Multi-dimensional mental health assessment method based on double-layer learning and module control

Abstract

The invention discloses a multi-dimensional psychological health assessment method based on double-layer learning and module control, which comprises module control weighting, a double-layer learning framework and bidirectional consistency constraint, and is characterized in that: module control weighting: based on a minimum dominating set, a control frequency and the status of different sub-items of an explicit region component table in a depression network, the sub-items of the explicit region component table are determined; differential weight assignment is realized; according to the double-layer learning framework, sub-items are predicted one by one through models, then a total score is obtained through weighted merging, multi-dimensional symptom information is reserved, and final overall judgment is connected; and bidirectional consistency constraint: adding a sub-item-total-branch coupling item into a loss function to ensure that local prediction and overall prediction are mutually matched, so that the interpretability and the practical value are enhanced. According to the method, subproject-level fine prediction and core entry identification are realized, and subproject prediction and overall score prediction are bidirectionally constrained in the same framework, so that local and overall inconsistency is prevented, and the interpretability and reliability of the method are improved.

Description

Technical Field

[0001] The present invention relates to a multi-dimensional mental health assessment method based on double-layer learning and module control, and belongs to the field of mental health assessment. Background Art

[0002] Depression is one of the most common and highly harmful mental disorders in contemporary society, having a profound impact on an individual's emotions, thinking, actions, and social functions. Currently in clinical practice, in traditional psychological or mental health assessments, many scales (such as PHQ-9, HAMD-17, etc.) often simply sum up the scores of each sub-item (entry) to form a total score to measure whether the subject has a depressive tendency or the degree of depression. However, such an approach has obvious limitations in multi-dimensional symptom analysis and refined assessment:

[0003] On the one hand, when different individuals have the same total score, the differences in specific sub-items may be extremely significant; some people may have higher scores in sleep disorders and appetite changes, while others may show more obvious manifestations in low mood and anhedonia. If we only rely on a total score to describe the entire depressive state, it is very easy to mask these heterogeneous differences, and there is also a lack of individual attention and explanatory ability for some key symptom entries.

[0004] On the other hand, the contribution degrees of different sub-items to the overall depressive risk are not the same. Some core symptoms or comorbid factors often have a profound impact in clinical practice, while other items are relatively less important. If we treat each sub-item equally or simply add them up when modeling, it is difficult to ensure sufficient attention to the "core nodes" in the symptom network, and the "peripheral nodes" will not be over-amplified.

[0005] In addition, in the scenario of simultaneously predicting multiple sub-items and the total score, there is often a problem that the sub-item model and the total score model are independently trained and do not cooperate with each other. If we train a separate model for each sub-item and then train another model for the total score, the results finally output by these two parts are very likely to be contradictory. For example, the sum of the sub-item prediction values does not match the overall score prediction value, making it difficult to determine which side to trust.

[0006] In recent years, with the development of digital technology and artificial intelligence, researchers have begun to pay attention to the feasibility of somatization features such as physiological signals or imaging information in mental health detection. Using deep learning or machine learning methods to mine these features helps to efficiently identify people with depression or potential risks.

[0007] But in specific practice, existing detection methods usually have difficulty overcoming the following limitations:

[0008] (1) Only output results in the form of "depressed / non-depressed" or "total score", ignoring the diverse and heterogeneous nature of depressive symptoms, which is not conducive to detailed analysis and subsequent intervention guidance.

[0009] (2) Lack of interpretability. Especially in scenarios with multiple sub-items and multiple dimensions, simply integrating all relevant features to predict a single score makes it difficult for professionals to clearly identify which symptom dimensions are more worthy of attention.

[0010] (3) Unable to balance the prediction consistency between the sub-item level and the overall level. Many studies handle these two issues separately, resulting in contradictions in the comprehensive output of the model.

[0011] At the same time, in recent years, the concept of Symptom Network has been widely applied in the research of affective disorders and mental health. Its main idea is that depression is not just a single factor or a certain latent variable, but a network formed by the mutual influence of various symptom nodes (such as insomnia, low mood, appetite changes, inattention, etc.). Some symptom nodes are more "central", strongly correlated with many nodes in the network, thus playing a key role in the overall depression risk. However, the application has not been fully expanded, and no corresponding in-depth analysis has been carried out on other modalities of data.

