Large model optimization method and system fused with medical knowledge graph

Abstract

The invention relates to the technical field of natural language processing and medical artificial intelligence, and discloses a large model optimization method and system fusing a medical knowledge graph, and the method comprises the steps: constructing a time fuzziness-confidence joint modeling network model, time information corresponding to a current chief complaint symptom, context chief complaint content and patient individual information are input, a normal estimation time prediction result is output, joint calibration is conducted on the normal estimation time prediction result based on causal path information in the medical knowledge graph, and time prior distribution on the graph side is obtained; and constructing joint distribution based on the normal estimation time prediction result and the time prior distribution at the map side, taking the expectation of the joint distribution as a time point prediction result, inputting the time point prediction result into a large model, outputting a structured suggestion, and feeding back the structured suggestion to the medical knowledge map for updating. Therefore, the problem that inference is wrong due to the fact that semantic dependence cannot be recognized or patient background factors are ignored is solved.

Description

Technical Field

[0001] The present invention relates to the technical fields of natural language processing and medical artificial intelligence, and particularly to a large model optimization method and system integrating a medical knowledge graph. Background Art

[0002] In the process of clinical diagnosis and treatment, the patient's chief complaint is the first-hand information for obtaining their condition. However, these chief complaints often contain a large amount of vague and uncertain information, especially time-related descriptions. For example, time expressions such as "about two or three days" and "about one week after surgery" are extremely common in natural language. Although these descriptions are somewhat understandable to human doctors, they lack operability and standardization for large model systems that rely on structured information for reasoning and advice.

[0003] Currently, mainstream methods mostly rely on template matching or rule-based time standardization strategies. Although certain effects can be achieved under standardized sentence patterns, for real chief complaint scenarios with arbitrary language expressions and complex contexts, reasoning errors often occur due to the inability to recognize semantic dependencies or the neglect of patient background factors. For example, the "fever" in "fever about one week during the postoperative recovery period" may be related to postoperative infection or other complication manifestations. Its time point is vague but extremely critical. This requires the system to not only understand the language but also comprehensively judge by combining medical knowledge, temporal relationships, and patient background. Summary of the Invention

[0004] The objective of the present invention is to establish a large model optimization system that can accurately understand fuzzy time expressions and perform reasoning by combining medical graph knowledge, which is the key path to solving the above problems.

[0005] To achieve the above objective, the present invention provides a large model optimization method integrating a medical knowledge graph, including: Extracting fuzzy time keywords and disease keywords from the chief complaint text; Based on the correspondence between the fuzzy time keywords and the disease keywords, extracting a subgraph from the medical knowledge graph and using a graph attention network to perform node representation learning on the subgraph, where the input is the keyword set in the chief complaint text and the output is the time information corresponding to the current chief complaint symptom; Constructing a time ambiguity-confidence joint modeling network model, where the input includes the time information corresponding to the current chief complaint symptom, the context chief complaint content, and the patient individual information, the output is the normal estimated time prediction result, and a joint loss function is constructed to train the time ambiguity-confidence joint modeling network model; Based on the causal path information in the medical knowledge graph, jointly calibrating the normal estimated time prediction result to obtain the time prior distribution on the graph side; Construct a joint distribution based on the normal estimation time prediction result and the time prior distribution on the atlas side, and use the expectation of the joint distribution as the time point prediction result; Input the time point prediction result into a large model to output structured suggestions, and feedback the structured suggestions to the medical knowledge atlas for updating.

[0006] Optionally, the extraction of fuzzy time keywords in the chief complaint text includes: Encode the chief complaint text into a sequence of word vectors; Build a pointer network, take the text sequence of the word vector sequence as input, and output the start and end positions of the fuzzy time phrase as the fuzzy time keywords.

[0007] Optionally, the use of a graph attention network for node representation learning of the subgraph includes: The goal of the graph attention network is to maximize the association between related disease or surgical procedure nodes and their recovery period nodes, and adopt a multi-task loss of node classification and edge type prediction.

[0008] Optionally, the joint calibration of the normal estimation prediction result based on the causal path information in the medical knowledge atlas to obtain the time prior distribution on the atlas side includes: Based on the normal estimation time prediction result, identify a set of candidate paths in the medical knowledge atlas that have a causal association with the target symptom in the current chief complaint, and extract the corresponding reference time distribution for the entities and relationship edges on each path in the candidate path set; Calculate the validity weight corresponding to the reference time distribution, and construct the time prior distribution on the atlas side based on the validity weight.

