A method and related apparatus for structured extraction of medical text based on field routing and context complexity assessment

Abstract

This application provides a method and related apparatus for structured extraction of medical text based on field routing and context complexity assessment. The method includes: S1. Obtaining the target medical text and a set of target fields; S2. Performing field routing decisions based on a field-engine capability mapping matrix; S3. Performing a complexity quantification assessment on the context of each field; S4. Performing authority extraction by a rule engine; S5. Supplementing extraction with a large language model and performing conflict resolution to obtain the target extraction result; S6. Performing terminology standardization mapping and outputting in a structured form. This application also provides related apparatus corresponding to the method, including devices, electronic devices, computer-readable storage media, and computer program products.

Description

Technical Field

[0001] This application relates to the field of natural language processing, and in particular to a method and related apparatus for structured extraction of medical text based on field routing and context complexity assessment. Background Technology

[0002] Structured extraction of medical text is a key technology that converts unstructured or semi-structured medical texts such as medical records, examination reports, and surgical records into standardized structured data. It is widely used in scenarios such as clinical decision support, medical research data collection, and medical big data analysis.

[0003] Existing methods for structured extraction of medical text suffer from the following problems: First, existing methods primarily employ a serial processing architecture of rule engines and large language models, applying the same processing flow to all fields without differentiating the routing to the most suitable inference engine based on field type. This results in simple structured fields (such as time and numerical values) being wasted on expensive large language models, while complex semantic fields (such as symptom descriptions) suffer from insufficient accuracy when processed by rule engines. Second, existing methods lack a quantitative evaluation mechanism for context complexity, failing to adaptively trigger supplementary extraction from large language models based on context complexity. This leads to the large language model being invoked even for simple contexts, resulting in wasted computational power, while complex contexts are not invoked, leading to insufficient semantic extraction capabilities. Third, existing methods treat the output of large language models and rule engine outputs equally or treat the output of large language models as the only source of information. The existing methods are post-processing refinement tools, failing to recognize the illusionary risk of large language models affecting structured data (time, values, dosage), resulting in the structured data extraction accuracy being affected by illusions; fourth, the confidence assessment of the existing methods is based only on the probability values ​​output by the inference engine itself, without introducing type consistency confidence based on field type, resulting in a lack of refinement in the confidence assessment at the field dimension; fifth, the output of the existing methods only includes the final extraction results, without retaining source information such as the source inference engine, confidence, and contextual complexity of each extraction result, resulting in downstream applications being unable to trace interpretability; sixth, the terminology output of the existing methods is the original text terminology, without performing terminology standardization mapping based on medical terminology standard libraries (such as ICD-10, SNOMED-CT, RxNorm), resulting in downstream applications being unable to directly use standardized codes. Summary of the Invention

[0004] To address the above technical issues, this application provides a method and related apparatus for structured extraction of medical text based on field routing and context complexity assessment.

[0005] Firstly, a method for structured extraction of medical text based on field routing and context complexity evaluation is provided, including the following steps: S1. Obtain the target medical text and the target field set. The target medical text is obtained by preprocessing clinical documents. The target field set includes multiple target fields, and each target field includes a field name and a field type. S2. Based on the field type of each target field, query the field-engine capability mapping matrix to perform field routing decisions and obtain the preferred inference engine corresponding to each target field. The field-engine capability mapping matrix includes the mapping relationship from field type to rule engine or large language model. S3. Perform a complexity quantification assessment based on syntactic multidimensional indicators on the context of each field in the target medical text to obtain a context complexity score. The syntactic multidimensional indicators include sentence length, number of nested clauses, number of negation expressions, and number of ambiguous pronouns. When the context complexity score is greater than a preset routing threshold, the preferred inference engine is upgraded to a large language model. S4. The rule engine performs authoritative extraction in advance to obtain rule extraction results. The rule extraction results include field values, the location of field values, and type consistency confidence. The type consistency confidence is calculated based on the degree of formal consistency between the field values ​​and the field types. S5. For fields where the preferred inference engine is a large language model or the type consistency confidence is lower than a preset confidence threshold, the large language model performs supplementary extraction to obtain supplementary extraction results. Based on the rule engine priority principle, conflict resolution is performed on the rule extraction results and the supplementary extraction results to obtain the target extraction results. The rule engine priority principle is that when the rule extraction results are inconsistent with the supplementary extraction results, the rule extraction results are used as the authority value. S6. Perform terminology standardization mapping on the target extraction results based on the medical terminology standard library to obtain standardized target extraction results. Output the standardized target extraction results and their source information in a structured form. The source information includes the source inference engine, type consistency confidence, and context complexity score of the target extraction results.

