Electronic medical record intelligent evaluation method based on complex quality control indexes

Abstract

The invention discloses an electronic medical record intelligent evaluation method based on complex quality control indexes, and relates to the field of medical information processing and artificial intelligence. The method comprises the steps that an original electronic medical record text is collected and preprocessed, and structured diagnosis and treatment information and an event sequence diagram are extracted; through prompt word chain construction and a semantic reasoning mechanism, a large language model is guided to intelligently evaluate complex quality control indexes in medical records. The complex quality control indexes comprise diagnosis basis sufficiency, treatment scheme rationality, key result and change record integrity, treatment measure integrity and causal relationship rationality. According to the method, technologies such as a medical knowledge graph, a graph neural network and semantic vector retrieval are utilized to realize external knowledge recall and causal reasoning support; a self-consistency reasoning mechanism, a self-reflection mechanism and a multi-model cross validation mechanism are introduced to improve the accuracy and credibility of an evaluation result; and finally, outputting a structured quality control report and a visual reasoning chain. According to the method, the intelligence and refinement level in a complex medical quality control task can be remarkably improved, high interpretability and practical value are achieved, and the method is suitable for application scenes such as hospital quality management, scientific research evaluation and medical document standardization.

Description

Technical Field

[0001] The present invention relates to the field of medical information processing and artificial intelligence technology, and more particularly to the field of intelligent evaluation methods for electronic medical records based on complex quality control indicators. Background Art

[0002] With the continuous advancement of medical informatization, electronic medical record systems have been widely used in medical institutions at all levels, becoming the core vehicle for recording the entire patient diagnosis and treatment process. However, the quality of electronic medical records remains a major challenge facing the medical industry. Common issues include insufficient diagnostic documentation, unreasonable treatment plans, missing records of key changes, unreflected disposition measures, and unclear causal logic of diagnosis and treatment. These issues not only affect the accuracy and continuity of medical services but also may lead to medical risks and even medical accidents. Therefore, how to effectively evaluate the complex and core quality control indicators in electronic medical records has become a key issue that needs to be addressed in the medical quality control system.

[0003] Currently, some medical quality control systems have attempted to introduce artificial intelligence technology to automatically review medical record content through methods such as rule matching and semantic recognition. However, due to the complex structure, diverse languages, and fragmented information of electronic medical records, traditional rule-based methods often have difficulty making accurate judgments when faced with issues such as consistency between diagnosis and treatment, matching disease progression with treatment measures, and the causal relationship between the chief complaint and the conclusion. To this end, we propose a unified intelligent evaluation method for five core indicators of complex medical quality control: diagnosis, treatment, record keeping, treatment, and causality. This method fully leverages the semantic understanding and reasoning capabilities of large language models, introduces an external knowledge enhancement mechanism, and combines it with a self-feedback optimization strategy to achieve efficient, accurate, and comprehensive electronic medical record quality control. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent evaluation method for electronic medical records based on complex quality control indicators to improve the accuracy, comprehensiveness and intelligence level of electronic medical record quality control, and assist hospitals in achieving more efficient and accurate medical quality management.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] S1: Obtain original electronic medical record texts from the hospital information system, covering various medical document information such as medical history, physical examination, laboratory examination, diagnosis, treatment, and prognosis, as basic data input for subsequent processing;

[0007] S2: Electronic medical record preprocessing: Convert the original electronic medical record text into a large language model input format suitable for quality control reasoning. It includes the following sub-steps:

[0008] S21: Extraction of structural units from medical record content: Using a method combining segmentation recognition with syntactic semantics, key structural units are extracted from electronic medical record texts, including chief complaint, history of present illness, physical examination, laboratory tests, diagnosis conclusion, treatment plan, etc., to form segmentation structure information and provide contextual support for prompt word construction;

[0009] S22: Time sequence mapping of medical events: Identify medical events with clear timelines in medical records, construct event node maps, and organize them in chronological order to provide a time sequence basis for subsequent quality control assessments;

[0010] S23: Organize the extracted structural units and event nodes into a unified structured expression format to support prompt word template filling, prompt word chain construction, and downstream large language model inference and evaluation;

[0011] S3: Conduct complex quality control indicator evaluation based on the medical records to be evaluated, including:

[0012] S31: A prompt-driven evaluation mechanism based on semantic prompt word construction is proposed for five complex quality control indicators: adequacy of diagnostic basis, rationality of treatment plan, completeness of key results and change records, completeness of treatment measures, and rationality of causal relationships. This mechanism uses the indicator semantic prompt construction module to combine complex indicator types, medical record data, and standard medical knowledge graphs to dynamically construct prompt words for specific quality control targets.

[0013] S32: Based on the indicator semantic prompt words, the information recall module is used to retrieve relevant knowledge in the external medical knowledge base to assist in quality control evaluation. The information recall module is based on semantic vector retrieval, including: using the medical semantic encoding model to vectorize the prompt words, and constructing a query vector in combination with the core diagnosis and treatment elements extracted from the medical records; performing similarity matching between the query vector and the vectorized medical literature, guidelines, cases and other knowledge items, supporting semantic sorting and multi-source fusion, and completing the recall of relevant knowledge; on this basis, dynamically constructing a medical knowledge graph subgraph, extracting diagnosis and treatment entities and their relationships in the medical records, modeling the causal relationship between entities through the graph neural network, and combining the graph reasoning model to further explore the treatment path and diagnosis and treatment logic. Finally, the inferred structured knowledge is used as a supplementary input to enhance the subsequent model's understanding of the rationality of diagnosis and treatment, treatment path, and other aspects;

[0014] S33: Construct a prompt word chain for a specific quality control task. This prompt word chain integrates the core diagnosis and treatment elements in the medical record data, the prompt word content corresponding to the quality control indicators, and related external knowledge. The prompt structure is organized according to preset rules. The structure includes: initial task prompts, supplementary instructions, output format constraints, and a prompt module that supports adaptive updates. The prompt word chain serves as input to guide the large language model to carry out multiple rounds of semantic reasoning and evaluation, thereby completing intelligent judgment of complex quality control indicators in medical records.

