Multi-stage electronic medical record generation method and system based on structured thinking chain

Abstract

The invention provides a multi-stage electronic medical record generation method and system based on a structured thinking chain, and relates to the technical field of computer data processing, and the method comprises the steps: generating at least one medical record draft intermediate state based on input original medical data through a structured thinking chain processing flow, the medical record draft intermediate state comprises at least one specific paper to be verified; starting a double-path knowledge calibration mechanism, and executing calibration on the intermediate state of the medical record draft, the calibration comprising: a forward knowledge enhancement path: in response to a specific disguise, retrieving associated evidence related to the specific disguise from one or more authoritative knowledge bases; and a backward fact checking path: utilizing a small reasoning model to carry out consistency judgment on the specific discourse and the associated evidence; and correcting the intermediate state of the medical record draft based on the result of the consistency judgment so as to generate a calibrated electronic medical record. According to the method, the fact accuracy and credibility are improved, and the structural integrity and logic self-consistency are ensured.

Description

Technical Field

[0001] This invention relates to the field of computer data processing technology, and particularly to medical information processing based on artificial intelligence. More specifically, it relates to a multi-stage electronic medical record generation method and system based on a structured thought chain. Background Technology

[0002] Currently, the creation of electronic health records (EHRs) is one of the core daily tasks of clinicians. Traditionally, doctors have completed these records by manually entering the data or using template-based systems. These methods are generally time-consuming, labor-intensive, and inefficient, consuming a significant amount of time that doctors could use for diagnosis, learning, and research.

[0003] In recent years, generative artificial intelligence technologies, represented by Large Language Models (LLM), have demonstrated enormous potential in natural language processing, text summarization, and content generation. Applying LLM to the automatic generation of doctor-patient dialogue summaries and clinical notes has become a research hotspot in this field, and is considered to have the potential to liberate doctors from tedious paperwork.

[0004] However, the application of large language models in the high-risk field of medicine faces an unresolved and fundamental technical obstacle: the hallucination problem.

[0005] First, the illusions of large language models manifest as factual errors. The model may generate information that sounds plausible but is actually false, completely fabricated, or unsupported by existing medical knowledge. For example, the model might fabricate a non-existent symptom, an incorrect lab result, or an outdated treatment plan in a medical record. In the medical field, such errors are unacceptable; they can directly lead to incorrect diagnoses, inappropriate treatments, and even delays in necessary care, posing a serious threat to patient safety.

[0006] Secondly, the output of large language models also suffers from structural deficiencies. Existing generative models primarily rely on statistical correlation rather than rigorous causal reasoning. This results in clinical notes that often lack coverage of the complex entity relationships and subtleties in real-world clinical scenarios. Specifically, the model may generate content that is structurally incomplete or logically inconsistent. For example, the model might arrive at a conclusion in the diagnosis section of a medical record but fail to provide (or omit) the symptoms or examination results necessary to support that diagnosis in the present illness or examination sections, or even produce contradictory descriptions between different parts of the medical record.

[0007] Furthermore, medical knowledge itself is complex, dynamic, and potentially conflicting. For example, new clinical guidelines may contradict older ones. Existing LLM systems generally lack a clear, controlled mechanism to detect and intelligently adjudicate these knowledge conflicts from different sources, resulting in unpredictable and unreliable outputs when faced with inconsistent evidence.

[0008] Therefore, a new technical architecture is urgently needed in this field. This architecture can no longer rely on the output of large language models, but must instead treat them as creative but unreliable draft generators. The architecture must include an external calibration and correction system to constrain and correct the output of the large language model. Specifically, this system must be able to guarantee the integrity and logical consistency of medical record data at the structural level (i.e., early in generation) (to address problem two); and proactively verify key arguments at the factual level (i.e., later in generation), and intelligently adjudicate in cases of conflicting evidence (to address problems one and three). Summary of the Invention

[0009] The purpose of this invention is to provide a multi-stage electronic medical record generation method and system based on a structured thinking chain to solve the problems pointed out in the background art.

[0010] In a first aspect, the multi-stage electronic medical record generation method based on a structured thought chain provided by embodiments of the present invention includes:

[0011] (a) Based on the input raw medical data, at least one intermediate state of medical record draft is generated through a structured thinking chain processing flow, wherein the intermediate state of medical record draft contains at least one specific statement to be verified;

[0012] (b) Initiate a dual-path knowledge calibration mechanism to perform calibration on the intermediate state of the medical record draft, the calibration including:

[0013] (b1) Forward knowledge enhancement path: In response to the specific assertion, retrieve relevant evidence related to the specific assertion from one or more authoritative knowledge bases; and

[0014] (b2) Backward fact-checking path: Using a small reasoning model, the consistency between the specific assertion and the related evidence is judged;

[0015] (c) Based on the result of the consistency judgment, the intermediate state of the medical record draft is corrected to generate a calibrated electronic medical record.

[0016] Optionally, before generating the intermediate state of the medical record draft in step (a), the method further includes:

[0017] A step-by-step thinking chain prompting strategy is adopted to guide the large language model to gradually extract information from the original medical data according to the preset medical paradigm in order to generate structured evidence intermediate state;

[0018] The preset medical paradigm includes the "one-story, five-history" medical paradigm.

[0019] Optionally, step (a) further includes:

[0020] The structured evidence intermediate state is logically matched and filled with a pre-set medical record template;

[0021] In the matching and filling process, the principle of not adding information is strictly enforced, and the introduction of new information other than the structured evidence intermediate state is prohibited in order to form a logical framework intermediate state.

[0022] Optionally, generating the intermediate state of the medical record draft in step (a) includes:

[0023] The logical framework intermediate state is combined with contextual data obtained from the patient's past medical records or laboratory systems to guide the large language model to convert the logical framework intermediate state into a natural language description, thereby forming the medical record draft intermediate state.

[0024] Optionally, the consistency determination in step (b2) specifically includes:

[0025] The small reasoning model performs a thought chain truncated verification, which only performs a binary consistency judgment on the specific assertion and the related evidence to determine whether the specific assertion is supported by the related evidence.

