Medical metadata processing and de-identification method and system for large medical models

Abstract

The present invention provides a method and system for processing and de-identifying medical metadata for large medical models, wherein the method comprises: compressing medical metadata based on an initial data distillation model to generate compressed data; training a data distillation model using the compressed data to generate optimized model parameters; continuing to compress the compressed data using the optimized model parameters; and iterating the compression step and the training step alternately until the model parameters converge. The present invention can realize efficient training of models for specific tasks with minimal data volume with almost no information loss, so that the model trained on the concentrated data set performs similarly to the model trained on the original data set. By combining with a differential model, de-privacy informationization can also be realized. The data compressed and reconstructed by the above method is versatile and can support the training of artificial intelligence models of different architectures.

Description

Technical Field

[0001] The present invention relates to the field of big data processing technology, and in particular to a metadata processing and de-identification method and system for a large medical model. Background Art

[0002] In the context of the rapid development of big data and artificial intelligence technology, large models have gradually shown their important clinical significance in the medical field, especially in auxiliary diagnosis, disease prediction and personalized treatment. Internationally, the ChatDoctor model is a large medical model developed by the Massachusetts Institute of Technology. This model can achieve natural dialogue with patients through natural language processing technology, thereby providing accurate medical consultation and services. This model was developed to address the limitations of existing large language models (such as ChatGPT) in medical knowledge. The accuracy of the model has been significantly improved by using 100,000 patient-doctor dialogues from online medical consultation platforms. In addition, the DISC-Med LLM model is a large medical model developed by Stanford University. This model can learn and understand a large amount of medical literature through deep learning technology, thereby providing doctors with more accurate diagnostic references. In order to build a high-quality supervised fine-tuning (SFT) dataset, the DISC-Med LLM model adopts three strategies: using medical knowledge graphs, reconstructing real-world conversations, and combining manually guided preference restatements. These strategies play a key role in training DISC-Med LLM, surpassing the performance of existing medical LLMs in single-round and multi-round consultation scenarios. In China, the Hua Tuo model is a large medical model developed by the Institute of Computing Technology of the Chinese Academy of Sciences. This model can learn and understand a large amount of medical imaging data through deep learning technology, thereby providing doctors with more accurate diagnostic references. The Hua Tuo model is not only based on the open source LLaMa-7B base model, but also integrates structured and unstructured medical knowledge from the Chinese Medical Knowledge Graph (C Me KG).

[0003] Medical metadata refers to the descriptive information of medical data, which helps to interpret, manage and use specific medical data. Medical metadata includes but is not limited to data source, collection time, data type, data quality, data format and standard, data relevance and dependency, and data security and privacy information. Efficient governance of medical metadata is the basic guarantee for the development of large medical models. Whether internationally or domestically, because medical metadata is single disease data and the amount of data is limited, many AI products developed based on the above medical models have the problem of products being applicable to a single scenario and a single function. Large-scale, high-quality medical data sets are the core foundation for the development of large medical models, so how to achieve standardization, homogenization, and systematic governance of multimodal data and multi-center data is imminent.

[0004] The training and construction of medical models require a large amount of data as support. The demand for large-scale data training directly brings about two problems: (1) excessive consumption of storage resources, computing resources, and training time; (2) easy leakage of privacy information. The existence of these objective problems has obviously hindered the sharing of medical data sets that can support the training of large medical models and further hindered the further development of related technologies. Although, in response to the above problems, a variety of data privacy protection sharing methods have been proposed in the prior art, such as differential privacy protection methods, which provide a rigorous mathematical framework for quantifying and controlling the risk of data privacy leakage in random queries or statistical analysis, balancing the needs of data analysis and privacy protection, and becoming an important privacy protection tool for applications involving sensitive data; however, integrating traditional differential privacy methods into model training will reduce the performance of the model and increase the cost of model training. Therefore, how to balance the sharing of high-quality medical data and data privacy protection is also a technical problem that needs to be solved. Summary of the invention

[0005] In view of the above problems, the present invention is proposed to provide a technical solution to overcome the above problems or at least partially solve the above problems.

