Search

CN-121999950-A - Quality evaluation and verification method for electronic medical record

CN121999950ACN 121999950 ACN121999950 ACN 121999950ACN-121999950-A

Abstract

The invention relates to a quality evaluation and verification method of an electronic medical record, which comprises the following steps of S1, cleaning, analyzing and standardizing a complete electronic medical record to obtain a standardized panoramic context of a patient and a text block set for vector retrieval, S2, constructing a mixed quality control knowledge base by an external medical knowledge base and a quality control rule set, S3, inputting the panoramic context of the patient, the mixed quality control knowledge base and a quality control point to be executed into a dynamic context manager, verifying the document sequence set by adopting a quality evaluation function, and outputting a structured defect report. The invention constructs an electronic medical record quality assessment and verification framework integrating dynamic context management, multi-source knowledge fusion and large language model reasoning capability, and has the core contribution that the context window limitation of LLM is effectively avoided by designing an exquisite context manager $\ mathcal { M $, so that the method can execute deep and interpretable logic examination on high-dimensional time-sequence EMR data.

Inventors

  • NONG BIN
  • JIA XUEMEI
  • LIU QING
  • JIANG HONGXIN
  • DENG YULIN
  • NI YA

Assignees

  • 西南财经大学天府学院

Dates

Publication Date
20260508
Application Date
20251216

Claims (7)

  1. 1. The electronic medical record quality evaluation and verification method is characterized by comprising the following steps of: S1, cleaning, analyzing and standardizing a complete electronic medical record, and reconstructing the complete electronic medical record into a unified time sequence event stream which takes a patient as a center to obtain a standardized patient panoramic context and a text block set for vector retrieval; s2, constructing a mixed quality control knowledge base by using an external medical knowledge base and a quality control rule set; S3, inputting the panoramic context of the patient, the mixed quality control knowledge base and the quality control points to be executed into a dynamic context manager, checking the text block set by adopting the mixed quality control knowledge base, and outputting a structured defect report.
  2. 2. The electronic medical record quality assessment and verification method according to claim 1, wherein step S2 comprises: analyzing the text quality control rule into an executable logic expression or a Datalog query; and slicing and vectorizing unstructured knowledge to construct a high-dimensional vector knowledge base.
  3. 3. The method for evaluating and verifying the quality of an electronic medical record according to claim 1, wherein in step S3, the dynamic context manager dynamically retrieves, screens and assembles information from the patient panoramic context and the mixed quality control knowledge base according to the quality control points to be executed to generate task-related, length-controlled dynamic contexts; And reasoning the dynamic context and the quality control points to be executed by adopting a large language model to obtain a structural defect report.
  4. 4. The electronic medical record quality assessment and verification method according to claim 3, wherein the generation of the dynamic context comprises Knowledge retrieval based on vector similarity: Mapping all document blocks in the normalized patient panoramic context and mixed quality control knowledge base to an n-dimensional vector space; Mapping the quality control points to be executed to the n-dimensional vector space; Retrieving Top-k most relevant context blocks by computing cosine similarity; timing context sliding window: Dividing the time sequence event stream into m overlapped windows W t ; Reasoning at the time t not only depends on the current window W t , but also depends on the compressed state abstract H t-1 ,H t at the time t-1 as a memory mechanism to transfer the history key information to the next window; recursive abstract and hierarchical context: The method comprises the steps of dividing a long file into leaf node blocks, recursively abstracting adjacent blocks to construct an abstract tree, reading only high-level abstract nodes when top-level quality control is executed, and reading lower-level abstract nodes when detail evidence is needed; When a large language model is adopted to infer a dynamic context and a quality control point to be executed, a quality control instruction Pq is a structured object and defined as: Pq=(Ro,T,C dyn ,CS,F out ), where Ro is role definition, T is specific quality control task description, cdyn is dynamic context generated by dynamic context manager, CS is constraint condition, and Fout is output format constraint.
  5. 5. The method for evaluating and verifying the quality of an electronic medical record according to claim 3 or 4, wherein the large language model obtains an original structured defect report through conditional probability generation, and then performs noise reduction and aggregation on the original structured defect report to finally be attributed to the structured defect report.
  6. 6. The method for evaluating and verifying the quality of an electronic medical record according to claim 1, wherein the structured defect report is verified manually after step S3.
  7. 7. The method for evaluating and verifying the quality of an electronic medical record according to claim 6, wherein the retrieval parameters of the large language model or the dynamic context manager are optimized by instruction fine tuning or reinforcement learning based by using data of verification feedback after manual verification.

