Data processing model training method and device and data processing method and device

Abstract

The embodiment of the invention provides a data processing model training method and device and a data processing method and device.The data processing model training method comprises the steps that reference knowledge data of a target domain is obtained, an initial model is trained according to the reference knowledge data, and a domain model of the target domain is obtained; determining a preset identification field corresponding to the target field, and performing multi-dimensional data sampling on the original sample data of the target field to obtain to-be-processed sample data; extracting a preset identification field and a corresponding field value from the to-be-processed sample data to obtain structured sample data corresponding to the to-be-processed sample data; inputting the structured sample data and the prompt text into a domain model to obtain a prediction result, and training the domain model according to the prediction result and a sample result corresponding to the to-be-processed sample data to obtain a data processing model; based on the prediction identification field, unified and complete structured sample data are extracted from the to-be-processed sample, and the problem of information loss is effectively avoided.

Description

TECHNICAL FIELD

[0001] Embodiments of the present specification relate to the technical field of Internet medical treatment, and in particular to a data processing model training method. One or more embodiments of the present specification also relate to a data processing model training apparatus, a data processing method, a data processing apparatus, a computing device, a computer-readable storage medium, and a computer program product. BACKGROUND

[0002] Coronary heart disease is the most important global disease burden and the leading cause of death. The harm of coronary heart disease mainly comes from serious adverse events caused by disease progression, such as sudden death, myocardial infarction, heart failure, etc. Therefore, strengthening the early accurate identification and individualized intervention of high-risk patients, and timely cutting off the event chain are the key to improving the efficiency of the prevention and treatment system.

[0003] At present, artificial intelligence is used to mine existing big data, realize the transformation from big data to clinical precise risk assessment and treatment decision, and the artificial intelligence model at the present stage is obtained by training the structured case data. However, in the absence of a systematic and perfect data standardization definition set, the structured case data mostly has the defects of insufficient data dimension and coarse granularity, so the artificial intelligence model obtained by training is not generalized enough and has a high illusion rate. Therefore, when the risk assessment and treatment decision are made based on the artificial intelligence model, there is a problem of inaccurate diagnosis. SUMMARY

[0004] Therefore, the embodiments of the present specification provide a data processing model training method, and one or more embodiments of the present specification also relate to a data processing model training apparatus, a data processing method, a data processing apparatus, a computing device, a computer-readable storage medium, and a computer program product to solve the technical defects in the prior art.

[0005] According to a first aspect of the embodiments of the present specification, a data processing model training method is provided, which includes: acquiring reference knowledge data of a target field, training an initial model according to the reference knowledge data, and obtaining a field model of the target field; determining a preset identification field corresponding to the target field, and performing multi-dimensional data sampling on original sample data of the target field to obtain to-be-processed sample data; extracting the preset identification field and the corresponding field value from the to-be-processed sample data to obtain structured sample data corresponding to the to-be-processed sample data; The structured sample data and prompt text are input into the domain model to obtain a prediction result, and the domain model is trained according to the prediction result and a sample result corresponding to the sample data to be processed to obtain a data processing model.

[0006] According to a second aspect of an embodiment of the present specification, a data processing model training apparatus is provided, including: The domain model obtaining module is configured to obtain reference knowledge data of a target domain, train an initial model according to the reference knowledge data, and obtain a domain model of the target domain. The data obtaining module is configured to determine a preset identification field corresponding to the target domain, and perform multi-dimensional data sampling on original sample data of the target domain to obtain sample data to be processed. The structured processing module is configured to extract the preset identification field and a corresponding field value from the sample data to be processed to obtain structured sample data corresponding to the sample data to be processed. The model training module is configured to input the structured sample data and prompt text into the domain model to obtain a prediction result, and train the domain model according to the prediction result and a sample result corresponding to the sample data to be processed to obtain a data processing model.

[0007] According to a third aspect of an embodiment of the present specification, a data processing method is provided, including: Determining sample data to be processed of a target domain and a corresponding preset identification field, extracting the preset identification field and a corresponding field value from the sample data to be processed to obtain structured data corresponding to the sample data to be processed. Inputting the structured data and the prompt text into a data processing model to obtain a data processing result, wherein the data processing model is obtained by training the data processing model.

[0008] According to a fourth aspect of an embodiment of the present specification, a data processing apparatus is provided, including: The structured processing module is configured to determine sample data to be processed of a target domain and a corresponding preset identification field, extract the preset identification field and a corresponding field value from the sample data to be processed to obtain structured data corresponding to the sample data to be processed. The result obtaining module is configured to input the structured data and the prompt text into a data processing model to obtain a data processing result, wherein the data processing model is obtained by training the data processing model.

[0009] According to a fifth aspect of an embodiment of the present specification, a computing device is provided, including: A memory and a processor. The memory is configured to store computer programs / instructions, and the processor is configured to execute the computer programs / instructions, which realize the steps of the data processing model training method and the data processing method.

[0010] According to a sixth aspect of an embodiment of the present specification, a computer readable storage medium is provided, which stores computer programs / instructions, which realize the steps of the data processing model training method and the data processing method when executed by a processor.

[0011] According to a seventh aspect of an embodiment of the present specification, a computer program product is provided, which includes computer programs / instructions, which realize the steps of the data processing model training method and the data processing method when executed by a processor.

[0012] The data processing model training method provided by one embodiment of the present specification, after the domain model is trained by using the reference knowledge data of the target domain, the domain model has professional knowledge in the domain, which lays a foundation for subsequent fine-grained learning; by performing multi-dimensional sampling on the original sample data of the target domain, more diverse sample data to be processed is obtained, which helps to improve the generalization ability of the final data processing model; based on the prediction recognition field corresponding to the target domain, unified, complete and fine-grained information is extracted from the sample to be processed to form structured sample data, which effectively avoids the information loss problem common in traditional structured methods; the structured sample data and the prompt text are input into the domain model, and the model is optimized by using the difference between the prediction result and the sample result, so that the obtained data processing model can understand the fine-grained features in the structured sample data and accurately capture the complex mapping relationship from the structured sample data to the sample result, thereby significantly improving the accuracy of the data processing result obtained by the data processing model in actual application, and enabling the data processing model to be applied to various data. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 is a flowchart of a data processing model training method provided by one embodiment of the present specification; Figure 2 is a flowchart of a data processing method provided by one embodiment of the present specification; Figure 3 is a process flowchart of a data processing model training method provided by one embodiment of the present specification; Figure 4a is a process diagram of structured processing provided by one embodiment of the present specification; Figure 4b is a process diagram of structured processing based on structured rules provided by one embodiment of the present specification; Figure 4c is an example diagram of structured data provided by an embodiment of the present specification; Figure 4d is a process diagram of training a structured model provided by an embodiment of the present specification; Figure 5 is a process diagram of data augmentation provided by an embodiment of the present specification; Figure 6 is a flow diagram of data sampling provided by an embodiment of the present specification; Figure 7 is a structural diagram of a data processing model training apparatus provided by an embodiment of the present specification; Figure 8 is a structural diagram of a data processing apparatus provided by an embodiment of the present specification; Figure 9 is a structural block diagram of a computing device provided by an embodiment of the present specification. DETAILED DESCRIPTION

[0014] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present specification. However, the present specification can be practiced without the specific details, other than in the examples described herein. Those of ordinary skill in the art, and others, can readily ascertain combinations and sub-combinations of the elements of the present specification that are suitable for particular applications without departing from the scope of the present specification. Thus, the following description is not intended to limit the present specification, but to present various examples of it. The description of the present specification is intended only to be illustrative, and is not intended to limit the scope of the present specification.

[0015] The terminology used in one or more embodiments of the present specification is for the purpose of describing particular embodiments only and is not intended to be limiting of one or more embodiments of the present specification. As used in one or more embodiments of the present specification and the accompanying claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in one or more embodiments of the present specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0016] It will be understood that, although the terms first, second, etc. can be used herein to describe various information, these terms are not intended to denote a temporal or chronological order. Rather, these terms are used only as a shorthand notation. For example, in one or more embodiments of the present specification, a first can be termed a second, and, similarly, a second can be termed a first, without departing from the scope of one or more embodiments of the present specification. As used herein, the term "if' can be interpreted to mean "when" or "upon" or "in response to determining" taking into account the context in which the term is used.

