Method and device for diagnosing depression and anxiety risk of high mobility personnel in judicial field

Abstract

The present application relates to the technical field of judicial informatization and psychological health assessment, in particular to a method and device for diagnosing depression and anxiety risk of high mobility personnel in the judicial field, which comprises: obtaining an evaluation data set of the personnel to be evaluated; inputting the evaluation data set into a pre-trained fusion analysis model to output the quantitative scores of depression and anxiety of the personnel to be evaluated; the fusion analysis model comprises a feature fusion layer and a multi-task risk assessment layer; the feature fusion layer is used to convert the evaluation data set into a unified feature representation to obtain a standardized data set; the multi-task risk assessment layer comprises a plurality of learners, which learn the standardized data set from different angles and synthesize the quantitative scores through a weighted voting method. The present application adopts an analysis method combining large models, deep learning and machine learning, which can mine deep correlations and non-linear complex features between multi-source data, significantly improving the accuracy and intelligence of risk identification.

Description

Technical Field

[0001] This invention relates to the fields of judicial informatization and mental health assessment technology, specifically to a method and device for diagnosing the risk of depression and anxiety among highly mobile personnel in the judicial field. Background Technology

[0002] As highly mobile and special individuals, criminal suspects in detention centers and those under community corrections are susceptible to changes in environment and status, leading to prominent psychological problems such as depression and anxiety. These issues directly impact prison security, judicial fairness, correctional effectiveness, and social reintegration. Systematic and professional diagnosis of depression and anxiety risks in these individuals is both necessary and urgent. For example, it ensures prison security, as excessive anxiety can disrupt prison order or increase the risk of recidivism; it facilitates effective correction and resocialization, reducing the risks associated with emotional problems; and it reflects humanitarian care and professional correctional requirements, helping to protect the right to mental health and achieve scientific case allocation and resource matching.

[0003] Currently, there are no patents directly addressing the diagnosis of depression and anxiety risks among special individuals in the judicial field. Existing indirectly related patents include: a big data-based method for classifying and managing community correction subjects and providing personalized correction (Publication No. CN116167001A), an intelligent adjustment system for community correction (Publication No. CN111563694A), and a smart correction system and terminal based on psychological assessment and analysis of community correction subjects (Publication No. CN117079772A). These technologies primarily focus on classification management, adjustment, or general psychological assessment, and have not yet achieved a systematic and professional diagnosis of depression and anxiety risks among highly mobile special individuals, making it difficult to effectively support the practical needs for precise intervention. Summary of the Invention

[0004] Based on the above technical background, this invention abandons the linear weighting method of traditional psychological scales and adopts an analysis method that combines large models, deep learning and machine learning to deeply mine and fuse multi-source data, providing a method and device for diagnosing depression and anxiety risk among highly mobile personnel in the judicial field, so as to improve the accuracy and robustness of diagnostic results.

[0005] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0006] In a first aspect, the present invention provides a method for diagnosing the risk of depression and anxiety among highly mobile personnel in the judicial field, comprising the following steps: Obtain the assessment dataset of the person to be assessed, wherein the assessment dataset includes at least one of demographic sociological data, psychological scale data, and daily check-in data; The assessment dataset is input into a pre-trained fusion analysis model, which outputs quantitative scores of depression and anxiety for the person being assessed. The fusion analysis model includes a feature fusion layer and a multi-task risk assessment layer; The feature fusion layer is used to convert the evaluation dataset into a unified feature representation to obtain a standardized dataset. The multi-task risk assessment layer includes several learners that learn from the standardized dataset from different perspectives and synthesize quantitative scores through a weighted voting method.

[0007] Furthermore, the steps performed by the feature fusion layer include: Robust standardization is performed on the numerical features in the evaluation dataset to obtain standardized numerical features; One-hot encoding is performed on the demographic category variables in the assessment dataset to obtain the category variable codes; The standardized numerical features and categorical variable codes are concatenated to obtain a standardized feature vector, which includes a demographic sociology feature vector, a psychological scale data vector, and a multimodal feature vector.

[0008] Furthermore, the learner is a weighted voting learner that integrates a multi-task neural network model, a gradient boosting decision tree, and logistic regression.

[0009] Furthermore, the model of the multi-task neural network includes: Shared hidden layer: consists of multiple fully connected layers, followed by batch normalization layers, LeakyReLU activation function layers, and Dropout layers, used to learn the common latent representations of depression and anxiety; Task-specific branch layer: includes a first branch subnetwork and a second branch subnetwork, used to output risk diagnosis results based on the common potential representation. The first branch subnetwork is used to output depression risk diagnosis results, and the second branch subnetwork is used to output anxiety risk diagnosis results. Each of the two branch subnetworks has independent network weight parameters. Output layer: Connects the first branch sub-network and the second branch sub-network respectively. Each branch sub-network is connected to a fully connected layer followed by a Sigmoid activation function to output the corresponding risk index and prediction confidence.

[0010] Furthermore, the loss function used to train the model of the multi-task neural network is: ,in, Loss due to depression prediction task To predict the loss of anxiety in the task, To prevent overfitting, an L2 weight decay term is used. For the difference regularization term used to optimize the intrinsic structure of multi-task learning, This is a task-specific regularization intensity hyperparameter.

[0011] Furthermore, the multi-task risk assessment layer synthesizes a quantitative score through a weighted voting method, specifically including the following steps: Calculate the current sample confidence weight for each learner; multiply the risk score output by each learner by the current sample confidence weight and then sum them to obtain the quantized score.

[0012] Furthermore, it also includes discrepancy arbitration processing on the quantified scores to verify the reliability of the model, specifically including: The standard deviation distribution of risk scores is obtained within a preset period, and the arbitration threshold is adjusted according to the changes in the standard deviation distribution of risk scores. The arbitration threshold is used to ensure that the disagreement arbitration standard is always consistent with the predictive characteristics of the current service population. The method for obtaining the standard deviation of risk scores includes: inputting the same standardized dataset into multiple learners respectively, obtaining the risk scores output by each learner respectively, and calculating the standard deviation of the output risk scores.

[0013] Furthermore, the evaluation dataset is analyzed to obtain a data quality score, which is used to guide data collection. This analysis includes the following steps: The evaluation dataset was subjected to response pattern analysis, reaction time analysis, and extreme response tendency analysis to obtain response pattern index, reaction time index, and extreme response tendency index, respectively. The overall quality score is obtained by multiplying the response pattern index, reaction time index, and extreme response tendency index by their respective weights and then summing them.

