A training method and system of a depression intervention model based on a deep knowledge graph

Abstract

The application discloses a kind of training method and system of depression intervention model based on deep knowledge graph, it is related to intelligent medical treatment and artificial intelligence technical field, the method is by obtaining the medical record text of patient, emotional language interaction data, sleep behavior data and physiological monitoring data.Data are carried out semantic analysis and entity relationship extraction, construct depression intervention knowledge graph, and embedding representation learning is carried out based on graph neural network, generates patient state vector.Patients state vector is input into intervention strategy generation model, and generates personalized intervention scheme prediction result.Based on prediction result, rehabilitation training is carried out, and intervention strategy is iteratively optimized through patient feedback.The method realizes the dynamic association of symptom, intervention and situation, and can continuously optimize personalized intervention scheme according to curative effect feedback, improve the rehabilitation effect of depression symptom and sleep disorder.

Description

Technical Field

[0001] This specification relates to the fields of smart healthcare and artificial intelligence technology. More specifically, this application relates to a training method and system for a depression intervention model based on deep knowledge graphs. Background Technology

[0002] Depressive disorders and their associated sleep problems have become a significant factor affecting public mental and physical health, with an increasing incidence rate and characterized by long course, high relapse rate, and significant individual differences. Current intervention methods for depression typically rely on physician experience to develop treatment plans or make intervention decisions based on single scale assessments, making it difficult to fully integrate information contained in multi-dimensional data such as patient medical records, daily communication, sleep behavior, and physiological state. Furthermore, while some systems have attempted to use machine learning models to assist in developing intervention strategies, most still employ traditional feature engineering or static model structures, lacking the ability to structurally express the complex relationships between depressive symptoms and failing to continuously model and dynamically update the efficacy feedback during the intervention process. In addition, when processing long-term patient follow-up data, current technologies often only perform simple data statistics or score predictions, failing to form a knowledge system with causal correlation and semantic reasoning capabilities, resulting in a lack of interpretability and continuous optimization mechanisms in the generation of intervention plans.

[0003] Therefore, how to construct a training method that can integrate multimodal data, dynamically express the relationship between symptoms and interventions, and continuously optimize strategies in actual rehabilitation training has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0004] The summary section introduces a series of simplified concepts, which will be further explained in detail in the detailed description section. This summary section is not intended to limit the key and essential technical features of the claimed technical solution, nor is it intended to determine the scope of protection of the claimed technical solution.

[0005] Firstly, this application proposes a training method for a depression intervention model based on deep knowledge graphs, including: Acquire multimodal data of the target patients, including medical record text, emotional language interaction data, sleep behavior data, and physiological monitoring data. Semantic parsing and entity relation extraction were performed on the above multimodal data to construct a knowledge graph for depression intervention. The knowledge graph for depression intervention includes symptom nodes, intervention measure nodes, situational factor nodes, and efficacy feedback nodes. Based on graph neural network, the above depression intervention knowledge graph is embedded and representation learned to obtain the patient state vector; The aforementioned patient state vectors are input into the intervention strategy generation model for identification, in order to generate personalized intervention plan prediction results for different patient states; Based on the predicted results of the above personalized intervention program, rehabilitation training is conducted for different patients.

[0006] In one feasible implementation, the above-mentioned semantic parsing and entity relation extraction of the multimodal data are used to construct a knowledge graph for depression intervention, including: Natural language processing operations were performed on the aforementioned medical record text and the aforementioned emotional language interaction data to identify depressive symptom entities, intervention behavior entities, and contextual semantic labels; Temporal feature encoding is performed on the above sleep behavior data and the above physiological monitoring data to generate quantitative indicator nodes corresponding to the above symptom entities; Based on a pre-defined medical knowledge ontology, the above-mentioned symptom entities, intervention behavior entities, and contextual semantic tags are mapped to form an initial graph structure. Based on the aforementioned efficacy feedback nodes, the relationships are updated with confidence weighting to generate the aforementioned knowledge graph of depression intervention.

[0007] In one feasible implementation, the above-mentioned relationship is updated with confidence weighting based on the above-mentioned efficacy feedback nodes to generate the above-mentioned depression intervention knowledge graph, including: Obtain the changes in emotional scores and sleep quality indicators before and after the intervention corresponding to the above-mentioned efficacy feedback nodes; The intervention effect weighting coefficient was calculated based on the above changes in emotion scores and sleep quality indicators. The association edge weights between the symptom nodes and the intervention measure nodes are dynamically adjusted based on the aforementioned intervention effect weight coefficients. The updated associated edge weights are written into the initial graph structure to generate the aforementioned knowledge graph of depression intervention with adaptive structure updates.

[0008] In one feasible implementation, the above-mentioned embedding representation learning based on the graph neural network on the aforementioned depression intervention knowledge graph to obtain the patient state vector includes: The node features and edge weights in the aforementioned depression intervention knowledge graph are vectorized and initialized to generate initial node representations; The adjacency relationships between the symptom nodes and the contextual factor nodes are weighted and aggregated based on the graph attention mechanism to obtain the first-layer graph representation; By combining the above-mentioned efficacy feedback nodes, the first-layer graph representation is trained under supervised constraints to obtain the updated graph embedding representation; The updated graph embedding representation is then subjected to a patient-level readout operation to generate the aforementioned patient state vector.

[0009] In one feasible implementation, the above-mentioned graph attention mechanism is used to perform weighted aggregation of the adjacency relationships between the symptom nodes and the contextual factor nodes to obtain a first-layer graph representation, including: Calculate the semantic similarity and temporal correlation between the above symptom nodes and the above contextual factor nodes to generate multidimensional attention weights; Based on the aforementioned multidimensional attention weights, the features of adjacent nodes are weighted and summed to obtain a local context representation; The above local context representation is input into a nonlinear activation function for feature transformation to obtain the node update vector; The update vectors of the above nodes are aggregated to form the first layer graph representation.

[0010] In one feasible implementation, the first-layer graph representation is subjected to supervised and constrained training in conjunction with the aforementioned therapeutic feedback nodes to obtain an updated graph embedding representation, including: Construct a graph embedding loss function with the above-mentioned efficacy feedback nodes as supervision labels, wherein the graph embedding loss function includes an intervention effect prediction error term and a node relationship preservation constraint term; The first layer graph representation above is trained by backpropagation algorithm to update parameters and obtain the optimized node embedding vector. The optimized node embedding vectors are regularized to enhance generalization ability across patient samples. Output the updated graph embedding representation after supervised constraint optimization.

