Preoperative decision support evaluation method for thyroid and breast disease based on large model and knowledge graph

Abstract

The invention belongs to the technical field of knowledge maps and large models, and discloses a breast and thyroid disease preoperative decision support evaluation method based on a large model and a knowledge map, and the method comprises the steps: obtaining the preoperative diagnosis and treatment data of a target patient, and obtaining the preoperative diagnosis and treatment data of the target patient based on a breast and thyroid disease diagnosis and treatment path knowledge map containing specific nodes and relationships; generating a path state vector containing a path identifier, a missing type and a stage code; generating a structured constraint prompt according to a knowledge graph crossing forbidding condition; inputting the two schemes into a trained large model to obtain decision candidate schemes and corresponding path sufficiency arguments; and finally, by calculating the path dependence intensity and the risk stability, screening out a final decision suggestion. According to the method, the problem of incomplete diagnosis and treatment information can be accurately solved, the accuracy, interpretability and risk controllability of preoperative decision making are improved, and clinical actual requirements are met.

Description

Technical Field

[0001] This application relates to the fields of knowledge graph and large model technology, and more specifically, to a method for preoperative decision support assessment of thyroid and breast disease based on large models and knowledge graphs. Background Technology

[0002] Preoperative assessment for thyroid and breast diseases is a crucial step in clinical diagnosis and treatment, directly impacting the rationality of surgical plans and treatment outcomes. However, in real-world clinical settings, highly incomplete diagnostic and treatment information is prevalent, posing numerous challenges to preoperative decision-making. The examination items collected from the same patient by different medical institutions and healthcare professionals vary significantly, and some key diagnostic and treatment indicators are often presented in vague natural language or missing from the medical record, making them difficult to use directly for accurate decision-making.

[0003] Existing preoperative assessment systems are mostly built on rules or statistical learning, assuming that the diagnosis and treatment information is complete, or using simple missing value imputation to handle information gaps. This leads to implicit distortions in decision-making logic, and the risk assessment results are highly sensitive to missing paths. Furthermore, they cannot provide reasonable explanations for decision-making basis under conditions of insufficient information.

[0004] In recent years, many large-scale model-related solutions have generated decision recommendations by supplementing information or through direct reasoning. They fail to structurally model the incompleteness of treatment pathways, treating missing data merely as ordinary null values ​​and ignoring the phased and evolving nature of these missing pathways and their varying directional impact on different decision-making options. Furthermore, large-scale model reasoning is prone to implicit pathway skipping, assuming key treatment steps have been completed. This results in superficially reasonable recommendations that violate clinical pathway constraints, making them difficult for clinicians to adopt and failing to meet actual treatment needs. Summary of the Invention

[0005] In response, this application provides a preoperative decision support assessment method for breast and thyroid diseases based on large models and knowledge graphs, in order to at least partially solve the above-mentioned technical problems.

[0006] This application provides a method for preoperative decision support assessment of breast and thyroid diseases based on large models and knowledge graphs, including the following methodological steps: S1, Obtain the preoperative diagnosis and treatment data of the target patient, and generate the path state vector of the target patient based on the knowledge graph of the breast and thyroid disease diagnosis and treatment path; the path state vector includes at least a path identifier, a missing type code, and a missing stage code. The missing type code indicates that the path is completely not executed, executed but with ambiguous results, or executed but with conflicting results. The missing stage code is generated based on the stage attributes of the nodes in the knowledge graph. The nodes of the knowledge graph include disease state nodes, examination behavior nodes, examination result state nodes, and decision nodes. The edges of the knowledge graph describe the sequential dependencies, optional conditions, and prohibited crossing conditions between nodes. The prohibited crossing conditions define the decision types that are prohibited from being generated when the state of the specified preceding node is not satisfied. S2, Based on the path state vector, extract and generate corresponding structured constraint hints; the structured constraint hints include a list of decision types currently prohibited from being generated, determined by combining the path identifier, missing type code, and missing stage code in the path state vector with the prohibited crossing conditions in the knowledge graph; S3, input the path state vector and the structured constraint hints into the trained large model, and the large model outputs at least one decision candidate and a path sufficiency demonstration for each decision candidate; the path sufficiency demonstration explains the core path on which the decision candidate depends and its completion status in the current path state; S4. Based on the path sufficiency demonstration, calculate the path dependency strength and risk stability of each decision candidate scheme, and select the final decision suggestion from the decision candidate schemes according to the calculation results; the path dependency strength is determined according to the proportion of completed paths in the core paths on which the decision candidate scheme depends, and the risk stability is determined according to the directional influence weight corresponding to the path missing that affects the decision candidate scheme.

[0007] In one possible embodiment, the method further includes defining stage attributes for nodes in the knowledge graph, wherein the stage attributes include a preoperative basic assessment stage, a preoperative enhanced assessment stage, and a decision confirmation stage; the prohibition on crossing conditions is associated with the stage attributes of the nodes.

[0008] In one possible embodiment, when the path state vector of the target patient indicates that a certain preceding path node is missing, if there is an alternative path based on the information compensation relationship in the knowledge graph and its path credibility reaches a preset threshold, then the prohibition on the corresponding decision type is partially lifted.

[0009] In one possible embodiment, generating a path state vector further includes determining a decision influence direction encoding based on a pre-generated path-decision relationship table, wherein the path-decision relationship table records the directional influence weight of each path identifier on different types of decision schemes.

[0010] In one possible embodiment, based on a historical case dataset, the differences in postoperative risk indicators corresponding to different types of decision-making schemes are statistically analyzed under the conditions of missing and non-missing specific paths. The differences in risk indicators are normalized and calibrated in conjunction with medical guideline constraints to obtain the directional influence weight.

[0011] In one possible embodiment, generating structured constraint hints specifically involves matching preceding path nodes from the knowledge graph that are incomplete and do not meet the dynamic release mechanism for prohibiting crossing conditions, based on the path state vector. The prohibition decision types associated with these preceding path nodes are then organized in a key-value pair structured format to generate the structured constraint hints.

[0012] In one possible embodiment, the large model generates an initial set of decision candidate solutions under the constraints of the structured constraint hints based on the path state vector and the structured constraint hints, and performs an initial score on each decision candidate solution; for each solution in the initial set of decision candidate solutions, the path sufficiency argument is generated.

