Case report electronic form automatic filling method and system based on doctor-patient conversation

Abstract

The application discloses a case report electronic form automatic filling method and system based on doctor-patient conversation, and the method comprises the following steps: obtaining doctor-patient conversation text under an emergency trauma scene; inputting the doctor-patient conversation text into an instance extraction module, combining a pre-constructed body part knowledge base and an injury condition category knowledge base, and extracting standardized body part information and injury condition category information by using a large language model; taking the standardized body part information and the injury condition category information as key values, calling a CRF form field knowledge base through a CRF form field extraction module, obtaining corresponding CRF form structures and field extraction rules; combining the doctor-patient conversation text, filling the CRF form structures according to the field extraction rules by using the large language model, and generating a structured CRF form. The application realizes full-process automation from doctor-patient conversation to the CRF form, and significantly improves the accuracy, timeliness and scene adaptability of data acquisition.

Description

Technical Field

[0001] This invention relates to the field of medical text processing technology, and in particular to a method and system for automatically filling out electronic forms for medical record reports based on doctor-patient dialogue. Background Technology

[0002] Emergency trauma is a common acute and critical illness scenario in clinical medicine, characterized by sudden onset, rapid progression, complex injuries, and extremely high requirements for timely information collection. Accurate and complete clinical data collection is the core foundation for developing treatment plans, assessing prognosis, and conducting clinical research for emergency trauma patients. The Case Report Form (CRF), as the core carrier of standardized clinical data collection, directly impacts the reliability of medical decisions and research data through its completed quality.

[0003] Currently, the generation and completion of CRF forms in emergency trauma scenarios mainly rely on two technical approaches: First, the traditional manual entry mode, where medical staff, after completing the doctor-patient consultation, manually organize and fill in pre-set CRF form templates based on memory or handwritten notes, then access the electronic medical record system. In this mode, CRF form templates are mostly generalized designs, covering basic fields such as patient information, injury location, injury mechanism, vital signs, and past medical history. Some medical institutions design more specific templates for different types of trauma. Second, a preliminary information-assisted mode. With the development of medical information technology, some medical institutions have introduced Speech-to-Text (STT) technology to convert doctor-patient dialogues into text data in real time, which medical staff then use to filter and enter CRF fields. A few other solutions attempt to use keyword matching technology to extract core information such as "trauma time," "cause of injury," and "symptom presentation" from the doctor-patient dialogue text and automatically fill in the corresponding fixed fields of the CRF form, reducing the workload of manual entry. Meanwhile, in the medical application of Natural Language Processing (NLP), related technologies have been initially implemented in scenarios such as electronic medical record structuring and medical literature retrieval, providing technical references for information extraction from doctor-patient dialogue texts.

[0004] Current technologies still have many shortcomings that urgently need to be addressed, specifically: 1. Insufficient real-time performance and high risk of information loss. In traditional manual data entry, medical staff must complete CRF entries during breaks in high-intensity emergency care, resulting in significant time delays. This makes it easy to miss crucial information in doctor-patient dialogues due to memory lapses or busy schedules. Even solutions incorporating speech-to-text technology often require subsequent manual editing and filling, failing to achieve synchronous integration between the dialogue process and CRF generation, thus failing to meet the "real-time data collection" needs of emergency trauma. 2. Poor accuracy and completeness of information extraction. Existing keyword-matching-based information extraction solutions can only identify simple keywords corresponding to fixed fields in the CRF, unable to handle complex language scenarios such as colloquial expressions, ellipses, and ambiguous sentences in emergency doctor-patient dialogues. Furthermore, they cannot identify implicit information in the dialogue, leading to insufficient completeness and accuracy of extracted information, affecting the quality of CRF data. 3. Poor form template adaptability, making it difficult to match complex injuries. Existing CRF templates are mostly pre-set, fixed formats that only cover general trauma scenarios. However, emergency trauma patients exhibit significant differences in their injuries, and the core fields of the CRF differ considerably depending on the specific injury. Current technology cannot dynamically adjust the field structure of the CRF form based on real-time injury information obtained during doctor-patient dialogues. This results in templates either having redundant fields that increase the workload of filling out forms, or missing key fields that fail to meet personalized data collection needs. 4. High cost and low efficiency of manual intervention. Whether using traditional manual data entry or voice-to-text assisted data entry, medical staff ultimately need to invest a significant amount of time in information verification, supplementation, organization, and field matching. In emergency trauma scenarios, medical staff are already under heavy workloads, and the additional CRF data entry work further exacerbates the shortage of manpower and reduces overall treatment efficiency. At the same time, human error is easily introduced during manual intervention, further affecting the reliability of CRF data. Summary of the Invention

[0005] The purpose of this invention is to address the problems of insufficient real-time performance, poor accuracy and completeness of information extraction, poor adaptability of form templates, and high cost and low efficiency of manual intervention in existing technologies. It provides a method and system for automatically filling out electronic forms for case reports based on doctor-patient dialogue. The system combines a two-stage CRF automated form filling system with a body part knowledge base, an injury category knowledge base, and a CRF form field knowledge base. The aim is to improve the accuracy, timeliness, and automation of CRF form extraction from doctor-patient dialogue in emergency trauma departments.