[0026] Optionally, the authoritative knowledge base in step (b1) includes:

[0027] The evidence-based medicine literature submodule is used for real-time retrieval of clinical guidelines and high-quality literature; and

[0028] The medical knowledge graph submodule is used to provide structured associations between diseases, symptoms, examinations, and treatments.

[0029] Optionally, the method further includes:

[0030] (d) Presenting the calibrated electronic medical record generated in step (c) to authorized users via the human-machine interface; and

[0031] (e) Visualize the key intermediate states generated by the structured thought chain in steps (a) and (b), and / or highlight the specific assertions that fail to be consistent and are corrected in step (b2).

[0032] Optionally, in step (a), before strictly implementing the no-information-addition principle during the matching and filling process, the following is also included:

[0033] Define a set of medical integrity constraints, which are encoded into a formal specification language and used to describe the structural and association rules that the intermediate state of the structured evidence must satisfy. The structural and association rules include at least entity integrity constraints and referential integrity constraints.

[0034] A formal verification engine is launched, and the structured evidence intermediate state is used as a model to be tested, so as to use the formal verification engine to mathematically prove whether the structured evidence intermediate state satisfies the medical integrity constraint.

[0035] In response to the formal verification engine outputting a verification failure result, wherein the verification failure result indicates that the structured evidence intermediate state violates the medical integrity constraint due to information omission or logical breakage, a correction protocol is activated.

[0036] The correction protocol is authorized to exempt the structured evidence intermediate state from the information non-increase principle when correcting the structured evidence intermediate state, so as to ensure that the structural integrity of the structured evidence intermediate state has been achieved before proceeding to the knowledge calibration in step (b).

[0037] Optionally, the consistency judgment in step (b2) further includes: in response to the small inference model detecting a conflict state in step (b2) where the specific assertion is simultaneously supported by at least one piece of related evidence and refuted by at least another piece of related evidence, suspending the binary consistency judgment and activating a revocable adjudication engine; the revocable adjudication engine is configured to access a set of dominant meta-rules for resolving the conflict state, and based on the dominant meta-rules, adjudicate a winning piece of evidence and one or more rejected pieces of evidence from the related evidence of the conflict;

[0038] The one or more rejected pieces of evidence are retained nondestructively as an immutable annotation associated with the calibrated electronic medical record;

[0039] And generate ruling source metadata, which records the advantage meta-rule on which the revocable ruling engine made its ruling, and bind the ruling source metadata to the immutable annotation; wherein the correction in step (c) is based on the winning evidence, while retaining the link to the immutable annotation.

[0040] Secondly, an embodiment of the present invention provides a multi-stage electronic medical record generation system based on a structured thought chain, comprising:

[0041] One or more processors; and

[0042] A memory storing computer-executable instructions that, when executed by the one or more processors, implement the method as described in the first aspect;

[0043] The system specifically includes:

[0044] (i) A chain-based reasoning and control module, adapted to perform step (a) to generate the intermediate state of the medical record draft;

[0045] (ii) A knowledge service module, adapted to perform step (b1) to retrieve the relevant evidence from the authoritative knowledge base for the forward knowledge enhancement path;

[0046] (iii) The chain reasoning and control module further includes a small reasoning model adapted to perform the (b2) step to perform the consistency judgment in the backward fact-checking path;

[0047] The chain reasoning and control module is also adapted to perform step (c) to correct the intermediate state of the medical record draft based on the result of the consistency judgment, so as to generate the calibrated electronic medical record.

[0048] The present invention has achieved the following beneficial effects:

[0049] Improving factual accuracy and credibility: This invention establishes a dual-path knowledge calibration mechanism, specifically a closed loop comprising forward knowledge enhancement and backward fact verification, which proactively verifies potential illusions or factual errors that may arise when a large language model generates draft medical records. By utilizing a small inference model and an authoritative knowledge base for consistency judgment, the completeness and factual accuracy of the final generated electronic medical record are improved, addressing the core security risks of LLM in medical applications identified in the background section.

[0050] Ensuring structural integrity and logical consistency: This invention introduces formal verification into the electronic medical record generation process. By encoding medical integrity constraints (such as entity integrity and referential integrity) into a formal specification language and using a verification engine to perform mathematical proofs early in the generation process (intermediate state of structured evidence), structural defects caused by information omissions or logical breaks are fundamentally eliminated, ensuring the inherent logical consistency of medical record data.

[0051] Achieving Advanced Conflict Resolution and Transparency: Addressing potential evidentiary conflicts between authoritative knowledge bases, this invention provides a revocable adjudication engine. This engine transcends simple binary judgments, intelligently adjudicating conflicting evidence based on pre-defined dominant meta-rules (such as the timeliness and authority of evidence). Simultaneously, it non-destructively preserves rejected evidence and records the source of the adjudication. This not only resolves knowledge conflict issues in complex medical scenarios but also enhances the system's transparency, auditability, and human-machine collaboration capabilities.

[0052] Balancing generation efficiency and process controllability: This invention employs a multi-stage process based on a structured thinking chain, breaking down the complex medical record generation task into several controllable stages, including information extraction, logic construction, language narration, and dual-path calibration. Through mechanisms such as the principle of non-incremental information and truncated verification of the thinking chain, the system's computational efficiency and process controllability are improved while ensuring generation quality, distinguishing it from the uncontrollable, black-box end-to-end generation models in the background technology.

[0053] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0054] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0055] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0056] Figure 1 This is a schematic diagram of a multi-stage electronic medical record generation method based on a structured thinking chain in an embodiment of the present invention;

[0057] Figure 2 This is a schematic diagram of a multi-stage electronic medical record generation system based on a structured thinking chain, as described in an embodiment of the present invention. Detailed Implementation

[0058] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0059] Figure 1 A flowchart of a multi-stage electronic medical record generation method based on a structured thought chain is provided for embodiments of this application, such as... Figure 1 As shown, the method includes:

[0060] (a) Based on the input raw medical data, at least one intermediate state of medical record draft is generated through a structured thinking chain processing flow, wherein the intermediate state of medical record draft contains at least one specific statement to be verified;

[0061] (b) Initiate a dual-path knowledge calibration mechanism to perform calibration on the intermediate state of the medical record draft, the calibration including:

[0062] (b1) Forward knowledge enhancement path: In response to the specific assertion, retrieve relevant evidence related to the specific assertion from one or more authoritative knowledge bases; and

[0063] (b2) Backward fact-checking path: Using a small reasoning model, the consistency between the specific assertion and the related evidence is judged;

[0064] (c) Based on the result of the consistency judgment, the intermediate state of the medical record draft is corrected to generate a calibrated electronic medical record.

