Medical record generation optimization method based on type differentiation

Abstract

The invention provides a medical record generation optimization method based on type differentiation, and belongs to the field of natural language processing, and the method comprises the steps: constructing a medical record type classification system, designing an exclusive generation model and a scene rule for different types of medical records such as a first disease course, a daily disease course and a stage knot, and combining with a clinical feedback dynamic optimization generation strategy, thereby achieving the purpose of optimizing the medical record generation. The medical record generation method is suitable for scenes such as clinical medical record writing, medical teaching medical record construction and medical AI auxiliary diagnosis and treatment systems of hospitals at all levels.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of natural language processing, large model application, medical application, and in particular to a medical record generation optimization method based on type differentiation. BACKGROUND

[0002] Medical record generation is an urgent problem under the current social background. Traditional medical record writing relies on manual input by doctors, which not only consumes a lot of working time of doctors, but also relies on pure manpower to achieve the standard requirements of medical records in terms of format, terminology uniformity and key information of patients on the same day.

[0003] There are four core pain points in the existing medical record generation technology: first, the type adaptation is single, the traditional method uses a general generation model to process all types of medical records, without designing a dedicated solution for type characteristics such as "first course of disease needs to highlight diagnosis basis" and "discharge summary needs to include discharge guidance", resulting in chaotic medical record structure (such as missing rescue time record in emergency medical record); second, the generation rule is universal, the current generation rule is based on a unified template (such as "including chief complaint, history of present illness, past medical history"), and the differences between surgical departments (which need to emphasize surgical conditions) and pediatric departments (which need to supplement growth and development history) lead to a disconnection between medical record content and department needs; third, the scene response is passive, it is unable to dynamically adjust the generation priority according to the diagnosis and treatment scene (such as daily course, stage summary, etc.), such as daily course still generated in the order of "chief complaint-history of present illness-past medical history", which cannot match the medical record standard; fourth, the quality evaluation is homogeneous, using a unified standard (such as field integrity) to evaluate the quality of all types of medical records, without designing differentiated evaluation dimensions for type characteristics such as "first course of disease needs to evaluate diagnosis logic" and "daily course needs to evaluate diagnosis and treatment consistency", resulting in a mismatch between quality evaluation and clinical actual needs. In addition, the existing technology does not establish a "medical record type-clinical scene-generation strategy" linkage mechanism: when the generation effect of a certain type of medical record is not good, only the model parameters can be adjusted as a whole, and it is unable to locate whether it is "type feature extraction deficiency" "department rule adaptation difference" or "scene priority setting unreasonable", resulting in low optimization efficiency. SUMMARY

[0004] In order to solve the above technical problems, the present application provides a medical record generation optimization method based on type differentiation.

[0005] The technical scheme of the present application is:

[0006] A medical record generation optimization method based on type differentiation, by constructing a medical record type classification system, designing exclusive generation models and scenario rules for different types of medical records such as first course, daily course and stage summary, combining with clinical feedback to dynamically optimize the generation strategy, finally improving the accuracy, standardization and scene adaptation ability of medical record generation, suitable for clinical medical record writing, medical teaching medical record construction, medical AI assisted diagnosis and treatment system and other scenes in hospitals at all levels. Specifically, the following steps are included:

[0007] 1) Unified management of raw data, efficient quality management of data through directory classification and directory file;

[0008] 2) Unified management of derived data, quality management of raw data tracking to derived data and support for user-selected classification extraction and construction of published data through directory classification and directory file;

[0009] 3) Unified management of published data, quality management of raw data tracking to published data and support for user-selected proportion construction of training data, validation data and test data through directory classification and directory file;

[0010] 4) Unified management of training data, quality management of raw data tracking to training data through directory classification and directory file;

[0011] 5) Unified management of validation data, quality management of raw data tracking to validation data through directory classification and directory file, and support for reasoning result verification according to data verification rules to realize training effect evaluation.

[0012] 6) Unified management of test data, quality management of raw data tracking to test data through directory classification and directory file, and support for reasoning result verification according to data verification rules to realize test effect evaluation.

