Rare disease data retrieval method and system

Abstract

The application provides a rare disease data retrieval method and system, belonging to the technical field of data retrieval, comprising: cloud data center construction infrastructure, real-time data processing and feature generation, blockchain distributed storage, retrieval information knowledge mapping, distributed retrieval and data extraction, result visualization display, which realizes the safe storage and tamper-proofing of rare disease data by constructing a blockchain database containing a PBFT consensus mechanism, ensuring data integrity and reliability; adopts a multi-algorithm fusion retrieval engine to realize the standardized integration of heterogeneous data, dynamic updating of knowledge graph, and accurate generation of keyword graph and retrieval label; through the two-level retrieval mechanism of knowledge mapping and rough retrieval and precise retrieval, the accuracy and efficiency of retrieval are improved; combined with the visual display module, the retrieval results are presented intuitively, meeting the efficient acquisition demand of rare disease data for clinical diagnosis and treatment and scientific research, breaking the data island and promoting cross-field resource sharing.

Description

Technical Field

[0001] This invention relates to the field of data retrieval technology, and in particular to a method and system for retrieving data on rare diseases. Background Technology

[0002] Rare diseases, also known as orphan diseases, refer to types of diseases with low incidence rates, most of which are genetic. Globally, the number of people suffering from rare diseases is enormous, but due to the often complex symptoms and lack of clinical experience, diagnosis and treatment are challenging. As societal awareness of rare diseases increases, relevant departments are paying more attention to them; however, existing rare disease data retrieval systems still cannot meet the growing needs of clinical diagnosis and research, exhibiting the following shortcomings: Existing rare disease data retrieval technologies have the following shortcomings: Lack of a universal rare disease database: Existing rare disease databases are built in a fragmented manner, lacking unified norms and standards. These databases are updated slowly and their content is not updated in a timely manner. Moreover, most of them are limited to a single field and lack cross-disciplinary integration. When doctors and researchers conduct rare disease diagnosis and treatment and research, they can often only obtain information through scattered literature resources and incomplete databases, making it difficult to form a systematic and professional knowledge system. Data silos exist: The diagnosis and treatment of rare diseases require interdisciplinary and comprehensive knowledge, including multiple fields such as genomics and clinical medicine. Existing rare disease databases are significantly lacking in professionalism and face the problem of data silos, making it difficult to share resources and results across fields, which limits progress in areas such as new drug development and treatment innovation. Difficulty in integrating heterogeneous data: Rare disease data from different data sources often have different formats and structures. Existing technologies often encounter difficulties in integrating these data, and cannot effectively convert these heterogeneous data into a unified format, thus affecting the construction of a general rare disease database and subsequent data analysis and application. Low retrieval efficiency: The retrieval algorithms used in existing technologies are relatively outdated and cannot efficiently process large-scale and complex rare disease data. During the retrieval process, existing technologies often consume a lot of computing resources, resulting in resource waste.

[0003] Therefore, this invention proposes a method and system for retrieving rare disease data. Summary of the Invention

[0004] This invention provides a method and system for retrieving rare disease data to solve the aforementioned technical problems.

[0005] This invention provides a method for retrieving rare disease data, comprising: S1: Cloud Data Center Infrastructure Construction: Construct a distributed storage blockchain and a distributed consensus blockchain with PBFT consensus mechanism, integrate them, and set up a distributed ledger to obtain a rare disease database; collect historical rare disease data and rare disease knowledge, and after preprocessing, construct heterogeneous data processing models, knowledge graph construction models, keyword graph generation models, and retrieval tag generation models based on CNN-GRPO-DPA, BERT-LSTM-CRF-SVM, CNN-cGAN, and LSTM-GCN-MLP fusion algorithms, respectively, and integrate them into a rare disease data retrieval engine; build a rare disease data retrieval platform including user login, search input, result visualization, and data download modules; S2: Real-time data processing and feature generation: Collect real-time rare disease data, preprocess it, and then call the heterogeneous data processing model to generate standardized real-time data; call the knowledge graph construction model to extract entities and relationships, and update to obtain a real-time rare disease knowledge graph; after clustering, call the keyword graph generation model to generate a real-time keyword graph, and call the search tag generation model to generate real-time search tags. S3: Blockchain Distributed Storage: Real-time rare disease data is sharded according to ICD-11 encoding, stored in a distributed storage blockchain, and smart contracts are called to generate transaction data, which is then synchronized to the distributed ledger after PBFT consensus. S4: Knowledge Mapping of Retrieved Information: Collect user retrieval information, extract entities and relationships after preprocessing, calculate the Euclidean distance with entities in the real-time knowledge graph, and map to obtain graph structure retrieval information; S5: Distributed retrieval and data extraction: The target keyword map is obtained by coarse retrieval in the ledger, the target storage address is obtained by fine retrieval, and the fragments are extracted and spliced ​​to obtain the target standardized real-time data. S6: Results visualization: The target data is displayed in topology maps and tables through the retrieval platform.

[0006] Preferably, a rare disease database is obtained, including: Identify several data servers in the storage partition and use an identity allocation mechanism to assign the data servers as storage nodes and consensus nodes; A distributed storage blockchain is constructed by connecting several storage nodes in a distributed manner, and a distributed consensus blockchain is constructed by connecting several consensus nodes in a distributed manner. By integrating distributed storage blockchain and distributed consensus blockchain, and setting up a distributed ledger, a rare disease database is obtained.

[0007] Preferred options include integrating rare disease data retrieval engines, such as: We collect some rare disease knowledge and heterogeneous historical rare disease data from different data sources, and preprocess them to obtain some preprocessed rare disease knowledge and some heterogeneous preprocessed historical rare disease data. Based on several heterogeneous preprocessed historical rare disease data, a heterogeneous data processing model is constructed using a fusion algorithm of deep learning and reinforcement learning, and several homogeneous standardized historical rare disease data are generated. Based on several standardized historical rare disease data, a knowledge graph construction model was constructed using natural language processing algorithms, resulting in a historical rare disease knowledge graph that includes several historical knowledge named entities and several historical knowledge entity relationships. Several standardized historical rare disease data are clustered to obtain several historical data clusters. Based on these historical data clusters, a keyword graph generation model is constructed using a deep learning algorithm to obtain the historical rare disease keyword graph for each historical data cluster. The historical rare disease keyword map is diffused to the corresponding historical data clusters to obtain several standardized historical rare disease data with historical rare disease keyword maps, and a retrieval tag generation model is constructed. By integrating heterogeneous data processing models, knowledge graph construction models, keyword graph generation models, and retrieval tag generation models, a rare disease data retrieval engine is obtained.

[0008] Preferably, a coarse search is performed in the ledger to obtain the target keyword map, a fine search is performed to obtain the target storage address, and fragments are extracted and concatenated to obtain the target standardized real-time data, including: Based on the real-time retrieval information after knowledge mapping, a coarse search is performed in several real-time rare disease keyword maps in the distributed ledger of the rare disease database to obtain the target real-time rare disease keyword map and several corresponding alternative real-time retrieval tags. Based on the real-time retrieval information after knowledge mapping, a fine search is performed among several candidate real-time retrieval tags to obtain the target real-time retrieval tag and the corresponding target real-time storage address. Based on several target real-time storage addresses, extract the corresponding target real-time rare disease data fragments from several storage nodes of the distributed storage blockchain of the rare disease database. Data is spliced ​​together based on the target real-time rare disease data fragments to obtain the corresponding target standardized real-time data.

[0009] Preferably, user search information is collected, preprocessed, and entities and relationships are extracted. The Euclidean distance between the entity and the real-time knowledge graph entity is calculated, and the graph structure search information is obtained by mapping, including: The rare disease data retrieval platform is used to collect user search information and preprocess the user search information to obtain preprocessed search information. The model is constructed using the knowledge graph of the rare disease data retrieval engine. Several real-time data named entities and several real-time data entity relationships are extracted from the preprocessed retrieval information, and the Euclidean distance between each data named entity and several real-time knowledge named entities in the real-time rare disease knowledge graph is obtained. The real-time knowledge named entity with the closest Euclidean distance is taken as the first target knowledge named entity of the real-time data named entity, and several real-time knowledge entity relations of the first target knowledge named entity are taken as target knowledge entity relations. The real-time knowledge named entity corresponding to the other side of each target knowledge entity relationship of the first knowledge named entity is used as the second target knowledge named entity. Map all first-target knowledge named entities, all target knowledge entity relationships, and all second-target knowledge named entities to the corresponding real-time data named entities to obtain the real-time data named entities after knowledge mapping. By integrating several real-time knowledge-mapped named entities and several real-time data entity relationships, graph structure retrieval information is obtained.

[0010] Preferably, the knowledge graph construction model is based on an algorithm derived from bidirectional Transformer encoder representation-long short-term memory network-conditional random field-support vector machine, and the knowledge graph construction model includes a word embedding module based on BERT algorithm, a semantic feature extraction module based on LSTM algorithm, a named entity output module based on CRF algorithm, and an entity relation output module based on SVM algorithm, which are connected in sequence.

[0011] Preferably, the retrieval label generation model is constructed based on the Long Short-Term Memory Network-Graph Convolutional Network-Multilayer Perceptron algorithm, and the retrieval label generation model includes a sequence feature extraction module based on the LSTM algorithm, a graph structure feature extraction module based on the GCN algorithm, and a retrieval label generation module based on the MLP algorithm. The sequence feature extraction module and the graph structure feature extraction module are both connected to the retrieval label generation module.

[0012] This invention provides a rare disease data retrieval system, comprising: The infrastructure construction module is used to build a distributed storage blockchain and a distributed consensus blockchain with PBFT consensus mechanism. After integration, a distributed ledger is set up to obtain a rare disease database. Historical rare disease data and rare disease knowledge are collected and preprocessed. Based on the fusion algorithms of CNN-GRPO-DPA, BERT-LSTM-CRF-SVM, CNN-cGAN, and LSTM-GCN-MLP, heterogeneous data processing models, knowledge graph construction models, keyword graph generation models, and retrieval tag generation models are built respectively, and integrated into a rare disease data retrieval engine. A rare disease data retrieval platform including user login, search input, result visualization, and data download modules is built. The data processing and feature generation module is used to collect real-time rare disease data, preprocess it, call the heterogeneous data processing model to generate standardized real-time data, call the knowledge graph construction model to extract entities and relationships, and update the real-time rare disease knowledge graph; after clustering, call the keyword graph generation model to generate a real-time keyword graph, and call the search tag generation model to generate real-time search tags. The distributed storage module is used to slice real-time rare disease data according to ICD-11 encoding, store it in a distributed storage blockchain, call smart contracts to generate transaction data, and synchronize it to the distributed ledger after PBFT consensus. The knowledge mapping module is used to collect user search information, extract entities and relationships after preprocessing, calculate the Euclidean distance with the entities in the real-time knowledge graph, and map the graph structure search information. The retrieval and data extraction module is used to perform a coarse search in the ledger to obtain the target keyword map, a fine search to obtain the target storage address, extract fragments and splice them to obtain the target standardized real-time data; The visualization module is used to display target data in topology maps and tables through the retrieval platform.

