Medical data extraction model training method, medical data extraction method and medium

Abstract

The invention relates to a medical data extraction model training method, a medical data extraction method and a medium, and relates to the technical field of medical information. The method comprises the following steps: capturing unstructured original medical data based on an RPA robot, identifying the unstructured original medical data based on a preset medical data extraction neural network model, and labeling the corresponding unstructured original medical data by taking the identified content as a label to obtain unstructured original medical data labeling data; performing auditing processing on the unstructured original medical data annotation data to obtain unstructured annotation data, and obtaining an unstructured data set based on the unstructured annotation data; and taking the unstructured data set as a training set, and performing training fine adjustment updating on a preset medical data extraction neural network model to obtain an updated medical data extraction neural network model. The medical data labeling efficiency is improved, and the development of a large model is promoted under the condition of less training data.

Description

Technical Field

[0001] The present invention relates to the field of medical information technology, and in particular to a medical data extraction model training, a medical data extraction method and a medium. Background Art

[0002] Medical data comes from a wide range of sources and is complex in type, including structured data such as electronic medical records (EMRs) and unstructured data such as medical images, doctors’ handwritten notes, and laboratory reports. The data is scattered and in different formats, which brings huge challenges to processing and analysis.

[0003] Currently, medical data annotation still relies mainly on manual operations, which is inefficient and error-prone, and cannot meet the growing demand for medical big data analysis. In addition, the lack of unified interfaces and standards between medical systems of different manufacturers further exacerbates the difficulty of data acquisition and integration. Summary of the invention

[0004] The technical problem to be solved by the present application is to provide a medical data extraction model training, a medical data extraction method and a medium, which has the characteristics of improving the efficiency of medical data annotation and promoting the development of large models when there is less training data.

[0005] In a first aspect, an embodiment provides a medical data extraction model training method, comprising: Capture unstructured raw medical data based on RPA robots; Identify the unstructured original medical data based on a preset medical data extraction neural network model, and use the identified content as a label to annotate the corresponding unstructured original medical data to obtain unstructured original medical data annotated data; Reviewing and processing the unstructured original medical data annotation data to obtain unstructured annotation data; Based on unstructured annotated data, obtain unstructured data sets; Based on the unstructured data set as a training set, the preset medical data extraction neural network model is trained, fine-tuned and updated to obtain an updated medical data extraction neural network model.

[0006] In one embodiment, it further includes: The RPA robot captures structured medical data and uses the captured content as labels to annotate the corresponding structured original medical data to obtain structured original medical data annotated data; Based on the structured original medical data annotation data, obtaining structured annotation data; Based on the structured annotated data, obtain the structured data set; The step of obtaining a training set based on the unstructured data set includes: The structured data set and the unstructured data set are integrated to obtain a training set.

[0007] In one embodiment, the step of obtaining structured annotated data based on the structured original medical data includes: The structured original medical data annotation data is reviewed and processed to obtain structured annotation data.

[0008] In one embodiment, the audit process includes: Determine whether the marking position is correct, if not, adjust the marking position; and / or, Determine whether the label is correct, if not, adjust the label.

[0009] In one embodiment, the unstructured raw medical data includes medical text data and medical image data.

[0010] In one embodiment, the obtaining of the unstructured data set includes: taking all the unstructured annotated data as the unstructured data set, or taking a portion of the preset unstructured annotated data as the unstructured data set.

[0011] In a second aspect, an embodiment provides a medical data extraction method, including: extracting structured and unstructured data from original medical data based on an RPA robot, wherein the method for extracting unstructured data includes: Capture unstructured raw medical data based on RPA robots; Identify the unstructured raw medical data based on a preset medical data extraction neural network model; Using the identified unstructured original medical data content as a label, annotating the corresponding unstructured original medical data to obtain unstructured original medical data annotated data; Reviewing the unstructured original medical data annotation data to determine the accuracy of the unstructured original medical data recognition, including determining the accuracy of extracting the recognition object based on the annotation position and / or determining the accuracy of the recognition content based on the label, and adjusting the annotation position and / or label as needed; if there is any adjustment, using the adjusted unstructured original medical data annotation data as the unstructured annotation data; Based on unstructured annotated data, obtain unstructured data sets; Using the unstructured labeled data in the unstructured data set as training data; The training data is used as input to train, fine-tune and update the preset medical data extraction neural network model to obtain an updated medical data extraction neural network model, and unstructured data in the original medical data is extracted based on the updated medical data extraction neural network model.