[0014] Furthermore, the weights of the indicators are automatically updated, and the update steps include: Obtain the initial training set for the indicator weights; The initial training set is updated in the first round based on a reinforcement learning algorithm; A new round of updates will be performed using the unsupervised distribution alignment method if any of the following conditions are met: The amount of expert feedback data has reached a new threshold; The data distribution in the dataset has shifted.

[0015] Secondly, the present invention also provides a diagnostic device for depression and anxiety risk of highly mobile personnel in the judicial field, including a data acquisition module and a data processing module; The data acquisition module acquires the assessment dataset of the person to be assessed, which includes at least one of demographic sociology data, psychological scale data, and daily check-in data. The data processing module inputs the assessment dataset into a pre-trained fusion analysis model and outputs the quantitative scores of depression and anxiety of the person to be assessed. The fusion analysis model includes a feature fusion layer and a multi-task risk assessment layer; The feature fusion layer is used to convert the evaluation dataset into a unified feature representation to obtain a standardized dataset. The multi-task risk assessment layer includes several learners that learn from the standardized dataset from different perspectives and synthesize quantitative scores through a weighted voting method.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention adopts an analysis method that combines large models, deep learning and machine learning, replacing the linear weighting mode of traditional psychological scales. It can uncover deep correlations and complex nonlinear features between multi-source data, realize continuous optimization and adaptive learning of diagnostic models, and significantly improve the accuracy and intelligence of risk identification. Attached Figure Description

[0017] Figure 1 This is a flowchart of a diagnostic method for depression and anxiety risk among highly mobile personnel in the judicial field, as described in Example 1. Figure 2 This is a schematic diagram of a diagnostic device for depression and anxiety risk among highly mobile personnel in the judicial field, as shown in Example 3. Figure 3 This refers to the overall architecture of the device in Example 3; Figure 4 This is a schematic diagram of the data flow of the data acquisition and aggregation module in Example 3; Figure 5 This is a structural diagram of the scale data quality verification module in Example 3; Figure 6 This is a three-level architecture diagram of the data analysis module in Example 3; Figure 7 This is a portion of the code for implementing the multi-task neural network in Example 4; Figure 8 This is a portion of the code used in Example 4 to calculate the confidence level of a multi-task neural network. Figure 9 This is a portion of the code for implementing the gradient boosting decision tree and confidence level in Example 4; Figure 10 This is a portion of the code implementing logistic regression and confidence level in Example 4; Figure 11 This is a portion of the code for the multi-model integrated multi-task risk prediction engine in Example 4. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to specific embodiments. However, this should not be construed as limiting the scope of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0019] In this invention, criminal suspects and community correction personnel are collectively referred to as special personnel. Special personnel have higher mobility in terms of physical environment and social identity, and their psychological state is more easily affected by changes in environment and identity. Therefore, they are more likely to develop psychological problems such as depression and anxiety, which may bring new risks to the special personnel and their environment. The stability and predictability of the psychological state of highly mobile individuals directly affect the effectiveness of supervision, safety correction, and social reintegration. Therefore, systematically and professionally diagnosing the risk of depression and anxiety in such individuals is of paramount necessity and urgency. Against this backdrop, this invention provides a method for diagnosing the risk of depression and anxiety in highly mobile individuals in the judicial field, comprising the following steps: obtaining an assessment dataset of the individual to be assessed, the assessment dataset including at least one of demographic sociological data, psychological scale data, and daily check-in data; inputting the assessment dataset into a pre-trained fusion analysis model, outputting quantitative scores of depression and anxiety for the individual to be assessed; the fusion analysis model includes a feature fusion layer and a multi-task risk assessment layer; the feature fusion layer is used to convert the assessment dataset into a unified feature representation to obtain a standardized dataset; the multi-task risk assessment layer includes several learners, which learn the standardized dataset from different perspectives and synthesize quantitative scores through a weighted voting method. The method of this invention, firstly, constructs multi-source heterogeneous information based on demographic sociological data, psychological scale data, daily check-in data, etc., and uses this multi-source heterogeneous information to diagnose the risk of depression and anxiety among highly mobile special personnel in the judicial field, breaking through the current business paradigm that mainly uses psychological scale data; secondly, it employs large-scale models, deep learning, and machine learning methods to analyze the data and obtain diagnostic results, surpassing the linear weighting method of psychological scales, making the analysis process more efficient. Furthermore, this invention also provides a diagnostic device for the risk of depression and anxiety among highly mobile personnel in the judicial field, including a data acquisition module and a data processing module; the data acquisition module acquires the assessment dataset of the person to be assessed, the assessment dataset including at least one of demographic sociological data, psychological scale data, and daily check-in data; the data processing module inputs the assessment dataset into a pre-trained fusion analysis model and outputs the quantitative scores of depression and anxiety of the person to be assessed, forming a diagnostic device that can be quickly deployed and adapted to multiple scenarios.

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0021] Example 1 A diagnostic method for depression and anxiety risk among highly mobile individuals in the judicial field, flowchart as follows: Figure 1 As shown, it includes the following steps: Obtain the assessment dataset of the person to be assessed, wherein the assessment dataset includes at least one of demographic sociological data, psychological scale data, and daily check-in data; The assessment dataset is input into a pre-trained fusion analysis model, which outputs quantitative scores of depression and anxiety for the person being assessed. The fusion analysis model includes a feature fusion layer and a multi-task risk assessment layer; The feature fusion layer is used to convert the evaluation dataset into a unified feature representation to obtain a standardized dataset. The multi-task risk assessment layer includes several learners that learn from the standardized dataset from different perspectives and synthesize quantitative scores through a weighted voting method.

[0022] The input data is X, and the risk index is Y. X and Y are correlated through a corresponding model (the model is a function that establishes a mapping relationship between X and Y). The prediction confidence score is obtained by comparing the difference between the predicted value and the actual value after the risk index is predicted; see Prediction Confidence Score Calculation for details. Risk factors are obtained by comparing the changes in the predicted risk values ​​under different combinations of input data. Risk factors refer to all input features that improve the model's predicted risk values; see Risk Factor Output for details. Outcome labels are used during the training of the fusion analysis model. In this technology, outcome labels are a batch of depression and anxiety samples that have been confirmed by other means. The threshold is determined based on data distribution and business needs and can be adjusted according to actual needs. Common threshold determination methods include: defining values ​​that deviate from the mean by more than 3 standard deviations as risk outliers; establishing a norm suitable for a specific population; or manually defining a threshold based on actual business needs. The risk index (also called a quantification score) is the model's predicted output.

