A method for electronic medical record quality control fact extraction and determination based on checking and backtracking

Abstract

This invention relates to the fields of medical information processing, natural language processing, and medical quality control, and provides a method for fact extraction and judgment in electronic medical record quality control based on verification and backtracking. It aims to solve problems existing in current electronic medical record quality control, such as missed fact extraction, insufficient evidence, poor end-to-end interpretability of judgments, and the easy propagation of errors to the rule-based reasoning stage. The method first structures and standardizes the quality control indicators and the original electronic medical record text. Second, it performs an initial full-scale fact extraction and generates feedback instructions through completeness and evidence sufficiency verification. Subsequently, it performs target-oriented backtracking extraction for missing or questionable facts, dynamically manages the results of multiple rounds of fact extraction, and finally outputs quality control conclusions, factual evidence chains, and reasoning chains based on logical rules. This invention is applicable to scenarios involving the automatic judgment, review, and verification of electronic medical record quality control indicators and medical quality management.

Description

Technical Field

[0001] This invention relates to the fields of medical information processing, natural language processing, and medical quality control technology, specifically to a method for extracting and judging electronic medical record quality control facts based on verification and backtracking, applicable to the extraction of quality control facts, evidence location, logical reasoning, and generation of quality control conclusions for electronic medical record texts. Background Technology

[0002] Medical quality is a core aspect of healthcare service management. Electronic medical records (EMRs), as the primary information carrier in the clinical diagnosis and treatment process, record a wealth of clinical information, including basic patient information, medical history, examination and test results, treatment measures, and disease progression. Conducting medical quality control based on EMRs provides crucial support for standardizing treatment practices, ensuring medical safety, verifying compliance with medical insurance policies, and evaluating healthcare management.

[0003] In existing technologies, the calculation and determination of electronic medical record quality control indicators typically rely on two main methods: one is manual review, where clinicians, medical record quality control personnel, or medical administrators read each medical record item by item and make determinations based on quality control rules; the other is automated processing based on rule engines or keyword templates, where engineers construct rules according to quality control standards and then perform matching calculations on structured or semi-structured data. The former suffers from low efficiency, reliance on experience for consistency, and high manual costs; the latter is less adaptable to differences in expression, contextual dependencies, information scattered across paragraphs, negative expressions, and implicit semantics in free text, and also has higher rule maintenance costs.

[0004] In recent years, large language models have demonstrated strong capabilities in text understanding, information extraction, and complex reasoning, enabling their application in electronic medical record quality control tasks. However, when quality control indicator definitions and original medical record texts are directly input into the model for end-to-end conclusion judgment, the following problems easily arise: First, in scenarios involving multiple conjunctive quality control indicators, the model may easily omit some factual items that should be verified; second, the model may make guesswork-based judgments based on experience when clear original text evidence is lacking; third, the model output often lacks a clear chain of evidence and reasoning, making it difficult to meet the traceability and verifiability requirements of medical quality control tasks; fourth, even if some current methods adopt a two-stage approach of "fact extraction - rule reasoning," they often treat fact extraction as a one-time process. Once there are omissions, errors, or insufficient evidence in fact extraction, subsequent rule reasoning will only propagate or even amplify the aforementioned errors, leading to a final misjudgment.

[0005] Therefore, there is an urgent need for an electronic medical record quality control fact extraction and judgment method that can perform fine-grained fact decomposition of quality control indicators for electronic medical record quality control tasks, and improve the completeness of fact extraction, sufficiency of evidence and reliability of final judgment through verification and backtracking mechanisms. Summary of the Invention

[0006] The purpose of this invention is to provide a method for extracting and judging facts for electronic medical record quality control based on verification and backtracking, which solves the problems of fact omission, insufficient evidence, unauditable end-to-end judgment, and the propagation of fact-level errors to the rule-level in existing electronic medical record quality control tasks, thereby improving the accuracy, interpretability, and verifiability of quality control conclusions.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for extracting and determining facts for electronic medical record quality control based on verification and backtracking, comprising the following steps:

[0008] Step 1: Input Organization and Standardization. Obtain the indicator descriptions and original electronic medical records corresponding to the quality control indicators to be judged. Perform structured parsing on the indicator descriptions to obtain a list of facts and logical rules. At the same time, clean, segment, and standardize the original electronic medical records to form input text suitable for subsequent evidence location and fact extraction.