[0009] Optionally, the input of the large model includes: the time point prediction result, chief complaint symptoms, surgical type, and patient background.

[0010] Optionally, the feedback of the structured suggestions to the medical knowledge atlas for updating includes: Aggregate each structured suggestion according to confidence buckets and then feedback it to the medical knowledge atlas to promote the continuous update and evolution of the atlas structure.

[0011] Optionally, the aggregation of each structured suggestion according to confidence buckets and then feedback it to the medical knowledge atlas to promote the continuous update and evolution of the atlas structure includes: Based on the time point and standard deviation in the structured suggestion, construct a corresponding confidence time interval; Model the fuzzy time expression in the confidence time interval as an attribute on the node or edge in the atlas structure to construct a structured time evolution relationship; Perform continuous update and evolution according to the described structured time evolution relationship.

[0012] On the other hand, this application also discloses a large model optimization system integrating a medical knowledge graph, including: Extraction module: used to extract fuzzy time keywords and disease keywords in the chief complaint text; Learning module: used to extract a subgraph from the medical knowledge graph based on the correspondence between the fuzzy time keywords and disease keywords, and use a graph attention network to perform node representation learning on the subgraph. Among them, the input is the keyword set in the chief complaint text, and the output is the time information corresponding to the current chief complaint symptom; Prediction training module: used to construct a time ambiguity-confidence joint modeling network model. Among them, the input includes the time information corresponding to the current chief complaint symptom, the context chief complaint content, and the patient's individual information, the output is the normal estimation time prediction result, and a joint loss function is constructed to train the time ambiguity-confidence joint modeling network model; Joint verification module: used to jointly calibrate the normal estimation time prediction result based on the causal path information in the medical knowledge graph to obtain the time prior distribution on the graph side; Prediction result module: construct a joint distribution based on the normal estimation time prediction result and the time prior distribution on the graph side, and use the expectation of the joint distribution as the time point prediction result; Result update module: input the time point prediction result into the large model to output a structured suggestion, and feedback the structured suggestion to the medical knowledge graph for update.

[0013] Compared with the prior art, the large model optimization method and system integrating a medical knowledge graph in the embodiments of the present invention have the beneficial effect that: a large model optimization system that can accurately understand fuzzy time expressions and perform reasoning in combination with medical graph knowledge can be used as a key path to solve the problem of reasoning errors caused by the inability to recognize semantic dependencies or ignoring patient background factors. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is a schematic flowchart of a large model optimization method integrating a medical knowledge graph in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0015] The following will further describe in detail the specific embodiments of the present invention in conjunction with the drawings and embodiments. The following embodiments are used to illustrate the present invention, but are not used to limit the scope of the present invention.

[0016] As Figure 1 shown, a large model optimization method integrating a medical knowledge graph in a preferred embodiment of an embodiment of the present invention includes the following steps: S101: Extract the fuzzy time keywords and disease keywords from the chief complaint text; It should be noted that in the actual clinical diagnosis process, patients often cannot provide accurate time due to memory blur or unclear symptom severity during expression. However, this is an important reference basis for subsequent judgment of whether there are pathological conditions such as postoperative infection and abnormal recovery. For example, in the chief complaint "About two or three days after knee replacement surgery, I began to feel the incision site getting hot", the "about" and "two or three days" are such fuzzy time complaints, which are real expressions that frequently appear in clinical consultations. In actual application scenarios, such fuzzy expressions may result in different diagnostic schemes due to semantic context differences. For example, there is a sequential dependence in semantics between "having a fever about one week after surgery" and "starting to have a fever about one week after surgery". Therefore, first automatically identify the fuzzy phrases describing the time interval from the patient's chief complaint text, and determine their starting and ending positions in the text sequence.

[0017] In an embodiment of the present application, starting from the patient's chief complaint text, the fuzzy time keywords are first found. Specifically, the patient's chief complaint text can be encoded into a sequence of word vectors , and a pointer network is established to receive the text sequence as input and output the starting and ending positions of the fuzzy time phrase. The pointer network uses BERT embeddings as the encoder to extract context representations , and uses a dual-pointer method to mark the starting and ending boundaries of the word vector sequence, while implementing two independent attention mechanisms to predict the starting position and the ending position respectively. When training the pointer network, the cross-entropy loss function is used, and the goal is to minimize the error between the model prediction and the manually marked position, so as to accurately intercept the fuzzy time phrase.

[0018] In addition, when extracting the fuzzy time keywords, it is also necessary to extract the disease keywords from the patient's chief complaint, such as "knee replacement" and "fever".