[0013] In one possible embodiment, the initial score is calculated by using the original probability scores of the decision candidate schemes predicted by the large model as a basis, and then correcting them by combining the directional influence weights corresponding to the missing paths affecting the scheme; the larger the absolute value of the directional influence weights, the greater the correction to the original probability score.

[0014] In another aspect, this application also provides a preoperative decision support and assessment system for breast and thyroid diseases based on large models and knowledge graphs, including: A path state vector generation module is used to acquire preoperative diagnosis and treatment data of the target patient and generate a path state vector of the target patient based on a knowledge graph of breast and thyroid disease diagnosis and treatment paths. The path state vector includes at least a path identifier, a missing type code, and a missing stage code. The missing type code indicates whether the path is completely unexecuted, executed but with ambiguous results, or executed but with conflicting results. The missing stage code is generated based on the stage attributes of nodes in the knowledge graph. The nodes of the knowledge graph include disease state nodes, examination behavior nodes, examination result state nodes, and decision nodes. The edges of the knowledge graph describe the sequential dependencies, optional conditions, and prohibited crossing conditions between nodes. The prohibited crossing conditions define the decision types that are prohibited from being generated when the state of a specified preceding node is not satisfied. The structured constraint hint generation module is used to extract and generate corresponding structured constraint hints based on the path state vector; the structured constraint hints include a list of decision types that are currently prohibited from being generated, determined by combining the path identifier, missing type code, and missing stage code in the path state vector with the prohibited crossing conditions in the knowledge graph; The large model decision generation module is used to input the path state vector and the structured constraint hints into the trained large model, and the large model outputs at least one decision candidate and a path sufficiency demonstration for each decision candidate; the path sufficiency demonstration explains the core path on which the decision candidate depends and its completion status in the current path state; The decision screening and output module is used to calculate the path dependency strength and risk stability of each decision candidate scheme based on the path sufficiency demonstration, and to select the final decision suggestion from the decision candidate schemes according to the calculation results; the path dependency strength is determined according to the proportion of completed paths in the core paths on which the decision candidate scheme depends, and the risk stability is determined according to the directional influence weight corresponding to the path missing that affects the decision candidate scheme.

[0015] This application also provides an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the preoperative decision support assessment method for breast and thyroid diseases based on large models and knowledge graphs as described above.

[0016] In another aspect, this application provides a storage medium storing computer program instructions that can be executed by a processor to implement the preoperative decision support assessment method for breast and thyroid diseases based on large models and knowledge graphs as described above.

[0017] Another aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the preoperative decision support assessment method for breast and thyroid diseases based on large models and knowledge graphs as described above.

[0018] This application constructs a knowledge graph of thyroid and breast disease diagnosis and treatment pathways with stage attributes and dynamic constraints, achieving a multi-dimensional and accurate characterization of missing treatment pathways. The generated pathway state vectors comprehensively reflect the type, stage, and direction of decision impact of missing pathways. Structured constraint prompts provide clear constraint boundaries for large-scale model reasoning, effectively avoiding implicit path crossings in large-scale models. The trained large-scale model can generate reliable decision candidate schemes and path sufficiency arguments under constraints. Combining a comprehensive score of path dependence strength and risk stability, the final decision recommendation with the weakest path dependence and the most stable risk is selected. This enables accurate, interpretable, and risk-controllable preoperative decision support for thyroid and breast disease in clinical scenarios with incomplete diagnosis and treatment information, significantly improving the scientific rigor and clinical applicability of preoperative decision-making. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0020] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of a preoperative decision support assessment method for thyroid and breast disease based on a large model and knowledge graph, provided as an embodiment of this application.

[0021] Figure 2 A schematic diagram illustrating the knowledge graph construction process for the diagnosis and treatment pathway of breast and thyroid diseases provided in this application embodiment.

[0022] Figure 3 This is a schematic diagram of the path state vector generation process provided in an embodiment of this application.

[0023] Figure 4 This is a schematic diagram of the structured constraint hint generation process provided in the embodiments of this application.

[0024] Figure 5 This is a schematic diagram of the large model structure provided for an embodiment of this application.

[0025] Figure 6 This is a schematic diagram of the structure of a preoperative decision support and assessment system for thyroid and breast disease based on a large model and knowledge graph, provided in an embodiment of this application.

[0026] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0028] It should be noted that all user information (including but not limited to user device information, user personal information, object information corresponding to device usage data, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, device usage data, etc.) involved in all embodiments of this application are information and data authorized by the user or fully authorized by all parties.

[0029] This method is applicable to preoperative decision support scenarios for thyroid and breast diseases, and is particularly suitable for clinical environments where there are discrepancies in the collection of diagnostic and treatment information in medical institutions at all levels, including primary hospitals, key examination indicators are ambiguous or missing, and treatment pathways do not strictly follow standard guidelines. Implementation can rely on existing HIS / EMR system data interfaces, possessing historical thyroid and breast disease case datasets containing complete and incomplete pathway cases, thyroid and breast disease clinical treatment guideline text data, and computing hardware resources capable of supporting large-scale model deployment. The data complies with relevant medical privacy protection regulations and has the basic computing power for path state quantification and large-scale model inference, ensuring efficient completion of the entire process of knowledge graph construction, path state encoding, and decision reasoning.

[0030] The implementation process of the method described in this application will be described in detail below with reference to specific embodiments. It should be noted that this embodiment is only used to explain this application and is not intended to limit the scope of protection of this application. Conventional adjustments or substitutions of each step by those skilled in the art without departing from the concept of this application should be included in the scope of protection of this application.

[0031] like Figure 1 As shown in the figure, this application discloses a schematic diagram of a preoperative decision support assessment method for breast and thyroid diseases based on a large model and knowledge graph, including the following method steps: S1, Obtain the preoperative diagnosis and treatment data of the target patient, and generate the path state vector of the target patient based on the knowledge graph of the breast and thyroid disease diagnosis and treatment path; the path state vector includes at least a path identifier, a missing type code, and a missing stage code. The missing type code indicates that the path is completely not executed, executed but with ambiguous results, or executed but with conflicting results. The missing stage code is generated based on the stage attributes of the nodes in the knowledge graph. The nodes of the knowledge graph include disease state nodes, examination behavior nodes, examination result state nodes, and decision nodes. The edges of the knowledge graph describe the sequential dependencies, optional conditions, and prohibited crossing conditions between nodes. The prohibited crossing conditions define the decision types that are prohibited from being generated when the state of the specified preceding node is not satisfied. S2, Based on the path state vector, extract and generate corresponding structured constraint hints; the structured constraint hints include a list of decision types currently prohibited from being generated, determined by combining the path identifier, missing type code, and missing stage code in the path state vector with the prohibited crossing conditions in the knowledge graph; S3, The path state vector and the structured constraint hints are input into the trained large model, and the large model outputs at least one decision candidate and a path sufficiency argument for each decision candidate; the path sufficiency argument represents the core path on which the decision candidate depends and its completion status in the current path state; S4. Based on the path sufficiency demonstration, calculate the path dependency strength and risk stability of each decision candidate scheme, and select the final decision suggestion from the decision candidate schemes according to the calculation results; the path dependency strength is determined according to the proportion of completed paths in the core paths on which the decision candidate scheme depends, and the risk stability is determined according to the directional influence weight corresponding to the path missing that affects the decision candidate scheme.