[0006] The technical solution adopted in this invention is:

[0007] The method for automatically filling out electronic case report forms based on doctor-patient dialogue includes the following steps:

[0008] Text Acquisition: Acquire the text of doctor-patient dialogues in emergency trauma scenarios;

[0009] The first stage of extraction involves inputting the doctor-patient dialogue text into the instance extraction module. Combining the pre-built body part knowledge base and injury category knowledge base, a large language model is used to extract standardized body part information and injury category information.

[0010] The second stage of extraction involves using standardized body part information and injury category information as keys. The CRF form field extraction module calls the CRF form field knowledge base to obtain the corresponding CRF form structure and field extraction rules. Combined with the doctor-patient dialogue text, the large language model is used to fill the CRF form structure according to the field extraction rules to generate a structured CRF form.

[0011] Furthermore, the methods for constructing the body part knowledge base and the injury category knowledge base include:

[0012] The medical standard terminology of the core target group is used as the top-level core node;

[0013] Construct a first-level association library associated with the top-level core node. The first-level association library includes a core thesaurus for mapping everyday spoken expressions, a core medical context association library for associating typical symptoms or injury types, and a sub-library for recording the subdivided parts or subdivided injuries of the core target objects.

[0014] For each node in the sub-library, a secondary association library is constructed. The secondary association library includes a sub-synonym library and a sub-medical context association library.

[0015] Furthermore, the CRF form field knowledge base adopts a modular, layered architecture, including:

[0016] The general field module is used to cover the core fields for all clinical scenarios;

[0017] Multiple body part-specific field modules, each corresponding to a body part, are used to store CRF fields specific to the corresponding body part; specifically, the multiple body part-specific field modules include head module, neck module, face module, chest module, back module, abdomen and perineum module, waist and kidney area module, and limb and pelvis module.

[0018] At least one injury category-specific field module, each injury category-specific field module corresponding to one injury category, used to store CRF fields specific to the corresponding injury category; specifically, it includes two injury category-specific modules: a skin and soft tissue module and a burn and scald module.

[0019] Each CRF field includes a field identifier, a standard Chinese name, a field type, a set of options, a default value, a thesaurus, and extraction rules.

[0020] Specifically, the field identifier serves as a unique identifier for interface and association with the hospital system, consistent with the hospital's definition, and designed using English abbreviations and semantic encoding rules; the standard Chinese name adopts a unified and standardized field name to eliminate naming ambiguity; the field type defines the field data format, supporting four core types: Boolean, single choice, multiple choice, and text; the option set defines standardized optional values ​​for single / multiple choice fields to ensure a consistent value range; the default value is the default fill value when there is no relevant clinical description, ensuring data integrity; the thesaurus associates synonyms, near-synonyms, and colloquial expressions for fields or options to achieve expression normalization; the extraction rules are used to clarify the judgment logic for extracting field values ​​from clinical dialogues / medical records, ensuring that the extraction process is repeatable and verifiable.

[0021] Furthermore, the thesaurus for each CRF field is generated as follows:

[0022] By analyzing CRF form fields, different expressions of the same field / option are extracted to form an original thesaurus;

[0023] The original thesaurus pool is deduplicated to remove semantically conflicting expressions, ensuring that the synonyms are semantically consistent with the standard fields / options;

[0024] The original deduplicated thesaurus pool is hierarchically classified to form thesaurus of corresponding types, and then summarized to form the thesaurus of corresponding CRF fields.

[0025] Specifically, for Boolean and text fields, a "Field → Synonym List" is constructed to obtain field-level synonyms; for single-select / multiple-select fields, an "Option → Synonym Map" is constructed to obtain option-level synonyms.

[0026] Specifically, the thesaurus for each CRF field supports the addition of synonyms based on new clinical scenarios, keeping the knowledge base up-to-date.

[0027] Furthermore, the implementation of the corresponding CRF form structure is determined as follows: based on the input body part information and injury category information, the fields of the corresponding general field module, body part specific field module and injury category specific field module are called from the CRF form field knowledge base and combined to form the CRF form structure to be filled in this time.

[0028] Furthermore, before processing with a large language model in both the first and second stage extraction, a prompt word engineering step is included. This step organizes the doctor-patient dialogue text, knowledge base information, and extraction targets / rules into instructions to guide the large language model to perform standardized extraction or structured completion tasks.

[0029] Furthermore, the extraction rules are designed based on the principle of "default when no description is provided, matching when a description is provided, and judgment when a description is ambiguous." The implementation of the extraction rules includes:

[0030] Boolean fields follow the rule of "explicit mention takes effect", and default values ​​are used if there is no explicit mention.