[0012] In summary, with the continuous progress of feature extraction technologies such as physiological signals and machine learning algorithms, numerous studies have attempted to use these new technologies to assist in mental health screening. However, in this field, the existing technologies still have obvious deficiencies in the following aspects:

[0013] (1) A large number of algorithms pay little attention to the details of the sub-items of the scale. They often simply combine all items into a total score and train a model to directly map from the feature data to this total score. Although this simplifies the process in some applications, it will completely eliminate the heterogeneity of different sub-symptoms, making it difficult to provide strong support for personalized intervention or accurate diagnosis.

[0014] (2) The structural associations between sub-items are not fully utilized. Some studies directly average the sub-item scores or add them in a fully connected layer, lacking the ability to distinguish between "core symptom nodes" and "peripheral symptom nodes", resulting in important items not receiving more attention in the model and allowing weakly related items in the network to cause noise interference to the results.

[0015] (3) The overall and local prediction results do not match. Some methods attempt to perform multi-task prediction on sub-items, but then train another model for the total score. Eventually, these two models may draw inconsistent conclusions during inference, thereby reducing the credibility and interpretability of the model.

[0016] Based on the above problems, the present invention proposes a technical solution that can accurately and uniformly model the multi-dimensional sub-items and the overall score of the scale with the assistance of feature data; the "module control network" is used to obtain the differential weights between sub-items, and the prediction of each sub-item and the prediction of the overall score are considered under the principle of "bidirectional consistency", significantly improving the clinical and application value of the method. Summary of the Invention

[0017] The object of the present invention is to provide a multi-dimensional mental health assessment method based on double-layer learning and module control. By combining a double-layer learner with module control weighting, it not only realizes fine prediction at the sub-item level and core item identification, but also enables the sub-item prediction and the overall score prediction to be bidirectionally constrained within the same framework, preventing inconsistencies between the local and the whole, and enhancing the interpretability and reliability of the method.

[0018] To achieve the above object, the technical solution adopted by the present invention is: a multi-dimensional mental health assessment method based on double-layer learning and module control, which jointly realizes the multi-dimensional depression detection of the scale through module control weighting, a double-layer learning framework, and bidirectional consistency constraints, specifically as follows:

[0019] Module control weighting: Based on the minimum dominating set and control frequency, explicitly distinguish the status of different sub-items of the scale in the depression network, and realize differential weight assignment;

[0020] Double-layer learning framework: First, predict each sub-item with a single model, and then weighted merge to obtain the total score, which not only retains multi-dimensional symptom information but also connects to the final overall judgment;

[0021] Bidirectional consistency constraint: Add a sub-item-total score coupling term to the loss function to ensure that the local prediction and the overall prediction cooperate with each other, thereby enhancing the interpretability and practical value.

[0022] Furthermore, the module control weighting includes three parts: the construction of the symptom-related network, the minimum dominating set and control frequency, and the generation of sub-item weights. Specifically:

[0023] (1) Construction of the symptom-related network

[0024] Regard all sub-items as nodes in the network; assume that the scale has J sub-items. To obtain the correlation degree of the sub-items in the overall depression structure, first construct a correlation matrix Where:

[0025] R ij =corr(y i ,y j ),

[0026] y i 、y jis the true score or measurement value of sub-item i, j, and corr(·,·) uses Pearson or Spearman correlation; then, the sparse inverse covariance estimation method is used to explore more direct interaction relationships and eliminate some redundant edges;

[0027] (2) Minimum dominating set and control frequency

[0028] The minimum dominating set is introduced to quantitatively measure the control or coverage of a sub-project on other nodes in the network. The definition is: is a dominating set, and finding the minimum dominating set is to find the set that makes D as small as possible, specifically using a random greedy search algorithm for multiple iterations;

[0029] Perform N rounds of randomization iterations, and obtain an approximate MDS in each round. r , r = 1, ..., N, for each node k, count how many rounds it appears in D, and record the number of appearances as n k , define the control frequency of the node:

[0030]

[0031] CF k The higher it is, the more often node k appears in the minimum dominating set, or has strong connections with more nodes, and its core position is more significant;

[0032] (3) Sub-project weight generation

[0033] The control frequency CF k After normalization to the interval [0,1], it can be regarded as the importance weight w of node k k .When summing up the predictions later, the contribution of the sub-items in the total score will be highlighted or weakened accordingly, thus achieving explainability weighting.

