Intelligent generation method and system of traditional chinese medicine nursing scheme based on relationship constraint gate and efficacy feedback adaptive update

Abstract

The application discloses a traditional Chinese medicine nursing scheme intelligent generation method and system based on relationship constraint gating and curative effect feedback adaptive updating. The method constructs a traditional Chinese medicine nursing knowledge graph containing multiple types of relationships; constructs a nursing demand subgraph according to patient information and generates a patient state vector; through a relationship constraint gating graph neural network, different relationships are subjected to differentiated propagation strategies for node representation learning, including positive enhancement propagation for suitable / enhanced relationships, attenuation propagation for restricted relationships, negative punishment propagation for conflict relationships and hard shielding propagation for taboo relationships; based on the learned representation, candidate measures are recalled and a combined scheme is generated, and the optimal scheme is selected by using a scoring function that fuses matching points, cooperation points, penalty points and risk points; and relationship weights, gating parameters and scoring parameters are adaptively updated according to curative effect feedback. The application can generate an individualized scheme that conforms to traditional Chinese medicine nursing constraints and is safe and interpretable, and has closed-loop optimization capability.

Description

Technical Field

[0001] This invention belongs to the field of medical information technology, specifically relating to a method and system for intelligent generation of TCM nursing plans based on relational constraint gating and adaptive updating of efficacy feedback. Background Technology

[0002] The development of TCM nursing plans typically requires a comprehensive consideration of the patient's chief complaint, present medical history, past medical history, tongue and pulse information, constitution type, syndrome characteristics, lifestyle, emotional state, and previous nursing experience. This involves a combination of individualized nursing measures across multiple dimensions, including dietary care, daily routine care, emotional well-being, functional exercises, and acupoint care. However, existing TCM nursing support systems suffer from the following main problems:

[0003] 1. Reliance on human experience or static rule bases. Most systems use manually defined rules, keyword matching, or template-based recommendations, which makes it difficult to express the complex relationships between "symptoms-syndrome-constitution-nursing measures-treatment feedback".

[0004] 2. Insufficient knowledge representation capabilities. Traditional relational databases or rule tables struggle to represent the multi-step, multi-type constraints, and dynamic feedback relationships in TCM nursing knowledge.

[0005] 3. Deep learning models lack constraints related to Traditional Chinese Medicine (TCM) nursing. While purely data-driven models can learn statistical patterns, they struggle to explicitly incorporate constraints such as appropriateness, enhancement, limitation, conflict, and contraindications in TCM nursing, potentially leading to inconsistent or unsafe nursing plans.

[0006] 4. Existing recommendation methods lack closed-loop optimization capabilities. Most existing methods are based on static training data and output fixed results, making it difficult to continuously update subsequent plans using information on changes in efficacy, compliance, and adverse reactions after nursing implementation.

[0007] 5. Directly applying general knowledge graph recommendation or graph neural network methods suffers from scenario mismatch. General recommendation methods typically treat entity relationships as homogeneous or weakly constrained edges, focusing primarily on similarity propagation or association strength propagation. This makes it difficult to handle hierarchical relationship constraints in TCM nursing scenarios, such as "recommended but limited in quantity," "combinable but with temporal restrictions," "suppressed under specific constitutions," and "hard-shielded due to clear contraindications." Furthermore, general feedback learning methods usually only uniformly correct the final output score, failing to differentiate the varying impacts of efficacy improvement, adverse reactions, and compliance on different relationship types. This makes it difficult to meet the requirements of TCM nursing protocol generation for constraint consistency, safety, and phased adjustment.

[0008] Therefore, a technical solution is needed that can uniformly express the positive and negative relationships and efficacy feedback in TCM nursing, and explicitly introduce relationship constraints, gating regulation, and adaptive update mechanisms during the training and generation process. Summary of the Invention

[0009] The purpose of this invention is to provide an intelligent generation method and system for TCM nursing plans based on relational constraint gating and adaptive updating of efficacy feedback, to solve the problems of insufficient knowledge representation ability, inability to handle complex relational constraints, and lack of closed-loop optimization in existing technologies. This invention can structurally express various types of relational and constraint information in TCM nursing, explicitly handle positive synergy and negative conflict relationships during the graph representation learning stage, generate individualized nursing plans that take into account matching, synergy, safety, and interpretability, and adaptively update relational weights, gating parameters, and plan scoring parameters based on efficacy feedback, forming a closed-loop optimization mechanism of "knowledge modeling—plan generation—efficacy feedback—adaptive updating".

[0010] To achieve the above objectives, the present invention adopts the following technical solution: a method for intelligently generating TCM nursing plans based on relational constraint gating and adaptive updating of efficacy feedback, comprising the following steps:

[0011] Step S1, Data Acquisition and Preprocessing: Acquire and process the patient's multimodal data, which includes at least symptoms, signs, physical condition, syndrome, and nursing goals;

[0012] Step S2, Knowledge Graph Construction: Construct a TCM nursing knowledge graph, which includes multiple nodes and multiple edges connecting the nodes. The nodes include at least symptom nodes, constitution nodes, syndrome nodes, and nursing intervention nodes. The edges represent the relationships between the nodes, including at least appropriate relationships, reinforcing relationships, restrictive relationships, conflict relationships, and taboo relationships.

[0013] Step S3, Demand State Encoding: Encode the acquired multimodal patient data into a patient demand state vector;

[0014] Step S4, Relationship Constraint Gated Graph Learning: The patient's demand state vector and the knowledge graph are input into a relationship constraint gated graph neural network to obtain the final representation of the nursing intervention nodes. The relationship constraint gated graph neural network employs differentiated propagation strategies for different types of relationships, specifically including: positive reinforcement propagation for appropriate and reinforcing relationships; decay propagation for restrictive relationships; negative penalty propagation for conflicting relationships; and hard shielding propagation for taboo relationships.

[0015] Step S5, Nursing Needs Subgraph Construction and Candidate Measure Recall: Map the nodes corresponding to the patient's need state vector to the knowledge graph to construct a nursing needs subgraph; based on the nursing needs subgraph, recall candidate nursing measures from the knowledge graph, and mask the recalled candidate nursing measures according to taboo relationships;

[0016] Step S6, Solution Generation: Combine the recalled candidate nursing measures to form multiple candidate solutions, and calculate the comprehensive score of each candidate solution based on the final representation of the nursing measure node. Select the solution with the highest comprehensive score as the optimal nursing solution output.

[0017] Step S7, Feedback Update: Obtain the efficacy feedback signal after the implementation of the nursing plan, and use the feedback signal to adaptively update the relation weights in the knowledge graph, the gating parameters of the relation constraint gating graph neural network, and the scoring parameters of the comprehensive plan score.