[0032] In some embodiments, please refer to Figure 2 , Figure 2 This is a schematic diagram illustrating the construction process of the knowledge graph for the diagnosis and treatment pathway of breast and thyroid diseases provided in this application embodiment. By extracting relevant nodes and relationships for the diagnosis and treatment of breast and thyroid diseases, and combining stage attributes and dynamic constraints, a knowledge graph for the diagnosis and treatment pathway with the ability to describe path reachability is constructed, providing a foundation for subsequent path state identification and decision constraints.

[0033] Specifically, in S201, core node extraction and definition are performed. Based on clinical treatment guidelines for breast and thyroid diseases and historical case data, structured extraction is conducted using natural language processing technology to identify core nodes, including: disease status nodes to represent the patient's disease diagnosis-related status, such as TI-RADS 4a, suspected ductal carcinoma in situ, etc.; examination behavior nodes to represent examination procedures without specific equipment, such as ultrasound, fine-needle aspiration, gene testing, etc.; examination result status nodes to represent the information output by the examination, including quantitative indicator ranges and qualitative descriptive conclusions; and decision nodes to represent preoperative decision conclusions, such as recommendations for observation, local resection, or total resection, etc. After node extraction, each node is assigned a unique identifier using an alphanumeric combination to ensure the uniqueness of each node in the knowledge graph.

[0034] In S202, each node in the knowledge graph is assigned a stage attribute. Stage attributes include the preoperative basic assessment stage, the preoperative enhanced assessment stage, and the decision confirmation stage, used to distinguish the functional positioning of the node in the diagnosis and treatment process. The preoperative basic assessment stage is a mandatory transitional stage for all cases, including nodes related to core basic examinations, such as thyroid ultrasound and basic breast imaging examinations. The preoperative enhanced assessment stage is a supplementary assessment stage for high-risk or insufficiently informed cases, including nodes related to targeted examinations, such as fine-needle aspiration biopsy and gene testing. The decision confirmation stage is a transitional stage where the final decision is made based on the results of preceding examinations, including nodes related to the summary and analysis of multi-dimensional results. The assignment of stage attributes uses a combination of rule matching and manual verification. First, stage matching rules are set based on treatment guidelines; for example, nodes requiring mandatory initial examinations as specified in the guidelines are assigned the preoperative basic assessment stage by default. Then, medical professionals conduct sampling verification of the assigned values ​​to ensure that the stage attributes are consistent with the actual clinical diagnosis and treatment process. For certain special nodes, such as fine needle biopsy nodes, their stage attributes can be dynamically adjusted. For low-risk breast nodule cases, after confirmation by medical personnel, they can be adjusted to non-mandatory stage nodes, and the adjustment record is stored synchronously in the node metadata.

[0035] In S203, define the relationships between nodes. These relationships include sequential dependencies, optional conditions, and prohibited crossing conditions. The definitions of each type are as follows: Sequential dependencies describe the execution order constraints between nodes; for example, a fine-needle aspiration conclusion node can only be valid if the ultrasound examination node is completed and certain characteristics exist. Optional conditions describe the non-mandatory execution attributes of nodes; for example, the gene testing node for low-risk nodules is not required by all treatment pathways. Prohibited crossing conditions define the decision types that are prohibited from being generated when the current node state is not met. These prohibited crossing conditions are associated with the node's stage attributes; for example, if the ultrasound examination node in the preoperative basic assessment stage is not completed, the generation of any decision type corresponding to a deterministic surgical decision node is prohibited.

[0036] The system includes a dynamic deregulation mechanism to prevent cross-condition configuration. When the target patient's path state vector indicates a missing preceding path node, if an alternative path exists and its confidence level reaches a preset threshold, the prohibition on the corresponding decision type is partially lifted. The determination of the alternative path is based on information compensation relationships in the knowledge graph. For example, certain imaging feature nodes can, to some extent, substitute for missing pathological information nodes, thus constituting an alternative path relationship. Path confidence is calculated based on the consistency between the alternative path and the original path in historical case data. The preset threshold must be set with reference to the risk control requirements of clinical diagnosis and treatment; for example, it can be configured as the lower limit of the 95% confidence interval for outcome consistency to ensure the safety and rationality of the dynamic deregulation mechanism.

[0037] In S204, a knowledge graph is constructed based on the above definitions. The extracted nodes, node stage attributes, and relationships between nodes are entered into the graph database to form a knowledge graph for the diagnosis and treatment pathway of breast and thyroid diseases. The graph database uses a database type that supports graph-structured data storage and efficient querying, such as Neo4j. After construction, medical professionals and technicians can jointly verify the completeness and accuracy of the knowledge graph. Verification includes whether the nodes fully cover the key aspects of breast and thyroid disease diagnosis and treatment, whether the relationships between nodes conform to clinical diagnosis and treatment logic, and whether the prohibited crossing conditions and dynamic release mechanisms are reasonable. Any parts that fail verification need to be corrected and optimized to ensure that the knowledge graph can provide reliable foundational support for subsequent steps.

[0038] In some embodiments, for step S1, this step analyzes the preoperative diagnosis and treatment data of the target patient based on the constructed knowledge graph of the diagnosis and treatment path for breast and thyroid diseases, identifies the missing path states and performs multi-dimensional encoding, generates a path state vector, and provides input data for subsequent generation of structured constraint prompts and large model reasoning.