[0013] Further,

[0014] Build type-specific generation model: design differentiated generation algorithm for different types of medical records (first course adopts diagnosis basis priority algorithm, stage summary adopts time axis integration algorithm), build type-specific sub-model, and ensure the accuracy of core field generation through type-specific training and joint fine-tuning;

[0015] Continuous optimization iteration: based on generation logs, evaluation results and customer feedback, update rule library (add / adjust department / scenario rules), fine-tune model parameters (optimize type-specific generation algorithm), verify optimization effect through test data set, form "optimization-verification-reoptimization" closed loop.

[0016] Further,

[0017] Among them, the management activities of original data, derived data, published data, training data, verification data and test data provide means to improve quality efficiency, solve the data quality management problem and data on-demand construction problem in medical record generation fine-tuning training.

[0018] Further,

[0019] The type-specific generation model is based on a Transformer architecture to build a medical record type-specific sub-model, shares a basic word vector layer, and differentially configures an encoding layer and a decoding layer: the decoding layer of the first course sub-model adopts a "diagnosis basis-complaints-present history" reverse generation order, the encoding layer of the stage summary sub-model adds a "time axis integration module", and the core field generation accuracy of each sub-model is ensured through type-specific training.

[0020] Further,

[0021] For medical application field, large model application field and natural language processing field, it is suitable for medical record generation (first course, stage summary, daily course, etc.) of all levels of hospitals, medical AI auxiliary diagnosis and treatment system, different scenes such as routine, emergency and postoperative follow-up.

[0022] Further,

[0023] It also includes a continuous optimization iteration step: based on the generation log (number of manual adjustments), the evaluation result (problem type distribution), customer feedback (doctor satisfaction), the rule library is updated (new medical record generation rules are added), the model parameters are fine-tuned (the generation algorithm is optimized), and the optimization effect is verified through the test data set (such as the improvement of the completeness of the first course diagnosis basis), forming a closed-loop iteration mechanism.

[0024] The beneficial effects of the present application are

[0025] Through the five technical innovations of type classification system, exclusive generation model, scenario-based rule library, dynamic adaptation mechanism and differential evaluation, the existing technical blank is filled. On this basis, the data set is optimized and managed to generate the management needs of original data, the inspection and optimization needs of derived data, and the tracking needs of these data. The present application realizes the continuous improvement of the data quality and training effect of medical record generation training by centrally storing, classifying managing, on-demand constructing and continuously tracking the original data, derived data, published data, training data, verification data and test data generated by medical record generation training. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 is the overall workflow schematic diagram of the present application;

[0027] Figure 2 is a unified management schematic diagram of original data;

[0028] Figure 3 is a unified management schematic diagram of derived data;

[0029] Figure 4 is a unified management schematic diagram of published data;

[0030] Figure 5 is a unified management schematic diagram of training data;

[0031] Figure 6 is a unified management schematic diagram of verification data;

[0032] Figure 7 is a unified management schematic diagram of test data. DETAILED DESCRIPTION

[0033] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments but not all embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0034] The present application relates to a type differentiation-based multi-scene medical record generation optimization method. In view of the pain points of traditional medical record generation technology, such as "type adaptation unification, generation rule generalization, scene response passivation, and quality evaluation homogenization", the method establishes a medical record type classification system, constructs a type-specific generation model, designs a scene-based rule library, realizes dynamic adaptive adjustment and differentiated quality evaluation, and solves problems such as "structure inconsistency, content loss, non-standard expression, and poor scene adaptation" in the generation of different types of medical records (such as first course of disease, daily course of disease, and stage summary). Through continuous optimization and generation training according to project requirements, the generation quality of medical records is improved, and the time required for daily medical record generation is shortened, forming a full-link solution technology field of "type classification-specialized generation-scene adaptation-quality evaluation-continuous optimization".

[0035] The present application aims to solve the problems of "type adaptation unification, generation rule generalization, scene response passivation, quality evaluation homogenization, and optimization linkage deficiency" in existing medical record generation technology. By establishing a medical record type classification system to clarify the core requirements of each type, constructing a type-specific generation model to improve generation accuracy, designing a scene-based rule library to adapt to department and scene differences, realizing dynamic adaptive adjustment to respond to changes in diagnosis and treatment scenes, and ensuring that differentiated quality evaluation results meet clinical needs, the generation quality of different types of medical records is ultimately improved, the generation time is shortened, the clinical adaptation ability is improved, and the individualized medical record generation needs of different departments and different scenes are met.