[0023] Furthermore, specific methods for obtaining the assessment dataset of the personnel to be assessed include constructing a comprehensive assessment dataset by fusing structured information with multimodal behavioral data, providing high-quality input for subsequent in-depth analysis.

[0024] The content of demographic and sociological data includes, but is not limited to, the cause of action, detention / correctional stage, age, education level, and family support status of individuals. Demographic and sociological data from different sources undergoes a unified standardization transformation to form structured fields, generating structured demographic and sociological data. As a specific implementation, this unified standardization transformation can be achieved through a pre-established internal data dictionary.

[0025] The data content of psychological scales includes standardized self-rating scales for depression and anxiety (such as PHQ-9, GAD-7, SDS, SAS, etc.). Since the scales belong to the traditional diagnostic criteria, the acquired data can be converted into structured data.

[0026] The daily check-in data includes short facial videos, photos, and audio clips collected on a scheduled or on-demand basis. This is special data in this field, generated during the management of specific personnel. Therefore, privacy protection and a minimum necessary declaration are required: the collected personal information, such as short facial videos, photos, and audio clips, is used only for behavioral analysis. The preservation of the feature data obtained from the analysis is a reasonable use of personal information, and the consent of the individuals involved was obtained at the time of collection.

[0027] The daily check-in data comes from diverse sources, necessitating standardization. The raw multimodal data R (video stream V, audio stream A) is transformed into a low-dimensional, structured, and quantifiable psychological and behavioral feature vector F using a series of lightweight algorithms. The processing adheres to privacy-preserving design principles: F = G(R), where G is a local feature extraction function. After completion, R can be securely erased.

[0028] Furthermore, the assessment datasets of the individuals to be assessed are stored and managed using a database. The database uses a relational database (such as MySQL) to store structured data (demographic information, scale scores) and an object storage or time series database to store unstructured multimodal raw data or feature data.

[0029] Furthermore, the fusion analysis model is the core part of the method of this invention. This model is a multi-level, interpretable fusion analysis model that receives the assessment dataset of the person to be assessed and outputs a quantified risk score of the person to be assessed on the dimensions of depression and anxiety. Based on the risk score, the risk level can also be assessed and matched with the corresponding key diagnostic criteria.

[0030] The fusion analysis model includes a feature fusion layer and a multi-task risk assessment layer.

[0031] The feature fusion layer is used to eliminate the differences in the dimensions and distribution of features from different sources and merge them into a unified feature representation; the multi-task risk assessment layer contains multiple basic learners, each of which learns data from different perspectives, and finally synthesizes the risk scores of depression and anxiety through a weighted voting method.

[0032] Furthermore, the steps performed by the feature fusion layer include: Robust standardization is performed on the numerical features in the evaluation dataset to obtain standardized numerical features. Robust standardization is a data preprocessing method whose core idea is to scale the data using statistics (median and interquartile range) that are insensitive to outliers instead of the traditional mean and standard deviation. As a specific implementation, robust standardization includes robustly standardizing numerical features (such as age, total scale score, fundamental standard deviation, etc.), calculated using the following formula: ,in The median of characteristic X, Let X be the quartile of characteristic X. It is a numerical feature after robust standardization.

[0033] One-hot encoding is performed on the demographic category variables (such as case type and education level) in the assessment dataset to obtain the category variable codes. Demographic categories refer to different group types that divide people according to demographic characteristics. These characteristics typically include age, gender, income, education level, occupation, marital status, family size, and geographical location. Simply put, demographic categories are a method of classifying people using objective, measurable, and relatively stable socio-demographic attributes. Demographic categorical variables are variables used in statistical analysis or data modeling to represent the demographic category to which an individual belongs. These variables take discrete, mutually exclusive group labels and are usually categorical variables. Their characteristics are: 1. Values ​​are finite categories, not continuous values ​​(e.g., not "age in years," but "age group: youth / middle-aged / elderly"). 2. They are usually nominal or ordinal variables: Nominal variables: There is no inherent order between categories, such as gender (male / female) and marital status. Ordinal variables: Categories have a natural order, such as education level (high school / undergraduate / graduate) or income level (low / middle / high). 3. Used for grouping, comparison, or control analysis: In tasks such as regression, classification, and clustering, they are often converted into dummy variables or labeled codes. Demographic categorical variables are variables that record demographic characteristics in discrete category form. Most machine learning models require converting the unordered categories in demographic categorical variables into one-hot encoding.

[0034] The standardized numerical features and categorical variable codes are concatenated to obtain a standardized feature vector. This standardized feature vector includes a demographic sociology feature vector, a psychological scale data vector, and a multimodal feature vector. The multimodal feature vector is obtained by converting daily check-in data. All processed features are then concatenated to form the final feature vector. , where P is the demographic sociology feature vector, S is the psychological scale data vector, and B is the preprocessed multimodal feature vector.

[0035] Furthermore, the multi-task risk assessment layer contains multiple basic learners, each learning data from different perspectives, and finally synthesizes a risk index for depression and anxiety through a weighted voting method.

[0036] Furthermore, the learner is a weighted voting learner that integrates a multi-task neural network model, a gradient boosting decision tree, and logistic regression.

[0037] Furthermore, the model of the multi-task neural network includes: Shared hidden layers: Typically 2-3 layers, including multiple fully connected layers. Following the fully connected layers are batch normalization layers, LeakyReLU activation function layers, and Dropout layers, used to learn the common latent representations of depression and anxiety. Specifically, the LeakyReLU activation function layer is a leaky rectified linear unit activation function layer; "Leaky" means leaky, and "ReLU" means rectified linear unit. Dropout layers are regularization layers used in neural networks. During training, they randomly set the output of some neurons to zero with a certain probability, preventing them from participating in the current forward and backward propagation; during the testing phase, the full network is used (usually with scaled weights). The purpose is to prevent complex co-adaptation relationships between neurons, thereby reducing overfitting and improving the model's generalization ability. The LeakyReLU activation function is used to alleviate gradient vanishing. Each layer is followed by BatchNormalization to accelerate convergence and Dropout layers (e.g., dropout rate = 0.3) to prevent overfitting, learning the common latent representations of depression and anxiety. Where F is the fused feature vector of dimension d, W1 is the weight of F, b1 is the bias, LeakyReLU is the activation function, BN is the BatchNormalization function to accelerate convergence, and Dropout is the function to prevent overfitting.