[0015] S34: Self-consistency reasoning, self-reflection checking, and multi-model cross-validation mechanisms are introduced to verify and optimize the results of five types of complex quality control assessments. The large language model performs quality control reasoning based on the prompt word chain and generates preliminary conclusions. After generating the preliminary conclusion, if the confidence level of the conclusion is lower than a preset threshold, the prompt word adaptive iteration mechanism is triggered to automatically adjust or reconstruct the prompt word chain and re-execute the reasoning process until a conclusion that meets the confidence requirements is obtained.

[0016] S35: Introducing a large-scale model reasoning interpretation module, for each complex quality control indicator evaluation process, automatically generates a reasoning chain description that includes diagnosis and treatment information references, prompt word semantic analysis, and judgment basis, and displays it in a visual manner to reveal the logical path and reasoning basis of the large-scale model quality control judgment, thereby enhancing the credibility and auditability of the evaluation conclusion;

[0017] S36: During the quality control evaluation process, for each type of complex quality control indicator, the following specific strategies are adopted to enhance the reasoning ability and evaluation accuracy of the large language model:

[0018] Sufficiency of diagnostic evidence: Construct a reasoning path diagram to track the "diagnosis support chain" in the medical record. Automatically generate counterexamples for missing parts to prompt rewriting. Use a large language model for "chain reasoning" and combine it with the medical knowledge graph to score the diagnosis support.

[0019] Rationality of treatment plans: A "disease-treatment pathway map constraint mechanism" is introduced to construct a "treatment pathway map" covering major diseases, which includes diagnostic nodes, treatment nodes, and guideline constraint relationships. During the evaluation process, it is determined whether there are "path deviations," "path jumps," or "violations of the recommended sequence." If there are any anomalies, personalized rationality judgments are also made based on patient characteristics and medication contraindications.

[0020] Completeness of key results and change records: Using the "cross-paragraph alignment-evolution trajectory modeling" method, time series modeling and clustering methods are used to analyze indicator change trends and record continuity. If there are logical breaks or omissions, prompts are automatically generated;

[0021] Completeness of treatment measures: Utilizing the medical event chain map to conduct "treatment response delay detection," the extracted medical event is evaluated to determine whether a response is given within a reasonable timeframe. This includes, for example, whether bleeding is stopped, whether oxygen is administered for dyspnea, etc., to detect issues such as "treatment delay," "no response," and "response mismatch."

[0022] Reasonable causal relationship: Construct a causal path and use a language model combined with a medical knowledge graph for forward and reverse reasoning. If there are incomplete paths or interruptions, such as mismatches between diagnosis and treatment, the causal relationship is judged to be unreasonable.

[0023] S4: Output a unified quality control report, including quantitative scoring, problem location and annotation, and structured output for five types of complex quality control indicators.

[0024] As a further solution of the present invention: the specific steps in S31 are as follows:

[0025] S311: Based on the type of complex quality control indicators to be evaluated, identify target categories from five categories: adequacy of diagnostic basis, rationality of treatment plan, completeness of key results and change records, completeness of treatment measures, and rationality of causal relationships;

[0026] S312: Matching preset prompt word templates for different indicator categories. The templates are designed based on a structured question framework, combining the logical structure of each indicator category with medical knowledge building elements, and defining standardized input slot variables and semantic framework constraints;

[0027] S313: Combining medical record data with the medical knowledge graph, content is extracted from multiple sources of data, including medical history, physical signs, examination results, treatment procedures, and disease progression, to dynamically fill variable slots in the prompt word template, ensuring semantic integrity and contextual coherence of the prompt word.

[0028] S314: Outputting natural language prompt words for specific indicators. The prompt words have medical interpretability and semantic reasoning capabilities and can be used as input prompt content for the large language model to perform quality control tasks;

[0029] S315: The prompt word generation process supports the embedding of chain thinking reasoning templates to guide the language model to perform multi-step quality control judgments and improve the reasoning depth and accuracy.

[0030] As a further solution of the present invention: the specific steps in S32 are as follows:

[0031] S321: Extract key diagnostic and treatment elements, including symptoms, diagnosis, and treatment options;

[0032] S322: Based on the prompt words and diagnosis and treatment elements, use the pre-trained medical semantic encoding model to vectorize them;

[0033] S323: performing similarity matching between the query vector and pre-vectorized literature, guidelines, or cases in an external medical knowledge base to obtain relevant knowledge fragments;

[0034] S324: Semantically sort the matching results and integrate candidate medical knowledge from multiple sources to remove redundancy.