[0039] Please see Figure 3 , Figure 3This is a schematic diagram of the path state vector generation process provided in this application embodiment. Specifically, it includes, in S301, preoperative diagnosis and treatment data acquisition and preprocessing. Preoperative diagnosis and treatment data of the target patient is acquired through the data interface of an existing HIS / EMR system. The data content includes a list of completed examination items, examination result text, a list of unexecuted examination items, patient basic information, and preliminary disease diagnosis results. The acquired diagnosis and treatment data is preprocessed. First, duplicate and invalid data are removed, such as duplicate records of the same examination result or incorrectly formatted examination item names. Then, natural language processing technology is used to perform structured parsing of the examination result text, extracting quantitative indicators and qualitative descriptions. For example, quantitative indicators such as nodule size and aspect ratio are extracted from ultrasound examination result text, and qualitative descriptions such as nodule properties are extracted from pathology examination result text. For examination result text that cannot be directly structured and parsed, it is marked as awaiting further analysis, providing a basis for subsequent missing type coding.

[0040] In S302, path status identification. Based on the preprocessed preoperative diagnostic and treatment data and the knowledge graph of thyroid and breast disease diagnostic and treatment paths, the path status of the target patient is identified. Path status includes completed paths, missing paths, and paths with uncertain results. For example, a completed path refers to the path corresponding to the examination items that the patient has performed and whose results are clear, that is, the completed examination items in the diagnostic and treatment data can be completely matched with the examination behavior nodes and examination result status nodes in the knowledge graph; a missing path refers to the path corresponding to the examination items that the patient has not performed, that is, the unperformed examination items in the diagnostic and treatment data correspond to the examination behavior nodes in the knowledge graph, but there is no corresponding examination result status node; an uncertain result path refers to the path corresponding to the examination items that the patient has performed but whose results cannot be clearly determined, that is, the examination result text still cannot be accurately matched with the examination result status nodes in the knowledge graph after preprocessing, including two situations: fuzzy results and conflicting results.

[0041] In S303, path status is encoded in multiple dimensions. Identified missing paths and paths with uncertain outcomes are encoded in multiple dimensions. The encoding content includes path identifier, missing type encoding, missing stage encoding, and decision impact direction encoding. The specific implementation of each encoding dimension is as follows: The path identifier uses the unique identifier of the corresponding path node in the knowledge graph to ensure that the code corresponds one-to-one with the path node in the knowledge graph, which facilitates subsequent association queries.

[0042] Missing type coding indicates the missing type of the path. Missing types include three categories: completely unexecuted, executed but with ambiguous results, and executed but with conflicting results, corresponding to coding values ​​T1, T2, and T3, respectively. The coding for completely unexecuted paths is based on a list of unexecuted examination items clearly recorded in the diagnostic data; that is, the examination behavior node corresponding to the path identifier has no relevant execution record, coded as T1. The coding for executed but ambiguous results paths is based on the fact that the examination result text, after preprocessing, cannot extract clear quantitative indicators or qualitative descriptions, such as an examination result describing the nodule's nature as undetermined, coded as T2. The coding for executed but conflicting results paths is based on the fact that the same indicator yields contradictory results from different examination items, such as ultrasound indicating benign nodule characteristics while fine-needle aspiration suggests suspected malignancy, coded as T3.

[0043] Missing stage codes are generated based on the stage attributes of nodes in the knowledge graph. Each path node in the knowledge graph has predefined stage attributes, including the preoperative basic assessment stage, the preoperative enhanced assessment stage, and the decision confirmation stage, corresponding to the code values ​​S1, S2, and S3, respectively. By matching the stage attributes of the nodes corresponding to the identified missing paths with paths where the outcome is uncertain, the corresponding missing stage code can be obtained. For example, if the stage attribute of a node corresponding to a missing path is the preoperative basic assessment stage, then its missing stage code is S1.

[0044] The direction of decision impact is encoded based on a pre-generated path-decision relationship table, which records the directional impact weights of each path identifier on different types of decision options. These directional impact weights are generated through the following steps: A historical case dataset was constructed, containing a large amount of preoperative treatment pathway data, decision-making scheme types, and postoperative risk indicator data for patients with breast and thyroid diseases. Postoperative risk indicators include recurrence rate and complication rate. Based on the historical case dataset, the differences in postoperative risk indicators corresponding to different types of decision-making schemes were statistically analyzed under the conditions of missing and not missing specific pathways. For example, the difference in postoperative recurrence rate between conservative and aggressive decision-making schemes when the pathological subtype pathway was missing, and the corresponding difference in recurrence rate when the pathological subtype pathway was not missing, were statistically analyzed. The statistically obtained differences in risk indicators were normalized using the min-max normalization method, mapping the differences in risk indicators to the interval [-1,1] to obtain the initial directional influence weights. The initial directional influence weights were calibrated in conjunction with medical guideline constraints. For example, if the medical guideline clearly indicates that the absence of a certain pathway increases the risk of the aggressive decision-making scheme, the directional influence weight of the aggressive decision-making scheme corresponding to that pathway was adjusted upward to avoid weight distortion caused by single-center data bias. Finally, the calibrated directional influence weights were obtained and stored in the path-decision relationship table.

[0045] The decision impact direction coding includes two categories: impact on the risk of conservative options and impact on the risk of aggressive options, corresponding to coding values ​​D1 and D2, respectively. The decision impact direction coding is determined based on the sign and magnitude of the directional impact weight corresponding to each path identifier in the path-decision relationship table. For example, if the directional impact weight of a path identifier corresponding to a conservative option is positive, it indicates that the absence of this path will increase the risk of the conservative option, and its decision impact direction coding is D1; ​​if the directional impact weight of the corresponding aggressive option is negative, it indicates that the absence of this path will reduce the risk of the aggressive option, and its decision impact direction coding is D2.

[0046] In S304, the path state vector is combined. The path identifier, missing type encoding, missing stage encoding, and decision influence direction encoding are combined to generate the target patient's path state vector. The expression for the path state vector is:

[0047] Where V represents the path state vector, I represents the path identifier, T represents the missing type code, S represents the missing stage code, and D represents the decision influence direction code. For example, if the path identifier of a path is P001, the missing type code is T1, the missing stage code is S1, and the decision influence direction code is D2, then its corresponding path state vector is [P001, T1, S1, D2]. All missing paths and paths with uncertain outcomes for the target patient need to generate corresponding path state vectors to form a set of path state vectors, which will serve as input data for subsequent steps.

[0048] In some embodiments, for step S2, based on the generated set of path state vectors, combined with the prohibited crossing conditions and dynamic release mechanisms in the knowledge graph of breast and thyroid disease diagnosis and treatment paths, a list of decision types that are prohibited from being generated is extracted, and structured constraint hints are generated in a structured format to provide clear constraint boundaries for large model reasoning.