[0034] Furthermore, the two-layer learning framework includes independent prediction of sub-items in the first layer and total score aggregation in the second layer.

[0035] Furthermore, in the independent prediction of sub-items, for each sub-item y j (j=1,…,J), respectively in a set of candidate models Select the best one from the above to predict the score of the sub-item; let the feature vector be X i Represents the comprehensive characteristics of the i-th subject, including:

[0036]

[0037] in Contains feature data of different modes. For each sub-item, let:

[0038]

[0039] Among them, CV_score represents the score quantifying the model effect during the cross-validation process; for the finally selected optimal model, the input X i predicts the estimate of the sub-item score

[0040] Furthermore, in the total score aggregation, to align with the common form of the total score of the scale, the weighted sum is calculated to obtain the overall depression prediction

[0041]

[0042] where w j That is, from the control frequency normalization weight, reflecting the relative importance of sub-item j in the symptom network.

[0043] Further, in the bidirectional consistency constraint, a bidirectional consistency constraint is introduced during the training and optimization process to couple the prediction of the sub-item and the total score under the same objective function.

[0044] Furthermore, denote y j,i as the true score of the i-th subject on sub-item j, as the true total depression score; the predicted values of each sub-item and the weighted sum obtained For all samples i = 1, …, N, define the comprehensive loss:

[0045]

[0046] where:

[0047] Constrains the prediction accuracy of each sub-item to make the sub-item as close as possible to the true value;

[0048] Takes into account the prediction error of the overall depression score;

[0049] Forces the overall weighted sum and to be consistent, thus realizing the coupling of the sub-item and the total score;

[0050] Introduces a correlation coefficient constraint to encourage the prediction sequence of each sub-item to be consistent with the true sequence at the linear correlation level and minimize random noise as much as possible;

[0051] During the specific training, adjust λ according to the data scale, the number of sub-items, and the priority attention target 1 , λ 2 , λ 3 , λ4 Ratio

[0052] Furthermore, all sub - project models and overall model parameters are optimized in a multi - round iterative or end - to - end manner.

[0053] The beneficial effects of the present invention are as follows:

[0054] (1) Multi - dimensional information fidelity: Fine - grained prediction of different sub - projects is achieved at the first layer, retaining the details and differences of various symptoms, providing a basis for subsequent personalized intervention.

[0055] (2) Stronger interpretability: With the help of modular control weighting, the importance of each sub - project in the depression network can be intuitively presented, and a clear and easy - to - understand weight distribution is output.

[0056] (3) Improvement of prediction accuracy and consistency: The bidirectional consistency constraint enhances both the overall prediction and the accuracy at the sub - project level.

[0057] The additional aspects and advantages of the present invention will be partly given in the following description, partly will become obvious from the following description, or will be understood through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 is the implementation schematic diagram of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0059] The present invention will be described in detail below with reference to the accompanying drawings.

[0060] As Figure 1 , a multi - dimensional mental health assessment method based on double - layer learning and module control, includes the following three key modules: (a) Module Control Weighting, (b) Dual - Layer Learning framework, and (c) Bidirectional Consistency. The complete process of multi - dimensional depression detection of the scale is realized through the cooperation of the three modules. The following is a specific introduction:

[0061] 1. Module Control Weighting

[0062] 1.1 Construction of symptom - related network

[0063] In this step, all sub - projects are regarded as nodes in the network. Assuming that the scale has J sub - projects (such as PHQ - 9 corresponding to 9 items), in order to obtain their association degree in the overall depression structure, a correlation matrix needs to be constructed first where

[0064] R ij =corr(y i ,y j ),

[0065] y i ,y j is the true score or measurement value of sub-items i and j, and corr(·,·) usually uses Pearson or Spearman correlation. Then, sparse inverse covariance estimation methods such as Graphical Lasso can be used to explore more direct interaction relationships and eliminate some redundant edges.