[0006] In any embodiment of this application, the field type includes at least one of the following: time field, numeric field, terminology field, descriptive field, and relational field.

[0007] In any embodiment of this application, the context complexity score is calculated based on the following formula: , Where C is the context complexity score, L is the normalized value of sentence length, N is the normalized value of the number of nested clauses, D is the normalized value of the number of negation expressions, P is the normalized value of the number of ambiguous pronouns, and α, β, γ and δ are the corresponding weights.

[0008] In any embodiment of this application, the type consistency confidence is calculated based on the formal consistency between the field value and the field type, which is fundamentally different from the probability confidence output by the large language model itself.

[0009] In conjunction with any embodiment of this application, the conflict resolution based on the rule engine priority principle is used to address the illusion risk of large language models for structured data.

[0010] In conjunction with any embodiment of this application, the medical terminology standard library includes at least ICD-10 disease classification codes, SNOMED-CT clinical medical terms, and RxNorm drug standard nomenclature; the results output in a structured form retain the source information of each extraction result to support interpretability and traceability.

[0011] In a second aspect, a medical text structure extraction device based on field routing and context complexity assessment is provided, comprising: a text and field set acquisition unit, a field routing decision unit, a context complexity assessment unit, a rule engine pre-extraction unit, a large model supplementation and conflict resolution unit, and a terminology standardization and traceability output unit, which respectively execute steps S1 to S6 of the method described in the first aspect.

[0012] Thirdly, an electronic device is provided, including a processor and a storage unit, the storage unit storing a computer program that, when executed by the processor, causes the electronic device to perform the method described in the first aspect.

[0013] Fourthly, a computer-readable storage medium is provided, storing a computer program that, when executed by a processor, causes the processor to perform the method described in the first aspect.

[0014] Fifthly, a computer program product is provided, comprising a computer program or instructions that, when executed on a computer, cause the computer to perform the method described in the first aspect.

[0015] It should be understood that the above general descriptions and subsequent specific descriptions are for illustrative and explanatory purposes only and do not impose any limitations on this application.