[0036] To achieve the above objectives, the technical scheme of the present application is as follows:

[0037] 1. Multi-dimensional medical record type classification system, based on three dimensions of diagnosis and treatment stage (admission, daily, stage), the medical record is divided into several categories (such as first course of disease, daily course of disease, stage summary), and the core fields (such as first course of disease including "chief complaint, history of present illness, past history") and expression specifications (such as drug use in psychiatric department should strictly follow the specifications) of each type of medical record are clear.

[0038] 2. Type-specific generation model, different types of medical records are designed with different generation algorithms: 1) First course of disease adopts "diagnosis basis priority generation algorithm", first extract key data of inspection / image to generate diagnosis basis, and then deduce chief complaint and history of present illness; 2) Stage summary adopts "time axis integration algorithm", integrates data in the order of "admission situation-diagnosis and treatment process-stage guidance"; 3) Daily course of disease adopts "emergency information priority generation algorithm", which prioritizes generating key information such as doctor's arrangement of medication and examination, and then supplements the details;

[0039] 3. Type differentiation quality evaluation system, different types of medical records are designed with special evaluation dimensions: 1) First course of disease evaluates the completeness and logical consistency of diagnosis basis; 2) Stage summary evaluates the coherence of diagnosis and treatment process; 3) Daily course of disease evaluates the accuracy of time record and the consistency of medication, etc. The specific steps are as follows:

[0040] Step 1: Construction of medical record type classification system

[0041] Raw data needs to support tracking to derived data according to directory and file to improve data quality, so by defining 《Medical Record Type Classification Standard.xml》 clear classification dimensions and specific types, realize the quality management and classification management of raw data:

[0042] 1) Directory classification management: classify each raw directory according to medical record type and hospital department, so as to extract as needed later.

[0043] 2) Directory naming rules:

[0044] (1) The root directory is named 【01-raw data】,

[0045] (2) Create the same named project collection directory under the above two directories to store the original doctor-patient dialogue data and the corresponding medical record data.

[0046] The naming rules of project collection directory: directory serial number + specific hospital + medical record type

[0047] For example:

[0048]

D0001-xx hospital-first course of disease

[0049]

D0002-yy Hospital-Daily Medical Report

[0050] (3) Under the above project collection directory, plan two subdirectories: [01-Patient Data] and [02-Medical Record Data].

[0051] 3) File naming rules: Bed number + Patient name + Date

[0052] For doctor-patient dialogue data and corresponding inpatient medical record data, files with the same name are stored separately in the project data collection directory, such as:

[0053] 【0001-xxx-20250615.txt】

[0054] 【0002-yyy-20250620.txt】

[0055] The advantage of doing this is that it allows for quick comparison of patient information and corresponding admission medical records by directory name and file name, thereby improving the efficiency of data evaluation and data optimization.

[0056] 4) Data Validation Rules: This section describes the composition and constraints of the original data files (patient-doctor dialogues and admission medical records), taking the initial course of illness as an example.

[0057] Present illness (not empty), personal history (not empty), family history (not empty), preliminary diagnosis (not empty), diagnostic basis (not empty), differential diagnosis (can be empty), treatment plan (not empty).

[0058] 5) According to the data validation rules, the original data can be validated to generate a "Validation Result.xml". The validation result is used to mark erroneous data, guide data quality improvement, and avoid misuse in the training process, which would affect the training quality.

[0059] This step can be found in [reference]. Figure 2 Unified management of raw data.

[0060] Step 2: Unified Management of Derived Data

[0061] Derived data includes various types of datasets derived from the original data, such as:

[0062] Derived data needs to support tracing back to the original data by directory and file for consistent data optimization. Additionally, derived data needs to support extracting one or more derived versions to build customized project datasets.

[0063] Therefore, the `datadefinition.xml` file is defined to implement quality management and classification management of derived data.

[0064] 1) Directory classification management: Maintain consistency with the directory classification method of the original data.

[0065] 2) Directory naming rules:

[0066] (1) The root directory of derived data is named 【02-derived data】

[0067] (2) Create project collection directories under the above two subdirectories, and the directory naming rule uses directory serial number + hospital department + medical record type description to track and trace the data source, such as:

[0068]

D0001-xx hospital_xx department-first course

[0069]

D0002-yy hospital_yy department-daily course

[0070] (3) Under the above project collection directory, plan two subdirectories: 【01-patient data】 and 【02-medical record data】.