[0066] 1.2 Minimum Dominating Set (MDS) and Control Frequency

[0067] After constructing the sub-project association network G = (V, E), in order to quantitatively measure the sub-project's "control or coverage of other nodes in the network", the concept of Minimum Dominating Set (MDS) can be introduced. Definition: is a dominating set when any node in the network either belongs to D or is connected to a node in D by an edge. Finding MDS means finding a set that makes D as small as possible. Since this problem is NP-hard on general graphs, a random greedy search algorithm can be used for multiple iterations:

[0068]

[0069] This process is repeated for N rounds of random iterations, and an approximate MDS is obtained in each round. r ,r=1,…,N, where r is round, and each round will have a minimum dominating set MDS; for each node k, count how many rounds it appears in D, and record the number of appearances as n k , define the control frequency of the node:

[0070]

[0071] CF k The higher it is, the more often node k appears in the minimum dominating set, or has strong connections with more nodes, and its core position is more significant.

[0072] 1.3 Sub-project weight generation

[0073] The control frequency CF k After normalization to the interval [0,1], it can be regarded as the importance weight w of node k k .When summing up the predictions later, we will highlight or weaken the contribution of the sub-items in the total score, thereby achieving explainability weighting.

[0074] 2. Dual-Layer Learning Framework

[0075] 2.1 The First Layer: Independent Prediction of Sub-items

[0076] In traditional methods, if we only care about a total score, we may directly map from physiological signal features to that score. However, the present invention emphasizes "fine-grained sub-item analysis", that is: for each sub-item y j (j = 1, …, J), the optimal one is selected respectively from a group of candidate models to predict the score of this sub-item. Let the feature vector be X i represent the comprehensive features of the i-th subject, including:

[0077]

[0078] where may include feature data of different modalities, such as human behavior data, imaging data, and physiological electrical signal data, etc. For each sub-item, we let:

[0079]

[0080] Here, CV_score represents the score quantifying the model effect in the cross-validation process (such as R 2 , F1-score or other regression metrics). The finally selected optimal model will take X i as input and predict the estimated score of the sub-item. Doing so can adapt to the situation where different sub-items have inconsistent sensitivities to different features.

[0081] 2.2 The Second Layer: Aggregation of Total Score

[0082] After obtaining the predicted values of each sub-item in the first layer, to align with the common form of the total score of the scale, we calculate the weighted sum to obtain the overall depression prediction

[0083]

[0084] where w j is the weight from control frequency normalization, reflecting the relative importance of sub-item j in the symptom network.

[0085] 3. Bidirectional Consistency

[0086] 3.1 Reasons for Constraint

[0087] If only the predictions at the sub - item level and the total - score level are trained separately, mismatches may occur: the sum of sub - item predictions conflicts with the separately trained "total - score prediction". To eliminate such inconsistencies, we introduce a bidirectional consistency constraint during the training and optimization process, coupling the sub - item and total - score predictions under the same objective function.

[0088] 3.2 Comprehensive Loss Function

[0089] Let \(y_{ij}\) j,i be the true score of the \(i\) - th subject on the \(j\) - th sub - item, and \(y_i\) be the true total depression score. The predicted value of each sub - item \(\hat{y}_{ij}\) and the weighted sum \(\hat{y}_i=\sum_{j = 1}^{m}w_j\hat{y}_{ij}\) For all samples \(i = 1,\ldots,N\), define the comprehensive loss:

[0090]

[0091] where:

[0092] constrains the prediction accuracy of each sub - item, making the sub - item as close as possible to the true value;

[0093] takes into account the prediction error of the overall depression score;

[0094] forces the overall weighted sum \(\hat{y}_i=\sum_{j = 1}^{m}w_j\hat{y}_{ij}\) and \(\hat{y}_i\) to be consistent, thus achieving the coupling of sub - items and the total score;

[0095] introduces a correlation coefficient constraint to encourage the predicted sequence of each sub - item to be consistent with the true sequence in terms of linear correlation, minimizing random noise as much as possible.