[0019] S102: Based on the correspondence between the fuzzy time keywords and the disease keywords, extract a subgraph from the medical knowledge graph and use the graph attention network to perform node representation learning on the subgraph, where the input is the set of keywords in the chief complaint text, and the output is the time information corresponding to the current chief complaint symptom; In an embodiment of the present application, a correspondence between the disease keywords (diseases or symptoms) and the fuzzy time keywords (fuzzy times) is established to extract a subgraph from the locally stored medical knowledge graph. For example, for the patient's chief complaint in the aforementioned example, the extracted keywords are {"surgery": "knee replacement surgery", "symptom": "incision site getting hot"}, and a subgraph is extracted from the complete medical knowledge graph according to the extracted keywords. The formed subgraph contains information about the surgical procedures, diseases, symptoms, and recovery periods directly or indirectly related to the current chief complaint.

[0020] As an example, after extracting the subgraph, a graph attention network is established for node representation learning. The input is keywords, and the output is the time period embedding. The goal of the graph attention network is to maximize the association between relevant disease or surgical procedure nodes and their recovery cycle nodes, and a multi-task loss of node classification and edge type prediction is adopted. Using the trained graph attention network, a graph structure is established on the medical knowledge graph. The nodes include diseases, surgical procedures, complications, recovery cycles, etc., and the edges represent temporal and causal relationships. Specifically, a set of keywords is extracted from the chief complaint , such as {"knee replacement", "fever"}, and relevant nodes are located on the local medical knowledge graph , and the obtained subgraph is extracted . The graph attention network outputs the time information corresponding to the current chief complaint symptoms by aggregating relevant node information. For example, based on "knee replacement surgery", "the recovery period is 3 - 7 days" is obtained, and again based on "postoperative fever", "it may occur within 2 - 4 days" is obtained.

[0021] S103: Construct a time ambiguity-confidence joint modeling network model. Among them, the input includes the time information corresponding to the current chief complaint symptoms, the context chief complaint content, and the patient's individual information, and the output is the normal estimated time prediction result, and a joint loss function is constructed to train the time ambiguity-confidence joint modeling network model; It should be noted that in actual clinical consultations, the vague expression of the patient's chief complaint essentially combines language ambiguity, event context dependence, and individual recovery speed differences, posing challenges to traditional structured reasoning systems. Therefore, this application proposes a time ambiguity-confidence joint modeling network (TFN) to map natural language vague expressions into refined time points with distribution information, and introduces patient individual variables to improve personalized accuracy, so as to calculate refined time based on vague time.

[0022] Among them, the time ambiguity-confidence joint modeling network model consists of two core branches, both of which are simple neural network structures. The model input includes the time information corresponding to the current chief complaint symptoms (vague time phrase ), the context chief complaint content , and the patient's individual information . The three parts are encoded into vectors using BERT and concatenated to obtain the input vector . Among them, the main branch outputs the time point when the symptom actually occurs , and the sub-branch outputs the standard deviation of the speculation result , and finally outputs a normal estimate with "explainable uncertainty" , that is, the normal estimated time prediction result.

[0023] Specifically, a joint loss function is constructed in the loss function stage, comprehensively considering prediction accuracy and standard deviation rationality. The main loss uses the Smooth-L1 function , which measures the error between the model output time point and the refined time point of manual annotation, emphasizes the precision control of the center point prediction, and is suitable for dealing with actual scenarios with certain annotation errors in the chief complaint. Among them, the secondary loss constructs a logarithmic likelihood loss for normal estimation. Based on the standard deviation (σ) of the time distribution output by the model, a normal distribution probability density function is constructed, and the negative logarithmic likelihood (NLL) loss of the true refined time point under this distribution is calculated. It can dynamically learn the uncertainty corresponding to different fuzzy expressions (e.g., "nearly" is more biased towards the lower prediction limit than "almost"), so that the model has the ability of "adjustable confidence interval". The mathematical form is . In high-sensitivity scenarios such as postoperative recovery time prediction, for example, the judgment of the anti-infection window, the secondary loss term can systematically output not only the predicted time point value, but also an interpretable prediction confidence range, thereby enhancing the system's support ability for clinical decisions. Then, the loss function of the TFN model is obtained as , where and are both hyperparameters. Exemplarily,[[]] takes 0.7,[[]] takes 0.3, paying more attention to the accuracy of time point prediction while taking uncertainty into account.[[]] takes 0.3,[[]] takes 0.7, paying more attention to the uncertainty of prediction, and is suitable for scenarios with high requirements for the risk window coverage rate. The system can flexibly balance between "prediction accuracy" and "confidence control" to adapt to the differences in time output requirements in different clinical scenarios (for example, preoperative evaluation pays attention to accuracy, and postoperative monitoring pays more attention to the risk window coverage rate).