[0018] Furthermore, in step S1, the acquired multimodal data originates from clinical case data, traditional Chinese medicine (TCM) knowledge data, and protocol and feedback data. The clinical case data includes TCM nursing-related electronic medical records, nursing assessment forms, nursing record forms, nursing plan forms, progress notes, discharge summaries, follow-up records, nursing execution records, and interim efficacy evaluation records. The TCM knowledge data includes TCM nursing textbooks, TCM nursing clinical pathways, expert consensus and industry guidelines, classic TCM works, and hospital internal TCM nursing standards. The protocol and feedback data includes historically used nursing protocols, expert-reviewed recommended protocols, nursing compliance records, efficacy improvement records, and adverse reaction records.

[0019] The data collection uses one nursing assessment cycle, one hospitalization period, or one follow-up cycle as a sample unit, and the patient's nursing process is divided into multiple stage samples according to time windows.

[0020] The method uses supervision tags derived from historically implemented and effective TCM nursing protocols, and / or standard protocols provided by senior nursing experts after reviewing historical cases. When the historical implementation protocol differs from the expert-reviewed protocol, a weighted fusion approach is used to construct the supervision tags.

[0021]

[0022] in, This represents the supervision label of the i-th sample. This represents the historical execution plan for the i-th sample. Let represent the expert review scheme for the i-th sample. , These are the weighting coefficients. .

[0023] Furthermore, in step S2, for any relation type Next node With nodes The initial relation weight of the relation edge is defined as follows:

[0024]

[0025] in, Representing entities , Normalized co-occurrence frequency values ​​in case studies and literature This indicates the normalized intensity value marked by experts. This represents the normalized value of historical efficacy support. The weighting coefficients are and satisfy the following conditions: ;

[0026] Furthermore, in step S7, the update of the relation weights adopts the following feedback update formula:

[0027]

[0028] in, This represents the updated relation weight. This indicates the relation weight before the update. This represents the update rate corresponding to relation type r. Represents an update mapping function specific to the relation type. This represents the feedback support function.

[0029] Furthermore, in step S4, the differentiated propagation strategy is implemented through a relation gating value, which is jointly determined by the relation type, the representation of adjacent nodes, and the patient's demand state vector; specifically, node j uses relation r to determine the relation gating value of node i. Calculated using the following formula:

[0030]

[0031] in, and Let be the l-th layer representations of nodes i and j, respectively, and q be the patient demand state vector. and Let be the gate weight matrix and gate bias vector corresponding to relation r, respectively, and σ be the Sigmoid activation function.

[0032] Furthermore, in step S4, the (l+1)th layer representation of node i Update using the following formula:

[0033]

[0034] in, It is a set of positive relationships that includes suitable relationships and enhancing relationships. It is a set of negative relations that includes restrictive, conflicting, and taboo relations. Let i be the set of neighbors of node i in relation r. To consider the effective gating value after taboo hard shielding, Let r be the transformation matrix corresponding to relation r in the l-th layer. Let l be the self-connection transformation matrix corresponding to the l-th layer. This is the normalization coefficient.

[0035] Furthermore, in step S5, the recalled candidate care measures are screened out based on contraindications, specifically as follows:

[0036] When the weight of the taboo relationship between any candidate nursing measure and any taboo node is greater than a preset taboo threshold, the nursing measure is removed from the candidate set; furthermore, in the learning of the relation constraint gating graph, for nursing measures that meet the taboo judgment conditions, their effective gating value is... It is set directly to zero to achieve hard-block propagation.

[0037] Furthermore, in step S6, the comprehensive score is calculated based on a weighted average of the matching score, synergy enhancement score, restriction penalty score, conflict penalty score, and risk penalty score.

[0038]

[0039] in, This indicates the overall score; The matching score is used to characterize the degree of matching between the candidate treatment and the patient's needs; The synergistic enhancement score is used to characterize the strength of the enhancing relationship between nursing interventions within the candidate protocol. The penalty score is used to characterize the penalty value of nursing measures that are in a restricted state in the candidate plan; The conflict penalty score is used to characterize the strength of the conflict relationship between nursing measures within the candidate options. Risk penalty score is used to characterize the strength of the association between nursing interventions and patient contraindications in candidate protocols; This refers to the scoring weighting coefficient;

[0040] The optimal nursing plan is a structured plan, which includes the following fields: target patient information, corresponding symptoms and constitution, list of nursing measures, frequency and timing of implementation, precautions, contraindications, and explanation path of the plan; the explanation path of the optimal nursing plan is presented in a chain structure of "symptoms / signs → syndrome / constitution → nursing measures → evidence supporting efficacy".

[0041] Furthermore, step S7 specifically includes:

[0042] The therapeutic feedback signal is decomposed into an improvement component, a compliance component, and an adverse reaction component;

[0043] Based on the positive changes in the improvement degree component and compliance component, increase the relation weights of relevant appropriate relations and enhanced relations in the knowledge graph;

[0044] Increase the relational weights of relevant limiting and conflicting relationships based on the occurrence or enhancement of adverse reaction components;

[0045] The gating parameters of the relation-constrained gating graph neural network are adjusted using gradient updates.

[0046] Based on the correlation between the efficacy feedback signal and the contribution of each scoring item, the scoring weight coefficient of the overall treatment plan score is adjusted.

[0047] Furthermore, the relation-constrained gated graph neural network is optimized using a joint loss function during the training phase. The joint loss function is:

[0048]

[0049] in, Losses due to relationship rebuilding To predict losses for weighted care plans, To punish losses in the event of conflict To avoid shielding the loss, To provide feedback on consistency loss, This is the loss weighting coefficient.

[0050] This invention also provides an intelligent system for generating TCM nursing plans to implement the above method, comprising:

[0051] The data acquisition and preprocessing module is used to acquire and process patients' multimodal data;

[0052] The knowledge graph construction module is used to build a knowledge graph for traditional Chinese medicine nursing.

[0053] The demand state encoding module is used to encode the acquired multimodal patient data into a patient demand state vector;

[0054] The relation constraint gating graph learning module is used to input the patient's demand state vector and the knowledge graph into the relation constraint gating graph neural network to obtain the final representation of the nursing intervention node; wherein, the relation constraint gating graph neural network adopts a differentiated propagation strategy for different types of relations;

[0055] The nursing needs subgraph construction module is used to map the nodes corresponding to the patient needs state vector to the knowledge graph to construct the nursing needs subgraph.

[0056] The candidate measure recall module is used to recall candidate nursing measures from the knowledge graph based on the nursing needs subgraph, and to screen the recalled candidate nursing measures according to the taboo relationship;

[0057] The solution generation module is used to combine the recalled candidate nursing measures to form multiple candidate solutions, and calculate the comprehensive score of each candidate solution based on the final representation of the nursing measure node, and select the solution with the highest comprehensive score as the optimal nursing solution output.

