CN-121981151-A - Medical intelligent agent generation method, response method and device

CN121981151ACN 121981151 ACN121981151 ACN 121981151ACN-121981151-A

Abstract

A medical intelligent agent generating method, a response method and a device. The medical agent includes a plurality of engines and a large model. The rule matching engine is used for matching the user problems with the rule chain, and medical conclusion and clinical guideline are obtained according to the matching result. The rule chain comprises a plurality of interrelated medical rules, the medical rules comprise condition conditions and corresponding medical conclusions, and the rule chain is obtained by performing conversion operation on structural medical knowledge of clinical guidelines. The document matching engine is used for matching the semantic similarity between the user problem and the entity relation extracted from the medical document, and determining the medical conclusion and the medical document set according to the matching result. The entity relationship is the correspondence between a disorder and a medical conclusion. The first large model is used for reasoning and obtaining corresponding medical conclusions according to symptoms contained in the user problems. The second largest model is used to determine final medical conclusions and basis from the output of the various engines and models. The processing also performs privacy protection on the related privacy data.

Inventors

LUO WEI
NING ZHAOYANG
DU HANMIN
LIU XISHAN
XIE WEIZHI
WANG KE

Assignees

浙江扁鹊健康科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260408

Claims (12)

1. The method for generating the medical intelligent agent comprises a rule matching engine, a document matching engine, a first large model and a second large model, wherein the second large model is used for determining a final medical conclusion and basis according to the output of each engine and model, and the method comprises the following steps: Performing conversion operation on structural medical knowledge corresponding to guideline texts in clinical guidelines to obtain a corresponding rule chain, and constructing a rule matching engine based on the rule chain to enable the rule matching engine to match a first user problem with the rule chain and obtain a corresponding first medical conclusion and a corresponding clinical guideline according to a matching result, wherein any one rule chain comprises a plurality of interrelated medical rules, and any one medical rule comprises disease conditions and corresponding medical conclusions; Respectively extracting a plurality of corresponding entity relations from a plurality of medical documents, and constructing a document matching engine based on the entity relations so that the document matching engine performs semantic similarity matching on a first user problem and the entity relations of the plurality of medical documents, and determining a second medical conclusion and a corresponding medical document set according to a matching result, wherein any entity relation is a corresponding relation between a disease and the medical conclusion; And carrying out fine adjustment on the first large model by utilizing historical diagnosis and treatment data containing symptoms and medical conclusions, so that the first large model after fine adjustment obtains a corresponding third medical conclusion according to the symptom reasoning contained in the first user problem.
2. The method of claim 1, the step of performing a conversion operation for structured medical knowledge in a clinical guideline corresponding to guideline origin, comprising: Directly converting the structured medical knowledge into a corresponding initial rule chain; converting non-uniform terms in the initial rule chain into uniform terms according to a knowledge graph to obtain term alignment rule chains, wherein the knowledge graph takes terms as nodes and comprises association relations between a plurality of non-uniform terms and the uniform terms; The rule chain is determined based on the term alignment rule chain.
3. The method of claim 2, wherein the step of determining the rule chain based on the term alignment rule chain comprises determining labeling information of any term alignment rule chain, wherein the labeling information comprises timeliness information, specialty information and authority information, and adding the labeling information into the term alignment rule chain to obtain the rule chain.
4. The method of claim 3, wherein the step of obtaining the corresponding first medical conclusion based on the matching result comprises selecting the first medical conclusion from a plurality of different medical conclusions based on labeling information of a rule chain in which the different medical conclusions are located when the matching result shows that the plurality of different medical conclusions are obtained.
5. The method of claim 1, the step of determining a second medical conclusion from the matching result comprising: When the matching result shows that a plurality of different medical conclusions are obtained, aiming at any medical conclusion, executing calculation operation based on journal influence factors and related similarity of medical documents corresponding to the medical conclusion to obtain scores of the medical conclusions; And taking the medical conclusion corresponding to the highest score as the second medical conclusion.
6. The method of claim 1, the step of fine-tuning the first large model using historical diagnostic data including conditions and medical conclusions, comprising: Inputting symptoms in the historical diagnosis and treatment data into the first large model, and determining corresponding reasoning medical conclusions according to the symptoms through the first large model; Determining a predictive loss based on differences between the inferred medical conclusions and medical conclusions in the historical diagnostic data; and fine tuning the first large model according to the prediction loss.