[0035] S325: Based on the extracted medical record entities and knowledge fragments, a local medical knowledge graph is constructed, entities and their relationships in the medical records are dynamically extracted, and entity relationships are modeled using a graph neural network;

[0036] S326: Introducing a graph reasoning model to analyze possible causal chains or diagnosis and treatment pathways and form structured reasoning knowledge;

[0037] S327: The candidate medical knowledge fragments generated by integrating semantic recall and graph reasoning are output as auxiliary evaluation references for subsequent reasoning and analysis.

[0038] As a further solution of the present invention: the specific steps in S34 are as follows:

[0039] S341: Perform self-consistent reasoning on five types of complex quality control evaluation indicators, generate candidate conclusions through multiple rounds of prompts, and vote based on the consistency of the results;

[0040] S342: Guide the large model to conduct Chain-of-Thought self-reflection checks to identify potential evaluation omissions or reasoning loopholes;

[0041] S343: Use multi-model cross-validation to compare the consistency of the evaluation results of multiple models under the same indicator;

[0042] S344: Based on the feedback results of various mechanisms, generate the final evaluation conclusion and its credibility score for each indicator; if the confidence of the conclusion is lower than the preset threshold, trigger the prompt word adaptive iteration mechanism to automatically adjust or reconstruct the prompt word chain and re-execute the quality control reasoning until a conclusion that meets the confidence requirements is obtained.

[0043] As a further solution of the present invention: the specific steps in S35 are as follows:

[0044] S351: Based on the intermediate reasoning results and prompt word input of the large language model during the execution of complex quality control judgments, a logical reasoning chain text is generated to explain the semantic deduction path behind the conclusion;

[0045] S352: Mark and cite key diagnostic and treatment elements in medical records and supporting medical knowledge recalled from external knowledge bases as the basis for reasoning and judgment;

[0046] S353: Using visualization components, the reasoning chain is converted into a graphical display, including the semantic decomposition of prompt words, diagnosis and treatment information nodes, causal judgment path, and conclusion output path, forming a complete and auditable reasoning chain diagram;

[0047] S354: Output the reasoning chain diagram and the final quality control conclusion together for clinical staff to review and provide feedback, thereby improving the comprehensibility and traceability of the evaluation results.

[0048] As a further solution of the present invention: the specific steps in S36 are as follows:

[0049] S361: The assessment of the adequacy of diagnostic evidence includes the following steps: constructing a reasoning path diagram based on medical record data, with the diagnosis conclusion as the terminal node, and the medical history, physical signs, test results, imaging findings, etc. extracted from the medical records as intermediate nodes, connecting edges according to medical causal logic to generate a diagnosis support chain; using a large language model to generate language counterexamples for missing or broken nodes in the path, to assist doctors in identifying insufficient information; and combining the diagnosis-evidence mapping relationship in the medical knowledge graph to perform semantic rationality judgment on each reasoning path. Finally, the chain reasoning model CoT framework generates a diagnosis support score and outputs a structured quality control result;

[0050] S362: The evaluation of the rationality of the treatment plan includes the following steps: constructing a disease-treatment pathway map based on standard diagnosis and treatment guidelines, which includes diagnosis nodes, treatment action nodes, path sequence edges, and guideline constraint labels; using a graph structure matching algorithm to compare the treatment pathway extracted from the medical record with the standard pathway map to determine whether there are abnormal patterns such as "path jump", "path reversal", "path separation", or "reverse implementation"; if there are abnormalities, further calling the patient feature nodes in the knowledge graph, using a large language model to determine the personalized rationality of the deviation from the path, and providing natural language prompts;

[0051] S363: The assessment of the integrity of key results and change records includes the following steps: First, uniformly extract the content related to numerical indicators in the medical record, align them across paragraphs based on semantic representation and position encoding technology, and normalize and match the descriptions of the same medical indicator at different time points; second, use time series modeling methods combined with clustering analysis algorithms to model the trend and construct the trajectory of indicator changes; if there are abnormal situations such as trajectory interruption, sudden change in numerical value, or sparse records, the logical break segments are automatically marked and natural language prompts are generated for reminders to improve the structural integrity of the medical record;

[0052] S364: The assessment of the integrity of treatment measures includes the following steps: constructing a medical event chain map, which uses clinical symptoms and sudden illness as trigger event nodes, and diagnosis and treatment operations as response nodes, and represents the response logic through edge connections; using a timestamp extraction method to mark the event occurrence time and response measure time, and based on the event-response time window model to determine whether there is a "delayed response", "missing response" or "response mismatch"; at the same time, calling the disease-treatment rule matching library in the knowledge graph to verify the semantic relevance between the response behavior and the triggering event, and generating a reasonable explanation text through a language model;

[0053] S365: The evaluation of the rationality of causal relationships includes the following steps: constructing a complete causal path diagram in the medical record through entity relationship extraction technology, which connects the chief complaint, examination, diagnosis, treatment, and results into a diagram according to time and causal logic; applying a two-way causal reasoning mechanism based on a large language model, combined with the medical knowledge graph to verify the path consistency, and automatically identifying if "path interruption", "causal loss" or "causal error reversal" is found; after the abnormal causal path is determined by the system, the language model is called to generate structured prompt content and rewriting suggestions to improve the logical consistency and medical explanation ability of the causal chain.

[0054] By adopting the above technical solution, compared with the prior art, the beneficial effects of the present invention are:

[0055] 1. Intelligent Assessment of Complex Quality Control Indicators: This paper proposes a prompt-driven assessment mechanism based on semantic prompt words. Combining structured medical record data, standard medical knowledge graphs, and the reasoning capabilities of large language models, it enables intelligent assessment of multiple complex quality control indicators, including diagnostic evidence, treatment plans, and key outcome records. This assessment mechanism is not only efficient and accurate, but also adaptively adjusts to clinical needs, significantly improving the intelligence of quality control assessment.