[0058] The feedback update module is used to obtain efficacy feedback signals after the implementation of the nursing plan, and to adaptively update the relation weights, the gating parameters of the relation constraint gating graph neural network, and the scoring parameters of the comprehensive plan score using the feedback signals.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] 1. This invention integrates relationships such as suitability, enhancement, limitation, conflict, contraindication, and feedback in TCM nursing knowledge into knowledge graph modeling, thereby improving the ability to express knowledge.

[0061] 2. This invention introduces a relational constraint gating mechanism into the graph neural network, which enables positive and negative relations to be propagated and learned in a differentiated manner, thereby improving the consistency and safety of nursing plans.

[0062] 3. This invention employs attenuation, negative penalty, and hard shielding for restriction, conflict, and taboo relationships, respectively, which differs from the uniform weighted propagation method in general graph recommendations.

[0063] 4. This invention incorporates the patient's current symptoms, physical condition, syndrome, nursing goals, and nursing stages into gating calculations to achieve patient-state-driven individualized solution generation.

[0064] 5. This invention comprehensively considers matching, synergy, constraints, conflicts, and risks by combining scoring functions, making the output scheme more structurally reasonable.

[0065] 6. This invention enables the system to continuously learn and perform closed-loop optimization by adaptively updating the relationship weights, gating parameters, and scoring parameters through efficacy feedback.

[0066] 7. The output results of this invention include an explanation path, which improves the interpretability and clinical acceptability of the nursing plan.

[0067] 8. This invention effectively avoids simplifying the generation of TCM nursing plans into a general recommendation problem, and improves the adaptability to specific nursing constraints. Attached Figure Description

[0068] Figure 1 This is a flowchart of the overall process of the intelligent generation method of TCM nursing plan based on relational constraint gating and adaptive update of efficacy feedback provided in the embodiments of the present invention.

[0069] Figure 2 This is a schematic diagram of the TCM nursing knowledge graph structure in an embodiment of the present invention;

[0070] Figure 3 This is a schematic diagram of the relation constraint gating graph learning process in an embodiment of the present invention;

[0071] Figure 4 This is a schematic diagram illustrating the generation of the optimal nursing plan in an embodiment of the present invention. Detailed Implementation

[0072] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0073] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0074] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0075] like Figure 1 As shown, this embodiment provides an intelligent generation method for TCM nursing plans based on relational constraint gating and adaptive updating of efficacy feedback, including the following steps:

[0076] Step S1, Data Acquisition and Preprocessing: Acquire and process the patient's multimodal data, which includes at least symptoms, signs, physical condition, syndrome, and nursing goals;

[0077] Step S2, Knowledge Graph Construction: Construct a TCM nursing knowledge graph, which includes multiple nodes and multiple edges connecting the nodes. The nodes include at least symptom nodes, constitution nodes, syndrome nodes, and nursing intervention nodes. The edges represent the relationships between the nodes, including at least appropriate relationships, reinforcing relationships, restrictive relationships, conflict relationships, and taboo relationships.

[0078] Step S3, Demand State Encoding: Encode the acquired multimodal patient data into a patient demand state vector;

[0079] Step S4, Relationship Constraint Gated Graph Learning: The patient's demand state vector and the knowledge graph are input into a relationship constraint gated graph neural network to obtain the final representation of the nursing intervention nodes. The relationship constraint gated graph neural network employs differentiated propagation strategies for different types of relationships, specifically including: positive reinforcement propagation for appropriate and reinforcing relationships; decay propagation for restrictive relationships; negative penalty propagation for conflicting relationships; and hard shielding propagation for taboo relationships.

[0080] Step S5, Nursing Needs Subgraph Construction and Candidate Measure Recall: Map the nodes corresponding to the patient's need state vector to the knowledge graph to construct a nursing needs subgraph; based on the nursing needs subgraph, recall candidate nursing measures from the knowledge graph, and mask the recalled candidate nursing measures according to taboo relationships;

[0081] Step S6, Solution Generation: Combine the recalled candidate nursing measures to form multiple candidate solutions, and calculate the comprehensive score of each candidate solution based on the final representation of the nursing measure node. Select the solution with the highest comprehensive score as the optimal nursing solution output.

[0082] Step S7, Feedback Update: Obtain the efficacy feedback signal after the implementation of the nursing plan, and use the feedback signal to adaptively update the relation weights in the knowledge graph, the gating parameters of the relation constraint gating graph neural network, and the scoring parameters of the comprehensive plan score.

[0083] The following section provides a further explanation of the relevant technical aspects of this method.

[0084] 1. Data Acquisition and Preprocessing

[0085] 1.1 Data Source

[0086] Acquire multimodal patient data to construct training sets and knowledge graphs; data sources should include at least the following three categories:

[0087] 1. Clinical case data

[0088] Electronic medical records related to traditional Chinese medicine nursing in the hospital's HIS / EMR system;

[0089] Nursing assessment form, nursing record form, nursing plan form;

[0090] Medical records, discharge summary, follow-up records;

[0091] Nursing care execution records and periodic efficacy evaluation records.

[0092] 2. Traditional Chinese Medicine Knowledge Data

[0093] Traditional Chinese Medicine Nursing Textbook;

[0094] Traditional Chinese Medicine nursing clinical pathway;

[0095] Expert consensus and industry guidelines;

[0096] Classic works of Traditional Chinese Medicine;

[0097] Hospital internal TCM nursing standards.

[0098] 3. Plan and Feedback Data

[0099] The nursing protocols actually used in history;

[0100] Recommended plan after expert review;

[0101] Nursing compliance record;

[0102] Record of therapeutic improvement;

[0103] Adverse reaction records.

[0104] Preferably, the data comes from real TCM nursing cases that have undergone desensitization processing, and is labeled and reviewed by senior nursing experts.

[0105] 1.2 Sample Unit Division

[0106] Data collection uses one nursing assessment cycle, one hospitalization period, or one follow-up cycle as a sample unit. For hospitalized patients, it is preferable to divide the nursing process into multiple phases based on time windows, with the time window length set as follows: sky, Tian or This allows the samples to reflect the dynamic changes in the nursing status.

[0107] No. A sample can be represented as:

[0108]

[0109] in:

[0110] Text features include chief complaint, symptom description, tongue appearance description, pulse appearance description, and nursing record text;

[0111] Numerical characteristics, including sleep score, pain score, appetite score, fatigue score, vital signs, etc.

[0112] Category characteristics, including gender, age group, body type, and type of comorbidities;

[0113] Syndrome label or diagnosis result;

[0114] : A set of labels or nursing interventions for the target nursing care program;

[0115] : Therapeutic feedback labels, including improvement, compliance, and adverse reaction information;

[0116] : The time window marker corresponding to the sample.

[0117] 1.3 Supervision of Label Sources

[0118] Supervision labels are preferably derived from one or a combination of the following two types of results:

[0119] 1. Historically implemented and effective TCM nursing protocols;

[0120] 2. The standard treatment plan is given by senior nursing experts after reviewing historical cases.