[0006] The present invention provides a method for processing and de-identifying medical metadata for a large medical model, the method comprising: compressing medical metadata based on an initial data distillation model to generate compressed data; training a data distillation model using the compressed data to generate optimized model parameters; continuing to compress the compressed data using the optimized model parameters; and alternately iterating the compression step and the training step until the model parameters converge.

[0007] Optionally, in the compression step and the training step, the same algorithm model is executed on the medical metadata and the model parameters respectively.

[0008] Optionally, compressing medical metadata based on an initial data distillation model to generate compressed data includes: compressing differential data between first modality medical metadata and second modality medical metadata based on the initial data distillation model; continuing to compress the compressed data using optimized model parameters, including: continuing to compress differential data between compressed first modality medical metadata and second modality medical metadata using optimized model parameters; the medical metadata of non-current task is the second modality medical metadata, and the medical metadata of the current task is the first modality medical metadata.

[0009] Optionally, the differential coefficient of the first modality hospital metadata is determined using a Beta random function.

[0010] Optionally, before compressing the medical metadata, the method also includes: classifying the original medical data into different types of modalities based on a pre-established medical metadata standard; collecting data based on a standard corresponding to the type of modality; extracting feature data from the collected data; performing privacy protection processing on the extracted feature data based on a hierarchical identification principle; labeling the feature data after privacy protection processing; quantitatively managing and storing the labeled feature data to generate medical metadata.

[0011] The present invention also provides a method for processing and de-identifying medical metadata for a large medical model, the method comprising: converting discrete text into a continuous feature space to generate text features; using the text features as conditional input, compressing the medical metadata based on an initial data distillation model to generate compressed data; training the data distillation model using the compressed data and text features to generate optimized model parameters; using the optimized model parameters to continue compressing the compressed data using the text features as conditional input; and iterating the compression step and the training step alternately until the model parameters converge.

[0012] Optionally, the discrete text is converted into a continuous feature space through a pre-trained text model.

[0013] Optionally, the method also includes: using the medical metadata before compression to iteratively train the medical large model, and obtaining the gradient value of each round as the first gradient value; using the medical metadata after compression to iteratively train the medical large model, and obtaining the gradient value of each round as the second gradient value; judging whether the first gradient value and the second gradient value of each round are similar, and if similar, judging that the effects of the medical metadata before and after compression on the training of the medical large model are similar.

[0014] Optionally, during each round of iterative training, a noise is added to the second gradient value to ensure that the compressed medical metadata reaches a predetermined privacy level.

[0015] The present invention provides a medical metadata processing and de-identification system for a large medical model, the system comprising: a distillation compression module, used to compress medical metadata based on an initial data distillation model to generate compressed data; a distillation training module, used to train a data distillation model using the compressed data to generate optimized model parameters; the distillation compression module continues to compress the compressed data using the optimized model parameters; the distillation compression module and the distillation training module alternately and iteratively perform compression operations and training operations until the model parameters converge.

[0016] The present invention provides a medical metadata processing and de-identification system for a large medical model, the system comprising: a text conversion module, used for converting discrete text into a continuous feature space to generate text features; a distillation compression module, used for taking text features as conditional input, compressing medical metadata based on an initial data distillation model, and generating compressed data; a distillation training module, using compressed data and text features to train a data distillation model to generate optimized model parameters; the distillation compression module uses the optimized model parameters to take text features as conditional input to continue compressing the compressed data; the distillation compression module and the distillation training module alternately and iteratively perform compression operations and training operations until the model parameters converge.

[0017] The present invention provides a computer device, which includes a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the medical metadata processing and de-identification method described above is executed.

[0018] The present invention can achieve efficient training of models for specific tasks with minimal data volume with almost no information loss, so that the model trained on the concentrated data set performs similarly to the model trained on the original data set. The data compressed and reconstructed by the above method is versatile and can support the training of artificial intelligence models of different architectures.