[0065] Please see Figure 1 This embodiment provides a core process for a multi-stage electronic medical record generation method based on a structured thinking chain. This process consists of... Figure 2 The system shown (e.g., deployed on a hospital's private cloud server) executes as follows: First, in step (a), the system receives raw medical data as input, such as audio-transcribed text from doctor-patient conversations or unstructured text from physical examination reports, and drives a large language model (as...) through a structured thought chain processing flow. Figure 2 As part of the chain-like reasoning and control module (i), a series of controlled reasoning steps are executed to generate at least one intermediate state of a medical record draft. This intermediate state is a natural language narrative draft, but internally (e.g., through metadata tagging) contains at least one specific assertion to be verified (e.g., a preliminary diagnosis or a description of a key physical sign). Then, the system automatically initiates a two-path knowledge calibration mechanism in step (b) to calibrate the intermediate state of the medical record draft. Figure 2The process is completed collaboratively by the knowledge service module (ii) and the small inference model (iii) in the chain reasoning and control module (i), and specifically includes two concurrent or sequential execution paths: (b1) a forward knowledge enhancement path, in which the knowledge service module (ii) actively retrieves relevant evidence related to the specific assertion from one or more authoritative knowledge bases (e.g., an internal medical knowledge graph or an external evidence-based medicine literature base) in response to the specific assertion mentioned above; and (b2) a backward fact-checking path, executed by the small inference model (iii) (whose computational cost is much lower than that of the large language model that generates the draft), which uses the small inference model to make one or more independent consistency judgments between the specific assertion generated in step (a) and the relevant evidence retrieved in step (b1); finally, in step (c), the chain reasoning and control module (i) collects the results of the consistency judgments generated in step (b2) (e.g., support, contradiction, or no relevance), and based on the judgment results, corrects the intermediate state of the medical record draft through preset correction logic (e.g., retention, deletion, or replacement) to generate an electronic medical record that is calibrated at the fact level.

[0066] In one embodiment, before generating the intermediate state of the medical record draft in step (a), the method further includes:

[0067] A step-by-step thinking chain prompting strategy is adopted to guide the large language model to gradually extract information from the original medical data according to the preset medical paradigm in order to generate structured evidence intermediate state;

[0068] The preset medical paradigm includes the "one-story, five-history" medical paradigm.

[0069] This embodiment further refines the early stage of step (a) in Embodiment 1. Specifically, before generating the intermediate state of the medical record draft, the system also includes a key preprocessing step, namely, employing a step-by-step thought chain prompting strategy, which is configured in... Figure 2In the chain reasoning and control module (i), the large language model is guided to extract information from the original medical data step by step and in a controlled manner according to a preset medical paradigm. This strategy does not require the large language model to generate all content at once, but rather decomposes the complex extraction task into multiple simple and independent sub-tasks through a series of carefully designed and sequential prompts, so as to ensure that the large language model focuses on specific information fragments at each step, thereby suppressing its hallucination tendency to the greatest extent. The preset medical paradigm reflects medical expertise. In this embodiment, it includes the "one-statement-five-history" medical paradigm, that is, the chain reasoning and control module (i) will sequentially prompt the large language model to extract the chief complaint (one statement), present illness history, past medical history, personal history, family history, and marital / menstrual history (five histories) from the original medical data (such as outpatient dialogue). The output generated by this structured and step-by-step extraction process is the intermediate state of structured evidence.

[0070] In one embodiment, step (a) further includes:

[0071] The structured evidence intermediate state is logically matched and filled with a pre-set medical record template;

[0072] In the matching and filling process, the principle of not adding information is strictly enforced, and the introduction of new information other than the structured evidence intermediate state is prohibited in order to form a logical framework intermediate state.

[0073] This embodiment further refines the intermediate stage of step (a) in embodiment one. Specifically, after generating the intermediate state of structured evidence in embodiment two (e.g., a set of key-value pairs containing fields such as chief complaint and present medical history), step (a) further includes: [The text abruptly ends here, so the translation stops as well.] Figure 2The chain reasoning and control module (i) executes to perform logical structure matching and filling of the structured evidence intermediate state with one or more pre-defined medical record templates (e.g., predefined admission record templates or SOAP note templates in a hospital information system (HIS)). This involves filling data items (e.g., fever for three days) from the structured evidence intermediate state into the corresponding fields (e.g., present medical history field) of the medical record template. In particular, during the matching and filling process, the system strictly adheres to a principle of no information addition. This principle is encoded as a deterministic verification logic in the chain reasoning and control module (i). This verification logic works by tracing data provenance. After filling is completed, the system reverse-verifies each piece of information filled into the template to ensure that it can find a unique and corresponding source in the original structured evidence intermediate state. This principle is configured to strictly prohibit (e.g., automatic deletion or marking) the introduction of any new information other than the structured evidence intermediate state during this process (e.g., symptoms or descriptions creatively added by the large language model during filling that do not exist in the original data). In this way, the system forms a logically closed logical framework intermediate state that contains only the original extracted information.

[0074] In one embodiment, generating the intermediate state of the medical record draft in step (a) includes:

[0075] The logical framework intermediate state is combined with contextual data obtained from the patient's past medical records or laboratory systems to guide the large language model to convert the logical framework intermediate state into a natural language description, thereby forming the medical record draft intermediate state.

[0076] This embodiment further refines the final stage of step (a) in Embodiment 1, namely the specific process of generating the intermediate state of the medical record draft. This process occurs after the intermediate state of the logical framework (i.e., a template filled with structured data) is generated in Embodiment 3, and includes: Figure 2 The chain reasoning and control module (i) merges the logical framework intermediate state with contextual data (such as the patient's allergy history, long-archived diagnoses, and recent test results) obtained in real time from the patient's past medical record database or laboratory information system (LIS) at the data level. Then, the chain reasoning and control module (i) takes this merged, information-rich structured data (rather than the original medical data) as input and guides the large language model (or a downstream model specifically for natural language stylization) again, instructing it to convert this logical framework intermediate state into a fluent natural language description that conforms to medical terminology. Since the task of the large language model at this time is no longer to extract or create facts, but only to polish and narrate a structured and constrained input, its illusion space is greatly compressed, thus forming a medical record draft intermediate state (i.e., a natural language draft that can be used for fact calibration in subsequent steps (b)).