[0121] When historical implementation plans differ from expert review plans, a weighted fusion approach can be used to construct oversight labels:

[0122]

[0123] in, This represents the supervision label of the i-th sample. This represents the historical execution plan for the i-th sample. Let represent the expert review scheme for the i-th sample. , These are the weighting coefficients. ,and Or it can be dynamically set based on the quality of cases.

[0124] 1.4 Construction of Positive and Negative Samples

[0125] For relationship reconstruction and solution prediction tasks, positive and negative samples need to be constructed:

[0126] 1) Positive relational samples: entity relation triples that are clearly present in case studies, literature, or norms;

[0127] 2) Negative relational samples: constructed through tail entity replacement, head entity replacement, or relational replacement.

[0128] 3) Positive sample of the plan: A combination of nursing measures that has been actually adopted and has achieved the expected therapeutic effect or has been confirmed by experts;

[0129] 4) Negative sample of the protocol: combinations that only partially overlap with the positive sample but have poor efficacy; combinations that contain conflicting or contraindications; combinations that are significantly mismatched with the patient's symptoms and physical condition.

[0130] 1.5 Handling Missing Items and Class Imbalance

[0131] For missing text or missing numerical features, one of the following strategies can be used:

[0132] 1) Rule completion;

[0133] 2) Mode / mean imputation;

[0134] 3) Missing mask vector representation;

[0135] 4) Neighbor-based completion based on similar cases.

[0136] To address the imbalance in nursing intervention label categories, methods such as category weight recalibration, oversampling, or focus loss weighting can be used for balancing.

[0137] 1.6 Annotated Content

[0138] The case data and knowledge data are annotated as follows:

[0139] 1. Entity annotation

[0140] –Symptom entity;

[0141] -Physical features;

[0142] –Physical entity;

[0143] –Certificate entity;

[0144] – Entities of nursing interventions;

[0145] – Prohibited items;

[0146] –Efficacy feedback entity.

[0147] 2. Relationship labeling

[0148] -Suitable relationship;

[0149] –Strengthen relationships;

[0150] -Restriction relationship;

[0151] -Conflict relationship;

[0152] - Forbidden relationship;

[0153] – Improve relationships;

[0154] - Feedback relationship.

[0155] 3. Supervision label

[0156] -Multi-label output of nursing interventions;

[0157] – Nursing plan combination label;

[0158] – Therapeutic efficacy rating label;

[0159] – Adverse reaction label.

[0160] 2. Construction of a Knowledge Graph for Traditional Chinese Medicine Nursing

[0161] 2.1 Definition of a Spectrum

[0162] The knowledge graph of traditional Chinese medicine nursing is defined as:

[0163]

[0164] in:

[0165] : A set of nodes;

[0166] : Edge set;

[0167] : A set of relation types;

[0168] Node feature matrix.

[0169] Each relation edge is represented as a triple:

[0170]

[0171] in:

[0172] : Head entity;

[0173] : Relationship type;

[0174] Tail entity.

[0175] 2.2 Node Types

[0176] Node types must include at least:

[0177] Symptom nodes;

[0178] Vital signs nodes;

[0179] Tongue-shaped nodes;

[0180] Pulse nodes;

[0181] Constitutional nodes;

[0182] Syndrome nodes;

[0183] Nursing target milestones;

[0184] Nursing intervention points;

[0185] Forbidden nodes;

[0186] Therapeutic effect feedback node.

[0187] 2.3 Relationship Types and Their Processing Methods

[0188] Relationship types must include at least:

[0189] Corresponding syndrome relationship;

[0190] Suitable relationship;

[0191] Strengthen relationships;

[0192] Restrictive relationship;

[0193] Conflict relationship;

[0194] Taboo relationship;

[0195] Improve relationships;

[0196] Feedback relationship.

[0197] In this invention, different relationships are treated differently as follows:

[0198] Relationship type Methods of dissemination Scoring method Filtering methods suitable Positive propagation Improve match score Unfiltered Enhance Positive propagation Enhance collaboration Unfiltered limit Attenuation propagation Increase mild punishment Conditional reservation conflict negative propagation Increase conflict penalty points Strong constraint filtering Taboo Hard shielding propagation Increase penalties for high-risk activities Directly remove Improvement / Feedback Participate in updates Adjust feedback score Do not filter directly

[0199] Therefore, this invention does not simply weight multiple relationship graphs and then perform ordinary propagation, but rather defines different propagation paradigms, scoring paradigms, and filtering paradigms based on the semantic differences of the relationships.

[0200] 2.4 Relationship Weight Calculation

[0201] For any relation type Next node With nodes The initial relation weight of the relation edge is defined as follows:

[0202]

[0203] in:

[0204] :entity , Normalized values ​​of co-occurrence frequency in case studies and literature;

[0205] Expert-marked intensity normalized values;

[0206] : Normalized value of historical efficacy support;

[0207] Weighting coefficients, satisfying .

[0208] The normalized value of co-occurrence frequency can be defined as:

[0209]

[0210] Historical support for treatment efficacy can be defined as:

[0211]

[0212] in:

[0213] :entity The number of times they co-occur;

[0214] : The number of samples whose symptoms improved after using this relationship to correspond to nursing interventions;

[0215] : The total number of samples corresponding to the nursing interventions using this relationship;

[0216] Smoothing constant.

[0217] 2.5 Map Construction Process

[0218] The knowledge graph construction process includes:

[0219] 1. Clean and standardize the original case texts and knowledge texts;

[0220] 2. Entity extraction based on terminology dictionaries, standard terminology databases, and entity recognition models;

[0221] 3. Extracting relationships between entities based on relation extraction models and rule templates;

[0222] 4. Normalize and remove duplicates from the extracted triples;

[0223] 5. Construct a heterogeneous graph according to preset node types and relationship types;

[0224] 6. Calculate the relation weight of each relation edge and write it into the graph;

[0225] 7. Encode the therapeutic effect feedback information into feedback relationships and write them into the atlas.

[0226] Figure 2 This is a schematic diagram of the knowledge graph structure for Traditional Chinese Medicine nursing. Figure 2 The diagram illustrates the relationship structure between symptom nodes, constitution nodes, tongue and pulse nodes, syndrome nodes, nursing intervention nodes, contraindication nodes, and efficacy feedback nodes in the TCM nursing knowledge graph. This includes syndrome-related relationships, appropriate relationships, enhancement relationships, restriction relationships, conflict relationships, contraindication relationships, improvement relationships, and feedback relationships.

[0227] 3. Node Feature Representation and Demand State Encoding

[0228] 3.1 Initial Characteristics of Nodes

[0229] The first in the atlas The initial representation of each node is defined as follows:

[0230]

[0231] in:

[0232] : The semantic encoding vector of the node name or description text;

[0233] Node type embedding vector;

[0234] Node attribute vector, including attributes such as frequency, risk level, applicable age group, and applicable physical condition.