[0013] Compared with the prior art, the beneficial effects of this application are as follows: By constructing a blockchain database with a PBFT consensus mechanism, secure storage and tamper-proof protection of rare disease data are achieved, ensuring data integrity and reliability. A multi-algorithm fusion retrieval engine enables standardized integration of heterogeneous data, dynamic updating of the knowledge graph, and precise generation of keyword graphs and search tags. A two-tiered retrieval mechanism combining knowledge mapping and coarse and fine searches improves retrieval accuracy and efficiency. Combined with a visualization module, retrieval results are presented intuitively, meeting the needs of clinical diagnosis and research for efficient acquisition of rare disease data, breaking down data silos, and promoting cross-domain resource sharing.

[0014] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings.

[0015] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a rare disease data retrieval method according to an embodiment of the present invention; Figure 2 This is a structural diagram of a rare disease data retrieval system according to an embodiment of the present invention. Detailed Implementation

[0017] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0018] This invention provides a method for retrieving rare disease data, such as... Figure 1 As shown, it includes: S1: Cloud Data Center Infrastructure Construction: Construct a distributed storage blockchain and a distributed consensus blockchain with PBFT consensus mechanism, integrate them, and set up a distributed ledger to obtain a rare disease database; collect historical rare disease data and rare disease knowledge, and after preprocessing, construct heterogeneous data processing models, knowledge graph construction models, keyword graph generation models, and retrieval tag generation models based on CNN-GRPO-DPA, BERT-LSTM-CRF-SVM, CNN-cGAN, and LSTM-GCN-MLP fusion algorithms, respectively, and integrate them into a rare disease data retrieval engine; build a rare disease data retrieval platform including user login, search input, result visualization, and data download modules; S2: Real-time data processing and feature generation: Collect real-time rare disease data, preprocess it, and then call the heterogeneous data processing model to generate standardized real-time data; call the knowledge graph construction model to extract entities and relationships, and update to obtain a real-time rare disease knowledge graph; after clustering, call the keyword graph generation model to generate a real-time keyword graph, and call the search tag generation model to generate real-time search tags. S3: Blockchain Distributed Storage: Real-time rare disease data is sharded according to ICD-11 encoding, stored in a distributed storage blockchain, and smart contracts are called to generate transaction data, which is then synchronized to the distributed ledger after PBFT consensus. S4: Knowledge Mapping of Retrieved Information: Collect user retrieval information, extract entities and relationships after preprocessing, calculate the Euclidean distance with entities in the real-time knowledge graph, and map to obtain graph structure retrieval information; S5: Distributed retrieval and data extraction: The target keyword map is obtained by coarse retrieval in the ledger, the target storage address is obtained by fine retrieval, and the fragments are extracted and spliced ​​to obtain the target standardized real-time data. S6: Results visualization: The target data is displayed in topology maps and tables through the retrieval platform.

[0019] In this embodiment, the PBFT consensus mechanism is a practical Byzantine fault-tolerant consensus mechanism, a fault-tolerant consensus algorithm suitable for distributed systems. Distributed storage blockchain refers to a blockchain structure formed by the distributed connection of several storage nodes, and distributed consensus blockchain refers to a blockchain structure formed by the distributed connection of several consensus nodes. The distributed ledger refers to a distributed database used to record all data operation records in the distributed storage blockchain and the distributed consensus blockchain. All nodes hold a copy of the ledger, ensuring data traceability.

[0020] In this embodiment, the CNN-GRPO-DPA fusion algorithm refers to a heterogeneous data processing algorithm formed by fusing Convolutional Neural Network (CNN), Group Relative Policy Optimization (GRPO), and Data Processing Algorithm (DPA). For two types of heterogeneous data—XML-formatted electronic medical record data and CSV-formatted gene testing data—CNN extracts deep features, GRPO generates processing strategies adapted to both types of data (XML parsing strategy and CSV data normalization strategy), and the format conversion algorithm in DPA converts both types of data into standardized JSON format data. Specifically, the CNN uses a structure of 3 convolutional layers + 2 pooling layers, the GRPO objective function is set to minimize the processing error rate and maximize processing efficiency, and the DPA includes sub-algorithms such as format conversion, Gaussian denoising, and magnitude unification, all implemented using the Python TensorFlow framework.

[0021] The BERT-LSTM-CRF-SVM fusion algorithm is a knowledge graph construction algorithm that integrates the Bidirectional Transformer Encoder Representation (BERT), Long Short-Term Memory Network (LSTM), Conditional Random Field (CRF), and Support Vector Machine (SVM). It can achieve named entity extraction and entity relationship recognition. For example, in standardized historical rare disease data such as Dravet syndrome patients with the c.1837C>T mutation in the SCN1A gene, manifested as fever-induced epilepsy, BERT generates word embedding vectors, LSTM extracts semantic features, CRF identifies named entities such as Dravet syndrome, SCN1A gene, c.1837C>T mutation, and fever-induced epilepsy, and SVM identifies entity relationships such as Dravet syndrome-pathogenic gene-SCN1A gene and Dravet syndrome-symptom-fever-induced epilepsy. Among them, BERT uses a base-uncased pre-trained model, LSTM has a hidden layer dimension of 256, the CRF transition matrix is ​​optimized through training, and SVM uses the RBF kernel function, implemented through HuggingFace's Transformers library and Scikit-learn library.

[0022] The CNN-cGAN fusion algorithm refers to a keyword graph generation algorithm that combines a convolutional neural network (CNN) with a conditional generative adversarial network (cGAN). It can generate hierarchical keyword graphs based on cluster features.

[0023] The LSTM-GCN-MLP fusion algorithm is a retrieval tag generation algorithm that integrates Long Short-Term Memory (LSTM), Graph Convolutional Network (GCN), and Multilayer Perceptron (MLP). It can generate retrieval tags based on data features and keyword graphs. For example, for standardized historical data with a keyword graph, LSTM extracts sequence features (e.g., a 3-year-old male patient, onset date January 2024), GCN extracts graph structure features from the keyword graph (e.g., core keyword-related keyword weights), and MLP fuses these two features to generate retrieval tags such as Dravet syndrome, SCN1A mutation, rare childhood diseases, and anti-epileptic drugs. Specifically, the LSTM uses a hidden layer dimension of 128, the GCN uses two graph convolutional layers, and the MLP contains three fully connected layers. The activation function is ReLU, and the algorithm is implemented using the TensorFlow framework.

[0024] The heterogeneous data processing model refers to a model built based on the CNN-GRPO-DPA fusion algorithm, used to convert heterogeneous rare disease data into homogeneous standardized data. For example, inputting XML-formatted medical record data from a hospital HIS system, CSV-formatted test data from a gene testing institution, and PDF text data from medical literature, the model uses CNN to extract features from each data point, GRPO to generate corresponding processing strategies, and the DPA algorithm to uniformly convert them into standardized JSON-formatted data with a consistent structure and format. Specifically, the model adopts a modular design, including a feature extraction module, a strategy generation module, and a data processing module. Model parameters are optimized through batch training, and the training dataset contains 100,000 heterogeneous rare disease data points.

[0025] A knowledge graph construction model refers to a model built based on a BERT-LSTM-CRF-SVM fusion algorithm, used to extract entities and relationships and construct a knowledge graph. For example, given 100,000 standardized historical rare disease data entries, this model extracts 10 categories of named entities, such as "disease," "gene," "symptom," and "drug," as well as 8 categories of entity relationships, such as "pathogenic gene," "accompanying symptoms," and "treatment drug," constructing a historical rare disease knowledge graph containing 50,000 nodes and 80,000 edges.

[0026] Keyword graph generation models refer to models built based on the CNN-cGAN fusion algorithm, used to generate keyword graphs for data clustering. During model training, manually annotated keyword graphs are used as labels, and cluster features are input in batches for training.

[0027] The retrieval tag generation model refers to a model built based on the LSTM-GCN-MLP fusion algorithm, used to generate retrieval tags for standardized data. For example, given standardized historical data with a keyword graph, the model generates 3-5 retrieval tags for each data point. For instance, the tags for a Dravet syndrome data point might be "rare disease," "epilepsy-related," "SCN1A mutation," and "childhood onset." Implementation: The model training loss function uses cross-entropy loss, the optimizer is the Adam optimizer, and the training dataset contains 80,000 labeled standardized data points.

[0028] In this embodiment, for example, a web interface is developed using Vue 3.0 + ElementPlus. Users log in to the platform with their account and password, and enter "infant fever epilepsy, SCN1A gene" in the search input box. The platform sends the search information to the cloud data center. After receiving the results, the platform displays the keyword map in a topology map and the detailed data in a table. Users can click the download button to export the data (supports CSV and PDF formats).

[0029] The user login module refers to the functional module that implements user authentication and access control. For example, when registering, users submit information such as name, institution, and identity (doctor / researcher). After the background review is approved, an account and password are assigned. When logging in, the user's identity is verified by the account, password, and verification code. Doctor users have data retrieval and download permissions, while researchers have additional data upload permissions.

[0030] The search input module refers to the functional module used to receive user search information, supporting input methods such as text input and keyword selection. For example, users can directly enter the text "3-year-old child, fever-induced epilepsy, suspected rare disease", or they can select keyword combinations such as disease type, symptoms, and genes through the drop-down menu for search.

[0031] The results visualization module is a functional module that displays search results in a visual format, supporting both topology graphs and tables. For example, the keyword graph of search results is displayed as a topology graph, with core keyword nodes being the largest and related keyword nodes having their size adjusted according to their matching degree; the thickness of the edges indicates the strength of the association. Detailed data is displayed in tables, including fields such as data ID, disease name, pathogenic gene, symptoms, drugs, data source, and timestamp.

[0032] The data download module refers to the functional module used to export search results data. For example, after a user selects the data items to be downloaded, chooses the export format (CSV, PDF), and clicks the download button, the system generates the corresponding format file and provides a download link. The CSV format contains all data fields, while the PDF format contains data details and keyword graphs.