[0077] In one embodiment, the consistency determination in step (b2) specifically includes:

[0078] The small reasoning model performs a thought chain truncated verification, which only performs a binary consistency judgment on the specific assertion and the related evidence to determine whether the specific assertion is supported by the related evidence.

[0079] The small inference model in this embodiment refers to a lightweight model that is functionally strictly limited and highly optimized for computational resources. Its design goals differ significantly from the large language model (LLM) used for content generation; it does not perform open-ended text generation but instead focuses on truncated verification of the thought chain in the backward fact-checking path. Its specific technical features and implementation methods may include:

[0080] The model's reasoning task is strictly truncated. It does not need to understand the full context of the medical record, nor does it perform complex multi-step causal reasoning. Its sole task is to receive a specific assertion (e.g., a diagnosis) and a piece of relevant evidence (e.g., a lab result or guideline provision), and efficiently perform consistent judgments in binary or ternary (e.g., supportive, contradictory, irrelevant) arguments.

[0081] In one specific embodiment, the model can be a Transformer-based model (such as BERT) that has been fine-tuned for Natural Language Inference (NLI) tasks. For example, a model pre-trained on a medical corpus (such as BioBERT or PubMedBERT) can be used as a base, and then supervised fine-tuned on a dedicated medical dataset containing (argument, evidence) → (support / contradiction) labels to focus on factual consistency classification.

[0082] Because its model has a much smaller number of parameters (e.g., likely in the range of 100 million to 1 billion, rather than the hundreds of billions of parameters in an LLM) than the large language models used to generate content, its computational cost (training and inference costs) is extremely low. This allows the system to perform fact-checking on a large number of specific arguments in medical record drafts in parallel and at scale, thus avoiding the inefficiencies and recursion illusions that can arise when using expensive large language models for self-verification.

[0083] In one embodiment, the authoritative knowledge base in step (b1) includes:

[0084] The evidence-based medicine literature submodule is used for real-time retrieval of clinical guidelines and high-quality literature; and

[0085] The medical knowledge graph submodule is used to provide structured associations between diseases, symptoms, examinations, and treatments.

[0086] This embodiment further illustrates the specific composition of the authoritative knowledge base in step (b1) of Embodiment 1. This knowledge base consists of... Figure 2 The knowledge service module (ii) is responsible for management, maintenance, and querying. It is not a single database, but a heterogeneous, multimodal knowledge system. In this embodiment, it includes at least two core sub-modules: (1) an evidence-based medicine literature sub-module, which connects in real time to external, recognized medical information sources, such as PubMed, the Cochrane Library, UpToDate Clinical Advisors, and the latest clinical guidelines and high-quality literature issued by the National Health Commission, through an application programming interface (API). The function of this sub-module is to provide the latest evidence-based (EBM) related evidence for the forward knowledge enhancement path; and ( 2) Medical Knowledge Graph Submodule: This submodule is an internally built and maintained structured knowledge base (e.g., implemented using graph database technologies such as Neo4j or JanusGraph). It stores a large number of medical entities (such as diseases, symptoms, drugs, and examination items) and their structured relationships (e.g., aspirin - [used for treatment] → coronary heart disease, or metformin - [contraindications exist] → renal insufficiency). The function of this submodule is to provide internal, logical evidence of association for the backward fact-checking path, used to verify the logical consistency of internal arguments in the intermediate state of the medical record draft (e.g., verifying whether there are contraindications between diagnosis and medication).

[0087] In one embodiment, the method further includes:

[0088] (d) Presenting the calibrated electronic medical record generated in step (c) to authorized users via the human-machine interface; and

[0089] (e) Visualize the key intermediate states generated by the structured thought chain in steps (a) and (b), and / or highlight the specific assertions that fail to be consistent and are corrected in step (b2).

[0090] This embodiment describes the human-computer interaction and result presentation stage of the present invention, which occurs after step (c) of Embodiment 1 (i.e., generating the calibrated electronic medical record). The method further includes: (d) presenting the calibrated electronic medical record generated in step (c) to an authorized user (e.g., an attending physician or medical records reviewer) on a human-computer collaboration interface (e.g., a dedicated web front-end interface integrated into the EHR system); and (e) in order to achieve transparency and explainability (ExplainableAI, XAI), the interface is configured to (1) visually display the structured thought chain generated in steps (a) and (b). Key intermediate states, for example, allowing users to click a button to view the structured evidence intermediate state extracted from the original dialogue (from Example 2), or to view the filled logical framework intermediate state (from Example 3); and / or (2) highlighting specific assertions that failed the consistency judgment in step (b2) (i.e., were judged as contradictory or unsupported) and were therefore corrected in step (c). For example, the system might display the illusion content initially generated by the large language model with a red strikethrough and highlight the correct content after the relevance evidence correction based on step (b1) with a green highlight, while providing a clickable icon so that the user can trace back to the relevance evidence itself on which the correction is based.

[0091] In one embodiment, step (a) further includes, before strictly adhering to the no-information-addition principle during the matching and filling process:

[0092] Define a set of medical integrity constraints, which are encoded into a formal specification language and used to describe the structural and association rules that the intermediate state of the structured evidence must satisfy. The structural and association rules include at least entity integrity constraints and referential integrity constraints.

[0093] A formal verification engine is launched, and the structured evidence intermediate state is used as a model to be tested, so as to use the formal verification engine to mathematically prove whether the structured evidence intermediate state satisfies the medical integrity constraint.

[0094] In response to the formal verification engine outputting a verification failure result, wherein the verification failure result indicates that the structured evidence intermediate state violates the medical integrity constraint due to information omission or logical breakage, a correction protocol is activated.

[0095] The correction protocol is authorized to exempt the structured evidence intermediate state from the information non-increase principle when correcting the structured evidence intermediate state, so as to ensure that the structural integrity of the structured evidence intermediate state has been achieved before proceeding to the knowledge calibration in step (b).