Description

Quality evaluation and verification method for electronic medical record Technical Field The invention belongs to the field of medical services, and particularly relates to a quality evaluation and verification method for electronic medical records. Background The quality of the electronic medical record (Electronic Medical Record, EMR) is used as a digital core carrier of medical information, and the accuracy of clinical decisions, the level of medical safety assurance and the availability of follow-up scientific research data are directly determined. EMR is essentially a complex data set of high dimension, heterogeneous, and strong temporal nature. However, medical record construction in clinical practice is subject to interference from subjective factors, cognitive load and workflow, resulting in quality defects such as information absences (Omission), logical backups (Contradiction), semantic ambiguities (Ambiguity) and normalized deviations (device). The traditional Quality Control (QC) paradigm of medical records, whether it is an expert system relying on manual auditing or a Rule-based Engine based on keyword matching, has its inherent limitations. The former is difficult to unify and costly, while the latter is limited to shallow text features and cannot be effectively examined for deep semantic logic (e.g., causal links of "diagnosis-treatment-prognosis"). Disclosure of Invention The invention aims to solve the technical problem of providing an electronic medical record quality evaluation and verification method which simulates the cognition process of clinical specialists and realizes multi-dimensional and automatic depth verification of EMR completeness (Completeness), consistency (Consistency), logics (Logicality) and standardability (company). In order to solve the problems, the technical scheme adopted by the invention is that the electronic medical record quality evaluation and verification method comprises the following steps: S1, cleaning, analyzing and standardizing a complete electronic medical record, and reconstructing the complete electronic medical record into a unified time sequence event stream which takes a patient as a center to obtain a standardized patient panoramic context and a text block set for vector retrieval; s2, constructing a mixed quality control knowledge base by using an external medical knowledge base and a quality control rule set; S3, inputting the panoramic context of the patient, the mixed quality control knowledge base and the quality control points to be executed into a dynamic context manager, checking a document sequence set by adopting a quality evaluation function, and outputting a structured defect report. Further, step S2 includes: analyzing the text quality control rule into an executable logic expression or a Datalog query; and slicing and vectorizing unstructured knowledge to construct a high-dimensional vector knowledge base. Further, in step S3, the dynamic context manager dynamically retrieves, screens and assembles information from the patient panoramic context and the mixed quality control knowledge base according to the quality control points to be executed, and generates a task-related, length-controlled dynamic context; And reasoning the dynamic context and the quality control points to be executed by adopting a large language model to obtain a structural defect report. Further, the generation of the dynamic context includes Knowledge retrieval based on vector similarity: Mapping all document blocks in the normalized patient panoramic context and mixed quality control knowledge base to an n-dimensional vector space; Mapping the quality control points to be executed to the n-dimensional vector space; Retrieving Top-k most relevant context blocks by computing cosine similarity; timing context sliding window: Dividing the time sequence event stream into m windows Wt with overlapping; reasoning at the time t not only depends on the current window Wt, but also depends on the compressed state abstracts Ht-1 and Ht at the time t-1 as a memory mechanism to transfer the history key information to the next window; recursive abstract and hierarchical context: The method comprises the steps of dividing a long file into leaf node blocks, recursively abstracting adjacent blocks to construct an abstract tree, reading only high-level abstract nodes when top-level quality control is executed, and reading lower-level abstract nodes when detail evidence is needed; When a large language model is adopted to infer a dynamic context and a quality control point to be executed, a quality control instruction Pq is a structured object and defined as: Pq=(Ro,T,Cdyn,CS,Fout), where Ro is role definition, T is specific quality control task description, cdyn is dynamic context generated by dynamic context manager, CS is constraint condition, and Fout is output format constraint. Further, the large language model obtains an original structured defect report through condition