[0011] In one feasible implementation, the above-mentioned patient state vector is input into the intervention strategy generation model for identification, in order to generate personalized intervention plan prediction results for different patient states, including: The patient state vector and the context features corresponding to the context factor nodes are fused and encoded to form the policy input features; The above-mentioned strategy input features are then input into a reinforcement learning-based intervention strategy generation model to generate a sequence of candidate intervention schemes. Based on the historical intervention effects corresponding to the above efficacy feedback nodes, the reward function is calculated for the above candidate intervention scheme sequence, and the parameters of the above intervention strategy generation model are updated. Output the predicted results of the above personalized intervention plan that match the current patient status.

[0012] In one feasible implementation, the above-mentioned input of the policy input features into a reinforcement learning-based intervention policy generation model to generate a sequence of candidate intervention schemes includes: The above policy input features are mapped to state vectors in the policy state space; Based on the policy network, the above state vector is sampled for action to generate multiple combinations of intervention actions; Based on the constraint rules corresponding to the above-mentioned situational factor nodes, the legality of the above-mentioned combination of intervention actions is screened to form candidate intervention paths; The candidate intervention paths are encoded in chronological order to generate the sequence of candidate intervention schemes.

[0013] In one feasible implementation, the above-mentioned reward function calculation for the candidate intervention program sequence based on the historical intervention effects corresponding to the above-mentioned efficacy feedback nodes, and parameter update of the above-mentioned intervention strategy generation model, includes: A multi-objective reward function is constructed based on the degree of symptom improvement, the extent of sleep quality improvement, and compliance indicators in the above-mentioned efficacy feedback nodes; The return values ​​of the above candidate intervention scheme sequences are calculated, and a policy gradient update signal is generated; The policy network parameters of the above intervention policy generation model are iteratively optimized based on the above policy gradient update signal; The output is the model generated by the above intervention strategy after parameter updates.

[0014] Secondly, this application proposes a training system for a depression intervention model based on deep knowledge graphs, used to execute the training method for the depression intervention model based on deep knowledge graphs described in the first aspect, including: The acquisition unit is used to acquire multimodal data of the target patient, including medical record text, emotional language interaction data, sleep behavior data, and physiological monitoring data. The construction unit is used to perform semantic parsing and entity relation extraction on the above multimodal data to construct a depression intervention knowledge graph, which includes symptom nodes, intervention measure nodes, situational factor nodes, and efficacy feedback nodes. The embedding unit is used to learn the embedding representation of the above-mentioned depression intervention knowledge graph based on a graph neural network to obtain the patient state vector. The identification unit is used to input the above-mentioned patient state vector into the intervention strategy generation model for identification, so as to generate personalized intervention plan prediction results for different patient states. The training unit is used to conduct rehabilitation training for different patients based on the predicted results of the above-mentioned personalized intervention plan.

[0015] In summary, compared to traditional depression intervention methods that rely on a single data source or fixed rules, this invention achieves a fully intelligent upgrade of the entire process from data acquisition, knowledge construction, state representation to intervention strategy generation by introducing a technical approach combining deep knowledge graphs and graph neural networks. By performing unified semantic parsing and relational modeling on medical record text, emotional language interaction data, sleep behavior data, and physiological monitoring data, this invention can simultaneously express symptom evolution, contextual influence, and efficacy feedback within the knowledge graph. This allows intervention decisions to move beyond single indicators and instead be based on comprehensive analysis of multi-dimensional information, significantly improving the accuracy of individual state identification. Furthermore, this invention utilizes graph neural networks for embedding representation learning, transforming complex graph structure information into computable patient state vectors. This enables the model to capture potential correlations and dynamic trends between symptoms, thereby improving the precision of intervention strategy matching. Finally, this invention introduces a reinforcement learning-based recognition mechanism during the intervention strategy generation process and continuously optimizes it in conjunction with efficacy feedback nodes. This allows the system to continuously adjust the intervention plan based on actual rehabilitation training effects, achieving a shift from static recommendation to dynamic adaptive intervention. This invention integrates the knowledge graph update process with a closed-loop rehabilitation training system, enabling training results to drive model iteration and enhance the system's long-term learning ability and clinical applicability. This invention improves personalization, interpretability, continuous optimization capabilities, and comprehensive intervention effects on accompanying problems such as sleep disorders.

[0016] The training method for the depression intervention model based on deep knowledge graphs proposed in this application, along with other advantages, objectives, and features of this application, will be partly reflected in the following description and partly understood by those skilled in the art through research and practice of this application. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit this specification. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A schematic diagram illustrating the training method of a depression intervention model based on deep knowledge graphs, provided in an embodiment of this application; Figure 2 This application provides a flowchart illustrating a method for constructing a knowledge graph for depression intervention. Figure 3 This is a flowchart illustrating a method for updating relationships with confidence weight based on efficacy feedback nodes, as provided in an embodiment of this application. Figure 4 This is a flowchart illustrating a method for obtaining a patient state vector, as provided in an embodiment of this application. Figure 5 This is a flowchart illustrating a method for obtaining a first-layer graph representation provided in an embodiment of this application. Figure 6 This is a flowchart illustrating a method for obtaining an updated graph embedding representation, as provided in an embodiment of this application. Figure 7 A flowchart illustrating a method for generating prediction results of personalized intervention plans for different patient states, provided in an embodiment of this application; Figure 8 This is a flowchart illustrating a method for generating a sequence of candidate intervention schemes, provided in an embodiment of this application. Figure 9 This is a flowchart illustrating a method for updating parameters of an intervention strategy generation model, as provided in an embodiment of this application. Figure 10 This is a structural diagram of a training system for a depression intervention model based on a deep knowledge graph, provided in an embodiment of this application. Detailed Implementation

[0018] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus. The technical solutions of the embodiments of this application will now be clearly and completely described in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0019] Please see Figure 1 This is a schematic diagram illustrating the training process of a depression intervention model based on deep knowledge graphs, as provided in an embodiment of this application. Specifically, it may include: S110. Obtain multimodal data of the target patient, wherein the multimodal data includes medical record text, emotional language interaction data, sleep behavior data, and physiological monitoring data; S120. Perform semantic parsing and entity relation extraction on the above multimodal data to construct a knowledge graph for depression intervention. The knowledge graph for depression intervention includes symptom nodes, intervention measure nodes, situational factor nodes, and efficacy feedback nodes. S130. Based on graph neural network, the above depression intervention knowledge graph is embedded and representation learned to obtain the patient state vector; S140. Input the above patient state vector into the intervention strategy generation model for identification, so as to generate personalized intervention plan prediction results for different patient states. S150. Based on the prediction results of the above-mentioned personalized intervention plan, rehabilitation training is carried out for different patients.

