically, the causal decoupling representation module is a variational autoencoder VAE architecture, and the parameterized conditional generation model is:

[0054] ;

[0055] in, , represents the encoder, Represents a decoder, represents the adjacency matrix between symptom nodes, Table latent variables Sufficient statistics; Represents the joint probability distribution of the conditional generation model, corresponding to the high-order semantic features of the interview samples , is a latent variable, is an independent noise variable, is the conditional variable, For a given latent variable , independent noise variables and condition variables Post-generation reconstruction data The conditional probability of For a given condition variable Latent variables are generated later and independent noise variables The conditional probability of .

[0056] Then the encoding and decoding process is:

[0057] ;

[0058] in, and represents independent noise terms with probability density, respectively and correspond, is the noise term The probability distribution of is the noise term Probability distribution of; Encoder: Reconstruct data from and condition variables The latent variables are encoded in and the noise term ; Decoder: transforms latent variables Decode to reconstruct data , and add the noise term .

[0059] Then, the generation model and inference model are expressed as follows:

[0060] ;

[0061] Generative Models Stated as: Given a latent variable And the noise term , generate reconstruction data Here we assume that the reconstructed data Depend on Generate and add the noise term . Inference Model Expressed as: Reconstructed data and condition variables Encoding to obtain latent variables and independent noise variables , and then through the independent noise term , and estimate the conditional posterior.

[0062] The joint probability of can be defined as:

[0063] ;

[0064] in, , the joint probability Stated as: In the condition variable Next, the latent variable and independent noise variables The joint distribution of Equal to the independent noise variable Distribution and latent variables Given a condition variable Distribution The product of .

[0065] The prior is based on the conditional variable The factorized Gaussian distribution of is:

[0066] ;

[0067] in, represents a Gaussian distribution with mean and variance, Represents a conditional variable The function of corresponds to the mean of the Gaussian distribution, Representation variables The function of , corresponding to the variance of the Gaussian distribution, and is any value, the distribution has two sufficient statistics, namely The mean and variance of , Parameters used to define the Gaussian distribution is the mean, is the variance. In this implementation, set and .

[0068] exist In the formula, this implementation uses a function and another function to parameterize the mean and variance of the prior distribution. This conditional prior distribution not only takes into account the latent variables The distribution of Dynamically adjust its parameters to capture richer contextual information.

[0069] After using VAE to obtain the latent symptom factors, the structural causal model (SCM) is used to understand the causal relationship between these symptoms. In this work, the linear SCM is used to model the causal relationship between these symptoms, and the calculation process is:

[0070] ;

[0071] in, represents the adjacency matrix of the causal graph, represents the transposed matrix of the adjacency matrix, independent noise variables It follows a normal distribution. represents the generated causal representation, Represents the identity matrix.

[0072] For example, through the above The formula calculates the causal representations between the eight symptoms of depression, and inputs these eight causal representations into the downstream depression estimation module for depression assessment.

[0073] The first loss function The construction is as follows:

[0074] ;

[0075] in, is the reconstruction loss of the causal decoupling representation, which means that given the latent variable After reconstructing the data The expected log probability of reconstruction, which measures the ability of the model to reconstruct the input data; Represents the KL loss of the posterior distribution and the prior distribution, corresponding to the posterior distribution With prior distribution The KL divergence between The negative sign before indicates that this part is the penalty term in the loss function; Represents the KL loss between the approximate posterior of the latent variable and the conditional prior distribution, corresponding to the posterior distribution With the prior distribution The KL divergence between the two also acts on the regularization part of VAE, the purpose is to make the latent variables With given conditions Consistent.

[0076] Represents data distribution Take expectations, Represents the latent variable Distribution Take expectations, Represents a given latent variable After that, reconstruct the data The logarithmic function of represents the KL divergence, Random noise representing the encoder parameterization The posterior distribution of Represents random noise The prior distribution of The latent variables representing the encoder parameterization The posterior distribution of Indicates a given condition Latent variables The prior distribution of and are the weight parameters used to adjust the KL divergence term.

[0077] When training the causal decoupling representation module through the first loss function, the causal decoupling representation module is trained by taking the high-order semantic features of multiple interview texts as the training set. The training stopping condition is that the loss function is fully converged or the set training rounds are reached. Then, the high-order semantic features of another multiple interview texts are used as the verification set to verify the trained causal decoupling representation module. When the loss of the verification set does not decrease, the training is terminated.

[0078] Step three, freeze the causal representations between depression symptoms and send them to the downstream depression estimation module, calculate the total causal effect between the nodes in the causal graph based on the causal representation and normalize them to obtain the normalized causal effect coefficient between the nodes, calculate each node representation based on the normalized causal effect coefficient, concatenate each node representation and send it to the ordinal regression module to output the total score prediction value of depression of the interview sample, so as to evaluate the severity of depression and construct a second loss function to train the downstream depression estimation module.