[0096] During specific training, \(\lambda_1\), \(\lambda_2\), \(\lambda_3\), \(\lambda_4\) 1 , \(\lambda_2\) 2 , \(\lambda_3\) 3 , \(\lambda_4\) 4 can be adjusted according to the data scale, the number of sub - items, and the priority attention goal. If more emphasis is placed on overall accuracy, \(\lambda_1\) 2 can be appropriately increased; if more emphasis is placed on the fitting effect of each sub - item, \(\lambda_2\), \(\lambda_3\) 1 , \(\lambda_4\) 4 can be adjusted larger.

[0097] 3.3 Optimization and Convergence

[0098] In the specific implementation, multi - round iteration or an end - to - end approach can be used to optimize the parameters of all sub - item models and the overall model. For example:

[0099] Repeat until convergence:

[0100] After several rounds, the optimal solution where the sub - project and the overall prediction tend to be unified.

[0101] The present invention will be described in detail below in conjunction with specific embodiments.

[0102] Embodiment 1 (Audio - video behavior data):

[0103] 1. Data collection and pre - processing

[0104] We first conduct a depression screening on a certain population. Each participant is required to fill out a psychological scale with 9 sub - projects (in this example, PHQ - 9, which can also be extended to other scales), and read a neutral article in front of the device for about 1 minute (audio and video are recorded synchronously, frame rate 30FPS, audio sampling rate 44.1kHz). The main features of the collected audio - video include:

[0105] x (audio) → (Mean pitch, pitch variance, fundamental frequency jitter, timbre MFCC, multi - band energy distribution... )

[0106] x (video) → (Facial expression intensity, blink frequency, eye movement range, head pose, facial key points... )

[0107] After recording, noise reduction and endpoint detection are performed on the audio part; face detection and expression recognition are performed on the video part, and the temporal mean or statistic is extracted as the final numerical feature. After excluding some samples with missing data or severe noise, the remaining valid data is used for subsequent modeling, where y j,i represents the true score of sample i on sub - project j of the scale (taking values such as 0 - 3, 0 - 4, etc.), and y d,i is its total score.

[0108] 2. Construct a symptom network and calculate the module control weight

[0109] The 9 - item scores of all samples are sorted into a matrix, and the correlation between sub - projects is statistically analyzed. Then, a sparse association structure is obtained through Graphical Lasso to form a 9 - node symptom network. Next, the random greedy algorithm is executed multiple times to search for the minimum dominating set. Each iteration selects the smallest subset that covers all nodes of the network, and records the number of occurrences n k of each node in different search results. Finally, the control frequency CF k = n k / N, and it is normalized to w k . For example, some core symptom nodes (such as "anhedonia", "depressed mood") appear multiple times in the MDS, and the weight can reach about 0.8, while "appetite change" may be only about 0.2.

[0110] 3. Design of a two - layer learning model for sub - items and total scores

[0111] (1) First layer: Independent prediction of sub - items

[0112] For each sub - item \(y\) j , perform cross - validation in the candidate model set (such as LR, RF, XGBoost, etc.), and select the model with the best \(R\) 2 or mean squared error as \(f\) model \((y\) j ). Given the input of audio - video fusion features \(X\) i , obtain For example:

[0113]

[0114] Processing each sub - item in this way enables the model to exhibit the optimal feature mapping ability for different dimensions of depressive symptoms.

[0115] (2) Second layer: Weighted aggregation and two - way consistency optimization

[0116] Based on the aforementioned \(w\) j , first generate a preliminary total - score estimate using simple weighted sum:

[0117]

[0118] To fully correct this process and align it with the true \(y\) d,i , we set a two - way consistency loss during training, comprehensively considering the fitting accuracy and correlation between sub - items and the total score. Apply the comprehensive loss to this system, including:

[0119]

[0120] where can also be regarded as a differentiable target output (if using a neural network or an end - to - end back - propagation structure), or alternately optimized in a "fix - update" manner in the training loop. This process ultimately approaches a solution that is coordinated and consistent at the sub - item level and the overall level.

