Metadata-based artificial intelligence model management system

By combining vector concatenation of continuous and discrete metadata with the fusion processing of medical image data, along with diffusion models and knowledge graphs, the problem of heterogeneous data source fusion in existing technologies has been solved, achieving efficient medical data management and prediction capabilities.

WO2026232768A1PCT designated stage Publication Date: 2026-11-12CHINA NAT INST OF STANDARDIZATION
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/CN2025/131523
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-08
Filing Date
2025-10-31
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Existing metadata management and AI model training methods struggle to efficiently integrate heterogeneous data sources when dealing with complex medical data. They neglect the richness and completeness of the data, leading to information loss and limited model expressive power, especially with significant performance bottlenecks in multimodal data scenarios.

Method used

By collecting continuous and discrete metadata, vector concatenation and linear mapping are performed to generate latent vectors of uniform length; combined with medical image data, a diffusion model is constructed using the ControlNet control module and U-Net network for inverse denoising training; the diffusion model is deployed for data standardization and parameter aggregation, and a knowledge graph is used to record operation logs and query statement transformations.

Benefits of technology

It enables refined processing of different types of metadata, enhances the model's generalization ability and prediction accuracy, improves the diversity and flexibility of medical data generation, and enhances the efficiency and transparency of data querying.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2025131523_12112026_PF_FP_ABST
    Figure CN2025131523_12112026_PF_FP_ABST
Patent Text Reader

Abstract

A metadata-based artificial intelligence model management system, comprising: a metadata processing module, configured to collect metadata that includes continuous metadata and discrete metadata, calculate the center point of each field of the continuous metadata and encode same into a label distribution vector, count the total number of categories of the discrete metadata, perform encoding to construct a mapping matrix, perform vector concatenation, and use a linear transformation matrix to perform linear mapping for transformation into a latent vector of a uniform length.
Need to check novelty before this filing date? Find Prior Art

Description

A metadata-based artificial intelligence model management system

[0001] This application claims priority to Chinese patent application No. CN202510585066.X, filed on May 8, 2025, entitled “An Artificial Intelligence Model Management System Based on Metadata”, the entire contents of which are incorporated herein by reference. Technical Field

[0002] This disclosure relates to the field of model management technology, and in particular to an artificial intelligence model management system based on metadata. Background Technology

[0003] Metadata, as the "data" that describes data, includes information such as data source, data type, data relationship, and data processing method. With the continuous advancement of big data technology, enterprises and research institutions are increasingly relying on efficient metadata management systems when processing massive amounts of information. Furthermore, in the medical field, prediction and analysis based on metadata are gradually being proposed and used. These systems can not only improve the efficiency of data query, but also reveal the deeper value hidden behind the data through the correlation analysis of metadata.

[0004] However, existing metadata management and AI model training methods still have certain shortcomings. Most current mainstream technologies focus on the processing and modeling of single-type metadata, and often fail to adequately refine the handling of differences between continuous and discrete metadata. In particular, it is difficult to efficiently capture the potential relationships between categories in medical information, resulting in information loss and limited model expressive power. In addition, existing methods usually ignore how to efficiently integrate heterogeneous data sources. Especially in the case of multimodal data in the medical field, model training is difficult to consider the richness and completeness of data at the same time. Therefore, existing technologies face significant performance bottlenecks and application limitations when dealing with complex data types and large-scale data. Summary of the Invention

[0005] According to some embodiments, a first aspect of this disclosure provides a metadata-based artificial intelligence model management system, comprising:

[0006] The metadata processing module collects metadata, including continuous and discrete metadata. It calculates the field centroid encoding for continuous metadata into a label distribution vector, counts the total number of categories for discrete metadata, encodes and constructs a mapping matrix, concatenates vectors, and uses a linear transformation matrix to perform linear mapping to transform them into latent vectors of uniform length.

[0007] The medical data processing module extracts the latent vectors of medical images and fuses them with unified latent vectors to generate conditional vectors. Based on external structural features, it introduces the ControlNet control module to generate control vectors, which are then fused with the image latent vectors to generate fused latent vectors. Noise is gradually added, and a diffusion model is built based on the U-Net network for reverse denoising training to generate restored latent vectors.

[0008] The model deployment module deploys and diffuses models on the client side to standardize data, aggregate model parameters, update the global model, and evaluate the global model.