7. The method of claim 1, further comprising performing the transformation operation for structured medical knowledge in the newly added clinical guideline that corresponds to guideline origin when a newly added clinical guideline exists, resulting in a corresponding newly added rule chain, updating the rule matching engine based on the newly added rule chain; When a new medical document exists, extracting a plurality of corresponding new entity relations from the new medical document, and updating the document matching engine based on the new entity relations; When the newly-increased historical diagnosis and treatment data exist, the newly-increased historical diagnosis and treatment data are utilized to continuously finely tune the first large model.
8. A user problem response method based on a medical agent comprises a rule matching engine, a document matching engine, a first large model and a second large model, wherein the method comprises the following steps: Matching a second user problem to be responded with a rule chain through the rule matching engine, and obtaining a corresponding first medical conclusion and a corresponding clinical guideline according to a matching result, wherein any one rule chain comprises a plurality of medical rules which are associated with each other, any one medical rule comprises a disease condition and a corresponding medical conclusion, and the rule chain is obtained by executing conversion operation on structural medical knowledge corresponding to guideline texts in the clinical guideline; Carrying out semantic similarity matching on the second user problem and entity relations of a plurality of medical documents through the document matching engine, and determining a second medical conclusion and a corresponding medical document set according to a matching result, wherein any entity relation is a corresponding relation between a disease and the medical conclusion; reasoning according to the symptoms contained in the second user problem through the first large model to obtain a corresponding third medical conclusion; Determining, by the second large model, a final medical conclusion and basis for the second user problem based on the first medical conclusion, the second medical conclusion and corresponding clinical guideline, and the third medical conclusion and corresponding medical document set.
9. The medical agent generating device comprises a rule matching engine, a document matching engine, a first large model and a second large model, wherein the second large model is used for determining a final medical conclusion and basis according to the output of each engine and model, and the device comprises: The rule engine construction module is configured to execute conversion operation on the structured medical knowledge corresponding to the guideline original text in the clinical guideline to obtain a corresponding rule chain, and construct a rule matching engine based on the rule chain to enable the rule matching engine to match a first user problem with the rule chain and obtain a corresponding first medical conclusion and a corresponding clinical guideline according to a matching result, wherein any one rule chain comprises a plurality of mutually-related medical rules, and any one medical rule comprises a disease condition and a corresponding medical conclusion; The document engine construction module is configured to respectively extract a plurality of corresponding entity relations from a plurality of medical documents, and construct a document matching engine based on the entity relations, so that the document matching engine carries out semantic similarity matching on a first user problem and the entity relations of the plurality of medical documents, and determines a second medical conclusion and a corresponding medical document set according to a matching result, wherein any entity relation is a corresponding relation between a disease and the medical conclusion; and the large model fine-tuning module is configured to fine-tune the first large model by utilizing historical diagnosis and treatment data containing symptoms and medical conclusions, so that the fine-tuned first large model obtains a corresponding third medical conclusion according to the symptom reasoning contained in the first user problem.
10. A user problem response device based on a medical agent, wherein the medical agent comprises a rule matching engine, a document matching engine, a first large model and a second large model, and the device comprises: The rule matching module is configured to match a second user problem to be responded with a rule chain through the rule matching engine, and obtain a corresponding first medical conclusion and a corresponding clinical guideline according to a matching result, wherein any one rule chain comprises a plurality of medical rules which are associated with each other, any one medical rule comprises a disease condition and a corresponding medical conclusion, and the rule chain is obtained by executing conversion operation on structural medical knowledge corresponding to guideline original text in the clinical guideline; the document matching module is configured to perform semantic similarity matching on the second user problem and entity relations of a plurality of medical documents through the document matching engine, and determine a second medical conclusion and a corresponding medical document set according to a matching result, wherein any entity relation is a corresponding relation between a disease and the medical conclusion; the experience reasoning module is configured to infer according to symptoms contained in the second user problem through the first large model to obtain a corresponding third medical conclusion; a conclusion determination module configured to determine, by the second large model, a final medical conclusion and basis for the second user problem based on the first medical conclusion, the second medical conclusion, and the corresponding clinical guideline, and the third medical conclusion and the corresponding set of medical documents.
11. A computer readable storage medium having stored thereon a computer program which, when executed in a computer, causes the computer to perform the method of any of claims 1-8.
12. A computing device comprising a memory having executable code stored therein and a processor, which when executing the executable code, implements the method of any of claims 1-8.