[0012] In one embodiment, the method for extracting unstructured data further includes: Capture structured raw medical data based on RPA robots; Based on the extracted structured medical data, the extracted content is used as a label to annotate the corresponding structured original medical data to obtain structured original medical data annotated data; Based on the structured original medical data annotation data, obtaining structured annotation data; Based on the structured annotated data, obtain the structured data set; The structured labeled data in the structured dataset is used as training data.

[0013] In one embodiment, the step of obtaining structured annotated data based on the structured original medical data includes: The structured original medical data annotation data is reviewed and processed to obtain structured annotation data.

[0014] In a third aspect, an embodiment provides a computer-readable storage medium, in which a program is stored, and the program can be loaded by a processor to execute the medical data extraction model training method described in any one of the above embodiments, and / or the medical data extraction method described in any one of the above embodiments.

[0015] The beneficial effects of the present invention are: On the one hand, RPA captures unstructured raw medical data, which reduces the manual collection process of raw data and greatly reduces labor costs. On the other hand, the preset medical data extraction neural network model is used to identify the unstructured raw medical data, and the identified content is used as a label to annotate the corresponding unstructured raw medical data to obtain unstructured raw medical data annotated data, which reduces the complex annotation process of the raw medical data. If there are problems with the annotation, the annotated unstructured raw medical data can be used to make quick adjustments and modifications, which greatly saves labor costs. Therefore, the current labor cost of building a data set can be greatly reduced. In the case of less training data, the development of large models can be promoted, and the problem of scattered medical data and non-uniform formats can be compatible. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flowchart of a medical data extraction model training method according to an embodiment of the present application; Figure 2 This is a flow chart of a method for extracting unstructured medical data according to an embodiment of the present application; Figure 3 This is a flowchart of a method for extracting unstructured medical data according to another embodiment of the present application. DETAILED DESCRIPTION

[0017] The present invention is further described in detail below by specific embodiments in conjunction with the accompanying drawings. Wherein similar elements in different embodiments adopt associated similar element numbers. In the following embodiments, many detailed descriptions are for making the present application better understood. However, those skilled in the art can easily recognize that some features can be omitted in different situations, or can be replaced by other elements, materials, methods. In some cases, some operations related to the present application are not shown or described in the specification, this is to avoid the core part of the present application being overwhelmed by too much description, and for those skilled in the art, it is not necessary to describe these related operations in detail, and they can fully understand the related operations according to the description in the specification and the general technical knowledge in the art.

[0018] In addition, the features, operations or characteristics described in the specification can be combined in any appropriate manner to form various implementations. At the same time, the steps or actions in the method description can also be interchanged or adjusted in a manner that is obvious to those skilled in the art. Therefore, the various sequences in the specification and the drawings are only for the purpose of clearly describing a certain embodiment and are not meant to be a required sequence, unless otherwise specified that a certain sequence must be followed.

[0019] To facilitate the description of the inventive concept of this application, the medical information extraction technology is briefly described below.

[0020] Although some current large models such as GPT and Qwen already have strong semantic understanding and generation capabilities, their application in the field of medical information service technology still faces the problem of insufficient professionalism. Currently, various large models applied in the field of medical information service technology still rely heavily on domain-specific high-quality datasets for fine-tuning. However, building these datasets is time-consuming and labor-intensive, which has become a key bottleneck restricting the development of large models.

