Intelligent depression screening method and system based on causal learning and interpretable federal mechanism

Abstract

The invention discloses an intelligent depression screening method and system based on causal learning and an interpretable federation mechanism, relates to the technical field of depression risk assessment, and solves the problems of feature redundancy, lack of privacy protection, insufficient interpretability and insufficient interpretation credibility of a machine learning model in the prior art. The method comprises the steps of preprocessing collected original high-dimensional data, performing feature screening by utilizing a CIIG algorithm, performing training and optimization by adopting multiple classifiers to screen out a final depression classification and recognition model with the best effect, and generating explainable diagnosis output conforming to causal constraints based on a causal topological structure and an SHAP contribution value. Meanwhile, a federal causal consistency updating mechanism is adopted to carry out multi-node cooperative training and structure maintenance on the model so as to reduce input parameters and maintain or improve recognition performance, and an interpretable diagnosis mechanism based on causal topology driving improves result interpretability; and federal learning that data heterogeneous distribution fusion of multiple medical institutions and original data are not out of the local is supported.

Description

Technical Field

[0001] This application relates to the field of depression risk assessment technology, and in particular to an intelligent screening method for depression based on causal learning and an interpretable federal mechanism. Background Technology

[0002] Current methods for assessing depression risk commonly use all items from psychological scales as model input, resulting in significant feature redundancy, high model training costs, and inconsistent performance across different populations and scenarios. Furthermore, the decision-making process of such methods lacks clear explanatory evidence, making them unreliable by clinicians.

[0003] Secondly, while deep learning models can automatically extract deep features, their internal structure is a typical black box, making it difficult to reveal the pathological logic behind their predictions. This makes it difficult to guarantee clinical safety and reliability, and significantly limits their practical application in medicine.

[0004] Medical data distribution varies significantly across institutions. Existing screening models are mostly trained on single-center samples, making them difficult to adapt to multi-center deployments, and their performance even drops sharply when applied across datasets. Furthermore, directly sharing raw data for joint training poses a risk of privacy breaches.

[0005] Although existing methods have attempted to use statistical interpretation tools such as SHAP to present feature contributions, these interpretations are usually based only on feature correlations and cannot guarantee consistency with clinical causality. This can easily lead to misleading interpretations and cannot provide causal evidence to support diagnosis and treatment. Summary of the Invention

[0006] The purpose of this application is to overcome the problems of feature redundancy, lack of privacy protection, insufficient interpretability, and insufficient interpretability in existing machine learning models, and to provide a causal-driven, minimalist, clinically interpretable, and scalable intelligent screening method and system for depression based on causal learning and interpretable federated mechanisms.

[0007] Firstly, a method for intelligent screening of depression based on causal learning and an explainable federated mechanism is provided, including: Construct the original high-dimensional dataset and perform preprocessing; The CIIG causal screening algorithm was used to screen causal features in the preprocessed original high-dimensional dataset to identify key causal features that have a direct causal effect on the state of depression. Multiple depression classification and recognition models were constructed based on different classifiers; Based on the key causal features, multiple depression classification and recognition models were trained and optimized, and the model with the best performance was selected as the final depression classification and recognition model through evaluation. After the final depression classification and identification model outputs the depression risk prediction results, an interpretable diagnostic output that conforms to causal constraints is generated based on the causal topology and SHAP contribution value. A federated causal consistency update mechanism is adopted to perform multi-node collaborative training and structure preservation of the model.

[0008] In some possible implementations, the original high-dimensional dataset includes scores for each item of a user's psychological scale, clinical characteristics, and depression level, and the preprocessing includes missing value imputation and standardization.

[0009] In some possible implementations, the CIIG causal filtering algorithm is used to filter causal features from the preprocessed original high-dimensional dataset, including: Using conditional independence tests to construct a network of variable relationships; The information geometric causal orientation discrimination method is used to identify causal association characteristics; Redundant features related to the causal relationship with depression were removed to obtain a set of key causal features directly related to depression.

[0010] In some possible implementations, the number of key causal features is 10-15.

[0011] In some possible implementations, the number of key causal features is 12.

