CN-121983217-A - Knowledge-enhanced reasoning chain verification-based patient information processing method

CN121983217ACN 121983217 ACN121983217 ACN 121983217ACN-121983217-A

Abstract

The invention provides a patient information processing method based on knowledge-enhanced reasoning chain verification, and relates to the technical field of medical data processing. The patient information processing method based on knowledge enhancement reasoning chain verification comprises the steps of constructing a first vector library containing meta index information and abstracts and a second vector library containing text semantic vectors, constructing a hierarchical search path, analyzing medical records, extracting structured information, reasoning according to a reasoning chain of sequence numbers, assertions types and assertions reasons, calculating a multidimensional weighted certainty score for each assertion of the reasoning chain, correcting the reasoning chain according to the scores, and finally outputting patient information processing results. The invention constructs a set of patient data reasoning method with interpretable, verifiable and repairable reasoning conclusion, which is used as an aid to reduce the misdiagnosis rate of doctors.

Inventors

ZHANG TAIYU
LI HANWEN
CAI HUA
YU DINGDING
Bai qi
ZHANG YIWEN
WANG HAORAN
XIA TIAN
CAO PEI
ZHAO SHUANG

Assignees

华院计算技术(上海)股份有限公司

Dates

Publication Date: 20260505
Application Date: 20260407

Claims (10)

1. A patient information processing method based on knowledge enhancement reasoning chain verification is characterized by comprising the following steps: S1, constructing a hierarchical medical knowledge base and a hierarchical retrieval path: Constructing a first vector library, wherein the first vector library comprises meta index information of medical books and semantic vectors of book abstracts; The hierarchical retrieval path is that meta-index information of books is retrieved from a first vector library, sub-text segments are retrieved from the partitioned text of the corresponding books and reordered, and the meta-index information is combined to obtain retrieval results; S2, analyzing medical record information, extracting structural information from the medical record information, and carrying out hierarchical retrieval when additional knowledge is judged to be needed; S3, reasoning according to a preset reasoning chain structure, wherein each link of the reasoning chain comprises a sequence number, a assertion, an assertion type and an assertion reason, wherein the assertion type comprises an evidence assertion, a diagnosis assertion and a strategy assertion; S4, carrying out hierarchical search on each assertion of the inference chain, and obtaining a deterministic score through weighted calculation, wherein weighted items comprise opposite semantic support degree, assertion reason consistency, assertion consistency and medical entity coverage rate; S5, repairing the inference chain by using the language model according to the retrieved knowledge for the assertion with the certainty score lower than the threshold value, and repeating the steps S4-S5 until the maximum iteration number is reached or no assertion with the certainty score lower than the threshold value exists; s6, outputting a patient information processing result according to the assertion.
2. The method according to claim 1, wherein in step S1, The meta index information comprises a title, an author and a publication year; the medical book block text is obtained by dividing the medical book into blocks based on the text length.
3. The method of claim 2, wherein the method of partitioning comprises: According to the analyzed mark down title element, the medical book text is segmented to obtain a paragraph element; Splitting paragraph elements with lengths larger than an upper threshold value by taking sentences as units and reorganizing according to the original sequence, and taking paragraphs with lengths smaller than the upper threshold value and the largest lengths as new paragraph elements; merging the paragraph element with the length smaller than the lower threshold value with the next paragraph element to obtain a new paragraph element; And splicing to enable the head and tail of the adjacent paragraph elements to have overlapping contents with set length values.
4. The method according to claim 1, wherein the structured information in step S2 includes basic information of the patient, complaint information, current medical history, examination information.
5. The method according to claim 1, wherein in the step S3, the setting of each assertion type includes: establishing a mapping relation between symptoms and diseases to be examined according to the complaint and current medical history examination information in the medical records; the diagnosis assertion is comprehensively inferred according to the evidence assertion, and the possible diseases of the patient are output; policy class assertions, based on diagnostic class assertions, giving follow-up treatment suggestions, including inspection or medication.
6. The method according to claim 1, wherein in the step S4, the calculation process of each item includes: the opposite semantic support degree is that knowledge base retrieval is respectively carried out on the assertion and the negative assertion generated aiming at the content of the assertion, the support degree of the retrieval content on the assertion and the negative assertion is respectively evaluated, and the relative support degree of the assertion relative to the negative assertion is further calculated; calculating the reason of the assertion and the similarity index of the text retrieved by the reason of the assertion, and taking the average value of the similarity indexes of the most relevant texts retrieved; asserting consistency, namely splicing the front assertion and the back assertion, inputting the assertions into a cross encoder, and calculating a consistency score by using an attention mechanism; and the coverage rate of the medical entity is obtained through statistics, wherein the coverage rate of the medical entity related to the current assertion in the knowledge base retrieval result is obtained.
7. The method of claim 6, wherein the medical entity comprises a disease, a symptom, a drug.
8. The method according to claim 1, wherein in step S6, the output is based on the diagnosis type, the reasoning result for the doctor to diagnose and refer to, and the proposal is based on the strategic assertion, including the disease to be examined and the index information to be examined.
9. The method according to claim 1, wherein the step S6 further comprises outputting detailed information of the inference chain.
10. A computer program product, characterized in that the computer program, when being executed by a processor, implements the steps of the method according to any one of claims 1-9.