[0049] Please see Figure 4 , Figure 4 This is a schematic diagram of the structured constraint prompt generation process provided in the embodiments of this application. Specifically, it includes, in S401, matching the state of the preceding path node. Each path state vector in the path state vector set is traversed, and the corresponding preceding path node is matched from the breast and thyroid disease diagnosis and treatment path knowledge graph based on the path identifier. The state requirements of the preceding path node and the associated prohibited crossing conditions are obtained. The state requirements of the preceding path node include whether it must be completed, the result threshold after completion, etc., and the associated prohibited crossing conditions include the decision type that is prohibited from being generated when the preceding path node does not meet the state requirements.

[0050] For each preceding path node, based on the missing type and missing stage codes in the path state vector, it is determined whether its current state meets the triggering condition for the prohibition of crossing. If the path state vector indicates that the preceding path node is completely unexecuted, executed but with ambiguous results, or executed but with conflicting results, and the preceding path node is a mandatory stage node, then the corresponding prohibition of crossing condition is triggered. If the preceding path node is a non-mandatory stage node, it is necessary to combine the risk assessment results of clinical diagnosis and treatment to determine whether the prohibition of crossing condition is triggered. For example, if a non-mandatory stage gene testing node is missing, and the patient's disease risk is high, then the prohibition of crossing condition for preventing the generation of aggressive decision types is triggered.

[0051] In S402, the dynamic release mechanism is verified. For preceding path nodes that trigger the prohibition-crossing condition, it is further verified whether the dynamic release mechanism is satisfied. Based on the information compensation relationship in the knowledge graph, the alternative path corresponding to the preceding path node is searched. If an alternative path exists, the path credibility of the alternative path is calculated. The path credibility of the alternative path is determined by the degree of consistency between the results of the alternative path and the original path in historical case data. For example, the accuracy rate of decision-making using imaging features to replace pathological information in historical cases is statistically analyzed, and this accuracy rate is the path credibility of the alternative path.

[0052] The calculated reliability of alternative paths is compared with a preset threshold. If the path reliability reaches or exceeds the preset threshold, the corresponding prohibition condition is partially lifted, allowing the generation of some decision types while prohibiting the generation of high-risk decision types. If the path reliability does not reach the preset threshold, the prohibition condition remains in effect, prohibiting the generation of all corresponding decision types. For example, when a pathological subtype path is missing, if the path reliability of the image feature alternative path reaches the preset threshold, the generation of local resection decision types is allowed while the generation of extended resection decision types is prohibited; if the path reliability does not meet the threshold, the generation of all surgery-related decision types is prohibited.

[0053] In S403, a list of prohibited decision types is compiled. Based on the results of the aforementioned pre-path node status matching and dynamic deregulation mechanism verification, a list of currently prohibited decision types is compiled. The list includes the name of the prohibited decision type, the corresponding triggering pre-path node, and the prohibition status (complete or partial prohibition). For example, if the prohibited decision type is extended resection, the corresponding triggering pre-path node is the pathological subtype examination node, and the prohibition status is complete prohibition; if the prohibited decision type is radical surgery, the corresponding triggering pre-path node is the gene testing node, and the prohibition status is partial prohibition, meaning it can be deregulated when the path credibility meets the standard.

[0054] In S404, the organization generates structured constraint hints. The compiled list of prohibited decision types is organized in a key-value pair structured format to generate structured constraint hints. An example of the structured constraint hint format is as follows: Prohibited decision type: Extended resection; Pre-triggered node: Pathological subtype examination node; Prohibited status: Completely prohibited. Prohibited decision type: radical surgery; triggering node: gene testing node; prohibition status: partial prohibition. The generation of structured constraint hints must ensure that the format is standardized and the information is clear, so that large models can quickly parse and follow the constraints, avoiding constraint failure due to ambiguous format. After generation, the structured constraint hints need to be validated to ensure that the information such as prohibited decision types and triggering prerequisite nodes is accurate and consistent with the constraints in the knowledge graph.

[0055] In some embodiments, for step S3, the generated set of path state vectors and structured constraint hints are input into the trained large model, which generates decision candidate schemes and corresponding path sufficiency arguments under constraints, providing a basis for the final decision suggestion selection.

[0056] In one embodiment, such as Figure 5 As shown, Figure 5 This is a schematic diagram of the large model structure provided in an embodiment of this application. The large model adopts an encoder-decoder structure based on the Transformer architecture, specifically including an input layer, an encoding layer, a cross-attention layer, a decoding layer, and an output layer.

[0057] The input layer receives the path state vector set and structured constraint cues, transforming them into vector forms that the model can process. The path state vector set is transformed into vectors through a combination of node embedding and encoding embedding, while the structured constraint cues are transformed into vectors through rule embedding. The encoding layer employs a multi-layer Transformer encoder, with each layer containing a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism captures the internal correlation information of the input data, while the feedforward neural network performs non-linear transformations on the encoded features. The number of encoding layers can be configured from 12 to 24, and the number of heads in the multi-head self-attention mechanism can be configured as follows. The number of heads ranges from 12 to 16, and the hidden layer dimension of the feedforward neural network can be configured from 2048 to 4096 dimensions. The cross-attention layer is used to fuse the path topology information in the knowledge graph with the feature vector output by the encoding layer, enabling the model to make full use of the constraints of the knowledge graph for reasoning. The decoding layer adopts a multi-layer Transformer decoder, which generates decision candidate schemes and path sufficiency arguments through autoregression. The number of layers in the decoder is consistent with that of the encoding layer, and the hidden layer dimension matches that of the encoding layer. The output layer adopts the softmax activation function to predict the probability of decision candidate schemes and output the text content of the path sufficiency argument.

[0058] The training process of the large model includes three stages: data preparation, model pre-training, and model fine-tuning. In the data preparation stage, historical case data, treatment guideline text data, and knowledge graph data related to preoperative diagnosis and treatment of breast and thyroid diseases are collected to construct the training dataset. The input of the training dataset includes a set of path state vectors for historical cases and corresponding structured constraint hints. The output includes the final decision-making plan adopted by clinicians and the rationality analysis of the path, i.e., reference text for the sufficiency demonstration of the path. The training dataset is divided into a training set, a validation set, and a test set, with a configurable ratio of 7:2:1 to ensure the effectiveness of the training process and the accuracy of model performance evaluation.