[0020] For example, the present application provides a training method for a depression intervention model based on deep knowledge graphs, which aims to transform multi-source heterogeneous information of patients into computable knowledge representations, and form interpretable and iterative personalized intervention plans with the support of graph reasoning and deep learning representations, thereby achieving synergistic rehabilitation of depressive state improvement and sleep disorder-related symptoms.

[0021] In step S110, multimodal data of the target patient is acquired, including medical record texts that reflect past medical history, medication history, comorbidities and clinical assessment conclusions, as well as emotional and language interaction data formed by the patient during follow-up dialogues, psychological scale questioning and answering, voice / text communication, etc. At the same time, sleep behavior data that reflects circadian rhythm and sleep structure and physiological monitoring data that reflects autonomic nervous activity and stress level are also acquired. The multimodal parallel acquisition method ensures the completeness and objectivity of the patient's state profile and provides a continuous and traceable data foundation for subsequent map construction.

[0022] In step S120, semantic parsing and entity relation extraction are performed on the aforementioned multimodal data. Symptom descriptions, triggering factors, coping strategies, and intervention behaviors in the medical record text and emotional language interaction data are identified as structured entities. Key indicators in sleep behavior data and physiological monitoring data are mapped as quantifiable graph attributes or nodes, thereby constructing a knowledge graph for depression intervention. In this graph, symptom nodes represent core manifestations such as low mood, loss of interest, accompanying anxiety, difficulty falling asleep, or early awakening. Intervention measure nodes represent actionable strategies such as psychological intervention, sleep hygiene education, cognitive behavioral training, exercise prescriptions, and medication adjustment suggestions. Contextual factor nodes represent important background factors affecting intervention effectiveness, such as stressful events, living environment, social support, and adherence constraints. Therapeutic feedback nodes carry feedback information such as changes in scales after intervention, the extent of improvement in sleep indicators, and changes in relapse risk. This allows the graph to not only possess static associations between symptoms, causes, and interventions but also to form an updatable knowledge loop driven by feedback.

[0023] In step S130, the aforementioned knowledge graph of depression intervention is embedded using a graph neural network. By message passing and aggregation of node features and edge relationships, information scattered across different nodes and relationships is fused into a unified low-dimensional vector space, thereby obtaining a patient state vector. This patient state vector can simultaneously encode the patient's symptom combination features, sleep problem patterns, situational constraints, and historical treatment feedback. It reflects not only the current severity but also individual differences in which interventions the patient is more sensitive to and in which situations they are more prone to fluctuations, providing a state representation foundation for subsequent strategy identification and decision-making.

[0024] In step S140, the aforementioned patient state vector is input into the intervention strategy generation model for identification, in order to generate personalized intervention plan prediction results for different patient states. The intervention strategy generation model can learn the benefit differences of different intervention actions in different states based on historical efficacy feedback, and complete the matching and combination of intervention plans by combining contextual factor constraints, thereby outputting executable intervention content, frequency, stage goals and precautions and other plan elements, so that the output results are both individualized and adapted, and have interpretable knowledge path support.

[0025] In step S150, rehabilitation training is conducted for different patients based on the prediction results of the above-mentioned personalized intervention plan. Specifically, patients can be organized to complete psychological training tasks, sleep behavior adjustment tasks, and daily behavior activation tasks according to the training cycle and stage goals defined in the plan. At the same time, compliance and efficacy feedback data are collected during the training process, so that the training results can be fed back to the efficacy feedback node and used for subsequent map updates and model iterations. In this way, the intervention strategy can be gradually optimized and the individualized accuracy can be improved during continuous use.

[0026] In summary, compared to traditional depression intervention methods that rely on a single data source or fixed rules, this invention achieves a fully intelligent upgrade of the entire process from data acquisition, knowledge construction, state representation to intervention strategy generation by introducing a technical approach combining deep knowledge graphs and graph neural networks. By performing unified semantic parsing and relational modeling on medical record text, emotional language interaction data, sleep behavior data, and physiological monitoring data, this invention can simultaneously express symptom evolution, contextual influence, and efficacy feedback within the knowledge graph. This allows intervention decisions to move beyond single indicators and instead be based on comprehensive analysis of multi-dimensional information, significantly improving the accuracy of individual state identification. Furthermore, this invention utilizes graph neural networks for embedding representation learning, transforming complex graph structure information into computable patient state vectors. This enables the model to capture potential correlations and dynamic trends between symptoms, thereby improving the precision of intervention strategy matching. Finally, this invention introduces a reinforcement learning-based recognition mechanism during the intervention strategy generation process and continuously optimizes it in conjunction with efficacy feedback nodes. This allows the system to continuously adjust the intervention plan based on actual rehabilitation training effects, achieving a shift from static recommendation to dynamic adaptive intervention. This invention integrates the knowledge graph update process with a closed-loop rehabilitation training system, enabling training results to drive model iteration and enhance the system's long-term learning ability and clinical applicability. This invention improves personalization, interpretability, continuous optimization capabilities, and comprehensive intervention effects on accompanying problems such as sleep disorders.

[0027] In one feasible implementation, such as Figure 2 As shown, step S120 above performs semantic parsing and entity relation extraction on the multimodal data to construct a knowledge graph for depression intervention, including: S1201. Perform natural language processing operations on the above-mentioned medical record text and the above-mentioned emotional language interaction data to identify depressive symptom entities, intervention behavior entities and contextual semantic labels. S1202. Perform time-series feature encoding on the above sleep behavior data and the above physiological monitoring data to generate quantitative indicator nodes corresponding to the above symptom entities. S1203. Based on the preset medical knowledge ontology, the above symptom entities, the above intervention behavior entities and the above contextual semantic tags are mapped to each other to construct an initial graph structure. S1204. Based on the above efficacy feedback nodes, the above relationships are updated with confidence weighting to generate the above depression intervention knowledge graph.

[0028] In one feasible implementation, such as Figure 3 As shown, step S1204 updates the above relationships with confidence weights based on the above efficacy feedback nodes, generating the above depression intervention knowledge graph, including: S12041. Obtain the changes in emotional scores and sleep quality indicators before and after intervention corresponding to the above-mentioned efficacy feedback nodes; S12042. Calculate the intervention effect weighting coefficient based on the above-mentioned changes in emotion scores and sleep quality indicators. S12043. Based on the above intervention effect weighting coefficient, dynamically adjust the association edge weights between the above symptom nodes and the above intervention measure nodes; S12044. Write the updated associated edge weights into the above initial graph structure to generate the above depression intervention knowledge graph with adaptive structure updates.

