A large model optimization method and system integrating medical knowledge graph

Abstract

The present invention relates to the fields of natural language processing and medical artificial intelligence technology, and discloses a large-scale model optimization method and system that integrates a medical knowledge graph. The method includes: constructing a temporal fuzziness-confidence joint modeling network model, wherein the input includes the time information corresponding to the current chief complaint, the context of the chief complaint, and individual patient information; the output is a normal estimated time prediction result; the normal estimated time prediction result is jointly calibrated based on the causal path information in the medical knowledge graph to obtain the time prior distribution on the graph side; a joint distribution is constructed based on the normal estimated time prediction result and the graph-side time prior distribution, and the expectation of the joint distribution is used as the time point prediction result. The time point prediction result is input into the large-scale model to output structured suggestions, which are then fed back into the medical knowledge graph for updating. This solves the problem of reasoning errors caused by failing to identify semantic dependencies or ignoring patient background factors.

Description

Technical Field

[0001] The present invention relates to the fields of natural language processing and medical artificial intelligence technology, and in particular to a large-scale model optimization method and system integrating medical knowledge graphs. Background Art

[0002] During clinical diagnosis and treatment, patient complaints are the primary source of information about their condition. However, these complaints often contain a significant amount of ambiguous and uncertain information, particularly when it comes to time. Time expressions such as "about two or three days" or "about a week after surgery" are extremely common in natural language. While these descriptions are understandable to human doctors, they lack operability and standardization for large-scale model systems that rely on structured information for reasoning and recommendations.

[0003] Currently, mainstream methods rely on template matching or rule-based temporal normalization strategies. While these strategies can achieve some success with standardized sentence structures, they often fail to identify semantic dependencies or ignore patient background factors in real-world complaints involving casual language and complex contexts, leading to inference errors. For example, the "fever" in the phrase "fever during the postoperative recovery period, approximately one week" could be related to a postoperative infection or a manifestation of other complications. The timing of this occurrence is ambiguous but crucial. This requires the system to not only understand the language but also to make a comprehensive judgment based on medical knowledge, temporal relationships, and patient background. Summary of the Invention

[0004] The purpose of this invention is to establish a large-scale model optimization system that can accurately understand fuzzy time expressions and combine medical atlas knowledge for reasoning, as a key path to solving the above-mentioned problems.

[0005] To achieve the above objectives, the present invention provides a large model optimization method integrating medical knowledge graph, comprising:

[0006] Extract fuzzy time keywords and disease keywords from the chief complaint text;

[0007] Based on the correspondence between the fuzzy time keywords and the symptom keywords, a subgraph is extracted from the medical knowledge graph and a graph attention network is used to learn node representations of 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 symptoms;

[0008] Constructing a temporal ambiguity-confidence joint modeling network model, wherein the input includes the time information corresponding to the current main complaint symptom, the context of the main complaint content, and the individual patient information, and the output is a normal estimation time prediction result, and constructing a joint loss function to train the temporal ambiguity-confidence joint modeling network model;

[0009] The normal estimation time prediction result is jointly calibrated based on the causal path information in the medical knowledge graph to obtain the time prior distribution on the graph side;

[0010] Constructing a joint distribution based on the normal estimation time prediction result and the time prior distribution on the atlas side, and taking the expectation of the joint distribution as the time point prediction result;

[0011] The time point prediction results are input into the big model to output structured suggestions, and the structured suggestions are fed back into the medical knowledge graph for updating.

[0012] Optionally, extracting fuzzy time keywords from the main complaint text includes:

[0013] Encode the main complaint text into a word vector sequence;

[0014] A pointer network is established, the word vector sequence text sequence is taken as input, and the starting and ending positions of the fuzzy time phrases are output as the fuzzy time keywords.

[0015] Optionally, the performing node representation learning on the subgraph using a graph attention network includes:

[0016] The goal of the graph attention network is to maximize the association between related disease or procedure nodes and their recovery period nodes, using a multi-task loss of node classification and edge type prediction.