[0012] In some possible implementations, the classifiers include, but are not limited to: random forest, support vector machine, K-nearest neighbor and LightGBM. Multiple depression classification and recognition models are trained through multi-fold cross-validation and the recognition results are output. The accuracy, AUC, sensitivity and specificity of each classifier are calculated. The best performing model is selected as the final depression classification and recognition model based on the accuracy, AUC, sensitivity and specificity.

[0013] In some possible implementations, after the final depression classification and identification model outputs a depression risk prediction result, an interpretable diagnostic output that conforms to causal constraints is generated based on the causal topology and SHAP contribution value, including: For any node i, its contribution value after fusion The definition is as follows:

[0014] Where S(i) is the original SHAP contribution value calculated by the model, and w(p) i i) is the parent node P i The causal influence weight of node i, if node i has multiple parent nodes, is the combined contribution value. Represented as:

[0015] in, This represents the k-th parent node of node i; The explanation of the collaboration mechanism satisfies the principle of priority of parent node contribution and the constraint of causal consistency; Among them, the principle of priority for parent node contribution is: if p i If it is an upstream causal characteristic, then it satisfies:

[0016] in, Indicates the parent node p i The contribution value after fusion; Causal consistency constraint: Contribution explanation follows the edge direction and propagation chain of the causal graph; Interpretable diagnostic outputs include: a ranking of feature contribution values ​​by magnitude and a symptom evolution chain presented in the form of a causal chain.

[0017] In some possible implementations, structural consistency verification of causal structures includes: Compare the graph edit distance between the locally uploaded cause-effect graph and the global cause-effect graph; If the graph editing distance is less than or equal to the preset threshold, the verification passes and the model parameters are aggregated. If the graph editing distance is greater than a preset threshold, causal structure alignment correction is performed to restore causal topological consistency and model parameter aggregation is performed. If causal structure alignment correction cannot be performed, model parameter aggregation is rejected.

[0018] Secondly, a depression intelligent screening system based on causal learning and an explainable federated mechanism is provided, including: The data acquisition module is used to construct the original high-dimensional dataset and perform preprocessing. The feature selection module is used to select causal features from the preprocessed original high-dimensional dataset using the CIIG causal selection algorithm, so as to select key causal features that have a direct causal effect on the state of depression. The model building module is used to build various depression classification and recognition models based on different classifiers; The training and evaluation module is used to train and optimize multiple depression classification and recognition models based on the key causal features, and select the model with the best performance as the final depression classification and recognition model through evaluation. The output module is used to generate interpretable diagnostic output that conforms to causal constraints based on the causal topology and SHAP contribution value after the final depression classification and identification model outputs the depression risk prediction results. The federated training module is used to perform multi-node collaborative training and structure preservation of the model using a federated causal consistency update mechanism.

[0019] Thirdly, a computer-readable storage medium is provided for program code executable by a device, the program code including steps for performing a method as described in any of the implementations of the first aspect above.

[0020] This application has the following beneficial effects: 1. This application achieves conditionally independent structure adaptive inference, information geometric causal direction discrimination, and network stability constraints and redundant feature removal through the integrated CIIG causal screening algorithm. It also completes the integrated linkage optimization of causal feature learning and feature selection in one calculation, which can reduce the number of input features from 90 to 10-15, significantly reduce data acquisition and training costs, and maintain or improve recognition performance, which is beneficial for large-scale rapid screening of depression.

[0021] 2. This application is based on a causal topology-driven interpretable diagnostic mechanism. Instead of simply using SHAP output, this application utilizes the constraint features of causal network topology to contribute interpretability, outputting a "symptom evolution chain" and a path for tracing the source of depressive state, such as: sleep disorder → anxiety disorder → low mood → severity of depression.

[0022] 3. This application constructs a federated causal learning framework with cross-center generalization capabilities, which supports the fusion of heterogeneous data distribution from multiple medical institutions and federated learning where the original data does not leave the local area. This ensures that the consistency of causal structure constraints maintains long-term reliability, thereby improving the security of local data. It is significantly better than easily failed models that rely solely on statistical learning and can be used for a long time without frequent retraining.