[0024] S104: Jointly calibrate the normal estimation time prediction result based on the causal path information in the medical knowledge graph to obtain the time prior distribution on the graph side; It should be noted that relying solely on the time ambiguity-confidence joint modeling network model will lead to misjudgment of time offset due to semantic ambiguity. In real electronic medical records, the specific meanings of a large number of fuzzy time phrases depend on other medical events in the chief complaint. If the model does not show the causal relationship between modeling events, it may mis-match "recently" as around the current time, ignoring "postoperative" as the starting point of reasoning, resulting in inference bias.

[0025] ​​​​On the other hand, due to the natural lack of robustness caused by the weak supervision signal, even if the TFN outputs a reasonable predicted time point, the basis for its prediction may not necessarily conform to clinical common sense or the law of path evolution. For example, in the chief complaint of "fever after postoperative infection", if the model does not explicitly capture the evolution chain of "postoperative → infection → fever", it may push the fever time too early or too late, reducing the actual credibility. Therefore, to enhance the semantic consistency and medical rationality constraint of the time ambiguity modeling results, a causal calibration module is introduced to jointly calibrate the fuzzy time point prediction results based on the causal path information in the medical knowledge graph.

[0026] Specifically, first, let the predicted time point output by the aforementioned time ambiguity-confidence joint modeling network model be , and its corresponding time probability distribution be , where represents the inferred time (in days) relative to the current chief complaint reference point, and this distribution can be a normal distribution, a skewed distribution, or a Gaussian mixture distribution. Then, based on the normal estimated time prediction result (chief complaint parsing result), the system identifies a set of candidate paths in the medical knowledge graph that are causally related to the target symptom in the current chief complaint. For each path , its corresponding reference time distribution is extracted according to the entities (such as "postoperative", "infection", "fever") and relationship edges (such as "causes", "secondary to") on the path, and its effectiveness weight is calculated. The weight can be comprehensively calculated based on the graph path length, edge weight confidence, entity similarity, etc., and satisfies the normalization condition . Accordingly, the time prior distribution on the graph side is constructed as .

[0027] S105: Construct a joint distribution based on the normal estimated time prediction result and the time prior distribution on the graph side, and use the expectation of the joint distribution as the time point prediction result; In this embodiment, after obtaining the prediction distribution of the time ambiguity-confidence joint modeling network model and the graph prior distribution, the causal calibration module uses a linear weighted fusion method to construct a joint distribution , where is an adjustable fusion parameter used to control the trade-off relationship between the model prediction and the graph prior, and can be set through validation set tuning or end-to-end learning. Finally, the system uses the expectation of the joint distribution as the final time point prediction result . represents the time probability distribution output by the time ambiguity-confidence joint modeling network (TFN) model, reflecting the prediction uncertainty of the model for the time point. Represents the time prior distribution obtained based on the causal path information of the medical knowledge graph, reflecting the constraints of the causal paths related to the current chief complaint in the knowledge graph on the time point.

[0028] S106: Input the time point prediction result into the large model to output structured suggestions, and feedback the structured suggestions to the medical knowledge graph for updating.

[0029] In an embodiment of the present application, after obtaining the time point prediction result (accurate time point), a large model (such as LLaMA or GPT, etc.) or a large model in the medical field is selected and an input containing four parts is constructed, namely the refined time point, the chief complaint symptoms, the surgical type, and the patient background (more inputs can also be added as constraints in practical applications). The large model outputs structured suggestions, such as "It is recommended to immediately conduct a postoperative infection assessment, including blood routine and C-reactive protein examinations, and initiate antibiotic treatment if necessary."

[0030] So far, the complete process from parsing the patient's chief complaint to refining the fuzzy time and then to the large model giving diagnostic suggestions has been completed. However, to improve the self-adaptability of the knowledge graph and the accuracy of time reasoning, it is necessary to perform corresponding updates on the local medical knowledge graph in the above process.

[0031] Specifically, a knowledge graph dynamic update module is added, which is used to feedback the results of each time refinement to the medical knowledge graph after binning and aggregating them according to the confidence level, promoting the continuous evolution of the graph structure.