[0016] In this application, compared with the prior art, the technical problem to be solved by this application is: to provide a differential routing mechanism based on field type in the structured extraction of medical text, to provide a quantitative evaluation mechanism based on context complexity, to handle the illusion risk of large language models for structured data, to provide type consistency confidence evaluation based on field type, to retain source information to support downstream interpretability traceability, and to provide terminology standardization mapping based on medical terminology standard library.The technical means employed in this application and their non-obvious aspects compared to existing technologies include: First, this application implements field-level routing decisions based on a field-engine capability mapping matrix. This design is not seen in existing methods employing a unified processing architecture (such as serial processing of rule engines-large models-high-precision models). The field-level routing mechanism enables simple structured fields (time / numerical values) to use the rule engine, while complex semantic fields (symptom descriptions) use the large language model, achieving a precise balance between computing power and accuracy. Second, this application drives adaptive routing based on a complex metric evaluation mechanism using syntactic multidimensional indicators (sentence length + nesting + negation + ambiguity). The design described herein is not found in existing methods that employ fixed serial processing or post-processing rule engines. The complex metric mechanism allows simple contexts to bypass the large language model, saving computational power, while complex contexts call the large language model to ensure accuracy. Third, this application employs a rule engine-first principle to resolve conflicts. This design differs from the "confidence threshold followed by rule engine post-processing" mechanism of existing methods (such as in reference document 2). Existing methods use the large language model as the authoritative value and then refine it with a rule engine, while this application uses the rule engine as the authoritative value and only supplements the large language model, directly addressing and resolving the large language model's limitations on structured data (time / value / dosage). The application addresses several key issues: First, it addresses the risk of hallucinations, which is not identified in existing methods. Second, it calculates type consistency confidence based on the formal consistency between field values ​​and field types. This design differs fundamentally from existing methods that rely on the probability output of the inference engine itself for confidence assessment. The type consistency confidence provides a refined evaluation of the field dimension, making subsequent conflict resolution decisions more accurate. Third, the structured output of this application retains source information for each extraction result, including the source inference engine, type consistency confidence, and context complexity score. This design is not seen in existing methods that only output the final extraction result. The source information is preserved... The application provides interpretable traceability for downstream clinical decision support systems; sixth, this application performs terminology standardization mapping based on medical terminology standard libraries such as ICD-10, SNOMED-CT, and RxNorm. This design is not seen in existing methods for outputting original text terms. The terminology standardization mapping allows downstream applications to directly use standardized codes, significantly improving clinical data interoperability; seventh, this application integrates field routing decisions, context complexity assessment, rule engine prioritization, and terminology standardization mapping into a complete medical text structure extraction technology system. This overall technical architecture is not seen in existing methods that only use some components. The related devices described in this application include devices, electronic devices, computer-readable storage media, and computer program products. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be explained below.

[0018] Figure 1 This is a flowchart illustrating a medical text structure extraction method based on field routing and context complexity assessment, provided as an embodiment of this application.

[0019] Figure 2 This is a schematic diagram of a medical text structure extraction device based on field routing and context complexity assessment, provided as an embodiment of this application.

[0020] Figure 3 This is a schematic diagram of the hardware architecture of an electronic device provided in an embodiment of this application. Detailed Implementation

[0021] To enable those skilled in the art to more fully understand the technical solution of this application, the technical solution of this application will be explained in detail and clearly with reference to the accompanying drawings.

[0022] Please see Figure 1 , Figure 1 This is a flowchart illustrating a medical text structure extraction method based on field routing and context complexity assessment, provided for an embodiment of this application. The method includes steps S1 to S6.

[0023] S1. Text and Field Set Acquisition: Acquire the target medical text and the target field set.

[0024] In this embodiment, the target medical text is obtained by preprocessing clinical documents. The preprocessing includes at least sentence segmentation, noise reduction, unifying full-width and half-width characters, and cleaning up garbled text. The clinical documents include, but are not limited to, electronic medical records, examination reports, surgical records, discharge summaries, or doctor-patient dialogue texts.

[0025] The target field set includes multiple target fields defined based on a medical terminology ontology, each target field including a field name and a field type. The medical terminology ontology can be based on existing standard definitions. Examples of the target field set definition include: {field name: "diagnosis", field type: "terminology"}; {field name: "systolic blood pressure", field type: "numerical"}; {field name: "diagnosis time", field type: "time"}; {field name: "chief complaint", field type: "descriptive"}.

[0026] S2. Field routing decision: Based on the field type of each target field in the target field set, query the field-engine capability mapping matrix, perform field routing decision, and obtain the preferred inference engine corresponding to each target field.