[0021] In view of this, an embodiment of the present application provides an automated medical data annotation method and medium based on RPA and a large model. In the scheme, on the one hand, unstructured original medical data is captured by RPA, which reduces the manual collection process of original data and greatly reduces labor costs. On the other hand, the unstructured original medical data is identified by a preset medical data extraction neural network model, and the identified content is used as a label to annotate the corresponding unstructured original medical data to obtain unstructured original medical data annotated data, thereby reducing the complex annotation process of the original medical data. If there are problems with the annotation, the already annotated unstructured original medical data can also be used to achieve rapid adjustment and modification, which greatly saves labor costs. Therefore, the current labor cost of constructing a data set can be greatly reduced. In the case of less training data, the development of large models can be promoted, and the problem of scattered and non-uniform formats of medical data can be compatible.

[0022] In one embodiment of the present application, a medical data extraction model training method is provided. Please refer to Figure 1 ,include: Step S10, capturing unstructured raw medical data based on the RPA robot.

[0023] By creating an RPA robot, you can capture unstructured raw medical data from various data sources (such as databases, medical devices, and file systems). The captured unstructured raw medical data can be various text information or image information.

[0024] In one embodiment, the unstructured raw medical data includes medical text data and medical image data.

[0025] Step S20, based on the preset medical data extraction neural network model, the unstructured original medical data is identified, and the identified content is used as a label to annotate the corresponding unstructured original medical data to obtain unstructured original medical data annotated data.

[0026] The preset medical data extraction neural network model can be a built-in neural network large model for calling. The preset medical data extraction neural network model is a pre-trained neural network model for extracting medical data, and some existing medical data extraction models can also be used. The applicant found in the study that in order to obtain a medical data extraction neural network model with high precision and accuracy, on the one hand, due to the dispersion and non-uniform format of medical data, it will bring huge challenges to the analysis, and on the other hand, due to the hugeness and complexity of medical data, as well as local characteristics, it requires a huge labeling work that is seriously time-consuming and labor-intensive. Therefore, the present application first obtains a preset medical data extraction neural network model, which can realize the preset degree of medical data extraction, so that the unstructured original medical data can be identified based on the preset medical data extraction neural network model, and based on the obtained identification content, the identified content can be used as a label to label the corresponding unstructured original medical data, and obtain the unstructured original medical data labeling data, so that the unstructured original medical data can be automatically labeled, thereby greatly reducing the workload of manual labeling.

[0027] Step S30, reviewing and processing the unstructured original medical data annotation data to obtain unstructured annotation data.

[0028] Based on the obtained unstructured original medical data annotation data, an audit process can be performed based on expert knowledge. In some embodiments, it includes determining whether the annotation position is correct (for example, the content paragraph or position of the identified content of the label in the original medical text, or the annotation position in the image data, etc.) and / or whether the label is correct (for example, if the label is a disease classification label, then whether the classification of the identified label is correct or accurate).

[0029] If you want to determine whether the marking position is correct, if the marking position is incorrect, adjust the marking position. If you want to determine whether the label is correct, if the label is incorrect, adjust the label, such as modifying the label category, adding a label category, etc.

[0030] In one embodiment, manual review is performed based on the visualization page, so that the annotation results can be quickly verified and corrected.

[0031] Step S40: acquiring an unstructured data set based on the unstructured annotated data.

[0032] In one embodiment, all unstructured annotated data are used as unstructured data sets to further train, fine-tune and update the preset medical data extraction neural network model.

[0033] In one embodiment, a portion of the preset unstructured annotated data is used as an unstructured data set to further train, fine-tune and update the preset medical data extraction neural network model. For example, in the review process, the unstructured annotated data with adjusted annotation positions and / or labels is often more representative, and the performance of the medical data extraction neural network model updated based on this portion of the annotated data is more significantly improved.

[0034] Step S50, based on the unstructured data set as a training set, the preset medical data extraction neural network model is trained, fine-tuned and updated to obtain an updated medical data extraction neural network model.

[0035] In one embodiment, the preset medical data extraction neural network model can be trained, fine-tuned and updated based on the LoRA (Low-Rank Adaptation) fine-tuning technology.