[0038] Task-specific branch layers: These include a first branch sub-network and a second branch sub-network, used to output risk diagnosis results based on the shared latent representations. The first branch sub-network outputs depression risk diagnosis results, and the second branch sub-network outputs anxiety risk diagnosis results. Each branch sub-network has independent network weight parameters. Specifically, each task-specific branch layer has 1-2 layers, based on the output of the shared layer. The network is divided into two independent subnetworks, corresponding to anxiety and depression tasks respectively, with each branch having independent weights. .

[0039] Output layer: Connects the first branch sub-network and the second branch sub-network respectively. Each branch sub-network is connected to a fully connected layer followed by a sigmoid activation function to output the corresponding risk index and prediction confidence. Specifically, each branch is connected to a fully connected layer followed by a sigmoid function to output the risk index. Simultaneously, it outputs the prediction confidence score Conf_nn. The calculation steps for the prediction confidence score Conf_nn output by the output layer include: first calculating the information entropy of the risk index, and then calculating the prediction confidence score based on the information entropy of the risk index. The formula for calculating the information entropy of the risk index is: The formula for calculating the prediction confidence score Conf_nn is: ,in =2 log(2).

[0040] The multi-task neural network model uses a composite loss function during training. The composite loss function is: ,in, Loss due to depression prediction task To predict the loss of anxiety in the task, To prevent overfitting, an L2 weight decay term is used. For the difference regularization term used to optimize the intrinsic structure of multi-task learning, The L2 regularization hyperparameter is used to prevent model overfitting. The task-specific regularization strength hyperparameter quantifies the degree of importance attached to the structural objective of "encouraging the difference between task-specific weights and shared weights" during model optimization. and These are user-adjustable hyperparameters that represent the task weights for depression and anxiety tasks in different scenarios. It is the task weight of the depression prediction task. It is the task weight for the anxiety prediction task.

[0041] L2 weight decay is a regularization technique used in machine learning models. It is defined as follows: In addition to the original loss function of the model, half the sum of the squares of all weight parameters multiplied by a positive coefficient is added as a penalty term, i.e.: ; in, λ is the model's weight parameter, where λ > 0, and λ is the weight decay coefficient (hyperparameter). This term is also called the L2 regularization term or ridge regression penalty term. It forces the weights to tend towards smaller values ​​(but not strictly zero), thereby limiting model complexity, suppressing overfitting, and improving generalization ability. The role of the L2 weight decay term is that, during gradient descent updates, this term is equivalent to proportionally reducing (decreasing) the weights in each iteration, hence the name "weight decay".

[0042] Furthermore, and Using binary cross-entropy, ,in As a true label for depression, This is the depression risk index predicted by the model. Similarly; furthermore, given that the cost of underreporting (classifying high-risk cases as low-risk) in the judicial field far outweighs the cost of false alarms, this device introduces a judicial scenario weighting factor. This forms a weighted binary cross-entropy. ,in, For the actual outcomes of anxiety and depression (real label). The model predicts anxiety and depression, where log is the logarithmic function. Here, N is an adjustable hyperparameter for high-risk samples, and N is the number of training samples. It represents the number of the training sample; a high value can be set. (like =3). Parameter regularization loss: , where θ is an element in the weight matrix W, which is composed of adjustable hyperparameters for high-risk samples. This step is to sum the squares of all θ for high-risk samples.

[0043] Specifically, this invention employs a differential regularization loss. To enable multi-task networks to learn shared representations while retaining sufficient task specificity to distinguish between depression and anxiety, two related but distinct psychological constructs, this device introduces a task-specific regularization term into the loss function. This term aims to reduce the weight matrix of task-specific layers. Shares weight matrix with the last layer The module calculates the structural similarity between the two vectorized structures and then penalizes it. ,in This means flattening the matrix into a vector, and this loss increases as the directions of the two weight vectors converge. The task-specific regularization strength hyperparameter quantifies the degree of importance placed on the structural objective of "encouraging the difference between task-specific weights and shared weights" during model optimization. A smaller... This enables the model to learn highly shared general representations, and larger The forced model develops more discriminative dedicated representation pathways for depression and anxiety risks. This design allows the device to flexibly adapt to different requirements for risk identification precision in various correctional settings through parameter configuration. By incorporating these pathways into the total loss for minimization, task-specific parameters are effectively encouraged to maintain a certain degree of orthogonality or low correlation with shared parameters, thereby learning more discriminative task-specific features based on shared parameters.

[0044] Furthermore, the multi-task neural network model also outputs risk factors, calculating the contribution of features to the diagnostic results based on the Shapley Additive Explanations (SHAP) algorithm to identify key risk factors. Specifically: assuming the object classification model has M features and the predicted value of the object classification model is f(x), then the decomposition formula for the SHAP value is: ,in, This is the baseline value for the object classification model, usually the average predicted value of the training set; the prediction result of the object classification model is decomposed into the contribution of each feature. For a sample x, the model's predicted value f(x) is decomposed into the baseline value. and the SHAP values ​​of each feature .

[0045] The SHAP value is calculated considering all possible feature combinations, measuring the contribution of each feature in different combinations. For a feature i, the formula for calculating its SHAP value is as follows: Where M{i} represents all possible combinations of participants, here referring to all possible combinations of features, and S is the set of features. It is a set containing feature i The profit (i.e., the predicted value). Let represent the marginal contribution of feature i to set S, and M be the number of features in the object classification model.

[0046] As a specific implementation, the learner can also be a gradient boosting decision tree, preferably using LightGBM (Gradient Boosting Framework), which supports multi-output regression and is computationally efficient. It is trained directly using features F, anxiety outcome labels, and depression outcome labels (anxiety and depression outcome labels can be samples already diagnosed with anxiety and depression by a hospital). LightGBM constructs a series of trees, with each tree learning the residuals predicted by all previous trees; the output... (Depression risk predictor) and (Anxiety risk prediction value), and feature importance Imp_lgb and main decision path Rule_lgb; based on feature importance Imp_lgb, aggregate into confidence Conf_lgb; use LightGBM's predict(..., pred_contrib=True) function to obtain the contribution value of each feature to the prediction of this sample; take the absolute value of the contribution value to obtain the feature importance metric a, and calculate the norm of a. d is the characteristic number.