[0032] In an embodiment of the present application, according to the time point output by the time refinement reasoning module and the standard deviation , a corresponding confidence time interval is constructed. is a parameter used to define the width of the confidence time interval, representing the time range extended forward and backward from the predicted time point . It determines the span of the confidence interval on the time axis. Exemplarily, can take 1 day. This time point and its confidence information constitute an important input for graph update, rather than using a static update method with a single time point value.

[0033] The fuzzy time expression is specifically modeled as the "time uncertainty" attribute on the nodes or edges in the graph, constructing a structured time evolution relationship, and establishing the following structured quadruple in the graph: ; where represents an event node (such as "knee joint replacement"), is a symptom node (such as "incision site fever"), is the original fuzzy time expression (such as "about two or three days"), It is the refined time point output by the TFN module (such as "the 3.5th day"). Such quadruples are stored as time semantic relation units in the knowledge graph, and the mapping edge of "fuzzy expression → refined time" is constructed. , accumulate the actual distributions of different fuzzy expressions in specific medical contexts, and use them as the dynamic sample pool for subsequent inference model training. Then, in terms of mapping edge weight update, the system adopts the following weighted smoothing mechanism: Among them, represents the updated edge weight, which is calculated by combining the old weight and the confidence of the current inference result, and is used to dynamically adjust the time semantic relationship in the knowledge graph. represents the old weight of the edge in the knowledge graph, which reflects the previous confidence in the relationship between fuzzy time expressions and refined time. is the confidence of the current inference result, which is obtained by the time fuzziness-confidence joint modeling network model through the negative log-likelihood function.

[0034] The working process of the present invention is as follows: extract the fuzzy time keywords and disease keywords in the chief complaint text; based on the corresponding relationship between the fuzzy time keywords and disease keywords, extract the subgraph from the medical knowledge graph and use the graph attention network to perform node representation learning on the subgraph, where the input is the keyword set in the chief complaint text and the output is the time information corresponding to the current chief complaint symptom; construct a time fuzziness-confidence joint modeling network model, where the input includes the time information corresponding to the current chief complaint symptom, the context chief complaint content, and the patient individual information, the output is the normal estimation time prediction result, and a joint loss function is constructed to train the time fuzziness-confidence joint modeling network model; perform joint calibration on the normal estimation time prediction result based on the causal path information in the medical knowledge graph to obtain the time prior distribution on the graph side; construct a joint distribution based on the normal estimation time prediction result and the time prior distribution on the graph side, and use the expectation of the joint distribution as the time point prediction result; input the time point prediction result into the large model to output the structured suggestion, and feedback the structured suggestion to the medical knowledge graph for update.

[0035] In summary, the embodiments of the present invention provide a large model optimization method integrating a medical knowledge graph. By establishing a large model optimization system that can accurately understand fuzzy time expressions and perform reasoning in combination with medical graph knowledge, it is the key path to solve the problem of reasoning errors caused by the inability to recognize semantic dependencies or ignoring patient background factors.

[0036] On the other hand, the present application also provides a large model optimization system integrating a medical knowledge graph, including: Extraction module: used to extract the fuzzy time keywords and disease keywords in the chief complaint text; Learning module: It is used to extract a subgraph from a medical knowledge graph based on the correspondence between fuzzy time keywords and disease keywords, and use a graph attention network to perform node representation learning on the subgraph. Among them, the input is the keyword set in the chief complaint text, and the output is the time information corresponding to the current chief complaint symptom; Prediction training module: It is used to construct a time ambiguity-confidence joint modeling network model. Among them, the input includes the time information corresponding to the current chief complaint symptom, the context chief complaint content, and the patient's individual information. The output is the normal estimation time prediction result, and a joint loss function is constructed to train the time ambiguity-confidence joint modeling network model; Joint verification module: It is used to jointly calibrate the normal estimation time prediction result based on the causal path information in the medical knowledge graph to obtain the time prior distribution on the graph side; Prediction result module: Based on the normal estimation time prediction result and the time prior distribution on the graph side, a joint distribution is constructed, and the expectation of the joint distribution is used as the time point prediction result; Result update module: The time point prediction result is input into a large model to output a structured suggestion, and the structured suggestion is fed back to the medical knowledge graph for update.

[0037] The above is only the preferred implementation mode of the present invention. It should be pointed out that for those of ordinary skill in the art of this technology, without departing from the counting principle of the present invention, several improvements and replacements can still be made, and these improvements and replacements should also be regarded as within the protection scope of the present invention.