[0071] File naming rules: consistent with the file naming rules of original data.

[0072] 3) Data verification rules: consistent with the file naming rules of original data.

[0073] 4) According to the data verification rules, the derived data can be verified to produce 《Verification Result.xml》, which is used to mark error data, guide data quality improvement, and avoid misuse in the training link to affect training quality.

[0074] This step can be referred to Figure 3 Unified management of derived data.

[0075] Step three, unified management of published data

[0076] Published data is a subset of derived data extracted and constructed according to user-selected data classification (business use).

[0077] Published data needs to support tracking management according to the release directory name and file name.

[0078] Therefore, define 《Data Definition.xml》 to achieve tracking management of test data:

[0079] 1) Directory naming rules:

[0080] (1) The root directory of published data is named 【03-published data】

[0081] (2) Create project release directories under the above root directory, and the directory naming rule uses release description + release date to track and trace the release data source, such as:

[0082]

xx hospital - daily course - 20250510

[0083]

yy hospital - first course - 20250610

[0084] (3) Under the project release directory, create two subdirectories: 【01-patient data】 and 【02-medical record data】 to store the doctor-patient conversation data and the corresponding medical records.

[0085] 2) File naming rules:

[0086] For doctor-patient conversation data and corresponding admission medical record data, use the same file naming method. The file naming rule is: directory number + bed number + patient name + date

[0087] For example:

[0088]

D0001-00001-xxx-20250809.txt

[0089]

D0001-00002-yyy-20250815.txt

[0090]

D0002-00003-zzz-20250621.txt

[0091] The advantage of this is that you can quickly compare and evaluate doctor-patient conversation data and corresponding patient medical record data according to the directory and file name, thereby improving data evaluation efficiency and query efficiency, and further improving data modification and optimization efficiency.

[0092] This step can refer to the data management in Figure 4 .

[0093] Step four, unified management of training data

[0094] Training data is a subset of data extracted from the same name release directory in proportion.

[0095] Training data needs to support tracking management according to the release directory name and file name.

[0096] Therefore, define 《data definition.xml》 to achieve tracking management of test data:

[0097] 1) Directory naming rules:

[0098] (1) The root directory of the release data is named 【04-training data】

[0099] (2) In the above root directory, inherit the project release directory of the published data, such as:

[0100]

xx hospital - daily course - 20250510

[0101]

yy hospital - first course - 20250610

[0102] (3) In the project release directory, inherit two subdirectories: 【01 - patient data】 and 【02 - medical record data】, respectively store the doctor-patient conversation data and the corresponding medical record data.

[0103] 2) File naming rules: consistent with the file naming rules of published data.

[0104] This step can be referred to Figure 5 , unified management of training data.

[0105] Step five, unified management of verification data

[0106] Verification data is a subset of data extracted from the same name release directory in proportion.

[0107] Verification data needs to support tracking management according to the release directory name and file name.

[0108] Therefore, define 《Data Definition.xml》 to realize the tracking management of test data:

[0109] 1) Directory naming rules:

[0110] (1) The root directory of verification data is named 【05 - verification data】

[0111] (2) In the above root directory, inherit the project release directory of the published data, such as:

[0112]

xx hospital - daily course - 20250510

[0113]

yy hospital - first course - 20250610

[0114] (3) In the project release directory, inherit two subdirectories: 【01 - patient data】, 【02 - medical record data】, and add 【03, medical record data (from reasoning result)】 to store doctor-patient conversation data, corresponding medical record data, and corresponding medical record-verification data reasoning results.

[0115] 2) File naming rules: consistent with the file naming rules of published data.

[0116] 3) Data verification rules: consistent with the data verification rules of the original data.

[0117] 4) In the training process, the inference results of the verification data are checked according to the data checking rules to generate a checking result.xml, which can be used to evaluate the current training effect and judge whether it is necessary to optimize and continue training.

[0118] This step can be seen in Figure 6 , unified management of verification data.

[0119] Step six, unified management of test data

[0120] Test data is a subset of data extracted in proportion from the same name release directory.