[0047] As a specific implementation, the learner can also be a logistic regression function, treating the risk score as a continuous value, using ridge regression, and applying L2 regularization to prevent the coefficient from being too large and improve stability. ,in Output (Depression risk predictor) and (Anxiety Risk Predictor) and Characteristic Coefficient The specific steps for calculating the confidence level Conf_lr include: After training the model, the coefficient vector is obtained. And its covariance matrix Cov(w); For new samples ,predict ; Calculate the variance of the predictions: ; Construct a 95% confidence interval: ; Map the upper and lower bounds of the interval to probability intervals using the sigmoid function. ;but The calculation formula is: ; Preferably, the outputs of the three models are combined using a weighted voting method to synthesize risk scores for depression and anxiety, specifically including the following steps: First, calculate the confidence weights of the current samples for each model. , ,in To standardize the confidence scores of the model, Conf_nn, Conf_lgb, and Conf_lr are normalized to the [0,1] interval. For temperature coefficient, The larger the value, the higher the weight of the high-confidence model. At the same time, prior performance weights are introduced, and the historical performance (AUC) of the model on the validation set is used as one of the benchmarks for weight calculation. The normalized AUC is used as the weight vector of confidence.

[0048] Secondly, calculate the weighted average risk index: , ;in, It is a predictor of depression risk. These are the predicted values ​​for anxiety risk; the subscripts represent the corresponding models. These are the weights of each model's risk prediction value, with the subscript representing the corresponding model. Finally, a disagreement arbitration method is used to obtain the risk scores for depression and anxiety. Among these, the arbitration threshold... It is not fixed, but rather possesses online adaptive capabilities. The system continuously monitors the standard deviation of all recent (e.g., the past month) forecasts. The system uses a rolling distribution, recalculating thresholds for recent data at fixed intervals (e.g., quarterly) based on percentiles. If the relative change between the new and old thresholds exceeds a preset range (e.g., 10%), it is automatically updated after confirmation by the system administrator. This method ensures that the disagreement arbitration standard remains consistent with the predicted characteristics of the current service population. Specifically, First, record the standard deviation of the three model-predicted risk indices for each sample. Secondly, drawing The empirical cumulative distribution function (ECDF); finally, It can be set to the 70-80 quantile of ECDF. It can be set to the 95th-99th quantile of ECDF.

[0049] Introducing result smoothing for high-discrepancy scenarios to prevent extreme output: When the mean standard deviation of the model predictions is greater than... When this occurs, it indicates that the decision-making process for this sample is somewhat ambiguous. and Shrinking towards the neutral value of 0.5 helps reduce the risk of misjudgment, that is... , Similarly, we should enhance the robustness of its application in the judicial field.

[0050] Risk level mapping: Combining preset thresholds (e.g., low risk level is less than 0.3, medium risk level is greater than or equal to 0.3 and less than or equal to 0.7, and high risk level is greater than 0.7) with the higher of the two, the final risk level is determined (e.g., "high risk of depression, medium risk of anxiety"). Depression risk and anxiety risk are assessed separately, and the model will output predicted values ​​for depression risk and anxiety risk respectively. The thresholds are determined based on data distribution and business needs and can be adjusted according to actual needs. Common threshold classification methods include: classifying values ​​deviating from the mean by more than 3 standard deviations as risk outliers; establishing a norm suitable for a specific population; or manually defining thresholds based on actual business needs.

[0051] Example 2 Based on Example 1, Example 2 further provides a method for analyzing the evaluation dataset, obtaining a data quality score for the evaluation dataset, which is used to guide data collection, and specifically includes the following steps: The evaluation dataset was subjected to response pattern analysis, reaction time analysis, and extreme response tendency analysis to obtain response pattern index, reaction time index, and extreme response tendency index, respectively. The overall quality score is obtained by multiplying the response pattern index, reaction time index, and extreme response tendency index by their respective weights and then summing them.

[0052] Response pattern analysis (detecting random responses and fixed-pattern responses) includes: Proportion of consecutive identical options in a long sequence: This measures whether participants selected the same option consecutively without reading the questions. It is measured in the response sequence of a scale of length N. ,in For each option value, find the longest consecutive substring of identical options. ,in The length of the longest consecutive substring of identical options; usually A value >0.4 can be considered high risk.

[0053] Answer sequence autocorrelation / periodicity: Detects the existence of regular answer patterns such as "yes-no-yes-no" or "1-2-3-1-2-3". Calculate the first-order difference sequence of the answer sequence R. ; Calculate the entropy of the difference sequence. ,in It is the difference value In sequence Frequency of occurrence in; Calculate the expected entropy of the answer difference sequence of length N. ,in For a single difference value Expected entropy , The number of options for each question. For difference Values The probability, , For all possible difference values; Calculate the relative entropy ratio ,like Much less than 1 (e.g.) If the answer is less than 0.5, it indicates that there is a strong regularity or pattern in the answer.

[0054] Response time analysis (detecting hasty or excessively delayed responses) specifically includes: Abnormal response time ratio: Identify questions with excessively short or long response times. Predefine a reasonable answering time interval [tmin, tmax]; record the actual answering time ti for each question; ,when >0.3, the reliability of the data is questionable.

[0055] Reaction time consistency: Detects any abnormal fluctuations in the overall answering pace. Records the actual answering time ti for each question; ,in and The mean and standard deviation of the time taken to answer the questions.

[0056] Analysis of extreme response tendencies Extreme option ratio: Detects whether there is a tendency to choose the most extreme option (e.g., always choosing "1 - strongly disagree" or "5 - strongly agree"). .

[0057] The above multiple indicators are weighted and integrated to form a comprehensive quality score, which is then divided into levels. ,in For the value of the j-th indicator, For normalization function, As the indicator weight, For high-quality data, The data is of medium quality. The data is of low quality. The data quality will be displayed on the device's interactive interface, and personnel will be alerted if the data is of low quality.

[0058] right The update follows these steps: Initialization phase Adopting domain knowledge-based fixed weights ; Data accumulation stage The system operates normally and randomly samples a portion of the data (especially data with an overall quality score in the middle range) and submits it to domain experts for manual quality scoring to form the initial training set; First update Once the training set reaches a certain size (e.g., 500 examples), a reinforcement learning algorithm based on expert feedback is initiated. Reinforcement learning is a type of machine learning algorithm whose core principle is to allow an agent to learn the optimal policy through trial and error in its interaction with the environment—that is, which action to choose in each state to maximize the long-term accumulated reward signal (usually expressed as a discounted cumulative reward). Unlike supervised learning, reinforcement learning does not have a "standard answer" label; instead, it learns autonomously through delayed and sparse reward signals. Typical algorithm examples include Q-learning, SARSA, DQN, PPO, and A3C. Specifically: regarding the loss function... ,in For the expert's final decision on the m-th sample, gradient descent is used to minimize... , .