[0017] Optionally, 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:

[0018] Based on the normal estimation time prediction result, a set of candidate paths causally associated with the target symptom in the current chief complaint is identified in the medical knowledge graph, and corresponding reference time distributions are extracted for entities and relationship edges on each path in the candidate path set;

[0019] 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.

[0020] Optionally, the large model input includes: the predicted results at the time point, main symptoms, surgery type and patient background.

[0021] Optionally, feeding back the structured suggestions to the medical knowledge graph for updating includes:

[0022] Each structured suggestion is aggregated into buckets according to confidence and then fed back into the medical knowledge graph to promote the continuous update and evolution of the graph structure.

[0023] Optionally, the step of aggregating each structured suggestion into buckets according to confidence and feeding it back into the medical knowledge graph to promote the continuous update and evolution of the graph structure includes:

[0024] Constructing corresponding confidence intervals based on the time points and standard deviations in the structured suggestions;

[0025] Modeling the fuzzy time expression in the confidence time interval as an attribute of a node or an edge in the graph structure to construct a structured time evolution relationship;

[0026] Continuously update and evolve according to the structured time evolution relationship.

[0027] On the other hand, this application also discloses a large model optimization system integrating medical knowledge graph, including:

[0028] Extraction module: used to extract fuzzy time keywords and symptom keywords from the chief complaint text;

[0029] A learning module is configured to extract a subgraph from the medical knowledge graph based on the correspondence between the fuzzy time keywords and the symptom keywords, and perform node representation learning on the subgraph using a graph attention network, wherein the input is a set of keywords in the chief complaint text, and the output is the time information corresponding to the current chief complaint symptoms;

[0030] Prediction training module: used to construct a temporal ambiguity-confidence joint modeling network model, wherein the input includes the time information corresponding to the current main complaint symptom, the context of the main complaint content and the individual patient information, the output is the normal estimation time prediction result, and a joint loss function is constructed to train the temporal ambiguity-confidence joint modeling network model;

[0031] Joint verification module: used to jointly calibrate the normal estimation time prediction results based on the causal path information in the medical knowledge graph to obtain the time prior distribution on the graph side;

[0032] Prediction result module: constructs a joint distribution based on the normal estimation time prediction result and the time prior distribution on the atlas side, and uses the expectation of the joint distribution as the time point prediction result;

[0033] Result update module: inputs the time point prediction results into the big model to output structured suggestions, and feeds the structured suggestions back to the medical knowledge graph for updating.

[0034] Compared with the existing technology, the embodiment of the present invention, a large model optimization method and system that integrates medical knowledge graphs, has the following beneficial effects: the large model optimization system can accurately understand fuzzy time expressions and combine medical graph knowledge for reasoning, which can serve as a key path to solving the difficult problem of reasoning errors caused by the inability to identify semantic dependencies or ignoring patient background factors. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flow chart of a large model optimization method integrating medical knowledge graph according to an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0037] like Figure 1 As shown, a large model optimization method integrating medical knowledge graphs according to a preferred embodiment of the present invention includes the following steps:

[0038] S101: extracting fuzzy time keywords and symptom keywords from the chief complaint text;

[0039] It should be noted that in the actual clinical diagnosis process, patients are often unable to provide accurate time due to vague memories or unclear symptoms, but this is an important reference for subsequent judgments on whether there are postoperative infections, abnormal recovery and other pathologies. For example, the "approximately" and "two or three days" in the chief complaint "I started to feel the incision getting hot about two or three days after knee replacement surgery" are real expressions that appear frequently in clinical consultations. In actual application scenarios, such ambiguous expressions may lead to different diagnostic solutions due to differences in semantic context. For example, "fever occurred about a week after surgery" and "fever started about a week after surgery" have a semantic dependency. Therefore, we first automatically identify the ambiguous phrases describing the time interval from the patient's chief complaint text and determine their starting and ending positions in the text sequence.