[0033] In this embodiment, real-time rare disease data refers to rare disease-related data collected in real time from different data sources, including clinical medical records, gene testing results, medical imaging reports, etc. For example, three data entries uploaded in real time by a hospital are: Patient A, female, 2 years old, recurrent seizures, gene testing shows SCN2A gene mutation (text type); Patient B, male, 5 years old, muscle atrophy, electromyography shows neurogenic damage (text + numerical type); Patient C, male, 1 year old, brain CT shows ventricular enlargement (image + text type).

[0034] Preprocessing refers to cleaning, denoising, and format conversion of real-time rare disease data to remove invalid and erroneous data, preparing it for subsequent processing. For example, for patient A's data, redundant characters such as "repeatedly" and "display" are removed, and "variation" is corrected to "mutation." For patient B's electromyography data, missing values ​​are filled using the mean imputation method, and noise is removed using Gaussian filtering. For patient C's CT images, they are converted to JPG format and normalized to 512×512 pixels. Implementation methods: Text data preprocessing uses jieba word segmentation and regular expression cleaning; numerical data preprocessing uses the Pandas library; and image data preprocessing uses the OpenCV library, all automated through Python scripts.

[0035] Standardized real-time data refers to real-time data that has been processed by a heterogeneous data processing model and has a unified format and consistent structure. For example, the three real-time data sets mentioned above were processed and uniformly converted into JSON format, containing fixed fields such as patient ID, gender, age, disease-related symptoms, examination results, and genetic information.

[0036] In this embodiment, an entity refers to a core element with specific meaning in real-time data (such as disease name, gene, symptom, drug), and a relation refers to the association between entities (such as pathogenic gene, accompanying symptom). For example, from the standardized data of patient A, the entities "patient A", "SCN2A gene", and "convulsions" are extracted, and the entity relations are "patient A-symptom-convulsions" and "patient A-pathogenic gene-SCN2A gene". Specifically, entities and relations are automatically extracted through a knowledge graph construction model. Entities are identified by unique IDs, and relations are encoded using standardized relation types (such as "symptoms" encoded as S001 and "pathogenic gene" encoded as G001).

[0037] A real-time rare disease knowledge graph refers to a dynamic knowledge graph formed by updating the historical knowledge graph with entities and relationships based on real-time data. For example, if the relationship "Dravet syndrome-SCN1A gene" already exists in the historical knowledge graph, and a new entity "Patient A-SCN2A gene-convulsions" is added for patient A, the knowledge graph is updated, adding an SCN2A gene node, a convulsion symptom node, and the corresponding relationship edges.

[0038] Clustering refers to using clustering algorithms to group standardized real-time data according to similarity, forming data clusters. For example, using the fuzzy C-means (FCM) clustering algorithm, 100 standardized real-time data points can be divided into 5 clusters, such as cluster 1 containing 25 epilepsy-related data points (including patient A's data), and cluster 2 containing 20 muscle atrophy-related data points (including patient B's data).

[0039] In this embodiment, the real-time keyword map refers to the keyword map generated for real-time data clustering. For example, the real-time keyword map of cluster 1 (epilepsy-related) is centered on "epilepsy-like seizures" and associated with keywords such as "SCN2A gene", "childhood onset", and "convulsions"; the keyword map of cluster 2 (muscle atrophy-related) is centered on "muscle atrophy" and associated with keywords such as "neurogenic damage" and "SMN1 gene".

[0040] In this embodiment, real-time retrieval tags refer to the retrieval tags generated for each piece of standardized real-time data. For example, the retrieval tags for patient A's standardized real-time data are "rare childhood disease", "epilepsy-like seizures", and "SCN2A gene mutation"; the data tags for patient B are "muscular atrophy", "neurogenic damage", and "rare disease".

[0041] In this embodiment, ICD-11 coding is achieved by parsing the disease name in real-time data, matching it with the ICD-11 coding library (the latest version of the ICD-11 coding table is deployed locally), and automatically assigning a code to the data. For example, the ICD-11 code corresponding to Dravet syndrome is 8A61.0.

[0042] In this embodiment, data sharding refers to dividing real-time rare disease data into several data segments according to ICD-11 encoding and generating replica shards to improve storage reliability and retrieval efficiency. For example, 500 real-time data entries encoded as 8A61.0 (Dravet syndrome) in ICD-11 are divided into 5 data shards (100 data entries per shard), and 2 replica shards are generated from each shard, resulting in a total of 15 shards (5 primary shards + 10 replica shards).

[0043] In this embodiment, the smart contract has a preset rule that when a storage node receives a data shard, it automatically extracts the storage address and timestamp to generate transaction data. After the transaction data is verified by the consensus node, it is synchronized to the distributed ledger. When a storage node receives a shard, the smart contract automatically executes the rule to generate transaction data containing the shard ID, storage address, timestamp, and node signature.

[0044] In this embodiment, transaction data refers to structured data that records data shard storage information and consensus information, used for consensus verification and ledger synchronization among blockchain nodes. For example, the transaction data of a certain data shard is {"Shard ID":"8A61.0_01_01","Storage Address":"node3: / data / 8A61.0 / 01","Timestamp":"2025-01-10 14:30:25","Node Signature":"xxx","Consensus Status":"Pending Consensus"}.

[0045] In this embodiment, the user search information refers to the information entered by the user through the search platform to query rare disease data, including text descriptions, keyword combinations, etc. For example, the user's search information is "3-year-old child, convulsions after fever, suspected SCN1A gene-related rare disease" or "muscle weakness, SMN1 gene mutation".

[0046] Knowledge mapping refers to mapping the entities and relationships of user-retrieved information to a real-time rare disease knowledge graph, generating standardized graph-structured retrieval information. Graph-structured retrieval information refers to the network structure information formed by integrating the entities and relationships after knowledge mapping.

[0047] In this embodiment, the target storage address refers to the storage path of the target data fragment corresponding to the target retrieval tag in the distributed storage blockchain.

[0048] In this embodiment, a topology graph refers to a visual chart that displays the target keyword map in a network structure. Nodes represent keywords, and edges represent the relationships between keywords. For example, in the topology graph of the target keyword map, "Dravet syndrome" is the core node (largest in size), "SCN1A gene," "post-febrile seizures," and "valproic acid" are first-level related nodes (second largest in size), and "childhood onset" and "epilepsy-like seizures" are second-level related nodes (smallest in size). The thickness of the edges indicates the strength of the association (e.g., the edge of "Dravet syndrome-SCN1A gene" is the thickest).

[0049] Tabular display refers to a visualization method that presents detailed information about standardized real-time data of a target in tabular form.

[0050] The beneficial effects of the above technical solution are as follows: By constructing a blockchain database with a PBFT consensus mechanism, secure storage and tamper-proof protection of rare disease data are achieved, ensuring data integrity and reliability; a multi-algorithm fusion retrieval engine enables standardized integration of heterogeneous data, dynamic updating of knowledge graphs, and accurate generation of keyword graphs and search tags; a two-level retrieval mechanism of knowledge mapping and coarse and fine retrieval improves the accuracy and efficiency of retrieval; and a visualization module presents retrieval results intuitively, meeting the needs of clinical diagnosis and research for efficient acquisition of rare disease data, breaking down data silos, and promoting cross-domain resource sharing.

[0051] This invention provides a method for retrieving rare disease data, resulting in a rare disease database, comprising: Identify several data servers in the storage partition and use an identity allocation mechanism to assign the data servers as storage nodes and consensus nodes; A distributed storage blockchain is constructed by connecting several storage nodes in a distributed manner, and a distributed consensus blockchain is constructed by connecting several consensus nodes in a distributed manner. By integrating distributed storage blockchain and distributed consensus blockchain, and setting up a distributed ledger, a rare disease database is obtained.

[0052] Preferred options include integrating rare disease data retrieval engines, such as: We collect some rare disease knowledge and heterogeneous historical rare disease data from different data sources, and preprocess them to obtain some preprocessed rare disease knowledge and some heterogeneous preprocessed historical rare disease data. Based on several heterogeneous preprocessed historical rare disease data, a heterogeneous data processing model is constructed using a fusion algorithm of deep learning and reinforcement learning, and several homogeneous standardized historical rare disease data are generated. Based on several standardized historical rare disease data, a knowledge graph construction model was constructed using natural language processing algorithms, resulting in a historical rare disease knowledge graph that includes several historical knowledge named entities and several historical knowledge entity relationships. Several standardized historical rare disease data are clustered to obtain several historical data clusters. Based on these historical data clusters, a keyword graph generation model is constructed using a deep learning algorithm to obtain the historical rare disease keyword graph for each historical data cluster. The historical rare disease keyword map is diffused to the corresponding historical data clusters to obtain several standardized historical rare disease data with historical rare disease keyword maps, and a retrieval tag generation model is constructed. By integrating heterogeneous data processing models, knowledge graph construction models, keyword graph generation models, and retrieval tag generation models, a rare disease data retrieval engine is obtained.

[0053] In this embodiment, a storage partition refers to a logical storage area in a cloud data center specifically designated for storing rare disease data, and a data server refers to a physical or virtual server with data storage and computing capabilities.

[0054] In this embodiment, the identity allocation mechanism is implemented using a Python script. Server performance is periodically (monthly) monitored and scores are updated. Node identities can be dynamically adjusted (when a storage node's performance improves to a score ≥ 80, it can be upgraded to a consensus node; when a consensus node's performance drops to a score < 80, it is downgraded to a storage node). Specific scoring criteria are shown in Table 1. Table 1 Scoring Criteria In this embodiment, rare disease knowledge refers to professional knowledge related to rare diseases, including disease definitions, causes, symptoms, diagnostic criteria, treatment plans, and gene associations, derived from medical literature, clinical guidelines, and authoritative databases. For example, rare disease knowledge includes "Dravet syndrome is an autosomal dominant epileptic encephalopathy caused by the SCN1A gene" and "Spinal muscular atrophy is a neuromuscular disease caused by mutations in the SMN1 gene, with the main symptom being muscle weakness," etc.

[0055] Heterogeneous historical rare disease data refers to historical rare disease-related data from different data sources with different data formats and structures, including but not limited to text-based medical record data, numerical gene testing data, and image-based medical imaging data.