[0029] For example, the purpose of step S120 is to unify the depression-related information that was originally scattered in medical record text, emotional language interaction, sleep behavior and physiological monitoring into computable graph elements (nodes / edges / weights), and further use efficacy feedback to perform closed-loop calibration of the graph relationship strength, so that the generated depression intervention knowledge graph can not only express the structural association between symptoms, situation and intervention, but also adaptively evolve with the accumulation of efficacy data, thereby providing a high signal-to-noise ratio structural prior for subsequent graph neural network embedding learning and intervention strategy recognition.

[0030] In S1201, natural language processing (NLP) operations are first performed on the medical record text and emotional language interaction data. These NLP operations may include medical entity recognition, negation and degree word parsing, time expression normalization, and relation extraction, in order to identify depressive symptom entities from the text and dialogue. Entities involved in intervention and contextual semantic tags .

[0031] To enhance the clinical usability of entities, this implementation can also generate an attribute vector for each entity. For example, the attribute vector for a symptom entity can be written as follows: ,in Indicates severity (derived from degree terms / scale mapping), freq Indicates frequency of occurrence, It indicates the duration, thus making the symptom entity not only a noun, but also carrying intensity information that can be used for subsequent calculations.

[0032] In step S1202, sleep behavior data and physiological monitoring data are encoded using temporal features to form quantitative indicator nodes that can be aligned with the aforementioned symptom entities. This is done within a preset time window. Internal extraction of sleep structural features Physiological stress characteristics And encode it as a quantitative indicator node. .in, This is the sleep latency period. The duration of wakefulness after falling asleep. Total sleep duration For sleep efficiency; For heart rate variability, Heart rate, Blood oxygen saturation This refers to the respiratory rate.

[0033] To enable the fusion of indices with different dimensions, this implementation method can perform normalization on each index, for example: in, The median. Interquartile range, This represents a truncation function used to reduce the interference of outliers on subsequent weight updates. Through the above encoding, an alignment relationship can be established between quantification indicator nodes and symptom entities; for example, associating early awakening / difficulty falling asleep symptoms with... By establishing connections based on indicators, we can provide an objective chain of evidence for subsequent graph reasoning.

[0034] In S1203, a relational mapping is performed between symptom entities, intervention behavior entities, and contextual semantic labels based on a pre-defined medical knowledge ontology (such as the hierarchy of depressive symptoms, classification of intervention measures, types of contextual factors, and their constraint rules) to construct an initial graph structure. For example, three types of key relational edges can be defined: symptom-context influence edges. Symptom-intervention effect And intervention-situational adaptation To ensure the initial graph has computable strength, this implementation assigns an initial weight to each edge. It can be determined by both ontology priors and data evidence, for example: in, For the Sigmoid function, Indicates the prior strength of the ontology (e.g., a guideline recommendation level mapping). Indicates symptoms With intervention Co-occurrence intensity in historical case / follow-up records These are trainable or configurable coefficients. This design allows the initial atlas to both inherit medical common sense and retain its adaptability to local population data distribution.

[0035] In S1204, the above relationships are updated with confidence weights based on the efficacy feedback nodes to generate the final knowledge graph of depression intervention. The key to this step is to explicitly convert the changes before and after intervention into edge-weighted update signals, enabling the graph to continuously correct which interventions are more effective for which symptoms in which situations. Therefore, when obtaining the changes in mood scores and sleep quality indicators before and after intervention in S12041, standardized improvement measures can be constructed respectively.

[0036] For example, the improvement in mood score can be defined as: The improvement in sleep quality can be defined as: in, and The scores on the emotion scale before and after the intervention are respectively (the higher the score, the more severe the condition, the more severe the condition can be; if the scales are opposite, the directions can be aligned). To prevent extremely small quantities with a denominator of zero; and For the first Normalized values ​​of several sleep indicators before and after the intervention. As the weight of sleep indicators, and This represents the number of sleep indicators. Through this construction, Reflects the proportion of improvement in subjective emotions. This reflects the strength of objective sleep structure improvement, thus providing dual-channel evidence for subsequent weighting coefficient calculation.

[0037] In S12042, to reflect the credibility and transferability of therapeutic feedback, this implementation method introduces an intervention effect weighting coefficient. It simultaneously considers the magnitude of improvement, compliance, and data quality, and can be exemplarily defined as: in, For coefficients; These are compliance indicators (such as training completion rate, daily check-in rate, and percentage of key task executions). This is an uncertainty penalty term used to characterize the reliability of the efficacy assessment. It can be derived from multi-source consistency, for example: in, The scaling-match coefficient between mood improvement and sleep improvement. Tolerance bandwidth. When mood improvement and sleep improvement corroborate each other, consistency cons are higher. Lower, thus making the weighting coefficient It is closer to a credible therapeutic effect; conversely, when the subjective and objective are significantly different, It will be suppressed to prevent erroneous feedback from causing the graph relationships to be over-reinforced.

[0038] In S12043, when dynamically adjusting the association weights between symptom nodes and intervention measure nodes based on the aforementioned intervention effect weight coefficients, to avoid weight oscillations and maintain long-term memory, this implementation can adopt a combined update mechanism of gated exponential sliding update coupled with contextual constraints. An exemplary weight update can be expressed as: in, For the first Symptoms during rotation update With intervention The right to the border, To update the step size gating coefficient, Candidate update values; As a reinforcing factor, it can be composed of both situational fit and the strength of symptom evidence, for example... ,in Indicates intervention The degree of suitability in the current context (such as the feasibility of a certain type of intervention when there is high work pressure and irregular work and rest). Indicates symptoms The strength of the evidence (determined by multimodal consistency); This is to intervene in risk items (such as side effect risk, excessive burden risk, or adverse reaction risk). This is the risk penalty coefficient. The gating coefficient... It can be further designed to adapt to changes in uncertainty, for example: in, The base learning rate is used. Updates are more conservative when feedback uncertainty is high, and more agile when feedback is reliable, thus significantly improving the stability and clinical safety of the atlas weight evolution.

[0039] In S12044, when the updated associated edge weights are written into the initial graph structure to form a knowledge graph for depression intervention that is updated adaptively, this implementation not only updates the edge weights, but also synchronously updates the traceable mapping from the efficacy feedback nodes to the edge weights, so as to facilitate subsequent interpretation and auditing.