[0121] Test data needs to support tracking management according to the release directory name and file name.

[0122] Therefore, define the data definition.xml to realize the tracking management of test data:

[0123] 1) Directory naming rules:

[0124] (1) The root directory of test data is named 【06-Test data】

[0125] (2) Under the above root directory, inherit the project release directory of the release data, such as:

[0126]

xx hospital-daily course-20250510

[0127]

yy hospital-first course-20250610

[0128] (3) Under the project release directory, inherit two subdirectories: 【01-patient data】, 【02-medical record data】, and add 【03-medical record data (from inference results)】 to store patient specific data, corresponding medical records, and corresponding medical record-test data inference results.

[0129] 2) File naming rules: consistent with the file naming rules of the release data.

[0130] 3) Data checking rules: consistent with the data checking rules of the original data.

[0131] 4) After training, the inference results of the test data are checked according to the data checking rules to generate a checking result.xml, which can be used to evaluate the current training effect and evaluate the training effect.

[0132] This step can be seen in Figure 7 , unified management of test data.

[0133] The above merely describes the preferred embodiments of the present application, which are used for explaining the technical solutions of the present application, and are not used for limiting the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A type-differentiated medical record generation optimization method, characterized by comprising the following steps: 1) Unified management of raw data, which realizes efficient quality management of data through directory classification and directory files; 2) Unified management of derived data, which realizes quality management of raw data tracking to derived data and supports user-selected classification extraction for building and publishing data through directory classification and directory files; 3) Unified management of published data, which realizes quality management of raw data tracking to published data and supports user-selected proportion for building training data, validation data, and test data through directory classification and directory files; 4) Unified management of training data, which realizes quality management of raw data tracking to training data through directory classification and directory files; 5) Unified management of validation data, which realizes quality management of raw data tracking to validation data through directory classification and directory files, and supports data verification rules for reasoning result verification to realize training effect evaluation; 6) Unified management of test data, which realizes quality management of raw data tracking to test data through directory classification and directory files, and supports data verification rules for reasoning result verification to realize test effect evaluation.

2. The method of claim 1, characterized by dividing medical records into several categories based on three dimensions of diagnosis and treatment stages, and clearly defining the core fields and expression specifications of each category of medical records.

3. The method of claim 2, characterized by constructing type-specific generation models: designing different generation algorithms for different types of medical records, using diagnosis-based priority algorithm for the first course of disease, time axis integration algorithm for stage summary, and emergency information priority generation algorithm for daily course of disease; constructing type-specific sub-models through type-specific training and joint fine-tuning to ensure the accuracy of core field generation.

4. The method of claim 3, characterized by extracting test / imaging data to generate diagnosis basis first, and then deducing chief complaint and history of present illness for the first course of disease; 2) integrating data in the order of "admission situation-diagnosis and treatment process-stage guidance" for stage summary; 3) prioritizing the generation of doctor's medication and examination information for daily course of disease, and then supplementing detailed content.

5. The method of claim 3 or 4, characterized by designing specific evaluation dimensions for different types of medical records: 1) evaluating the completeness and logical consistency of diagnosis basis for the first course of disease; 2) evaluating the coherence of diagnosis and treatment process for stage summary; 3) evaluating the accuracy of time record and the consistency of medication for daily course of disease.

6. The method of claim 5, characterized by continuous optimization iteration, based on generation logs, evaluation results, and customer feedback, updating rule library and fine-tuning model parameters, verifying optimization effect through test data set, forming a "optimization-verification-reoptimization" closed loop.

7. The method of claim 1, characterized in that the management activities of raw data, derived data, published data, training data, validation data, and test data respectively provide means to improve quality efficiency, solve the data quality management problem and data on-demand construction problem in medical record generation fine-tuning training. ​ ​ ​ ​ ​ ​ ​ 8. The method of claim 3, wherein, the type-specific generation model is built based on a Transformer architecture, and shares a basic word vector layer but has different coding and decoding layers: the decoding layer of the first course sub-model adopts a reverse generation order of "diagnosis basis-complaints-present history", and the coding layer of the stage summary sub-model adds a "time axis integration module", and the accuracy of the generation of core fields of each sub-model is ensured through type-specific training.

Images