Multi-path data search fuzzy recommendation method

Abstract

The invention discloses a multi-path data search fuzzy recommendation method, and relates to the technical field of computer processing, and the method comprises the following steps: S01, analyzing fuzzy semantics input by a user through a pre-trained BioBERT model, and generating a structured query logic containing an entity type and a relation path; s02, constructing a dynamic multi-dimensional relation graph based on the obtained structured query logic, synchronously querying a relational database and a non-relational graph database, and returning a candidate set; s03, aggregating encryption parameters by adopting a federated learning framework to generate a global model; s04, generating a recommendation list based on the real-time physiological indexes of the patient and the diagnosis stage; and S05, capturing the latest literature through a medical knowledge updating engine, and updating graph nodes. According to the method, the problems that a traditional method depends on manual maintenance and lacks dynamic and context understanding are solved, and adaptive support for complex medical scenes is realized by combining relational and non-relational databases, a federal learning framework and a semantic analysis technology.

Description

Technical Field

[0001] The present invention relates to the technical field of computer processing, and particularly to a multi-channel data search and fuzzy recommendation method. Background Art

[0002] Physicians on outpatient duty obtain the test reports of patients and determine the diseases in combination with the pathological characteristics of the patients, and then prescribe medications accordingly. Limited by the medical practice experience and personal ability of doctors, different doctors often have different views and medication methods for the same disease. Especially for newly independent outpatient physicians, due to their limited medical practice experience, they often need to consult hospital internship tutors or clinic directors for some difficult pathologies.

[0003] Therefore, in the prior art, historical pathological data is collected and a knowledge graph is created, so that when data is input, relevant experience information can be given for reference. Related technologies: such as the authorized patent in China, publication number CN112836058A, which discloses a medical knowledge graph establishment method and device, and a medical knowledge graph query method and device.

[0004] In the prior art including the above patents, the update of the knowledge graph depends on manual maintenance, and it is difficult to dynamically integrate the latest medical research results. Moreover, the recommended rules or the collaborative filtering recommendation method lack the understanding of the medical context and cannot adapt to complex diagnosis scenarios. In summary, how to solve this problem is expected to be well solved. Summary of the Invention

[0005] In view of the above technical problems, the technical solution adopted by the present invention is a multi-channel data search and fuzzy recommendation method, and the method includes the following steps: S01. Parse the fuzzy semantics input by the user through a pre-trained BioBERT model to generate a structured query logic including entity types and relationship paths; S02. Build a dynamic multi-dimensional relationship graph based on the obtained structured query logic, query the relational database and the non-relational graph database synchronously, and return a candidate set; S03. Adopt a federated learning framework to aggregate encrypted parameters to generate a global model; S04. Generate a recommendation list based on the patient's real-time physiological indicators and diagnosis stage; S05. Grab the latest literature through a medical knowledge update engine to update the graph nodes; S06. Integrate the results using a dynamic weight algorithm and return them to the client.

[0006] Preferably, the generation of the structured query logic in step S01 includes: S11. Use the BioBERT model to map the input statement s i to the medical entity space and output a probability distribution: , where BioBERT(s i ) is the feature vector of the input statement s i encoded by the BioBERT model, W q is the query classification matrix, b q is the bias term, and s i is the magic term semantic statement input by the user; S12. Obtain clinical trial data in real time and use a Transformer encoder to extract the feature vector of the treatment method.

[0007] Preferably, the construction of the dynamic multi-dimensional relationship graph in step S02 includes: S21. Mine the association weight between disease d and treatment plan t based on the graph attention network: , where is the learnable attention mechanism parameter vector, W is the weight matrix of the graph neural network, and N d is the set of neighbor nodes of disease node d; S22. Extract symptom time series data from the electronic health record and predict the disease evolution path through LSTM: , where h t is the LSTM hidden state at time step t, x t is the input symptom feature at time step t, h t-1 is the hidden state at time step t-1, is the predicted probability of the disease evolution path, Wy is the weight matrix of the output layer, and by is the bias term of the output layer.

[0008] Preferably, the federated learning framework in step S03 includes: S31. Inject differential Gaussian noise into the local parameter update of each medical institution node M i : ; where is the local parameter update amount of the i-th medical institution, is the gradient clipping function, is the gradient of the loss function of the i-th node, takes 1; S32. Use the Paillier homomorphic encryption algorithm for Δwi Encrypted aggregation: , where ΔW enc is the aggregated result of the encrypted global parameters, and N is the total number of medical institutions participating in federated learning.

[0009] Preferably, the generation of the recommended list in S04 includes: S41. Through the real-time physiological index P t and the filtering options in the diagnosis stage S, the filtering conditions are: , where r i is the i-th candidate solution in the recommended list, Match(r i , P t ) is the matching degree between the solution r i and the real-time physiological index P t , θ1 = 0.7, and Stage(r i ) is the diagnosis stage applicable to the solution r i .

[0010] S42. Perform cross-modal association on the structured diagnosis record D struct and the unstructured imaging report D image : , where Score(r j ) is the cross-modal association score, BERT(D struct ) is the BERT encoding vector of the structured diagnosis record, and ResNet(D image) ) is the ResNet feature vector of the unstructured imaging report.

[0011] Preferably, the medical knowledge update in step S05 includes: S51. Obtain the input literature materials and slice them to obtain entity relationships; S52. Extract the entity relationship triples (h, r, t), and update the confidence of the graph node u through entity linking: , where Conf(u) is the updated confidence of the graph node u, Confold(u) is the original confidence of the node u, is the TF-IDF weight of the node u extracted from the literature, and ; S53. Dynamically adjust the graph topology based on the node degree, and delete the isolated nodes with degrees lower than the threshold θ2.