[0040] For example, a set of evidence can be maintained for each edge. And record the update evidence of this round simultaneously during writing. This ensures that the final graph meets the requirements of interpretable edge weights, traceable updates, and verifiable evidence. The resulting knowledge graph for depression intervention can continuously absorb new feedback and automatically adjust relationship strength during ongoing patient follow-up. It also provides graph structure inputs that better reflect the actual efficacy distribution for the graph neural network embedding learning in subsequent step S130, thereby improving the personalization and accuracy of the entire depression intervention model training and recognition output.

[0041] In one feasible implementation, such as Figure 4 As shown, step S130 above uses a graph neural network to perform embedding representation learning on the aforementioned depression intervention knowledge graph to obtain a patient state vector, including: S1301. Perform vectorization initialization on the node features and edge weights in the above depression intervention knowledge graph to generate initial node representations; S1302. Based on the graph attention mechanism, the adjacency relationships between the above symptom nodes and the above contextual factor nodes are weighted and aggregated to obtain the first-layer graph representation; S1303. Combine the above-mentioned therapeutic feedback nodes to perform supervised and constrained training on the first-layer graph representation to obtain the updated graph embedding representation. S1304. Perform a patient-level readout operation on the updated graph embedding representation to generate the patient state vector.

[0042] In one feasible implementation, such as Figure 5 As shown, step S1302 above uses a graph attention mechanism to perform weighted aggregation of the adjacency relationships between the symptom nodes and the contextual factor nodes to obtain a first-layer graph representation, including: S13021. Calculate the semantic similarity and temporal correlation between the above symptom nodes and the above contextual factor nodes, and generate multidimensional attention weights. S13022. Based on the above multidimensional attention weights, the features of adjacent nodes are weighted and summed to obtain the local context representation; S13023. Input the above local context representation into a nonlinear activation function to perform feature transformation, and obtain the node update vector; S13024. The above node update vectors are aggregated to form the first layer graph representation.

[0043] In one feasible implementation, such as Figure 6 As shown, step S1303 above, combined with the above-mentioned efficacy feedback node, performs supervised and constrained training on the above-mentioned first-layer graph representation to obtain an updated graph embedding representation, including: S13031. Construct a graph embedding loss function with the above-mentioned efficacy feedback nodes as supervision labels, wherein the above-mentioned graph embedding loss function includes an intervention effect prediction error term and a node relationship preservation constraint term; S13032. Based on the backpropagation algorithm, the parameters of the first layer graph representation above are updated and trained to obtain the optimized node embedding vector; S13033. The optimized node embedding vector is regularized to enhance the generalization ability across patient samples. S13034. Output the updated graph embedding representation after supervised constraint optimization.

[0044] For example, in this feasible implementation, the purpose of step S130 is to transform the depression intervention knowledge graph generated in step S120 into a vectorized representation that the model can directly process, so that the subsequent intervention strategy generation model can comprehensively identify the patient's current depressive state, objective evidence of sleep / physiology, and situational constraints based on the patient's state vector. To this end, this implementation uses a graph neural network as the core, and forms an interpretable, trainable, and iteratively optimized embedding learning process through node and edge initialization, attention-weighted aggregation, efficacy feedback supervision and constraints, and patient-level readout.

[0045] First, in step S1301, the node features and edge weights in the depression intervention knowledge graph are vectorized and initialized to generate initial node representations. For example, for any node... (These can be symptom nodes, intervention node, situational factor node, or efficacy feedback node), construct its initial feature vector. This can include entity type embeddings, symptom intensity attributes, sleep / physiological quantitative indicators, time window statistics, etc. For either side to give its border rights and edge type embedding Merge into edge features To reflect the prior constraints of edge weights on message passing, this implementation can introduce edge weights as a gating term into the edge-aware enhancement of node initialization, defining the initial node as: in, For nodes The initial node representation, For a trainable parameter matrix, For nodes The set of adjacent nodes, This is the edge weight gating function.

[0046] For example, the edge weight gating function can be taken as... ,in This is a smoothing constant used to avoid gradient vanishing caused by small edge weights. This allows high-confidence relations to have a stronger impact on node representations during the initialization phase, enhancing the efficiency of using graph priors in embedding learning.

[0047] Subsequently, in step S1302, the adjacency relationships between symptom nodes and contextual factor nodes are weighted and aggregated based on a graph attention mechanism to obtain a first-layer graph representation. Specifically, in S13021, to simultaneously characterize semantic proximity and temporal co-occurrence strength, this embodiment constructs multi-dimensional attention weights, treating semantic similarity and temporal correlation as two independent but fusionable attention channels. For example, symptom nodes can be defined... Contextual Factor Nodes The semantic similarity is: in, The node semantic vectors are obtained by the text semantic encoder; the temporal correlation is defined as follows: in, and These are the time of symptom onset and the time of the situational event, respectively. The time decay coefficient controls the penalty strength of the time difference on the correlation. Based on the two types of scores mentioned above, a multi-dimensional attention weight vector is generated in S13021. Furthermore, it can be compressed into final attention coefficients through trainable fusion gating: in, For trainable fusion coefficients, The confidence edge weights for the symptom-context edge are given. Indicates at symptom nodes The neighborhood is normalized to ensure that the attention weights are comparable and stable. Unlike the traditional method of calculating attention using only node features, this implementation explicitly couples the temporal correlation and graph confidence edge weights into the attention calculation. This allows the aggregation process to simultaneously satisfy the consistency of medical semantics, the consistency of follow-up time, and the prior consistency formed by efficacy feedback, resulting in stronger creativity and clinical interpretability.

[0048] In S13022, the local context representation is obtained by weighted summation of the features of adjacent nodes based on the above multidimensional attention weights, which can be exemplarily written as: in, Symptom nodes Local context representation, This is a trainable transformation matrix. In S13023, the local context representation is input to the nonlinear activation function for feature transformation to obtain the node update vector, which can be exemplarily written as: in, Update the vector for the node. This represents vector concatenation. For a trainable parameter matrix, This is a non-linear activation function (e.g., ReLU or GELU). In S13024, the update vectors of each node are aggregated to form the first-layer graph representation. For example, the update results of the symptom subgraph and the context subgraph can be combined to form this representation. ,in This corresponds to the step of completing the weighted aggregation to obtain the first-level graph representation, which is a set of graph nodes.