[0040] In one embodiment of the present application, we first start from the patient's main complaint text to find fuzzy time keywords. Specifically, the patient's main complaint text can be encoded into a word vector sequence , establish a pointer network to receive text sequences as input and output the start and end positions of fuzzy time phrases. The pointer network uses BERT embedding as the encoder to extract context representation , uses a dual-pointer approach to mark the start and end boundaries of a word vector sequence, while implementing two independent attention mechanisms to predict the start and end positions, respectively. The pointer network is trained using a cross-entropy loss function, aiming to minimize the error between the model's predictions and the manually annotated positions, thereby accurately capturing ambiguous temporal phrases.

[0041] In addition, extracting fuzzy time keywords also requires extracting disease keywords from the patient's chief complaint, such as "knee replacement" and "fever".

[0042] S102: Based on the correspondence between fuzzy time keywords and symptom keywords, a subgraph is extracted from the medical knowledge graph and a graph attention network is used to learn node representations for the subgraph. The input is a set of keywords in the chief complaint text, and the output is the time information corresponding to the current chief complaint symptoms.

[0043] In one embodiment of the present application, a correspondence is established between disease keywords (diseases or symptoms) and fuzzy time keywords (fuzzy time) to extract subgraphs 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","symptom":"incision fever"}. Based on the extracted keywords, a subgraph is extracted from the complete medical knowledge graph. The resulting subgraph contains information about the surgical procedure, disease, symptoms, and recovery period that are directly or indirectly related to the current chief complaint.

[0044] As an example, after extracting the subgraph, a graph attention network is established to learn node representations, with the input being keywords and the output being time segment embeddings. The goal of the graph attention network is to maximize the association between the relevant disease or procedure nodes and their recovery period nodes, using a multi-task loss of node classification and edge type prediction. Using the trained graph attention network, a graph structure is established on the medical knowledge graph. The nodes include diseases, procedures, complications, recovery periods, etc., and the edges represent time sequence and causal relationships. Specifically, a set of keywords is extracted from the main complaint. , such as {"knee replacement", "fever"}, in the local medical knowledge graph Locate the relevant nodes and extract the subgraph The graph attention network aggregates relevant node information to output the time information corresponding to the current main symptom. For example, based on "knee replacement surgery", it can obtain "recovery period is 3-7 days", and for "postoperative fever", it can obtain "may occur within 2-4 days".

[0045] S103: Constructing a temporal ambiguity-confidence joint modeling network model, wherein the input includes the time information corresponding to the current chief complaint, the context of the chief complaint, and the individual patient information, and the output is a normal estimation time prediction result, and constructing a joint loss function to train the temporal ambiguity-confidence joint modeling network model;

[0046] It should be noted that in actual clinical consultations, the fuzzy expressions of patients' chief complaints inherently combine linguistic ambiguity, event context dependency, and individual differences in recovery speed, posing a challenge to traditional structured reasoning systems. To this end, this application proposes a temporal fuzziness-confidence joint modeling network (TFN), which maps fuzzy natural language expressions into refined time points with distribution information. It also introduces individual patient variables to improve personalized accuracy, allowing refined time to be calculated based on fuzzy time.

[0047] The temporal fuzziness-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 main complaint (fuzzy time phrase ), context of the main complaint , patient individual information Three parts, the same is encoded into a vector using BERT and concatenated to obtain the input vector The main branch outputs the time when the symptoms actually occurred. , the sub-branch outputs the standard deviation of the inference result , the final output is a normal estimate with "explainable uncertainty" That is, the normal estimation time prediction result.