[0079] The downstream depression estimation module includes a graph causal effect network and an ordinal regression module. The calculation method of the graph causal effect network is as follows:

[0080] First, a causal graph is constructed based on the causal representation. In the causal graph, the nodes correspond to the causal representations obtained from the above records, and the edges between the nodes are the total causal effects, where the total causal effect The calculation is as follows:

[0081] ;

[0082] in, represents the identity matrix, An adjacency matrix representing a causal graph.

[0083] Secondly, the softmax function is used to normalize the total causal effect:

[0084] ;

[0085] in, Representation Node and nodes The total causal effect between Indicated in Direction is determined by the softmax function Normalize it, Representation Node To one of its neighbor nodes The causal effect of Representation Node and nodes The total causal effect between Representation Node neighbor nodes.

[0086] When the normalized causal effect coefficients are obtained, these coefficients will be used to calculate the linear combination of the features corresponding to them, and the node representation will be output based on the multi-head causal effect mechanism. At this time, the node representation output by each node is as follows.

[0087] ;

[0088] in, represents the node representation, Representation Node The neighbor nodes of Indicated in Node in layer To Node The causal effect coefficient, Indicates The weight matrix of the layer causal effects, Representation Node The corresponding causal representation, is the activation function, Represents the total number of layers of causal effect mechanisms.

[0089] In particular, when the prediction layer adopts a multi-head causal effect mechanism, the average strategy is used to return the node representation of each node:

[0090] ;

[0091] in, Representation Node The neighbor nodes of for The node index in .

[0092] After passing through the graph causal effect network, each node representation is concatenated and sent to the ordinal regression module. The goal of ordinal regression is to predict the value of an ordered variable, and it is more concerned with capturing the relative order of the output variable. The regression layer of the ordinal regression module consists of (N-1) binary classifiers, each of which outputs a corresponding depression prediction value. All depression prediction values ​​are added together to obtain the total depression score prediction value, thereby obtaining the depression assessment result.

[0093] No. The predicted empirical probability of a binary classifier is derived as follows:

[0094] ;

[0095] Therefore, the total score prediction for depression was defined as follows:

[0096] ;

[0097] in, It represents the total score prediction value of depression, which is also a ranking index. It is an integer in the range of {1,2,..., N}, where N is a set value. In this embodiment, N=24, so Belongs to 0-24, requiring 23 binary classifiers, It represents the concatenated representation after concatenating all node representations. Representation of splicing Through the The depression prediction value generated by the binary classifier, Indicates Concatenated representations in binary classifiers The probability of being predicted as class 1, Representation of splicing In the The corresponding category labels in the binary classifiers are is the activation function, Representation of splicing The linear transformation of represents the weight matrix, Indicates The bias term of a binary classifier.

[0098] The second loss function The construction is as follows:

[0099] ;

[0100] in, represents the weight matrix, Table offset term, represents the total number of interview samples, Indicates The weights of the binary classifiers, Indicates splicing representation, Representation of splicing The linear transformation of Representation of splicing In the The corresponding category labels in the binary classifier.

[0101] When the causal decoupling representation module is trained by the first loss function, the downstream depression estimation module is trained by taking the causal representation between depression symptoms as the training set. The training stopping condition is that the loss function fully converges or reaches the set training rounds. Then, the trained downstream depression estimation module is verified by taking the causal representation between other depression symptoms as the verification set. When the loss of the verification set does not decrease, the training is terminated.

[0102] This example uses the trained depression assessment model to perform the depression estimation task. Its input is the interview dialogue text and the score of each symptom, and the output result is the depression estimation score. At the same time, the adjacency matrix and node representation can be output to explain the causal relationship between symptoms and provide certain insights into psychological mechanisms.

[0103] According to steps one to three, this embodiment proposes a depression assessment model for assessing the severity of depression. The depression assessment model will also learn the causal relationship and dynamic evolution process between the eight symptoms of depression, which to a certain extent provides the psychological pathogenesis of depression; specifically, the depression assessment model first learns the causal representation of depressive symptoms, and then uses a graph causal effect network to simultaneously model symptom nodes and causal relationships, thereby achieving an accurate assessment of the severity of depression.

[0104] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for assessing the severity of depression based on causal representation learning, characterized in that: The interview text between the subject and the psychologist and the score of each depression symptom are input into the trained depression assessment model, and the depression estimation score is output to assess the severity of the subject's depression; The training process of the depression assessment model is as follows: Step 1: Obtain M semi-structured interview samples that contain both depression assessment sub-item scores and total scores to form a training data set, and perform feature extraction on the interview samples in the training data set to obtain high-order semantic features of each interview sample; Step 2: The high-order semantic features of each interview text are sequentially fed into the causal decoupling representation module, the high-order semantic features are encoded based on the encoder, the decoupled potential symptom representation of depression is learned through the supervision signal, the causal relationship between the potential symptom representations is established based on the causal discovery algorithm, the potential symptom representation is fed into a directed acyclic graph and the causal relationship between the potential symptom representations is combined to obtain the causal representation between the depression symptoms, and the first loss function is constructed to train the causal decoupling representation module; Step three, freeze the causal representations between depression symptoms and send them to the downstream depression estimation module, calculate the total causal effect between the nodes in the causal graph based on the causal representation and normalize them to obtain the normalized causal effect coefficient between the nodes, calculate each node representation based on the normalized causal effect coefficient, concatenate each node representation and send it to the ordinal regression module to output the total score prediction value of depression of the interview sample, so as to evaluate the severity of depression and construct a second loss function to train the downstream depression estimation module.