[0009] Step 2: Initial Full Fact Extraction. Based on the aforementioned fact list, a first full scan of the electronic medical records is performed. A structured judgment result is given for each fact, and the corresponding evidence fragments and explanations are output to form an initial fact set.

[0010] Step 3: Completeness and Evidence Sufficiency Verification. The initial set of facts is verified to identify uncovered facts, facts with missing fields, facts with insufficient evidence, and facts where the evidence does not semantically match the facts, and corresponding feedback instructions are generated.

[0011] Step 4: Goal-oriented backtracking extraction. Based on the feedback instructions, a targeted secondary scan is performed only on the relevant paragraphs of the electronic medical record for the target facts that are missing, questionable, or lack sufficient evidence, to supplement evidence and update the factual status.

[0012] Step 5: Dynamic Fact Set Management. This involves unified management of initial fact results and backtracking update results, maintaining the final fact state table, and recording and processing fact conflicts, consistency issues, and update trajectories.

[0013] Step 6: Rule-based reasoning and conclusion generation. Substitute the final factual state table into the corresponding logical rules for merging and reasoning, outputting the final judgment result of the quality control indicators, as well as the factual evidence chain and rule-based reasoning chain corresponding to the judgment result.

[0014] 2. As a further aspect of the present invention: the specific steps of input organization and normalization in step 1 are as follows:

[0015] (1) Read the structured description information of the quality control indicator to be judged. The structured description information includes at least the indicator name, indicator definition, facts list and logical rules.

[0016] (2) Decompose the indicator definition into several fine-grained facts that can be independently determined, and denote the set of facts as:

[0017]

[0018] in, Indicates the first One fact to be determined.

[0019] (3) Construct a corresponding fact pattern template for each fact. The fact pattern template shall include at least: fact name, target semantics, allowed values, evidence requirements and suggested search clues.

[0020] (4) Preprocess the original electronic medical record text. The preprocessing includes at least text cleaning, standardization of synonyms, unification of inspection item names, preservation of paragraph boundaries and division of the smallest evidence unit, so that the subsequent model can complete the evidence location in the segmented context.

[0021] (5) The preprocessing results of electronic medical records and the fact pattern template are combined to form the input of the fact extraction module, thereby providing a unified data foundation for the subsequent initial full scan.

[0022] 3. As a further aspect of the present invention: the specific steps for the initial full-scale fact extraction in step 2 are as follows:

[0023] (1) Based on the fact pattern template and electronic medical record text, the large language model is called to perform a full scan of all facts for the first time, and a structured result is output for each fact.

[0024] (2) The output of each fact should include at least the fact judgment value and the original evidence, where the first fact is the first fact judgment value. The results of extracting these facts can be expressed as follows:

[0025]

[0026] in, Indicates the first The value of the judgment of the facts, This represents the original evidence fragment corresponding to the judgment value.

[0027] (3) The factual judgment value The three-valued state set is used for representation, namely:

[0028]

[0029] Among them, True indicates that there is clear evidence in the medical record to support the establishment of the fact; False indicates that there is clear evidence in the medical record to support the invalidity of the fact, or that the fact is not clearly established; Not Sure indicates that no clear evidence was found in the medical record to support the establishment or invalidity of the fact.

[0030] (4) The principle of “evidence first” is adopted in the initial full-scale fact extraction process, that is, the numerical values, examination items, examination completion records or clinical descriptions that appear clearly in the original medical records are extracted first; if there are only vague clues, approximate statements or speculative descriptions, they should not be directly judged as True, but should be output as Not Sure.

[0031] (5) Summarize the structured results corresponding to all facts to form an initial set of facts:

[0032]

[0033] 4. As a further aspect of the present invention: the specific steps for verifying the completeness and sufficiency of evidence in step 3 are as follows:

[0034] (1) Perform a completeness check on each fact in the initial fact set to determine whether there are any uncovered facts, facts with missing fields, or facts that have not returned valid evidence.

[0035] (2) Perform an evidence sufficiency check on each fact in the initial fact set to determine whether the evidence semantically matches the target fact and whether the following errors exist:

[0036] One approach is to replace target inspection with non-target inspection;

[0037] Secondly, vague mentions were mistakenly judged as completed.