[0048] Specifically, a joint loss function is constructed in the loss function stage, and the prediction accuracy and standard deviation rationality are comprehensively considered. The main loss adopts the Smooth-L1 function. , measures the error between the model output time point and the manually annotated refined time point, emphasizes the accuracy control of the center point prediction, and is suitable for actual scenarios where there are certain annotation errors in the main complaint. Among them, the secondary loss constructs the log-likelihood loss for the normal estimation, constructs the normal distribution probability density function based on the standard deviation (σ) of the time distribution of the model output, and calculates the negative log-likelihood (NLL) loss of the true refined time point under this distribution. It can dynamically learn the uncertainty corresponding to different fuzzy expressions (such as "nearly" is more biased towards the lower limit of the prediction than "almost"), so that the model has the ability of "adjustable confidence interval", and the mathematical form is In highly sensitive scenarios such as postoperative recovery time prediction, such as the determination of the anti-infection window, the secondary loss term can enable the system to output not only the predicted time point value, but also an interpretable prediction confidence range, thereby enhancing the system's support for clinical decision-making. The loss function of the TFN model is then obtained as ,in and All are hyperparameters, for example, Take 0.7, Taking 0.3, more emphasis is placed on the accuracy of time point predictions while taking uncertainty into account. Take 0.3, A value of 0.7 places greater emphasis on prediction uncertainty and is suitable for scenarios requiring high risk window coverage. The system flexibly balances prediction accuracy with confidence control, adapting to varying time output requirements in different clinical scenarios (for example, preoperative assessment prioritizes accuracy, while postoperative monitoring focuses on risk window coverage).

[0049] S104: Jointly calibrate the normal estimation time prediction results based on the causal path information in the medical knowledge graph to obtain the time prior distribution on the graph side;

[0050] It's important to note that relying solely on a network model that combines temporal ambiguity and confidence can lead to misjudgments of time offsets due to semantic ambiguity. In real electronic medical records, the specific meanings of many ambiguous time phrases depend on other medical events in the chief complaint. If the model doesn't demonstrate causal relationships between modeled events, it might mismatch "recent" events with those around the current time, ignoring the "postoperative" starting point for reasoning, leading to biased inferences.

[0051] On the other hand, due to the inherent lack of robustness caused by weak supervisory signals, even if the TFN outputs a reasonable predicted time point, its prediction basis may not necessarily conform to clinical common sense or the laws of path evolution. For example, in the case of a chief complaint of "fever after postoperative infection," if the model does not explicitly capture the evolutionary chain of "postoperative → infection → fever," the fever time may be pushed too early or too late, reducing actual credibility. Therefore, to achieve enhanced semantic consistency and medical rationality constraints for the results of temporal fuzziness modeling, 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.

[0052] Specifically, first, let the predicted time point output by the aforementioned temporal fuzziness-confidence joint modeling network model be , and its corresponding time probability distribution is ,in Represents the inferred time relative to the current chief complaint reference point (in days). The distribution can be normal, skewed, or Gaussian mixture. Then, based on the normal estimation time prediction results (chief complaint analysis results), the system identifies a set of candidate paths in the medical knowledge graph that are causally related to the target symptoms in the current chief complaint. For each path , extract the corresponding reference time distribution based on the entities on the path (such as "postoperative", "infection", "fever") and the relationship edges (such as "leading to", "secondary to") , and calculate its effectiveness weight The weight can be calculated based on the graph path length, edge weight confidence, entity similarity, etc., to meet the normalization conditions Based on this, the time prior distribution of the constructed graph side is .

[0053] S105: 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;

[0054] In this embodiment, after obtaining the predicted distribution of the temporal fuzziness-confidence joint modeling network model and the graph prior distribution, the causal calibration module uses a linear weighted fusion method to construct the joint distribution ,in, It is an adjustable fusion parameter used to control the trade-off between model prediction and graph priors, which can be set by validation set tuning or end-to-end learning. The expectation of the final time point prediction result . It represents the temporal probability distribution output by the Temporal Fuzziness-Confidence Joint Modeling Network (TFN) model, reflecting the uncertainty of the model's prediction of time points. It represents the time prior distribution obtained based on the causal path information of the medical knowledge graph, reflecting the constraints of the causal path related to the current chief complaint on the time point in the knowledge graph.