[0059] During the model pre-training phase, the large model is pre-trained based on collected diagnostic and treatment guideline text data and knowledge graph data, enabling the model to learn professional knowledge and path constraint logic in the field of breast and thyroid disease diagnosis and treatment. Pre-training employs a combination of masked language modeling tasks and path relationship prediction tasks. The masked language modeling task enhances the model's understanding of diagnostic and treatment texts, while the path relationship prediction task improves the model's mastery of path constraint relationships in the knowledge graph.

[0060] The path relationship prediction task specifically involves sampling node pairs from the knowledge graph, such as disease node-inspection node and inspection node-decision node, and randomly masking one type of relationship, such as sequential dependency or conditional optionality. The model is required to predict the masked relationship type based on the context. Alternatively, given a decision node, the model is required to select its necessary preceding path nodes from the candidate set.

[0061] The number of pre-training iterations can be configured to 10 to 20 rounds, the learning rate can be configured to 1e-5 to 5e-5, and an adaptive moment estimation optimizer is used for parameter updates.

[0062] In the model fine-tuning phase, based on the constructed training and validation sets, the pre-trained large model is fine-tuned to adapt it to the specific task of preoperative decision support for breast and thyroid diseases. During fine-tuning, the difference between the decision candidate schemes predicted by the large model and the path sufficiency arguments and the labeled results in the training dataset is used as the loss function. The loss function is a weighted sum of the cross-entropy loss function and the mean squared error loss function. The cross-entropy loss function is used to optimize the prediction accuracy of the decision candidate schemes, and the mean squared error loss function is used to optimize the text generation quality of the path sufficiency arguments. The number of fine-tuning iterations can be configured to 20 to 30 rounds, and the learning rate adopts a cosine annealing scheduling strategy. The initial learning rate can be configured to 5e-6 to 1e-5. Fine-tuning is stopped when the loss function value of the validation set does not decrease for 5 consecutive rounds to avoid model overfitting. After fine-tuning, the model performance is evaluated using a test set. Evaluation metrics include the accuracy of the decision candidate schemes, the logical coherence and accuracy of the path sufficiency arguments, etc. Only after the evaluation meets the standards can the model be used in the subsequent inference process.

[0063] The set of path state vectors and structured constraint hints are input into a trained large model. The large model first generates an initial set of decision candidate solutions under the constraints of the structured constraint hints, based on the input data and its learned domain knowledge and path constraint logic. During generation, the large model adheres to the prohibited decision types requirements in the structured constraint hints and must not generate decision types explicitly prohibited in the list. Simultaneously, the large model performs an initial score on each generated decision candidate solution. The initial score is calculated based on the original probability score predicted by the large model for the decision candidate solution, and then adjusted by incorporating the directional influence weights corresponding to the missing paths affecting that solution.

[0064] The formula for calculating the initial score is:

[0065] Where S represents the initial score of the decision candidate, This represents the original probability score of the decision candidate predicted by the large model, where n represents the number of missing paths affecting that decision candidate. This represents the directional impact weight corresponding to the missing k-th path. According to this formula, the larger the absolute value of the directional impact weight, the greater the correction to the original probability score. In other words, the more significant the impact of missing paths on the decision candidate, the lower the initial score, ensuring that the initial score reflects the degree of impact of missing paths on the decision.

[0066] After generating the initial set of decision candidate solutions, an initial scoring threshold is set. The initial scoring threshold should be set with reference to the risk tolerance of clinical decision-making. For example, it can be configured to be 0.3 to 0.5. Decision candidate solutions with initial scores higher than this threshold are selected to form the initial set of decision candidate solutions, ensuring that the decision candidate solutions in the set have a certain degree of reliability.

[0067] In one embodiment, for each option in the initial set of decision candidate options, the large model generates a corresponding path sufficiency argument.

[0068] This generation process leverages the text generation capabilities of a finely tuned large model. Its inputs include the current decision candidate, a set of path state vectors, and the core path topology information associated with the chosen option retrieved from a knowledge graph. The model autoregressively generates coherent text containing a list of core dependent paths, completion status, and impact analysis via a decoder. To improve the accuracy and standardization of the arguments, the path sufficiency arguments in the training dataset follow a three-part structure: "listing core paths -> judging completion status -> analyzing the impact of missing paths," allowing the model to learn this generation paradigm.

[0069] The sufficiency of the path argument needs to explain the core paths on which the candidate decision solution depends and their completion status in the current path state. Specifically, it includes three parts: a list of core dependent paths, the completion status of each core dependent path, and an analysis of the impact of missing paths on the decision solution.

[0070] The core dependency path list is determined based on the sequential dependencies in the knowledge graph and the associated paths of decision nodes. For example, the core dependency paths for a local resection decision include ultrasound examination and preliminary pathological diagnosis. The completion status of each core dependency path is determined based on information in the path state vector set, i.e., completed, missing, or with uncertain results. The impact analysis of missing paths on the decision is based on directional influence weights, explaining whether missing paths increase or decrease the risk of the decision and the degree of impact. The generation of path sufficiency arguments must ensure clear logic and sufficient evidence, providing clear basis for doctors to understand the rationality of the decision.

[0071] In some embodiments, for step S4, this step is based on path sufficiency argumentation, calculates the path dependence strength and risk stability of each decision candidate, selects the optimal final decision recommendation through comprehensive scoring, and supports dynamic decision updates after the diagnosis and treatment data is updated.

[0072] Path dependency strength is determined based on the proportion of completed paths in the core paths upon which the decision candidate depends, and is calculated using the following formula:

[0073] Where D represents the path dependence strength, This indicates the number of completed paths in the core path upon which the decision candidate solution depends. This represents the total number of core paths upon which the decision candidate solution depends. The path dependency strength ranges from 0 to 1; a higher value indicates a higher degree of completion of the core paths upon which the decision candidate solution depends, and thus a higher reliability of the decision.

[0074] Risk stability is determined based on the directional impact weight corresponding to the path missing that affects the candidate decision solution, and the calculation formula is as follows:

[0075] Where R represents risk stability, and n represents the number of missing paths affecting the candidate solution for this decision. This represents the directional impact weight corresponding to the missing k-th path. The risk stability value ranges from 0 to 1; a larger value indicates that the missing path has a smaller impact on the risk of the decision candidate, and the risk of the decision is more stable.