[0038] Third, the test results were misinterpreted as indicating that the inspection items had been completed;

[0039] Fourth, it misquotes text fragments that are irrelevant to the current facts as evidence.

[0040] (3) To measure fact coverage, a fact coverage rate can be defined. for:

[0041]

[0042] in, For indicator functions, when the first A value of 1 is assigned if there is valid evidence for a fact; otherwise, a value of 0 is assigned.

[0043] (4) The verification module marks the problematic facts and generates a set of facts to be traced back:

[0044]

[0045] (5) Generate a feedback instruction based on the set of facts to be traced back. The feedback instruction includes at least the name of the facts to be traced back, the suggested search paragraph, the suggested search clues, and the requirements for the form of the target evidence.

[0046] (6) The verification module does not directly rewrite the initial fact set, but transmits the feedback instruction to the backtracking extraction stage to ensure that the fact correction process has explicit constraints and traceability.

[0047] 5. As a further aspect of the present invention: the specific steps of the target-oriented backtracking extraction in step 4 are as follows:

[0048] (1) Receive feedback instructions generated by the verification module and perform targeted secondary scanning only on the target facts in the set of facts to be traced back, without repeatedly processing non-target facts that have been clearly supported by evidence.

[0049] (2) Based on the search clues corresponding to each target fact, priority is given to finding evidence in the relevant medical record paragraphs, including but not limited to auxiliary examinations, test results, admission records, medical records, preoperative records and imaging reports.

[0050] (3) If new clear evidence is found during the backtracking process, the judgment value and evidence of the corresponding fact are updated; if no new clear evidence is found, the original state of the fact remains unchanged, or it remains Not Sure.

[0051] Let the first The set of facts after the cycle is traced back is:

[0052]

[0053] If the current iteration does not produce any factual updates compared to the previous iteration, then:

[0054]

[0055] when At that point, stop backtracking.

[0056] To avoid invalid loops, a maximum number of backtracking iterations can be preset. When satisfied The backtracking process ends at that point.

[0057] (6) Record the reasons for updating each fact during the backtracking extraction process, including the initial judgment, verification opinion, backtracking evidence and final update result, so as to construct a complete fact update trajectory.

[0058] 6. As a further aspect of the present invention: the specific steps of dynamic fact set management in step 5 are as follows:

[0059] (1) Summarize the initial full-scale fact extraction results and the backtracking extraction and update results to form the final fact status table.

[0060] (2) When multiple pieces of evidence appear in different paragraphs for the same fact, priority shall be determined according to the clarity of the evidence, semantic consistency with the target fact, and timeline.

[0061] (3) If there is conflicting evidence for the same fact and a decision cannot be made based on the clarity of the evidence or the order of time, the status of the fact shall be downgraded to Not Sure, while retaining the conflicting evidence to avoid being forcibly assigned a value during the rule reasoning stage.

[0062] (4) Represent the final state of each fact as follows:

[0063]

[0064] in, This represents the final judgment value. Indicates the evidence ultimately adopted. The update log indicates this fact.

[0065] (5) Form a set of final facts from all final facts:

[0066]

[0067] This final set of facts will serve as the sole source of input facts for the rule-based reasoning phase.

[0068] 7. As a further aspect of the present invention: the specific steps for rule reasoning and conclusion generation in step 6 are as follows:

[0069] (1) Read the final set of facts And the corresponding logical rule expressions, which logically map each fact state.

[0070] (2) When a conservative judgment strategy is adopted for quality control tasks, Not Sure is treated as a non-met requirement to ensure the security of the conclusion and the consistency of the audit.

[0071] For conjunction rules, if the rule expression is:

[0072]

[0073] The final conclusion is deemed compliant only if all facts meet the requirements; if any key fact is not met, the final conclusion is deemed non-compliant.

[0074] (4) For disjunctive rules, if the rule expression is:

[0075]

[0076] If any one of the facts meets the requirements, the final conclusion can be determined to be compliant; if none of the facts meet the requirements, the final conclusion can be determined to be non-compliant.

[0077] (5) For the general rule-based reasoning process, the final conclusion can be written as:

[0078]

[0079] in, This represents a rule merging function executed according to logical rules. This indicates the final quality control judgment result.