[0023] 4. This application constructs a fully implementable clinical-grade intelligent screening system architecture for depression, forming a complete industrialization path from data input → causal screening → classification and identification → interpretable output → cross-institutional collaborative deployment: a modular system (data collection / causal screening / intelligent identification / interpretation and display), compatible with hospital workstations / self-service terminals / remote diagnosis and treatment apps, and supporting online model monitoring and performance self-recovery strategies. Attached Figure Description

[0024] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0025] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart of the intelligent screening method for depression based on causal learning and an explainable federal mechanism, as described in Embodiment 1 of this application; Figure 2 This is a schematic diagram of the causal network structure of depression in the intelligent screening method for depression based on causal learning and an explainable federated mechanism in Embodiment 1 of this application; Figure 3 This is a flowchart of model training in the intelligent screening method for depression based on causal learning and an interpretable federal mechanism according to Embodiment 1 of this application; Figure 4 This is a structural block diagram of the intelligent depression screening system based on causal learning and an explainable federal mechanism, as described in Embodiment 2 of this application. Figure 5 This is a hardware structure diagram of medical deployment in Embodiment 2 of this application.

[0027] Figure label: 100. Data Acquisition Module; 200. Feature Selection Module; 300. Model Building Module; 400. Training and Evaluation Module; 500. Output Module; 600. Federated Training Module. Detailed Implementation

[0028] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. Example 1

[0029] like Figure 1 As shown, Embodiment 1 of this application relates to an intelligent screening method for depression based on causal learning and an explainable federated mechanism, comprising: S100. Construct the original high-dimensional dataset and perform preprocessing.

[0030] Specifically, the system collects scores for each item on a scale filled out by users, along with clinical characteristics such as age and gender, and depression levels, to form a raw high-dimensional dataset. Data preprocessing operations, including missing value imputation and standardization, are then performed. In this embodiment, the data is collected from psychological scales, such as the SCL-90 scale, along with raw data and clinical characteristic data. Each user's completed scale score is... Clinical features are The target label Y represents the level of depression (0: normal, 1: mild, 2: moderate, 3: severe).

[0031] Preprocessing includes: 1) Missing value imputation: Imput missing values ​​of SCL-90 scale items using KNN imputation or multiple imputation techniques.

[0032] 2) Standardization process: Standardize all features according to the following formula:

[0033] in, These are the original eigenvalues. It is the mean of the i-th feature in the training set. Let be the standard deviation of the i-th feature in the training set. X represents the standardized feature values, and X is the original feature set.

[0034] This standardization step ensures that each feature has a mean of 0 and a standard deviation of 1, in order to eliminate the influence of dimensions and improve the stability of model training.

[0035] This standardization step ensures that each feature has a mean of 0 and a standard deviation of 1, in order to eliminate the influence of dimensions and improve the stability of model training.

[0036] After preprocessing the original high-dimensional dataset, the dataset is partitioned: the preprocessed dataset is divided... The dataset is divided into a 70% training set and a 30% validation set.

[0037] Dataset middle, The standardized SCL-90 scale score features are represented; quantitative variables (such as physiological measurements) in clinical feature C are also standardized in the same way, while categorical variables (such as gender, past medical history, etc.) are coded using One-Hot or Label; the target label Y is the result of the depression status grading, which belongs to discrete category data, so it is not standardized.

[0038] S200. The CIIG causal screening algorithm is used to screen the causal features of the preprocessed original high-dimensional dataset to identify key causal features that have a direct causal effect on the state of depression.

[0039] For psychological scales, such as the SCL-90 scale with 90 items, the CIIG (Causal Inference Information Geometry) causal screening algorithm proposed in this application is used: 1) Construct a network of variable relationships using conditional independence tests; 2) Identify causal relationship features using the information geometric causal direction discrimination method; 3) Remove redundant features that have no causal relationship with depression.

[0040] Ultimately, 10–15 key psychological indicators that have a direct causal relationship with depressive state were selected, with 12 being preferred.

[0041] The CIIG causal screening algorithm proposed in this application is the first to combine conditional independent structure learning, information geometric causal direction determination and network stability constraints. By preserving the core causal features, it significantly reduces the input dimension without compromising recognition performance, giving the model the advantages of interpretability, strong generalization and efficient deployment.

[0042] In this embodiment, the specific implementation steps of the CIIG causal screening algorithm are as follows: First, construct a variable relationship network: for each pair of features... and Perform conditional independence tests based on hypothesis testing methods. Determine features and Determine if the two variables are independent given other conditions S. If they are conditionally independent, delete the edge.