[0059] Continuous learning phase The system has entered a stable operating period, and the update trigger condition is any of the following: Experts reported that the amount of data had reached a new threshold (e.g., 500 new cases). The system detected a high-quality data pool. ( The distributions of the indicators in >0.8 show a significant drift compared to their distributions in the full data pool (e.g., statistically significant differences at the 99% confidence interval). Here, the distribution of each indicator includes the data distribution itself, such as the shift in the mean. Therefore, the unsupervised distribution alignment method should be followed. Update parameters, among which , Let j be the mean of the j-th indicator in the entire data pool. For the j-th indicator in The mean of the middle, To combine standard deviations. Unsupervised distribution alignment refers to the process of transforming or mapping the data distributions of the source domain and the target domain as closely as possible, without class labels (or with the target domain being completely unlabeled), so that a model trained in the source domain can be directly or after fine-tuning applied to the target domain. In short: unsupervised distribution alignment is about "bringing" the distributions of two domains closer together, using only feature-level distribution information, without target domain labels.

[0060] Weight smoothing and constraints To prevent overfitting or drastic fluctuations, the weights are exponentially smoothed compared to the previous generation after each update. and apply a non-negative ( ) to the weights. ) and the summation is 1 ( () constraints.

[0061] Example 3 Please refer to Figure 2 , Figure 2 This is a schematic diagram of a diagnostic device for depression and anxiety risk among highly mobile individuals in the judicial field, provided as an embodiment of this application.

[0062] The diagnostic device for depression and anxiety risk among highly mobile personnel in the judicial field (90) may include: The data acquisition module 91 is used to acquire the assessment dataset of the person to be assessed, which includes at least one of demographic sociological data, psychological scale data, and daily check-in data.

[0063] The data processing module 92 inputs the assessment dataset into a pre-trained fusion analysis model and outputs the quantitative scores of depression and anxiety of the person to be assessed. The fusion analysis model includes a feature fusion layer and a multi-task risk assessment layer; The feature fusion layer is used to convert the evaluation dataset into a unified feature representation to obtain a standardized dataset. The multi-task risk assessment layer includes several learners that learn from the standardized dataset from different perspectives and synthesize quantitative scores through a weighted voting method.

[0064] It should be understood that the various modules of the diagnostic device 90 for the risk of depression and anxiety of highly mobile judicial personnel provided in the above embodiments are only illustrated by the division of the functional modules in the above description when conducting natural disaster risk early warning. In actual applications, the above functions can be assigned to different functional modules as needed. That is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above.

[0065] The functional modules in the above embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of the embodiments of this application.

[0066] As a specific embodiment, the depression and anxiety risk diagnosis device 90 for highly mobile personnel in the judicial field includes a data acquisition and aggregation module, a scale data quality verification module, a data analysis module, and a diagnostic result output module, realizing a closed loop of "data acquisition - data governance - data analysis - diagnostic output". The overall architecture of the device is as follows: Figure 3 As shown.

[0067] The data acquisition and aggregation module is the front-end and foundation of this invention, responsible for compliantly and efficiently collecting raw data related to the risk of depression and anxiety among individuals from multi-source, heterogeneous judicial supervision scenarios. It constructs a comprehensive assessment dataset by fusing structured information and multimodal behavioral data, providing high-quality input for subsequent in-depth analysis. The data acquisition and aggregation module consists of three parts: a data interface unit, a local acquisition terminal, and a data aggregation center. A schematic diagram of the data flow of the data acquisition and aggregation module is shown below. Figure 4 As shown.

[0068] The specific data collection and aggregation schemes for each data source are as follows: Population sociology information (structured data): Data content includes, but is not limited to, the person's case details, detention / correctional stage, age, education level, and family support status; Data collection method: Interact with existing information systems of judicial institutions through predefined secure application programming interfaces (Secure APIs) or dedicated data middleware; One-way synchronization and data masking: A one-way data synchronization strategy is adopted, reading only necessary information from external systems to ensure no reverse writing and avoid interfering with core judicial processes. Sensitive fields are masked before transmission (e.g., using numbers instead of names). Data standardization: Establish an internal data dictionary to standardize and transform heterogeneous data formats from different systems into structured fields.

[0069] Psychological scale data (structured data) Data content: Standardized self-rating scales for depression and anxiety (such as PHQ-9, GAD-7, SDS, SAS, etc.); Data collection method: Local data collection: Personnel complete the questionnaire by using the touch screen interactive terminal provided with this device and following the guided interface; Mobile data collection: For personnel authorized to use mobile phones, the information is collected through a dedicated encrypted mobile app that is installed under authorization; Guidance and Compliance: Before the test begins, clearly inform the test purpose, data usage, and rights via voice or text to ensure that the test is conducted with informed consent; Encrypted Transmission and Aggregation: Data collected by the mobile app is transmitted to the device's server via HTTPS with end-to-end encryption, where it is centrally stored and managed by the data aggregation center. Daily check-in data (multimodal unstructured data) Data content: Short-duration facial videos, photos, and audio clips collected on a scheduled or on-demand basis; Data collection method: Local data collection: Data is collected at fixed locations in the business scenario using the high-definition camera and noise-canceling microphone array integrated into the device; Mobile data collection: By authorizing a mobile app, the phone's camera and microphone can be used to collect data at a specified time or when the correctional personnel report for duty. Privacy protection and minimum necessity: The data collection process clearly indicates that video / audio will only be used for behavioral analysis. Localized preliminary processing is employed (e.g., extracting facial motion units and speech spectrum features only on the device), and the original media files can be set to be automatically deleted at a set period, saving only the feature data obtained from the analysis. Environmental adaptability: The hardware design takes into account the complex environment of the monitoring site, such as lighting and background noise, to ensure the basic quality of data usability.

[0070] Preliminary localization processing methods for daily check-in data: Processing Overview: The goal of localization processing is to transform the raw multimodal data R (video stream V, audio stream A) into a low-dimensional, structured, and quantifiable psycho-behavioral feature vector F at the data source end through a series of lightweight algorithms. The processing follows privacy-preserving design principles, namely: F=G(R), where G is the local feature extraction function, and R can be securely erased after completion.

[0071] Facial video processing module: Based on quantitative analysis of temporal action units and motion dynamics. Facial action unit intensity temporal sequence extraction: A real-time AU intensity regression model based on a lightweight backbone network such as MobileNetV3 is adopted. The input is the aligned facial region image It (frame t), and the output is the intensity values ​​aut(i)∈[0,5] of N core AUs, where 0 indicates no occurrence and 5 indicates maximum intensity: [aut(1),aut(2),...,aut(N)]=AU-Net(It), i∈{1,2,...,N}; Temporal feature quantization: For the intensity temporal sequence {au1(i),au2(i),...,auT(i)} of each AU, the following structured features are calculated: average intensity and frequency of occurrence: , ,in The intensity threshold, For indicator functions, (condition) = 1 if condition = True, (condition) = 0 if condition = False; Intensity dynamics: Asymmetry index: .