[0235] 3.2 Patient Demand State Vector

[0236] The symptoms, signs, constitution, syndromes, and nursing goals of the target patients are coded to construct a demand state vector:

[0237]

[0238] in:

[0239] Symptom encoding vector;

[0240] : Vital feature encoding vector;

[0241] Physical constitution encoding vector;

[0242] Syndrome encoding vector;

[0243] Nursing goal encoding vector;

[0244] Nursing stage encoding vector.

[0245] By introducing nursing phase coding vectors, the same nursing intervention can be activated to different degrees in the acute, recovery, or consolidation phases.

[0246] 4. Learning Relationship Constraint Gated Graphs

[0247] 4.1 Relationship Gating Mechanism

[0248] For different relationship types, a relationship constraint gating mechanism is introduced. Differentiated propagation strategies are implemented through relationship gating values. The relationship gating value is jointly determined by the relationship type, the representation of adjacent nodes, and the patient's demand state vector.

[0249] Specifically, nodes Through relationships For nodes Relationship Gating Value Calculated using the following formula:

[0250]

[0251] in:

[0252] :relation The lower gate value;

[0253] , : The l-th level representation of node i and node j;

[0254] , :relation The corresponding gate weight matrix and gate bias vector;

[0255] : Patient demand state vector;

[0256] : Sigmoid activation function.

[0257] Therefore, different patients in different states can obtain different gating strengths for the same relation edge, realizing patient state-driven adaptive relation activation.

[0258] 4.2 Relationship Differentiation Gating Strategy

[0259] To avoid treating all negative relationships uniformly, this invention adopts the following strategy for different relationship types:

[0260] 1. Appropriate Relationship Gating

[0261] Positive reinforcement gating is used; the larger the gating value, the more suitable the nursing intervention is for dissemination and recall in the current patient condition.

[0262] 2. Enhance relationship gating

[0263] The use of combined enhanced gating not only affects the propagation of individual nodes, but also the subsequent combined collaborative scores.

[0264] 3. Restriction Relationship Gating

[0265] Attenuation gating is used to reduce the propagation intensity without directly removing candidate measures, which is suitable for nursing scenarios where "it is available but should be downgraded in priority".

[0266] 4. Conflict Gating

[0267] A negative penalty gating system is adopted, which participates in both propagation suppression and combined scoring penalties; when the gating value exceeds the conflict threshold, the combination of related measures is directly judged as infeasible.

[0268] 5. Taboo Relationship Gating

[0269] Using hard-shielded gating, if the taboo condition is met, the effective propagation value of the corresponding measure node is directly set to zero:

[0270]

[0271] This layered strategy is one of the key differences between this invention and the general attention / gating mechanism.

[0272] 4.3 Relationship Propagation Mechanism

[0273] Divide the set of relations into a set of positive relations. and negative relation set :

[0274]

[0275]

[0276] For the first Layered graph propagation, the (l+1)th layer representation of node i Update using the following formula:

[0277]

[0278] in:

[0279] :node In relationship The set of neighbors below;

[0280] : No. Layer Relationship The corresponding transformation matrix;

[0281] : Normalization coefficient;

[0282] : No. The self-connection transformation matrix corresponding to the layer;

[0283] : Effective gating value after taboo hard shielding;

[0284] : A set of positive relationships that include suitable relationships and enhancing relationships;

[0285] : A set of negative relationships that include restrictive, conflicting, and taboo relationships.

[0286] 4.4 Conflict Threshold Control

[0287] For conflicting relationships, set a conflict threshold. If a candidate measure is effective... satisfy:

[0288]

[0289] This measure determines that a conflicting relationship is incompatible during the candidate combination generation stage and does not proceed to subsequent scoring steps. This mechanism allows conflicting relationships to simultaneously serve the functions of "soft punishment" and "hard filtering".

[0290] 4.5 Nursing interventions indicate learning

[0291] go through After gating propagation based on hierarchical relationships, the final representation of the nursing intervention node is obtained:

[0292]

[0293] in, This represents any nursing intervention node. This representation simultaneously encodes structural information, relational constraints, and patient status-related information from the TCM nursing knowledge graph.

[0294] Figure 3 This is a schematic diagram of the learning process for a relation-constrained gating graph. Figure 3 The present invention illustrates a relation gating calculation module that uses the patient state vector, positive relation set, and negative relation set as inputs. It achieves positive reinforcement propagation of appropriate and reinforcing relations, decay propagation of restrictive relations, negative penalty propagation of conflict relations, and hard shielding propagation of taboo relations through a relation propagation mechanism, thereby completing the node representation update and outputting the nursing measure representation.

[0295] 5. Training Objectives and Training Process

[0296] Relation-constrained gated graph neural networks are optimized using a joint loss function during the training phase.

[0297] 5.1 Losses in Relationship Reconstruction

[0298] To ensure that the graph relation structure can be effectively learned by the model, the relation reconstruction loss is defined as:

[0299]

[0300] in:

[0301] : Set of positive sample relations;

[0302] : Set of negative sampling relations;

[0303] :relation The bilinear mapping matrix.

[0304] 5.2 Nursing plan predicts losses

[0305] Treating nursing intervention prediction as a multi-label classification task, and defining prediction probabilities for all nursing interventions. The predicted loss from the nursing care plan is:

[0306]

[0307] in:

[0308] Total number of nursing interventions;

[0309] : No. The true label of each nursing intervention;

[0310] : The predicted probability output by the model.

[0311] To alleviate category imbalance, a weighted approach can be further adopted:

[0312]

[0313] in, For the first Category weights for different types of nursing interventions.

[0314] 5.3 Conflict Constraint Penalty Loss

[0315] For conflict care measures on the set The conflict penalty loss is defined as:

[0316]

[0317] in: For conflict measures; This is the conflict intensity coefficient.

[0318] 5.4 Taboo against shielding losses

[0319] To strengthen the hard constraint of taboo relations, taboo-shielding loss is defined for taboo measures:

[0320]

[0321] in:

[0322] A set of nursing interventions that meet the contraindications;

[0323] Taboo strength coefficient.

[0324] 5.5 Feedback Consistency Loss

[0325] To ensure that model training is consistent with historical efficacy feedback, a feedback consistency loss is introduced:

[0326]

[0327] in:

[0328] Model for the first Predicted value of therapeutic improvement for each sample;

[0329] Actual therapeutic effect and degree of improvement;

[0330] Total number of samples.

[0331] 5.6 Joint Loss Function

[0332] The joint loss function is defined as follows:

[0333]

[0334] in, Losses due to relationship rebuilding To predict losses for weighted care plans, To punish losses in the event of conflict To avoid shielding the loss, To provide feedback on consistency loss, This is the loss weighting coefficient.