[0036] LoRA fine-tuning technology is an efficient method for fine-tuning deep learning models. It uses low-rank matrix decomposition technology to fine-tune pre-trained large models to better adapt to tasks in specific fields. Compared with traditional fine-tuning methods, LoRA can achieve rapid adaptation on smaller data sets while retaining the general capabilities of pre-trained models, which is particularly suitable for applications such as the medical field that require a lot of domain knowledge.

[0037] The fine-tuning method of the neural network model based on the LoRA model can be implemented using the existing technical methods. For example, if the preset medical data extraction neural network model of the present application is the neural network architecture of ReNet, the LoRA module can be added to certain convolutional layers of ResNet. These modules are mainly used to update the low-rank matrix of weights. Then set the fine-tuning parameters, determine the rank parameters of LoRA (to determine the size of the low-rank matrix), and at the same time, set the fine-tuning learning rate. Since the parameters updated by LoRA are relatively small, the learning rate can be appropriately set smaller. In each training batch, the training data is input into the medical data extraction neural network model, and the loss function of the output and annotation is calculated. For classification tasks, the cross entropy loss function is commonly used; for regression tasks, such as predicting physiological indicators, the mean square error loss function is commonly used. The parameters and in the LoRA module are updated by the back propagation algorithm, while most of the other parameters of the preset medical data extraction neural network model remain frozen.

[0038] Based on any of the above embodiments, on the one hand, RPA is used to capture unstructured raw medical data, which reduces the manual collection process of raw data and greatly reduces labor costs. On the other hand, the unstructured raw medical data is identified through a preset medical data extraction neural network model, and the identified content is used as a label to annotate the corresponding unstructured raw medical data to obtain unstructured raw medical data annotated data, which reduces the complex annotation process of the raw medical data. If there are problems with the annotation, the annotated unstructured raw medical data can also be used to achieve rapid adjustment and modification, which greatly saves labor costs. Therefore, the solution of any of the above embodiments can greatly reduce the current labor cost of constructing a data set, promote the development of large models when there is less training data, and can be compatible with the problem of scattered medical data and non-uniform formats.

[0039] The medical data extraction model training method based on any one of the above embodiments can obtain a more accurate unstructured medical data extraction model while greatly reducing the workload of human annotation.

[0040] In one embodiment, in order to obtain a more accurate unstructured medical data extraction model or a model that can extract both unstructured data and structured medical data, the medical data extraction model training method also includes: based on the RPA robot to capture structured medical data, and use the captured content as labels to annotate the corresponding structured original medical data to obtain structured original medical data annotated data. Based on the structured original medical data annotated data, structured annotated data is obtained, and based on the structured annotated data, a structured data set is obtained. In this way, the unstructured data set and the structured data set can be integrated to obtain a training set.

[0041] In one embodiment, the structured original medical data annotation data is reviewed and processed to obtain structured annotation data, including determining whether the annotation position is correct, and if not, adjusting the annotation position; and / or determining whether the label is correct, and if not, adjusting the label.

[0042] Similarly, for the acquisition of structured data sets, in one embodiment, all structured annotated data are used as structured data sets. In one embodiment, a preset portion of structured annotated data is used as an unstructured data set. For example, in the review process, the structured annotated data whose annotation position and / or label have been adjusted is often more representative, and the performance of the medical data extraction neural network model updated based on this portion of annotated data is more significantly improved.

[0043] A medical data extraction method provided in an embodiment of the present application includes: extracting structured and unstructured data from original medical data based on an RPA robot.

[0044] Please refer to Figure 2 In one embodiment, a method for extracting unstructured medical data includes: Step S1001, capturing unstructured raw medical data based on the RPA robot.

[0045] Step S1002, identifying unstructured raw medical data based on a preset medical data extraction neural network model.

[0046] Step S1003: Use the identified unstructured original medical data content as a label to annotate the corresponding unstructured original medical data to obtain unstructured original medical data annotated data.

[0047] Step S1004, review and process the unstructured original medical data annotation data.