Description

Knowledge-enhanced reasoning chain verification-based patient information processing method Technical Field The invention relates to the technical field of medical data processing, in particular to a patient information processing method based on knowledge-enhanced reasoning chain verification. Background Most of the existing medical data processing methods directly use the input of a user or the medical record information as the input of a retriever, and input the retrieved content into a large model to output a diagnosis result, and the mode has the problems that: 1. The search content is unstable, and the user input generally has spoken language characteristics and the expression of the proprietary books in the knowledge base exists in and out. When medical record information is used, the information types are multiple, and the information comprises information such as main complaints, medical history, test indexes and the like, so that the problem of noise in medical knowledge directly recalled is solved. 2. Hallucination risk when the knowledge of the recall is insufficient, the large model is prone to generating a clinically contradictory hallucinogenic diagnosis. In the medical diagnostic system, however, the accuracy is required to be high. 3. The interpretability and verifiability are insufficient, namely the model directly outputs a diagnosis conclusion, and the model lacks a verifiable evidence chain for asserting one by one, so that an expert is difficult to review diagnosis reasons. Thus, there is a need for an integrated method that combines knowledge base of medical books, inference chain generation and validation, large model diagnostic capabilities for providing more reliable, interpretable and verifiable auxiliary diagnostic conclusions. Disclosure of Invention Aiming at the defects existing in the prior art, the invention provides a patient information processing method based on knowledge-enhanced inference chain verification, in particular to a patient information processing method based on inference chain, retrieval enhancement generation (RETRIEVAL-Augmented Generation, RAG) and inference chain verification. In a first aspect, the present invention provides a patient information processing method based on knowledge-enhanced inference chain verification, including the steps of: S1, constructing a hierarchical medical knowledge base and a hierarchical retrieval path: Constructing a first vector library, wherein the first vector library comprises meta index information of medical books and semantic vectors of book abstracts; The hierarchical retrieval path is that meta-index information of books is retrieved from a first vector library, sub-text segments are retrieved from the partitioned text of the corresponding books and reordered, and the meta-index information is combined to obtain retrieval results; S2, analyzing medical record information, extracting structural information from the medical record information, and carrying out hierarchical retrieval when additional knowledge is judged to be needed; S3, reasoning according to a preset reasoning chain structure, wherein each link of the reasoning chain comprises a sequence number, a assertion, an assertion type and an assertion reason, wherein the assertion type comprises an evidence assertion, a diagnosis assertion and a strategy assertion; S4, carrying out hierarchical search on each assertion of the inference chain, and obtaining a deterministic score through weighted calculation, wherein weighted items comprise opposite semantic support degree, assertion reason consistency, assertion consistency and medical entity coverage rate; S5, repairing the inference chain by using the language model according to the retrieved knowledge for the assertion with the certainty score lower than the threshold value, and repeating the steps S4-S5 until the maximum iteration number is reached or no assertion with the certainty score lower than the threshold value exists; s6, outputting a patient information processing result according to the assertion. In a second aspect, the invention provides a computer program product which, when executed by a processor, implements the steps of the method of the first aspect. The method collects a large number of medical professional books, performs treatments such as blocking and vectorization, and constructs a medical knowledge base. And generating an inference chain containing an inference conclusion, a conclusion type and an inference basis based on the information of the patient, such as the complaint and the like by using the large model, mapping the inference conclusion and the inference basis to a medical knowledge base for consistency verification, and carrying out completion and repair of the inference chain, thereby realizing a patient data inference method with interpretable, verifiable and repairable inference conclusion, and reducing the misdiagnosis rate of doctors as an aid. Drawings FIG. 1 is a f