[0009] The query module analyzes the uncertainty of the global model, defines the knowledge graph, records model operation data, uploads operation logs, and transforms user-input queries for data retrieval.

[0010] In one optional embodiment of this disclosure, the step of performing vector concatenation, using a linear transformation matrix to perform a linear mapping to transform the vectors into latent vectors of uniform length, includes:

[0011] For each field of continuous metadata, calculate the center point of each field based on the variable range and score range of the field's sample values;

[0012] Construct a Gaussian kernel label distribution function to encode its values ​​into a vector, and map the input value x of the field to a label distribution vector;

[0013] For each field of discrete metadata, count the number of unique values ​​of all samples in the field as the total number of categories C, and perform embedding encoding to construct a mapping matrix with dimensions C×d, where d is a fixed embedding dimension;

[0014] Each row represents the embedding vector of a category, resulting in the embedding vector representation of each discrete variable;

[0015] The concatenated label distribution vector and the embedded vector representation of discrete metadata are used. The output vector of all continuous variables has dimension p, and the embedded vector of all discrete variables has dimension d. This is based on the total number of continuous variables N. c and the total number of discrete variables N d Define the dimension D after concatenation;

[0016] The encoded vectors of all fields are concatenated sequentially into a single vector, and then a linear transformation matrix is ​​used to linearly map and transform it into a latent vector of uniform length.

[0017] In one optional embodiment of this disclosure, the generation of the control vector, which is fused with the unified latent vector of metadata to generate a condition vector, includes:

[0018] Based on the corresponding medical image data of clinical data, the image data is paired with the unified latent vector of metadata, and the image latent vector of medical image data is extracted using a pre-trained image encoder and fused with the unified latent vector of metadata to generate a conditional vector.

[0019] To determine the external structural features of medical image data, including keypoints, labels, and structural information, a ControlNet module is introduced. This module takes the conditional vector and external structural features as input to generate a control vector z. p And it is fused with the image latent vector to generate a fused latent vector;

[0020] The forward diffusion process is based on the fusion latent vector by gradually adding noise;

[0021] A diffusion model is constructed based on the U-Net network, and inverse denoising training is performed to recover the latent vector of the original image from the noisy state. The mean squared error loss (MSE) is used to train the model to predict noise.

[0022] The Adam optimizer is used for gradient descent optimization to update the model's parameters. The iteration stops when the model's loss no longer decreases significantly during continuous iterations.

[0023] The restored latent vector is generated by the trained diffusion model based on the noise input and condition vector generation conditions in the latent space.

[0024] In one optional embodiment of this disclosure, the aggregation model parameters update the global model, including:

[0025] Collect customer metadata based on the client terminal, standardize the metadata of each client using the Meta-Calibrator, deploy the diffusion model and train the model using local data;

[0026] For each client, a weighting coefficient is determined, and a weighted average method is used to aggregate the model parameters of all clients. The global model is then updated based on the global model parameters.

[0027] In one optional embodiment of this disclosure, the evaluation of the global model includes,

[0028] Perform FID and SSIM evaluations on the global model;

[0029] The difference threshold is determined by the sum of the mean and standard deviation of the historical distribution difference values ​​between the generated image and the real image. If the distribution difference value is less than or equal to the difference threshold, it is determined that the original features of the image are preserved.

[0030] The structural threshold is determined by the sum of the mean and standard deviation of the historical structural differences between the generated image and the real image. If the structural difference is greater than or equal to the structural threshold, the original features of the image are preserved.

[0031] In one optional embodiment of this disclosure, the analysis of model uncertainty in the global model includes,

[0032] A resampling mechanism (LAS) is introduced for the global model. The latent vector and conditional vector of the input image are randomly initialized multiple times, and the corresponding resampling restored latent vector is obtained. The average predicted value and average standard deviation are calculated to generate an uncertainty score.

[0033] The scoring threshold is determined by the sum of the mean and two standard deviations of the historical uncertainty scores. If the uncertainty score is less than the scoring threshold, it is judged to be of low uncertainty.

[0034] In one optional embodiment of this disclosure, the collected metadata includes continuous metadata and discrete metadata, including,

[0035] By traversing each column of the original clinical data table, the value type and value distribution range of each column are read. If the field value is a continuous numeric type, it is marked as continuous metadata; if the field value is a discrete category label, it is marked as discrete metadata. Each row in the clinical data table is used as a sample.