[0019] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above technical solution of the present invention and its objectives, features and advantages more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0021] Figure 1 The process of the method for standardization and high-quality governance of multimodal medical metadata is shown;

[0022] Figure 2 A flowchart of a method for processing and de-identifying medical metadata for a large medical model according to an embodiment is shown;

[0023] Figure 3 A flowchart of a method for processing and de-identifying medical metadata for a large medical model according to a second embodiment is shown;

[0024] Figure 4 A flowchart of the verification phase of the medical metadata processing and de-identification method for a medical large model is shown;

[0025] Figure 5 The process of uploading intermediate data to the blockchain during the generation of medical metadata datasets is shown;

[0026] Figure 6 The medical metadata processing method and the verification principle of the original medical metadata dataset are shown;

[0027] Figure 7 A schematic diagram of the structure of a computer device for executing the medical metadata processing and de-identification method for a large medical model proposed in the present invention is shown. DETAILED DESCRIPTION

[0028] The exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present invention and to enable the scope of the present invention to be fully communicated to those skilled in the art.

[0029] The present invention mainly revolves around the idea of ​​data set distillation. In the present invention, data set distillation refers to the use of a large-scale data set to generate a smaller-scale data set, so that it still retains most of the information of the original data set. Specifically, there is a small difference between the performance of the machine learning model trained on the small-scale data set generated by data set distillation and the machine learning model trained on the original data set. This feature reduces the cost of training data transmission, storage and analysis. In addition to maintaining the characteristics of model performance, data set distillation is also conducive to privacy protection, because the generated small-scale data set is visually different from the original data set, and the model trained on the generated small-scale data set has better resistance to privacy attacks such as member reasoning attacks and model reverse attacks. The present invention innovatively proposes to use a non-reversible low-dimensional space embedding method and mathematical model of multivariate heterogeneous data to achieve the concentration and reconstruction of large medical data sets and generate small-scale data sets, thereby providing a data basis for training and fine-tuning various large medical models. The method proposed by the present invention can ensure that the original data cannot be reconstructed from the concentrated and reconstructed small-scale data set or from the model trained using the small-scale data set, thereby protecting the privacy security of the data subject.

[0030] Before data distillation, it is necessary to optimize the metadata governance process of medical metadata, because medical metadata comes from many sources, has many types and modalities, and is updated frequently. The key links of medical metadata governance include standardized collection, data extraction, storage, quality control, privacy processing, annotation and storage of medical metadata. The means of optimizing the governance process of medical metadata include sharing and exchanging metadata based on unified standards, realizing a universal and extended metadata quality control framework based on unified specifications, and realizing an intelligent system for metadata full-process governance based on a unified system.

[0031] As a specific implementation, Figure 1 As shown, the present invention proposes a method for standardizing and high-quality governance of multimodal medical metadata, the method comprising:

[0032] S11. Classify raw medical data into different types of modalities based on pre-established medical metadata standards;

[0033] S12. Collect data according to the standard corresponding to the type of modality;

[0034] S13. Extracting feature data from the collected data;

[0035] S14. Perform quality control on the extracted feature data according to a standardized intelligent process;

[0036] S15. De-privacy processing of quality-controlled data based on the principle of hierarchical identification;

[0037] S16. labeling the feature data after privacy removal;

[0038] S17. Quantitatively manage and store the labeled feature data.

[0039] In the above implementation, medical metadata standards are pre-established for each modality of medical data. During the processing of medical metadata, data is first collected in accordance with the standards, and data extraction is performed based on the knowledge-based understanding of the collected data; then, the extracted data is quality audited (including data denoising) based on the audit process, and the data that meets the quality requirements is graded and de-identified, thereby achieving data privacy processing; then, the privacy-processed data is machine-assisted labeled to complete data labeling; the privacy-processed data is quantitatively managed, thereby establishing a high-quality medical specialty data set.