[0027] In this embodiment, the field type includes at least one of the following: time field, numeric field, terminology field, descriptive field, and relational field. The field-engine capability mapping matrix includes at least the following mapping relationships: The time-type field is preferably selected from the rule engine. The selection is based on the fact that time expressions have strong structured features, and the rule engine can accurately match based on regular expressions with zero risk of illusion. The numerical fields are preferably selected from the rule engine. The selection is based on the fact that the combination of numerical values ​​and units has strong structured features, and the rule engine has better accuracy than large language models. The terminology field is preferably selected from either the rule engine or the large language model, dynamically based on the medical terminology dictionary hit rate. When the dictionary hits, the rule engine is preferred; when the dictionary does not hit, the large language model is preferred. The descriptive fields are preferably derived from the large language model. This selection is based on the fact that natural language descriptions have strong semantics, and the large language model has better semantic understanding capabilities than the rule engine. The relational field is preferably selected from the large language model, and the selection is based on the fact that entity relation extraction requires complex semantic reasoning.

[0028] The field-level routing decision mechanism described herein is the core innovation of this application, and it differs fundamentally from the conventional steps of existing methods that employ a unified processing architecture. This field-level routing ensures that different field types receive the most suitable inference engine processing, achieving a precise balance between computational power and accuracy.

[0029] S3. Context complexity assessment: Perform a complexity quantification assessment based on syntactic multidimensional indicators on the context of each field in the target medical text to obtain a context complexity score.

[0030] In this embodiment, the context complexity score is calculated based on the following formula: , Where C is the context complexity score, L is the normalized value of sentence length, N is the normalized value of the number of nested clauses, D is the normalized value of the number of negation expressions, P is the normalized value of the number of ambiguous pronouns, α, β, γ and δ are the corresponding weights, and α+β+γ+δ=1. The values ​​of the corresponding weights are respectively α from 0.2 to 0.4, β from 0.2 to 0.4, γ from 0.1 to 0.3 and δ from 0.1 to 0.3.

[0031] When the context complexity score is greater than a preset routing threshold, the preferred inference engine is upgraded to a large language model, and the value range of the preset routing threshold is from 0.4 to 0.7. The adaptive routing upgrade mechanism based on context complexity quantification assessment is another core innovation point of this application, enabling simple contexts to bypass the large language model to save computing power and complex contexts to call the large language model to ensure accuracy.

[0032] S4. Rule engine pre-extraction: The rule engine performs authoritative extraction pre-emptively to obtain a rule extraction result.

[0033] In this embodiment, the rule engine performs extraction on the target medical text based on regular rules and a medical dictionary. The rule extraction result includes a field value, the location of the field value, and a type consistency confidence level.

[0034] The type consistency confidence level is calculated based on the formal consistency degree between the field value and the field type. The formal consistency degree is calculated according to the following rules: For time-type fields, when the field value conforms to a time regular expression (such as "YYYY-MM-DD", "YYYY年MM月DD日"), the type consistency confidence level is greater than or equal to 0.9; For numeric fields, when the field value conforms to a numeric regular expression and is within a reasonable physiological range (such as systolic blood pressure between 60 and 250), the type consistency confidence level is greater than or equal to 0.9; For term-type fields, when the field value hits a medical dictionary (such as a disease dictionary, a symptom dictionary, a drug dictionary), the type consistency confidence level is greater than or equal to 0.8.

[0035] The type consistency confidence level based on the formal consistency between the field value and the field type is essentially different from the confidence level assessment based on the probability output of the inference engine itself. The type consistency confidence level provides a refined assessment at the field dimension.

[0036] S5. Large model supplementation and conflict resolution: For fields where the preferred inference engine is a large language model or the type consistency confidence level is lower than a preset confidence threshold, the large language model performs supplementary extraction to obtain a supplementary extraction result, and conflict resolution is performed on the rule extraction result and the supplementary extraction result based on the rule engine pre-emption principle to obtain a target extraction result.

[0037] In this embodiment, the large language model performs the supplementary extraction based on a preset prompt template, and the preset prompt template includes a task description, a list of target field types, and an output format constraint. The large language model includes, but is not limited to, a general large language model based on the Transformer architecture or a medical large language model fine-tuned for the medical field.