Description

Medical intelligent agent generation method, response method and device Technical Field One or more embodiments of the present disclosure relate to the field of medical artificial intelligence, and in particular, to a method for generating a medical agent, a method for responding to the medical agent, and a device for generating the medical agent. Background The medical intelligent agent is also called as medical AI (Artificial intelligence) intelligent agent, is an AI system with certain autonomy, thinking and movement, has strong data processing capability and learning capability, and can perform complex reasoning and integration according to knowledge sources, thereby assisting in executing medical tasks. As a digital partner of a doctor, the doctor can deeply participate in a clinical decision process and provide decision support for the doctor. Medical science has very strong stringency and specialty, and medical scenes also have complexity. In an auxiliary medical setting, medical agents are increasingly required to be able to provide doctors with more stringent, more trusted auxiliary information. Medical intelligent agents also need to effectively protect input private data and prevent the private data from being revealed. At present, an improved scheme is desired to provide a medical agent by which more rigorous and more reliable medical assistance information is obtained. Disclosure of Invention One or more embodiments of the present disclosure describe a method, and an apparatus for generating a medical agent, so as to provide a medical agent, through which medical auxiliary information that is more strict and more reliable is obtained. The specific technical scheme is as follows. In a first aspect, an embodiment provides a method for generating a medical agent, where the medical agent includes a rule matching engine, a document matching engine, a first large model, and a second large model, where the second large model is used to determine a final medical conclusion and basis according to outputs of the engines and models, and the method includes: Performing conversion operation on structural medical knowledge corresponding to guideline texts in clinical guidelines to obtain a corresponding rule chain, and constructing a rule matching engine based on the rule chain to enable the rule matching engine to match a first user problem with the rule chain and obtain a corresponding first medical conclusion and a corresponding clinical guideline according to a matching result, wherein any one rule chain comprises a plurality of interrelated medical rules, and any one medical rule comprises disease conditions and corresponding medical conclusions; Respectively extracting a plurality of corresponding entity relations from a plurality of medical documents, and constructing a document matching engine based on the entity relations so that the document matching engine performs semantic similarity matching on a first user problem and the entity relations of the plurality of medical documents, and determining a second medical conclusion and a corresponding medical document set according to a matching result, wherein any entity relation is a corresponding relation between a disease and the medical conclusion; And carrying out fine adjustment on the first large model by utilizing historical diagnosis and treatment data containing symptoms and medical conclusions, so that the first large model after fine adjustment obtains a corresponding third medical conclusion according to the symptom reasoning contained in the first user problem. In one implementation, the step of performing a conversion operation for structured medical knowledge in a clinical guideline corresponding to guideline origin includes: Directly converting the structured medical knowledge into a corresponding initial rule chain; converting non-uniform terms in the initial rule chain into uniform terms according to a knowledge graph to obtain term alignment rule chains, wherein the knowledge graph takes terms as nodes and comprises association relations between a plurality of non-uniform terms and the uniform terms; The rule chain is determined based on the term alignment rule chain. In one implementation mode, the step of determining the rule chain based on the term alignment rule chain comprises the steps of determining labeling information of any term alignment rule chain, wherein the labeling information comprises timeliness information, specialty information and authority information, and adding the labeling information into the term alignment rule chain to obtain the rule chain. In one implementation, the step of obtaining the corresponding first medical conclusion according to the matching result includes selecting the first medical conclusion from a plurality of different medical conclusions according to labeling information of a rule chain in which the different medical conclusions are located when the matching result shows that the plurality