[0055] S106: Input the time point prediction results into the big model to output structured suggestions, and feed the structured suggestions back to the medical knowledge graph for updating.

[0056] In one embodiment of the present application, after obtaining a time point prediction result (precise time point), a large model (such as LLaMA or GPT) or a large medical model is selected and constructed with four input components: the refined time point, the presenting symptom, the type of surgery, and the patient's background (more inputs can be added as constraints in actual applications). The large model outputs structured recommendations, such as "It is recommended to immediately conduct a postoperative infection evaluation, including blood routine and C-reactive protein testing, and initiate antibiotic treatment if necessary."

[0057] At this point, the entire process from analyzing patient complaints to refining fuzzy temporal information, and finally to the large-scale model providing diagnostic recommendations, has been completed. However, to improve the adaptability and temporal reasoning accuracy of the knowledge graph, the local medical knowledge graph needs to be updated accordingly during this process.

[0058] Specifically, a knowledge graph dynamic update module is added to aggregate the results of each time refinement into buckets according to confidence and feed them back to the medical knowledge graph, thereby promoting the continuous evolution of the graph structure.

[0059] In one embodiment of the present application, the time point output by the time refinement reasoning module is and standard deviation , and then construct the corresponding confidence interval . It is a parameter used to define the width of the confidence interval, which means the confidence interval from the predicted time point The time range to which the confidence interval extends forward and backward. It determines the span of the confidence interval on the time axis, for example, The time point and its confidence information constitute important inputs for graph updates, rather than static updates using a single time point value.

[0060] Expressing fuzzy time As the "time uncertainty" attribute of nodes or edges in the graph, it is specially modeled to build a structured time evolution relationship and establish the following structured quadruple in the graph: ;in, represents an event node (e.g., "knee replacement"), Symptom nodes (such as "wound fever"), For primitive fuzzy time expressions (such as "about two or three days"), The refined time point output by the TFN module (such as "day 3.5"). This type of quadruple is stored as a temporal semantic relationship unit in the knowledge graph, and a mapping edge of "fuzzy expression → refined time" is constructed. , the actual distribution of different fuzzy expressions in specific medical contexts is accumulated and used as a dynamic sample pool for subsequent inference model training. Then, in terms of updating the mapping edge weights, the system adopts the following weighted smoothing mechanism:

[0061]

[0062] in, Represents the updated edge weight, which is calculated by combining the old weight and the confidence of the current reasoning result, and is used to dynamically adjust the temporal semantic relationship in the knowledge graph. Represents the old weight of the edge in the knowledge graph, reflecting the previous confidence in the relationship between the fuzzy time expression and the refined time. is the confidence of the inference result, which is obtained by the negative log-likelihood function calculated by the temporal ambiguity-confidence joint modeling network model.

[0063] The working process of the present invention is as follows: extracting fuzzy time keywords and symptom keywords from the chief complaint text; based on the correspondence between fuzzy time keywords and symptom keywords, extracting a subgraph from the medical knowledge graph and using a graph attention network to perform node representation learning on the subgraph, wherein the input is a set of keywords in the chief complaint text, and the output is the time information corresponding to the current chief complaint symptoms; constructing a time fuzziness-confidence joint modeling network model, wherein the input includes the time information corresponding to the current chief complaint symptoms, the contextual chief complaint content and the patient's individual information, and the output is a normal estimation time prediction result, and constructing a joint loss function to train the time fuzziness-confidence joint modeling network model; based on the causal path information in the medical knowledge graph, the normal estimation time prediction result is jointly calibrated to obtain the time prior distribution on the graph side; constructing a joint distribution based on the normal estimation time prediction result and the time prior distribution on the graph side, and taking the expectation of the joint distribution as the time point prediction result; inputting the time point prediction result into the large model to output structured suggestions, and feeding the structured suggestions back to the medical knowledge graph for updating.