[0031] In single-choice / multiple-choice fields, configure field extraction rule descriptions for each optional field and add the field extraction rule descriptions to the thesaurus. Priority: clinically explicit field terminology description / numerical description > thesaurus matching > default value; handling without description: if not explicitly mentioned, fill in the default value;

[0032] Text fields: Use the rule of "fill in if the part / information is clear, leave blank if not";

[0033] Special scenario rules: For rapid diagnosis and treatment scenarios such as emergency rooms, "simplified judgment rules" are designed (such as the "consciousness" field: being able to have normal conversation is judged as "conscious", without the need for complex evaluation).

[0034] An automated electronic form filling system for medical records based on doctor-patient dialogue includes:

[0035] The knowledge base construction module is used to build and store knowledge bases for body parts, injury categories, and CRF form fields.

[0036] The instance extraction module connects to the knowledge base construction module. It receives doctor-patient dialogue text and, based on the body part knowledge base and injury category knowledge base, calls a large language model to extract standardized body part and injury category information.

[0037] The CRF form field extraction module connects the instance extraction module and the knowledge base construction module. Based on the output of the instance extraction module, it calls the CRF form field knowledge base to determine the form structure and rules, and calls the large language model to generate the final CRF structured form.

[0038] Furthermore, the knowledge base construction module includes a body part and injury category knowledge base module and a CRF form field knowledge base module.

[0039] Specifically, the body parts and injury categories knowledge base module analyzes the existing CRF forms in the emergency trauma department and currently includes eight body parts: head, neck, face, chest, back, abdomen and perineum, waist and kidney area, and limbs and pelvis, as well as two injury categories: skin and soft tissue injuries and burns.

[0040] The present invention adopts the above technical solution and has the following beneficial effects compared with the prior art:

[0041] 1) Improve the accuracy and robustness of information extraction by combining body part knowledge base, injury category knowledge base and CRF form field knowledge base with LLM, especially in complex language scenarios such as colloquial expressions, elliptical sentences and ambiguous sentences in emergency doctor-patient dialogue.

[0042] 2) By using a two-stage CRF automated form filling method that first identifies body parts and injury categories and then extracts the corresponding CRF form fields, the timeliness, accuracy, and form template adaptability of CRF form extraction are improved.

[0043] 3) The method of this invention automates the entire process from doctor-patient dialogue text to CRF form field extraction, greatly reducing the cost of manual intervention. Attached Figure Description

[0044] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments;

[0045] Figure 1 This is a schematic diagram illustrating the principle architecture of the automatic form filling method for electronic case reports based on doctor-patient dialogue of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0047] This invention constructs a body part knowledge base and an injury category knowledge base to extract entities from doctor-patient dialogue text, accurately identifying the patient's body parts and injury categories. It then calls upon the corresponding form structure for secondary identification to extract the corresponding CRF form fields. The specific implementation process of this invention includes a body part knowledge base module, an injury category knowledge base module, an entity extraction module, a CRF form field knowledge base module, and a CRF form field extraction module.

[0048] like Figure 1 As shown, this invention discloses a specific embodiment of a method for automatically filling out electronic medical record forms based on doctor-patient dialogue, including:

[0049] Text Acquisition: Acquire the text of doctor-patient dialogues in emergency trauma scenarios;

[0050] The first stage of extraction involves inputting the doctor-patient dialogue text into the instance extraction module. Combining the pre-built body part knowledge base and injury category knowledge base, a large language model is used to extract standardized body part information and injury category information.

[0051] The second stage of extraction involves using standardized body part information and injury category information as keys. The CRF form field extraction module calls the CRF form field knowledge base to obtain the corresponding CRF form structure and field extraction rules. Combined with the doctor-patient dialogue text, the large language model is used to fill the CRF form structure according to the field extraction rules to generate a structured CRF form.

[0052] Furthermore, the methods for constructing the body part knowledge base and the injury category knowledge base include:

[0053] Using medical standard terminology as the top-level core node;

[0054] Construct a first-level association library associated with the top-level core node. The first-level association library includes a core thesaurus for mapping everyday spoken expressions, a core medical context association library for associating typical symptoms or injury types, and a sub-library for recording specific parts or injuries.

[0055] For each node in the sub-library, a secondary association library is constructed. The secondary association library includes a sub-synonym library and a sub-medical context association library.

[0056] Specifically, the methods for constructing a body part knowledge base include:

[0057] The core body parts standard terminology is used as the top-level core node;

[0058] Construct a first-level association library associated with the top-level core node. The first-level association library includes a core thesaurus for mapping everyday spoken expressions, a core medical context association library for associating typical disease or injury types, and a sub-library for recording the subdivided parts of the core body parts.