[0040] The above method can assist doctors in multimodal data labeling, and at the same time build complete medical data collection, quality control and management standards, laying a data foundation for subsequent data set distillation, and ultimately providing strong support for the research and application of large medical models.

[0041] In one aspect of the present invention, a method for processing and de-identifying medical metadata for a large medical model is provided. Figure 2 As shown, the method includes:

[0042] S21. compressing the medical metadata based on the initial data distillation model to generate compressed data;

[0043] S22. Use the compressed data to train the data distillation model to generate optimized model parameters;

[0044] S23. Continue to compress the compressed data using the optimized model parameters;

[0045] S24. The compression step and the training step are iterated alternately until the model parameters converge.

[0046] The above-mentioned medical metadata processing and de-identification method for large medical models can refine large training data sets into small data sets for specific medical data processing tasks or medical data of specific modalities, and can use the small data sets to train the model, which can achieve the effect of training the original data set, and can also achieve the technical effect of privacy protection. The present invention can realize efficient training of models for specific tasks with minimal data volume with almost no information loss, so that the model trained on the concentrated data set performs similarly to the model trained on the original data set. The data compressed and reconstructed by the above method is versatile and can support the training of artificial intelligence models of different architectures.

[0047] As a specific implementation, the model for jointly modeling model parameters and medical metadata is as follows:

[0048]

[0049] Among them, the parameter Θ are model parameters, is the original data, t is the compressed data. By continuously alternating and iterating the optimization, we can obtain Related highly compressed data t , t The amount of data is much smaller than the actual data The amount of data.

[0050] Because the modalities of medical metadata are different, the distillation optimization objectives, training methods, parameterization methods, and learning methods of sample labels are also different, and the present invention does not impose any restrictions on this.

[0051] In order to prevent the training samples from being deduced from the compressed data or the trained medical model and avoid the leakage of privacy information, the present invention innovatively proposes a method for the association distillation of medical multimodal data, and the association of multimodal data can also improve the distillation efficiency. Integrating the probability control of the differential model into the data distillation model can realize the compression processing of multimodal medical metadata, and can achieve controllable multimodal data privacy protection that can resist model reverse attacks. Based on the initial data distillation model, the medical metadata is compressed to generate compressed data, including: based on the initial data distillation model, the differential data of the first modality medical metadata and the second modality medical metadata is compressed; using the optimized model parameters to continue to compress the compressed data, including: using the optimized model parameters to continue to compress the differential data of the compressed first modality medical metadata and the second modality medical metadata; the medical metadata of the non-current task is the second modality medical metadata, and the medical metadata of the current task is the first modality medical metadata.

[0052] As a specific implementation method, the following formulas (3)-(5) are used to replace the above formula (2) to compress data. t Optimization.

[0053]

[0054] In the above algorithm, and These are two types of medical metadata.

[0055] Formula (3) introduces non-current tasks based on formula (2): Data Verify and optimize the compressed data. Formula (4) and Formula (5) use the Beta random function to construct the Vicinity distribution of the verification data, ensuring that the compressed data contains not only the information of the current task, but also the information of other tasks, so as to achieve multimodal data combined with the task model, and to prevent the reconstruction of original data or sensitive information from the trained application model to the greatest extent.

[0056] The above implementation adopts a meta-learning-based method, that is, the optimization goal is set to reduce the loss of the model trained and converged on the synthetic dataset on the original dataset; at the same time, a proxy loss-based method is adopted, such as aligning the weights or output feature distributions of the models trained on the synthetic dataset and the original dataset respectively. Both of the above methods regard the synthetic dataset as an optimizable weight and reduce the corresponding loss of such methods by optimizing the weight. Through the above technical means, the overhead of dataset transmission, storage and model training can be reduced to achieve high-quality sharing of medical data.