[0076] The final decision recommendation is selected based on the highest comprehensive score, which is a weighted sum of path dependence strength and risk stability, calculated using the following formula:

[0077] Where F represents the overall score, The weighting coefficient represents the strength of path dependence, ranging from 0 to 1. It can be adjusted according to the focus of clinical diagnosis and treatment, for example, when more emphasis is placed on path integrity. It can be configured to 0.6; when a greater emphasis is placed on risk stability, It can be configured to 0.4.

[0078] After calculating the comprehensive score of each decision candidate, the candidate options are ranked from highest to lowest comprehensive score, and the candidate option with the highest comprehensive score is selected as the final decision recommendation. If multiple decision candidate options have the same and highest comprehensive score, the final determination must be made by a physician, taking into account the detailed content of the pathway sufficiency demonstration, to ensure the rationality and applicability of the decision recommendation.

[0079] This method constructs a knowledge graph of thyroid and breast disease (MTD) diagnosis and treatment pathways with stage attributes and dynamic constraints, achieving a multi-dimensional and accurate characterization of missing pathways. The generated pathway state vectors comprehensively reflect the type, stage, and direction of decision impact of missing pathways. Structured constraint prompts provide clear constraint boundaries for large-scale model inference, effectively avoiding implicit path crossings. The trained large-scale model can generate reliable decision candidate solutions and path sufficiency arguments under constraints. Combining a comprehensive score of path dependence strength and risk stability, the final decision recommendation with the weakest path dependence and the most stable risk is selected. This enables accurate, interpretable, and risk-controllable preoperative decision support for MTD in clinical scenarios with incomplete diagnosis and treatment information, significantly improving the scientific rigor and clinical applicability of preoperative decision-making.

[0080] It should be noted that although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. On the contrary, the steps depicted in the flowchart can be performed in a different order. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0081] Please see Figure 6 , Figure 6 This is a schematic diagram of the structure of a preoperative decision support and assessment system for breast and thyroid diseases based on a large model and knowledge graph, provided in an embodiment of this application. As shown in the figure, the system includes: The path state vector generation module 601 is used to acquire preoperative diagnosis and treatment data of the target patient and generate a path state vector of the target patient based on the knowledge graph of the diagnosis and treatment path for breast and thyroid diseases. The path state vector includes at least a path identifier, a missing type code, and a missing stage code. The missing type code indicates that the path is completely unexecuted, executed but with ambiguous results, or executed but with conflicting results. The missing stage code is generated based on the stage attributes of nodes in the knowledge graph. The nodes of the knowledge graph include disease state nodes, examination behavior nodes, examination result state nodes, and decision nodes. The edges of the knowledge graph describe the sequential dependencies, optional conditions, and prohibited crossing conditions between nodes. The prohibited crossing conditions define the decision types that are prohibited from being generated when the state of a specified preceding node is not satisfied. The structured constraint hint generation module 602 is used to extract and generate corresponding structured constraint hints based on the path state vector; the structured constraint hints include a list of decision types that are currently prohibited from being generated, determined by combining the path identifier, missing type code, and missing stage code in the path state vector with the prohibited crossing conditions in the knowledge graph; The large model decision generation module 603 is used to input the path state vector and the structured constraint hints into the trained large model, and the large model outputs at least one decision candidate and a path sufficiency demonstration for each decision candidate; the path sufficiency demonstration explains the core path on which the decision candidate depends and its completion status in the current path state; The decision screening and output module 604 is used to calculate the path dependency strength and risk stability of each decision candidate scheme based on the path sufficiency demonstration, and to select the final decision suggestion from the decision candidate schemes according to the calculation results; the path dependency strength is determined according to the proportion of completed paths in the core paths on which the decision candidate scheme depends, and the risk stability is determined according to the directional influence weight corresponding to the path missing that affects the decision candidate scheme.

[0082] Those skilled in the art will clearly understand that the technical solutions of the embodiments of this application can be implemented by means of software and / or hardware. In this specification, "unit" and "module" refer to software and / or hardware that can independently complete or cooperate with other components to complete a specific function, wherein the hardware may be, for example, a field-programmable gate array (FPGA), an integrated circuit (IC), etc.

[0083] Each processing unit and / or module in the embodiments of this application can be implemented by an analog circuit that implements the functions described in the embodiments of this application, or by software that executes the functions described in the embodiments of this application.

[0084] Please see Figure 7 It shows a schematic diagram of the structure of an electronic device according to an embodiment of this application, which can be used to implement... Figure 1 The method in the illustrated embodiment. (As shown) Figure 7 As shown, the electronic device may include: The system includes at least one processor 701, at least one network interface 704, a user interface 703, a memory 705, and at least one communication bus 702. The communication bus 702 is used to enable connection and communication between the components. The user interface 703 may include buttons, and optionally include a standard wired or wireless interface. The network interface 704 may include, but is not limited to, a Bluetooth module, an NFC module, a Wi-Fi module, etc.

[0085] The processor 701 may include one or more processing cores and connect to various parts within the electronic device 700 via various interfaces and lines. It implements various functions and data processing of the device 700 by running or executing instructions, programs, code sets, or instruction sets stored in the memory 705, and by accessing data in the memory 705. Optionally, the processor 701 may be implemented using at least one hardware form of DSP, FPGA, or PLA. The processor 701 may also integrate one or more combinations of CPU, GPU, and modem.

[0086] The memory 705 may include random access memory (RAM) or read-only memory (ROM). Optionally, the memory 705 includes a non-transitory computer-readable medium for storing instructions, programs, code, code sets, or instruction sets. The memory 705 may be divided into a program storage area and a data storage area, wherein the program storage area can be used to store instructions for implementing an operating system and instructions for implementing the foregoing method embodiments; the data storage area can be used to store data related to the relevant method embodiments. The memory 705 may also be at least one storage device located remotely from the processor 701. Figure 7 As shown, the memory 705, which serves as a computer storage medium, may contain an operating system, a network communication module, a user interface module, and program instructions.

[0087] In particular, the methods and / or embodiments in this application can be implemented as computer software programs. For example, the embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. When the computer program is executed by processor 701, the functions defined in the methods of this application are performed.

[0088] Another embodiment of this application provides a storage medium storing computer program instructions thereon, which can be executed by a processor to implement the methods and / or technical solutions of any one or more embodiments of this application.