[0335] 5.7 Training Process

[0336] The training phase includes the following steps:

[0337] 1) Collect case data, knowledge data, and feedback data;

[0338] 2) Perform desensitization, cleaning, standardization, time window segmentation, and sample construction;

[0339] 3) Complete entity annotation, relationship annotation, and scheme label annotation;

[0340] 4) Generate a relation sample set and a scheme sample set based on the positive and negative sample construction rules;

[0341] 5) Construct a knowledge graph of traditional Chinese medicine nursing and calculate the relationship weights;

[0342] 6) Initialize node features and patient state coding model;

[0343] 7) Training the relation-constraint gated graph neural network;

[0344] 8) Joint optimization relationship reconstruction loss, scheme prediction loss, conflict penalty loss, taboo shielding loss, and feedback consistency loss;

[0345] 9) Save knowledge graph parameters, gating parameters, and scheme prediction parameters.

[0346] 6. Nursing Needs Subgraph Construction and Candidate Response Recall

[0347] 6.1 Construction of Nursing Needs Subgraph

[0348] For the target patient, their symptoms, signs, constitution, syndrome, and nursing goals are mapped to a set of nodes in a graph. And construct a nursing needs subgraph:

[0349]

[0350] in:

[0351] It consists of the patient input node and its neighboring extended nodes;

[0352] : The set of corresponding relation edges.

[0353] Preferably, a one- or multi-hop neighborhood expansion is performed centered on the demand node, incorporating relevant symptom nodes, nursing measure nodes, contraindication nodes, and efficacy feedback nodes.

[0354] 6.2 Single Measure Matching Score

[0355] For any nursing intervention node The matching of its value relative to the current patient's needs is defined as follows:

[0356]

[0357] in:

[0358] : Demand node representation;

[0359] Nursing intervention nodes are indicated;

[0360] : Path weight between demand node and nursing measure node.

[0361] according to Nursing interventions are ranked, and the top few candidate interventions are recalled from different nursing dimensions to form a candidate intervention set.

[0362] 6.3 Taboos and Shielding During the Recall Phase

[0363] For candidate nursing interventions that meet the contraindications If it is related to any taboo node satisfy:

[0364]

[0365] If the weight of the taboo relationship between a candidate nursing measure and any taboo node is greater than a preset taboo threshold, then the nursing measure is directly removed from the candidate set during the candidate recall phase. Furthermore, in relational constraint gating graph learning, for nursing measures that meet the taboo judgment conditions, their effective gating value... It is directly set to zero to achieve hard shielding propagation. This step is used to prevent tabu measures from being processed only in the final output stage, thereby improving inference efficiency and security.

[0366] 7. Nursing plan combination generation and scoring

[0367] 7.1 Candidate Combination Generation

[0368] The candidate measures are categorized into dietary care, emotional regulation, daily life regulation, functional exercise, and acupoint care, and a set of candidate solutions is constructed based on preset combination rules. .

[0369] 7.2 Combination matching points

[0370] For candidate solutions Combination matching is defined as follows:

[0371]

[0372] 7.3 Synergistic Enhancement

[0373] The synergistic enhancement component among nursing interventions in the plan is defined as follows:

[0374]

[0375] in, Indicates nursing measures and The weight of strengthening relationships between them.

[0376] 7.4 Limiting Penalty Points

[0377] For nursing interventions in the "available but need to be downgraded" state, define a limitation penalty score:

[0378]

[0379] in, To limit the relational gating value, To limit the weight of relationships.

[0380] 7.5 Conflict Penalty Points

[0381] The conflict penalty score in the scheme is defined as follows:

[0382]

[0383] in, Indicates nursing measures and The weight of the conflict relationship between them.

[0384] 7.6 Risk Penalty Score

[0385] Let the set of contraindication factors for patients be . The risk penalties for the proposed solution are categorized as follows:

[0386]

[0387] in, Indicates nursing measures With taboo nodes Weight of taboo relationships between them.

[0388] 7.7 Overall Score

[0389] Candidate solutions Overall rating Defined as:

[0390]

[0391] in:

[0392] Matching score: used to characterize the degree of matching between candidate solutions and patient needs;

[0393] Synergistic enhancement score, used to characterize the strength of the enhancing relationship between nursing measures within the candidate protocol;

[0394] : Restriction penalty score, used to characterize the penalty value of nursing measures that are in a restricted state in the candidate plan;

[0395] Conflict penalty score: used to characterize the strength of conflict relationships between nursing measures within the candidate program;

[0396] Risk penalty score: used to characterize the strength of the association between nursing interventions and patient contraindications in candidate protocols;

[0397] : Scoring weighting coefficient;

[0398] Ultimately, the optimal nursing care plan is as follows:

[0399]

[0400] Solutions with a comprehensive score below the threshold or that violate strong taboo constraints will be eliminated.

[0401] 7.8 Structured Output

[0402] The optimal care plan output is preferably a structured care plan, which includes at least the following fields:

[0403] 1) Target patient information;

[0404] 2) Corresponding symptoms and physical condition;

[0405] 3) Recommended nursing goals;

[0406] 4) List of nursing interventions;

[0407] 5) Implementation frequency and duration;

[0408] 6) Precautions;

[0409] 7) Prohibitions and Reminders;

[0410] 8) Solution explanation path.

[0411] The optimal nursing care plan interpretation path is presented in a chain structure of "symptoms / signs → syndrome / constitution → nursing measures → evidence supporting efficacy".

[0412] Figure 4 This is a schematic diagram of the process of generating the optimal nursing plan. Figure 4 This invention illustrates the process of starting from a nursing needs subgraph, recalling candidate nursing measures, generating candidate combinations, calculating matching scores, synergy enhancement scores, restriction penalty scores, conflict penalty scores, and risk penalty scores, and outputting the optimal nursing plan after filtering by conflict / contraindications.

[0413] 8. Adaptive updates driven by therapeutic feedback

[0414] 8.1 Definition of Feedback Signal

[0415] After the nursing plan was implemented, the following efficacy feedback signals were collected:

[0416] 1) Symptom improvement rate;

[0417] 2) Improvement in vital signs;

[0418] 3) Nursing compliance;

[0419] 4) Adverse reactions;

[0420] 5) Expert review results;

[0421] 6) Patient satisfaction.

[0422] Encode the feedback signal into a feedback function. It is used to update relation weights and model parameters.

[0423] 8.2 Construction of Feedback Support Function

[0424] Relationship to any nursing intervention Construct the feedback support function:

[0425]

[0426] in:

[0427] : The degree of improvement of symptoms / signs after the implementation of nursing interventions;

[0428] Nursing compliance score;

[0429] Adverse reaction intensity;

[0430] Expert review and support level;

[0431] This is the feedback fusion coefficient.