[0043] Then calculate the causal direction: for any two feature variables where a dependency was determined in the above steps. and The information theory criterion based on Kullback-Leibler (KL) divergence is used to determine the causal direction.

[0044] The symbols are defined as follows: : Feature variables The marginal probability distribution; : Feature variables The marginal probability distribution; Given features Features under conditions The conditional probability distribution; Given features Features under conditions The conditional probability distribution; (2) Formula for calculating KL divergence For any two probability distributions P(x) and Q(x), the KL divergence is defined as:

[0045] If the variable is discrete, it can be rewritten as a summation:

[0046] The smaller the KL value, the closer the two distributions are, meaning that the information loss is smaller when using Q to approximate P.

[0047] (3) Principle of Causal Direction Determination Calculate the information loss for the two possible directions separately: ; ; like This indicates that... For the reason, When the result is less information loss, the causal direction is determined. Otherwise, the causal direction is determined as follows: .

[0048] It is best to perform feature filtering: remove redundant features that have no causal relationship with depression Y, and finally obtain the set of features directly related to depression.

[0049] CIIG algorithm pseudocode: Input: X (SCL-9090 dimensional input) Output: Z (Set of key causal features) 1:G← Initialization of a fully connected graph 2: For each node, the adjacency degree k ≤ 4. 3: For any pair of nodes (x, y): 4: If the size of the condition set |S|≤2 and CI(x,y|S) passes through, then retain the edge. 5: For each edge e(x→y): 6: Calculate KL(P(y|x)||Q(y)) and KL(P(x|y)||Q(x)). 7: If the KL difference ≥ τ (= 0), determine the direction. 8: Calculate the causal stability index and remove weak edges. 9: Output a causal stable graph and select nodes with in-degree > 0 to form Z.

[0050] Here, the symbol CI(x, y | S) represents "the conditional independence test of x and y given a condition set S".

[0051] Test based on conditional mutual information formula:

[0052] when When the significance threshold is set to 0.01 to 0.05, CI(x,y|S) is considered to pass, meaning that x and y can be considered independent under the condition set S, and the corresponding edge is deleted from the graph G. Otherwise, if The two are considered to be dependent on each other, so the edge is retained.

[0053] like Figure 2 The diagram shown illustrates the causal network structure of depression, with symbol definitions and parameter descriptions: S: Conditioning set, which contains the set of other variables used to test conditional independence, and |S|≤2.

[0054] |S|: The size of the condition set S, i.e., the number of variables it contains.

[0055] CI(x,y|S): represents a conditional independence test for variables x and y given a condition set S.

[0056] e(x→y): The edge between nodes x and y in the graph. If the edges are determined to be independent, the edges are deleted.

[0057] S300. Construct various depression classification and recognition models based on different classifiers, including but not limited to: Random Forest, Support Vector Machine, K-Nearest Neighbors and LightGBM.

[0058] S400. Based on the key causal features, various depression classification and recognition models are trained and optimized respectively, and the model with the best performance is selected as the final depression classification and recognition model through evaluation.

[0059] Specifically, such as Figure 3As shown, the key features selected in step S200 are input into the depression classification and recognition model (including but not limited to random forest, support vector machine, K-nearest neighbors and LightGBM), and the model is trained using multi-fold cross-validation.

[0060] For each classifier, the accuracy Acc is calculated using the following formula:

[0061] Where TP is the number of true positives, TN is the number of true negatives, FP is the number of false positives, and FN is the number of false negatives.

[0062] Model optimization: The hyperparameters of the classifier are tuned using methods such as grid search or Bayesian optimization to achieve optimal model performance.

[0063] Evaluation and Selection: Different classification models are evaluated on the validation set, and the best-performing model is selected to output the depression risk level. And calculate various performance indicators (accuracy, AUC, sensitivity, specificity, etc.).

[0064] Testing using a real clinical dataset (sample size of no less than 3000 cases) showed that: Using the causal features selected in this embodiment to train the classification model can improve the accuracy from 85% to over 94%. Input features are reduced by more than 85%, and computation latency is reduced by approximately 70%. This demonstrates that the application offers significant performance improvements and deployment friendliness.