[0038] The conflict resolution based on the rule engine's priority principle is executed according to the following rules: First, when the field value exists in the rule extraction result and the type consistency confidence is greater than or equal to the preset confidence threshold, the target extraction result adopts the rule extraction result; Second, when the field value does not exist in the rule extraction result or the type consistency confidence is lower than the preset confidence threshold, the target extraction result adopts the supplementary extraction result; Third, when the rule extraction result and the supplementary extraction result exist simultaneously but are inconsistent, the target extraction result adopts the rule extraction result as the authoritative value, and the supplementary extraction result is recorded as a reference in the source information. The preset confidence threshold ranges from 0.6 to 0.8.

[0039] The rule engine-first priority principle is one of the core innovations of this application, and it is fundamentally different from the "confidence threshold followed by rule engine post-processing" mechanism of existing methods (such as prior art CN121502646A): existing methods use the large language model as the authoritative value and then refine it with the rule engine, while this application uses the rule engine as the authoritative value and only supplements the large language model. The rule engine-first priority principle directly addresses and solves the illusion risk problem of the large language model in structured data (time / value / dosage). This illusion risk problem originates from the occasional errors of the large language model in generating structured data, and this illusion risk problem has not been identified in existing medical text extraction methods.

[0040] S6. Terminology Standardization and Traceability Output: Based on the medical terminology standard library, perform terminology standardization mapping on the target extraction results to obtain standardized target extraction results, and output the standardized target extraction results and their source information in a structured form.

[0041] In this embodiment, the medical terminology standard library includes at least ICD-10 disease classification codes, SNOMED-CT clinical medical terms, and RxNorm drug standard nomenclature. The terminology standardization mapping includes at least the following mappings: mapping disease names in the target extraction results to ICD-10 codes (e.g., "hypertension" mapped to "I10"); mapping symptom names to SNOMED-CT codes; and mapping drug names to RxNorm codes.

[0042] The structured format includes at least one of JSON, XML, and key-value pairs. The structured format also includes a pending review status marker; when the type consistency confidence level is lower than the pending review threshold (ranging from 0.5 to 0.7), the target extraction result is marked as pending review, prompting the annotator or doctor to conduct a manual review.

[0043] The results are output in a structured format, retaining source information for each target extraction result. This source information includes at least the source inference engine, type consistency confidence score, and context complexity score of the target extraction result. This interpretability traceability based on source information enables downstream clinical decision support systems to assess the reliability of each extraction result, a feature not found in existing methods that only output the final extraction result.

[0044] Clinical Application Example 1: In a structured inpatient medical record scenario, the method performs structured extraction on the admission record text "Patient is a 65-year-old male, admitted due to recurrent chest pain for 3 days, with a 5-year history of hypertension, currently taking amlodipine 5mg once daily". S2 field routing decision yields: Gender (terminal) → Rule Engine; Age (numerical) → Rule Engine; Chief Complaint (descriptive) → Large Language Model; Past Illnesses (terminal) → Rule Engine + Dictionary; Medication (terminal + numerical combination) → Rule Engine + Large Language Model dual extraction. S3 context complexity scoring calculates C=0.35 for the chief complaint "recurrent chest pain for 3 days" (below the 0.5 threshold), so it is not upgraded to the Large Language Model; for the past history sentence, due to the ambiguous pronoun "past," C=0.55 (exceeding the threshold), the Large Language Model is upgraded to supplement it. In the S4-S5 conflict resolution, the drug "Amlodipine" was matched by the rule engine from the drug dictionary, with a type consistency confidence of 0.92. The large language model also extracted "Amlodipine", which is consistent, so the rule engine result is adopted. The output after S6 terminology standardization mapping is: {"Past Diseases":[{"name":"Hypertension","icd10":"I10","source":"Rule Engine","confidence":0.95}],"Medication":[{"name":"Amlodipine","rxnorm":"17767","dose":"5mg","frequency":"Once daily","source":"Rule Engine","confidence":0.92}]}.