[0017] Furthermore, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0018] First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0019] LLM: Large Language Model, a pre-trained language model with an extremely large number of parameters, typically consisting of billions or more parameters, and a neural network model capable of performing various natural language processing tasks.

[0020] CPT: Continue PreTrain, an LLM training technique that continues pre-training.

[0021] SFT: Supervised Fine-Tuning, an LLM training technique for supervised fine-tuning. LoRA: Low-Rank Adaptation of Large Language Models, is a low-rank adaptation technique for fine-tuning LLMs.

[0022] ENTDA: A data augmentation method for generating text based on entities.

[0023] Risk assessment and treatment decisions for high-risk patients with coronary artery disease are primarily based on clinical trial evidence. However, due to limitations in data dimensionality and population representativeness, their extrapolation and accuracy are generally questioned. In recent years, with the accumulation of massive amounts of daily clinical data and breakthroughs in basic research methods, the quantity and dimensionality of data used to construct risk assessment and treatment decision-making models have exploded.

[0024] It is generally accepted that using artificial intelligence to mine existing big data will help build a more comprehensive and accurate integrated risk assessment and treatment decision-making system. However, progress in this area has been very slow to date. Specifically, a key link in realizing the transformation from big data to precise clinical risk assessment and treatment decision-making lies in the practical application of diagnostic models. Existing diagnostic models lack a systematic and complete set of standardized data definitions and professionally labeled datasets for model training. Moreover, most existing datasets suffer from insufficient data dimensionality and coarse granularity, and the correlations between data lack breadth and depth. Therefore, existing diagnostic models generally suffer from limited application scenarios and high illusion rates.

[0025] This specification provides a data processing model training method, and also relates to a data processing model training device, a data processing method, a data processing apparatus, a computing device, a computer-readable storage medium, and a computer program product, which will be described in detail in the following embodiments.

[0026] See Figure 1 , Figure 1 A flowchart of a data processing model training method provided in one embodiment of this specification is shown, which specifically includes the following steps.

[0027] Step 102: Obtain reference knowledge data for the target domain, train an initial model based on the reference knowledge data, and obtain a domain model for the target domain.

[0028] The target domain can be understood as the specific professional field that the initial model needs to focus on mastering, such as fields like medicine, education, and law. Reference knowledge data can be understood as validated professional knowledge data within the target domain, including various knowledge data such as professional classics, academic literature, and guidelines. The initial model can be understood as a general-purpose model with basic language comprehension capabilities; while possessing broad language processing capabilities, it lacks in-depth domain-specific knowledge. The domain model can be understood as an expert-level model acquired through training, capable of deeply understanding domain knowledge.

[0029] Specifically, professional knowledge data in the target domain is acquired and used as training samples to continue pre-training on a general-based initial model. The initial model continuously absorbs professional knowledge data through a self-supervised learning mechanism, thereby injecting authoritative professional knowledge data into the initial model. This significantly improves the accuracy of the domain model's responses in the target domain and effectively suppresses the illusion phenomenon caused by knowledge gaps.

[0030] Step 104: Determine the preset identification field corresponding to the target domain, and perform multi-dimensional data sampling on the original sample data of the target domain to obtain the sample data to be processed.

[0031] Among them, the preset identification field refers to the set of key information elements predefined in the target domain. The preset identification field can be entities and / or attributes contained in the original sample data. For example, when the original sample data is case data, the preset identification field can include medication orders, examination item names, examination results, examination time, blood pressure, etc.

[0032] Raw sample data can be understood as an unprocessed initial data set within the target domain, characterized by diverse sources and heterogeneous formats; multi-dimensional data sampling can be understood as the process of selecting data from multiple dimensions, which may include data source, data result, time, etc.; sample data to be processed can be understood as representative and balanced data obtained after sampling processing.

[0033] Specifically, by systematically sorting out the knowledge system of the target domain, the entities and / or attribute elements that need to be extracted are identified, thereby obtaining standard and comprehensive preset identification fields, which provide a clear basis for subsequent structured processing; in fact, by determining a complete set of preset identification fields, the integrity and accuracy of information extracted from unstructured text can be ensured.

[0034] Taking the medical field as an example, preset identification fields are used to structure case data. Therefore, based on a specific medical knowledge system, key information elements are extracted from different modules contained in the case data. For example, case data includes modules such as medical order information, examination information, laboratory test information, surgical records, progress notes, admission records, and discharge summaries. Preset identification fields defined in the medical order information module may include fields such as medication orders, Holter monitoring, chest X-rays, and thrombolytic therapy. Preset identification fields defined in the examination information module may include fields such as examination item name, examination result, and examination time. Preset identification fields defined in the laboratory test information module may include fields such as examination item name, result value, result value description, reference range, and examination time. These preset identification fields provide a complete information extraction framework for subsequently establishing structured case data. As shown in Table 1, Table 1 illustrates examples of multiple preset identification fields in the medical field provided in the embodiments of this specification.

[0035]

[0036] Table 1 During the data sampling process, the distribution characteristics of the original sample data in different dimensions are considered to avoid model defects caused by data sampling bias. In the implementation process, multiple factors such as the data source, result, and time of the original sample data are taken into account, and data sampling is carried out from multiple dimensions to ensure that the sample data to be processed can cover various key scenarios and solve the problem of data imbalance. This multi-dimensional sampling strategy helps to improve the diversity and representativeness of the sample data to be processed, laying the foundation for building a data processing model with strong generalization ability.

[0037] By defining preset recognition fields corresponding to the target domain, it is possible to perform unified and standardized structured processing on raw data from different sources and in different formats in the target domain. Furthermore, through a multi-dimensional data sampling strategy, the final sample data to be processed can be comprehensive and representative, providing a reliable data foundation for subsequent model training.

[0038] In one or more embodiments described herein, the original sample data of the target domain is first classified to obtain data of different dimensions. Thus, by sampling from data of different dimensions, multi-dimensional data sampling can be achieved to obtain sample data to be processed. Specific implementation methods are as follows: Multi-dimensional data sampling is performed on the original sample data of the target domain to obtain sample data to be processed, including: The original sample data of the target domain is divided into preset dimensions to obtain original sample data of multiple preset dimensions; Data is sampled from the original sample data of the multiple preset dimensions to obtain the sample data to be processed.

[0039] Here, the preset dimensions can be understood as data classification dimensions predefined based on domain knowledge. These preset dimensions reflect the essential characteristics and inherent patterns of the data corresponding to the target domain. Raw sample data can be understood as the initial, unprocessed dataset collected from the target domain. Data sampling can be understood as the process of extracting representative samples from the raw sample data of each dimension.

[0040] Specifically, the original sample data corresponding to the target domain is analyzed, and multiple preset dimensions are identified within the original sample data. The original sample data is then segmented based on these preset dimensions. For example, if the original sample data in the target domain is medical case data, the preset dimensions could be hospital origin, diagnostic pathway, patient demographic characteristics (including age, gender, etc.), etc. By sampling the original sample data contained within each preset dimension, diverse sample data across multiple preset dimensions is ensured for processing.

[0041] Taking the medical field as an example, when sampling data from multiple dimensions for original case data (i.e., original sample data), several key factors need to be considered. These multiple preset dimensions include data source dimension, diagnosis result dimension, demographic characteristic dimension, diagnosis path dimension, and feature integrity dimension.