[0065] S500: After the final depression classification and identification model outputs the depression risk prediction results, an interpretable diagnostic output that conforms to causal constraints is generated based on the causal topology and SHAP contribution value.

[0066] To improve the interpretability of depression screening models in medical scenarios, this application proposes an interpretive synergy mechanism based on causal topology. After the model generates the original SHAP contribution values, this application further introduces causal weight modulation, so that the feature contribution interpretation not only reflects statistical importance but also the strength of causal influence. Specifically, for any node i, its fused contribution value... The definition is as follows:

[0067] Where S(i) is the original SHAP contribution calculated by the model, w(p) i i) is the parent node P i The causal influence weight of node i is derived from the causal edge weights obtained during the CIIG causal learning phase of this application. If a node has multiple parent nodes, the contribution value after fusion is... It can be represented as:

[0068] Among them, symbols This represents the k-th parent node of node i.

[0069] Based on the above formula, the interpretation mechanism of this application satisfies the following constraints: (1) Parent node contribution priority principle: If it is an upstream cause characteristic, then it must be:

[0070] Among them, symbols Indicates the parent node p i The fused contribution value is the value by which upstream nodes in the causal chain are preferentially represented in the explanation, in order to conform to the clinical etiological transmission logic.

[0071] (2) Causal consistency constraint: The contribution explanation follows the edge direction and propagation chain of the causal graph, and the output results can truly reflect the path of disease development, rather than the influence of isolated features.

[0072] (3) Traceable and interpretable output: This application presents the merged contribution results through a causal path structure, and the output includes: 1) Ranking by the magnitude of feature contribution values; 2) The “symptom evolution chain” presented in the form of a causal chain.

[0073] This enables clinicians to trace the source of increased risk and provide targeted intervention recommendations, avoiding the problems of existing black-box models being difficult to verify and unreliable.

[0074] S600 employs a federated causal consistency update mechanism to perform multi-node collaborative training and structure preservation of the model.

[0075] To address the issue of causal structure corruption in multi-center training, this application proposes a federated causal consistency constraint mechanism. In this application, each participating node only uploads its local causal structure and parameter update information, without sharing the original data. While aggregating model parameters, the server performs structural consistency verification on the uploaded causal graph. Specifically, the server first compares the graph edit distance (GED) between the local causal graph and the global causal graph. When the GED value is less than or equal to a threshold δ (δ≤0.2), the local update is considered not to corrupt the global causal topology, and normal aggregation can proceed. When the GED value exceeds the threshold, causal structure alignment correction is performed, including adjusting edge directions and removing abnormal causal edges, to restore causal topology consistency. If the deviation is too large to be corrected, the update request is rejected to avoid introducing erroneous causal relationships that would reduce the overall interpretability and generalization of the model.

[0076] In this application, federated learning is primarily used to achieve collaborative updates of causal structures and model parameters across multiple institutions or nodes without sharing the original data, thereby ensuring data privacy and model consistency. Its implementation includes the following three core components: (1) Local modeling and causal structure extraction: Each participating node (such as different medical institutions, terminal devices, or computing nodes) utilizes its own dataset locally. Independently train the model to obtain the local causal graph structure. and its corresponding parameter set .

[0077] During local training, the node performs the following operations: 1) Extract the set of key local causal features using the CIIG causal screening algorithm; 2) Construct a local causal topology graph; 3) Train a depression screening model based on local samples.

[0078] Once completed, only upload a summary of the causal structure (such as an adjacency matrix or graph feature encoding) and model parameters, without uploading any original patient data.

[0079] (2) Central server causal consistency aggregation mechanism: After receiving the local causal graphs from each node, the central server executes the causal consistency constraint aggregation mechanism.

[0080] To prevent system instability caused by local causal topological conflicts, the central server first calculates the graph edit distance (GED) between the local graph and the global graph:

[0081] in: : Cause-effect graph uploaded by local nodes; The global cause-effect graph currently stored on the central server; The cost of editing a node or edge; Edit operation sequence.

[0082] like (The threshold is usually set to 0.2), then the local update is considered to be consistent with the global causal structure, and parameter aggregation can be performed; like Then, topology alignment correction is performed, including: 1) Adjust the direction of the edges to eliminate conflicts; 2) Delete abnormal causal edges; 3) Perform structural calibration on key nodes.