[0056] Preprocessed rare disease knowledge refers to structured and standardized rare disease knowledge after preprocessing, which facilitates subsequent knowledge graph construction. For example, the rare disease knowledge before preprocessing is "Dravet syndrome, also known as severe myoclonic epilepsy of infancy, usually occurs before the age of 1, caused by mutation of the SCN1A gene, with symptoms including fever-induced seizures and psychomotor developmental delay". After preprocessing, it is converted into structured knowledge: {"Disease Name":"Dravet Syndrome","Alternative Name":"Severe Myoclonic Epilepsy of Infancy","Age of Onset":"Before 1 Year","Cause of Disease":"SCN1A Gene","Symptoms":["Fever-Induced Seizures","Psychomotor Developmental Delay"]}.

[0057] Historical knowledge named entities refer to core elements with specific meanings extracted from preprocessed rare disease knowledge, including disease names, gene names, symptoms, drugs, mutation sites, etc. For example, historical knowledge named entities extracted from rare disease knowledge include Dravet syndrome, and each entity is assigned a unique ID (e.g., the disease entity ID is prefixed with D, and the gene entity ID is prefixed with G).

[0058] Historical knowledge entity relationships refer to the associations between named entities in historical knowledge, including relationship types such as pathogenic genes, accompanying symptoms, treatment drugs, and age of onset. For example, historical knowledge entity relationships include Dravet syndrome - pathogenic gene - SCN1A gene, etc., and each relationship type is assigned a unique code (e.g., pathogenic gene is coded as G001, accompanying symptoms are coded as S001).

[0059] A historical rare disease knowledge graph refers to a network-like knowledge structure built upon named entities and their relationships based on historical knowledge. It can intuitively display the connections between entities related to rare diseases. For example, a historical rare disease knowledge graph contains 50,000 nodes (including 2,000 disease nodes, 3,000 gene nodes, 10,000 symptom nodes, 5,000 drug nodes, and 30,000 other nodes) and 80,000 edges (relationships). For instance, the "Dravet syndrome" node is connected to the "G001-SCN1A gene" node via the "pathogenic gene" edge, and to the "S001-fever-induced epileptic seizure" node via the "accompanying symptom" edge.

[0060] Historical data clusters refer to datasets formed by grouping standardized historical rare disease data according to similarity using clustering algorithms. Data within the same cluster have high similarity (such as the same disease, similar symptoms, or the same pathogenic gene). For example, using the fuzzy C-means (FCM) clustering algorithm, 80,000 standardized historical rare disease data are divided into 20 clusters, each corresponding to one or a class of rare diseases. For instance, cluster 1 contains 4,000 data related to Dravet syndrome, cluster 2 contains 3,500 data related to spinal muscular atrophy, and cluster 3 contains 3,000 data related to phenylketonuria.

[0061] The diffusion of historical rare disease keyword maps refers to linking the keyword maps generated for historical data clusters to each standardized historical rare disease data point within that cluster. This allows each data point to carry cluster-level keyword features, supporting the generation of subsequent search tags. For example, if historical data cluster 1 contains 4000 standardized historical data points for Dravet syndrome, the historical rare disease keyword map (ID: KG-H-001) of this cluster is diffused to the metadata fields of each data point through a "cluster ID-data ID" mapping relationship. That is, each data point adds a "keyword map ID: KG-H-001" field, and the corresponding keyword map can be accessed through this ID.

[0062] The beneficial effects of the above technical solution are as follows: by rationally dividing storage partitions and data server identities, a distributed storage and consensus blockchain are constructed, and combined with a distributed ledger, a decentralized and highly fault-tolerant rare disease database architecture is formed, improving the reliability and scalability of data storage. Based on four core models constructed by the fusion of multiple algorithms, efficient integration of heterogeneous data, structured construction of knowledge graphs, hierarchical generation of keyword graphs, and accurate matching of search tags are achieved. The rare disease data retrieval engine formed through modular integration has end-to-end data processing and retrieval capabilities, significantly improving the standardization rate of rare disease data and the accuracy of retrieval feature generation.

[0063] This invention provides a rare disease data retrieval method, which obtains a target keyword map through coarse retrieval in a ledger, obtains the target storage address through fine retrieval, extracts fragments and concatenates them to obtain standardized real-time target data, including: Based on the real-time retrieval information after knowledge mapping, a coarse search is performed in several real-time rare disease keyword maps in the distributed ledger of the rare disease database to obtain the target real-time rare disease keyword map and several corresponding alternative real-time retrieval tags. Based on the real-time retrieval information after knowledge mapping, a fine search is performed among several candidate real-time retrieval tags to obtain the target real-time retrieval tag and the corresponding target real-time storage address. Based on several target real-time storage addresses, extract the corresponding target real-time rare disease data fragments from several storage nodes of the distributed storage blockchain of the rare disease database. Data is spliced ​​together based on the target real-time rare disease data fragments to obtain the corresponding target standardized real-time data.

[0064] In this embodiment, the real-time retrieval information after knowledge mapping refers to the graph structure information formed after the user's retrieval information has been preprocessed, entity relationship extracted, and mapped to the knowledge graph. It includes standardized entities (such as diseases and genes), entity relationships, and association weights, and is the core basis for retrieval matching. For example, the user's original retrieval information "2-year-old child, fever and convulsions, SCN1A gene abnormality" will be transformed into the following graph structure retrieval information after knowledge mapping: Nodes (entities) include "2-year-old child" (type: patient characteristic, weight 1.0), "fever and convulsions" (type: symptom, weight 0.95), "SCN1A gene" (type: gene, weight 0.98), and "Dravet syndrome" (type: disease, mapping association weight 0.92); Edges (relationships) include "2-year-old child - symptom - fever and convulsions" (weight 0.95), "fever and convulsions - associated disease - Dravet syndrome" (weight 0.92), and "Dravet syndrome - pathogenic gene - SCN1A gene" (weight 0.98).

[0065] In this embodiment, for example, the distributed ledger stores 100 real-time rare disease keyword maps (covering Dravet syndrome, spinal muscular atrophy, etc.). Using the core entities "Dravet syndrome", "SCN1A gene", and "febrile seizures" in the retrieved information after knowledge mapping as search keywords, the core keywords and related keywords of the keyword maps in the ledger are traversed to select 15 target real-time rare disease keyword maps that contain at least one core search entity and have a correlation degree ≥ 0.6 (e.g., KG-R-001: Dravet syndrome real-time map, correlation degree 0.92; KG-R-005: infantile epilepsy real-time map, correlation degree 0.75). The 60 candidate search tags corresponding to these maps are extracted (e.g., "Dravet syndrome-SCN1A" "febrile seizures-children"). This is considered a coarse search.

[0066] Alternative real-time search tags refer to search tags that are initially related to the search information and extracted from the target real-time rare disease keyword map during the coarse search phase. For example, alternative search tags extracted from the target real-time map KG-R-001 (Dravet syndrome) include "Dravet syndrome-SCN1A-2 years old", "Dravet syndrome-febrile seizures-valproic acid", "Dravet syndrome-autosomal dominant inheritance", and "Dravet syndrome-mental retardation", for a total of 12 tags.

[0067] Precise retrieval refers to the retrieval process that uses real-time retrieval information after knowledge mapping as the basis for accuracy, calculates semantic similarity among candidate real-time retrieval tags, filters out target real-time retrieval tags with a relevance of ≥0.85, and extracts the target real-time storage address corresponding to the tag. For example, the core feature combination of the retrieved information after knowledge mapping is "Dravet syndrome-SCN1A gene-2 ​​years old-febrile seizures". The semantic similarity between 60 candidate tags and this combination is calculated (using the cosine similarity algorithm, based on the word vector distance between the tag and the retrieved feature): among which "Dravet syndrome-SCN1A-2 years old" has a similarity of 0.93, "Dravet syndrome-febrile seizures-valproic acid" has a similarity of 0.88, and "infantile epilepsy-SCN1A-febrile" has a similarity of 0.72. Twelve target real-time retrieval tags with a similarity ≥0.85 are selected, and the 15 target real-time storage addresses corresponding to these tags (such as "node2: / data / real / 8A61.0 / 03" and "node5: / data / real / 8A61.0 / 07") are extracted, which are the target real-time storage addresses.

[0068] A target real-time rare disease data shard refers to a fragment containing a portion of the target standardized real-time rare disease data, extracted from the storage nodes of a distributed storage blockchain based on the target real-time storage address. Each shard is an independent encrypted file. For example, based on the target storage address "node3: / data / real / ICD-8A61.0 / 20250110_03.shard", the shard file (80MB in size, AES-256 encrypted) is extracted from node3. After decryption, it contains 20 target data entries (each in JSON format, including patient ID, disease information, gene testing results, etc.). The shard index records the position of the shard in the complete dataset (e.g., "Total shards 5, current shard 3, data ID range: R20250110_051-R20250110_070").

[0069] Data sharding and splicing refers to the process of merging, deduplicating, and completing multiple extracted target real-time rare disease data shards according to their shard indices to form a complete target-standardized real-time rare disease dataset. For example, extracting 5 target data shards (corresponding to shards 3, 4, 5, 1, and 2 at storage addresses node3, node5, and node7), first sorting them by shard indices (1-5), then reading the index file of each shard to determine the data ID range, deleting duplicate data during merging (keeping the latest timestamp version, such as data IDR20250110_055 which exists in both shards 3 and shard 4, keeping the version with the newest timestamp in shard 4), and completing missing fields (e.g., if the "drug recommendation" field is missing in a data entry, filling it with "Refer to the Dravet syndrome treatment guidelines, valproic acid recommended"), and finally splicing them together to obtain 100 complete target-standardized real-time rare disease datasets.

[0070] The beneficial effects of the above technical solution are: the two-level retrieval mechanism of coarse and fine retrieval greatly improves retrieval efficiency and accuracy; the distributed fragmentation extraction based on the target storage address ensures the efficiency and reliability of data extraction; and the deduplication, completion and format verification in the data fragmentation and splicing process ensure the integrity and standardization of the final target data.

[0071] This invention provides a rare disease data retrieval method, which collects user search information, extracts entities and relationships after preprocessing, calculates the Euclidean distance with entities in a real-time knowledge graph, and maps the data to obtain graph-structured retrieval information, including: The rare disease data retrieval platform is used to collect user search information and preprocess the user search information to obtain preprocessed search information. The model is constructed using the knowledge graph of the rare disease data retrieval engine. Several real-time data named entities and several real-time data entity relationships are extracted from the preprocessed retrieval information, and the Euclidean distance between each data named entity and several real-time knowledge named entities in the real-time rare disease knowledge graph is obtained. The real-time knowledge named entity with the closest Euclidean distance is taken as the first target knowledge named entity of the real-time data named entity, and several real-time knowledge entity relations of the first target knowledge named entity are taken as target knowledge entity relations. The real-time knowledge named entity corresponding to the other side of each target knowledge entity relationship of the first knowledge named entity is used as the second target knowledge named entity. Map all first-target knowledge named entities, all target knowledge entity relationships, and all second-target knowledge named entities to the corresponding real-time data named entities to obtain the real-time data named entities after knowledge mapping. By integrating several real-time knowledge-mapped named entities and several real-time data entity relationships, graph structure retrieval information is obtained.