[0036] In one optional embodiment of this disclosure, defining the knowledge graph record model operation data refers to using the DataHub architecture to define the triple structure of the knowledge graph KG, recording the model's input / output mapping, data source, processing steps, model parameters, and execution environment information.

[0037] In one optional embodiment of this disclosure, the upload operation log refers to recording operation logs for each operation step of the model and storing them as structured metadata.

[0038] In one optional embodiment of this disclosure, the step of converting the user-input query and performing data query includes:

[0039] The BERT large model is used to convert user-input queries into SPARQL statements to query metadata in the knowledge graph;

[0040] Based on the user's query, the system will return detailed information on the corresponding model path, inference conditions, and input data.

[0041] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 is a schematic diagram of the structure of the metadata-based artificial intelligence model management system in Example 1.

[0044] Figure 2 is a flowchart of the metadata-based artificial intelligence model management system in Example 1. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this disclosure clearer, the following detailed description of a metadata-based artificial intelligence model management system, through embodiments and in conjunction with the accompanying drawings, is provided. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this disclosure.

[0046] The serial numbers assigned to components in this document, such as "first" and "second," are used solely to distinguish the described objects and have no sequential or technical meaning. Unless otherwise specified, the terms "connection" and "linkage" used in this disclosure include both direct and indirect connections (linkages). It should be understood that the terms "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are used solely for the convenience of describing this disclosure and for simplification. They do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this disclosure.

[0047] In this disclosure, unless otherwise expressly specified and limited, "above" or "below" the second feature can mean that the first and second features are in direct contact, or that the first and second features are in indirect contact through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0048] Example 1, referring to Figures 1 and 2, is the first embodiment of this disclosure. This embodiment provides a metadata-based artificial intelligence model management system, including the following steps:

[0049] S1. Collect metadata including continuous metadata and discrete metadata. Calculate the field center point encoding of continuous metadata into a label distribution vector. Calculate the total number of categories of discrete metadata, encode and construct a mapping matrix, concatenate the vectors, and use a linear transformation matrix to perform linear mapping to transform it into a latent vector of uniform length.

[0050] Preferably, the collected metadata includes continuous metadata and discrete metadata, including,

[0051] By traversing each column of the original clinical data table, the value type and value distribution range of each column are read. If the field value is a continuous numeric type, it is marked as continuous metadata; if the field value is a discrete category label, it is marked as discrete metadata. Each row in the clinical data table is used as a sample.

[0052] By traversing the original clinical data tables and classifying them according to the value type and value distribution range of each column, the data can be effectively and automatically labeled as continuous or discrete metadata. This helps to clearly identify the data structure and provides accurate data type identification for subsequent data processing, modeling, and analysis, ensuring consistency and efficiency in the processing process. At the same time, it helps the model to adopt appropriate processing methods according to the characteristics of different types of data, thereby improving the model's performance and accuracy.

[0053] Furthermore, vector concatenation is performed, using a linear transformation matrix to linearly map and transform the vectors into latent vectors of uniform length, including...

[0054] For each field of continuous metadata, the centroid of each field is calculated based on the variable range and score range of the field's sample values, as shown below:

[0055] Where, μ i Let represent the center point of the i-th Gaussian kernel, n represent the number of segments, b represent the maximum value of the variable range, and a represent the minimum value of the variable range.

[0056] Construct a Gaussian kernel label distribution function to encode its values ​​as a vector, mapping the input value x of the field to a label distribution vector, represented as:

[0057] Among them, v i This represents the label distribution vector of the i-th field, where σ represents the smoothing factor, which can be calculated using the formula... This is determined to ensure consistent coverage width across the distribution.

[0058] For each field of discrete metadata, count the number of unique values ​​of all samples in the field as the total number of categories C, and perform embedding encoding to construct a mapping matrix with dimensions C×d, where d is a fixed embedding dimension;

[0059] Each row represents the embedding vector of a category, thus obtaining the embedding vector representation of each discrete variable. For each sample, the category index j is obtained by looking up the table, and the j-th row is extracted as the embedding vector.