[0059] For each node in the sub-library, a secondary association library is constructed. The secondary association library includes a sub-synonym library and a sub-medical context association library.

[0060] Furthermore, the methods for constructing the injury category knowledge base include:

[0061] The core injury category standard terminology serves as the top-level core node;

[0062] Construct a first-level association library associated with the top-level core node. The first-level association library includes a core thesaurus for mapping everyday spoken expressions, a core medical context association library for associating typical symptoms or injury types, and a sub-library for recording subdivided injuries of core injury categories.

[0063] For each node in the sub-library, a secondary association library is constructed. The secondary association library includes a sub-synonym library and a sub-medical context association library.

[0064] Furthermore, the CRF form field knowledge base adopts a modular, layered architecture, including:

[0065] The general field module is used to cover the core fields for all clinical scenarios;

[0066] Multiple body part-specific field modules, each corresponding to a body part, are used to store CRF fields specific to that body part; specifically, there are 8 body part-specific modules, including head module, neck module, face module, chest module, back module, abdomen and perineum module, waist and kidney area module, and limb and pelvis module.

[0067] At least one injury category-specific field module, each injury category-specific field module corresponding to one injury category, used to store CRF fields specific to the corresponding injury category; specifically, it includes two injury category-specific modules: a skin and soft tissue module and a burn and scald module.

[0068] Each CRF field includes a field identifier, a standard Chinese name, a field type, a set of options, a default value, a thesaurus, and extraction rules.

[0069] Specifically, the field identifier serves as a unique identifier for interface and association with the hospital system, consistent with the hospital's definition, and designed using English abbreviations and semantic encoding rules; the standard Chinese name adopts a unified and standardized field name to eliminate naming ambiguity; the field type defines the field data format, supporting four core types: Boolean, single choice, multiple choice, and text; the option set defines standardized optional values ​​for single / multiple choice fields to ensure a consistent value range; the default value is the default fill value when there is no relevant clinical description, ensuring data integrity; the thesaurus associates synonyms, near-synonyms, and colloquial expressions for fields or options to achieve expression normalization; the extraction rules are used to clarify the judgment logic for extracting field values ​​from clinical dialogues / medical records, ensuring that the extraction process is repeatable and verifiable.

[0070] Furthermore, the thesaurus for each CRF field is generated as follows:

[0071] By analyzing CRF form fields, different expressions of the same field / option are extracted to form an original thesaurus;

[0072] The original thesaurus pool is deduplicated to remove semantically conflicting expressions, ensuring that the synonyms are semantically consistent with the standard fields / options;

[0073] The original deduplicated thesaurus pool is hierarchically classified to form thesaurus of corresponding types, and then summarized to form the thesaurus of corresponding CRF fields.

[0074] Specifically, for Boolean and text fields, a "Field → Synonym List" is constructed to obtain field-level synonyms; for single-select / multiple-select fields, an "Option → Synonym Map" is constructed to obtain option-level synonyms.

[0075] Specifically, the thesaurus for each CRF field supports the addition of synonyms based on new clinical scenarios, keeping the knowledge base up-to-date.

[0076] Furthermore, the implementation of the corresponding CRF form structure is determined as follows: based on the input body part information and injury category information, the fields of the corresponding general field module, body part specific field module and injury category specific field module are called from the CRF form field knowledge base and combined to form the CRF form structure to be filled in this time.

[0077] Furthermore, before processing with a large language model in both the first and second stage extraction, a prompt word engineering step is included. This step organizes the doctor-patient dialogue text, knowledge base information, and extraction targets / rules into instructions to guide the large language model to perform standardized extraction or structured completion tasks.

[0078] Furthermore, the extraction rules are designed based on the principle of "default when no description is provided, matching when a description is provided, and judgment when a description is ambiguous." The implementation of the extraction rules includes:

[0079] Boolean fields follow the rule of "explicit mention takes effect", and default values ​​are used if there is no explicit mention.

[0080] In single-choice / multiple-choice fields, configure field extraction rule descriptions for each optional field and add the field extraction rule descriptions to the thesaurus. Priority: clinically explicit field terminology description / numerical description > thesaurus matching > default value; handling without description: if not explicitly mentioned, fill in the default value;

[0081] For example, in the Trauma Index Circulatory Indicator, the single-choice fields are: "Normal", "Systolic blood pressure <100mmHg or heart rate >100 beats / min"; the default value is "Normal". After adding the field extraction rule description to the thesaurus, the thesaurus is: "Normal": ["Normal blood pressure", "Normal heart rate", "No abnormal blood pressure", "No abnormal heart rate"], "Systolic blood pressure <100mmHg or heart rate >100 beats / min": ["Low blood pressure", "Fast heart rate", "Panic", "Fast heartbeat"].