[0057] The images in the small-scale synthetic dataset generated by dataset distillation are visually different from those in the original dataset, and the model trained on the small-scale synthetic dataset has good robustness to member inference attacks; other studies have shown that it is difficult for attackers to infer the training samples of the original dataset through the model trained on the small-scale synthetic dataset. Based on the above characteristics of dataset distillation technology, this project will focus on studying the application of dataset distillation technology in multimodal data privacy protection and high-quality sharing.

[0058] For the distillation and concentration of multimodal medical image metadata, most image-based dataset distillation methods cannot be directly applied due to the discrete characteristics of text data. In the present invention, the discrete text data is mapped to a continuous space through a word embedding method, and then the dataset distillation is performed through an image-based dataset distillation method.

[0059] As a specific implementation method, Figure 3 As shown, the medical metadata processing and de-identification method for a large medical model includes the following steps:

[0060] S31. Convert discrete text into continuous feature space to generate text features;

[0061] S32. Taking the text features as conditional input, compressing the medical metadata based on the initial data distillation model to generate compressed data;

[0062] S33. Train the data distillation model using the compressed data and text features to generate optimized model parameters;

[0063] S34. Using the optimized model parameters to take the text features as conditional input to continue compressing the compressed data;

[0064] S35. The compression step and the training step are iterated alternately until the model parameters converge.

[0065] Specifically, a pre-trained text model (such as the BERT model) is used to convert discrete text into a continuous feature space, and then the text features are used as conditional input to control the calculation model parameters. And compressed data By optimizing the compressed image data using text control, you can obtain compressed data that is more conducive to the current task.

[0066] As a specific implementation method, the formula for generating text features is as follows:

[0067]

[0068] By modifying the model formula (2) and formula (3) mentioned above, we get formula (7) and (8), which are as follows:

[0069]

[0070] In order to verify that the compressed medical metadata has the same effect as the uncompressed medical metadata in training the model. Figure 4 As shown, the medical metadata processing and de-identification method for a large medical model proposed by the present invention also includes:

[0071] S41. Iteratively train the model using the medical metadata before compression to obtain a gradient value of each round as a first gradient value;

[0072] S42. Iteratively train the model using the compressed medical metadata to obtain a gradient value of each round as a second gradient value;

[0073] S43. Determine whether the first gradient value and the second gradient value of each round are similar;

[0074] S44. Under similar circumstances, it is determined that the effects of medical metadata before and after compression on model training are similar.

[0075] In step S42, during each round of iterative training, a noise is added to the second gradient value to achieve a predetermined privacy level for the compressed medical metadata.

[0076] Specifically, when given a Datasets of data = {( , )} 1( and Representing data and labels respectively), in order to achieve privacy protection, it is necessary to use data set distillation to generate a distillation set = {( , )} 1 , where < . and Should satisfy the approximate distribution, and the model exist The training results on The training effect on When, the data set The privacy leakage should be within a pre-set privacy level ( , ) and Respectively in the data set and The model trained on , then the data distillation objective can be expressed as follows:

[0077]

[0078] in represents the loss function, such as the cross entropy function. The optimization of formula (9) can be converted into the desired and In each iteration t, similar gradients are generated, namely

[0079]

[0080] Among them, Grad ( ( ); )express ( ) in the dataset About model parameters Gradient ( ( ); )express ( ) Regarding the model parameters on the dataset D gradient.

[0081] Preferably, in each iteration, the Gaussian mechanism is used to achieve a preset privacy level ( , ), that is, for the gradient Grad in formula (10) ( ( ); ) plus a noise, and then optimize formula (10) to iteratively obtain the distilled data set , and then use the obtained distilled dataset Update model weights To obtain the model weights for the next iteration +1.

[0082] In privacy auditing, the actual privacy protection effect of the large model trained with the compressed medical metadata can be estimated through methods such as member inference attack. Figure 5 As shown in FIG. 1 , noise is added to a part of the compressed data set S to generate a data set D', and the data sets D' and D are sampled respectively to obtain samples M(D') and M(D), and the samples M(D') and M(D) are used to verify whether the predetermined privacy level is reached. As a specific algorithm, the judgment Is it greater than or equal to , if the judgment is yes, the privacy verification is passed.