[0089] In the above embodiments, the descriptions of each embodiment have different focuses. Parts not described in detail in a certain embodiment can be referred to in the relevant descriptions of other embodiments. The above descriptions are merely preferred embodiments of this application and explanations of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to the technical solutions formed by specific combinations of the above technical features, but should also cover other technical solutions formed by arbitrary combinations of the above technical features or their equivalent features without departing from the inventive concept.

Claims

1. A preoperative decision support assessment method for breast and thyroid disease based on large models and knowledge graphs, characterized in that, include: S1, Obtain the preoperative diagnosis and treatment data of the target patient, and generate the path state vector of the target patient based on the knowledge graph of the breast and thyroid disease diagnosis and treatment path; the path state vector includes at least a path identifier, a missing type code, and a missing stage code. The missing type code indicates that the path is completely not executed, executed but with ambiguous results, or executed but with conflicting results. The missing stage code is generated based on the stage attributes of the nodes in the knowledge graph. The nodes of the knowledge graph include disease state nodes, examination behavior nodes, examination result state nodes, and decision nodes. The edges of the knowledge graph describe the sequential dependencies, optional conditions, and prohibited crossing conditions between nodes. The prohibited crossing conditions define the decision types that are prohibited from being generated when the state of the specified preceding node is not satisfied. S2, Based on the path state vector, extract and generate corresponding structured constraint hints; the structured constraint hints include a list of decision types currently prohibited from being generated, determined by combining the path identifier, missing type code, and missing stage code in the path state vector with the prohibited crossing conditions in the knowledge graph; S3, The path state vector and the structured constraint hints are input into the trained large model, and the large model outputs at least one decision candidate and a path sufficiency demonstration for each decision candidate; the path sufficiency demonstration explains the core path on which the decision candidate depends and its completion status in the current path state; S4. Based on the path sufficiency demonstration, calculate the path dependency strength and risk stability of each decision candidate scheme, and select the final decision suggestion from the decision candidate schemes according to the calculation results; the path dependency strength is determined according to the proportion of completed paths in the core paths on which the decision candidate scheme depends, and the risk stability is determined according to the directional influence weight corresponding to the path missing that affects the decision candidate scheme.

2. The method for preoperative decision support assessment of thyroid and breast disease based on large models and knowledge graphs according to claim 1, characterized in that, It also includes defining stage attributes for nodes in the knowledge graph, wherein the stage attributes include the preoperative basic assessment stage, the preoperative enhanced assessment stage, and the decision confirmation stage; the prohibition on crossing conditions are associated with the stage attributes of the nodes.

3. The method for preoperative decision support assessment of thyroid and breast disease based on large models and knowledge graphs as described in claim 1, characterized in that: When the path state vector of the target patient indicates that a certain preceding path node is missing, if there is an alternative path based on the information compensation relationship in the knowledge graph and its path credibility reaches a preset threshold, then the prohibition on the corresponding decision type is partially lifted.

4. The method for preoperative decision support assessment of thyroid and breast disease based on large models and knowledge graphs according to claim 1, characterized in that, The process of generating a path state vector also includes determining the direction encoding of decision influence based on a pre-generated path-decision relationship table, which records the directional influence weight of each path identifier on different types of decision schemes.

5. The method for preoperative decision support assessment of thyroid and breast disease based on large models and knowledge graphs according to claim 4, characterized in that, Based on historical case datasets, the differences in postoperative risk indicators under different decision-making schemes with and without specific pathways were statistically analyzed. The differences in risk indicators were normalized and calibrated in conjunction with medical guideline constraints to obtain the directional influence weights.

6. The method for preoperative decision support assessment of thyroid and breast disease based on large models and knowledge graphs according to claim 1, characterized in that, The structured constraint hint is generated by matching the preceding path nodes in the knowledge graph that are incomplete and do not meet the dynamic release mechanism of the prohibited crossing condition from the path state vector. The prohibition decision types associated with these preceding path nodes are organized in a key-value pair structured format to generate the structured constraint hint.

7. The method for preoperative decision support assessment of thyroid and breast disease based on large models and knowledge graphs according to claim 1, characterized in that, The large model generates an initial set of decision candidate solutions under the constraints of the structured constraint hints based on the path state vector and the structured constraint hints, and performs an initial score on each decision candidate solution; For each of the initial set of decision candidate solutions, generate the path sufficiency argument.

8. The method for preoperative decision support assessment of thyroid and breast disease based on large models and knowledge graphs according to claim 7, characterized in that, The initial score is calculated by taking the original probability score of the decision candidate scheme predicted by the large model as the basis and correcting it by combining the directional influence weights corresponding to the missing paths affecting the scheme; the larger the absolute value of the directional influence weight, the greater the correction of the original probability score.

9. A preoperative decision support and assessment system for breast and thyroid diseases based on large models and knowledge graphs, characterized in that, include: A path state vector generation module is used to acquire preoperative diagnosis and treatment data of the target patient and generate a path state vector of the target patient based on a knowledge graph of breast and thyroid disease diagnosis and treatment paths. The path state vector includes at least a path identifier, a missing type code, and a missing stage code. The missing type code indicates whether the path is completely unexecuted, executed but with ambiguous results, or executed but with conflicting results. The missing stage code is generated based on the stage attributes of nodes in the knowledge graph. The nodes of the knowledge graph include disease state nodes, examination behavior nodes, examination result state nodes, and decision nodes. The edges of the knowledge graph describe the sequential dependencies, optional conditions, and prohibited crossing conditions between nodes. The prohibited crossing conditions define the decision types that are prohibited from being generated when the state of a specified preceding node is not satisfied. The structured constraint hint generation module is used to extract and generate corresponding structured constraint hints based on the path state vector; the structured constraint hints include a list of decision types that are currently prohibited from being generated, determined by combining the path identifier, missing type code, and missing stage code in the path state vector with the prohibited crossing conditions in the knowledge graph; The large model decision generation module is used to input the path state vector and the structured constraint hints into the trained large model, and the large model outputs at least one decision candidate and a path sufficiency demonstration for each decision candidate; the path sufficiency demonstration explains the core path on which the decision candidate depends and its completion status in the current path state; The decision screening and output module is used to calculate the path dependency strength and risk stability of each decision candidate scheme based on the path sufficiency demonstration, and to select the final decision suggestion from the decision candidate schemes according to the calculation results; the path dependency strength is determined according to the proportion of completed paths in the core paths on which the decision candidate scheme depends, and the risk stability is determined according to the directional influence weight corresponding to the path missing that affects the decision candidate scheme.

10. An electronic device, characterized in that, include: At least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 8.

Images