[0056] 2. Multi-source integration of semantic recall and graph reasoning: By combining the semantic recall module with graph reasoning technology, relevant information from external medical knowledge bases can be effectively retrieved and integrated, further improving the accuracy and reliability of the evaluation process. In addition, the graph neural network's modeling of medical record entity relationships can accurately reveal the causal relationships in the disease diagnosis and treatment process, helping doctors better understand the diagnosis and treatment logic.

[0057] 3. Adaptive Iteration Mechanism and Model Self-Validation: This invention incorporates self-consistent reasoning, self-reflection, and multi-model cross-validation mechanisms to perform multiple rounds of optimization on complex quality control assessment results, ensuring the reliability of the conclusions. When the confidence level of the assessment conclusion falls below a preset threshold, the system automatically adjusts the prompt word chain and re-infers until a reliable assessment result is reached, significantly improving the accuracy and credibility of the assessment.

[0058] 4. Reasoning Chain Visualization and Result Interpretation: This invention uses a large-scale model reasoning interpretation module to automatically generate a logical path containing diagnosis and treatment information references and reasoning chain descriptions, and displays this in a visual manner, helping clinical staff better understand and review the evaluation conclusions. This visual display not only improves the traceability of evaluation results, but also enhances the transparency and auditability of the evaluation process.

[0059] 5. Comprehensive Processing of Complex Medical Scenarios: Through multi-step reasoning and evaluation, this invention enables comprehensive and in-depth analysis of complex medical information in medical records, including assessment of the completeness and rationality of diagnostic support, treatment pathways, and dispositions. Compared to traditional quality control and assessment methods, this invention provides more efficient, accurate, and intelligent assessment capabilities in more complex medical scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to make the purpose, technical solutions and advantages of the invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings, in which:

[0061] Figure 1 This is a schematic diagram of the overall process of the electronic medical record intelligent evaluation method provided by the present invention;

[0062] Figure 2 It is a complex quality control evaluation flow chart provided by the present invention;

[0063] Figure 3 This is an evaluation strategy diagram for five complex quality control indicators provided by the present invention;

[0064] Figure 4 It is a case diagram for evaluating the adequacy of diagnostic evidence constructed by the present invention;

[0065] Figure 5 It is a prompt case diagram for evaluating the rationality of the treatment plan constructed by the present invention. DETAILED DESCRIPTION

[0066] The present invention will be described in further detail below with reference to the accompanying drawings and implementation examples so that those skilled in the art can more clearly understand the technical solution of the present invention, but this does not limit the scope of protection of the present invention.

[0067] like Figure 1 As shown, the present invention provides an intelligent evaluation method for electronic medical records based on complex quality control indicators, comprising the following steps:

[0068] S1: Obtain original electronic medical record texts from the hospital information system, covering various medical document information such as medical history, physical examination, laboratory examination, diagnosis, treatment, and prognosis, as basic data input for subsequent processing;

[0069] S2: Electronic medical record preprocessing: Convert the original electronic medical record text into a large language model input format suitable for quality control reasoning. It includes the following sub-steps:

[0070] S21: Extraction of structural units from medical record content: Using a method combining segmentation recognition with syntactic semantics, key structural units are extracted from electronic medical record texts, including chief complaint, history of present illness, physical examination, laboratory tests, diagnosis conclusion, treatment plan, etc., to form segmentation structure information and provide contextual support for prompt word construction;

[0071] S22: Time sequence mapping of medical events: Identify medical events with clear timelines in medical records, construct event node maps, and organize them in chronological order to provide a time sequence basis for subsequent quality control assessments;

[0072] S23: Organize the extracted structural units and event nodes into a unified structured expression format to support prompt word template filling, prompt word chain construction, and downstream large language model inference and evaluation;

[0073] S3: Conduct complex quality control indicator evaluation based on the medical records to be evaluated, including:

[0074] S31: A prompt-driven evaluation mechanism based on semantic prompt word construction is proposed for five complex quality control indicators: adequacy of diagnostic basis, rationality of treatment plan, completeness of key results and change records, completeness of treatment measures, and rationality of causal relationships. This mechanism uses the indicator semantic prompt construction module to combine complex indicator types, medical record data, and standard medical knowledge graphs to dynamically construct prompt words for specific quality control targets.

[0075] S32: Based on the indicator semantic prompt words, the information recall module is used to retrieve relevant knowledge in the external medical knowledge base to assist in quality control evaluation. The information recall module is based on semantic vector retrieval, including: using the medical semantic encoding model to vectorize the prompt words, and constructing a query vector in combination with the core diagnosis and treatment elements extracted from the medical records; performing similarity matching between the query vector and the vectorized medical literature, guidelines, cases and other knowledge items, supporting semantic sorting and multi-source fusion, and completing the recall of relevant knowledge; on this basis, dynamically constructing a medical knowledge graph subgraph, extracting diagnosis and treatment entities and their relationships in the medical records, modeling the causal relationship between entities through the graph neural network, and combining the graph reasoning model to further explore the treatment path and diagnosis and treatment logic. Finally, the inferred structured knowledge is used as a supplementary input to enhance the subsequent model's understanding of the rationality of diagnosis and treatment, treatment path, and other aspects;

[0076] S33: Construct a prompt word chain for a specific quality control task. This prompt word chain integrates the core diagnosis and treatment elements in the medical record data, the prompt word content corresponding to the quality control indicators, and related external knowledge. The prompt structure is organized according to preset rules. The structure includes: initial task prompts, supplementary instructions, output format constraints, and a prompt module that supports adaptive updates. The prompt word chain serves as input to guide the large language model to carry out multiple rounds of semantic reasoning and evaluation, thereby completing intelligent judgment of complex quality control indicators in medical records.