[0049] Next, in step S1303, supervised training is performed on the first-layer graph representation in conjunction with the efficacy feedback nodes to obtain an updated graph embedding representation. Unlike unsupervised objectives (such as adjacency reconstruction) alone, this embodiment introduces efficacy feedback as a supervisory signal into the embedding learning, making the embedding space sensitive to the effectiveness of the intervention, thereby improving the usability of subsequent policy identification. In S13031, a graph embedding loss function is constructed with efficacy feedback nodes as supervisory labels. This loss function includes an intervention effect prediction error term and a node relationship preservation constraint term. For example, the overall improvement amount in the efficacy feedback can be used as the supervisory label. and define the predicted value For the intervention effect estimation obtained from the embedding readout: in, These are the convergent embeddings of symptom-related nodes, intervention nodes, and context nodes, respectively, with MLP() being the multilayer perceptron prediction head. The intervention effect prediction error term can then be defined as: Meanwhile, to avoid the supervision signal from destroying the semantics of the graph structure, this implementation introduces a node relationship preservation constraint, which ensures that nodes corresponding to high-confidence edges remain close in the embedding space, while low-confidence edges remain relatively separate. An exemplary definition is: in, For the set of positive edges of the graph, For the set of negative sampling edges, The interval threshold is used to separate nodes that should not be connected. Therefore, the graph embedding loss function can be written as: in, These are the loss weight coefficients. The supervision labels for the above loss function come from the efficacy feedback nodes, allowing the embedding learning to directly serve the intervention effectiveness modeling. Simultaneously, edge weights are utilized. Weighting the strength of relationships allows confidence updates driven by therapeutic feedback to influence the embedded spatial structure in a controllable manner.

[0050] In S13032, the first-layer graph representation is trained by backpropagation algorithm to update parameters and obtain the optimized node embedding vector, i.e., by minimizing the above... Update the attention fusion coefficients, transformation matrix, and prediction head parameters to obtain the optimized embedding. In S13033, the optimized node embedding vector is regularized to enhance generalization across patient samples. An exemplary approach is to use a combination of unit ball normalization and discard perturbation regularization, defined as follows: in, To prevent extremely small division by zero, random dropout noise is applied to the node embeddings during the training phase to improve robustness. Finally, in S13034, the updated graph embedding representation optimized by supervised constraints is output. This corresponds to the step of obtaining the updated graph embedding representation through supervised training.

[0051] Finally, in step S1304, a patient-level readout operation is performed on the updated graph embedding representation to generate a patient state vector. For example, to ensure that the readout results simultaneously reflect symptom dominance and context modulation, this implementation may employ symptom-context dual-channel gated readout to form the patient-level vector, defined as follows: in, For the set of symptom nodes, For the set of context nodes, To read out the weights; further define the gating coefficients: And based on this, a patient state vector is generated: in, Let the patient state vector be... For element-wise multiplication, This is the trainable parameter matrix. The above formula is not a simple average convergence, but explicitly distinguishes between the symptom subspace and the context subspace, and adaptively determines whether the current patient state is more driven by symptoms or more influenced by the context through gating coefficients, thereby making the output patient state vector more sensitive and stable for the identification of individualized intervention strategies.

[0052] This implementation completes the embedding representation learning process from the knowledge graph of depression intervention to the patient state vector, and lays an interpretable and optimizable feature foundation for the subsequent step S140 to identify intervention strategies based on the patient state vector.

[0053] In one feasible implementation, such as Figure 7 As shown, step S140 above identifies the patient state vector input into the intervention strategy generation model to generate personalized intervention plan prediction results for different patient states, including: S1401. The above patient state vector and the context features corresponding to the above situational factor nodes are fused and encoded to form the strategy input features; S1402. Input the above-mentioned strategy input features into the intervention strategy generation model based on reinforcement learning to generate a sequence of candidate intervention schemes; S1403. Based on the historical intervention effects corresponding to the above-mentioned efficacy feedback nodes, calculate the reward function for the above-mentioned candidate intervention scheme sequence, and update the parameters of the above-mentioned intervention strategy generation model. S1404. Output the predicted results of the above personalized intervention plan that match the current patient status.

[0054] In one feasible implementation, such as Figure 8 As shown, step S1402 above inputs the above-mentioned policy input features into the reinforcement learning-based intervention policy generation model to generate a sequence of candidate intervention schemes, including: S14021. Map the above policy input features into state vectors in the policy state space. S14022. Based on the policy network, the above state vector is sampled for action to generate multiple combinations of intervention actions; S14023. Based on the constraint rules corresponding to the above-mentioned situational factor nodes, the legality of the above-mentioned combination of intervention actions is screened to form candidate intervention paths; S14024. Encode the above candidate intervention paths in chronological order to generate the above candidate intervention scheme sequence.

[0055] In one feasible implementation, such as Figure 9 As shown, step S1403 above calculates the reward function for the candidate intervention program sequence based on the historical intervention effects corresponding to the above efficacy feedback nodes, and updates the parameters of the above intervention strategy generation model, including: S14031. Construct a multi-objective reward function based on the degree of symptom improvement, the extent of sleep quality improvement, and compliance indicators in the above-mentioned efficacy feedback nodes; S14032. Calculate the return value for the above candidate intervention scheme sequence and generate a strategy gradient update signal; S14033. Based on the above policy gradient update signal, iteratively optimize the policy network parameters of the above intervention policy generation model; S14034. Output the above intervention strategy generation model after parameter update.

[0056] For example, in this feasible implementation, step S140, based on the patient state vector obtained in step S130, introduces executability constraints of contextual factors and long-term benefit information from efficacy feedback. A closed-loop identification process is achieved through a reinforcement learning-based intervention strategy generation model: candidate solution generation—legitimacy screening—multi-objective reward evaluation—strategy parameter iteration, thereby outputting a personalized intervention plan prediction result matching the current patient state. Unlike traditional one-time template recommendations, this implementation describes intervention decision-making as a sequential decision problem, enabling the model to simultaneously consider short-term symptom relief, sleep improvement, and patient compliance, and continuously optimize the generation strategy driven by historical efficacy feedback.