[0072] In this embodiment, for example, the user's original search information is "5-year-old boy, progressive muscle weakness, SMN1 gene positive, disease type: neuromuscular disease, data source: tertiary hospital". After preprocessing, the preprocessed search information is "5-year-old boy, progressive muscle weakness, SMN1 gene positive, neuromuscular disease, tertiary hospital".

[0073] In this embodiment, real-time data entity relationship extraction refers to the identification of the association relationships between extracted real-time data named entities through the SVM relationship classification module of the knowledge graph model. The relationship type must be consistent with the relationship type of the real-time rare disease knowledge graph (such as "patient-symptom" or "disease-gene"). For example, based on the extracted real-time data named entities, the SVM model identifies four types of real-time data entity relationships: "5-year-old boy-symptom-progressive muscle weakness" (patient characteristic-symptom relationship, confidence 0.95); "neuromuscular disease-associated gene-SMN1 gene" (disease-gene relationship, confidence 0.92); "5-year-old boy-disease-neuromuscular disease" (patient characteristic-disease relationship, confidence 0.88); "SMN1 gene-attribute-positive" (gene-detection result relationship, confidence 0.90). Each relationship contains four elements: head entity ID, relationship type, tail entity ID, and confidence.

[0074] In this embodiment, the Euclidean distance is calculated using the linalg.norm function from the NumPy library, which has a computational complexity of O(n). This is an existing technique and can be used directly.

[0075] In this embodiment, the first target knowledge named entity refers to the knowledge entity in the real-time rare disease knowledge graph that has the smallest Euclidean distance (highest semantic similarity) to a certain real-time data named entity. The target knowledge entity relationship refers to the existing associations between the first target knowledge named entity and other knowledge entities in the real-time rare disease knowledge graph. For example, the knowledge entity relationships of the first target knowledge named entity "Spinal muscular atrophy (DIS-K-023)" in the real-time knowledge graph include: "Spinal muscular atrophy - Pathogenic gene - SMN1 gene (GEN-K-105)" (relationship type: pathogenic gene, weight 0.98); "Spinal muscular atrophy - Core symptom - Muscle weakness (SYM-K-089)" (relationship type: core symptom, weight 0.95), etc. These relationships are all considered target knowledge entity relationships. The second target knowledge named entity refers to the knowledge entity on the other side of the target knowledge entity relationship that is opposite to the first target knowledge named entity.

[0076] In this embodiment, for example, the real-time data named entity "Neuromuscular Disease" (ID: DATA-DIS-001) is mapped to the entity as follows: {"Original Entity ID": "DATA-DIS-001", "Original Entity Name": "Neuromuscular Disease", "First Target Knowledge Entity ID": "DIS-K-023", "First Target Knowledge Entity Name": "Spinal Muscular Atrophy", "Association List": [{"Relationship Type": "Pathogenic Gene", "Weight": 0.98,"Second Target Knowledge Entity ID": "GEN-K-105", "Second Target Knowledge Entity Name": "SMN1 Gene"}, {"Relationship Type": "Core Symptom", "Weight": 0.95,"Second Target Knowledge Entity ID": "SYM-K-089", "Second Target Knowledge Entity Name": "Muscle Weakness"}], "Mapping Confidence": 0.92} (The mapping confidence is the inverse transformation of the Euclidean distance).

[0077] Real-time retrieval information after knowledge mapping in graph structures refers to the integration of named entities and relationships between real-time data entities after all knowledge mappings, forming a mesh structure information containing nodes (entities), edges (relationships), and weights (confidence).

[0078] The beneficial effects of the above technical solution are: based on the entity relationship extraction of the knowledge graph construction model, combined with semantic similarity matching of Euclidean distance, the accurate mapping between real-time data entities and knowledge graph entities is realized. Through the hierarchical expansion of the first target entity, target relationship and second target entity, the dimensions of the retrieved information are enriched, and a structured graph structure retrieval information is formed, which greatly improves the accuracy of subsequent distributed retrieval.

[0079] This invention provides a rare disease data retrieval method, which obtains a historical rare disease keyword map for each historical data cluster, including: The semantic vector of a single standardized historical rare disease data point within a historical data cluster is extracted using a BERT pre-trained model, and the local feature vector of the cluster is obtained. By statistically analyzing historical data, we can identify the proportion of rare disease main types, the number of overlapping pathogenic genes, and the frequency of occurrence of core symptoms within clusters. We can then construct a global statistical feature matrix and obtain a global statistical feature vector. The number of edges and hierarchical relationships associated with cluster entities in the historical rare disease knowledge graph are retrieved to construct a knowledge association feature matrix; the standard deviation of the knowledge association feature matrix is ​​calculated, and the matrix elements are scaled and corrected based on the standard deviation to obtain the knowledge association feature vector; Attention mapping and weighted concatenation are performed on the cluster local feature vector, global statistical feature vector, and knowledge association feature vector to obtain the cluster fusion feature vector; Construct and train a keyword graph generation model; Using manually annotated rare disease cluster-level keywords as labels, the cluster fusion feature vector is input into the keyword graph generation model for training; The trained model is applied to each cluster to generate a hierarchical set of core keywords and related keywords, and the matching value between keywords is calculated. Based on the matching value, a topological keyword network centered on the core keywords is constructed, thus obtaining the keyword map of historical rare diseases.

[0080] In this embodiment, the present invention employs a medical pre-trained model (biobert-base-cased-v1.1), which is suitable for semantic feature extraction from rare disease text data. For example, for a single standardized historical data entry in historical data cluster 1 (Dravet syndrome) stating "Female patient, 6 months old, experienced generalized tonic-clonic seizures due to fever; gene testing showed a heterozygous mutation of SCN1A gene c.1837C>T, diagnosed as Dravet syndrome," the BERT pre-trained model converts this text into a 256-dimensional semantic vector (e.g., [-0.12, 0.35, ..., 0.28]), which contains core semantic information such as "Dravet syndrome," "SCN1A gene," and "febrile seizures."

[0081] In this embodiment, the cluster local feature vector refers to the vector that reflects the local textual semantic features of the cluster by averaging the BERT semantic vectors of all single standardized historical rare disease data within the historical data cluster. The vector dimension is consistent with the semantic vector of a single data point. For example, historical data cluster 1 contains 4,000 Dravet syndrome data points. Each data point generates a 256-dimensional semantic vector. The mean of the corresponding dimensions of all vectors is calculated, and finally, a 256-dimensional cluster local feature vector is obtained. This vector centrally reflects the common textual semantics of the data within the cluster (such as the high-frequency semantics of "Dravet syndrome", "SCN1A gene", and "febrile epilepsy").

[0082] In this embodiment, the global statistical feature matrix refers to a two-dimensional matrix (rows: statistical indicators, columns: indicator values ​​or categories) constructed by statistically analyzing global statistical indicators such as the proportion of rare disease main types, the number of overlapping pathogenic genes, and the frequency of core symptoms within historical data clusters. This matrix is ​​used to quantify the global data distribution characteristics of the clusters. For example, the global statistical feature matrix of historical data cluster 1 (Dravet syndrome) is shown in Table 2: Table 2 Global Statistical Feature Matrix In this embodiment, the construction of the global statistical feature vector is divided into two steps: matrix element normalization: statistical values ​​(such as the number of data entries) are normalized using the maximum value (x_norm=x / max_x), and the proportion / frequency is directly retained in the range [0,1]; matrix expansion: the two-dimensional matrix is ​​converted into a one-dimensional vector according to the index order - index value order; if the matrix dimension is too high (such as >50), it is reduced to 50 dimensions by PCA, and the vector is normalized by L2 after generation to ensure that it is consistent with the format of the local feature vector.

[0083] In this embodiment, the knowledge association feature matrix refers to a two-dimensional matrix (rows: knowledge association indicators, columns: entity type or relation type) constructed by retrieving knowledge association indicators such as the number of edges and hierarchical relationships associated with entities in the historical rare disease knowledge graph that are clustered with historical data. This matrix is ​​used to quantify the association strength between cluster data and the knowledge graph. For example, the core entities of historical data cluster 1 (Dravet syndrome) are "Dravet syndrome" (disease), "SCN1A gene" (gene), and "fever-induced epilepsy" (symptom). The knowledge association feature matrix is ​​shown in Table 3. Table 3 Knowledge Association Feature Matrix In this embodiment, the steps for constructing the knowledge association feature vector are as follows: calculate the standard deviation of the matrix; if the standard deviation is 0 (all elements are the same), then the scaling factor is set to 1; element scaling correction (x_corr=x / standard deviation); matrix expansion (same as the global statistical feature matrix); L2 normalization; if the vector dimension is too high (e.g., >30), then reduce the dimension to 30 through PCA to ensure coordination with the dimensions of other feature vectors (local 256 dimensions, global 11 dimensions, knowledge association 30 dimensions, total dimension 297 dimensions).

[0084] In this embodiment, for example, the cluster local feature vector V_local (256 dimensions, weights) 0.6), Global Statistical Feature Vector V_global (11-dimensional, weighted) 0.25), Knowledge-related Feature Vector V_knowledge (30-dimensional, weighted) 0.15), its weighting is: V_local_weighted = V_local × ,V_global_weighted=V_global× ,V_knowledge_weighted=V_knowledge× The splicing process involves splicing the three weighted vectors in sequence to obtain a cluster fusion feature vector of 256+11+30=297 dimensions: V_fusion=[V_local_weighted[0],...,V_local_weighted

[255] ,V_global_weighted[0],...,V_global_weighted

[10] ,V_knowledge_weighted[0],...,V_knowledge_weighted

[29] ].

[0085] In this embodiment, the manually labeled rare disease cluster hierarchical keywords refer to a set of keywords with a hierarchical structure labeled by medical experts (such as rare disease clinicians and geneticists) for clustering historical data. These keywords include core keywords (cluster core features) and related keywords (cluster secondary features), and are used as training labels for the keyword graph generation model.