[0045] Clinical Application Example 2: In a multi-center clinical research data acquisition scenario, the method performs structured extraction on the surgical record text "laparoscopic anterior resection of rectal cancer + total mesorectal resection + terminal ileostomy". After S2 routing, the surgical name (terminal type) is matched with SNOMED-CT by the rule engine; the surgical approach (relational type) is matched with the large language model; and the additional operations (descriptive type) are matched with the large language model. The final output is a standardized SNOMED-CT code used for multi-center data aggregation and statistical analysis, and the source inference engine and confidence level of each extracted item are labeled to support traceability.

[0046] Please see Figure 2 , Figure 2A schematic diagram of a medical text structure extraction device based on field routing and context complexity evaluation provided in this application embodiment. The medical text structure extraction device 1 based on field routing and context complexity evaluation includes: Text and field set acquisition unit 11 is used to acquire target medical text and target field set.

[0047] Field routing decision unit 12 is used to perform field routing decisions based on the field type query field-engine capability mapping matrix.

[0048] The context complexity evaluation unit 13 is used to perform a complexity quantification evaluation of the context based on syntactic multidimensional metrics.

[0049] The rule engine front-end extraction unit 14 is used for authoritative extraction performed by the rule engine front-end.

[0050] The large model supplementation and conflict resolution unit 15 is used to supplement and extract conflict resolution based on the rule engine's priority principle.

[0051] The terminology standardization and traceability output unit 16 is used to perform terminology standardization mapping based on the medical terminology standard library and output it in a structured form that retains the source information.

[0052] Please see Figure 3 , Figure 3 This is a schematic diagram of the hardware architecture of an electronic device provided in an embodiment of this application. The electronic device 2 includes a processor 21, a memory 22, an input device 23, and an output device 24. The processor 21, the memory 22, the input device 23, and the output device 24 are connected in communication via a bus.

[0053] The method described in this application can be implemented based on hardware, software, or a combination of hardware and software.

Claims

1. A medical text structured extraction method based on field routing and context complexity evaluation, characterized in that It includes the following steps: S1. Obtain the target medical text and the target field set. The target medical text is obtained by preprocessing clinical documents. The target field set includes multiple target fields, and each target field includes a field name and a field type. S2. Based on the field type of each target field, query the field-engine capability mapping matrix to perform field routing decisions and obtain the preferred inference engine corresponding to each target field. The field-engine capability mapping matrix includes the mapping relationship from field type to rule engine or large language model. S3. Perform a complexity quantification assessment based on syntactic multidimensional indicators on the context of each field in the target medical text to obtain a context complexity score. The syntactic multidimensional indicators include sentence length, number of nested clauses, number of negation expressions, and number of ambiguous pronouns. When the context complexity score is greater than a preset routing threshold, the preferred inference engine is upgraded to a large language model. S4. The rule engine performs authoritative extraction in the front end to obtain rule extraction results. The rule extraction results include field values, the position of the field values, and type consistency confidence. The type consistency confidence is calculated based on the degree of formal consistency between the field values ​​and the field types. S5. For fields where the preferred inference engine is a large language model or the type consistency confidence level is lower than a preset confidence threshold, supplementary extraction is performed by the large language model to obtain supplementary extraction results. Based on the rule engine priority principle, conflict resolution is performed on the rule extraction results and the supplementary extraction results to obtain the target extraction results. The rule engine priority principle is that when the rule extraction results are inconsistent with the supplementary extraction results, the rule extraction results are used as the authority value. S6. Perform terminology standardization mapping on the target extraction results based on the medical terminology standard library to obtain standardized target extraction results. Output the standardized target extraction results and their source information in a structured form. The source information includes the source inference engine, type consistency confidence, and context complexity score of the target extraction results.

2. The method according to claim 1, characterized in that... The field types include at least one of time-type fields, numerical fields, terminology fields, descriptive fields, and relational fields; the field-engine capability mapping matrix includes at least the following: the time-type fields are preferably the rule engine, the numerical fields are preferably the rule engine, the terminology fields are preferably the rule engine or the large language model, the descriptive fields are preferably the large language model, and the relational fields are preferably the large language model.