[0072] Head movement and gaze dynamics characteristics: A general 3D facial landmark model is predefined: ; A lightweight face landmark detection model is used to detect the coordinates of the corresponding 2D image landmarks in the current input frame It. ; A linear solution is obtained using the EPNP algorithm, which is suitable for edge computing devices. Where si is a non-zero scaling factor, and K is the camera intrinsic parameter matrix. Let be the camera extrinsic parameter matrix to be determined; Based on matrix R, calculate the pitch angle Pitch(θ) = arcsin(-R31), the yaw angle Yaw(φ) = atan2(R32 / cosθ, R33 / cosθ), and the roll angle Roll(ψ) = atan2(R21 / cosθ, R11 / cosθ). Range of motion: ; Eye contact rate: ,in The angle between the line of sight and the camera axis. The threshold for the line-of-sight angle (e.g., 15°).

[0073] Speech Processing Module: Quantitative Analysis Based on Acoustic and Prosodic Models Fundamental frequency and energy profile analysis The fundamental frequency of each frame of speech is extracted using the YAAPT algorithm, which is based on autocorrelation. (Unit: Hz); Fundamental frequency range and dynamics: , ; Analysis of speech rate and pause patterns Rate of articulation: ,in For the estimated number of syllables, Total effective voice duration; Pause Index: ,in For the total length of all pauses, This represents the total length of the speech segment.

[0074] Generate unified multimodal feature vectors .

[0075] Data storage and management (device comes with its own server) Hardware foundation: This device integrates a high-performance embedded server or minicomputer with ample storage space and computing resources; Database design: Use relational databases (such as MySQL) to store structured data (demographic information, scale scores), and use object storage or time series databases to store unstructured multimodal raw data or feature data; Full disk encryption and access control: The server hard drive uses full disk encryption technology and establishes a strict role-based access control (RBAC) method; Audit logs: All data access, modification, and deletion operations are recorded in detailed audit logs; Offline and Backup: This device supports offline operation under network isolation and has the ability to perform regular data backup and recovery.

[0076] Scale data quality verification module This module sits between the data acquisition and aggregation module and the data analysis module, acting as a portal and filter for data quality. It receives raw questionnaire responses and applies a series of algorithms and statistical indicators to automatically score and categorize the quality of each questionnaire. The verification results are attached to the data as metadata tags for subsequent differential processing by the analysis module. The module structure is as follows: Figure 5 As shown.

[0077] Data analysis module The data analysis module is the decision-making center of this device. It receives quality-verified structured data (demographics, scale scores, multimodal behavioral characteristics) from upstream sources and outputs an individual's risk level, quantitative score, and key diagnostic criteria in the dimensions of depression and anxiety through a multi-level, interpretable fusion analysis model. The core design of this module lies in fusion, interpretation, and adaptation to legal scenarios. Module Architecture and Data Processing Flow: This module adopts a three-level architecture of "feature layer fusion → multi-task risk assessment → dynamic calibration," as follows... Figure 6 As shown.

[0078] The diagnostic results output module is the decision-making interface and action hub of this device. It receives the final risk index, level, arbitration results and multi-source evidence from the data analysis module. Through intelligent report generation, hierarchical early warning and visualization, and interfaces with external systems, it realizes the structured, operable and process-oriented output of risk diagnosis conclusions, forming a complete "diagnosis-early warning-intervention" business closed loop.

[0079] Report generation engine: This device automatically generates a diagnostic report that includes basic information about the subjects, core conclusions, key multimodal evidence and data quality, interpretation of risk factors, and dynamic trends under multiple diagnoses. Tiered early warning and visualization interface This device triggers different levels of alerts based on the risk level and pushes them to different user roles: Tiered early warning: Level 1 Warning (High Risk): If at least one of depression and anxiety is at high risk, or the data quality is low, the device will automatically send a pop-up window and SMS warning to the user. Level 2 warning (medium risk): No high risk and data quality is not low, depression and anxiety are at least medium risk, and the device generates a tracking task in the user's to-do list. Level 3 alert (low risk): Other situations that do not fall under Level 1 or Level 2 alerts. The device will notify the user that the subject has recently completed a diagnosis and can view the full report.

[0080] Visual portrait: Individual profile: Each participant is provided with a personal dashboard that integrates a risk index meter, multimodal feature radar chart, and historical trend chart; Group profiling: Provides a panoramic view of the subject group, showing the distribution of people at each risk level, risk heat map, and early warning and handling progress dashboard.

[0081] Intervention recommendations Based on an intervention knowledge base and rule mapping, diagnostic conclusions are mapped to specific, actionable intervention recommendations.

[0082] Judicial system interface This device provides an interface to the judicial system, ensuring that key fields of diagnostic results can be embedded into existing judicial workflows.

[0083] Example 4 The above technologies, especially the multi-task risk prediction engine based on multi-model integration, were validated using real data from the judicial field: 1. Data from over 38,000 individuals with special circumstances in the judicial field were extracted, and after standardization, hot processing, and feature splicing, a feature vector F containing 161 features (i.e., d=161) and two outcome labels (anxiety risk and depression risk) were formed. 2. The training and test sets were divided in an 8:2 ratio, resulting in a training set of 30,000+ and a test set of 7,000+. 3. Train the multi-task neural network (MT-NN), gradient boosting decision tree (LightGBM), and logistic regression (LR) respectively, and calculate the confidence scores; Figure 7 This is a portion of the code that implements a multi-task neural network. Figure 8 This is a portion of the code that implements the confidence calculation of a multi-task neural network. Figure 9 This is a portion of the code that implements gradient boosting decision trees and confidence levels. Figure 10 This is part of the code that implements logistic regression and confidence levels.

[0084] 4. Construct a multi-model integrated, multi-task risk prediction engine. Figure 11 This section showcases a portion of the code for a multi-model integrated, multi-task risk prediction engine.

[0085] 5. The prediction performance of the multi-model ensemble was compared with that of the three individual models mentioned above. The results are shown in Table 1: Table 1: Comparison of the performance of multi-model ensemble prediction and single-model independent prediction

[0086] As shown in Table 1, the multi-model integration exhibits the most robust overall performance with no obvious weaknesses and the best F1 score. This means that it achieves the optimal balance between "not underreporting high-risk individuals" and "not falsely reporting low-risk individuals," avoiding the risk that a single model may fail for certain individuals and enhancing the reliability of decision-making in the judicial field.