[0432] Furthermore, different relation types employ different update directions:

[0433] For appropriate and enhanced relationships, if Increasing the value of the relation increases its weight.

[0434] For restrictive, conflicting, and contraindication relationships, if adverse reactions or decreased efficacy occur, the weight of the corresponding negative relationship will be increased.

[0435] If adverse reactions decrease and therapeutic effects improve, the strength of the limiting relationship can be reduced.

[0436] 8.3 Relationship Weight Update

[0437] For relationships Next node The relation weights are defined, and a feedback update formula is defined to update the relation weights:

[0438]

[0439] in:

[0440] Relationship weights before update;

[0441] Updated relationship weights;

[0442] Relationship type The corresponding update rate;

[0443] : Feedback support function;

[0444] : Relation-type specific update mapping function.

[0445] Preferably:

[0446]

[0447] This design allows different relation types to have different update directions for the same feedback signal.

[0448] 8.4 Gating Parameter Update

[0449] Gating parameters Perform gradient updates:

[0450]

[0451] in:

[0452] Learning rate;

[0453] Feedback consistency loss or its extended forms.

[0454] The gating parameters of the relation-constrained gating graph neural network are adjusted using gradient updates.

[0455] 8.5 Scoring parameters are updated adaptively.

[0456] Based on the actual efficacy of historical output protocols, the weights of the overall protocol score are adaptively updated:

[0457]

[0458] in:

[0459] : No. Category scoring item weights;

[0460] Update step size;

[0461] : The incremental feedback on the contribution of the corresponding scoring item to the therapeutic effect.

[0462] For example, when the synergistic enhancement score is significantly positively correlated with improved efficacy, it increases... When conflict penalty scores are significantly positively correlated with the incidence of adverse reactions, increasing... .

[0463] Impact path after 8.6 update

[0464] Once the efficacy feedback is updated, it will have at least the following impacts on the next round of reasoning:

[0465] 1) The weights of appropriate and enhanced relationships alter the recall order of candidate measures;

[0466] 2) The weights of restrictive, conflicting, and taboo relationships alter the propagation inhibition strength and filtering threshold;

[0467] 3) Gating parameters alter the activation intensity of the same measure in different patient states;

[0468] 4) The scoring parameters change the ranking result of the final combination scheme.

[0469] Therefore, feedback updates are not an independent post-processing step, but rather directly affect the entire process of generating subsequent solutions.

[0470] This embodiment also provides an intelligent TCM nursing plan generation system for implementing the above method, including:

[0471] The data acquisition and preprocessing module is used to acquire and process patients' multimodal data;

[0472] The knowledge graph construction module is used to build a knowledge graph for traditional Chinese medicine nursing.

[0473] The demand state encoding module is used to encode the acquired multimodal patient data into a patient demand state vector;

[0474] The relation constraint gating graph learning module is used to input the patient's demand state vector and the knowledge graph into the relation constraint gating graph neural network to obtain the final representation of the nursing intervention node; wherein, the relation constraint gating graph neural network adopts a differentiated propagation strategy for different types of relations;

[0475] The nursing needs subgraph construction module is used to map the nodes corresponding to the patient needs state vector to the knowledge graph to construct the nursing needs subgraph.

[0476] The candidate measure recall module is used to recall candidate nursing measures from the knowledge graph based on the nursing needs subgraph, and to screen the recalled candidate nursing measures according to the taboo relationship;

[0477] The solution generation module is used to combine the recalled candidate nursing measures to form multiple candidate solutions, and calculate the comprehensive score of each candidate solution based on the final representation of the nursing measure node, and select the solution with the highest comprehensive score as the optimal nursing solution output.

[0478] The feedback update module is used to obtain efficacy feedback signals after the implementation of the nursing plan, and to adaptively update the relation weights, the gating parameters of the relation constraint gating graph neural network, and the scoring parameters of the comprehensive plan score using the feedback signals.

[0479] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0480] This embodiment also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method.

[0481] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0482] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0483] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0484] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0485] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for intelligently generating TCM nursing plans based on relational constraint gating and adaptive updating of efficacy feedback, characterized in that, Includes the following steps: Step S1, Data Acquisition and Preprocessing: Acquire and process the patient's multimodal data, which includes at least symptoms, signs, physical condition, syndrome, and nursing goals; Step S2, Knowledge Graph Construction: Construct a TCM nursing knowledge graph, which includes multiple nodes and multiple edges connecting the nodes. The nodes include at least symptom nodes, constitution nodes, syndrome nodes, and nursing intervention nodes. The edges represent the relationships between the nodes, including at least appropriate relationships, reinforcing relationships, restrictive relationships, conflict relationships, and taboo relationships. Step S3, Demand State Encoding: Encode the acquired multimodal patient data into a patient demand state vector; Step S4, Relationship Constraint Gated Graph Learning: The patient's demand state vector and the knowledge graph are input into a relationship constraint gated graph neural network to obtain the final representation of the nursing intervention nodes. The relationship constraint gated graph neural network employs differentiated propagation strategies for different types of relationships, specifically including: positive reinforcement propagation for appropriate and reinforcing relationships; decay propagation for restrictive relationships; negative penalty propagation for conflicting relationships; and hard shielding propagation for taboo relationships. Step S5, Nursing Needs Subgraph Construction and Candidate Measure Recall: Map the nodes corresponding to the patient's need state vector to the knowledge graph to construct a nursing needs subgraph; based on the nursing needs subgraph, recall candidate nursing measures from the knowledge graph, and mask the recalled candidate nursing measures according to taboo relationships; Step S6, Solution Generation: Combine the recalled candidate nursing measures to form multiple candidate solutions, and calculate the comprehensive score of each candidate solution based on the final representation of the nursing measure node. Select the solution with the highest comprehensive score as the optimal nursing solution output. Step S7, Feedback Update: Obtain the efficacy feedback signal after the implementation of the nursing plan, and use the feedback signal to adaptively update the relation weights in the knowledge graph, the gating parameters of the relation constraint gating graph neural network, and the scoring parameters of the comprehensive plan score.

2. The intelligent generation method for TCM nursing plans based on relational constraint gating and adaptive updating of therapeutic effect feedback as described in claim 1, characterized in that, In step S1, the acquired multimodal data comes from clinical case data, traditional Chinese medicine (TCM) knowledge data, and protocol and feedback data. The clinical case data includes TCM nursing-related electronic medical records, nursing assessment forms, nursing record forms, nursing plan forms, progress notes, discharge summaries, follow-up records, nursing execution records, and interim efficacy evaluation records. The TCM knowledge data includes TCM nursing textbooks, TCM nursing clinical pathways, expert consensus and industry guidelines, classic TCM works, and hospital internal TCM nursing standards. The protocol and feedback data includes historically used nursing protocols, expert-reviewed recommended protocols, nursing compliance records, efficacy improvement records, and adverse reaction records. The data collection uses one nursing assessment cycle, one hospitalization period, or one follow-up cycle as a sample unit, and the patient's nursing process is divided into multiple stage samples according to time windows. The method uses supervision tags derived from historically implemented and effective TCM nursing protocols, and / or standard protocols provided by senior nursing experts after reviewing historical cases. When the historical implementation protocol differs from the expert-reviewed protocol, a weighted fusion approach is used to construct the supervision tags. in, This represents the supervision label of the i-th sample. This represents the historical execution plan for the i-th sample. This represents the expert review scheme for the i-th sample. , These are the weighting coefficients. .