Claims

1. A large model optimization method integrating a medical knowledge graph, characterized in that, It includes the following steps: Extract the fuzzy time keywords and disease keywords in the chief complaint text; Based on the corresponding relationship between the fuzzy time keywords and disease keywords, extract a subgraph from the medical knowledge graph and use a graph attention network to perform node representation learning on the subgraph. Among them, the input is the keyword set in the chief complaint text, and the output is the time information corresponding to the current chief complaint symptom; Construct a time ambiguity-confidence joint modeling network model. Among them, the input includes the time information corresponding to the current chief complaint symptom, the context chief complaint content, and the patient's individual information. The output is the normal estimation time prediction result, and a joint loss function is constructed to train the time ambiguity-confidence joint modeling network model; Based on the causal path information in the medical knowledge graph, jointly calibrate the normal estimation time prediction result to obtain the time prior distribution on the graph side; Construct a joint distribution based on the normal estimation time prediction result and the time prior distribution on the graph side, and use the expectation of the joint distribution as the time point prediction result; Input the time point prediction result into a large model to output structured suggestions, and feedback the structured suggestions to the medical knowledge graph for updating.

2. The large model optimization method integrating a medical knowledge graph according to claim 1, characterized in that, The extraction of fuzzy time keywords in the chief complaint text includes: Encode the chief complaint text into a sequence of word vectors; Build a pointer network, use the text sequence of the word vector sequence as the input, and output the start and end positions of the fuzzy time phrase as the fuzzy time keywords.

3. The optimization method of the large model integrating the medical knowledge graph according to claim 1, wherein, The use of a graph attention network to perform node representation learning on the subgraph includes: The goal of the graph attention network is to maximize the association between relevant disease or surgical procedure nodes and their recovery cycle nodes, and adopt a multi-task loss of node classification and edge type prediction.

4. A large model optimization method integrating a medical knowledge graph according to claim 1, characterized in that The joint calibration of the normal estimation prediction result based on the causal path information in the medical knowledge graph to obtain the time prior distribution on the graph side includes: Based on the normal estimation time prediction result, identify a set of candidate paths in the medical knowledge graph that have a causal association with the target symptom in the current chief complaint. For each entity and relationship edge on each path in the candidate path set, extract the corresponding reference time distribution; Calculate the validity weight corresponding to the reference time distribution, and construct the time prior distribution on the graph side based on the validity weight.

5. The optimization method of a large model integrating a medical knowledge graph according to claim 1, characterized in that, The input of the large model includes: the time point prediction result, the chief complaint symptom, the surgical type, and the patient background.

6. The optimization method of the large model integrating a medical knowledge graph according to claim 1, characterized in that The feedback of the structured suggestions to the medical knowledge graph for updating includes: Aggregate each structured suggestion by confidence bucket and feedback it to the medical knowledge graph to promote the continuous update and evolution of the graph structure.

7. A large model optimization method integrating a medical knowledge graph according to claim 6, characterized in that The aggregation of each structured suggestion by confidence bucket and feedback to the medical knowledge graph to promote the continuous update and evolution of the graph structure includes: Based on the time point and standard deviation in the structured suggestion, construct a corresponding confidence time interval; Model the fuzzy time expression in the confidence time interval as an attribute on the node or edge in the graph structure, and construct a structured time evolution relationship; Perform continuous update and evolution according to the structured time evolution relationship.

8. A large model optimization system integrating a medical knowledge graph, characterized in that It includes: Extraction module: used to extract fuzzy time keywords and disease keywords from the chief complaint text; Learning module: used to extract a subgraph from the medical knowledge graph based on the corresponding relationship between the fuzzy time keywords and the disease keywords, and use a graph attention network to perform node representation learning on the subgraph, where the input is the keyword set in the chief complaint text, and the output is the time information corresponding to the current chief complaint symptom; Prediction training module: used to construct a time ambiguity-confidence joint modeling network model, where the input includes the time information corresponding to the current chief complaint symptom, the context chief complaint content, and the patient individual information, the output is the normal estimated time prediction result, and a joint loss function is constructed to train the time ambiguity-confidence joint modeling network model; Joint verification module: used to perform joint calibration on the normal estimated time prediction result based on the causal path information in the medical knowledge graph to obtain the time prior distribution on the graph side; Prediction result module: constructs a joint distribution based on the normal estimated time prediction result and the time prior distribution on the graph side, and takes the expectation of the joint distribution as the time point prediction result; Result update module: inputs the time point prediction result into a large model to output a structured suggestion, and feeds back the structured suggestion to the medical knowledge graph for updating.

Images