[0121] 4. Specific operations of two - way consistency optimization during training

[0122] In practice, the following iterative process can be adopted:

[0123]

[0124] Iterate continuously until the loss decreases slowly and the gap between the sub-items and the total score meets the set threshold. This can ensure that the final output can accurately depict each sub-item, and also make the sum of the predictions of these sub-items (under the weighted constraint) tend to be consistent with the overall depression prediction.

[0125] 5. Comparison of Implementation Effects and Result Display

[0126] On the test set, the method of the present invention is compared with the following baselines:

[0127] Baseline A: Only perform a single total score regression (audio-visual → y d ), without considering the details of the sub-items;

[0128] Baseline B: Multi-task training for all sub-items and the total score, but without the "bidirectional consistency" constraint, simply sharing features;

[0129] Baseline C: Train the sub-items separately and then average and combine them, without network weights (w j = 1 / J).

[0130] The experimental results show that:

[0131] (1) The present invention has significantly improved compared with the baselines in terms of indicators such as the R 2 and mean square error of the total score prediction;

[0132] (2) The average prediction error of each sub-item is significantly lower than that of the baselines, indicating that the first-layer independent training + optimal model selection plays a positive role;

[0133] (3) Observing the bidirectional consistency can improve the correlation between the combined and the true total score y d , and at the same time, the weighted sum of the sub-item prediction values is no longer contradictory to the separate overall prediction;

[0134] (4) By checking w j , it can be intuitively seen that some entries have a greater impact on the total depression score, which also coincides with clinical experience; entries with too small weights mean that their correlation in this population is relatively weak.

[0135] The specific experimental results are shown in the following table:

[0136] Method MAE RMSE <![CDATA[R 2 <!-- 8 -->]]> SVM 1.95 3.02 0.0523 Random Forest 1.95 3.37 -0.1811 Ours 1.93 2.83 0.0003

[0137] Example Two (Brain Imaging Data):

[0138] 1. The research team recruited a number of subjects. In addition to filling out a certain depression scale, they also underwent brain imaging scans (resting-state fMRI and event-related EEG recordings) under a specific experimental paradigm to obtain inter-regional functional connectivity, EEG waveform features, etc. Finally, a multi-dimensional neuroimaging feature vector X was formed for each subject i。

[0139] 2. Similarly, consider the sub-item scores of the scale as nodes, and use methods such as correlation and Graphical Lasso to estimate the network structure. Then, obtain the control frequency of each sub-item through repeated searches of the minimum dominating set, and further obtain the sub-item weight w i 。In this scenario, some items (such as those related to emotion regulation) may obtain larger weights by presenting higher aggregation through imaging features.

[0140] 3. In the first layer, for each sub-item y j independently train the best prediction model starting from brain imaging features (random forest, gradient boosting tree, or even deep networks for imaging features such as CNN / 3D-CNN can be used); in the second layer, after obtaining in the form of weighted aggregation, make the sum of sub-item scores and the overall score maintain dynamic coordination through bidirectional consistency loss. The measured results show that this framework can effectively highlight the connection patterns or brain region signals in brain imaging that are most sensitive to specific sub-items, improve the consistency between the total score and each sub-item, and have both interpretability and refinement.

[0141] The above shows and describes the basic principles, main features, and advantages of the present invention. Those of ordinary skill in the art should understand that the above embodiments do not limit the protection scope of the present invention in any form. Any technical solutions obtained by means of equivalent replacement and the like fall within the protection scope of the present invention. The parts not involved in the present invention are the same as the prior art or can be implemented by the prior art.

Claims

1. A multi-dimensional mental health assessment method based on double-layer learning and module control, characterized in that: The multi-dimensional depression detection of the scale is achieved through module control weighting, a two-layer learning framework, and bidirectional consistency constraints, as follows: Module control weighting: Based on the minimum dominating set and control frequency, it explicitly distinguishes the status of different sub-items of the scale in the depression network and realizes differentiated weight assignment; Two-layer learning framework: First, the model predicts the sub-items one by one, and then weights and combines them to get the total score, which not only retains the multi-dimensional symptom information, but also connects the final overall judgment; Bidirectional consistency constraint: Add sub-item-total score coupling terms to the loss function to ensure that local predictions and overall predictions cooperate with each other, thereby enhancing interpretability and practical value.