[0060] The concatenated label distribution vector and the embedded vector representation of discrete metadata are used. The output vector of all continuous variables has dimension p, and the embedded vector of all discrete variables has dimension d. This is based on the total number of continuous variables N. c and the total number of discrete variables N d Define the concatenated dimension D as: D = N c ·p+N d ·d;

[0061] The encoded vectors of all fields are concatenated sequentially into a single vector. A linear transformation matrix is ​​then used to linearly map this vector to a uniform-length latent vector. This vector represents the metadata conditions and will be passed into the AI ​​model's control path, represented as follows:

[0062] in, This represents the original vector after concatenating all metadata, with a length of D. It means, v m The unified latent vector representation of metadata is given, where W represents the linear transformation matrix, which can be initialized using a standard normal distribution and then fixed during training.

[0063] By performing Gaussian kernel label distribution encoding for continuous metadata and embedding encoding for discrete metadata, the full range of variables can be covered through detailed segmentation, avoiding information loss caused by single numerical input. This allows the model to learn the potential patterns of the data more accurately when processing these features. For discrete data, embedding encoding effectively reduces dimensionality while preserving category information by mapping each category to a dense vector space, thereby improving the model's generalization ability and prediction accuracy.

[0064] By using a unified metadata latent vector, different types of features can be fused into the same latent space, while ensuring the comparability of features. The relative relationships between features can be captured by the model, thus providing the model with more comprehensive conditional inputs. This latent vector representation not only improves the integration of data but also enhances the model's comprehensive understanding of various input features, thereby further improving the performance of generation or prediction tasks.

[0065] S2: Extract the latent vector of the medical image and fuse it with the unified latent vector to generate a conditional vector. Based on the external structural features, the ControlNet control module is introduced to generate a control vector, which is then fused with the image latent vector to generate a fused latent vector. Noise is gradually added and a diffusion model is built based on the U-Net network. Reverse denoising training is performed to generate the restored latent vector.

[0066] Preferably, a control vector is generated and fused with the unified latent vector of metadata to generate a condition vector, including:

[0067] Based on the corresponding medical image data from the clinical data, the image data is paired with the unified latent vector of the metadata. A pre-trained image encoder is used to extract the image latent vector from the medical image data, and this vector is fused with the unified latent vector of the metadata to generate a conditional vector, represented as: z c =γ·z i +(1+γ)·v m ;

[0068] Among them, z c Let z represent the conditional vector. i γ represents the latent vector of the image output by the image encoder, and γ represents the fusion coefficient, which is determined based on historical experience.

[0069] To determine the external structural features of medical image data, including keypoints, labels, and structural information, a ControlNet module is introduced. This module takes the conditional vector and external structural features as input to generate a control vector z. p It is then fused with the image latent vector to generate a fused latent vector, denoted as: z co =β·z i +(1-β)·z p ;

[0070] Among them, z co β represents the fusion latent vector, and β represents the control weight, which is obtained by training and optimization based on calibrated training data;

[0071] Based on the fusion of latent vectors, a forward diffusion process is performed by gradually adding noise, making each image latent vector more blurred, as shown below:

[0072] Among them, z co , t and z co , t-1 Let a represent the fusion latent vectors at time steps t and t-1, respectively. t Let represent the attenuation factor at step t (determined based on historical experience), ∈ represent standard Gaussian noise with a covariance equal to the identity matrix;

[0073] A diffusion model is constructed based on the U-Net network, and inverse denoising training is performed to recover the latent vectors of the original image from the noisy state. The mean squared error loss (MSE) is used to train the model to predict noise, as shown below:

[0074] Among them, L n Indicates the calculation of loss, ∈ θ This represents the U-Net network model used for noise prediction;

[0075] The Adam optimizer is used for gradient descent optimization to update the model's parameters. The iteration stops when the model's loss no longer decreases significantly during continuous iterations.

[0076] Based on the noise input in the latent space and the conditional vector generation conditions, the restored latent vector is generated by the trained diffusion model and reconstructed into image data by the decoder.

[0077] By pairing medical data with unified latent vectors of metadata, the generated conditional vectors can provide the model with more comprehensive input information, enabling the model to consider both the latent features of medical data and the individualized metadata of patients. By controlling the conditional vectors through the ControlNet module, the model's fine-grained control over medical data can be further enhanced. Especially in medical data generation tasks, controlling the influence of external structural features can help the model generate medical data that is more in line with medical knowledge, such as accurately reconstructing the anatomical structure of specific parts or adjusting certain features of medical data (such as lesion areas). Through optimized control weights, the model can more accurately adjust the generation process based on feedback from training data, ensuring that the generated images are more in line with clinical needs.