[0082] Text fields: Use the rule of "fill in if the part / information is clear, leave blank if not";

[0083] Special scenario rules: For rapid diagnosis and treatment scenarios such as emergency rooms, "simplified judgment rules" are designed (such as the "consciousness" field: being able to have normal conversation is judged as "conscious", without the need for complex evaluation).

[0084] An automated electronic form filling system for medical records based on doctor-patient dialogue includes:

[0085] The knowledge base construction module is used to build and store knowledge bases for body parts, injury categories, and CRF form fields.

[0086] The instance extraction module connects to the knowledge base construction module. It receives doctor-patient dialogue text and, based on the body part knowledge base and injury category knowledge base, calls a large language model to extract standardized body part and injury category information.

[0087] The CRF form field extraction module connects the instance extraction module and the knowledge base construction module. Based on the output of the instance extraction module, it calls the CRF form field knowledge base to determine the form structure and rules, and calls the large language model to generate the final CRF structured form.

[0088] Furthermore, the knowledge base construction module includes a body part and injury category knowledge base module and a CRF form field knowledge base module.

[0089] Specifically, the body parts and injury categories knowledge base module analyzes the existing CRF forms in the emergency trauma department and currently includes eight body parts: head, neck, face, chest, back, abdomen and perineum, waist and kidney area, and limbs and pelvis, as well as two injury categories: skin and soft tissue injuries and burns.

[0090] In emergency trauma scenarios, patients often use colloquial expressions, ellipses, and ambiguous statements, which can easily lead to inefficient doctor-patient communication and inaccurate transmission of medical information. This invention aims to construct a structured knowledge base of body parts and injury categories, connecting "patient's everyday expressions → standard medical terminology → clinical symptom association," to assist emergency departments in quickly and accurately identifying conditions and improving communication efficiency.

[0091] First, the body part knowledge base adopts a hierarchical and relational architecture as its core structure. The "standard terminology for body parts" serves as the top-level core node, extending downwards to three primary relational libraries: a core thesaurus, a core medical context relational library, and a sub-part library. The "sub-part library" acts as a secondary core node, further deriving two secondary relational libraries: a sub-part thesaurus and a sub-part medical context relational library, forming a hierarchical mapping structure of "core node → primary relation → secondary relation." The top-level core node is the core body part node, using the standard medical terminology of eight body parts as the unified anchor point of the knowledge base to ensure the standardization of medical expressions. The primary relational libraries connect to "everyday expressions + clinical needs," building three parallel relational libraries around the core node: the core thesaurus, the core medical context relational library, and the sub-part library.

[0092] The core thesaurus collects common colloquialisms, dialect expressions, and simplified terms for key body parts (e.g., "head" corresponds to "head," "brain," "skull," and "top of the head"). This core thesaurus automatically matches "patient's spoken language" with medical terminology, solving the problem of "medical staff not understanding the words patients use."

[0093] The core medical context association database associates typical symptoms, injury types, and clinical complaints corresponding to core body parts (e.g., "neck" corresponds to "neck sprain, cervical spine injury, neck hematoma, tracheal injury, carotid artery injury"). As a supplement to the core thesaurus, the core medical context association database fills in gaps in the core thesaurus (e.g., the core thesaurus for "neck" includes "neck, neck, neck, neck," but if "tracheal injury" is mentioned, the core medical context association database can supplement this by associating "neck").

[0094] The sub-part database breaks down core parts into more specific sub-parts, clearly defining the standard medical terminology for each sub-part (e.g., "face" is broken down into "eyes, nose, mouth"). The sub-part database refines the precision of part descriptions, addressing the issue of how a patient's statement of "nasal bone fracture" can be linked to "face."

[0095] The secondary association database provides a precise mapping of sub-parts. As a sub-node of the "sub-part database," it supplements the two types of association databases: the sub-part synonym database and the sub-part medical context database. The sub-part synonym database collects everyday expressions for sub-parts (e.g., "mouth" corresponds to "mouth, oral cavity, lips"), further refining local expressions (e.g., mentioning "mouth" first associates with "mouth," then with "face"). The sub-part medical context database associates specific diseases and injuries corresponding to the sub-parts (e.g., "mouth" corresponds to "lip laceration," "tooth injury," "oral bleeding"). As a supplement to the sub-part synonym database, the sub-part medical context database addresses unresolved cases (e.g., if the sub-part synonym database for "mouth" is "mouth, oral cavity, lips," mentioning "tooth injury" will supplement it with "mouth," then "face").

[0096] Secondly, the injury category knowledge base also adopts a hierarchical and relational architecture as its core structure. With the "injury category standard terminology" as the top-level core node, it extends downward to three types of first-level relational databases: core thesaurus, core medical context relational database, and sub-injury database. Among them, the "sub-injury database" serves as a secondary core node, further deriving two types of second-level relational databases: sub-injury thesaurus and sub-injury medical context relational database, forming a hierarchical mapping structure of "core node → first-level relation → second-level relation".