[0077] S34: Self-consistency reasoning, self-reflection checking, and multi-model cross-validation mechanisms are introduced to verify and optimize the results of five types of complex quality control assessments. The large language model performs quality control reasoning based on the prompt word chain and generates preliminary conclusions. After generating the preliminary conclusion, if the confidence level of the conclusion is lower than a preset threshold, the prompt word adaptive iteration mechanism is triggered to automatically adjust or reconstruct the prompt word chain and re-execute the reasoning process until a conclusion that meets the confidence requirements is obtained.

[0078] S35: Introducing a large-scale model reasoning interpretation module, for each complex quality control indicator evaluation process, automatically generates a reasoning chain description that includes diagnosis and treatment information references, prompt word semantic analysis, and judgment basis, and displays it in a visual manner to reveal the logical path and reasoning basis of the large-scale model quality control judgment, thereby enhancing the credibility and auditability of the evaluation conclusion;

[0079] S36: During the quality control evaluation process, for each type of complex quality control indicator, the following specific strategies are adopted to enhance the reasoning ability and evaluation accuracy of the large language model:

[0080] Sufficiency of diagnostic evidence: Construct a reasoning path diagram to track the "diagnosis support chain" in the medical record. Automatically generate counterexamples for missing parts to prompt rewriting. Use a large language model for "chain reasoning" and combine it with the medical knowledge graph to score the diagnosis support.

[0081] Rationality of treatment plans: A "disease-treatment pathway map constraint mechanism" is introduced to construct a "treatment pathway map" covering major diseases, which includes diagnostic nodes, treatment nodes, and guideline constraint relationships. During the evaluation process, it is determined whether there are "path deviations," "path jumps," or "violations of the recommended sequence." If there are any anomalies, personalized rationality judgments are also made based on patient characteristics and medication contraindications.

[0082] Completeness of key results and change records: Using the "cross-paragraph alignment-evolution trajectory modeling" method, time series modeling and clustering methods are used to analyze indicator change trends and record continuity. If there are logical breaks or omissions, prompts are automatically generated;

[0083] Completeness of treatment measures: Utilizing the medical event chain map to conduct "treatment response delay detection," the extracted medical event is evaluated to determine whether a response is given within a reasonable timeframe. This includes, for example, whether bleeding is stopped, whether oxygen is administered for dyspnea, etc., to detect issues such as "treatment delay," "no response," and "response mismatch."

[0084] Reasonable causal relationship: Construct a causal path and use a language model combined with a medical knowledge graph for forward and reverse reasoning. If there are incomplete paths or interruptions, such as mismatches between diagnosis and treatment, the causal relationship is judged to be unreasonable.

[0085] S4: Output a unified quality control report, including quantitative scoring, problem location and annotation, and structured output for five types of complex quality control indicators.

[0086] like Figure 4 As shown, the present invention illustrates a diagnostic evidence adequacy assessment method based on reasoning path construction and chain thinking interpretation, which is used to perform multi-level automatic reasoning and quality assessment on structured medical record data. The method mainly includes the following steps:

[0087] Structured information extraction: Extract key structured medical information from electronic medical record text, including chief complaint, medical history, physical examination, test results, imaging findings, etc., and standardize it into data nodes that can be used for reasoning and graph building;

[0088] Reasoning Path Diagram Construction: With the diagnosis as the endpoint of the reasoning path, structured evidence information, including medical history, physical signs, laboratory test results, and imaging findings, is extracted from electronic medical records as intermediate nodes in the path. Each node is connected by directed edges with medical causal relationships to construct a complete reasoning path diagram.

[0089] Missing node identification and language counterexample generation: If there are missing nodes, logical breaks, or jumps in the path, the large language model is used to generate language counterexamples or questioning explanations to help indicate insufficient evidence chains, such as: "The lack of chest X-ray results cannot support the diagnosis of 'pneumonia'." This process facilitates model self-review and assists doctors in quality control.

[0090] Semantic rationality assessment: Combined with the standard "evidence → diagnosis" path relationship in the medical knowledge graph, the constructed reasoning path is compared and tested to identify whether there are logical errors, illegal jumps, or insufficient evidence confidence, and the consistency of the path semantics is evaluated;

[0091] Diagnostic support scoring: Using a chain-thinking reasoning framework, each reasoning path is gradually explained and scored on a scale of 0 to 1. A comprehensive assessment is conducted based on the completeness of evidence, path rationality, and medical semantic consistency to form a quantitative diagnostic support result.

[0092] Structured output generation: The final output includes the following: diagnostic conclusions, key support paths, language counterexamples, semantic consistency judgments, diagnostic support scores, and evaluation reasons, forming structured evaluation results that can be used for automatic quality control and review.