[0042] For example, in terms of data source, multiple case data can be obtained by sampling from different hospitals at different ratios (e.g., sampling data from three medical institutions in a ratio of 5:3:1). In terms of diagnostic results, the original case data can be divided into positive and negative cases based on whether the diagnosis is positive or negative. This allows for sampling from positive and negative cases at different ratios (e.g., 6:4) to obtain case data with different diagnostic results. Furthermore, case sampling can be based on the severity of the disease. In terms of demographic characteristics, case data can be stratified based on attributes such as age, gender, and region to ensure diverse case data. In terms of diagnostic pathways, diseases can be determined according to different diagnostic pathways such as "symptom-examination-diagnosis" and "risk factors-screening-diagnosis." Therefore, sampling can be conducted along different diagnostic pathways to obtain data based on those pathways. Based on the established case data, and considering the patient's symptoms, examination information, risk factors, and lifestyle characteristics in the original case data, different characteristics can be categorized into necessary characteristics, important characteristics, and auxiliary characteristics according to their importance (e.g., for coronary heart disease, symptoms such as chest pain and shortness of breath, and risk factors such as hypertension and diabetes are necessary characteristics; symptoms such as palpitations, fatigue, and dizziness, and imaging examination information such as echocardiography and coronary angiography are important characteristics; symptoms such as nausea and sweating, and the generation methods such as smoking and diet are auxiliary characteristics). Case data is sampled based on these different types of characteristics. It should be noted that when sampling case data based on different types of characteristics, necessary characteristics must be 100% covered, meaning that all sampled case data must contain necessary characteristics. For important characteristics, it is sufficient to ensure that at least 80% of the sampled case data contains important characteristics. For auxiliary characteristics, it is sufficient to ensure that at least 60% of the sampled case data contains auxiliary characteristics.

[0043] By sampling the original case data through the aforementioned preset dimensions to obtain the sample data to be processed, the final sample data to be processed can not only represent the diagnosis and treatment level of different medical institutions, but also cover the clinical characteristics of various patients, while ensuring the integrity of medical characteristics.

[0044] The data processing model training method provided in the embodiments of this specification can maintain the overall distribution pattern of the original sample data and effectively cover data in more dimensions as much as possible through a multi-dimensional sampling strategy, thus providing a comprehensive and balanced data foundation for subsequent data processing model training.

[0045] Step 106: Extract the preset identification field and its corresponding field value from the sample data to be processed to obtain the structured sample data corresponding to the sample data to be processed.

[0046] Here, "sample data to be processed" refers to representative data obtained after multi-dimensional sampling processing. This sample data is typically unstructured or semi-structured. Field values ​​can be understood as the specific information content in the sample data corresponding to a preset identification field. Structured sample data can be understood as structured data formed through systematic field extraction.

[0047] By using the preset identification fields corresponding to the target domain, information is extracted from the sample data to be processed, thereby realizing the unified and standardized extraction of information distributed in different modules and formats. The field values ​​corresponding to the preset identification fields in the sample data to be processed are obtained, and the result sample data corresponding to the sample data to be processed is obtained by recombining the preset identification fields and the extracted field values.

[0048] By extracting fields from the sample data to be processed, we can not only achieve structured processing of the sample data, but more importantly, establish semantic relationships between data elements to form structured sample data with clear semantic meaning. This provides high-quality input for subsequent data processing model training and realizes the transformation from data to structured knowledge.

[0049] For example, in the analysis scenario of case data corresponding to coronary heart disease, specific values ​​corresponding to preset identification fields are extracted from the sample data to be processed, which includes case modules such as medical order information and examination information; for example, the field value corresponding to the "medication prescribed" field is extracted as "drug a" from the medical order information, and the field value corresponding to the "examination item" field is extracted as "electrocardiogram" and the specific timestamp corresponding to the "examination time" field from the examination information.

[0050] In one or more embodiments of this specification, information extraction from the sample data to be processed is achieved through the collaborative operation of a rule engine and a structured model. Specific implementation methods are described below: Extracting the preset identification field and its corresponding field value from the sample data to be processed to obtain the structured sample data corresponding to the sample data to be processed includes: The preset identification field and its corresponding field value are extracted from the sample data to be processed using a structured model and / or structured rules to obtain the structured sample data corresponding to the sample data to be processed, wherein the structured model is obtained based on augmented data training.

[0051] Among them, the structured model can be understood as an information extraction model, which is used to extract the field values ​​corresponding to the preset identification fields from non-normalized and complex data; the structured rules can be understood as rules for extracting deterministic information based on predefined syntax rules, expressions, etc., which are suitable for data processing with regular format; augmented data can be understood as data enhanced through data amplification, noise injection, semantic reconstruction and other techniques, which are used to improve the generalization ability and robustness of the structured model.

[0052] Specifically, for sample data with standardized format and language, rule-based processing methods using regular expressions and pattern matching can be used to extract the field values ​​corresponding to preset recognition fields in the sample data, ensuring extraction efficiency and accuracy. For complex scenarios requiring semantic understanding and contextual reasoning, structured models trained with augmented data can be used for processing. Through this dual information extraction of structured rules and structured models, the stability advantage of structured rules in standardized scenarios can be preserved, while the adaptability of structured models in complex contexts can be leveraged, forming a complementary information extraction approach.

[0053] For example, when performing structured processing on case data, for regular data in examination information, such as "blood pressure: 120 / 80 mmHg", regular expression rules can be used to directly extract numerical information; while for medical records that require in-depth understanding, such as "the patient had a myocardial infarction three years ago and recently experienced chest pain again", a structured model is used to identify the complex field of "date of last myocardial infarction".

[0054] In other words, structured models are mainly used to extract fields from data that cannot be directly structured by structured rules, have low levels of text structure, complex semantics, or non-standard writing. For example, for fields that are related to understanding (such as smoking history and last myocardial infarction date), structured models extract them by reasoning based on the context of the case. For fields with inconsistent formats (such as lesion site group, target lesion site group, and aspirin group), the text semantics of these fields are complex and the level of structure is low, so structured models can be used for extraction.

[0055] The data processing model training method provided in the embodiments of this specification performs structured processing on different types of sample data to be processed through structured rules and structured models. Furthermore, the structured rules and structured models can be combined to ensure that uniform and standardized structured data is extracted efficiently and accurately from various types of sample data to be processed.

[0056] In one or more embodiments of this specification, before performing structured processing on the sample data to be processed, data preprocessing is required, such as data cleaning and format standardization, so as to perform efficient and accurate structured processing based on the preprocessed target sample data. Specific implementation methods are described below: Before extracting the preset identification field and its corresponding field value from the sample data to be processed using a structured model and / or structured rules, the method further includes: The sample data to be processed is preprocessed to obtain the target sample data; Extracting the preset identification field and its corresponding field value from the sample data to be processed using a structured model and / or structured rules includes: Using the extracted words in the structured rules, the preset identification field and its corresponding field value are extracted from the target sample data, and / or the target sample data is input into the structured model to obtain the preset identification field and its corresponding field value.

[0057] Data preprocessing can be understood as cleaning, standardizing, and formatting the sample data to be processed, aiming to improve data quality and consistency; structured rules contain predefined extraction vocabularies and pattern matching logic, suitable for processing standardized text content; target sample data can be understood as data that has reached the quality standard after data preprocessing, providing standardized input for subsequent information extraction.

[0058] Specifically, the sample data to be processed may contain problems such as garbled characters, extra line breaks, illegal characters, and duplicate values. By preprocessing the sample data, these problems can be solved, and standardized target sample data that can be directly used for information extraction can be obtained.

[0059] In the stage of extracting rules from target sample data using structured rules, structured fields are matched and extracted from the target sample data based on the preset extraction words and pattern matching rules in the structured rules, that is, the preset recognition fields and their corresponding field values ​​are obtained. In the stage of extracting models from target sample data using structured models, the preprocessed target sample data is input into the trained structured model, and complex semantic analysis is performed through the model to obtain the field values ​​corresponding to the preset recognition fields.

[0060] The data processing model training method provided in the embodiments of this specification can obtain high-quality target sample data by preprocessing the sample data to be processed. In this way, when extracting information from the target sample data and obtaining structured sample data, the coverage and accuracy of information extraction can be significantly improved, thereby providing high-quality structured sample data and reducing the cost of manual annotation.

[0061] In one or more embodiments of this specification, the dataset is expanded through data augmentation to address the problems of insufficient or low-quality training data. The structured model described above is then trained based on the augmented data. Specifically, given the text to be labeled and the corresponding initial labeling results in the initial data, data augmentation is performed on the initial labeling results to obtain augmented labeling results. While ensuring the quality of the augmented labeling results, the augmented labeling results and the corresponding predicted augmented text are determined as target data. Thus, augmented data is obtained based on the initial data and the target data. The augmented data is obtained through training using the following steps: Obtain the text to be annotated and the corresponding initial annotation result from the initial data, wherein the initial annotation result includes the preset annotation fields of the text to be annotated and the corresponding field values; Data augmentation processing is performed on the field values ​​in the initial annotation results to obtain augmented annotation results; The enhanced annotation results are input into the text generation model to obtain the predicted enhanced text, and the predicted enhanced text is input into the information extraction model to obtain the predicted annotation results. If the predicted annotation result and the enhanced annotation result are consistent, the enhanced annotation result and the corresponding predicted enhanced text are determined as target data, and the enhanced data is obtained based on the initial data and the target data.