[0080] (6) While outputting the final judgment result, the judgment status of each fact, the corresponding evidence fragment, the logical expression expansion process and the rule operation path are output simultaneously to form a complete chain of evidence and reasoning.

[0081] (III) Beneficial Effects

[0082] Compared with the prior art, the present invention has the following beneficial effects:

[0083] First, this invention breaks down the electronic medical record quality control task into a phased process of "fact extraction - rule reasoning", separating the uncertainty of language understanding from the certainty of rule execution, and reducing the black box nature of end-to-end judgment.

[0084] Second, by introducing a verification and backtracking mechanism, this invention transforms fact extraction from a one-time generation process into a closed-loop process that is "verifiable, backtrackable, and updatable," which can significantly improve fact coverage and evidence sufficiency, and reduce fact omissions and misjudgments.

[0085] Third, this invention addresses the issues of multiple updates and fact conflicts through dynamic fact set management, making the factual state that finally enters the rule-based reasoning stage more stable, consistent, and controllable.

[0086] Fourth, this invention strictly requires that conclusions be based on locatable original evidence and can simultaneously output the evidence chain and reasoning chain, which facilitates medical quality control audits, result verification, and subsequent error analysis.

[0087] This invention has good adaptability to electronic medical record quality control scenarios with multiple conditions, multiple paragraphs, and diverse expressions. It is applicable to various medical quality control indicators such as the completion rate of inspection items, the standardization of medical record writing, and the compliance of diagnosis and treatment processes. Attached Figure Description

[0088] Figure 1 This is a flowchart illustrating the steps of a method for extracting and determining facts for electronic medical record quality control based on verification and backtracking, as described in this invention.

[0089] Figure 2 This is a flowchart of the overall system of the electronic medical record quality control fact extraction and judgment method based on verification and backtracking according to the present invention. Detailed Implementation

[0090] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0091] Example 1: Overall Flow and Implementation of the Method

[0092] Combination Figure 1 and Figure 2 The method in this embodiment includes the following steps:

[0093] Step 1: Constructing Input Data. Input data consists of two parts: structured information on the indicator side and the original electronic medical record text. The structured information on the indicator side includes at least: indicator name, indicator definition, indicator description, facts list, and logical rules. The original electronic medical record text is the patient's medical record text to be quality controlled, including but not limited to chief complaint, present illness, past medical history, auxiliary examinations, test results, medical orders, admission records, and discharge records. In this embodiment, the quality control indicators are first decomposed into facts, converting the original indicator definitions into executable fact pattern templates. Each fact pattern template includes a fact name, allowed values, evidence requirements, and key search clues. Simultaneously, basic preprocessing is performed on the original electronic medical record text, including text cleaning, standardization of special symbols, normalization of item names, paragraph segmentation, and division into minimum evidence units.

[0094] Step 2: Perform initial full fact extraction. Input the fact pattern template and medical record text generated in Step 1 into the fact extraction module, and the model will perform an initial scan of all facts. In this step, each fact should return the following structured fields:

[0095] (1) Fact name;

[0096] (2) Value, i.e., the value of factual judgment;

[0097] (3) Evidence, that is, the original evidence that supports the judgment;

[0098] (4) Reason, optional brief explanation of the attribution.

[0099] This step requires prioritizing the extraction of examination items, test results, or treatment descriptions explicitly mentioned in the medical record. For facts not explicitly mentioned in the original text, inferences based on experience are not allowed; instead, the output should be "Cannot be determined" or "No clear evidence found." For example, regarding the fact that "24-hour urine protein quantification was completed," if the original text contains "24-hour urine protein 2.6g" or a semantically equivalent expression, then this fact is considered true. If only "urine protein positive" appears but does not explicitly state 24-hour urine protein quantification, it cannot be directly substituted and should remain as "Cannot be determined."

[0100] Step 3: Perform verification. Input the initial fact results obtained in Step 2 into the verification module. The verification module mainly performs two types of tasks:

[0101] First, integrity verification. This checks whether there are any uncovered fact items or incomplete structured output fields.

[0102] Secondly, the sufficiency of evidence needs to be verified. This involves determining whether the presented evidence is sufficient to support the current factual determination and whether there are issues such as evidence mismatch, semantic substitution, or over-inference.