[0093] like Figure 5 As shown in the figure, the present invention uses an example to demonstrate a method for constructing prompt words for medical quality control. The method aims to generate structured, semantically clear, and reasonable prompt words for the "rationality of treatment plan" in clinical quality control indicators, thereby improving the large language model's ability to understand and judge medical texts. The method mainly includes the following steps:

[0094] Determine target quality control indicators: The system receives user input or preset quality control indicator types, such as "reasonableness of treatment plan." This indicator is usually used to determine whether the treatment measures received by the patient are consistent with their diagnosis, symptoms, and examination results, and whether they meet the recommended standards of clinical guidelines;

[0095] Matching prompt word templates: Based on the indicator type, the system retrieves or calls the corresponding prompt word template. The template is in the form of structured slots, such as "[Diagnosis]: {diagnosis}, etc.", and also embeds task instructions, such as "Please determine whether the current treatment plan is reasonable and provide analysis basis and suggestions."

[0096] Filling structured medical record information: The system automatically matches the patient's relevant structured data content, including diagnosis results, symptoms, laboratory or imaging test results, actual treatment plans, etc., and dynamically fills it into the prompt word template;

[0097] Generate prompt text: Combine the completed templates into a complete, semantically clear, and contextually coherent prompt. The system can automatically add guiding words, such as "Please analyze step by step to see if it is reasonable," to enhance the logical structure and task orientation of the prompt.

[0098] Embedded chain thinking guidance: To improve the model's reasoning accuracy, chain prompts such as "step-by-step analysis" and "basis → analysis → conclusion" are added to the prompts to guide the large model to think step by step, such as judging whether the treatment direction is accurate based on the diagnosis, analyzing whether the drug used is a first-line recommended drug, etc.

[0099] Output format standardization: The prompt word text output by the system has result format constraints;

[0100] Knowledge guidance and reference prompts: Reference knowledge prompts can be added, such as "Please refer to the "Guidelines for the Diagnosis and Treatment of Acute Pancreatitis (2020)" or related PubMed literature" to enhance the medical consistency and authority of the model generation results.

[0101] This implementation achieves a comprehensive assessment of electronic medical record quality through multiple approaches, including semantic understanding, knowledge enhancement, and the integration of reasoning mechanisms. This effectively addresses the inability of traditional rule-based systems to handle complex clinical semantics. The method can be deployed in hospital quality control centers, medical records offices, and medical departments, or as an independent third-party intelligent quality control platform. It has practical application value in multiple scenarios, including inpatient medical record quality control, discharge record sampling, key case review, and medical safety incident investigation.

[0102] Specific embodiments of the present invention have been described above with reference to the accompanying drawings. However, those skilled in the art will appreciate that various modifications and substitutions may be made to the specific embodiments of the present invention without departing from the spirit and scope of the present invention. Such modifications and substitutions are intended to fall within the scope of the claims.