[0064] In summary, the present invention provides a large-scale model optimization method that integrates medical knowledge graphs. By establishing a large-scale model optimization system that accurately understands fuzzy temporal expressions and integrates medical knowledge graphs for reasoning, this system serves as a key path to addressing the problem of reasoning errors caused by failing to identify semantic dependencies or ignoring patient context.

[0065] On the other hand, this application also provides a large model optimization system integrating medical knowledge graph, including:

[0066] Extraction module: used to extract fuzzy time keywords and symptom keywords from the chief complaint text;

[0067] Learning module: Based on the correspondence between fuzzy time keywords and symptom keywords, it extracts subgraphs from the medical knowledge graph and uses the graph attention network to learn node representations for the subgraphs. The input is a set of keywords in the chief complaint text, and the output is the time information corresponding to the current chief complaint symptoms.

[0068] Prediction training module: used to build a temporal ambiguity-confidence joint modeling network model. The input includes the time information corresponding to the current chief complaint, the context of the chief complaint, and the individual patient information. The output is the normal estimation time prediction result. A joint loss function is constructed to train the temporal ambiguity-confidence joint modeling network model.

[0069] Joint verification module: used to jointly calibrate the normal estimation time prediction results based on the causal path information in the medical knowledge graph to obtain the time prior distribution on the graph side;

[0070] Prediction result module: Constructs a joint distribution based on the normal estimation time prediction result and the time prior distribution on the atlas side, and uses the expectation of the joint distribution as the time point prediction result;

[0071] Result update module: The time point prediction results are input into the large model to output structured suggestions, and the structured suggestions are fed back to the medical knowledge graph for updating.

[0072] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary counting personnel in this technical field, several improvements and substitutions can be made without departing from the counting principle of the present invention. These improvements and substitutions should also be considered within the scope of protection of the present invention.

Claims

1. A large model optimization method integrating medical knowledge graph, characterized in that: The following steps are involved: Extract fuzzy time keywords and disease keywords from the chief complaint text; Based on the correspondence between the fuzzy time keywords and the symptom keywords, a subgraph is extracted from the medical knowledge graph and a graph attention network is used to learn node representations of 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 symptoms; A time fuzziness-confidence joint modeling network model is constructed, wherein the input includes the time information corresponding to the current main complaint symptom, the context main complaint content and the patient's individual information, and 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. The time fuzziness-confidence joint modeling network model consists of two core branches, both of which are simple neural network structures, wherein the main branch outputs the time point when the symptom actually occurs, and the secondary branch outputs the standard deviation of the inferred result, and outputs the normal estimation time prediction result. A joint loss function is constructed in the loss function stage. The main loss uses the Smooth-L1 function to measure the error between the model output time point and the manually annotated refined time point, emphasizing the accuracy control of the center point prediction. The secondary loss constructs the log-likelihood loss for the normal estimation. Based on the standard deviation of the time distribution of the model output, the normal distribution probability density function is constructed, and the negative log-likelihood loss of the true refined time point under this distribution is calculated to dynamically learn the uncertainty corresponding to different fuzzy expressions. In the highly sensitive scenario of postoperative recovery time prediction, the secondary loss term can systematically output not only the time point prediction value, but also an interpretable prediction confidence range. The normal estimation time prediction results are jointly calibrated based on the causal path information in the medical knowledge graph to obtain the time prior distribution on the graph side, including: based on the normal estimation time prediction results, identifying a set of candidate paths that are causally associated with the target symptom in the current chief complaint in the medical knowledge graph, and extracting corresponding reference time distributions for entities and relationship edges on each path in the candidate path set; calculating the validity weight corresponding to the reference time distribution, and constructing the time prior distribution on the graph side based on the validity weight; Constructing a joint distribution based on the normal estimation time prediction result and the time prior distribution on the atlas side, and taking the expectation of the joint distribution as the time point prediction result; The time point prediction results are input into the big model to output structured suggestions, and the structured suggestions are fed back into the medical knowledge graph for updating.