[0057] In step S1401, the patient state vector and the context features corresponding to the situational factor nodes are fused and encoded to form the policy input features. The patient state vector... Readouts from graph embeddings are used to characterize the patient's depressive symptom structure, sleep / physiological evidence, and sensitivity to treatment efficacy. Contextual features. This is obtained by embedding the set of situational factor nodes and aggregating rule attributes, and is used to represent constraints such as current work pressure, work-rest regularity, social support, available resources, and risk taboos. To enhance creativity and avoid information dilution caused by simple splicing, this implementation method can adopt "gated-interactive" fusion coding. An exemplary strategy input feature can be defined as: in, Input features for the strategy, This represents vector concatenation. This represents element-wise multiplication. For the Sigmoid function, For a trainable parameter matrix, To fuse gated vectors, an interaction term is introduced. By integrating with gating, the model can automatically learn which symptom features need to be emphasized in the current context, thereby providing a more stable state representation for subsequent policy generation.

[0058] In step S1402, the policy input features are input into a reinforcement learning-based intervention policy generation model to generate a sequence of candidate intervention schemes. Specifically, in S14021, the policy input features are mapped to state vectors in the policy state space, which can be implemented, for example, through a state encoder: in, For a moment The state vector, The state encoding function can encode time information such as intervention stage and cycle progress, thus giving the strategy a phased nature. In S14022, action sampling is performed on the state vector based on the policy network to generate multiple intervention action combinations. To enhance creativity and reflect the multidimensional prescription attributes of medical intervention, this implementation designes the actions as a multi-branch structure, corresponding to intervention type, intensity, and frequency, and adopts a differentiable branch sampling mechanism. An exemplary definition is as follows: in, For parameters The policy network, This indicates the type of intervention (such as cognitive behavioral training, sleep hygiene, exercise prescription, etc.). Indicates the intensity of the intervention (e.g., training duration / difficulty level). This indicates the frequency of the action (e.g., the number of times it is performed daily / weekly). This branch action structure naturally gives candidate solutions executable prescription elements, avoiding the output of only abstract suggestions.

[0059] In S14023, the legality of intervention action combinations is screened based on the constraint rules corresponding to the contextual factor nodes to form candidate intervention paths. To avoid black-box strategies outputting unexecutable or high-risk solutions, this implementation introduces a rule-constrained projection operator. Project the action combination onto the feasible set In this context, an example can be represented as: in, The set of possible actions is determined by both contextual features and taboo rules. This is a safe action after legitimacy screening. Through this projective screening, the model retains the learning ability of the policy network while meeting clinical controllability and safety constraints. Subsequently, in S14024, the candidate intervention paths are encoded in chronological order to generate a sequence of candidate intervention plans, which can be exemplarily written as: in, That is, the sequence of candidate intervention protocols. The length of the planning period.

[0060] Next, in step S1403, reward functions are calculated for candidate pre-planned sequences based on historical intervention effects corresponding to efficacy feedback nodes, and parameters of the intervention strategy generation model are updated so that the strategy can continuously approach the optimal intervention combination based on historical data and online feedback. Specifically, in S14031, a multi-objective reward function is constructed based on the degree of symptom improvement, the magnitude of sleep quality improvement, and compliance indicators in the efficacy feedback nodes. The reward function proposed in this embodiment represents the four-element coupling of "efficacy-sleep-compliance-risk" and introduces edge-weighted confidence as a reward credibility modulation term. An exemplary immediate reward can be defined as: in, For a moment The amount of symptom improvement To improve sleep quality, As a compliance indicator, The risk and cost associated with the action; For dynamic weights, The weighted temperature coefficient; The confidence modulation factor. Used for graph edge weights The average confidence level of this action on the knowledge graph evidence chain is obtained by aggregation; For action The corresponding symptom-intervention evidence edge set. The reward weight of the above reward function is not a fixed hyperparameter, but is adaptively modulated by the confidence of the knowledge graph, so that intervention actions with more sufficient evidence obtain more stable gradient signals in policy learning, while actions with greater feedback noise have their impact on updates automatically suppressed.

[0061] In S14032, the reward value is calculated for the candidate intervention sequence and a policy gradient update signal is generated. A discounted reward can be used as an example. in, From time Initial cumulative returns, This is a discount factor. To reduce variance and improve convergence stability, this implementation may further introduce a value network baseline. The advantage function is obtained as follows: in, For parameters The value network. In S14033, the policy network parameters are iteratively optimized based on the policy gradient update signal. An exemplary update objective is: Furthermore, a pruning method optimized using a proximal strategy can be employed to enhance training stability. in, The cropping threshold, These are the parameters for the previous round of strategy. Finally, in S14034, the intervention strategy generation model with updated parameters is output to complete the parameter update step.

[0062] Finally, in step S1404, the predicted personalized intervention plan that matches the current patient status is output. For example, the plan with the highest expected return can be selected as the output from the sequence of candidate intervention plans: And the optimal solution Decode the intervention into executable intervention prescription elements (intervention type, intensity, frequency, stage goals and risk warnings), thereby enabling the prediction and output of personalized intervention plans for different patient conditions.

[0063] Through the above process, this implementation method organically combines knowledge graph embedding representation, context-constrained controllable screening, and efficacy feedback-driven reinforcement learning updates, enabling the intervention strategy generation model to output personalized and executable intervention sequences, and to continuously absorb efficacy feedback and adaptively optimize in long-term use, thereby significantly improving the accuracy and stability of depression intervention and sleep-related symptom improvement.

[0064] like Figure 10 As shown, this application proposes a training system 10 for a depression intervention model based on deep knowledge graphs, used to execute the training method for the depression intervention model based on deep knowledge graphs described in the first aspect, including: The acquisition unit 101 is used to acquire multimodal data of the target patient, wherein the multimodal data includes medical record text, emotional language interaction data, sleep behavior data and physiological monitoring data; Construction unit 102 is used to perform semantic parsing and entity relation extraction on the above multimodal data to construct a depression intervention knowledge graph, wherein the above depression intervention knowledge graph includes symptom nodes, intervention measure nodes, situational factor nodes and efficacy feedback nodes; Embedding unit 103 is used to perform embedding representation learning on the above-mentioned depression intervention knowledge graph based on graph neural network to obtain patient state vector; The identification unit 104 is used to input the above-mentioned patient state vector into the intervention strategy generation model for identification, so as to generate personalized intervention plan prediction results for different patient states. Training unit 105 is used to conduct rehabilitation training for different patients based on the prediction results of the above-mentioned personalized intervention plan.

[0065] The above 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.