[0062] The initial data can be understood as the original training samples annotated by domain experts, containing the text to be annotated and the corresponding structured initial annotation results. Data augmentation is used to generate new annotation combinations by transforming, rewriting, or optimizing the field values ​​in the annotation results. The text generation model is used to reverse-engineer the structured annotation results into text paragraphs that conform to the form of natural language expression. The information extraction model is used to re-extract structured information from the generated text.

[0063] Specifically, the process involves acquiring the text to be annotated by domain experts and the initial annotation results. Field-level data augmentation is then performed on the initial annotation results to generate new annotation combinations. For example, adding, deleting, or modifying field values ​​in the initial annotation results yields augmented annotation results. A trained text generation model is then used to convert the augmented annotation results into natural language text (i.e., predicted augmented text). Furthermore, an information extraction model is used to reverse-analyze the predicted augmented text to verify its consistency with the original augmented annotation results. If the predicted annotation results obtained through the information extraction model are consistent with the augmented annotation results, target data is obtained based on the generated predicted augmented text and the augmented annotation results. Combining the initial data annotated by domain experts and the target data obtained through augmentation, augmented data for training the structured model is obtained. Based on more abundant and higher-quality augmented data, the generalization ability and performance of the structured model can be improved.

[0064] Taking the lesion site field of coronary heart disease as an example, the initial data includes the angiography text in the surgical record (i.e., the text to be annotated) and the corresponding initial annotation results. For example, the initial annotation results include the field value "site a" corresponding to the lesion site field and the field value "90%" corresponding to the degree of stenosis of the lesion vessel. During the data augmentation stage, new augmented annotation results are generated by modifying and expanding the field values ​​such as "site a" and "90%". For example, the augmented annotation results include the field value "site b" corresponding to the lesion site field and the field value "70%" corresponding to the degree of stenosis of the lesion vessel field.

[0065] The text generation model generates corresponding angiographic text (i.e., predicted enhanced text) based on the enhanced annotation results; the information extraction model then parses the lesion site information and the degree of stenosis of the lesion vessel from the predicted enhanced text to obtain the field values ​​corresponding to the lesion site field and the degree of stenosis of the lesion vessel field (i.e., the predicted annotation results). The predictive annotation results are compared with the annotation results for consistency. If they are consistent, new target data is determined based on the predicted enhanced text and the enhanced annotation results. Then, by combining the initial data and the target data, the enhanced data after data augmentation can be obtained.

[0066] The data processing model training method provided in the embodiments of this specification can significantly improve the diversity and coverage of training data through data augmentation, effectively solving the problem of scarce labeled data in the medical field. At the same time, consistency verification ensures the quality and reliability of the target data, providing sufficient data support for building high-performance structured models.

[0067] In one or more embodiments of this specification, before obtaining augmented data through data augmentation, a text generation model and an information extraction model are trained to ensure the quality of the target data when these two models are applied to obtain the target data. Specific implementation methods are described below: Before inputting the enhanced annotation results into the text generation model to obtain predicted enhanced text, and inputting the predicted enhanced text into the information extraction model to obtain predicted annotation results, the process further includes: The initial annotation results are input into the initial generation model to obtain the predicted generated text. The initial generation model is then trained based on the predicted generated text and the text to be annotated to obtain the text generation model. The text to be labeled is input into the initial extraction model to obtain the predicted extraction result. The initial extraction model is then trained based on the predicted extraction result and the initial labeling result to obtain the information extraction model.

[0068] The initial annotation results can be understood as structured annotation data provided by domain experts, containing the standard answers of preset recognition fields and their corresponding field values; the initial generation model can be understood as a pre-trained model with basic text generation capabilities, whose task is to convert structured data into natural language text that conforms to the domain's expression habits; the initial extraction model can be understood as a model built on the pre-trained model, responsible for accurately identifying and extracting preset structured field information from unstructured text.

[0069] Specifically, the initial annotation results corresponding to the text to be annotated are used as input and fed into an initial generation model with basic text generation capabilities to obtain the natural language description (i.e., predicted generated text) corresponding to the initial annotation results. By comparing the predicted generated text with the text to be annotated, the model parameters are optimized, enabling the trained text generation model to master the ability to convert structured data into domain text.

[0070] When training the initial extraction model, the text to be labeled is input into the initial extraction model to obtain the structured field information corresponding to the text to be labeled (i.e., the predicted extraction result). The predicted extraction result is compared with the initial labeling result labeled by domain experts to learn how to enable the trained information extraction model to master the ability to extract structured field information from domain text.

[0071] After thorough training, the text generation model and the information extraction model possess high-quality text generation and information extraction capabilities, respectively.

[0072] The data processing model training method provided in the embodiments of this specification can not only improve the performance of the text generation model and the information extraction model in the professional field through training, but also establish semantic consistency between text generation and information extraction, providing a reliable technical guarantee for subsequent high-quality data augmentation. Thus, data augmentation can effectively solve the bottleneck problem of scarce training data in the professional field.

[0073] In one or more embodiments of this specification, the sample data to be processed contains lengthy sample data with extensive detailed descriptions. However, due to the length limitations of the input fields in the model, the lengthy sample data can be replaced with corresponding structured sample data, thereby significantly reducing the model input and making the feature information of this lengthy sample data more prominent. Specific implementation methods are described below: Extracting the preset identification field and its corresponding field value from the sample data to be processed to obtain the structured sample data corresponding to the sample data to be processed includes: From the sample data to be processed, determine the redundant sample data, determine the preset structured fields and corresponding field values ​​from the redundant sample data, and obtain the structured sample data corresponding to the redundant sample data, wherein the preset structured fields are determined from the preset identification fields; The redundant sample data in the sample data to be processed is replaced with the structured sample data corresponding to the redundant sample data to obtain the structured sample data corresponding to the sample data to be processed.

[0074] Among them, redundant sample data can be understood as text fragments in the sample data to be processed that are verbose, have low information density but contain important features; preset structured fields can be understood as fields related to redundant sample data that are selected from preset recognition fields.

[0075] Specifically, the process identifies lengthy sample data from the sample data to be processed, and based on domain knowledge, selects pre-defined structured fields related to the lengthy sample data from a pre-defined identification field system. Then, based on these pre-defined structured fields, the field values ​​corresponding to the pre-defined structured fields are accurately extracted from the lengthy sample data to obtain the structured sample data corresponding to the lengthy sample data. The extracted structured data is then used to replace the original lengthy sample data in the sample data to be processed, thus obtaining the structured sample data corresponding to the sample data to be processed. This achieves refined processing of the lengthy sample data, thereby significantly optimizing the information density and processing efficiency of the data while maintaining the integrity of key information.

[0076] In medical settings, case data (i.e., sample data to be processed) typically includes examination information modules and test information modules. The data in these modules are usually lengthy and verbose report texts. Therefore, the data in these modules can be identified as lengthy sample data, and the structured sample data corresponding to the lengthy sample data can be obtained. In this way, the structured sample data corresponding to the lengthy sample data can replace the original lengthy report text.

[0077] The data processing model training method provided in this manual replaces the lengthy sample data in the sample data to be processed with structured data. This can compress the examination and test results that originally required thousands of characters to describe into a set of key information of dozens of characters. This avoids exceeding the model context length limit and makes important features more prominent. This refinement process can significantly improve the domain model's focus on key pathological features, enhance the accuracy and interpretability of diagnostic reasoning, and at the same time greatly reduce the consumption of computing resources.

[0078] Step 108: Input the structured sample data and prompt text into the domain model to obtain the prediction result. Train the domain model based on the prediction result and the sample result corresponding to the sample data to be processed to obtain the data processing model.