[0097] Furthermore, the core of the aforementioned instance extraction module is based on doctor-patient dialogue text, combined with a knowledge base and a Large Language Model (LLM), to standardize the extraction of body parts and injury categories. The input to the instance extraction module is the doctor-patient dialogue text, including the patient's self-reported symptoms, the doctor's consultation content, and a description of the patient's condition; this serves as the raw material for subsequent information extraction. The instance extraction module incorporates two professional knowledge bases—the aforementioned body parts and injury location knowledge bases—to provide standardized references, supporting the standardization of the extraction results.

[0098] The instance extraction module constructs precise instructions for the LLM through prompt word engineering, serving as a bridge between the "raw text + knowledge base" and the LLM. It takes the doctor-patient dialogue text as the content to be processed, while embedding standardized expressions and classification standards from the "body part knowledge base" and the "injury category knowledge base," organizing this information into clear instructions that define the LLM's extraction goals and standards. The instance extraction module inputs the instructions constructed through prompt word engineering into the large-scale language model (LLM). The LLM, referring to the specifications of the two knowledge bases, matches the recognized colloquial expressions into standardized results, identifying content related to body parts and injuries from the doctor-patient dialogue text. The LLM extracts standardized results from the doctor-patient dialogue that conform to the knowledge base specifications: body parts and injury categories.

[0099] Furthermore, the aforementioned CRF form field knowledge base module, based on the emergency trauma department, constructs a CRF form field knowledge base using a layered modular architecture. It is divided into five core layers: "Field Classification Layer, Core Attribute Definition Layer, Synonym Mapping Layer, Extraction Rule Engine Layer, and Default Value Configuration Layer." It also includes two major field clusters: "General Field Set" and "Specialized Field Set," forming a knowledge base system covering all dimensions of the emergency trauma department, with standardized data collection and automated extraction logic. Its core architectural design goal is to solve the technical problems of inconsistent expression, inaccurate extraction, non-standard handling of missing values, and insufficient coverage of specific scenarios during clinical data collection, achieving accurate mapping and automated filling of doctor-patient dialogue data into CRF form fields.

[0100] The CRF form field knowledge base adopts a modular, layered architecture, allowing for flexible access to corresponding module fields based on specific clinical scenarios. This balances general applicability with specialized adaptability, removes irrelevant and redundant fields, and improves the efficiency and accuracy of CRF field extraction. Based on the emergency trauma scenario, the knowledge base is divided into general modules, body part-specific modules, and injury category-specific modules, achieving full coverage of scenarios, body parts, and injury categories.

[0101] The general module covers the core fields that need to be recorded in all clinical scenarios, including past medical history, allergy history, surgical history, trauma index (circulatory / respiratory), consciousness, etc.

[0102] The body part-specific modules are divided according to the type of body part in the CRF form, including eight body part-specific modules: head, neck, face, chest, back, abdomen and perineum, waist and kidney area, and limbs and pelvis. Each module focuses on the specific fields of the corresponding body part.

[0103] The injury category modules are divided according to the injury category type in the CRF form, including two injury category modules: the skin and soft tissue module and the burn module. Each model focuses on the specific fields of the corresponding injury category.

[0104] The core advantage of modularization lies in the ability to flexibly call corresponding module fields according to specific clinical scenarios, balancing universality and specialization, removing irrelevant and redundant fields, and improving the efficiency and accuracy of CRF field extraction.

[0105] Each module of the CRF form field knowledge base contains 7 core attributes, forming a standardized field structure to ensure the uniqueness and interpretability of the fields. (1) Field Identifier: A unique identifier used to connect with the hospital system. It is consistent with the hospital's definition and is designed using English abbreviations and semantic coding rules. (2) Standard Chinese Name: A unified field name to eliminate naming ambiguity. (3) Field Type: Defines the field data format and supports 4 core types: Boolean, single choice, multiple choice, and text. (4) Option Set: Defines standardized optional values ​​for single / multiple choice fields to ensure a unified range of values. (5) Default Value: The default fill value when there is no relevant clinical description to ensure data integrity. (6) Thesaurus: Synonyms, near-synonyms, and colloquial expressions associated with fields or options to achieve expression normalization. (7) Extraction Rules: Clarify the judgment logic for extracting field values ​​from clinical dialogues / medical records to ensure that the extraction process is repeatable and verifiable.

[0106] The thesaurus in the CRF form field knowledge base is the core of achieving normalization of diverse clinical expressions, and adopts the strategy of "layered construction + scenario adaptation". The thesaurus analyzes the CRF form fields, extracts different expressions of the same field / option, and forms the original thesaurus pool; the original thesaurus pool is classified in layers: (1) for Boolean and text fields, "field → thesaurus list" is constructed to obtain field-level synonyms; (2) for single-choice / multiple-choice fields, "option → thesaurus mapping table" is constructed to obtain option-level synonyms; (3) the original thesaurus pool is deduplicated and semantically conflicting expressions are removed to ensure that the synonyms are semantically consistent with the standard field / option; (4) the thesaurus supports the addition of synonyms according to new clinical scenarios to maintain the timeliness of the knowledge base.