[0087] Generate a sample report for one of the samples, as shown below: Psychological Risk Diagnosis and Assessment Report Sample ID: 0 [Overall Assessment Conclusion] Final risk level: High risk Depression risk index: 0.291 Anxiety Risk Index: 0.787 Arbitration Status: MEDIUM_DISAGREEMENT Model divergence: 0.233 [Model Council Voting Details] Multi-task neural networks: - Depression prediction: 0.012 - Anxiety prediction: 0.865 - Confidence level: 0.669 - Voting weight: 0.408 LightGBM: - Depression predictor: 0.250 - Anxiety prediction: 0.487 - Confidence level: 0.000 - Voting weight: 0.260 Logistic Regression: - Depression predictor: 0.667 - Anxiety prediction: 0.928 - Confidence level: 0.365 - Voting weight: 0.332 Intervention Recommendations - Initiate the security monitoring protocol immediately - 24-hour key focus The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A diagnostic method for the risk of depression and anxiety among highly mobile personnel in the judicial field, characterized by: Includes the following steps: Obtain the assessment dataset of the person to be assessed, wherein the assessment dataset includes at least one of demographic sociological data, psychological scale data, and daily check-in data; The assessment dataset is input into a pre-trained fusion analysis model, which outputs quantitative scores of depression and anxiety for the person being assessed. The fusion analysis model includes a feature fusion layer and a multi-task risk assessment layer; The feature fusion layer is used to convert the evaluation dataset into a unified feature representation to obtain a standardized dataset. The multi-task risk assessment layer includes several learners that learn from the standardized dataset from different perspectives and synthesize quantitative scores through a weighted voting method.

2. The diagnostic method for depression and anxiety risk among highly mobile personnel in the judicial field as described in claim 1, characterized in that, The steps performed by the feature fusion layer include: Robust standardization is performed on the numerical features in the evaluation dataset to obtain standardized numerical features; One-hot encoding is performed on the demographic category variables in the assessment dataset to obtain the category variable codes; The standardized numerical features and categorical variable codes are concatenated to obtain a standardized feature vector, which includes a demographic sociology feature vector, a psychological scale data vector, and a multimodal feature vector.

3. The diagnostic method for depression and anxiety risk among highly mobile personnel in the judicial field as described in claim 1, characterized in that, The learner is a weighted voting learner that integrates a multi-task neural network model, a gradient boosting decision tree, and logistic regression.

4. The diagnostic method for depression and anxiety risk among highly mobile personnel in the judicial field as described in claim 3, characterized in that, The model of the multi-task neural network includes: Shared hidden layer: consists of multiple fully connected layers, followed by batch normalization layers, LeakyReLU activation function layers, and Dropout layers, used to learn the common latent representations of depression and anxiety; Task-specific branch layer: includes a first branch subnetwork and a second branch subnetwork, used to output risk diagnosis results based on the common potential representation. The first branch subnetwork is used to output depression risk diagnosis results, and the second branch subnetwork is used to output anxiety risk diagnosis results. Each of the two branch subnetworks has independent network weight parameters. Output layer: Connects the first branch sub-network and the second branch sub-network respectively. Each branch sub-network includes a fully connected layer followed by a Sigmoid activation function to output the corresponding risk index and prediction confidence.

5. The diagnostic method for depression and anxiety risk among highly mobile personnel in the judicial field as described in claim 4, characterized in that, The loss function used to train the model of the multi-task neural network is: ,in, Loss due to depression prediction task To predict the loss of anxiety in the task, To prevent overfitting, an L2 weight decay term is used. For the difference regularization term used to optimize the intrinsic structure of multi-task learning, This is a task-specific regularization intensity hyperparameter.

6. The diagnostic method for depression and anxiety risk among highly mobile personnel in the judicial field as described in claim 1, characterized in that, The multi-task risk assessment layer synthesizes quantitative scores through a weighted voting method, specifically including the following steps: Calculate the current sample confidence weight for each learner; multiply the risk score output by each learner by the current sample confidence weight and then sum them to obtain the quantized score.

7. The diagnostic method for depression and anxiety risk among highly mobile personnel in the judicial field as described in claim 6, characterized in that, It also includes a disagreement arbitration process for the quantified scores to verify the reliability of the model, specifically including: The standard deviation distribution of risk scores is obtained within a preset period, and the arbitration threshold is adjusted according to the changes in the standard deviation distribution of risk scores. The arbitration threshold is used to ensure that the disagreement arbitration standard is always consistent with the predictive characteristics of the current service population. The method for obtaining the standard deviation of risk scores includes: inputting the same standardized dataset into multiple learners respectively, obtaining the risk scores output by each learner respectively, and calculating the standard deviation of the output risk scores.

8. The diagnostic method for depression and anxiety risk among highly mobile personnel in the judicial field as described in any one of claims 1-7, characterized in that, The evaluation dataset is analyzed to obtain a data quality score, which is used to guide data collection. This analysis includes the following steps: The evaluation dataset was subjected to response pattern analysis, reaction time analysis, and extreme response tendency analysis to obtain response pattern index, reaction time index, and extreme response tendency index, respectively. The overall quality score is obtained by multiplying the response pattern index, reaction time index, and extreme response tendency index by their respective weights and then summing them.

9. The diagnostic method for depression and anxiety risk among highly mobile personnel in the judicial field as described in claim 8, characterized in that, The weights of the indicators are automatically updated, and the update steps include: Obtain the initial training set for the indicator weights; The initial training set is updated in the first round based on a reinforcement learning algorithm; A new round of updates will be performed using the unsupervised distribution alignment method if any of the following conditions are met: The amount of expert feedback data has reached a new threshold; The data distribution in the dataset has shifted.

10. A diagnostic device for depression and anxiety risk among highly mobile personnel in the judicial field, characterized in that, It includes a data acquisition module and a data processing module; The data acquisition module acquires the assessment dataset of the person to be assessed, which includes at least one of demographic sociology data, psychological scale data, and daily check-in data. The data processing module inputs the assessment dataset into a pre-trained fusion analysis model and outputs the quantitative scores of depression and anxiety of the person to be assessed. The fusion analysis model includes a feature fusion layer and a multi-task risk assessment layer; The feature fusion layer is used to convert the evaluation dataset into a unified feature representation to obtain a standardized dataset. The multi-task risk assessment layer includes several learners that learn from the standardized dataset from different perspectives and synthesize quantitative scores through a weighted voting method.

Images