[0103] For example, if a fact requires "coagulation function test completed," but the medical record only shows "renal function-related biochemical indicators," then it cannot be considered that the coagulation function test has been completed; the verification module should identify this as insufficient evidence or a missing target test. When a missing or questionable item is found, the verification module generates a feedback instruction. The feedback instruction can take the form of: "Please specifically search for clear evidence regarding a certain fact in the auxiliary examination section, test results section, and preoperative record; if not found, please leave it as indeterminate."

[0104] Step 4: Perform retrospective extraction. Based on the feedback instructions output in Step 3, perform a targeted secondary scan only for missing or questionable facts. This step does not repeat the processing of facts already supported by sufficient evidence, but focuses on supplementing and correcting target facts. In this step, the search is prioritized based on the medical record paragraphs, search clues, and evidence format requirements specified in the feedback instructions, thereby reducing invalid scans and improving the efficiency of fact completion. For example, for the fact that "ultrasound examination (including the shape and size of both kidneys, ureters, and bladder) was completed," the retrospective search can prioritize paragraphs containing keywords such as "B-ultrasound," "color Doppler ultrasound," "urinary system ultrasound," "both kidneys," "bladder," and "ureter." If no corresponding examination is found, the fact is kept false or cannot be determined, and "no clear ultrasound examination evidence found" is recorded.

[0105] Step 5: Dynamically manage fact states. Since some facts may yield different results in the initial extraction and backtracking extraction, unified management of fact states across multiple rounds is necessary. In this step, the system maintains the final fact state table and records the following:

[0106] (1) Initial state of fact;

[0107] (2) Whether it is marked by the verification module;

[0108] (3) Feedback instructions;

[0109] (4) Whether the state is updated after backtracking;

[0110] (5) Final adoption status;

[0111] (6) The corresponding evidence and the source of the evidence paragraph.

[0112] If there is conflicting evidence for the same fact, priority is determined based on the clarity of the evidence, consistency of context, and chronological order. If it is still impossible to determine, it is downgraded to "cannot be determined" to avoid the propagation of errors to the rule-based reasoning stage.

[0113] Step 6: Execute logical rule reasoning. Input the final fact state table formed in Step 5 into the rule reasoning module and substitute it with the logical rule expression. The reasoning module supports AND, OR, and nested rules, and generates the final conclusion according to the preset quality control strategy. The output includes:

[0114] (1) Logic state cleaning results;

[0115] (2) The result of parsing the logical expression;

[0116] (3) Boolean operation process;

[0117] (4) Final quality control conclusion.

[0118] For conservative judgment strategies in quality control scenarios, "cannot be judged" can be treated as "does not meet requirements" to ensure the safety of the conclusions.

[0119] For example, when a certain conjunction rule is:

[0120]

[0121] If any one of these facts is false, the final conclusion is that it does not conform.

[0122] Example 2: Implementation of Prompt Words

[0123] In this embodiment, to ensure the uniformity and programmable parsing of the output structure of the fact extraction, verification, backtracking, and logical reasoning processes, corresponding prompt word templates are designed. The implementation method of the initial fact extraction prompt words corresponding to step 1 is as follows:

[0124] The input includes a Facts Schema template and electronic medical record text. The output should be in JSON format, with each fact returning two core fields: Value and Evidence. Hint constraints include:

[0125] (1) Explicit extraction takes precedence;

[0126] (2) Blindly extrapolating from inspections that are not clearly documented is prohibited;

[0127] (3) Implicit reasoning is limited to the reasonable induction of clinical conditions;

[0128] (4) The original text evidence must be returned.

[0129] A sample template could be expressed as: "This is a medical fact extraction task. Based on the given list of facts to be extracted, extract the corresponding information from the electronic medical record. For each fact, determine its value and provide the original textual evidence. If not explicitly mentioned, mark it as Not Sure. The output must be in JSON format, with each Fact containing a Value and Evidence."

[0130] The implementation method for the verification prompt words corresponding to step 2 is as follows:

[0131] The input includes: initial fact extraction results and metric definitions. The output includes: sufficiency status, a list of missing facts, and feedback instructions. Its main objective is:

[0132] (1) Identify missing facts;

[0133] (2) Identifying facts with insufficient evidence;

[0134] (3) Determine which facts have a key impact on the final conclusion;

[0135] (4) Generate search guidance instructions for subsequent backtracking.