[0083] If the deviation is too large to be corrected, the update request for that node will be rejected in order to maintain the stability of the global structure.

[0084] (3) Parameter aggregation and global update: The model parameters of the nodes that pass the consistency verification The weighted aggregation process will begin.

[0085] The aggregation rule is:

[0086] in, The node weights can be dynamically adjusted based on the number of samples, data quality, or GED consistency.

[0087] After aggregation is complete, the server generates a new global model. And distribute it back to each node.

[0088] Each node periodically repeats the process of "local update → upload summary → aggregate verification" to achieve multi-center collaborative learning.

[0089] The pseudocode is as follows: #Input: G_local (local causal graph), G_global (current global graph) #Output: Updated global graph G_global GED = GraphEditDistance(G_local, G_global) if GED <= δ: # δ = 0.2 G_global = AggregateStructure(G_global, G_local) else: G_c = AlignStructure(G_local, G_global) if GraphEditDistance(G_c, G_global)<= δ: G_global = AggregateStructure(G_global, G_c) else: RejectUpdate(G_local).

[0090] In this embodiment, the accuracy of intelligent identification of depression is improved by about 10%, and the AUC is improved by more than 0.05, significantly improving the reliability of the diagnostic results; the number of input features is reduced from 90 to 10-15, significantly reducing the data collection and training costs, which is conducive to large-scale rapid screening; based on the combination of causal topology and SHAP analysis, the output can be interpreted in accordance with collective clinical cognition, improving the acceptability and reference value for doctors.

[0091] This application proposes a CIIG causal screening algorithm. Unlike existing step-by-step causal learning, the CIIG (Causal Inference Information Geometry) algorithm of this invention achieves: conditionally independent structure adaptive inference, information geometric causal direction discrimination, network stability constraints and redundant feature removal (non-statistical correlation screening), and completes integrated linkage optimization of causal feature learning and feature selection in a single computation. Secondly, this application does not simply use SHAP output, but utilizes the causal network topology to constrain feature interpretability, outputting a "symptom evolution chain" and a path for tracing the source of depressive state, such as: sleep disorder → anxiety disorder → low mood → severity of depression. Furthermore, through causal screening, psychological scales, such as the SCL-90, are compressed into 10–15 core causal features, while maintaining or improving recognition performance, as shown in Table 1 below. Table 1:

[0092] This application also features a federated causal learning framework with cross-center generalization capabilities, supporting: heterogeneous data fusion from multiple medical institutions, federated learning where the original data remains local, and long-term reliability maintained by causal structure consistency constraints. This is significantly superior to easily failing models that rely solely on statistical learning, allowing for long-term use without frequent retraining. It also presents a fully deployable, clinical-grade intelligent screening system architecture for depression, forming a complete industrialization path from data input → causal screening → classification and identification → interpretable output → cross-institutional collaborative deployment: a modular system (data acquisition / causal screening / intelligent identification / interpretive display), compatible with hospital workstations / self-service terminals / remote diagnosis and treatment apps, and supporting online model monitoring and performance self-recovery strategies.

[0093] Example 2 like Figure 4 As shown, Embodiment 2 of this application relates to an intelligent screening system for depression based on causal learning and an explainable federated mechanism, comprising: Data acquisition module 100 is used to construct the original high-dimensional dataset and perform preprocessing; The feature selection module 200 is used to perform causal feature selection on the preprocessed original high-dimensional dataset using the CIIG causal selection algorithm, so as to select key causal features that have a direct causal effect on the state of depression. Model building module 300 is used to build various depression classification and recognition models based on different classifiers; The training and evaluation module 400 is used to train and optimize multiple depression classification and recognition models based on the key causal features, and select the best performing model as the final depression classification and recognition model through evaluation. The output module 500 is used to generate an interpretable diagnostic output that conforms to causal constraints based on the causal topology and SHAP contribution value after the final depression classification and identification model outputs the depression risk prediction results. Output module 500 specifically includes the following modules: The feature list display module shows the key psychological features involved in the diagnosis of depression and their names, allowing clinicians to quickly identify factors of concern. This module supports sorting by contribution and allows users to expand to view detailed feature definitions.