3. The method according to claim 1, characterized in that... The context complexity score is calculated based on the following formula: ， Where C is the context complexity score, L is the normalized value of sentence length, N is the normalized value of the number of nested clauses, D is the normalized value of the number of negation expressions, P is the normalized value of the number of ambiguous pronouns, α, β, γ and δ are the corresponding weights, and α+β+γ+δ=1. The values ​​of the corresponding weights are respectively α from 0.2 to 0.4, β from 0.2 to 0.4, γ from 0.1 to 0.3, and δ from 0.1 to 0.3; the value range of the preset routing threshold is 0.4 to 0.

7.

4. The method according to claim 1, characterized in that... The type consistency confidence score is calculated based on the degree of formal consistency between the field value and the field type. The degree of formal consistency is calculated according to the following rules: for time-type fields, the type consistency confidence score is greater than or equal to 0.9 when the field value conforms to a time regular expression; for numeric fields, the type consistency confidence score is greater than or equal to 0.9 when the field value conforms to a numeric regular expression and is within a reasonable physiological range; for terminology-type fields, the type consistency confidence score is greater than or equal to 0.8 when the field value matches a medical dictionary. The preset confidence threshold ranges from 0.6 to 0.

8.

5. The method according to claim 1, characterized in that... The conflict resolution based on the rule engine's prioritization principle is executed according to the following rules: when the field value exists in the rule extraction result and the type consistency confidence is greater than or equal to the preset confidence threshold, the target extraction result adopts the rule extraction result; when the field value does not exist in the rule extraction result or the type consistency confidence is lower than the preset confidence threshold, the target extraction result adopts the supplementary extraction result; when the rule extraction result and the supplementary extraction result exist simultaneously but are inconsistent, the target extraction result adopts the rule extraction result as the authority value, and the supplementary extraction result is recorded as a reference in the source information; the rule engine's prioritization principle is used to handle the illusion risk of large language models on structured data.

6. The method according to claim 1, characterized in that... The medical terminology standard library includes at least ICD-10 disease classification codes, SNOMED-CT clinical medical terms, and RxNorm drug standard nomenclature; the terminology standardization mapping includes at least mapping disease names in the target extraction results to ICD-10 codes, symptom names to SNOMED-CT codes, and drug names to RxNorm codes; the structured form includes at least one of JSON, XML, and key-value pairs; the structured form also includes a pending review status marker, marking the target extraction result as pending review when the type consistency confidence is lower than the pending review threshold; the pending review threshold ranges from 0.5 to 0.7; the standardized target extraction result is also used for at least one of the following downstream applications: automatic electronic medical record filling, clinical decision support systems, medical research data analysis, and medical insurance expense review.

7. A medical text structure extraction device based on field routing and context complexity evaluation, characterized in that... ,include: A text and field set acquisition unit is used to perform step S1 as described in claim 1; A field routing decision unit, configured to perform step S2 as described in claim 1; The context complexity evaluation unit is used to perform step S3 as described in claim 1; The rule engine pre-extraction unit is used to perform step S4 as described in claim 1; A large language model supplementation and conflict resolution unit is used to perform step S5 as described in claim 1; The terminology standardization and traceability output unit is used to perform step S6 as described in claim 1.

8. An electronic device, characterized in that... The electronic device includes: a processor and a storage unit for storing computer program code containing computer instructions, wherein when the processor executes these instructions, the electronic device performs the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that... The computer-readable storage medium stores a computer program, the computer program containing program instructions that, when executed by a processor, cause the processor to perform the method described in any one of claims 1 to 6.

10. A computer program product, characterized in that... The computer program product comprises a computer program or instructions that, when executed on a computer, cause the computer to perform the method described in any one of claims 1 to 6.

Images