[0136] A sample template could be: "You are a rigorous medical quality control officer. Please examine the following extracted list of facts, identify any missing or uncertain facts, and determine whether these missing items are absolutely necessary for the final conclusion. Output JSON, including is_sufficient, missing_keys, and feedback_instruction."

[0137] The implementation method for backtracking and extracting prompts in step 3 is as follows:

[0138] The input includes a list of missing facts, feedback instructions, and electronic medical record text. The output should only return the updated results for the target facts being traced. Its main constraints are:

[0139] (1) Only processed facts that have been marked;

[0140] (2) Prioritize finding evidence in the specified paragraphs;

[0141] (3) If no clear evidence is found, it shall not be forcibly judged as true;

[0142] (4) The output still uses the same JSON structure.

[0143] A sample template could be expressed as: "You are performing a backtracking extraction of medical record information. There are several missing or uncertain facts. The goal is to find the corresponding evidence based on the feedback instructions. Please carefully read the electronic medical record and try to find the missing information. Output JSON, containing only the updated facts."

[0144] The implementation method for the logical reasoning prompts corresponding to step 4 is as follows:

[0145] The input includes a final fact table and logical rules. The output includes the reasoning process and the final conclusion. This prompt emphasizes that the model strictly adheres to AND / OR Boolean logic and performs uniform state cleaning for "Not Sure". For example, it could require logical state cleaning before parsing the logical expression, followed by step-by-step Boolean operations, and finally outputting only "Yes" or "No" conclusions to form a final result suitable for medical quality control audits.

[0146] Example 3: Implementation Method for Determining Quality Control Indicators for IgA Nephropathy

[0147] The following describes the implementation process of the method of the present invention using a specific electronic medical record quality control case. The quality control indicator in this embodiment is: "IgA Nephropathy Quality Control Indicator - Completion Rate of Preoperative Examinations for Kidney Biopsy Patients". This indicator is defined as the proportion of kidney biopsy patients who complete all relevant preoperative examinations within 2 weeks. Required preoperative examinations include:

[0148] (1) Complete blood count and urinalysis;

[0149] (2) Liver and kidney function, coagulation function;

[0150] (3) Screening for infectious diseases, including hepatitis B, hepatitis C, syphilis, and HIV;

[0151] (4) Complement C3, Immunoglobulin IgA, Blood type;

[0152] (5) 24-hour urine protein quantification;

[0153] (6) Ultrasound examination, including the shape and size of both kidneys, ureters and bladder.

[0154] Based on this, the indicator is broken down into the following six facts: Fact 1: Complete blood and urine routine tests are completed; Fact 2: Liver and kidney function and coagulation function tests are completed; Fact 3: Infectious disease screening is completed; Fact 4: Complement C3, immunoglobulin IgA, and blood type tests are completed; Fact 5: 24-hour urine protein quantification is completed; Fact 6: Ultrasound examination is completed.

[0155] The corresponding electronic medical record text is as follows: "Auxiliary examinations: Urinalysis: Protein (+++), Red blood cells 265 / μL, 24-hour urine protein 2.6g, urine Bence Jones protein (-). Blood biochemistry: Albumin 40.2g / L, Serum creatinine 58.2μmol / L, Urea 4.78mmol / L, Uric acid 335μmol / L, Cystatin C 0.8mg / L, eGFR-EPI 114.9mL / min, Total cholesterol 6.29mmol / L, Triglycerides 1.2mmol / L, Blood glucose 5.38mmol / L, Glycated hemoglobin 4.8%." In this embodiment, the initial fact extraction in step 2 is performed first, and the following results are obtained:

[0156] Fact 1: True; Evidence: "Urinalysis: protein (+++), red blood cells 265 / μL; blood biochemistry test mentioned multiple blood indicators"; Fact 2: Cannot be determined; Evidence: "Blood biochemistry test mentioned multiple liver and kidney function indicators, but did not mention coagulation function"; Fact 3: False; Evidence: "No specific items for infectious disease screening were mentioned"; Fact 4: False; Evidence: "Complement C3, immunoglobulin IgA, and blood type test were not mentioned"; Fact 5: True; Evidence: "24-hour urine protein 2.6g"; Fact 6: False; Evidence: "No ultrasound examination was mentioned".