[0096] This embodiment details a crucial structural integrity assurance mechanism implemented in step (a) of Embodiment 1, prior to the strict adherence to the no-information-addition principle during the matching and filling process of Embodiment 3. The purpose of this mechanism is to immediately perform a rigorous mathematical check on the structured evidence intermediate state early in step (a), immediately after it has been extracted from the raw medical data (as in Embodiment 2), to ensure its structural integrity and self-consistency, thereby preventing data contamination from missing information or logical breaks in subsequent generation and calibration processes. This mechanism is based on a core insight (see Background Art): large language models not only produce factual errors (addressed by Embodiment 9) but also structural errors, such as generating data lacking complex entity relationships or logically incoherent data. This mechanism is designed to address this structural problem, and its complete working principle comprises the following sequential steps:

[0097] First, the system's knowledge engineers and medical expert team predefine a set of medical integrity constraints. These constraints are hard rules that must be met to describe a structurally qualified intermediate state of evidence. These constraints are not related to specific facts (such as a patient's blood glucose level), but only to the structure and relationships of the data. Based on data integrity theory, these constraints are concretized in this embodiment into at least two key database theory constraints: The first is entity integrity constraints, which require that each key medical entity (e.g., a diagnosis entry, a drug entry) must have a unique, non-null identifier (e.g., a primary key). This ensures the referenceability and uniqueness of data, preventing the large language model from generating ambiguous, duplicate, or unreferenceable data entities. The second is referential integrity constraints, which are the core of this mechanism. They enforce that the referencing relationships between entities must be valid. For example, it stipulates that the diagnostic basis referenced by each medication record (as a sub-table) in the structured evidence intermediate state (e.g., a diagnosis ID as a foreign key) must point to a diagnostic record that actually exists in that intermediate state (the primary key in the parent table). Similarly, the symptom or examination result basis referenced by each diagnostic record must also point to a symptom or examination result record that actually exists in that intermediate state. Mathematically, this constraint guarantees that there will be no logical breaks between data, for example, preventing the large language model from generating a diagnosis without a source (i.e., a diagnostic conclusion without any symptom or examination support).

[0098] Secondly, in order for these constraints to be strictly enforced by a computer, they must be encoded into a formal specification language. A formal specification language is a precise symbolic system based on mathematics (such as set theory and first-order logic), whose core value lies in eliminating the ambiguity of natural language rules. In this embodiment, the system employs a specification language based on set theory and predicate logic (e.g., conceptually similar to TLA+ or Znotation). For example, the aforementioned referential integrity constraint can be precisely expressed as a state predicate, which in TLA+ is called an invariant. This predicate declares: for any medication entity in the set of all medication entities in the structured evidence intermediate state, the value of its diagnostic reference ID attribute must be a member of the set of all diagnostic ID attribute values ​​in the diagnostic entity set. In this way, all medical integrity constraints are translated into a set of mathematical assertions that can be parsed and proven by the machine. For example, in a concrete implementation of a formal specification language, a rule used to verify referential integrity constraints (i.e., ensuring that each medication record has a valid diagnostic basis) can be encoded as the following predicate logic pseudocode:

[0099] / / --- Formal specification language example: Constraint R-01 ---

[0100] / / Constraint ID: RC-001 (Referential integrity: Diagnosis - Medication)

[0101] / / Rule description: Each medication record in the structured evidence intermediate state,

[0102] / / The diagnostic ID it references (Diagnosis_Ref_ID),

[0103] / / Must exist in the primary key ID of the Diagnoses collection.

[0104] / / 1. Define State Variables

[0105] Context: StructuredEvidence / / Model to be tested

[0106] / / 2. Define the set of valid diagnostic IDs

[0107] LET Diagnoses_IDs_Set = { d.ID | d IN StructuredEvidence.Diagnoses}

[0108] / / 3. Define Invariants / Mathematical Assertion

[0109] / / The formal verification engine will prove whether this assertion is true.

[0110] ASSERT FOR ALL m IN StructuredEvidence.Medications:

[0111] / / Check 1: References cannot be null (a reflection of entity integrity)

[0112] (m.Diagnosis_Ref_ID != NULL)

[0113] AND

[0114] / / Check 2: References must be valid (a reflection of referential integrity)

[0115] (m.Diagnosis_Ref_ID IN Diagnoses_IDs_Set)

[0116] Next, during the execution of step (a), once the large language model of Example 2 generates a structured evidence intermediate state (e.g., a JSON or XML object logically representing a state), the system immediately initiates a formal verification engine. This engine is a software tool that functions similarly to a model checker. The system inputs the newly generated structured evidence intermediate state as the model to be tested into the engine, along with the encoded formal reduction (i.e., the set of mathematical constraints mentioned above) as canonical input. The core function of the formal verification engine is not testing, but proof. It exhaustively explores all data states and relationships represented by the model to be tested (e.g., through reachability analysis or state space search), attempting to mathematically prove whether the model satisfies the set of medical integrity constraints in all cases. The system then responds to one of two possible outcomes. If the proof is successful (i.e., no violations are found), indicating that the structured evidence intermediate state is structurally complete and self-consistent, the process continues to the matching and filling step of Example 3. But more importantly, the system responds to the formal verification engine by outputting a verification failure result. A validation failure result means the engine has found a counterexample, namely a specific data state or data item, that violates at least one medical integrity constraint. This validation failure result precisely indicates that the structured evidence intermediate state violates the medical integrity constraint due to missing information or logical breaks. For example, the engine reports: Referential integrity constraint R-01 failed: Diagnosis reference ID Diag-101 was found at the medication entity Med-007, but Diag-101 was not found in the diagnosis entity set. Once this validation failure result is received, the system activates a correction protocol. This correction protocol is an automated, rule-based repair protocol, similar in concept to Automated Program Repair in software engineering. Finally, the correction protocol is executed, its core feature being that it is authorized by the system to exempt the structured evidence intermediate state from the principle of no information addition. This exemption is logically necessary: ​​because the validation failure is precisely due to missing information (e.g., the large language model omits the diagnosis Diag-101), the only way to repair it is to add or correct the missing information. The correction protocol will automatically perform preset repair actions based on the defects indicated by the counterexamples, such as: (1) prompting the large language model again and instructing it to specifically search for the diagnosis Diag-101 related to Med-007 in the original medical data and add it to the diagnosis entity set; or (2) if it cannot be found, marking the diagnosis reference ID of Med-007 as invalid or pending verification.After performing the repair action, the protocol resubmits the corrected structured evidence intermediate state to the formal verification engine, repeating this proof-failure-repair cycle until the formal verification engine outputs a successful verification result. Through this series of steps (defining constraints, encoding, starting the engine, proving, handling failures, activating the repair protocol, exempting and repairing), the present invention ensures that the structural integrity of the structured evidence intermediate state is achieved before proceeding to the knowledge calibration in step (b). This mechanism fundamentally guarantees that the data operated on in subsequent fact-checking (Example 9) is structurally reliable and self-consistent, thus forming the first key line of defense for the security of the present invention.