[0083] Verifying whether the training and fine-tuning of medical models using highly compressed data sets is credible is a technical problem that needs to be further solved. The present invention is based on a trusted verification mechanism for the computing process of zero-knowledge proof to ensure the correctness and credibility of the large-model fine-tuning training process after the medical data has been processed through processes such as data distillation and differential privacy, and combines blockchain technology to realize the verification process of the entire life cycle of medical data.

[0084] First, among the methods mentioned above for standardizing and high-quality governance of multimodal medical metadata, Figure 5 As shown in the figure, the intermediate data of each link is stored on the blockchain, so that each link can be verified to ensure the reliability of the data. The intermediate data of each link is stored as a data fingerprint, which can realize the true traceability of the large model training data and ensure the authenticity and integrity of medical data throughout its life cycle.

[0085] Secondly, if Figure 6 As shown, in the medical metadata processing and de-identification method for the medical big model mentioned above, each time the optimized model parameters are generated, a hash value is generated based on the model parameters and submitted to the blockchain.

[0086] According to the type of data distillation, the corresponding on-chain verification contract is selected; the task parameters are generated according to the selected on-chain verification contract, and the hash value is generated based on the model parameters, including: generating a global weighted hash value according to the task parameters and the model weight. When the model parameters converge, the hash value of the compressed data finally generated is submitted to the blockchain.

[0087] The above-mentioned technical means combined with blockchain evidence storage technology can ensure that medical data is processed through predetermined processes such as data distillation and differential privacy, thereby further ensuring the correctness and credibility of the large model fine-tuning training process.

[0088] The present invention provides a medical metadata processing and de-identification system for a large medical model, the system comprising: a distillation compression module, used to compress medical metadata based on an initial data distillation model to generate compressed data; a distillation training module, used to train a data distillation model using the compressed data to generate optimized model parameters; the distillation compression module continues to compress the compressed data using the optimized model parameters; the distillation compression module and the distillation training module alternately and iteratively perform compression operations and training operations until the model parameters converge.

[0089] As a better way of the above system, the distillation compression module compresses the differential data of the first modality medical metadata and the second modality medical metadata based on the initial data distillation model, and uses the optimized model parameters to continue to compress the differential data of the compressed first modality medical metadata and the second modality medical metadata; the medical metadata of the non-current task is the second modality medical metadata, and the medical metadata of the current task is the first modality medical metadata. In this embodiment, the probability control method of the differential model is integrated into the iterative optimization process of the data distillation model, so as to achieve controllable multi-data modality privacy protection that can resist model reverse attacks.

[0090] The present invention provides a medical metadata processing and de-identification system for a large medical model.

[0091] The system includes: a text conversion module, which is used to convert discrete text into a continuous feature space to generate text features; a distillation compression module, which is used to use the text features as conditional input, compress the medical metadata based on the initial data distillation model, and generate compressed data; a distillation training module, which uses the compressed data and text features to train the data distillation model to generate optimized model parameters; the distillation compression module uses the optimized model parameters to use the text features as conditional input to continue compressing the compressed data; the distillation compression module and the distillation training module alternately and iteratively perform compression operations and training operations until the model parameters converge.

[0092] The above system corresponds to the medical metadata processing and de-identification method for large medical models. The above method is implemented in the form of software function modules. The specific details are similar and will not be repeated here.

[0093] like Figure 7 As shown, the computer device includes a processor 710, a communication interface 720 and a memory 730, wherein the processor 710, the communication interface 720 and the memory 730 are connected via a communication bus 740, and a computer program is stored in the memory 730. When the computer program is executed by the processor, the processor completes the aforementioned medical metadata processing and de-identification method for the medical large model.