[0079] The prompt text can be understood as instructional text designed to guide the model to perform a specific task, mainly for defining the model role, describing the task, and constraining the output format; the domain model can be understood as a basic model pre-trained with domain knowledge, possessing basic cognitive abilities in the professional domain; the prediction result can be understood as the output result generated by the domain model based on the structured sample data corresponding to the sample data to be processed and the prompt text, while the sample result represents the diagnostic result corresponding to the sample data to be processed.

[0080] Structured sample data and carefully designed prompt text are input into the domain model to guide it in generating predictive outputs that meet the task requirements. The loss value is calculated by comparing the prediction results with the sample results, and the parameters of the domain model are adjusted based on the loss value. This allows the model to gradually grasp the mapping relationship from structured features to the final decision, resulting in a data processing model with adjusted parameters.

[0081] In the scenario of auxiliary diagnosis of coronary heart disease, the diagnostic decision model (i.e., the data processing model) is trained by inputting case data (excluding examination and test information), structured examination and test information, and prompt text "Please give a diagnosis and treatment suggestion based on the case information" into the domain model of the medical field. The diagnostic suggestions output by the domain model are compared with the diagnostic solutions given by medical experts.

[0082] The data processing model training method provided in the embodiments of this specification can significantly improve the sensitivity and reasoning ability of the data processing model to structured information, ensuring both the accuracy of diagnostic decisions and enhancing the clinical interpretability of the output results. The final data processing model can effectively support clinical decision-making, significantly improving diagnostic and treatment efficiency while maintaining high accuracy.

[0083] In one or more embodiments of this specification, to ensure the effectiveness and accuracy of the data processing model, continuous optimization of the data processing model is required. That is, after obtaining the data processing model, it is necessary to continuously adjust the data processing model using newly added data in the target domain. Specific implementation methods are described below: After obtaining the data processing model, the following is also included: At preset time intervals, new data in the target domain is acquired, and the preset identification fields and corresponding field values ​​are extracted from the new data to obtain structured sample data corresponding to the new data. The data processing model is adjusted based on the structured sample data corresponding to the newly added data.

[0084] Among them, newly added data can be understood as data that is continuously generated in the target domain and reflects the latest status. It has the characteristics of strong timeliness and the possibility of data distribution drift. The preset time period can be understood as the model update cycle set based on project needs and technical considerations, which can be flexibly configured according to the actual situation.

[0085] Specifically, after obtaining the data processing model, the system can periodically collect new data in the target domain and perform structured processing according to the established preset identification field system to ensure the uniformity and compatibility of the data format. In fact, when new data is obtained, it can be used as the original sample data mentioned above, and then the new data can be processed based on the above processing steps to obtain the structured sample data corresponding to the new data and adjust the data processing model.

[0086] In practical applications, new data in the target domain can be collected daily, the new data can be structured, and the structured sample data corresponding to the new data can be stored in the database. Based on a preset model update cycle (e.g., one month), incremental structured sample data corresponding to the new data can be obtained from the database for incremental training of the data processing model, so as to optimize the parameters and update the knowledge of the data processing model.

[0087] The data processing model training method provided in the embodiments of this specification significantly improves the adaptability of the data processing model to environmental changes through a dynamic periodic update mechanism, effectively overcomes the problem of model performance decay over time, ensures the accuracy and timeliness of the model output results, and provides a reliable guarantee for continuous and stable decision support.

[0088] See Figure 2 , Figure 2 A flowchart of a data processing method provided in one embodiment of this specification is shown, which specifically includes the following steps.

[0089] Step 202: Determine the data to be processed in the target domain and the corresponding preset identification fields, extract the preset identification fields and their corresponding field values ​​from the data to be processed, and obtain the structured data corresponding to the data to be processed.

[0090] Specifically, after obtaining the data to be processed in the target domain, the field values ​​corresponding to the preset identification fields corresponding to the target domain are extracted from the data to be processed to obtain the structured data corresponding to the data to be processed. Subsequently, the obtained structured data is processed using a data processing model to obtain the data processing results.

[0091] In one or more embodiments of this specification, when redundant data exists in the data to be processed, a predicted structured field related to the redundant data is determined from a preset identification field. This allows the acquisition of structured data corresponding to the redundant data based on the preset structured field, and the acquisition of structured data corresponding to the data to be processed is then achieved through replacement. Specific implementation methods are described below: Extracting the preset identification field and its corresponding field value from the data to be processed to obtain the structured data corresponding to the data to be processed, including: From the data to be processed, determine the redundant data, and from the redundant data, determine the preset structured fields and their corresponding field values ​​to obtain the structured data corresponding to the redundant data, wherein the preset structured fields are determined from the preset identification fields; The redundant data in the data to be processed is replaced with the structured data corresponding to the redundant data to obtain the structured data corresponding to the data to be processed.

[0092] For specific implementation methods, please refer to the above embodiments, which are not limited here.

[0093] Step 204: Input the structured data and the prompt text into the data processing model to obtain the data processing result, wherein the data processing model is obtained through training the above data processing model.

[0094] Specifically, in a medical setting, the data to be processed is the data of cases to be diagnosed. The structured data corresponding to the data of cases to be diagnosed, along with the prompt text "Please provide a diagnostic conclusion and treatment suggestion based on the case information," are input into the diagnostic decision model (i.e., the data processing model) to obtain the diagnostic results and treatment methods (i.e., the data processing results) corresponding to the data of cases to be diagnosed.

[0095] The data processing method provided in the embodiments of this specification can realize online diagnosis based on case data and output diagnostic results and treatment methods by utilizing a data processing model. Specifically, based on the defined preset recognition fields, it can perform real-time structured processing on original case data from multiple sources and in multiple formats. Through structured processing, the problems of lengthy case data and unclear important features can be solved, thereby improving the accuracy of data processing results.

[0096] See Figure 3 , Figure 3 A flowchart illustrating the processing procedure of a data processing model training method provided in one embodiment of this specification is shown.

[0097] Taking a data processing model as an example of a diagnostic decision-making model for coronary heart disease, this paper details the training method of this data processing model. Specifically, the case data is structured according to the knowledge system corresponding to coronary heart disease. The structured data obtained by the structured processing contains semantic features. Multi-dimensional data sampling is performed on the structured data, and the diagnostic decision-making model for coronary heart disease is trained based on the sampled data.

[0098] In practice, to address the shortcomings of existing datasets, such as insufficient data dimensions, coarse granularity, and a lack of breadth and depth in the correlation between data, a knowledge system corresponding to coronary heart disease was jointly constructed in collaboration with coronary heart disease experts. This knowledge system was used to determine which case modules needed to be identified for coronary heart disease cases, and which entities / attributes (i.e., fields) needed to be extracted from each case module. For example, based on the knowledge system, the case modules included in coronary heart disease cases were determined to include modules such as medical orders, examination information, laboratory information, surgical records, medical records, medical record cover sheet, admission record, and discharge summary. The entities / attributes corresponding to each case module were also determined, resulting in Table 1 above.

[0099] Once the knowledge system is established, features (i.e., field values) of specific fields can be extracted from case data, enabling the extraction of coronary heart disease features from case data and providing training data for subsequent model training. Specifically, for case data from different medical institutions and in different formats, the knowledge system is used to structure entities, attributes, etc., normalizing the data at the entity and attribute levels.

[0100] See Figure 4a , Figure 4aA schematic diagram of a structured processing procedure provided in one embodiment of this specification is shown.

[0101] Data preprocessing is performed on the original case data (i.e., the original sample data in the above embodiments). After data processing, two structuring methods can be used to obtain the structured data corresponding to the original case data: rule structuring (implemented through structured rules) and model structuring (implemented through structured models).

[0102] For a detailed explanation of the specific implementation of rule structuring, please refer to [link / reference]. Figure 4b , Figure 4b This diagram illustrates a process for structured processing based on structured rules, as provided in one embodiment of this specification. Specifically, the original text is first segmented into multiple short sentences using a specific delimiter for sentence-by-sentence processing. Next, data preprocessing is performed, using regular expressions and specific rules from the medical field to clean and eliminate invalid data, ensuring data quality. Then, regular expressions are used to match and extract key information, thereby locating and extracting target fields (such as rw1, rw2, ..., rwn). After processing each short sentence, the same operation is performed on the next sentence, repeating this process until all short sentences have been processed. Finally, the extracted fields are integrated and output in a structured data format.