3. The intelligent generation method for TCM nursing plans based on relational constraint gating and adaptive updating of efficacy feedback as described in claim 1, characterized in that, In step S2, for any relation type Next node With nodes The initial relation weight of the relation edge is defined as follows: in, Representing entities , Normalized values ​​of co-occurrence frequency in case studies and literature This indicates the normalized intensity value marked by experts. This represents the normalized value of historical efficacy support. The weighting coefficients are and satisfy the following conditions: ; Furthermore, in step S7, the update of the relation weights adopts the following feedback update formula: in, This indicates the updated relation weight. This indicates the relation weight before the update. This represents the update rate corresponding to relation type r. Represents an update mapping function specific to the relation type. This represents the feedback support function.

4. The intelligent generation method for TCM nursing plans based on relational constraint gating and adaptive updating of therapeutic effect feedback as described in claim 1, characterized in that, In step S4, the differentiated propagation strategy is implemented through a relation gating value, which is determined by the relation type, the representation of adjacent nodes, and the patient's demand state vector. Specifically, node j uses relation r to determine the relation gating value of node i. Calculated using the following formula: in, and Let be the l-th layer representations of nodes i and j, respectively, and q be the patient demand state vector. and Let be the gate weight matrix and gate bias vector corresponding to relation r, respectively, and σ be the Sigmoid activation function.

5. The intelligent generation method for TCM nursing plans based on relational constraint gating and adaptive updating of therapeutic effect feedback as described in claim 1, characterized in that, In step S4, the (l+1)th layer representation of node i Update using the following formula: in, It is a set of positive relationships that includes suitable relationships and enhancing relationships. It is a set of negative relations that includes restrictive, conflicting, and taboo relations. Let i be the set of neighbors of node i in relation r. To consider the effective gating value after taboo hard shielding, Let r be the transformation matrix corresponding to relation r in the l-th layer. Let l be the self-connection transformation matrix corresponding to the l-th layer. This is the normalization coefficient.

6. The intelligent generation method for TCM nursing plans based on relational constraint gating and adaptive updating of efficacy feedback as described in claim 1, characterized in that, In step S5, the recalled candidate care measures are screened out based on contraindications, specifically as follows: When the weight of the taboo relationship between any candidate nursing measure and any taboo node is greater than a preset taboo threshold, the nursing measure is removed from the candidate set; furthermore, in the learning of the relation constraint gating graph, for nursing measures that meet the taboo judgment conditions, their effective gating value is... It is set directly to zero to achieve hard-block propagation.

7. The intelligent generation method for TCM nursing plans based on relational constraint gating and adaptive updating of efficacy feedback as described in claim 1, characterized in that, In step S6, the comprehensive score is calculated based on a weighted average of the matching score, synergy enhancement score, restriction penalty score, conflict penalty score, and risk penalty score. in, This indicates the overall score; The matching score is used to characterize the degree of matching between the candidate treatment and the patient's needs; The synergistic enhancement score is used to characterize the strength of the enhancing relationship between nursing interventions within the candidate protocol. The penalty score is used to characterize the penalty value of nursing measures that are in a restricted state in the candidate plan; The conflict penalty score is used to characterize the strength of the conflict relationship between nursing measures within the candidate options. Risk penalty score is used to characterize the strength of the association between nursing interventions and patient contraindications in candidate protocols; This refers to the scoring weighting coefficient; The optimal nursing plan is a structured plan, which includes the following fields: target patient information, corresponding symptoms and constitution, list of nursing measures, implementation frequency and implementation time, precautions, contraindications and explanation path of the plan; the explanation path of the optimal nursing plan is presented in a chain structure of "symptoms / signs → syndrome / constitution → nursing measures → evidence supporting efficacy".

8. The intelligent generation method for TCM nursing plans based on relational constraint gating and adaptive updating of efficacy feedback as described in claim 1, characterized in that, Step S7 specifically includes: The therapeutic feedback signal is decomposed into an improvement component, a compliance component, and an adverse reaction component; Based on the positive changes in the improvement degree component and compliance component, increase the relation weights of relevant appropriate relations and enhanced relations in the knowledge graph; Increase the relational weights of relevant limiting and conflicting relationships based on the occurrence or enhancement of adverse reaction components; The gating parameters of the relation-constrained gating graph neural network are adjusted using gradient updates. Based on the correlation between the efficacy feedback signal and the contribution of each scoring item, the scoring weight coefficient of the overall treatment plan score is adjusted.

9. The intelligent generation method for TCM nursing plans based on relational constraint gating and adaptive updating of therapeutic effect feedback as described in claim 1, characterized in that, The relation-constrained gated graph neural network is optimized using a joint loss function during the training phase. The joint loss function is as follows: in, Losses due to relationship rebuilding To predict losses for weighted care plans, To punish losses in the event of conflict To avoid shielding the loss, To provide feedback on consistency loss, This is the loss weighting coefficient.

10. A smart system for generating TCM nursing plans to implement the method as described in any one of claims 1-9, characterized in that, include: The data acquisition and preprocessing module is used to acquire and process patients' multimodal data; The knowledge graph construction module is used to build a knowledge graph for traditional Chinese medicine nursing. The demand state encoding module is used to encode the acquired multimodal patient data into a patient demand state vector; The relation constraint gating graph learning module is used to input the patient's demand state vector and the knowledge graph into the relation constraint gating graph neural network to obtain the final representation of the nursing intervention node; wherein, the relation constraint gating graph neural network adopts a differentiated propagation strategy for different types of relations; The nursing needs subgraph construction module is used to map the nodes corresponding to the patient needs state vector to the knowledge graph to construct the nursing needs subgraph. The candidate measure recall module is used to recall candidate nursing measures from the knowledge graph based on the nursing needs subgraph, and to screen the recalled candidate nursing measures according to the taboo relationship; The solution generation module is used to combine the recalled candidate nursing measures to form multiple candidate solutions, and calculate the comprehensive score of each candidate solution based on the final representation of the nursing measure node, and select the solution with the highest comprehensive score as the optimal nursing solution output. The feedback update module is used to obtain efficacy feedback signals after the implementation of the nursing plan, and to adaptively update the relation weights, the gating parameters of the relation constraint gating graph neural network, and the scoring parameters of the comprehensive plan score using the feedback signals.

Images