[0117] In one embodiment, the consistency judgment in step (b2) further includes: in response to the small inference model detecting a conflict state in step (b2) where the specific assertion is simultaneously supported by at least one piece of related evidence and refuted by at least another piece of related evidence, suspending the binary consistency judgment and activating a revocable adjudication engine; the revocable adjudication engine is configured to access a set of dominant meta-rules for resolving the conflict state, and based on the dominant meta-rules, adjudicate a winning piece of evidence and one or more rejected pieces of evidence from the related evidence of the conflict;

[0118] The one or more rejected pieces of evidence are retained nondestructively as an immutable annotation associated with the calibrated electronic medical record;

[0119] And generate ruling source metadata, which records the advantage meta-rule on which the revocable ruling engine made its ruling, and bind the ruling source metadata to the immutable annotation; wherein the correction in step (c) is based on the winning evidence, while retaining the link to the immutable annotation.

[0120] This embodiment details an advanced conflict resolution mechanism activated during consistency judgment performed by the small inference model of Embodiment 5 in step (b2) of Embodiment 1. The mechanism is activated when the small inference model detects a conflict state during its thought chain truncation verification. A conflict state is defined as a complex scenario where a specific assertion (e.g., a recommendation in a draft medical record to use drug A) is simultaneously supported by at least one piece of relevant evidence (e.g., guideline G1 in the knowledge base supports the use of drug A) and refuted by at least another piece of relevant evidence (e.g., the latest study L1 or guideline G2 in the knowledge base refutes or prohibits the use of drug A). This situation is very common in the medical field because medical knowledge itself is constantly evolving and may be controversial. In standard, classical logic-based systems, this "A and not A" state is an intractable contradiction that can lead to system crashes or return unreliable, undefined results. This invention recognizes that such conflicts are valuable medical information that must be handled properly, rather than simply ignored or flagged as errors. Therefore, this invention is designed with the following working principle to address this problem:

[0121] First, in response to the small inference model detecting the aforementioned conflict state in step (b2), the binary consistency judgment is paused. At this point, simply outputting support or contradiction is no longer appropriate. The system (specifically...) Figure 2The chain-reasoning and control module (i) immediately activates a revocable adjudication engine. This engine is a specialized software module whose underlying logic is not classical logic, but non-monotonic logic and revocable reasoning. The core idea of ​​non-monotonic logic is that adding new information (e.g., a new piece of evidence) to the knowledge base can lead to the overturning or invalidation (i.e., revocation) of old, derived conclusions. This perfectly aligns with the reality in medical practice where new guidelines overturn old guidelines and new research overturns old theories. The revocable adjudication engine is configured to access a set of dominant meta-rules for resolving conflict states. Meta-rules are the core of this mechanism; they are rules about rules or knowledge about knowledge. Their role is not to judge the facts themselves (e.g., whether Guideline G1 is correct), but to determine which fact rule (i.e., related evidence) has higher priority or advantage in a specific context. In this embodiment, the engine is configured with a complete set of dominant meta-rules predefined by medical experts and knowledge engineers to handle conflicts. This set of meta-rules includes at least the following: First, the Lex Posterior rule (i.e., later evidence prevails over earlier evidence), which states that when two pieces of related evidence (e.g., two clinical guidelines) are of equal authority but have contradictory conclusions, the evidence published later prevails, and the evidence published earlier is rejected. Second, the Lex Superior rule (i.e., higher-level evidence prevails over lower-level evidence), which states that when two pieces of related evidence reach contradictory conclusions, evidence from a higher-authority source (e.g., national-level clinical guidelines issued by the National Health Commission or high-quality systematic reviews) prevails, and evidence from a lower-authority source (e.g., single case reports, small-sample observational studies, or individual expert opinions) prevails. The evidence is rejected; third, the specificity preponderance rule (i.e., special rules prevail over general rules) states that when two pieces of related evidence reach contradictory conclusions, evidence with more specific and relevant conditions (e.g., medication guidelines for patients with type 2 diabetes and severe renal insufficiency) prevails, while evidence with broader conditions (e.g., general guidelines for all patients with type 2 diabetes) is rejected; fourth, the data type preponderance rule states that objective, first-hand data obtained directly from the patient's current medical activity (e.g., current lab results or imaging reports) takes precedence over historical data (e.g., lab reports from three months ago) or general descriptions in textbooks (e.g., typical presentations of the disease). For example, in the specific implementation of the revocable adjudication engine, when the engine accesses the preponderance rule to resolve conflict states, a "timeliness preponderance rule (i.e., Lex Posterior)" for adjudicating timeliness conflicts can be encoded as the following IF-THEN logical pseudocode:

[0122] / / --- Example of a dominant meta-rule: MR-001 ---

[0123] / / Meta-rule ID: MR-001 (Timeliness Advantage Meta-rule)

[0124] / / Description: When two pieces of evidence are of equal authority, the evidence published later prevails.

[0125] / / Functions called by the engine

[0126] FUNCTION Adjudicate_Timeliness (Evidence E_Support, Evidence E_Deny):

[0127] / / Prerequisite: Only when the two pieces of evidence are of the same level of authority

[0128] / / Only the "timeliness" meta-rule is activated.