[0094] The feature contribution value visualization module, based on SHAP (Shapley Additive Explanation) calculation results, displays the positive and negative contributions of each feature to the diagnostic results through horizontal bar charts, line charts, or color gradients. This helps doctors intuitively determine which features are most crucial for predicting depression.

[0095] The sample-level interpretation module supports displaying the basis for the model's predictions for each sample, such as the user's specific psychological symptoms and their corresponding impact, enabling personalized interpretation capabilities. This facilitates clinical judgment and intervention recommendations by doctors.

[0096] The prediction results and risk grading module displays the sample's depression risk level (e.g., normal, mild, moderate to severe) based on the final classification model output. It supports intuitive color or icon labeling for rapid clinical interpretation.

[0097] The user interaction and configuration module provides a visual interface for interactive functions, including: (1) switching between different samples to view interpretation results; (2) adjusting thresholds; and (3) exporting analysis reports. This improves system usability and clinical workflow friendliness.

[0098] The Federated Training Module 600 is used to perform multi-node collaborative training and structure preservation of the model using a federated causal consistency update mechanism.

[0099] For example, such as Figure 5 As shown, the hardware architecture of this application in a medical deployment scenario includes, but is not limited to, the following modules: The input terminal, equipped with functions such as questionnaire completion and user interaction, can be used as a hospital workstation, outpatient self-service machine, tablet, mobile app, or remote psychological screening terminal. It is used to collect SCL-90 data and necessary clinical information completed by the user.

[0100] The data acquisition and transmission module supports real-time transmission of acquired data to the server via intranet / extranet connection, employing an encrypted transmission protocol to ensure data security.

[0101] The medical server / cloud computing node, the main execution carrier of the algorithm in this application, has the following built-in features: (1) Causal Feature Screening Unit (CIIG); (2) Intelligent Diagnostic Model; and (3) Model Interpretability Analysis Module. The server can be deployed in a hybrid manner, such as private cloud, edge computing node, or public cloud, depending on the scale of the hospital.

[0102] The data storage module stores user psychological scales, model training update logs, and other data. It complies with medical data privacy regulations and supports encryption, access control, and audit trails.

[0103] The front-end display and doctor terminal are used as a clinical interface for physicians, displaying depression risk classification results, interpretable analytical information (feature contribution, etc.), and historical data comparison and trend analysis.

[0104] The update and maintenance module supports online updates, corrections, and performance monitoring of the system model, ensuring the long-term reliability of diagnostic results.

[0105] Advantages: This deployment architecture features modularity, scalability, and high security, facilitating rapid deployment and widespread clinical application of the proposed method across multi-tiered healthcare institutions. It should be noted that other specific implementations of the intelligent depression screening system based on causal learning and interpretable federated mechanisms in this embodiment can be found in the specific implementations of the intelligent depression screening method based on causal learning and interpretable federated mechanisms described above. To avoid redundancy, they will not be repeated here.

[0106] Example 3 This application relates to a computer-readable storage medium in Embodiment 3, which is used for program code executed by a device, the program code including steps for performing a method as described in any implementation of Embodiment 1 of this application; The computer-readable storage medium may be a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM); the computer-readable storage medium may store program code, and when the program stored in the computer-readable storage medium is executed by a processor, the processor is used to perform the steps of the method in any of the implementations of Embodiment 1 of this application.

[0107] The above are merely preferred embodiments of this application; however, the scope of protection of this application is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in this application, based on the technical solution and its improved concept, should be covered within the scope of protection of this application.

Claims

1. A method for intelligent screening of depression based on causal learning and an explainable federated mechanism, characterized in that, include: Construct the original high-dimensional dataset and perform preprocessing; The CIIG causal screening algorithm was used to screen causal features in the preprocessed original high-dimensional dataset to identify key causal features that have a direct causal effect on the state of depression. Multiple depression classification and recognition models were constructed based on different classifiers; Based on the key causal features, multiple depression classification and recognition models were trained and optimized, and the model with the best performance was selected as the final depression classification and recognition model through evaluation. After the final depression classification and identification model outputs the depression risk prediction results, an interpretable diagnostic output that conforms to causal constraints is generated based on the causal topology and SHAP contribution value. A federated causal consistency update mechanism is adopted to perform multi-node collaborative training and structure preservation of the model.