[0048] The review process includes: judging the accuracy of the recognition of unstructured original medical data, including judging the accuracy of extracting the recognition object based on the annotation position and / or judging the accuracy of the recognition content based on the label, and adjusting the annotation position and / or label as needed; if there is any adjustment, the adjusted unstructured original medical data annotation data will be used as unstructured annotation data.

[0049] Step S1005: acquiring an unstructured data set based on the unstructured annotation data.

[0050] Step S1006: Use the unstructured labeled data in the unstructured data set as training data.

[0051] Step S300, using the training data as input to train, fine-tune and update the preset medical data extraction neural network model to obtain an updated medical data extraction neural network model.

[0052] Step S400, extracting unstructured data from the original medical data based on the updated medical data extraction neural network model.

[0053] Please refer to Figure 3 In another embodiment, the method for extracting unstructured medical data includes: Step S1001, capturing unstructured raw medical data based on the RPA robot.

[0054] Step S1002, identifying unstructured raw medical data based on a preset medical data extraction neural network model.

[0055] Step S1003: Use the identified unstructured original medical data content as a label to annotate the corresponding unstructured original medical data to obtain unstructured original medical data annotated data.

[0056] Step S1004, review and process the unstructured original medical data annotation data.

[0057] The review process includes: judging the accuracy of the recognition of unstructured original medical data, including judging the accuracy of extracting the recognition object based on the annotation position and / or judging the accuracy of the recognition content based on the label, and adjusting the annotation position and / or label as needed; if there is any adjustment, the adjusted unstructured original medical data annotation data will be used as unstructured annotation data.

[0058] Step S1005: acquiring an unstructured data set based on the unstructured annotation data.

[0059] Step S1006: Use the unstructured labeled data in the unstructured data set as training data.

[0060] Step S2001, capturing structured original medical data based on the RPA robot.

[0061] Step S2002: Based on the extracted structured medical data, the extracted content is used as a label to annotate the corresponding structured original medical data to obtain structured original medical data annotated data.

[0062] Step S2003, annotating the data based on the structured original medical data to obtain structured annotated data.

[0063] Step S2004: acquiring a structured data set based on the structured annotation data.

[0064] Those skilled in the art will appreciate that all or part of the structured annotation data may be selected based on demand to form a structured data set.

[0065] In one embodiment, the structured original medical data annotation data is audited to obtain the structured annotation data. The audit process may include: the audit process includes: determining whether the annotation position is correct, if not, adjusting the annotation position; and / or determining whether the label is correct, if not, adjusting the label.

[0066] Step S2005: Use the structured annotated data in the structured data set as training data.

[0067] In this way, the unstructured annotated data in the unstructured data set and the structured annotated data in the structured data set can be integrated into training data.

[0068] Step S300, using the training data as input to train, fine-tune and update the preset medical data extraction neural network model to obtain an updated medical data extraction neural network model.

[0069] Step S400, extracting unstructured data from the original medical data based on the updated medical data extraction neural network model.

[0070] The medical data extraction method based on any of the above-mentioned embodiments, while effectively reducing the cost of manual annotation of training data sets, promotes the development of large models and improves the optimization and processing capabilities of the models, thereby realizing a cycle of data collection → preprocessing → intelligent annotation → review → optimization, and continuously enhancing the closed-loop iterative mechanism of annotation efficiency and model performance.

[0071] In one embodiment of the present application, a computer-readable storage medium is provided, on which a program is stored. The stored program includes a method that can be loaded by a processor and process any of the above embodiments.

[0072] Those skilled in the art will appreciate that all or part of the functions of the various methods in the above-mentioned embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above-mentioned embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, and the storage medium can include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to implement the above-mentioned functions. For example, the program is stored in the memory of the device, and when the program in the memory is executed by the processor, all or part of the above-mentioned functions can be implemented. In addition, when all or part of the functions in the above-mentioned embodiments are implemented by computer programs, the program can also be stored in a storage medium such as a server, another computer, disk, optical disk, flash disk or mobile hard disk, and can be downloaded or copied and saved in the memory of the local device, or the system of the local device is updated, and when the program in the memory is executed by the processor, all or part of the functions in the above-mentioned embodiments can be implemented.