[0107] The extraction rules in the CRF form field knowledge base are designed based on the principle of "default when no description is available, matching when a description is available, and judgment when a description is ambiguous." Specifically, Boolean fields adopt the rule of "explicit mention takes effect," and default values ​​are used when no explicit mention is available; single / multiple selection fields configure field extraction rule descriptions for each optional field and add them to the thesaurus, with priority: clinically explicit field terminology / numerical description > thesaurus matching > default value; no description is handled: default value is filled if no explicit mention is available; text fields adopt the rule of "fill when specific location / information is available, leave blank if none is available"; special scenario rules are designed for rapid diagnosis and treatment scenarios such as emergency rooms, with "simplified judgment rules" (e.g., for the "consciousness" field: being able to converse normally is judged as "conscious," without complex evaluation).

[0108] The CRF form field knowledge base solves the problem of heterogeneous CRF form fields across different institutions by unifying field attribute definitions and naming conventions, enabling seamless integration of multi-center data. Its modular design allows extraction to call only relevant fields, eliminating irrelevant and redundant fields and improving extraction efficiency by 90%. The CRF form field knowledge base's thesaurus covers diverse clinical expressions, with clear and executable extraction rules, reducing human judgment errors and achieving a field extraction accuracy of ≥90% in emergency scenarios. Its modular design covers general scenarios, specific scenarios for different body parts and injury categories, flexibly adapting to different needs such as clinical research, emergency treatment, and burn management. The default value settings in the CRF form field knowledge base avoid the "no description means empty" problem, ensuring CRF form data integrity of ≥98%. The CRF form field knowledge base supports module addition, thesaurus supplementation, and rule optimization, adapting to the development of medical technology and changes in clinical scenarios.

[0109] Furthermore, the aforementioned CRF form field extraction module adopts a structured generation process with interconnected stages. The core is to use the "body part + injury category" extracted by the instance extraction module as input, and generate a standardized CRF form through collaboration between the knowledge base and the Large Language Model (LLM).

[0110] The CRF form field extraction module uses the "body part + injury category" extracted by the instance extraction module as input to determine the specific CRF form. The module calls the CRF form structure knowledge base, which uses the input "body part + injury category" as the key to match the corresponding CRF form framework template. The CRF form field knowledge base pre-stores the specific field details of CRF forms for various scenarios. It outputs a list of specific fields and specifications for this scenario to the "prompt word project," clarifying the detailed content items that need to be extracted and filled in. The prompt word project uses the doctor-patient dialogue text as the raw text, combining the form framework and corresponding specific field list and specifications from the CRF form structure knowledge base to organize the information into clear instructions, clarifying the target fields and extraction rules for LLM. After receiving the instructions from the prompt word project, LLM organizes and fills in the information based on the raw information of the corresponding scenario, according to the field structure and extraction rules required by the prompt words, completing the structured organization of the form content. After processing, LLM directly outputs a CRF structured form that meets the scenario requirements.

[0111] The CRF form field extraction module uses the CRF form field knowledge base to clarify the "framework + details" requirements of the form, and then guides the LLM to complete the structured generation of the content through prompt word engineering, finally obtaining a usable standardized CRF form.

[0112] The present invention adopts the above technical solution and has the following beneficial effects compared with the prior art:

[0113] 1) Improve the accuracy and robustness of information extraction by combining body part knowledge base, injury category knowledge base and CRF form field knowledge base with LLM, especially in complex language scenarios such as colloquial expressions, elliptical sentences and ambiguous sentences in emergency doctor-patient dialogue.

[0114] 2) By using a two-stage CRF automated form filling method that first identifies body parts and injury categories and then extracts the corresponding CRF form fields, the timeliness, accuracy, and form template adaptability of CRF form extraction are improved.

[0115] 3) The method of this invention automates the entire process from doctor-patient dialogue text to CRF form field extraction, greatly reducing the cost of manual intervention.

[0116] Obviously, the described embodiments are only a part of the embodiments of this application, not all of them. Without conflict, the embodiments and features in the embodiments of this application can be combined with each other. The components of the embodiments of this application described and illustrated herein can generally be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of this application is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

Claims

1. A method for automatically filling out electronic case report forms based on doctor-patient dialogue, characterized in that: Includes the following steps: Text Acquisition: Acquire the text of doctor-patient dialogues in emergency trauma scenarios; The first stage of extraction involves inputting the doctor-patient dialogue text into the instance extraction module. Combining the pre-built body part knowledge base and injury category knowledge base, a large language model is used to extract standardized body part information and injury category information. The second stage of extraction involves using standardized body part information and injury category information as keys. The CRF form field extraction module calls the CRF form field knowledge base to obtain the corresponding CRF form structure and field extraction rules. Combined with the doctor-patient dialogue text, the large language model is used to fill the CRF form structure according to the field extraction rules to generate a structured CRF form.