[0094] As a specific application scenario, the present invention is applicable to chest disease diagnosis systems and liver disease screening and diagnosis systems. At present, the country has established the "National Hepatobiliary Disease Standard Database", "National Multi-center Emergency Imaging Database", and "Lung Lesion Multimodal Imaging Medical Artificial Intelligence Standard Dataset". The data in the above databases include clinical data and imaging data. These clinical data and imaging data can be divided into different types of diseases. When the data from different sources and different types of diseases are annotated, they are based on unified image segmentation and annotation standards. In particular, when annotating image data, tools that meet the annotation quality (in terms of data operation, annotation operation, and business organization) are used. The display of the annotation tool should use a medical professional grayscale display, which needs to meet the DICOM standard and have a resolution of not less than 2M pixels. On the basis of standard annotation, the present invention can perform data distillation and privacy removal based on the data set provided by the above database, so as to generate high-standard, multi-dimensional metadata standards and 10,000 multi-label data sets, so as to support the needs of research and development and application including clinical medical research and artificial intelligence medical devices with high-quality data sets.

[0095] In the description provided herein, a large number of specific details are described. However, it is understood that embodiments of the present invention can be practiced without these specific details. In some instances, well-known methods, structures and techniques are not shown in detail so as not to obscure the understanding of this description.

[0096] Similarly, it should be understood that in order to streamline the present invention and aid in understanding one or more of the various inventive aspects, in the above description of exemplary embodiments of the present invention, the various features of the present invention are sometimes grouped together into a single embodiment, figure, or description thereof. However, this disclosed method should not be interpreted as reflecting the following intention: that the claimed invention requires more features than the features explicitly recited in each claim. More specifically, as reflected in the claims below, inventive aspects lie in less than all the features of the individual embodiments disclosed above. Therefore, the claims that follow the specific embodiment are hereby expressly incorporated into the specific embodiment, with each claim itself serving as a separate embodiment of the present invention.

[0097] It should be noted that the above embodiments illustrate the present invention rather than limit the present invention, and those skilled in the art may design alternative embodiments without departing from the scope of the appended claims. In the claims, any reference symbol between brackets shall not be construed as a limitation to the claims.

Claims

1. A method for processing and de-identifying medical metadata for a large medical model, characterized in that: The method includes: Compressing the medical metadata based on the initial data distillation model to generate compressed data, including: compressing the differential data of the first modality medical metadata and the second modality medical metadata based on the initial data distillation model, the medical metadata of the non-current task is the second modality medical metadata, and the medical metadata of the current task is the first modality medical metadata; The process of generating the medical metadata includes: The original medical data is divided into different types of modalities according to the pre-established medical metadata standards; data is collected according to the standards corresponding to the types of modalities; feature data is extracted from the collected data; privacy is removed from the extracted feature data according to the hierarchical identification principle; the feature data after privacy removal is annotated; the annotated feature data is quantitatively managed and stored to generate medical metadata; Use the compressed data to train the data distillation model and generate optimized model parameters; Continuing to compress the compressed data using the optimized model parameters, including: continuing to compress differential data between the compressed first modality medical metadata and the second modality medical metadata using the optimized model parameters; The compression step and the training step are iterated alternately until the model parameters converge. In the compression step and the training step, the same algorithm model is executed on the medical metadata and the model parameters, respectively.

2. The method for processing and de-identifying medical metadata for a large medical model according to claim 1, characterized in that: The differential coefficients of the first modality medical metadata were determined using the Beta random function.