2. The method for assessing the severity of depression based on causal representation learning according to claim 1, characterized in that: In step 3, the representation of each node is calculated based on the normalized causal effect coefficient, specifically: The normalized causal effect coefficient is used as a linear combination to calculate the representation of each node. At this time, the node representation output by each node is as follows: in, represents the node representation, Representation Node The neighbor nodes of Indicates Layer Causal Effect Node To Node The causal effect coefficient, Indicates The weight matrix of the layer causal effects, Representation Node The corresponding causal representation, is the activation function, represents the total number of layers of causal effect mechanisms; When the prediction layer adopts a multi-head causal effect mechanism, the average strategy is used to return the node representation of each node: in, Representation Node The neighbor nodes of for The node index in .

3. The method for assessing the severity of depression based on causal representation learning according to claim 2, characterized in that: In step 3, each node representation is concatenated and sent to the ordinal regression module. Specifically, the regression layer of the ordinal regression module is composed of (N-1) binary classifiers. Each binary classifier outputs a corresponding depression prediction value. All depression prediction values ​​are added together to obtain the total score prediction value of depression, thereby obtaining the depression assessment result. in, represents the total score prediction value of depression, It represents the concatenated representation after concatenating all node representations. Representation of splicing Through the The depression prediction value generated by the binary classifier, Indicates Concatenated representations in binary classifiers The probability of being predicted as class 1, Representation of splicing In the The corresponding category labels in the binary classifiers are is the activation function, Representation of splicing The linear transformation of represents the weight matrix, Indicates The bias term of a binary classifier.

4. The method for assessing the severity of depression based on causal representation learning according to claim 1, characterized in that: In step 2, the first loss function The construction is as follows: in, is the reconstruction loss of the causal decoupled representation, represents the KL loss of the posterior distribution and the prior distribution, represents the KL loss between the approximate posterior of the latent variable and the conditional prior distribution, Represents data distribution Take expectations, Represents the latent variable Distribution Take expectations, Represents a given latent variable After that, reconstruct the data The logarithmic function of represents the KL divergence, Random noise representing the encoder parameterization The posterior distribution of Represents random noise The prior distribution of The latent variables representing the encoder parameterization The posterior distribution of Indicates a given condition Latent variables The prior distribution of and are the weight parameters used to adjust the KL divergence term and reconstruct the data Corresponding to the high-order semantic features of the interview text.

5. The method for assessing the severity of depression based on causal representation learning according to claim 3, characterized in that: In step 3, the second loss function The construction is as follows: in, represents the weight matrix, Table offset term, represents the total number of interview samples, Indicates The weights of the binary classifiers, Indicates splicing representation, Representation of splicing The linear transformation of Representation of splicing In the The corresponding category labels in the binary classifier.

6. The method for assessing the severity of depression based on causal representation learning according to claim 1, characterized in that: In step 1, feature extraction is performed on the interview samples in the training data set to obtain the high-level semantic features of each interview sample, specifically: Encode the conversation sentences in each interview sample based on the BERT encoder to obtain a text vector; Based on the temporal convolutional network, the contextual information between the text vectors of each sentence is captured to obtain the high-order semantic features of each interview sample.

7. A depression severity assessment system based on causal representation learning, characterized in that: The interview text between the subject and the psychologist and the score of each depression symptom are input into the trained depression assessment model, and the depression estimation score is output to assess the severity of the subject's depression; The depression assessment model includes a feature extraction module, a causal decoupling representation module and a downstream depression estimation module. The training process of the depression assessment model is as follows: The feature extraction module is used to extract features from interview samples in a training data set to obtain high-order semantic features of each interview sample, wherein the training data set is M semi-structured interview samples that include both depression assessment sub-item scores and total scores; The causal decoupling representation module is used to decouple the high-order semantic features of each input interview text, encode the high-order semantic features based on the encoder, learn the decoupled potential symptom representation of depression through the supervision signal, establish the causal relationship between the potential symptom representations based on the causal discovery algorithm, send the potential symptom representation to the directed acyclic graph and obtain the causal representation in combination with the causal relationship between the potential symptom representations, and construct a first loss function to train the causal decoupling representation module; The downstream depression estimation module is used to evaluate the severity of depression based on the causal representations between the frozen depression symptoms, calculate the total causal effect between the nodes in the causal graph based on the causal representations and normalize them to obtain the normalized causal effect coefficients between the nodes, calculate each node representation based on the normalized causal effect coefficients, concatenate each node representation and send it to the ordinal regression module to output the total score prediction value of depression of the interview sample, thereby evaluating the severity of depression and constructing a second loss function to train the downstream depression estimation module.

Images