2. The intelligent screening method for depression based on causal learning and an explainable federated mechanism according to claim 1, characterized in that, The original high-dimensional dataset includes scores for each item of the user's psychological scale, clinical characteristics, and depression level. The preprocessing includes missing value imputation and standardization.

3. The intelligent screening method for depression based on causal learning and an explainable federated mechanism according to claim 1, characterized in that, The CIIG causal filtering algorithm is used to filter causal features in the preprocessed original high-dimensional dataset, including: Using conditional independence tests to construct a network of variable relationships; The information geometric causal orientation discrimination method is used to identify causal association characteristics; Redundant features related to the causal relationship with depression were removed to obtain a set of key causal features directly related to depression.

4. The intelligent screening method for depression based on causal learning and an explainable federated mechanism according to claim 3, characterized in that, The number of key causal features is 10-15.

5. The intelligent screening method for depression based on causal learning and an explainable federated mechanism according to claim 3, characterized in that, The number of key causal features is 12.

6. The intelligent screening method for depression based on causal learning and interpretable federated mechanisms according to any one of claims 1-5, characterized in that, The classifiers include, but are not limited to, random forest, support vector machine, K-nearest neighbor and LightGBM. Multiple depression classification and recognition models are trained through multi-fold cross-validation and the recognition results are output. The accuracy, AUC, sensitivity and specificity of each classifier are calculated. The best performing model based on accuracy, AUC, sensitivity and specificity is selected as the final depression classification and recognition model.

7. The intelligent screening method for depression based on causal learning and an explainable federated mechanism according to claim 6, characterized in that, After the final depression classification and identification model outputs depression risk prediction results, an interpretable diagnostic output that conforms to causal constraints is generated based on the causal topology and SHAP contribution value, including: For any node i, its contribution value after fusion The definition is as follows: Where S(i) is the original SHAP contribution value calculated by the model, and w(p) i i) is the parent node P i The causal influence weight of node i, if node i has multiple parent nodes, is the combined contribution value. Represented as: in, This represents the k-th parent node of node i; The explanation of the collaboration mechanism satisfies the principle of priority of parent node contribution and the constraint of causal consistency; Among them, the principle of priority for parent node contribution is: if p i If it is an upstream causal characteristic, then it satisfies: in, Indicates the parent node p i The contribution value after fusion; Causal consistency constraint: Contribution explanation follows the edge direction and propagation chain of the causal graph; Interpretable diagnostic outputs include: a ranking of feature contribution values ​​by magnitude and a symptom evolution chain presented in the form of a causal chain.

8. The intelligent screening method for depression based on causal learning and an interpretable federated mechanism according to claim 1 or 7, characterized in that, The federal causal consistency update mechanism includes: Compare the graph edit distance between the locally uploaded cause-effect graph and the global cause-effect graph; If the graph editing distance is less than or equal to the preset threshold, the verification passes and the model parameters are aggregated. If the graph editing distance is greater than a preset threshold, causal structure alignment correction is performed to restore causal topological consistency and model parameter aggregation is performed. If causal structure alignment correction cannot be performed, model parameter aggregation is rejected.

9. A depression intelligent screening system based on causal learning and an explainable federated mechanism, characterized in that, include: The data acquisition module is used to construct the original high-dimensional dataset and perform preprocessing. The feature selection module is used to select causal features from the preprocessed original high-dimensional dataset using the CIIG causal selection algorithm, so as to select key causal features that have a direct causal effect on the state of depression. The model building module is used to build various depression classification and recognition models based on different classifiers; The training and evaluation module is used to train and optimize multiple depression classification and recognition models based on the key causal features, and select the model with the best performance as the final depression classification and recognition model through evaluation. The output module is used to generate interpretable diagnostic output that conforms to causal constraints based on the causal topology and SHAP contribution value after the final depression classification and identification model outputs the depression risk prediction results. The federated training module is used to perform multi-node collaborative training and structure preservation of the model using a federated causal consistency update mechanism.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is for program code executed by the device, the program code including steps for performing the method as described in any one of claims 1-8.

Images