[0078] By performing a forward diffusion process on the fused latent vectors and adding noise, the model learns how to gradually recover from a fuzzy state, thus enabling it to generate higher-quality medical data when faced with unknown or fuzzy inputs. The introduction of the diffusion model can also further enhance the diversity of generated medical data and allow the model to generate different versions of medical data, thereby enhancing the model's exploratory capabilities.

[0079] The diffusion model built on the U-Net network can gradually recover clear image latent vectors from noise through reverse denoising training. The model can accurately predict image details during the denoising process while maintaining an efficient training process. The use of the Adam optimizer ensures that the model parameters can be updated stably, thereby accelerating convergence and improving the model's performance.

[0080] The trained diffusion model can generate a restored latent vector based on the noise input and conditional vector, and convert it into image data through the decoder. By merging medical image data and metadata to generate conditional vectors, the individualized information of patients can be effectively introduced into the image generation process. The introduction of the control module (ControlNet) allows the image generation process to take into account external structural features, and can precisely control specific anatomical structures during the generation process, effectively reducing the distortion of the generated images and improving the medical credibility of the images.

[0081] Introducing a progressive denoising diffusion process into the model results in greater diversity of medical data generated in each inference. The introduction of noise enables the model to generate diverse latent vector representations, thereby outputting different versions of medical data. This enhances the model's flexibility in practical clinical applications, especially in disease progression prediction, personalized medicine, and simulation of different clinical trials, providing more prediction options.

[0082] S3: The client deploys a diffusion model to standardize data, aggregates model parameters to update the global model, and evaluates the global model.

[0083] Preferably, the aggregated model parameters update the global model, including,

[0084] Collect customer metadata based on the client terminal, standardize the metadata of each client using the Meta-Calibrator, deploy the diffusion model and train the model using local data;

[0085] For each client, a weighting coefficient is determined, and a weighted average method is used to aggregate the model parameters of all clients. The global model is then updated based on the global model parameters, as follows:

[0086] Where θ g This represents the global model parameters, where M represents the total number of clients, and D represents the global model parameters. i Let θ represent the amount of data for client i, D represent the total amount of data, and θ represent the total amount of data for client i. i This represents the parameters of the i-th client model, including the weights and biases of the diffusion model. As model weights for different clients.

[0087] By using a metadata calibrator, we can ensure that all clients' data is processed under the same standard, thereby improving data uniformity and consistency. Deploying a diffusion model and training it with local data allows each client to personalize its model based on its own local data, improving the model performance of each client. Using a weighted average method to aggregate the model parameters of each client helps to better balance the influence of each client.

[0088] Furthermore, the global model is evaluated, including,

[0089] Perform FID and SSIM evaluations on the global model;

[0090] The difference threshold is determined by the sum of the mean and standard deviation of the historical distribution difference values ​​between the generated image and the real image. If the distribution difference value is less than or equal to the difference threshold, it is determined that the original features of the image are preserved.

[0091] The structural threshold is determined by the sum of the mean and standard deviation of the historical structural differences between the generated image and the real image. If the structural difference is greater than or equal to the structural threshold, the original features of the image are preserved.

[0092] By using FID and SSIM to evaluate the differences between generated and real images, the improvement in image quality and fidelity can be effectively measured. FID evaluation can quantify the distributional differences between generated and real images, while SSIM evaluation focuses on the structural similarity of images.

[0093] S4 analyzes the uncertainty of the global model, defines a knowledge graph to record model operation data, uploads operation logs, and converts user-input queries for data querying.

[0094] Preferably, the uncertainty of the global model analysis includes,

[0095] A resampling mechanism (LAS) is introduced for the global model. The latent vector and conditional vector of the input image are randomly initialized multiple times. The output of the model under different initialization conditions is obtained through multiple sampling, thereby obtaining the uncertainty information of the result, obtaining the corresponding resampling to restore the latent vector, and calculating the average predicted value and average standard deviation to generate an uncertainty score.

[0096] The scoring threshold is determined by the sum of the mean and two standard deviations of the historical uncertainty scores. If the uncertainty score is less than the scoring threshold, it is judged to be of low uncertainty.

[0097] By introducing a repeated sampling mechanism (LAS), the model can obtain the output under different initialization conditions by randomly initializing the latent vector and conditional vector of the input image multiple times, thereby better reflecting the uncertainty of the generated result;

[0098] By calculating the average predicted value and standard deviation of the restored latent vector obtained from multiple samplings, an uncertainty score can be generated for each output, which can effectively identify which generated results have high stability and which have large uncertainty.