[0129] / / (If the levels are different, it will be handled by the "Authority" meta-rule MR-002)

[0130] IF E_Support.Authority_Level == E_Deny.Authority_Level:

[0131] / / Core adjudication logic

[0132] IF E_Support.Publish_Date > E_Deny.Publish_Date:

[0133] / / Updated evidence from the supporting party

[0134] RETURN AdjudicationResult(

[0135] Winner: E_Support,

[0136] Vetoed: E_Deny,

[0137] Reason_Meta_Rule: "MR-001" )

[0139] ELSE IF E_Deny.Publish_Date > E_Support.Publish_Date:

[0140] / / Update of evidence by the denying party

[0141] RETURN AdjudicationResult(

[0142] Winner: E_Deny,

[0143] Vetoed: E_Support,

[0144] Reason_Meta_Rule: "MR-001" )

[0146] ELSE:

[0147] / / The dates are also the same, so this rule cannot be applied.

[0148] RETURN CanNotAdjudicate

[0149] ELSE:

[0150] / / Due to differences in authority, this rule does not apply.

[0151] RETURN CanNotAdjudicate

[0152] Next, upon activation, the revocable adjudication engine takes the current conflict state (i.e., the set of relevant evidence for a specific claim and all conflicts) as input, and then argues for the evidence in these conflicts based on the dominant meta-rules. The engine traverses its dominant meta-rule base, searching for the first meta-rule that can resolve the current conflict. For example, if the engine finds that Guideline G1 was published in 2015 and Study L1 was published in 2023, it will match and apply the timeliness dominant meta-rule. The engine then adjudicates one winning piece of evidence (i.e., Study L1) and one or more rejected pieces of evidence (i.e., Guideline G1) from the relevant evidence in the conflict. In this revocable reasoning process, the rejected evidence (Guideline G1) can be considered a refutational argument because it directly presents a conclusion contrary to Study L1; while the timeliness dominant meta-rule constitutes a weakening refutation of this refutational argument, that is, it weakens the probative value of the conclusion by pointing out the outdatedness of Guideline G1 (although not its content itself). Through this meta-rule-based argumentation, Study L1 ultimately becomes a guaranteed conclusion.

[0153] Subsequently, the system performs a crucial transparency and traceability operation: it non-destructively preserves one or more rejected pieces of evidence. This means the system does not delete the evidence of Guideline G1, as this would be dangerous in medical (and legal) terms, erasing the context of the decision. Instead, the system stores Guideline G1 and its sources (such as title, authors, publication date, etc.) as an immutable annotation associated with the calibrated electronic medical record (e.g., in a dedicated audit log table in the database, ensuring its immutability). Simultaneously, the system generates decision source metadata. This is a critical step that ensures interpretability. Decision source metadata is a structured piece of information (e.g., an XML or JSON fragment) that clearly records the dominant meta-rule on which the revocable decision engine made its decision (e.g., record rule ID: Meta-Rule-01, rule name: Timeliness Dominance Meta-Rule), and all the evidence involved in the decision (the winning study L1 and the rejected Guideline G1). The system then binds the decision source metadata to the immutable annotation.

[0154] Finally, upon completion of the consistency assessment in step (b2), the engine returns the conclusion of the winning evidence (e.g., the conclusion of study L1, which refutes the original specific assertion) to step (c) of Example 1. Therefore, the correction in step (c) is based on the winning evidence (e.g., the system might revise a draft medical record, removing the assertion recommending drug A and possibly replacing it with a new assertion supported by the winning evidence), while preserving links to immutable annotations. When authorized users of Example 7 (e.g., doctors) see this correction on the human-machine interface, they can not only see the result of the correction but also click the link to view the complete metadata of the decision source, understanding why the system made this decision (e.g., seeing that the system rejected it because study L1 (2023) is more up-to-date than guideline G1 (2015)...), achieving true human-machine collaboration, intelligent transparency, and decision traceability.

[0155] Figure 2 A schematic diagram of a multi-stage electronic medical record generation system based on a structured thought chain is provided for embodiments of this application, such as... Figure 2 As shown, the system includes:

[0156] One or more processors; and

[0157] A memory storing computer-executable instructions that, when executed by the one or more processors, implement the method described above.

[0158] The system specifically includes:

[0159] (i) A chain-based reasoning and control module, adapted to perform step (a) to generate the intermediate state of the medical record draft;

[0160] (ii) A knowledge service module, adapted to perform step (b1) to retrieve the relevant evidence from the authoritative knowledge base for the forward knowledge enhancement path;

[0161] (iii) The chain reasoning and control module further includes a small reasoning model adapted to perform the (b2) step to perform the consistency judgment in the backward fact-checking path;

[0162] The chain reasoning and control module is also adapted to perform step (c) to correct the intermediate state of the medical record draft based on the result of the consistency judgment, so as to generate the calibrated electronic medical record.

[0163] Please see Figure 2This embodiment provides a multi-stage electronic medical record generation system based on a structured thought chain. This system is a combination of hardware and software that executes the methods of the aforementioned embodiments (e.g., Embodiment 7). At the hardware level, the system includes (but is not limited to) one or more processors (e.g., a central processing unit (CPU) and a graphics processing unit (GPU) for parallel computing) deployed in a server or cloud computing environment, as well as memory (e.g., volatile RAM and non-volatile memory such as a solid-state drive (SSD)). Computer-executable instructions (i.e., program code) are stored in the memory. When one or more processors execute these instructions, the method of Embodiment 7, including steps (a), (b1), (b2), (c), (d), and (e), is implemented. At the software and functional module level, the system specifically includes: (i) a chain-based reasoning and control module, which is the system's central scheduling center and core logic layer, suitable for executing step (a), including coordinating the large language model to generate structured evidence intermediate states according to the one-statement-five-history paradigm (Example 2), performing formal verification (Example 8), performing matching and filling according to the principle of no information increment (Example 3), and guiding the large language model to generate medical record draft intermediate states (Example 4); (ii) a knowledge service module, which is the system's knowledge platform, suitable for executing step (b1), i.e., responding to the request of the chain-based reasoning and control module (i), and providing forward knowledge enhancement paths from authoritative knowledge bases (such as the evidence-based medicine literature submodule and medical knowledge graph submodule in Example 6). (iii) The chain reasoning and control module (i) also includes a small reasoning model, which is the system's fact checker. It is adapted to perform step (b2), that is, to perform consistency judgment in the backward fact-checking path (such as the thought chain truncation verification in Example 5), and to activate the revocable adjudication engine when a conflict state is detected (Example 9); wherein, the chain reasoning and control module (i) is also adapted to perform step (c), that is, to correct the intermediate state of the medical record draft based on the final judgment result of the small reasoning model (iii) (or the revocable adjudication engine) to generate a calibrated electronic medical record, and to perform steps (d) and (e) through the human-machine collaboration interface (Example 7) to present the results and key intermediate states to the authorized user.