Claims

1. An intelligent evaluation method for electronic medical records based on complex quality control indicators, characterized by: The steps include: S1: Obtain original electronic medical record texts from the hospital information system, covering various medical document information such as medical history, physical examination, laboratory examination, diagnosis, treatment, and prognosis, as basic data input for subsequent processing; S2: Electronic medical record preprocessing: Convert the original electronic medical record text into a large language model input format suitable for quality control reasoning. It includes the following sub-steps: S21: Extraction of structural units from medical record content: Using a method combining segmentation recognition with syntactic semantics, key structural units are extracted from electronic medical record texts, including chief complaint, history of present illness, physical examination, laboratory tests, diagnosis conclusion, treatment plan, etc., to form segmentation structure information and provide contextual support for prompt word construction; S22: Time sequence mapping of medical events: Identify medical events with clear timelines in medical records, construct event node maps, and organize them in chronological order to provide a time sequence basis for subsequent quality control assessments; S23: Organize the extracted structural units and event nodes into a unified structured expression format to support prompt word template filling, prompt word chain construction, and downstream large language model inference and evaluation; S3: Conduct complex quality control indicator evaluation based on the medical records to be evaluated, including: S31: A prompt-driven evaluation mechanism based on semantic prompt word construction is proposed for five complex quality control indicators: adequacy of diagnostic basis, rationality of treatment plan, completeness of key results and change records, completeness of treatment measures, and rationality of causal relationships. This mechanism uses the indicator semantic prompt construction module to combine complex indicator types, medical record data, and standard medical knowledge graphs to dynamically construct prompt words for specific quality control targets. S32: Based on the indicator semantic prompt words, the information recall module is used to retrieve relevant knowledge in the external medical knowledge base to assist in quality control evaluation. The information recall module is based on semantic vector retrieval, including: using the medical semantic encoding model to vectorize the prompt words, and constructing a query vector in combination with the core diagnosis and treatment elements extracted from the medical records; performing similarity matching between the query vector and the vectorized medical literature, guidelines, cases and other knowledge items, supporting semantic sorting and multi-source fusion, and completing the recall of relevant knowledge; on this basis, dynamically constructing a medical knowledge graph subgraph, extracting diagnosis and treatment entities and their relationships in the medical records, modeling the causal relationship between entities through the graph neural network, and combining the graph reasoning model to further explore the treatment path and diagnosis and treatment logic. Finally, the inferred structured knowledge is used as a supplementary input to enhance the subsequent model's understanding of the rationality of diagnosis and treatment, treatment path, and other aspects; S33: Construct a prompt word chain for a specific quality control task. This prompt word chain integrates the core diagnosis and treatment elements in the medical record data, the prompt word content corresponding to the quality control indicators, and related external knowledge. The prompt structure is organized according to preset rules. The structure includes: initial task prompts, supplementary instructions, output format constraints, and a prompt module that supports adaptive updates. The prompt word chain serves as input to guide the large language model to carry out multiple rounds of semantic reasoning and evaluation, thereby completing intelligent judgment of complex quality control indicators in medical records. S34: Self-consistency reasoning, self-reflection checking, and multi-model cross-validation mechanisms are introduced to verify and optimize the results of five types of complex quality control assessments. The large language model performs quality control reasoning based on the prompt word chain and generates preliminary conclusions. After generating the preliminary conclusion, if the confidence level of the conclusion is lower than a preset threshold, the prompt word adaptive iteration mechanism is triggered to automatically adjust or reconstruct the prompt word chain and re-execute the reasoning process until a conclusion that meets the confidence requirements is obtained. S35: Introducing a large-scale model reasoning interpretation module, for each complex quality control indicator evaluation process, automatically generates a reasoning chain description that includes diagnosis and treatment information references, prompt word semantic analysis, and judgment basis, and displays it in a visual manner to reveal the logical path and reasoning basis of the large-scale model quality control judgment, thereby enhancing the credibility and auditability of the evaluation conclusion; S36: During the quality control evaluation process, for each type of complex quality control indicator, the following specific strategies are adopted to enhance the reasoning ability and evaluation accuracy of the large language model: Sufficiency of diagnostic evidence: Build a reasoning path diagram to track the "diagnosis support chain" in the medical record. Automatically generate counterexamples for missing parts to prompt rewriting. Use a large language model for "chain reasoning" and, with the assistance of the medical knowledge graph, score the diagnosis support. Treatment plan rationality: A "disease-treatment pathway map constraint mechanism" is introduced to construct a "treatment pathway map" covering major diseases, which includes diagnostic nodes, treatment nodes, and guideline constraint relationships. During the evaluation process, it determines whether there are "path deviations," "path jumps," or "violations of the recommended sequence." If there are any anomalies, personalized rationality judgments are also made based on patient characteristics and medication contraindications. Completeness of key results and change records: Using the "cross-paragraph alignment-evolution trajectory modeling" method, time series modeling and clustering methods are used to analyze indicator change trends and record continuity. If there are logical breaks or omissions, prompts are automatically generated. Completeness of treatment measures: Utilizing the medical event chain map to conduct "treatment response delay detection," the extracted medical event is evaluated to determine whether a response is provided within a reasonable timeframe. This includes, for example, whether bleeding is stopped and whether oxygen is administered for dyspnea. This allows for detection of issues such as "treatment delay," "no response," and "response mismatch." Reasonable causal relationship: Construct a causal path and use a language model combined with a medical knowledge graph for forward and reverse reasoning. If there are incomplete paths or interruptions, such as mismatches between diagnosis and treatment, the causal relationship is judged to be unreasonable. S4: Output a unified quality control report, including quantitative scoring, problem location and annotation, and structured output for five types of complex quality control indicators.

2. The method according to claim 1, characterized in that The extraction of medical record content structural units in S21 is achieved by combining segmentation rules with syntactic and semantic analysis. First, paragraphs are divided based on the sentence features and formatting clues in the electronic medical record text, and then the core semantic structure of each paragraph is identified using syntactic analysis methods, including subject-verb-object relationships and key modifying components. The medical semantic template and contextual information are then combined to determine the paragraph type, thereby achieving automatic labeling of structural units such as chief complaint, current medical history, physical examination, laboratory examination, diagnostic conclusion, and treatment plan, thereby forming a structured medical record expression for downstream prompt word construction and large model reasoning tasks.

3. The method according to claim 1, characterized in that In S31, the implementation of the indicator semantic prompt generator includes the following sub-steps: S311: Based on the type of complex quality control indicators to be evaluated, identify target categories from five categories: adequacy of diagnostic basis, rationality of treatment plan, completeness of key results and change records, completeness of treatment measures, and rationality of causal relationships; S312: Matching preset prompt word templates for different indicator categories. The templates are designed based on a structured question framework, combining the logical structure of each indicator category with medical knowledge building elements, and defining standardized input slot variables and semantic framework constraints; S313: Combining medical record data with the medical knowledge graph, content is extracted from multiple sources of data, including medical history, physical signs, examination results, treatment procedures, and disease progression, to dynamically fill variable slots in the prompt word template, ensuring semantic integrity and contextual coherence of the prompt word. S314: Outputting natural language prompt words for specific indicators. The prompt words have medical interpretability and semantic reasoning capabilities and can be used as input prompt content for the large language model to perform quality control tasks; S315: The prompt word generation process supports the embedding of chain thinking reasoning templates to guide the language model to perform multi-step quality control judgments and improve the reasoning depth and accuracy.

4. The method according to claim 1, wherein The information recall module in S32 specifically includes: S321: Extract key diagnostic and treatment elements, including symptoms, diagnosis, and treatment options; S322: Based on the prompt words and diagnosis and treatment elements, use the pre-trained medical semantic encoding model to vectorize them; S323: performing similarity matching between the query vector and pre-vectorized literature, guidelines, or cases in an external medical knowledge base to obtain relevant knowledge fragments; S324: Semantically sort the matching results and integrate candidate medical knowledge from multiple sources to remove redundancy. S325: Based on the extracted medical record entities and knowledge fragments, a local medical knowledge graph is constructed, entities and their relationships in the medical records are dynamically extracted, and entity relationships are modeled using a graph neural network; S326: Introducing a graph reasoning model to analyze possible causal chains or diagnosis and treatment pathways and form structured reasoning knowledge; S327: The candidate medical knowledge fragments generated by integrating semantic recall and graph reasoning are output as auxiliary evaluation references for subsequent reasoning and analysis.