[0086] In this embodiment, the hierarchical set of core keywords and related keywords refers to the set of keywords with a hierarchical structure generated after the keyword graph generation model is applied to historical data clustering. The core keywords are located at the highest level (level 1), and the related keywords are divided into different levels (level 2, level 3, etc.) according to their association strength with the core keywords, which is consistent with the manually labeled hierarchical keyword structure. For example, the hierarchical set generated by the model for historical data cluster 1 is as follows: core keyword "Dravet syndrome" (level 1, matching value 1.0); first-level associated keywords "pathogenic gene - SCN1A" (level 2, matching value 0.93), "symptoms - fever-induced epilepsy" (level 2, matching value 0.89), "age of onset - 0-1 year" (level 2, matching value 0.86); second-level associated keywords "mutation site - c.1837C>T" (level 3, matching value 0.82), "treatment drug - valproic acid" (level 3, matching value 0.79), "inheritance mode - autosomal dominant" (level 3, matching value 0.76). This set has an overlap of ≥90% with the manually labeled hierarchical structure and keywords.

[0087] In this embodiment, the matching value between keywords refers to the quantified semantic similarity between the core keyword and related keywords, and the value ranges from [0,1]. The larger the value, the stronger the association between the keywords. It is calculated based on the cosine similarity of the word vectors of the keywords. For example, the cosine similarity of the word vectors of the core keyword "Dravet syndrome" and the related keyword "pathogenic gene-SCN1A" is 0.93 (matching value 0.93); the cosine similarity of "pathogenic gene-SCN1A" and "mutation site-c.1837C>T" is 0.88 (matching value 0.88); the cosine similarity of "symptom-fever-induced epilepsy" and "age of onset-0-1 year" is 0.65 (matching value 0.65, no edge connection is constructed due to the weak association); the matching values ​​between all keywords constitute a matching value matrix, which is used to construct the edge weights of the keyword graph.

[0088] Topological keyword network (historical rare disease keyword map) refers to a network topology structure centered on a core keyword, with related keywords distributed around it according to hierarchy and matching value, and connected by weighted edges, forming a network topology that intuitively reflects the relationship and strength between keywords. For example, in the topological keyword network of historical data cluster 1, the core keyword "Dravet syndrome" is located at the center (the largest node size, with a diameter of 20px); the related keywords of level 2, "pathogenic gene - SCN1A", "symptoms - fever-induced epilepsy", and "age of onset - 0-1 year old", surround the core (node ​​diameter 15px), with edge weights of 0.93, 0.89, and 0.86 respectively (edge ​​thickness is positively correlated with weight, 0.93 corresponds to an edge width of 3px, 0.89 corresponds to 2.8px, and 0.86 corresponds to 2.6px); the related keywords of level 3, "mutation site - c.1837C>T", "treatment drug - valproic acid", and "inheritance mode - autosomal dominant", surround the keywords of level 2 (node ​​diameter 12px), with edge weights of 0.88 (connecting "pathogenic gene - SCN1A"), 0.79 (connecting "Dravet syndrome"), and 0.76 (connecting "Dravet syndrome"), forming a complete topological network.

[0089] The beneficial effects of the above technical solution are as follows: Local semantic features of the cluster are extracted through the BERT pre-trained model, combined with global statistical features and knowledge association features to ensure the integrity and representativeness of the features; a self-scaling attention mechanism is adopted to achieve weighted fusion of multiple features, highlighting the contribution of important features and improving feature quality; a hierarchical keyword set and topological keyword network are generated based on the manually annotated hierarchical keyword training model, and the manual annotation accurately reflects the core features and relationships of the cluster, constructing a structured and reusable keyword graph for each historical data cluster, providing a training basis for the generation of subsequent real-time data keyword graphs.

[0090] This invention provides a method for retrieving rare disease data, which involves constructing and training a keyword graph generation model, including: Obtain cluster fusion feature vector ,Will The input self-scaling attention mechanism assigns feature weights, wherein the weight calculation formula for the self-scaling attention mechanism is: ,in, For feature dimensions; The weight matrix is ​​a learnable weight matrix; It is a dynamic scaling factor based on the sparsity of rare disease characteristics, and for The reciprocal of the frequency of occurrence of the j1-th dimension feature in historical clinical data of rare diseases; based on right Weighted mapping of features across all dimensions yields key feature representations. ,in, This represents the j1-th dimension of the cluster fusion feature vector; We receive key feature representations and use 3×3 mixed-size convolutional kernels to extract local semantic association features between features. ,in, For convolution kernel; This is a convolution operation; For bias; It is the ReLU activation function; Given a historical rare disease knowledge graph corresponding to the input cluster entities, graph structure association features are extracted through two graph convolutional layers. The graph convolution operation is as follows: ,in, Let be the set of neighbors of node i; For the first Features of layer node j; For the first Features of layer node j; , These are the sizes of the neighbor sets of node i and node j, respectively, i.e., the number of neighbors; For the first Learnable weight matrix of layer graph convolutional layer; For the first Learnable bias vectors for layered graph convolutional layers; Based on the clinical occurrence rate and standard deviation of the knowledge association feature matrix between nodes in the historical rare disease knowledge graph The optimized Laplacian matrix is ​​obtained by correcting the Laplacian matrix L corresponding to the associated features of the graph structure. ,in, Let be the co-occurrence frequency of node i and node j in clinical cases of rare diseases; use Optimize the output features of the graph convolutional layer to obtain graph structure association features. ; by , Using discrete-continuous variable hierarchical structure as the conditional input to the generator, rare disease keywords are generated: Discrete variables Define the core keyword categories for the main types of rare diseases, where C is the number of core keyword categories; With continuous variables Define the correlation between subtypes, pathogenic genes, symptom-related keywords and core keywords, and It is positively correlated with the clinical incidence rate of the corresponding rare disease subtype; A discriminator is set between the generator and the feature input, and the discriminator adopts the WGAN-GP loss function to verify the semantic matching degree between the generated keywords and the rare disease cluster data. The WGAN-GP loss function is used to... for: ,in, For discriminators; To generate keywords; To accurately label keywords; For interpolation samples; This is the penalty coefficient; For mathematical expectation operators; Gradient operator; It is the square of the L2 norm.

[0091] In this embodiment, the learnable weight matrix In self-scaling attention mechanisms, this refers to the method used to fuse cluster feature vectors. The matrix represents the linear transformation of features across each dimension. The matrix parameters are continuously optimized through model training to adapt to the feature transformation requirements of the keyword graph generation task. The dimensions are: feature dimension d × feature dimension d (e.g., 297 × 297). It is a 297×297 matrix, where the 100th row corresponds to The linear transformation weights of the 100th dimension feature (rare mutation sites) are given, and the transformation weights of the 200th dimension feature (fever symptoms) are given in the 200th row.

[0092] In this embodiment, for example, a frequency statistics table is constructed to collect statistics. The frequency of each dimension of features in 100,000 historical clinical data of rare diseases is shown in Table 4: Table 4 Frequency of Occurrence It should be noted that, =1.0 / (frequency+ ), Normalization (scaling to the [1,100] interval) is required to prevent extreme values ​​from affecting the weight calculation. frequency is the frequency of occurrence.

[0093] The 3×3 mixed-size convolution kernel K refers to the kernel used to extract key feature representations. The convolutional kernel for local semantic association between features has a size of 3×3 (width×height). The mixed size refers to using both 3×3 and 1×1 convolutional kernels in the convolutional layer (1×1 is used for dimensionality compression, and 3×3 is used for local feature extraction) to balance feature extraction capability and computational efficiency.

[0094] Local semantic association features This refers to the representation of key features using 3×3 mixed-size convolutional kernels. After convolution, the extracted local semantic associations between features (such as the association between "disease-gene" and "gene-mutation site") are in compressed dimensions (e.g., 128 dimensions), containing collaborative semantic information between features. Specifically: The calculation process is "convolution operation → ReLU activation → batch normalization". The convolution operation uses PyTorch's conv2d function (stride=1), the ReLU activation function is nn.ReLU(), and the batch normalization uses nn.BatchNorm1d() (to accelerate training and prevent overfitting). The dimension is 128, and it is stored as a PyTorch tensor (shape=(1,128)). The local semantic feature extraction time is <30ms.

[0095] In this embodiment, the present invention employs two graph convolutional layers to progressively deepen the extraction of graph structure features. The graph convolutional layers are implemented based on the GCNConv class. The first layer has an input feature dimension of entity feature dimension (e.g., 64 dimensions) and an output dimension of 128 dimensions; the second layer has an input dimension of 128 dimensions and an output dimension of 256 dimensions; the neighbor set of the graph convolution operation... The adjacency matrix is ​​obtained from the knowledge graph (the adjacency matrix is ​​stored in a sparse matrix format to save memory); ReLU activation function and dropout (rate=0.2) are used between layers to prevent overfitting.

[0096] In this embodiment, the Laplacian matrix L refers to the matrix used to describe the graph structure, defined as L=DA, where D is the degree matrix (a diagonal matrix, Dii is the degree of node i), and A is the adjacency matrix (Aij=1 indicates that node i is connected to node j, otherwise it is 0), used in the graph convolutional layer to smooth node features and enhance the stability of the graph structure. For example, the adjacency matrix A of 3 nodes (Dravet syndrome D, SCN1A gene G, fever symptom S) in the historical rare disease knowledge graph is: The degree matrix D is: Then the Laplace matrix L = DA. This matrix reflects the connection relationships and degree distribution between nodes.

[0097] In this embodiment, The optimization process is as follows: graph convolution output features → and Dot product → ReLU activation → L2 normalization.

[0098] In this embodiment, the value of C is determined by medical experts in conjunction with the ICD-11 rare disease classification (C=10, covering the main rare disease categories); the mapping relationship between c1 and the core keyword categories is stored in a JSON configuration file (e.g., c1=2: epileptic encephalopathy, core keyword candidate set: Dravet syndrome, West syndrome, Lennox-Gastaut syndrome); the generator converts c1 into a 32-dimensional vector through an embedding layer and fuses it with the conditional input.

[0099] In this embodiment, The value is determined by querying a rare disease incidence database, which is updated quarterly. The mapping relationship with the incidence rate is expressed using a linear function ( =min(1.0, incidence rate / 0.1), ensuring the incidence rate is greater than 10%. =1.0, when the incidence rate is <1%. ≥0.1); the generator will pass through a fully connected layer (1-dimensional) Integration with other characteristics, such as the clinical incidence of "classical Dravet syndrome" (SCN1A gene heterozygous mutation) among the subtypes of Dravet syndrome being 0.08 (accounting for 8% of Dravet syndrome cases), then the associated keyword "classical" and the core keyword "Dravet syndrome" are related. =0.8.