[0099] Furthermore, the knowledge graph is defined to record model operation data, including,

[0100] The DataHub architecture is used to define the triple structure of the knowledge graph (KG), which records the model's input / output mapping, data source, processing steps, model parameters, and execution environment information. Each triple contains: entity (such as dataset, model, inference result), attribute (such as data type, calculation method, model state), and relation (such as the mapping relationship between the model and the dataset).

[0101] By using the DataHub architecture to define the triple structure of the Knowledge Graph (KG), it is possible to systematically record and manage the model's input / output mapping, data sources, processing steps, model parameters, and execution environment information. This structured information storage method provides a clear framework for data traceability, management, and analysis, ensuring the transparency of data flow and model operation. The triple form makes the definition of each entity, attribute, and relationship clearer, facilitating cross-system or team collaboration and information sharing, thereby improving the traceability, maintainability, and efficiency of the workflow.

[0102] Furthermore, uploading operation logs refers to recording operation logs for each operation step of the model and storing them as structured metadata for auditing and review.

[0103] All metadata is indexed and associated with specific training processes, inference results, or abnormal events. Using the graph query interface, users can quickly find all operation history related to a certain model or dataset.

[0104] By using the DataHub architecture to define the triple structure of the knowledge graph (KG), it is possible to systematically record and manage the model's input / output mapping, data sources, processing steps, model parameters, and execution environment information, ensuring the transparency of data flow and model operation. The triple form makes the definition of each entity, attribute, and relationship clearer, facilitating cross-system or team collaboration and information sharing, thereby improving the traceability, maintainability, and efficiency of the workflow.

[0105] Furthermore, the user-input query is transformed and data is retrieved, including:

[0106] The BERT large model is used to convert user-input queries into SPARQL statements to query metadata in the knowledge graph;

[0107] Based on the user's query, the system will return detailed information on the corresponding model path, inference conditions, and input data;

[0108] By leveraging the BERT large model to convert user queries into SPARQL statements, more natural and intelligent user interaction can be achieved. This allows users to obtain information related to metadata in the knowledge graph through simple queries. This automated conversion improves the flexibility and efficiency of queries, enabling the system to quickly return relevant model paths, inference conditions, and detailed information on input data according to user needs, thereby optimizing the application of knowledge graphs and the user experience.

[0109] In summary, this disclosure, through Gaussian kernel label distribution encoding of continuous metadata and encoding of discrete metadata, can cover the full range of variables through meticulous segmentation. A unified metadata latent vector ensures that different types of features can be fused into the same latent space, while guaranteeing feature comparability. The introduction of a progressive denoising diffusion process into the model results in higher diversity of medical data generated in each inference. The introduction of noise enables the model to generate diverse latent vector representations, enhancing the model's flexibility in practical clinical applications, especially in disease progression prediction, personalized medicine, and simulation of different clinical trials, providing more prediction options.

[0110] It should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure and are not intended to limit it. Although this disclosure has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of this disclosure without departing from the spirit and scope of the technical solutions of this disclosure, and all such modifications and substitutions should be covered within the scope of the claims of this disclosure.

Claims

1. A metadata-based artificial intelligence model management system, wherein, include: The metadata processing module collects metadata, including continuous and discrete metadata. It calculates the field centroid encoding for continuous metadata into a label distribution vector, counts the total number of categories for discrete metadata, encodes and constructs a mapping matrix, concatenates vectors, and uses a linear transformation matrix to perform linear mapping to transform them into latent vectors of uniform length. The medical data processing module pairs the image data with the unified latent vector of the metadata based on the corresponding medical image data of the clinical data, and uses a pre-trained image encoder to extract the image latent vector of the medical image data, and fuses it with the unified latent vector of the metadata to generate a conditional vector. The external structural features of medical image data are determined, including key points, labels and structural information of the image. The ControlNet control module is introduced to take the conditional vector and external structural features as input to generate a control vector z_p, which is then fused with the image latent vector to generate a fused latent vector. The forward diffusion process is based on the fusion latent vector by gradually adding noise; A diffusion model is constructed based on the U-Net network, and inverse denoising training is performed to recover the latent vector of the original image from the noisy state. The mean squared error loss (MSE) is used to train the model to predict noise. The Adam optimizer is used for gradient descent optimization to update the model's parameters. The iteration stops when the model's loss no longer decreases significantly during continuous iterations. Based on the noise input and conditional vector generation conditions in the latent space, the restored latent vector is generated through the trained diffusion model. The model deployment module deploys and diffuses models on the client side to standardize data, aggregate model parameters, update the global model, and evaluate the global model. The query module analyzes the uncertainty of the global model, defines the knowledge graph, records model operation data, uploads operation logs, and transforms user-input queries for data retrieval.