2. According to claim 1, a multi-dimensional mental health assessment method based on double-layer learning and module control is characterized in that: The module control weighting includes the construction of symptom-related network, the generation of minimum dominating set and control frequency and sub-item weights, specifically: (1) Construction of symptom-related network All sub-items are regarded as nodes in the network; assuming that the scale has J sub-items, in order to obtain the degree of association between the sub-items in the overall depression structure, the correlation matrix is ​​first constructed. in: R ij =corr(and i ,and j ), y i ,y j is the true score or measurement value of sub-item i, j, and corr(·,·) uses Pearson or Spearman correlation; then, the sparse inverse covariance estimation method is used to explore more direct interaction relationships and eliminate some redundant edges; (2) Minimum dominating set and control frequency The minimum dominating set is introduced to quantitatively measure the control or coverage of a sub-project on other nodes in the network. The definition is: is a dominating set, and finding the minimum dominating set is to find the set that makes D as small as possible, specifically using a random greedy search algorithm for multiple iterations; Perform N rounds of randomization iterations, and obtain an approximate MDS in each round. r , r = 1, ..., N, for each node k, count how many rounds it appears in D, and record the number of appearances as n k , define the control frequency of the node: CF k The higher it is, the more often node k appears in the minimum dominating set, or has strong connections with more nodes, and its core position is more significant; (3) Sub-project weight generation The control frequency CF k After normalization to the interval [0,1], it can be regarded as the importance weight w of node k k .When summing up the predictions later, the contribution of the sub-items in the total score will be highlighted or weakened accordingly, thus achieving explainability weighting.

3. According to claim 1, a multi-dimensional mental health assessment method based on double-layer learning and module control is characterized in that: The two-layer learning framework includes independent prediction of sub-items in the first layer and total score aggregation in the second layer.

4. A multi-dimensional mental health assessment method based on double-layer learning and module control according to claim 3, characterized in that: In the sub-item independent prediction, for each sub-item y j (j=1,…,J), respectively in a set of candidate models Select the best one among them to predict the score of the sub-item; set up The eigenvector is X i Represents the comprehensive characteristics of the i-th subject, including: in Contains feature data of different modes. For each sub-item, let: Among them, CV_score represents the score of the quantitative model effect during the cross-validation process; the optimal model finally selected will be input into X i Predict the estimated scores of the sub-items 5. A multi-dimensional mental health assessment method based on double-layer learning and module control according to claim 4, characterized in that: In the total score aggregation, in order to align with the commonly used scale total score format, the weighted sum is calculated to obtain the overall depression prediction where w j That is, it comes from the normalized weight of the control frequency, reflecting the relative importance of sub-item j in the symptom network.

6. According to claim 1, a multi-dimensional mental health assessment method based on double-layer learning and module control is characterized in that: In the bidirectional consistency constraint, a bidirectional consistency constraint is introduced during the training and optimization process to couple the prediction of the sub-items and the total score under the same objective function.

7. A multi-dimensional mental health assessment method based on double-layer learning and module control according to claim 6, characterized in that: Remember j,i is the actual score of the i-th subject on sub-item j, is the actual total depression score; the predicted value of each sub-item And the weighted summation For all samples i=1,…,N, define the comprehensive loss: in: Constrain the prediction accuracy of each sub-item to make the sub-item as close to the true value as possible; Taking into account the prediction error of the overall depression score; Force overall weighted sum and Keep it consistent so that the sub-items are coupled with the overall score; Introduce correlation coefficient constraints to encourage each sub-project's predicted sequence to be consistent with the actual sequence at the linear correlation level, and minimize random noise; During specific training, the ratios of λ1, λ2, λ3, and λ4 are adjusted according to the data scale, the number of sub-projects, and the priority targets.

8. A multi-dimensional mental health assessment method based on double-layer learning and module control according to claim 7, characterized in that: Use multiple rounds of iterations or an end-to-end approach to optimize all sub-project models and the overall model parameters.

Images