[0157] The verification process in step 3 is then executed. The verification module identifies insufficient evidence for Fact 2, namely, although renal function-related biochemical indicators are present in the medical record, no clear coagulation function test results are found. Therefore, a feedback instruction is generated: "Please specifically search for evidence regarding coagulation function tests in the medical record."

[0158] Next, step 4, backtracking, is performed. After a targeted second scan of the medical record text based on the feedback instructions, no new clear evidence regarding coagulation function tests was found; therefore, Fact 2 remains in an indeterminate state. The remaining fact states remain unchanged.

[0159] In step 5, the final fact status is managed uniformly, forming the final fact status table as follows: Fact 1: True; Fact 2: Cannot be determined; Fact 3: False; Fact 4: False; Fact 5: True; Fact 6: False.

[0160] Finally, perform the rule reasoning in step 6.

[0161] The overall logical rule corresponding to this indicator is the conjunction of six facts, which can be expressed as:

[0162]

[0163] Substituting the final fact into the above equation, we get:

[0164]

[0165] Therefore, the final quality control conclusion of this embodiment is: not all relevant preoperative examinations were completed, i.e., the judgment result is No.

[0166] This embodiment demonstrates that when the original electronic medical record only clearly shows urinalysis, 24-hour urine protein quantification, and some blood biochemical indicators, but no clear evidence of infectious disease screening, complement C3, immunoglobulin IgA, blood typing, coagulation function testing, or ultrasound examination is found, this invention can stably output evidence-based quality control conclusions through a closed-loop mechanism of "initial extraction—verification—backtracking—rule-based reasoning," while simultaneously preserving the factual state and reasoning process, thus avoiding the model from making affirmative conclusions directly based on experience in scenarios with insufficient evidence.

Claims

1. A method for extracting and determining facts for electronic medical record quality control based on verification and backtracking, characterized in that, Includes the following steps: S1. Obtain the indicator description and electronic medical record text corresponding to the quality control indicator to be judged, perform structured parsing on the indicator description to obtain the fact list and logical rules, and clean, segment and standardize the electronic medical record text to form input text for subsequent evidence location and fact extraction. S2. Based on the fact list, call the large language model to perform a full scan of the original electronic medical record for the first time, give a structured judgment result for each fact, and output the evidence fragments and explanations corresponding to each fact to form an initial fact result set; S3. Perform completeness and evidence sufficiency checks on the initial set of fact results, identify uncovered facts, facts with missing fields, facts with insufficient evidence, and facts whose evidence does not match the semantics of the facts, and generate corresponding feedback instructions; S4. Based on the feedback instructions, perform a targeted secondary scan in the relevant paragraphs of the original electronic medical record only for the target facts that are missing, questionable, or lack sufficient evidence, to supplement evidence and update the factual results. S5. The initial fact results and the backtracking update results are managed in a unified manner, the final fact status table is maintained, and fact conflicts, consistency and update trajectories are recorded and processed to form a final fact result set; S6. Substitute the final set of factual results into the logical rules for merging and reasoning, and output the final judgment result of the quality control indicators, as well as the factual evidence chain and rule reasoning chain corresponding to the final judgment result.

2. The method for extracting and determining facts for electronic medical record quality control based on verification and backtracking as described in claim 1, characterized in that, In step S1, the structured parsing of the indicator description includes: reading the indicator name, indicator definition, fact list, and logical rules; breaking down the indicator definition into several fine-grained facts that can be independently determined; constructing a fact pattern template for each fine-grained fact, wherein the fact pattern template includes at least the fact name, target semantics, allowed values, evidence requirements, and search clues; and the standardization processing of the original electronic medical record includes text cleaning, standardization of synonym expressions, unification of inspection item names, preservation of paragraph boundaries, and division of the smallest evidence unit.

3. The method for extracting and determining facts for electronic medical record quality control based on verification and backtracking as described in claim 1, characterized in that, In step S2, the structured judgment result of each fact includes at least a fact identifier, a fact judgment value, a fragment of original evidence, and an evidence description. The fact judgment value includes three states: true, false, and uncertain. True means that there is clear evidence in the medical record to support the fact being true; false means that there is clear evidence in the medical record to support the fact being false or clearly incomplete; and uncertain means that no clear evidence sufficient to support the fact being true or false has been found in the medical record.