2. The method for automatically filling out electronic case report forms based on doctor-patient dialogue according to claim 1, characterized in that, The methods for constructing body part knowledge bases and injury category knowledge bases include: The medical standard terminology of the core target group is used as the top-level core node; Construct a first-level association library associated with the top-level core node. The first-level association library includes a core thesaurus for mapping everyday spoken expressions, a core medical context association library for associating typical symptoms or injury types, and a sub-library for recording the subdivided parts or subdivided injuries of the core target objects. For each node in the sub-library, a secondary association library is constructed. The secondary association library includes a sub-synonym library and a sub-medical context association library.

3. The method for automatically filling out electronic case report forms based on doctor-patient dialogue according to claim 1, characterized in that, The CRF form field knowledge base adopts a modular, layered architecture, including: The general field module is used to cover the core fields for all clinical scenarios; Multiple body part-specific field modules, each corresponding to a body part, are used to store CRF fields specific to the corresponding body part; At least one injury category-specific field module, each injury category-specific field module corresponds to one injury category, and is used to store CRF fields specific to the corresponding injury category; Each CRF field includes a field identifier, a standard Chinese name, a field type, an option set, a default value, a thesaurus, and extraction rules. The field identifier serves as a unique identifier for interface with the hospital system. The standard Chinese name adopts a unified and standardized field terminology. The field type defines the field data format. The option set defines standardized optional values. The default value provides a default fill value when there is no clinical description. The thesaurus associates the field or option with synonyms, near-synonyms, and colloquial expressions. The extraction rules are used to clarify the judgment logic for extracting field values ​​from clinical dialogues / medical records.

4. The method for automatically filling out electronic case report forms based on doctor-patient dialogue according to claim 3, characterized in that, The module includes several body part-specific fields, such as head, neck, face, chest, back, abdomen and perineum, waist and kidney area, and limbs and pelvis; the module also includes a category-specific field for injuries, such as skin and soft tissue and burns.

5. The method for automatically filling out electronic case report forms based on doctor-patient dialogue according to claim 1, characterized in that, The thesaurus for each CRF field is generated as follows: By analyzing CRF form fields, different expressions of the same field / option are extracted to form an original thesaurus; The original thesaurus is deduplicated to remove semantically conflicting expressions; The original deduplicated thesaurus pool is hierarchically classified to form thesaurus of corresponding types, and then summarized to form the thesaurus of corresponding CRF fields.

6. The method for automatically filling out electronic case report forms based on doctor-patient dialogue according to claim 1, characterized in that, The implementation of the corresponding CRF form structure is as follows: based on the input body part information and injury category information, the fields of the corresponding general field module, body part specific field module and injury category specific field module are called from the CRF form field knowledge base and combined to form the CRF form structure to be filled in this time.

7. The method for automatically filling out electronic case report forms based on doctor-patient dialogue according to claim 1, characterized in that, Before processing with a large language model in both the first and second stages of extraction, a prompt word engineering step is included. This step organizes the doctor-patient dialogue text, knowledge base information, extraction targets, or extraction rules into execution instructions to guide the large language model to perform standardized extraction or structured filling tasks.

8. The method for automatically filling out electronic case report forms based on doctor-patient dialogue according to claim 1, characterized in that, The extraction rules are designed based on the principle of "default when no description is provided, matching when a description is provided, and judgment when a description is ambiguous".

9. An automatic form filling system for electronic medical record reports based on doctor-patient dialogue, employing the automatic form filling method for electronic medical record reports based on doctor-patient dialogue as described in any one of claims 1 to 8, characterized in that, The system includes: The knowledge base construction module is used to build and store knowledge bases for body parts, injury categories, and CRF form fields. The instance extraction module connects to the knowledge base construction module. It receives doctor-patient dialogue text and, based on the body part knowledge base and injury category knowledge base, calls a large language model to extract standardized body part and injury category information. The CRF form field extraction module connects the instance extraction module and the knowledge base construction module. Based on the output of the instance extraction module, it calls the CRF form field knowledge base to determine the form structure and rules, and calls the large language model to generate the final CRF structured form.

10. The automatic form filling system for electronic case reports based on doctor-patient dialogue according to claim 9, characterized in that, The knowledge base construction module includes a body part and injury category knowledge base module and a CRF form field knowledge base module. The body part and injury category knowledge base module is based on the analysis of existing CRF forms in the emergency trauma department, including eight body parts: head, neck, face, chest, back, abdomen and perineum, waist and kidney area, and limbs and pelvis, as well as two injury categories: skin and soft tissue and burns.

Images