2. A large model optimization method integrating medical knowledge graph according to claim 1, characterized in that: The extraction of fuzzy time keywords in the main complaint text includes: Encode the main complaint text into a word vector sequence; A pointer network is established, the word vector sequence text sequence is taken as input, and the starting and ending positions of the fuzzy time phrases are output as the fuzzy time keywords.

3. A large model optimization method integrating medical knowledge graph according to claim 1, characterized in that: The performing node representation learning on the subgraph using a graph attention network includes: The goal of the graph attention network is to maximize the association between related disease or procedure nodes and their recovery period nodes, using a multi-task loss of node classification and edge type prediction.

4. A large model optimization method integrating medical knowledge graph according to claim 1, characterized in that: The large model input includes: the predicted results at the time point, the main symptoms, the type of surgery and the patient background.

5. The large model optimization method integrating medical knowledge graph according to claim 1 is characterized in that: Feeding back the structured suggestions to the medical knowledge graph for updating includes: Each structured suggestion is aggregated into buckets according to confidence and then fed back into the medical knowledge graph to promote the continuous update and evolution of the graph structure.

6. A large model optimization method integrating medical knowledge graph according to claim 5, characterized in that: The structured suggestions are aggregated into buckets based on confidence and fed back into the medical knowledge graph to promote the continuous update and evolution of the graph structure. This includes: Constructing corresponding confidence intervals based on the time points and standard deviations in the structured suggestions; Modeling the fuzzy time expression in the confidence time interval as an attribute of a node or an edge in the graph structure to construct a structured time evolution relationship; Continuously update and evolve according to the structured time evolution relationship.

7. A large model optimization system integrating medical knowledge graph, characterized by: include: Extraction module: used to extract fuzzy time keywords and symptom keywords from the chief complaint text; A learning module is configured to extract a subgraph from the medical knowledge graph based on the correspondence between the fuzzy time keywords and the symptom keywords, and perform node representation learning on the subgraph using a graph attention network, wherein the input is a set of keywords in the chief complaint text, and the output is the time information corresponding to the current chief complaint symptoms; Prediction training module: used to construct a time fuzziness-confidence joint modeling network model, wherein the input includes the time information corresponding to the current main complaint symptom, the context main complaint content and the patient's individual information, and 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, wherein the time fuzziness-confidence joint modeling network model consists of two core branches, both of which are simple neural network structures, wherein the main branch outputs the time point when the symptom actually occurs, and the secondary branch outputs the standard deviation of the inferred result, and outputs the normal estimation time prediction result. In the loss function stage A joint loss function is constructed. The main loss uses the Smooth-L1 function to measure the error between the model output time point and the manually annotated refined time point, emphasizing the accuracy control of the center point prediction. Among them, the secondary loss constructs the log-likelihood loss for the normal estimation. Based on the standard deviation of the time distribution of the model output, the normal distribution probability density function is constructed, and the negative log-likelihood loss of the true refined time point under this distribution is calculated. This is used to dynamically learn the uncertainty corresponding to different fuzzy expressions. In the highly sensitive scenario of postoperative recovery time prediction, the secondary loss term can systematically output not only the time point prediction value, but also an interpretable prediction confidence range. Joint verification module: used to jointly calibrate the normal estimation time prediction results based on the causal path information in the medical knowledge graph to obtain the time prior distribution on the graph side, including: based on the normal estimation time prediction results, identifying a set of candidate paths in the medical knowledge graph that are causally associated with the target symptom in the current chief complaint, extracting the corresponding reference time distribution for the entities and relationship edges on each path in the candidate path set; calculating the validity weight corresponding to the reference time distribution, and constructing the time prior distribution on the graph side based on the validity weight; Prediction result module: constructs a joint distribution based on the normal estimation time prediction result and the time prior distribution on the atlas side, and uses the expectation of the joint distribution as the time point prediction result; Result update module: inputs the time point prediction results into the big model to output structured suggestions, and feeds the structured suggestions back to the medical knowledge graph for updating.

Images