[0100] The discriminator refers to the component in cGAN used to distinguish generated keywords. With real-name labeled keywords Binary classification networks, such as generators that generate keywords. =“Dravet syndrome → SCN1A gene → c.1837C>T mutation”, actual keywords =“Dravet syndrome → SCN1A gene → fever-induced epilepsy” (manually annotated), discriminator input and The word vectors are used to output D( ) = 0.3 (close to 0, judged as false), D( ) = 0.9 (close to 1, judged as true), through loss function feedback generator optimization, finally making D( =0.8 (close to the actual keyword score).

[0101] In this embodiment, the interpolation sample ,and The gradient follows a uniform distribution U[0,1]; the gradient is calculated using the torch.autograd.grad function, and the gradient norm is calculated using torch.norm; the penalty coefficient... Set it to 10 (experiments have shown that this value has the best training stability in the rare disease keyword generation task).

[0102] The beneficial effects of the above technical solution are as follows: By dynamically allocating feature weights through a self-scaling attention mechanism, the contribution of core features such as rare mutations and highly co-occurring nodes is highlighted; based on a two-layer graph convolution combined with an optimized Laplacian matrix, the clinical structural associations of the knowledge graph are accurately captured, making the graph structure features more consistent with clinical practice; the cGAN discrete-continuous mixed variable generation strategy is adopted to ensure the hierarchy and clinical relevance of the generated keywords; the WGAN-GP loss function solves the training instability problem through gradient penalty, improves the quality of keyword graph generation, and ensures the accuracy and efficiency of retrieval matching.

[0103] This invention provides a rare disease data retrieval system, such as... Figure 2 As shown, it includes: The infrastructure construction module is used to build a distributed storage blockchain and a distributed consensus blockchain with PBFT consensus mechanism. After integration, a distributed ledger is set up to obtain a rare disease database. Historical rare disease data and rare disease knowledge are collected and preprocessed. Based on the fusion algorithms of CNN-GRPO-DPA, BERT-LSTM-CRF-SVM, CNN-cGAN, and LSTM-GCN-MLP, heterogeneous data processing models, knowledge graph construction models, keyword graph generation models, and retrieval tag generation models are built respectively, and integrated into a rare disease data retrieval engine. A rare disease data retrieval platform including user login, search input, result visualization, and data download modules is built. The data processing and feature generation module is used to collect real-time rare disease data, preprocess it, call the heterogeneous data processing model to generate standardized real-time data, call the knowledge graph construction model to extract entities and relationships, and update the real-time rare disease knowledge graph; after clustering, call the keyword graph generation model to generate a real-time keyword graph, and call the search tag generation model to generate real-time search tags. The distributed storage module is used to slice real-time rare disease data according to ICD-11 encoding, store it in a distributed storage blockchain, call smart contracts to generate transaction data, and synchronize it to the distributed ledger after PBFT consensus. The knowledge mapping module is used to collect user search information, extract entities and relationships after preprocessing, calculate the Euclidean distance with the entities in the real-time knowledge graph, and map the graph structure search information. The retrieval and data extraction module is used to perform a coarse search in the ledger to obtain the target keyword map, a fine search to obtain the target storage address, extract fragments and splice them to obtain the target standardized real-time data; The visualization module is used to display target data in topology maps and tables through the retrieval platform.

[0104] The beneficial effects of the above technical solution are as follows: By constructing a blockchain database with a PBFT consensus mechanism, secure storage and tamper-proof protection of rare disease data are achieved, ensuring data integrity and reliability; a multi-algorithm fusion retrieval engine enables standardized integration of heterogeneous data, dynamic updating of knowledge graphs, and accurate generation of keyword graphs and search tags; a two-level retrieval mechanism of knowledge mapping and coarse and fine retrieval improves the accuracy and efficiency of retrieval; and a visualization module presents retrieval results intuitively, meeting the needs of clinical diagnosis and research for efficient acquisition of rare disease data, breaking down data silos, and promoting cross-domain resource sharing.

[0105] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for retrieving rare disease data, characterized in that, include: S1: Cloud Data Center Infrastructure Construction: Construct a distributed storage blockchain and a distributed consensus blockchain with PBFT consensus mechanism, integrate them, and set up a distributed ledger to obtain a rare disease database; collect historical rare disease data and rare disease knowledge, and after preprocessing, construct heterogeneous data processing models, knowledge graph construction models, keyword graph generation models, and retrieval tag generation models based on CNN-GRPO-DPA, BERT-LSTM-CRF-SVM, CNN-cGAN, and LSTM-GCN-MLP fusion algorithms, respectively, and integrate them into a rare disease data retrieval engine; build a rare disease data retrieval platform including user login, search input, result visualization, and data download modules; S2: Real-time data processing and feature generation: Collect real-time rare disease data, preprocess it, and then call the heterogeneous data processing model to generate standardized real-time data; call the knowledge graph construction model to extract entities and relationships, and update to obtain a real-time rare disease knowledge graph; after clustering, call the keyword graph generation model to generate a real-time keyword graph, and call the search tag generation model to generate real-time search tags. S3: Blockchain Distributed Storage: Real-time rare disease data is sharded according to ICD-11 encoding, stored in a distributed storage blockchain, and smart contracts are called to generate transaction data, which is then synchronized to the distributed ledger after PBFT consensus. S4: Knowledge Mapping of Retrieved Information: Collect user retrieval information, extract entities and relationships after preprocessing, calculate the Euclidean distance with entities in the real-time knowledge graph, and map to obtain graph structure retrieval information; S5: Distributed retrieval and data extraction: The target keyword map is obtained by coarse retrieval in the ledger, the target storage address is obtained by fine retrieval, and the fragments are extracted and spliced ​​to obtain the target standardized real-time data. S6: Results Visualization: The target data is displayed in topology maps and tables through the retrieval platform; This includes the integration into a rare disease data retrieval engine, comprising: We collected some rare disease knowledge and heterogeneous historical rare disease data from different data sources, and preprocessed them to obtain some preprocessed rare disease knowledge and some heterogeneous preprocessed historical rare disease data. Based on several heterogeneous preprocessed historical rare disease data, a heterogeneous data processing model is constructed using a fusion algorithm of deep learning and reinforcement learning, and several homogeneous standardized historical rare disease data are generated. Based on several standardized historical rare disease data, a knowledge graph construction model was constructed using natural language processing algorithms, resulting in a historical rare disease knowledge graph that includes several historical knowledge named entities and several historical knowledge entity relationships. Several standardized historical rare disease data are clustered to obtain several historical data clusters. Based on these historical data clusters, a keyword graph generation model is constructed using a deep learning algorithm to obtain the historical rare disease keyword graph for each historical data cluster. The historical rare disease keyword map is diffused to the corresponding historical data clusters to obtain several standardized historical rare disease data with historical rare disease keyword maps, and a retrieval tag generation model is constructed. By integrating heterogeneous data processing models, knowledge graph construction models, keyword graph generation models, and retrieval tag generation models, a rare disease data retrieval engine is obtained. Among them, the historical rare disease keyword map of each historical data cluster is obtained, including: The semantic vector of a single standardized historical rare disease data point within a historical data cluster is extracted using a BERT pre-trained model, and the local feature vector of the cluster is obtained. By statistically analyzing historical data, we can identify the proportion of rare disease main types, the number of overlapping pathogenic genes, and the frequency of occurrence of core symptoms within clusters. We can then construct a global statistical feature matrix and obtain a global statistical feature vector. The number of edges and hierarchical relationships associated with cluster entities in the historical rare disease knowledge graph are retrieved to construct a knowledge association feature matrix; the standard deviation of the knowledge association feature matrix is ​​calculated, and the matrix elements are scaled and corrected based on the standard deviation to obtain the knowledge association feature vector; Attention mapping and weighted concatenation are performed on the cluster local feature vector, global statistical feature vector, and knowledge association feature vector to obtain the cluster fusion feature vector; Construct and train a keyword graph generation model; Using manually annotated rare disease cluster-level keywords as labels, the cluster fusion feature vector is input into the keyword graph generation model for training; The trained model is applied to each cluster to generate a hierarchical set of core keywords and related keywords, and the matching value between keywords is calculated. Based on the matching value, a topological keyword network centered on the core keywords is constructed, thus obtaining the historical rare disease keyword map. The construction and training of the keyword graph generation model includes: Obtain cluster fusion feature vector ,Will The input self-scaling attention mechanism assigns feature weights, wherein the weight calculation formula for the self-scaling attention mechanism is: ,in, For feature dimensions; The weight matrix is ​​a learnable weight matrix; It is a dynamic scaling factor based on the sparsity of rare disease characteristics, and for The reciprocal of the frequency of occurrence of the j1-th dimension feature in historical clinical data of rare diseases; based on right Weighted mapping of features across all dimensions yields key feature representations. ,in, This represents the j1-th dimension of the cluster fusion feature vector; We receive key feature representations and use 3×3 mixed-size convolutional kernels to extract local semantic association features between features. ,in, For convolution kernel; This is a convolution operation; For bias; It is the ReLU activation function; Given a historical rare disease knowledge graph corresponding to the input cluster entities, graph structure association features are extracted through two graph convolutional layers. The graph convolution operation is as follows: ,in, Let be the set of neighbors of node i; For the first Features of layer node j; For the first Features of layer node j; , These are the sizes of the neighbor sets of node i and node j, respectively, i.e., the number of neighbors; For the first Learnable weight matrix of layered graph convolutional layers; For the first Learnable bias vectors for layered graph convolutional layers; Based on the clinical occurrence rate and standard deviation of the knowledge association feature matrix between nodes in the historical rare disease knowledge graph The optimized Laplacian matrix is ​​obtained by correcting the Laplacian matrix L corresponding to the associated features of the graph structure. ,in, Let be the co-occurrence frequency of node i and node j in clinical cases of rare diseases; use Optimize the output features of the graph convolutional layer to obtain graph structure association features. ; by , Using discrete-continuous variable hierarchical structure as the conditional input to the generator, rare disease keywords are generated: Discrete variables Define the core keyword categories for the main types of rare diseases, where C is the number of core keyword categories; With continuous variables Define the correlation between subtypes, pathogenic genes, symptom-related keywords and core keywords, and It is positively correlated with the clinical incidence rate of the corresponding rare disease subtype; A discriminator is set between the generator and the feature input, and the discriminator adopts the WGAN-GP loss function to verify the semantic matching degree between the generated keywords and the rare disease cluster data. The WGAN-GP loss function is used to... for: ,in, For discriminators; To generate keywords; To accurately label keywords; For interpolation samples; This is the penalty coefficient; For mathematical expectation operators; Gradient operator; It is the square of the L2 norm.