[0103] The following details the steps of data preprocessing. Through data integration, structured, unstructured, and semi-structured data from different sources are collected. Text techniques (such as regular expressions) are used to extract key information from unstructured text, and data statistical techniques are applied to define data cleaning strategies and rules. The core data cleaning stage is carried out according to the preset data cleaning strategies and rules to ensure that the processed data conforms to the norms and logic of the medical field. Specific cleaning operations include outlier handling and deduplication to eliminate noise and ensure data consistency and accuracy.

[0104] Rule-based structuring primarily targets cases with a certain standard of language structure and a high degree of structure, using regular expressions and rules to extract information such as examination information and test results.

[0105] The specific implementation of model structuring is explained. Model structuring mainly targets fields of data that cannot be directly structured by rules, have low text structuring, complex semantics, or non-standard writing. These fields include two categories: (1) Understanding fields: such as smoking history, last myocardial infarction date, etc. These fields need to be extracted based on the context of the case. (2) Inconsistent format fields: such as lesion site group, target lesion site group, aspirin group, etc. The text format on which these fields depend is inconsistent. The text semantics on which these fields depend are complex and the structuring is low. Model extraction is used.

[0106] See Figure 4c , Figure 4c An example diagram of structured data provided in one embodiment of this specification is shown. In training a structured model, data augmentation is performed on the original data (i.e., the initial data in the above embodiment) to obtain augmented labeled data (i.e., the augmented data in the above embodiment). Based on the obtained augmented data, a large language model is trained using methods such as SFT / LoRA to obtain a structured model.

[0107] In the application phase of the structured model, patient medical records are received, and the structured model is used to structure data from different sources and in different formats based on the knowledge system, thereby obtaining the structured data corresponding to the medical records.

[0108] Once the structured rules are defined and a structured model is trained, structured data corresponding to the cases can be obtained using two structured processing methods: rule-based structuring and model-based structuring. (See [link to relevant documentation]). Figure 4d , Figure 4d This specification illustrates a process for training a structured model according to one embodiment.

[0109] The following section details the process of data augmentation (ENTDA). See [link to relevant documentation]. Figure 5 , Figure 5 This document illustrates a data augmentation process according to an embodiment of this specification. First, expert-annotated data is obtained. Specifically, the text data corresponding to a field is given to a coronary artery disease expert, who then annotates the value of the corresponding field, thus obtaining the annotation result. For example, taking the lesion site field as an example, this field depends on the angiography text from the surgical record module in the case. The angiography text (i.e., the text to be annotated) is given to the coronary artery disease expert, who then returns the structured field value of the lesion site field (i.e., the initial annotation result).

[0110] The text generation entity model (i.e., the information extraction model in the above embodiment) is trained based on the expert-annotated data (text to be annotated - initial annotation result), and the data is inverted (initial annotation result - text to be annotated) to train the entity generation text model (i.e., the text generation model in the above embodiment).

[0111] Starting from the initial annotation results, the distribution of entity numbers in the initial annotation results is statistically analyzed. Based on the distribution, entity labels are added, deleted, or modified to obtain preliminary data for augmentation (i.e., augmented annotation results). The augmented annotation results are then input into an entity-to-text model to generate text, resulting in predicted augmented text. The predicted augmented text is then input into a text-to-entity model to generate entities, resulting in predicted annotation results. The predicted annotation results and the augmented annotation results are compared to filter the data. If they match, target data is obtained based on the augmented annotation results and the predicted augmented text. The target data and the expert-annotated data are then merged to obtain augmented data after data expansion. Based on the augmented data, a large language model (LLM) is trained using methods such as SFT, LoRA, and multi-task learning to obtain a structured model.

[0112] With structured data available, a data processing model is trained. Specifically, this begins with hierarchical data sampling of the structured data, thereby obtaining diverse data from multiple levels and dimensions. See also... Figure 6 , Figure 6 This specification illustrates a data sampling process according to one embodiment. The goal of data sampling is to ensure that the samples represent the real clinical environment (data representativeness), balance the sample distribution of different diagnostic results (sample balance), ensure the coverage of key diagnostic features (feature integrity), and take into account the time variation of medical practice (time stability).

[0113] In the first layer, sampling is conducted stratified by hospital. Specifically, each hospital is sampled according to the proportion of cases. For example, if three hospitals have case data of 50,000:30,000:10,000, then the hospital sampling ratio is 5:3:1.

[0114] In the second stratification, patient sampling is performed based on the patient's diagnostic results, categorizing cases into positive and negative cases. For example, 45% of positive cases are diagnosed with confirmed coronary artery disease (CAD), 15% are diagnosed as suspected CAD, 30% are diagnosed as excluding CAD, and 10% are diagnosed as excluding cardiovascular disease. Stratification can also be based on disease severity. For instance, among the cases diagnosed with CAD, 50% of cases are selected from acute coronary syndromes (such as elevation myocardial infarction and unstable angina) and stable CAD (such as stable angina, asymptomatic myocardial ischemia, and microvascular myocardial ischemia), respectively. Based on demographic characteristics, the sample can be divided into three age groups: under 45 years old (15%), 45 to 65 years old (50%), and over 65 years old (35%). In terms of gender, the proportions of males and females are 55% and 45%, respectively. In terms of geographical distribution, the sample covers first-tier cities (40%), second-tier cities (35%), and third-tier and below cities (25%), thus comprehensively reflecting the distribution of different population backgrounds.

[0115] In the third layer, sampling is based on the characteristics of the knowledge system, specifically according to the diagnostic pathway. Based on the diagnostic pathway coverage requirements, for the "symptom-examination-diagnosis" pathway, it must cover the diagnostic pathway of "chest pain-ECG examination-myocardial infarction," the diagnostic pathway of "dyspnea-ultrasound-heart failure," and the diagnostic pathway of "asymptomatic-physical examination-discovery of coronary artery disease." Furthermore, for the "risk factors-screening-diagnosis" pathway, the focus is on covering the diagnostic pathway of "hypertension and diabetes-coronary artery examination-coronary artery disease," and it is necessary to cover the diagnostic pathway of "family history-stress test-coronary artery disease." Based on the requirement for completeness of key features, differentiated coverage targets are set for clinical features of different importance: essential features must achieve 100% coverage, including major symptoms such as chest pain and dyspnea, key examinations such as electrocardiogram and myocardial enzymes, and core risk factors such as hypertension, diabetes, and dyslipidemia; important features are required to achieve 80% or more coverage, covering secondary symptoms such as palpitations, fatigue, and dizziness, imaging examinations such as echocardiography and coronary angiography, and laboratory indicators such as blood lipids, blood glucose, and coagulation function; auxiliary features are set with a coverage ratio of no less than 60%, including other symptoms such as nausea and sweating, special examinations such as radionuclide imaging and CT, and lifestyle factors such as smoking, exercise, and diet, thereby systematically ensuring the comprehensiveness and representativeness of the research data.

[0116] In the fourth layer, sampling is based on the time dimension. Specifically, according to the requirements of time dimension sampling, the data for each year is divided into three independent subsets according to the time window: the training set contains 80% of the total data volume for each year and is used for model building and parameter learning; the validation set and the test set each account for 10% of the data volume for each year. The former is used for model tuning and performance evaluation, and the latter is used for independent verification of the final model effect.

[0117] During data sampling, sampling is conducted according to the different levels and dimensions mentioned above, and the sample quality is assessed to obtain high-quality samples. Specifically, the hospital is first selected and quotas are allocated; then, the selected patients are stratified and screened; next, feature integrity checks are carried out; then, sample quality assessment is conducted, and finally, samples that meet the requirements are determined.

[0118] In this embodiment, three levels are set for sample size planning: a minimum sample size of 20,000 cases, a target sample size of 50,000 cases, and a maximum sample size of 100,000 cases. Regarding sample quality control, a data integrity threshold is set, requiring a core feature missing rate of less than 10%. Label accuracy must be reviewed by experts to ensure consistency greater than 95%. Simultaneously, the data must be timely, covering the entire treatment process from admission to discharge. To address the data imbalance problem, various sampling strategies, including oversampling and undersampling, are employed. For oversampling, synthetic minority class samples, adaptive synthetic sampling, and reverse generation of input text by modifying label entities to enhance the training set are used, combined with knowledge-based guided sampling. For undersampling, random undersampling, removal of boundary noise samples, and selective sampling methods based on clinical significance are employed to comprehensively optimize data distribution and improve model training performance.