Claims

1. A training method for a depression intervention model based on deep knowledge graphs, characterized in that, include: Acquire multimodal data of the target patient, wherein the multimodal data includes medical record text, emotional language interaction data, sleep behavior data, and physiological monitoring data; Semantic parsing and entity relation extraction are performed on the multimodal data to construct a depression intervention knowledge graph, wherein the depression intervention knowledge graph includes symptom nodes, intervention measure nodes, situational factor nodes, and efficacy feedback nodes; The patient state vector is obtained by embedding representation learning of the depression intervention knowledge graph based on graph neural network; The patient state vector is input into the intervention strategy generation model for identification, so as to generate personalized intervention plan prediction results for different patient states; Rehabilitation training is conducted for different patients based on the prediction results of the personalized intervention program.

2. The training method for the depression intervention model based on deep knowledge graphs according to claim 1, characterized in that, The step of performing semantic parsing and entity relation extraction on the multimodal data to construct a knowledge graph for depression intervention includes: Natural language processing operations are performed on the medical record text and the emotional language interaction data to identify depressive symptom entities, intervention behavior entities, and contextual semantic labels; The sleep behavior data and the physiological monitoring data are encoded with temporal features to generate quantitative indicator nodes corresponding to the symptom entities; Based on a pre-defined medical knowledge ontology, the symptom entities, the intervention behavior entities, and the contextual semantic tags are mapped to form an initial graph structure. The above relationships are updated with confidence weights based on the efficacy feedback nodes to generate the depression intervention knowledge graph.

3. The training method for the depression intervention model based on deep knowledge graphs according to claim 2, characterized in that, The step of updating the above relationships with confidence weights based on the efficacy feedback nodes to generate the depression intervention knowledge graph includes: Obtain the changes in emotional scores and sleep quality indicators before and after the intervention corresponding to the therapeutic feedback nodes; The intervention effect weighting coefficient is calculated based on the change in the emotion score and the change in the sleep quality index. The association edge weights between the symptom node and the intervention measure node are dynamically adjusted based on the intervention effect weight coefficient. The updated associated edge weights are written into the initial graph structure to generate the depression intervention knowledge graph with adaptive structure updates.

4. The training method for the depression intervention model based on deep knowledge graphs according to claim 1, characterized in that, The process of embedding and learning the depression intervention knowledge graph based on a graph neural network to obtain a patient state vector includes: The node features and edge weights in the depression intervention knowledge graph are vectorized and initialized to generate initial node representations; The adjacency relationships between the symptom nodes and the contextual factor nodes are weighted and aggregated based on the graph attention mechanism to obtain the first-layer graph representation; The first layer graph representation is trained under supervised constraints using the therapeutic effect feedback node to obtain an updated graph embedding representation; A patient-level readout operation is performed on the updated graph embedding representation to generate the patient state vector.

5. The training method for the depression intervention model based on deep knowledge graphs according to claim 4, characterized in that, The graph attention mechanism performs weighted aggregation of the adjacency relationships between the symptom nodes and the contextual factor nodes to obtain a first-layer graph representation, including: Calculate the semantic similarity and temporal correlation between the symptom node and the contextual factor node to generate multidimensional attention weights; The local context representation is obtained by weighting and summing the features of adjacent nodes based on the multidimensional attention weights. The local context representation is input into a nonlinear activation function for feature transformation to obtain the node update vector; The node update vectors are aggregated to form the first layer graph representation.

6. The training method for the depression intervention model based on deep knowledge graphs according to claim 4, characterized in that, The step of supervising and constraining the training of the first-layer graph representation in conjunction with the therapeutic effect feedback node to obtain the updated graph embedding representation includes: Construct a graph embedding loss function with the efficacy feedback node as the supervision label, wherein the graph embedding loss function includes an intervention effect prediction error term and a node relationship preservation constraint term; The first layer graph representation is trained by backpropagation algorithm to update parameters and obtain the optimized node embedding vector. The optimized node embedding vector is regularized to enhance its generalization ability across patient samples. Output the updated graph embedding representation after supervised constraint optimization.

7. The training method for the depression intervention model based on deep knowledge graphs according to claim 1, characterized in that, The step of inputting the patient state vector into the intervention strategy generation model for identification, in order to generate personalized intervention plan prediction results for different patient states, includes: The patient state vector is fused and encoded with the context features corresponding to the context factor nodes to form the strategy input features; The strategy input features are input into a reinforcement learning-based intervention strategy generation model to generate a sequence of candidate intervention schemes; The reward function is calculated for the candidate intervention scheme sequence based on the historical intervention effects corresponding to the efficacy feedback nodes, and the parameters of the intervention strategy generation model are updated. Output the predicted results of the personalized intervention plan that match the current patient status.

8. The training method for the depression intervention model based on deep knowledge graphs according to claim 7, characterized in that, The step of inputting the policy input features into a reinforcement learning-based intervention policy generation model to generate a sequence of candidate intervention schemes includes: The policy input features are mapped to state vectors in the policy state space; The state vector is sampled based on a policy network to generate multiple combinations of intervention actions. Based on the constraint rules corresponding to the contextual factor nodes, the combination of intervention actions is screened for legality to form candidate intervention paths; The candidate intervention paths are encoded in chronological order to generate the candidate intervention scheme sequence.

9. The training method for the depression intervention model based on deep knowledge graphs according to claim 7, characterized in that, The step of calculating a reward function for the candidate intervention sequence based on the historical intervention effects corresponding to the efficacy feedback nodes, and updating the parameters of the intervention strategy generation model, includes: A multi-objective reward function is constructed based on the degree of symptom improvement, the extent of sleep quality improvement, and compliance indicators in the therapeutic feedback nodes. The return value is calculated for the candidate intervention sequence, and a policy gradient update signal is generated; The policy network parameters of the intervention policy generation model are iteratively optimized based on the policy gradient update signal. Output the intervention strategy generation model after parameter updates.

10. A training system for a depression intervention model based on deep knowledge graphs, used for the training method of the depression intervention model based on deep knowledge graphs as described in any one of claims 1 to 9, characterized in that, include: The acquisition unit is used to acquire multimodal data of the target patient, wherein the multimodal data includes medical record text, emotional language interaction data, sleep behavior data, and physiological monitoring data; The construction unit is used to perform semantic parsing and entity relation extraction on the multimodal data to construct a depression intervention knowledge graph, wherein the depression intervention knowledge graph includes symptom nodes, intervention measure nodes, situational factor nodes, and efficacy feedback nodes; An embedding unit is used to perform embedding representation learning on the depression intervention knowledge graph based on a graph neural network to obtain a patient state vector; The identification unit is used to input the patient state vector into the intervention strategy generation model for identification, so as to generate personalized intervention plan prediction results for different patient states. The training unit is used to conduct rehabilitation training for different patients based on the prediction results of the personalized intervention plan.

Images