2. The rare disease data retrieval method according to claim 1, characterized in that, Access to rare disease databases, including: Identify several data servers in the storage partition and use an identity allocation mechanism to assign the data servers as storage nodes and consensus nodes; A distributed storage blockchain is constructed by connecting several storage nodes in a distributed manner, and a distributed consensus blockchain is constructed by connecting several consensus nodes in a distributed manner. By integrating distributed storage blockchain and distributed consensus blockchain, and setting up a distributed ledger, a rare disease database is obtained.

3. The rare disease data retrieval method according to claim 2, characterized in that, A coarse search of the ledger yields the target keyword map, a fine search reveals the target storage address, and fragments are extracted and concatenated to obtain standardized real-time target data, including: Based on the real-time retrieval information after knowledge mapping, a coarse search is performed in several real-time rare disease keyword maps in the distributed ledger of the rare disease database to obtain the target real-time rare disease keyword map and several corresponding alternative real-time retrieval tags. Based on the real-time retrieval information after knowledge mapping, a fine search is performed among several candidate real-time retrieval tags to obtain the target real-time retrieval tag and the corresponding target real-time storage address. Based on several target real-time storage addresses, extract the corresponding target real-time rare disease data fragments from several storage nodes of the distributed storage blockchain of the rare disease database. Data is spliced ​​together based on the target real-time rare disease data fragments to obtain the corresponding target standardized real-time data.

4. The rare disease data retrieval method according to claim 1, characterized in that, The system collects user search information, preprocesses it to extract entities and relationships, calculates the Euclidean distance between the entity and the real-time knowledge graph entity, and maps it to obtain graph structure search information, including: The rare disease data retrieval platform is used to collect user search information and preprocess the user search information to obtain preprocessed search information. The model is constructed using the knowledge graph of the rare disease data retrieval engine. Several real-time data named entities and several real-time data entity relationships are extracted from the preprocessed retrieval information, and the Euclidean distance between each data named entity and several real-time knowledge named entities in the real-time rare disease knowledge graph is obtained. The real-time knowledge named entity with the closest Euclidean distance is taken as the first target knowledge named entity of the real-time data named entity, and several real-time knowledge entity relations of the first target knowledge named entity are taken as target knowledge entity relations. The real-time knowledge named entity corresponding to the other side of each target knowledge entity relationship of the first knowledge named entity is used as the second target knowledge named entity. Map all first-target knowledge named entities, all target knowledge entity relationships, and all second-target knowledge named entities to the corresponding real-time data named entities to obtain the real-time data named entities after knowledge mapping. By integrating several real-time knowledge-mapped named entities and several real-time data entity relationships, graph structure retrieval information is obtained.

5. The rare disease data retrieval method according to claim 1, characterized in that, The knowledge graph construction model is based on an algorithm derived from bidirectional Transformer encoder representation-long short-term memory network-conditional random field-support vector machine. The knowledge graph construction model includes a word embedding module based on BERT algorithm, a semantic feature extraction module based on LSTM algorithm, a named entity output module based on CRF algorithm, and an entity relation output module based on SVM algorithm, which are connected in sequence.

6. The rare disease data retrieval method according to claim 1, characterized in that, The retrieval label generation model is constructed based on the Long Short-Term Memory Network-Graph Convolutional Network-Multilayer Perceptron algorithm, and includes a sequence feature extraction module based on the LSTM algorithm, a graph structure feature extraction module based on the GCN algorithm, and a retrieval label generation module based on the MLP algorithm. The sequence feature extraction module and the graph structure feature extraction module are both connected to the retrieval label generation module.

7. A rare disease data retrieval system, characterized in that, include: The infrastructure construction module is used to build a distributed storage blockchain and a distributed consensus blockchain with PBFT consensus mechanism. After integration, a distributed ledger is set up to obtain a rare disease database. Historical rare disease data and rare disease knowledge are collected and preprocessed. Based on the fusion algorithms of CNN-GRPO-DPA, BERT-LSTM-CRF-SVM, CNN-cGAN, and LSTM-GCN-MLP, heterogeneous data processing models, knowledge graph construction models, keyword graph generation models, and retrieval tag generation models are built respectively, and integrated into a rare disease data retrieval engine. A rare disease data retrieval platform including user login, search input, result visualization, and data download modules is built. The data processing and feature generation module is used to collect real-time rare disease data, and after preprocessing, it calls the heterogeneous data processing model to generate standardized real-time data. The knowledge graph is used to build a model to extract entities and relationships, and then updated to obtain a real-time rare disease knowledge graph. After clustering, the keyword graph generation model is called to generate a real-time keyword graph, and the search tag generation model is called to generate a real-time search tag. The distributed storage module is used to slice real-time rare disease data according to ICD-11 encoding, store it in a distributed storage blockchain, call smart contracts to generate transaction data, and synchronize it to the distributed ledger after PBFT consensus. The knowledge mapping module is used to collect user search information, extract entities and relationships after preprocessing, calculate the Euclidean distance with the entities in the real-time knowledge graph, and map the graph structure search information. The retrieval and data extraction module is used to perform a coarse search in the ledger to obtain the target keyword map, a fine search to obtain the target storage address, extract fragments and splice them to obtain the target standardized real-time data; The visualization module is used to display target data in topology maps and tables through the retrieval platform; This includes the integration into a rare disease data retrieval engine, comprising: We collected some rare disease knowledge and heterogeneous historical rare disease data from different data sources, and preprocessed them to obtain some preprocessed rare disease knowledge and some heterogeneous preprocessed historical rare disease data. Based on several heterogeneous preprocessed historical rare disease data, a heterogeneous data processing model is constructed using a fusion algorithm of deep learning and reinforcement learning, and several homogeneous standardized historical rare disease data are generated. Based on several standardized historical rare disease data, a knowledge graph construction model was constructed using natural language processing algorithms, resulting in a historical rare disease knowledge graph that includes several historical knowledge named entities and several historical knowledge entity relationships. Several standardized historical rare disease data are clustered to obtain several historical data clusters. Based on these historical data clusters, a keyword graph generation model is constructed using a deep learning algorithm to obtain the historical rare disease keyword graph for each historical data cluster. The historical rare disease keyword map is diffused to the corresponding historical data clusters to obtain several standardized historical rare disease data with historical rare disease keyword maps, and a retrieval tag generation model is constructed. By integrating heterogeneous data processing models, knowledge graph construction models, keyword graph generation models, and retrieval tag generation models, a rare disease data retrieval engine is obtained. Among them, the historical rare disease keyword map of each historical data cluster is obtained, including: The semantic vector of a single standardized historical rare disease data point within a historical data cluster is extracted using a BERT pre-trained model, and the local feature vector of the cluster is obtained. By statistically analyzing historical data, we can identify the proportion of rare disease main types, the number of overlapping pathogenic genes, and the frequency of occurrence of core symptoms within clusters. We can then construct a global statistical feature matrix and obtain a global statistical feature vector. The number of edges and hierarchical relationships associated with cluster entities in the historical rare disease knowledge graph are retrieved to construct a knowledge association feature matrix; the standard deviation of the knowledge association feature matrix is ​​calculated, and the matrix elements are scaled and corrected based on the standard deviation to obtain the knowledge association feature vector; Attention mapping and weighted concatenation are performed on the cluster local feature vector, global statistical feature vector, and knowledge association feature vector to obtain the cluster fusion feature vector; Construct and train a keyword graph generation model; Using manually annotated rare disease cluster-level keywords as labels, the cluster fusion feature vector is input into the keyword graph generation model for training; The trained model is applied to each cluster to generate a hierarchical set of core keywords and related keywords, and the matching value between keywords is calculated. Based on the matching value, a topological keyword network centered on the core keywords is constructed, thus obtaining the historical rare disease keyword map. The construction and training of the keyword graph generation model includes: Obtain cluster fusion feature vector ,Will The input self-scaling attention mechanism assigns feature weights, wherein the weight calculation formula for the self-scaling attention mechanism is: ,in, For feature dimensions; The weight matrix is ​​a learnable weight matrix; It is a dynamic scaling factor based on the sparsity of rare disease characteristics, and for The reciprocal of the frequency of occurrence of the j1-th dimension feature in historical clinical data of rare diseases; based on right Weighted mapping of features across all dimensions yields key feature representations. ,in, This represents the j1-th dimension of the cluster fusion feature vector; We receive key feature representations and use 3×3 mixed-size convolutional kernels to extract local semantic association features between features. ,in, For convolution kernel; This is a convolution operation; For bias; It is the ReLU activation function; Given a historical rare disease knowledge graph corresponding to the input cluster entities, graph structure association features are extracted through two graph convolutional layers. The graph convolution operation is as follows: ,in, Let be the set of neighbors of node i; For the first Features of layer node j; For the first Features of layer node j; , These are the sizes of the neighbor sets of node i and node j, respectively, i.e., the number of neighbors; For the first Learnable weight matrix of layered graph convolutional layers; For the first Learnable bias vectors for layered graph convolutional layers; Based on the clinical occurrence rate and standard deviation of the knowledge association feature matrix between nodes in the historical rare disease knowledge graph The optimized Laplacian matrix is ​​obtained by correcting the Laplacian matrix L corresponding to the associated features of the graph structure. ,in, Let be the co-occurrence frequency of node i and node j in clinical cases of rare diseases; use Optimize the output features of the graph convolutional layer to obtain graph structure association features. ; by , Using discrete-continuous variable hierarchical structure as the conditional input to the generator, rare disease keywords are generated: Discrete variables Define the core keyword categories for the main types of rare diseases, where C is the number of core keyword categories; With continuous variables Define the correlation between subtypes, pathogenic genes, symptom-related keywords and core keywords, and It is positively correlated with the clinical incidence rate of the corresponding rare disease subtype; A discriminator is set between the generator and the feature input, and the discriminator adopts the WGAN-GP loss function to verify the semantic matching degree between the generated keywords and the rare disease cluster data. The WGAN-GP loss function is used to... for: ,in, For discriminators; To generate keywords; To accurately label keywords; For interpolation samples; This is the penalty coefficient; For mathematical expectation operators; Gradient operator; It is the square of the L2 norm.

Images