[0119] Given the training data obtained through sampling, a phased training strategy is employed to construct a diagnostic decision-making model for coronary heart disease (CHD). In the pre-training phase, CHD-related knowledge data is extracted from professional books, literature, and guidelines, and dirty data cleaning and invalid information filtering are performed. Injecting professional knowledge data into the model reduces the illusion of generated content. In the supervised fine-tuning phase, the data obtained from the aforementioned multi-stage sampling is used. The input of each data entry consists of three parts: prompts for role definition, task description, and format constraints; the original case text; and structured examination and testing data. By converting the examination and testing data into structured text, the input length is compressed and key features are highlighted. The model's output is the structured diagnostic result for the corresponding case, including the diagnosis conclusion, diagnostic basis, and treatment plan. In short, by continuing pre-training to learn CHD knowledge, and then through supervised fine-tuning to establish an accurate mapping from patient symptoms, medical history, and other characteristics to diagnostic results and treatment plans, while mastering a standardized output format, a diagnostic decision-making model for CHD is obtained.

[0120] In fact, to ensure the timeliness of data and the continuous optimization capability of the model, this embodiment adopts a daily update frequency at the data update level, supplementing the database with new cases through structured processing to ensure timely synchronization and accumulation of case data; at the model update level, a monthly update strategy is adopted, and the diagnostic decision model is iteratively trained based on incremental new cases, thereby continuously improving the accuracy and clinical adaptability of the model.

[0121] In the online prediction phase, the input cases are first structured based on the knowledge system, extracting coronary heart disease-related features from the cases. Specifically, based on the constructed coronary heart disease knowledge system, the input cases are structured, which can be achieved through rule-based structuring or model-based structuring, as detailed in the above embodiments, which will not be elaborated here. Subsequently, in the diagnostic decision processing, the structured examination and test information, along with the cleaned case text and prompt text, are concatenated to form a complete input. This input is then fed into the diagnostic decision model for prediction, and the model output is obtained. The output is then parsed to generate a structured diagnostic result that includes whether a diagnosis has been made, severity assessment, etiology analysis, and treatment recommendations, which is finally returned to the user.

[0122] The data processing model training method provided in this specification integrates case data from multiple data sources, broadly covering different disease characteristics and patient types. Furthermore, by constructing a coronary heart disease knowledge system, it develops an entity / attribute structuring method combining structured rules and a structured model. This enables real-time structured processing of original cases from multiple sources and in various formats, overcoming the reliance on pre-structured data in traditional methods. Simultaneously, the system supports daily updates of case data and monthly updates of the diagnostic decision-making model, ensuring data timeliness and continuous model optimization.

[0123] Training the model based on multi-source and multi-dimensional case data makes the trained model more generalizable. Furthermore, the structured processing can effectively solve the problems of lengthy original case data and lack of key features, resulting in a significant reduction in the hallucination rate of the trained model. In contrast, existing solutions are mostly limited to data from a single hospital and lack the guidance of a professional knowledge system, which leads to insufficient generalization of their models and a high hallucination rate.

[0124] Corresponding to the above method embodiments, this specification also provides embodiments of a data processing model training device. Figure 7 A schematic diagram of a data processing model training apparatus according to one embodiment of this specification is shown. Figure 7 As shown, the device includes: The domain model acquisition module 702 is configured to acquire reference knowledge data of the target domain, train an initial model based on the reference knowledge data, and obtain the domain model of the target domain. The data acquisition module 704 is configured to determine the preset identification field corresponding to the target domain, and to perform multi-dimensional data sampling on the original sample data of the target domain to obtain sample data to be processed; The structured processing module 706 is configured to extract the preset identification field and the corresponding field value from the sample data to be processed, and obtain the structured sample data corresponding to the sample data to be processed. The model training module 708 is configured to input the structured sample data and prompt text into the domain model, obtain prediction results, train the domain model based on the prediction results and the sample results corresponding to the sample data to be processed, and obtain a data processing model.

[0125] Optionally, the structured processing module 706 is further configured to: The preset identification field and its corresponding field value are extracted from the sample data to be processed using a structured model and / or structured rules to obtain the structured sample data corresponding to the sample data to be processed, wherein the structured model is obtained based on augmented data training.

[0126] Optionally, the structured processing module 706 is further configured to: Obtain the text to be annotated and the corresponding initial annotation result from the initial data, wherein the initial annotation result includes the preset annotation fields of the text to be annotated and the corresponding field values; Data augmentation processing is performed on the field values ​​in the initial annotation results to obtain augmented annotation results; The enhanced annotation results are input into the text generation model to obtain the predicted enhanced text, and the predicted enhanced text is input into the information extraction model to obtain the predicted annotation results. If the predicted annotation result and the enhanced annotation result are consistent, the enhanced annotation result and the corresponding predicted enhanced text are determined as target data, and the enhanced data is obtained based on the initial data and the target data.

[0127] Optionally, the structured processing module 706 is further configured to: The initial annotation results are input into the initial generation model to obtain the predicted generated text. The initial generation model is then trained based on the predicted generated text and the text to be annotated to obtain the text generation model. The text to be labeled is input into the initial extraction model to obtain the predicted extraction result. The initial extraction model is then trained based on the predicted extraction result and the initial labeling result to obtain the information extraction model.

[0128] Optionally, the structured processing module 706 is further configured to: The sample data to be processed is preprocessed to obtain the target sample data; Using the extracted words in the structured rules, the preset identification field and its corresponding field value are extracted from the target sample data, and / or the target sample data is input into the structured model to obtain the preset identification field and its corresponding field value.

[0129] Optionally, the structured processing module 706 is further configured to: From the sample data to be processed, determine the redundant sample data, determine the preset structured fields and corresponding field values ​​from the redundant sample data, and obtain the structured sample data corresponding to the redundant sample data, wherein the preset structured fields are determined from the preset identification fields; The redundant sample data in the sample data to be processed is replaced with the structured sample data corresponding to the redundant sample data to obtain the structured sample data corresponding to the sample data to be processed.

[0130] Optionally, the data acquisition module 704 is further configured to: The original sample data of the target domain is divided into preset dimensions to obtain original sample data of multiple preset dimensions; Data is sampled from the original sample data of the multiple preset dimensions to obtain the sample data to be processed.

[0131] The device further includes: an update module configured to determine new data in the target domain at preset time intervals, extract the preset identification field and its corresponding field value from the new data to obtain structured sample data corresponding to the new data; and adjust the data processing model based on the structured sample data corresponding to the new data.

[0132] The above is an illustrative scheme of a data processing model training device according to this embodiment. It should be noted that the technical solution of this data processing model training device and the technical solution of the data processing model training method described above belong to the same concept. For details not described in detail in the technical solution of the data processing model training device, please refer to the description of the technical solution of the data processing model training method described above.

[0133] Corresponding to the above method embodiments, this specification also provides data processing apparatus embodiments. Figure 8 A schematic diagram of the structure of a data processing apparatus according to one embodiment of this specification is shown. Figure 8 As shown, the device includes: The structured processing module 802 is configured to determine the data to be processed in the target domain and the corresponding preset identification field, extract the preset identification field and the corresponding field value from the data to be processed, and obtain the structured data corresponding to the data to be processed. The result acquisition module 804 is configured to input the structured data and the prompt text into a data processing model to obtain a data processing result, wherein the data processing model is obtained through training the aforementioned data processing model.

[0134] Optionally, the structured processing module 802 is further configured to: From the data to be processed, determine the redundant data, and from the redundant data, determine the preset structured fields and their corresponding field values ​​to obtain the structured data corresponding to the redundant data, wherein the preset structured fields are determined from the preset identification fields; The redundant data in the data to be processed is replaced with the structured data corresponding to the redundant data to obtain the structured data corresponding to the data to be processed.