3. A method for processing and de-identifying medical metadata for a large medical model, The characteristic is that Convert discrete text into continuous feature space to generate text features; Taking the text features as conditional input, compressing the medical metadata based on the initial data distillation model to generate compressed data, including: compressing the differential data of the first modality medical metadata and the second modality medical metadata based on the initial data distillation model, the medical metadata of the non-current task is the second modality medical metadata, and the medical metadata of the current task is the first modality medical metadata; Use compressed data and text features to train the data distillation model and generate optimized model parameters; Using the optimized model parameters to take the text features as conditional input to further compress the compressed data, including: using the optimized model parameters to further compress the differential data of the compressed first modality medical metadata and the second modality medical metadata; The compression step and the training step are iterated alternately until the model parameters converge; The process of generating the medical metadata includes: The original medical data are divided into different types of modalities according to the pre-established medical metadata standards; data collection is performed according to the standards corresponding to the types of modalities; feature data is extracted from the collected data; privacy is removed from the extracted feature data according to the hierarchical identification principle; the feature data after privacy removal is labeled; the labeled feature data is quantitatively managed and stored to generate medical metadata.

4. The method for medical metadata processing and de-identification of a large medical model according to claim 3, characterized in that: The discrete text is converted into a continuous feature space through a pre-trained text model.

5. The method for processing and de-identifying medical metadata for a large medical model according to claim 1 or 3, characterized in that: The method further includes: Iteratively train the large medical model using the medical metadata before compression, and obtain the gradient value of each round as the first gradient value; Iteratively train the large medical model using compressed medical metadata to obtain the gradient value of each round as the second gradient value; It is determined whether the first gradient value and the second gradient value of each round are similar. If they are similar, it is determined that the effects of the medical metadata before and after compression on training the large medical model are similar.

6. The method for medical metadata processing and de-identification of a large medical model according to claim 5, characterized in that: During each round of iterative training, a noise is added to the second gradient value to achieve a predetermined privacy level for the compressed medical metadata.

7. A medical metadata processing and de-identification system for a large medical model, characterized in that: The system includes: A distillation compression module is used to compress medical metadata based on an initial data distillation model to generate compressed data, including: compressing differential data between the first modality medical metadata and the second modality medical metadata based on the initial data distillation model, wherein the medical metadata of a non-current task is the second modality medical metadata, and the medical metadata of the current task is the first modality medical metadata; The medical metadata generation process includes: classifying the original medical data into different types of modalities according to a pre-established medical metadata standard; collecting data according to a standard corresponding to the type of modality; extracting feature data from the collected data; performing privacy protection processing on the extracted feature data according to a hierarchical identification principle; labeling the privacy-protected feature data; quantitatively managing and storing the labeled feature data to generate medical metadata; The distillation training module is used to train the data distillation model using compressed data to generate optimized model parameters; The distillation compression module continues to compress the compressed data using the optimized model parameters, including: continuing to compress the differential data of the compressed first modality medical metadata and the second modality medical metadata using the optimized model parameters; The distillation compression module and the distillation training module alternately and iteratively perform compression operations and training operations until the model parameters converge.

8. A medical metadata processing and de-identification system for a large medical model, characterized in that: The system includes: The text conversion module is used to convert discrete text into continuous feature space and generate Text features; A distillation compression module is used to use text features as conditional input, compress medical metadata based on an initial data distillation model, and generate compressed data, including: compressing differential data between the first modality medical metadata and the second modality medical metadata based on the initial data distillation model, the medical metadata of the non-current task is the second modality medical metadata, and the medical metadata of the current task is the first modality medical metadata; The medical metadata generation process includes: classifying the original medical data into different types of modalities according to a pre-established medical metadata standard; collecting data according to a standard corresponding to the type of modality; extracting feature data from the collected data; performing privacy protection processing on the extracted feature data according to a hierarchical identification principle; labeling the privacy-protected feature data; quantitatively managing and storing the labeled feature data to generate medical metadata; The distillation training module uses compressed data and text features to train the data distillation model and generate optimized model parameters; The distillation compression module uses the optimized model parameters to take the text features as conditional input to continue compressing the compressed data, including: using the optimized model parameters to continue compressing the differential data of the compressed first modality medical metadata and the second modality medical metadata; The distillation compression module and the distillation training module alternately and iteratively perform compression operations and training operations until the model parameters converge.

9. A computer device comprising a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the medical metadata processing and de-identification method according to any one of claims 1 to 6 is executed.

Images