2. The metadata-based artificial intelligence model management system as described in claim 1, wherein, The process of concatenating vectors, using a linear transformation matrix to perform a linear mapping to transform them into latent vectors of uniform length, includes... For each field of continuous metadata, calculate the center point of each field based on the variable range and score range of the field's sample values; Construct a Gaussian kernel label distribution function to encode its values ​​into a vector, and map the input value x of the field to a label distribution vector; For each field of discrete metadata, count the number of unique values ​​of all samples in the field as the total number of categories C, and perform embedding encoding to construct a mapping matrix with dimensions C×d, where d is a fixed embedding dimension; Each row represents the embedding vector of a category, resulting in the embedding vector representation of each discrete variable; The concatenated label distribution vector and the embedded vector representation of discrete metadata are combined. The output vector of all continuous variables has a dimension of p, and the embedded vector of all discrete variables has a dimension of d. The concatenated dimension D is defined based on the total number of continuous variables N_c and the total number of discrete variables N_d. The encoded vectors of all fields are concatenated sequentially into a single vector, and then a linear transformation matrix is ​​used to linearly map and transform it into a latent vector of uniform length.

3. The metadata-based artificial intelligence model management system as described in claim 2, wherein, The aggregation model parameters update the global model, including, Collect customer metadata based on the client terminal, standardize the metadata of each client using the Meta-Calibrator, deploy the diffusion model and train the model using local data; For each client, a weighting coefficient is determined, and a weighted average method is used to aggregate the model parameters of all clients. The global model is then updated based on the global model parameters.

4. The metadata-based artificial intelligence model management system as described in claim 3, wherein, The evaluation of the global model includes, Perform FID and SSIM evaluations on the global model; The difference threshold is determined by the sum of the mean and standard deviation of the historical distribution difference values ​​between the generated image and the real image. If the distribution difference value is less than or equal to the difference threshold, it is determined that the original features of the image are preserved. The structural threshold is determined by the sum of the mean and standard deviation of the historical structural differences between the generated image and the real image. If the structural difference is greater than or equal to the structural threshold, the original features of the image are preserved.

5. The metadata-based artificial intelligence model management system as described in claim 4, wherein, The analysis of model uncertainty in the global model includes, A resampling mechanism (LAS) is introduced for the global model. The latent vector and conditional vector of the input image are randomly initialized multiple times, and the corresponding resampling restored latent vector is obtained. The average predicted value and average standard deviation are calculated to generate an uncertainty score. The scoring threshold is determined by the sum of the mean and two standard deviations of the historical uncertainty scores. If the uncertainty score is less than the scoring threshold, it is judged to be of low uncertainty.

6. The metadata-based artificial intelligence model management system as described in claim 5, wherein, The collected metadata includes continuous metadata and discrete metadata, including, By traversing each column of the original clinical data table, the value type and value distribution range of each column are read. If the field value is a continuous numeric type, it is marked as continuous metadata; if the field value is a discrete category label, it is marked as discrete metadata. Each row in the clinical data table is used as a sample.

7. The metadata-based artificial intelligence model management system as described in claim 6, wherein, The definition of a knowledge graph, which records model operation data, refers to using the DataHub architecture to define the triple structure of the knowledge graph KG, recording the model's input / output mapping, data source, processing steps, model parameters, and execution environment information.

8. The metadata-based artificial intelligence model management system as described in claim 7, wherein, The uploaded operation log refers to recording operation logs for each operation step of the model and storing them as structured metadata.

9. The metadata-based artificial intelligence model management system as described in claim 8, wherein, The process of converting the user-input query and performing data retrieval includes: The BERT large model is used to convert user-input queries into SPARQL statements to query metadata in the knowledge graph; Based on the user's query, the system will return detailed information on the corresponding model path, inference conditions, and input data.