[0135] The above is an illustrative scheme of a data processing apparatus according to this embodiment. It should be noted that the technical solution of this data processing apparatus and the technical solution of the data processing method described above belong to the same concept. For details not described in detail in the technical solution of the data processing apparatus, please refer to the description of the technical solution of the data processing method described above.

[0136] Figure 9 A structural block diagram of a computing device 900 according to one embodiment of this specification is shown. The components of the computing device 900 include, but are not limited to, a memory 910 and a processor 920. The processor 920 is connected to the memory 910 via a bus 930, and a database 950 is used to store data.

[0137] The computing device 900 also includes an access device 940, which enables the computing device 900 to communicate via one or more networks 960. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 940 may include one or more of any type of wired or wireless network interface (e.g., a network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.

[0138] In one embodiment of this specification, the above-described components of the computing device 900 and Figure 9 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 9 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0139] The computing device 900 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 900 can also be a mobile or stationary server.

[0140] The processor 920 is used to execute the following computer program / instruction, which, when executed by the processor, implements the steps of the above-mentioned data processing model training method and data processing method.

[0141] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, the computing device embodiments are basically similar to the data processing model training method and data processing method embodiments, so the description is relatively simple. Relevant parts can be referred to the descriptions of the data processing model training method and data processing method embodiments.

[0142] An embodiment of this specification also provides a computer-readable storage medium storing a computer program / instructions that, when executed by a processor, implement the steps of the above-described data processing model training method and data processing method.

[0143] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, the computer-readable storage medium embodiments are relatively simple in description because they are fundamentally similar to the data processing model training method and data processing method embodiments; relevant parts can be referred to the descriptions of the data processing model training method and data processing method embodiments.

[0144] An embodiment of this specification also provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the above-described data processing model training method and data processing method.

[0145] The above is an illustrative scheme of a computer program product according to this embodiment. It should be noted that the technical solution of this computer program product belongs to the same concept as the technical solutions of the data processing model training method and the data processing method described above. For details not described in detail in the technical solution of the computer program product, please refer to the description of the technical solutions of the data processing model training method and the data processing method described above.

[0146] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0147] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0148] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0149] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0150] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A data processing model training method, comprising: Obtain reference knowledge data for the target domain, train an initial model based on the reference knowledge data, and obtain a domain model for the target domain; Determine the preset identification field corresponding to the target domain, and perform multi-dimensional data sampling on the original sample data of the target domain to obtain the sample data to be processed; Extract the preset identification field and its corresponding field value from the sample data to be processed to obtain the structured sample data corresponding to the sample data to be processed; The structured sample data and prompt text are input into the domain model to obtain prediction results. The domain model is then trained based on the prediction results and the sample results corresponding to the sample data to be processed to obtain a data processing model.

2. The method as described in claim 1, wherein extracting the preset identification field and its corresponding field value from the sample data to be processed to obtain structured sample data corresponding to the sample data to be processed includes: The preset identification field and its corresponding field value are extracted from the sample data to be processed using a structured model and / or structured rules to obtain the structured sample data corresponding to the sample data to be processed, wherein the structured model is obtained based on augmented data training.

3. The method as described in claim 2, wherein the augmented data is obtained through training via the following steps: Obtain the text to be labeled and the corresponding initial labeling results from the initial data, where, The initial annotation result includes the preset annotation fields and corresponding field values ​​of the text to be annotated; Data augmentation processing is performed on the field values ​​in the initial annotation results to obtain augmented annotation results; The enhanced annotation results are input into the text generation model to obtain the predicted enhanced text, and the predicted enhanced text is input into the information extraction model to obtain the predicted annotation results. If the predicted annotation result and the enhanced annotation result are consistent, the enhanced annotation result and the corresponding predicted enhanced text are determined as target data, and the enhanced data is obtained based on the initial data and the target data.

4. The method as described in claim 3, further comprising, before inputting the enhanced annotation result into the text generation model to obtain the predicted enhanced text, and inputting the predicted enhanced text into the information extraction model to obtain the predicted annotation result: The initial annotation results are input into the initial generation model to obtain the predicted generated text. The initial generation model is then trained based on the predicted generated text and the text to be annotated to obtain the text generation model. The text to be labeled is input into the initial extraction model to obtain the predicted extraction result. The initial extraction model is then trained based on the predicted extraction result and the initial labeling result to obtain the information extraction model.

5. The method as described in claim 2, before extracting the preset identification field and its corresponding field value from the sample data to be processed using a structured model and / or structured rules, further includes: The sample data to be processed is preprocessed to obtain the target sample data; Extracting the preset identification field and its corresponding field value from the sample data to be processed using a structured model and / or structured rules includes: Using the extracted words in the structured rules, the preset identification field and its corresponding field value are extracted from the target sample data, and / or the target sample data is input into the structured model to obtain the preset identification field and its corresponding field value.

6. The method according to any one of claims 1-5, wherein the preset identification field and its corresponding field value are extracted from the sample data to be processed to obtain structured sample data corresponding to the sample data to be processed, comprising: From the sample data to be processed, determine the redundant sample data, determine the preset structured fields and corresponding field values ​​from the redundant sample data, and obtain the structured sample data corresponding to the redundant sample data, wherein the preset structured fields are determined from the preset identification fields; The redundant sample data in the sample data to be processed is replaced with the structured sample data corresponding to the redundant sample data to obtain the structured sample data corresponding to the sample data to be processed.

7. The method according to any one of claims 1-5, wherein multi-dimensional data sampling is performed on the original sample data of the target domain to obtain sample data to be processed, comprising: The original sample data of the target domain is divided into preset dimensions to obtain original sample data of multiple preset dimensions; Data is sampled from the original sample data of the multiple preset dimensions to obtain the sample data to be processed.

8. The method according to any one of claims 1-5, further comprising, after obtaining the data processing model: At each preset time interval, new data in the target domain is determined, and the preset identification field and its corresponding field value are extracted from the new data to obtain the structured sample data corresponding to the new data. The data processing model is adjusted based on the structured sample data corresponding to the newly added data.

9. A data processing method, comprising: Determine the data to be processed in the target domain and the corresponding preset identification fields, extract the preset identification fields and their corresponding field values ​​from the data to be processed, and obtain the structured data corresponding to the data to be processed; The structured data and the prompt text are input into the data processing model to obtain the data processing result, wherein the data processing model is obtained by training the data processing model of any one of claims 1-8.

10. The method of claim 9, wherein extracting the preset identification field and its corresponding field value from the data to be processed to obtain structured data corresponding to the data to be processed includes: From the data to be processed, determine the redundant data, and from the redundant data, determine the preset structured fields and their corresponding field values ​​to obtain the structured data corresponding to the redundant data, wherein the preset structured fields are determined from the preset identification fields; The redundant data in the data to be processed is replaced with the structured data corresponding to the redundant data to obtain the structured data corresponding to the data to be processed.

11. A data processing model training device, comprising: The domain model acquisition module is configured to acquire reference knowledge data of the target domain, train an initial model based on the reference knowledge data, and obtain the domain model of the target domain. The data acquisition module is configured to determine the preset identification field corresponding to the target domain, and to perform multi-dimensional data sampling on the original sample data of the target domain to obtain the sample data to be processed; The structured processing module is configured to extract the preset identification field and its corresponding field value from the sample data to be processed, thereby obtaining the structured sample data corresponding to the sample data to be processed. The model training module is configured to input the structured sample data and prompt text into the domain model, obtain prediction results, and train the domain model based on the prediction results and the sample results corresponding to the sample data to be processed, thereby obtaining a data processing model.

12. A data processing apparatus, comprising: The structured processing module is configured to determine the data to be processed in the target domain and the corresponding preset identification fields, extract the preset identification fields and their corresponding field values ​​from the data to be processed, and obtain the structured data corresponding to the data to be processed. The result acquisition module is configured to input the structured data and the prompt text into a data processing model to obtain a data processing result, wherein the data processing model is obtained by training the data processing model of any one of claims 1-8.

13. A computing device, comprising: Memory and processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions, which, when executed by the processor, implement the steps of the method according to any one of claims 1 to 10.

14. A computer-readable storage medium storing a computer program / instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 10.

15. A computer program product comprising a computer program / instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 10.

Images