[0164] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A multi-stage electronic medical record generation method based on structured thought chain, characterized in that, The method comprises the following steps: (a) generating at least one medical record draft intermediate state containing at least one specific assertion to be verified based on input raw medical data through a structured thinking chain processing flow; (b) starting a double-path knowledge calibration mechanism to perform calibration on the medical record draft intermediate state, which comprises: (b1) a forward knowledge enhancement path: in response to the specific assertion, retrieving associated evidence related to the specific assertion from one or more authoritative knowledge bases; and (b2) a backward fact verification path: using a small reasoning model to judge the consistency of the specific assertion and the associated evidence; (c) based on the results of the consistency judgment, revising the medical record draft intermediate state to generate a calibrated electronic medical record.

2. The method of claim 1, wherein, Before generating the medical record draft intermediate state in the (a) step, it further comprises: adopting a step-by-step thinking chain prompting strategy to guide the large language model to extract information from the raw medical data according to the preset medical paradigm to generate a structured evidence intermediate state; wherein the preset medical paradigm comprises a five-history medical paradigm.

3. The method of claim 2, wherein, The (a) step further comprises: logically matching and filling the structured evidence intermediate state with a pre-set medical record template; wherein, in the matching and filling process, the principle of no information increase is strictly implemented, and no new information other than the structured evidence intermediate state is introduced to form a logical framework intermediate state.

4. The method of claim 3, wherein, The (a) step of generating the medical record draft intermediate state comprises: combining the logical framework intermediate state with the context data obtained from the patient's past medical records or laboratory system, guiding the large language model to convert the logical framework intermediate state into a natural language narrative to form the medical record draft intermediate state.

5. The method of claim 4, wherein, The consistency judgment in the (b2) step specifically comprises: The small reasoning model performs a thinking chain truncation verification, which only performs a binary consistency judgment on the specific assertion and the associated evidence to determine whether the specific assertion is supported by the associated evidence.

6. The method of claim 5, wherein, The authoritative knowledge base in the (b1) step comprises: an evidence-based medicine literature submodule for real-time retrieval of clinical guidelines and high-quality literature; and a medical knowledge graph submodule for providing structured associations between diseases, symptoms, examinations and treatments.

7. The method of claim 6, wherein, The method further comprises: (d) presenting the calibrated electronic medical record generated in the (c) step to authorized users on a human-computer collaborative interface; and (e) visualizing the key intermediate states produced by the structured thinking chain in the (a) step and the (b) step, and / or highlighting the specific assertion that fails the consistency judgment in the (b2) step and is revised.

8. The method of claim 3, wherein, In the (a) step, before strictly implementing the principle of no information increase in the matching and filling process, it further comprises: defining a set of medical integrity constraints, which are encoded as a formal specification language and are used to describe the structure and association rules that the structured evidence intermediate state must satisfy, including at least entity integrity constraints and reference integrity constraints; activating a correction protocol in response to the formal verification engine outputting a verification failure result, wherein the verification failure result indicates that the structured evidence intermediate state violates the medical integrity constraints due to missing information or logical breakage; wherein the correction protocol is authorized to be exempted from the information non-increase principle constraint in rectifying the structured evidence intermediate state to ensure that the structural level integrity of the structured evidence intermediate state has been achieved before entering the knowledge calibration of the (b) step. The consistency judgment in the (b2) step further comprises: in response to the small reasoning model detecting a conflict state in which the specific conclusion is supported by at least one associated evidence and denied by at least another associated evidence in the (b2) step, suspending the binary consistency judgment and activating an overruleable decision engine; the overruleable decision engine is configured to access a set of advantage meta-rules for resolving the conflict state, and based on the advantage meta-rules, decide a winning evidence and one or more defeated evidences from the conflicting associated evidences; 9. The method of claim 5, wherein, non-destructively preserving the one or more defeated evidences as an immutable annotation associated with the calibrated electronic medical record; and generating decision source metadata that records the advantage meta-rules on which the overruleable decision engine makes the decision, and binding the decision source metadata to the immutable annotation; wherein the correction in the (c) step is based on the winning evidence while preserving the link to the immutable annotation. comprise:

10. A multi-stage electronic medical record generation system based on structured thought chains, characterized in that, one or more processors; and a memory having computer-executable instructions stored thereon that, when executed by the one or more processors, implement the method of claim 7; The system specifically comprises: (i) a chain reasoning and control module adapted to perform the (a) step to generate the medical record draft intermediate state; (ii) a knowledge service module adapted to perform the (b1) step to retrieve the associated evidences for the forward knowledge enhancement path from the authoritative knowledge base; (iii) the chain reasoning and control module further comprises a small reasoning model adapted to perform the (b2) step to perform the consistency judgment in the backward fact checking path; wherein the chain reasoning and control module is further adapted to perform the (c) step to correct the medical record draft intermediate state based on the result of the consistency judgment to generate the calibrated electronic medical record. comprise: one or more processors; and a memory having computer-executable instructions stored thereon that, when executed by the one or more processors, implement the method of claim 7; The system specifically comprises: (i) a chain reasoning and control module adapted to perform the (a) step to generate the medical record draft intermediate state; (ii) a knowledge service module adapted to perform the (b1) step to retrieve the associated evidences for the forward knowledge enhancement path from the authoritative knowledge base; (iii) the chain reasoning and control module further comprises a small reasoning model adapted to perform the (b2) step to perform the consistency judgment in the backward fact checking path; wherein the chain reasoning and control module is further adapted to perform the (c) step to correct the medical record draft intermediate state based on the result of the consistency judgment to generate the calibrated electronic medical record.

Images