[0073] The above specific examples are used to illustrate the present invention, which is only used to help understand the present invention and is not intended to limit the present invention. For those skilled in the art, according to the concept of the present invention, some simple deductions, modifications or substitutions can be made.

Claims

1. A medical data extraction model training method, characterized in that: include: Capture unstructured raw medical data based on RPA robots; Identify the unstructured original medical data based on a preset medical data extraction neural network model, and use the identified content as a label to annotate the corresponding unstructured original medical data to obtain unstructured original medical data annotated data; Reviewing and processing the unstructured original medical data annotation data to obtain unstructured annotation data; Based on unstructured annotated data, obtain unstructured data sets; Based on the unstructured data set as a training set, the preset medical data extraction neural network model is trained, fine-tuned and updated to obtain an updated medical data extraction neural network model.

2. The medical data extraction model training method according to claim 1, characterized in that: Also includes: The RPA robot captures structured medical data and uses the captured content as labels to annotate the corresponding structured original medical data to obtain structured original medical data annotated data; Based on the structured original medical data annotation data, obtaining structured annotation data; Based on the structured annotated data, obtain the structured data set; The step of obtaining a training set based on the unstructured data set includes: The structured data set and the unstructured data set are integrated to obtain a training set.

3. The medical data extraction model training method according to claim 2, characterized in that: The step of obtaining structured annotated data based on the structured original medical data includes: The structured original medical data annotation data is reviewed and processed to obtain structured annotation data.

4. The medical data extraction model training method according to claim 1 or 3, characterized in that: The review process includes: Determine whether the marking position is correct, if not, adjust the marking position; and / or, Determine whether the label is correct, if not, adjust the label.

5. The medical data extraction model training method according to claim 1, characterized in that: The unstructured original medical data includes medical text data and medical image data.

6. The medical data extraction model training method according to claim 1, characterized in that: The obtaining of the unstructured data set includes: taking all the unstructured annotated data as the unstructured data set, or taking a part of the preset unstructured annotated data as the unstructured data set.

7. A medical data extraction method, characterized in that: include: The RPA robot extracts structured and unstructured data from the original medical data. The unstructured data extraction method includes: Capture unstructured raw medical data based on RPA robots; Identify the unstructured raw medical data based on a preset medical data extraction neural network model; Using the identified unstructured original medical data content as a label, annotating the corresponding unstructured original medical data to obtain unstructured original medical data annotated data; Reviewing the unstructured original medical data annotation data to determine the accuracy of the unstructured original medical data recognition, including determining the accuracy of extracting the recognition object based on the annotation position and / or determining the accuracy of the recognition content based on the label, and adjusting the annotation position and / or label as needed; if there is any adjustment, using the adjusted unstructured original medical data annotation data as the unstructured annotation data; Based on unstructured annotated data, obtain unstructured data sets; Using the unstructured labeled data in the unstructured data set as training data; The training data is used as input to train, fine-tune and update the preset medical data extraction neural network model to obtain an updated medical data extraction neural network model, and unstructured data in the original medical data is extracted based on the updated medical data extraction neural network model.

8. The medical data extraction method according to claim 7, characterized in that: Other methods for extracting unstructured data include: Capture structured raw medical data based on RPA robots; Based on the extracted structured medical data, the extracted content is used as a label to annotate the corresponding structured original medical data to obtain structured original medical data annotated data; Based on the structured original medical data annotation data, obtaining structured annotation data; Based on the structured annotated data, obtain the structured data set; The structured labeled data in the structured dataset is used as training data.

9. The medical data extraction method according to claim 8, characterized in that: The step of obtaining structured annotated data based on the structured original medical data includes: The structured original medical data annotation data is reviewed and processed to obtain structured annotation data.

10. A computer-readable storage medium, characterized in that: The medium stores a program, which can be loaded by a processor and execute the medical data extraction model training method as described in any one of claims 1 to 6, and / or the medical data extraction method as described in any one of claims 7 to 9.

Images