4. The method for extracting and determining facts for electronic medical record quality control based on verification and backtracking as described in claim 3, characterized in that, Step S2 employs an evidence-first extraction constraint, prioritizing the extraction of explicitly stated numerical values, examination items, examination completion records, or clinical descriptions from the original electronic medical record as the basis for judgment. When only vague clues, approximate statements, or speculative descriptions exist, the result is not directly determined to be valid, but rather the uncertain state is output.

5. The method for extracting and determining facts for electronic medical record quality control based on verification and backtracking as described in claim 1, characterized in that, The integrity check in step S3 is used to determine whether there are uncovered facts, missing key output fields, or no valid evidence returned. The evidence sufficiency check in step S3 is used to determine whether the original evidence fragment matches the target fact semantically, and to identify at least one of the following situations: substituting a non-target check for a target check, misjudging a vague mention as completed, misjudging a description of the test result as completed, or incorrectly citing a text fragment unrelated to the current fact as evidence.

6. The method for extracting and determining facts for electronic medical record quality control based on verification and backtracking as described in claim 5, characterized in that, The feedback instruction includes at least the name of the fact to be traced, the set of facts to be traced, the suggested search paragraph, the suggested search clues, and the target evidence form requirements. The verification in step S3 does not directly rewrite the initial set of fact results, but drives step S4 to perform target-oriented backtracking extraction through the feedback instruction.

7. The method for extracting and determining facts for electronic medical record quality control based on verification and backtracking as described in claim 1, characterized in that, In step S4, evidence is preferentially searched in medical record paragraphs related to the target fact. These related paragraphs include at least one of auxiliary examinations, test results, admission records, progress notes, preoperative records, and imaging reports. When new, definitive evidence is found, the corresponding fact's judgment value and evidence are updated. When no new, definitive evidence is found, the fact remains unchanged or in an uncertain state. Let the first... The set of factual results after the cycle is If the current iteration does not produce any factual updates compared to the previous iteration, then the condition is satisfied. Alternatively, backtracking may terminate when the preset maximum number of backtracking rounds is reached.

8. The method for extracting and determining facts for electronic medical record quality control based on verification and backtracking as described in claim 1, characterized in that, In step S5, when multiple pieces of evidence appear in different paragraphs for the same fact, priority is determined according to the clarity of the evidence, semantic consistency with the target fact, and timeline. When there is conflicting evidence for the same fact and a determination cannot be made based on the clarity of the evidence or the timeline, the status of the fact is downgraded to an uncertain state, while retaining the conflicting evidence and the corresponding update log.

9. The method for extracting and determining facts for electronic medical record quality control based on verification and backtracking as described in claim 8, characterized in that, The final result of each fact formed in step S5 is represented as follows: ,in, This represents the final judgment value. Indicates the evidence ultimately adopted. This represents the update log of the fact; the final result of all facts constitutes the final fact result set. Furthermore, the final set of factual results is the sole source of factual input when performing rule-based reasoning in step S6.

10. The method for extracting and determining facts for electronic medical record quality control based on verification and backtracking as described in claim 1, characterized in that, The logical rules in step S6 include at least conjunction rules, disjunction rules, and nested combination rules. When a conservative decision-making strategy is adopted, uncertain states are treated as unmet requirements. When the logical rule is a conjunction rule and any key fact does not meet the requirements, the quality control indicator is determined to be non-compliant. When the logical rule is a conjunction rule and all key facts meet the requirements, the quality control indicator is determined to be compliant. When the logical rule is a disjunction rule and any fact meets the requirements, the quality control indicator is determined to be compliant. When the logical rule is a disjunction rule and all facts do not meet the requirements, the quality control indicator is determined to be non-compliant.

11. A method for extracting and determining facts for electronic medical record quality control based on verification and backtracking, as described in claim 1 or 10, characterized in that, The output of step S6 includes at least the judgment results of each fact and the corresponding evidence table, the logical expression expansion process, the rule operation path, and the final quality control conclusion, so as to form a verifiable chain of evidence and a chain of reasoning.

Images