5. The method according to claim 1, characterized in that The evaluation result verification and optimization step in S34 includes: S341: Perform self-consistent reasoning on five types of complex quality control evaluation indicators, generate candidate conclusions through multiple rounds of prompts, and vote based on the consistency of the results; S342: Guide the large model to conduct Chain-of-Thought self-reflection checks to identify potential evaluation omissions or reasoning loopholes; S343: Use multi-model cross-validation to compare the consistency of the evaluation results of multiple models under the same indicator; S344: Based on the feedback results of various mechanisms, generate the final evaluation conclusion and its credibility score for each indicator; if the confidence of the conclusion is lower than the preset threshold, trigger the prompt word adaptive iteration mechanism to automatically adjust or reconstruct the prompt word chain and re-execute the quality control reasoning until a conclusion that meets the confidence requirements is obtained.

6. The method according to claim 1, characterized in that The large model reasoning and explanation module in S35 includes the following steps: S351: Based on the intermediate reasoning results and prompt word input of the large language model during the execution of complex quality control judgments, a logical reasoning chain text is generated to explain the semantic deduction path behind the conclusion; S352: Mark and cite key diagnostic and treatment elements in medical records and supporting medical knowledge recalled from external knowledge bases as the basis for reasoning and judgment; S353: Using visualization components, the reasoning chain is converted into a graphical display, including the semantic decomposition of prompt words, diagnosis and treatment information nodes, causal judgment path, and conclusion output path, forming a complete and auditable reasoning chain diagram; S354: Output the reasoning chain diagram and the final quality control conclusion together for clinical staff to review and provide feedback, thereby improving the comprehensibility and traceability of the evaluation results.

7. The method according to claim 1, characterized in that The S36 adopts the following strategies for each type of complex quality control indicator: S361: The assessment of the adequacy of diagnostic evidence includes the following steps: constructing a reasoning path diagram based on medical record data, with the diagnosis conclusion as the terminal node, and the medical history, physical signs, test results, imaging findings, etc. extracted from the medical records as intermediate nodes, connecting edges according to medical causal logic to generate a diagnosis support chain; using a large language model to generate language counterexamples for missing or broken nodes in the path, to assist doctors in identifying insufficient information; and combining the diagnosis-evidence mapping relationship in the medical knowledge graph to perform semantic rationality judgment on each reasoning path. Finally, the chain reasoning model CoT framework generates a diagnosis support score and outputs a structured quality control result; S362: The evaluation of treatment plan rationality includes the following steps: constructing a disease-treatment pathway map based on standard diagnosis and treatment guidelines, which includes diagnosis nodes, treatment action nodes, path sequence edges, and guideline constraint labels; using a graph structure matching algorithm to compare the treatment pathway extracted from the medical record with the standard pathway map to determine whether there are abnormal patterns such as "path jumps," "path reversals," "path departures," or "reverse implementation"; if an abnormality exists, further invoking the patient feature nodes in the knowledge graph, using a large language model to determine the personalized rationality of the deviation from the path, and providing natural language prompts; S363: The assessment of the integrity of key results and change records includes the following steps: First, uniformly extract the content related to numerical indicators in the medical record, align them across paragraphs based on semantic representation and position encoding technology, and normalize and match the descriptions of the same medical indicator at different time points; second, use time series modeling methods combined with clustering analysis algorithms to model the trend and construct the trajectory of indicator changes; if there are abnormal situations such as trajectory interruption, sudden change in numerical value, or sparse records, the logical break segments are automatically marked and natural language prompts are generated for reminders to improve the structural integrity of the medical record; S364: The assessment of the integrity of treatment measures includes the following steps: constructing a medical event chain map, which uses clinical symptoms and sudden illness as trigger event nodes and diagnosis and treatment operations as response nodes, and represents the response logic through edge connections; using a timestamp extraction method to mark the event occurrence time and response measure time, and based on the event-response time window model to determine whether there is a "delayed response", "missing response" or "response mismatch"; at the same time, calling the disease-treatment rule matching library in the knowledge graph to verify the semantic relevance between the response behavior and the triggering event, and generating a reasonable explanation text through a language model; S365: The evaluation of the rationality of causal relationships includes the following steps: constructing a complete causal path diagram in the medical record through entity relationship extraction technology, which connects the chief complaint, examination, diagnosis, treatment, and results into a diagram according to time and causal logic; applying a two-way causal reasoning mechanism based on a large language model, combined with the medical knowledge graph to verify the path consistency, and automatically identifying if "path interruption", "cause missing" or "causal error reversal" is found; after the abnormal causal path is determined by the system, the language model is called to generate structured prompt content and rewriting suggestions to improve the logical consistency and medical explanation ability of the causal chain.

8. The method according to claim 1, characterized in that The S4 unified quality control report output module includes: quantitative scoring, problem location and labeling according to five quality control indicators: adequacy of diagnostic basis, rationality of treatment, completeness of key result records, completeness of disposal measures and rationality of causal relationships, and output of a structured quality control result report.

Images