[0012] Preferably, the dynamic weight algorithm in step S06 includes: S61. Calculate weights based on data timeliness source credibility and data from user feedback : where = 0.5, = 0.3, = 0.2; S62. Dynamically adjust the exposure weight according to the user click-through rate: , where CTR j is the click-through rate of recommended item c j , = 0.1.

[0013] The present invention has at least the following beneficial effects: 1. Through the pre-trained BioBERT model and Transformer encoder, high-precision parsing of the user's fuzzy semantics is realized. Combining with real-time PubMed clinical trial data, the ambiguity of medical terms is effectively eliminated, and the accuracy of the generated structured query logic is significantly improved.

[0014] 2. The graph attention network of the dynamic multi-dimensional relationship graph and the time series LSTM model can mine the high-order associations between diseases and treatment plans and the evolution laws of symptoms, improve the data association accuracy, and thus support in-depth reasoning in complex medical scenarios.

[0015] 3. In the federated learning framework, differential Gaussian noise injection and Paillier homomorphic encryption technology, while protecting the data privacy of each medical institution, through the AUC dynamic weight aggregation mechanism, realize the efficient update of the global model.

[0016] 4. The dynamic weight algorithm fuses multi-dimensions of timeliness, credibility, and user feedback, introduces dynamic exposure adjustment based on click-through rate feedback, and based on the medical knowledge update engine based on TF-IDF and node confidence decay mechanism, automatically eliminates outdated knowledge, making the system faster from literature scraping to graph updating, so as to make the recommended reference data scientific and timely. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0018] Figure 1 Flow chart of a multi - path data search fuzzy recommendation method provided in the first embodiment of the present invention; Figure 2 Flow chart of S01 provided in the first embodiment of the present invention; Figure 3 Flow chart of S02 provided in the first embodiment of the present invention; Figure 4 Flow chart of S03 provided in the first embodiment of the present invention; Figure 5 Flow chart of S04 provided in the first embodiment of the present invention; Figure 6 Flow chart of S05 provided in the first embodiment of the present invention. Detailed implementation manners

[0019] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present invention.

[0020] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above - mentioned drawings are used to distinguish similar objects, and do not necessarily need to describe a specific order or sequence. It should be understood that such used data can be interchanged under appropriate circumstances so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non - exclusive inclusion. For example, a process, method, system, product or server including a series of steps or units does not necessarily have to be limited to those clearly listed steps or units, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0021] Embodiment 1:

[0022] This embodiment provides a multi - path data search fuzzy recommendation method, and the method includes the following steps, as Figure 1 shown: S01. Parse the fuzzy semantics input by the user through a pre - trained BioBERT model to generate a structured query logic including entity types and relationship paths; Specifically, as Figure 2 shown, the generation of the above - mentioned structured query logic includes: S11. Use the BioBERT model to map the input statement s i to the medical entity space and output a probability distribution: , where BioBERT(s i ) is the feature vector of the input statement s i encoded by the BioBERT model, W q is the query classification matrix, b q is the bias term, and s i is the magic term semantic statement input by the user; S12. Obtain clinical trial data in real time, and use a Transformer encoder to extract the feature vector of the treatment method.

[0023] Furthermore, the above also includes the disambiguation processing of fuzzy semantics, so that the fuzzy semantics for constructing and generating the structured query logic including entity types and relationship paths can be sufficiently identified. The steps include: S13. Segment the statement s i , label the part-of-speech (such as noun, verb), and identify the candidate positions of key entities; S14. Extract the local context window (such as the first n words before and after) related to the entity relationship e in the statement s i , and retain the syntactic dependency relationship; S15. Use a pre-trained language model to generate the dynamic word vectors of the statement s i and the entity relationship e, and capture the context-sensitive semantics. For the polysemous forms of the entity relationship e (such as "apple" referring to a company or a fruit), generate candidate sense vectors through sub-word segmentation (WordPiece) or an external knowledge base (such as WordNet); S16. Calculate the cosine similarity, weighted, and weighted similarity between the local context vector of s i and the word / entity vector of e.

[0024] In the above technology, the fuzzy semantics input by the user (such as "targeted drugs for treating advanced lung cancer") is first deeply semantically parsed by a pre-trained BioBERT model. BioBERT is fine-tuned based on a medical domain corpus, and extracts the context features of the input statement through a multi-layer Transformer architecture, identifying the medical entities (such as "lung cancer", "targeted drugs") and implicit relationships (such as "treatment", "advanced") therein. Subsequently, the model maps the semantics to a structured query logic, including entity type annotation (disease, drug, stage) and association path (disease → treatment method → drug). At the same time, the system obtains the latest clinical trial data in real time through the PubMed API, and uses a Transformer decoder to parse the effectiveness and side effect conclusions of the treatment method, supplementing the dynamic data source of the query logic.

[0025] And use the pre-trained BioBERT model (a variant of BERT fine-tuned on medical literature) to encode the user input statement, and output the probability distribution of each word belonging to a specific medical entity (such as disease, drug). The Softmax function maps the high-dimensional vector to the entity category probability to solve the ambiguity problem in fuzzy semantics.

[0026] Obtain the clinical trial abstracts in real time through the PubMed API, and use the Transformer decoder to perform attention modeling on "treatment means - efficacy conclusion". The query vector of the decoder focuses on the treatment method, and the key-value pair extracts the efficacy data to generate an efficacy feature vector for enhancing the clinical relevance of the query logic.

[0027] S02. Construct a dynamic multi-dimensional relationship graph based on the obtained structured query logic, synchronously query the relational database and the non-relational graph database, and return the candidate set; Specifically, as Figure 3 shown, the construction of the above dynamic multi-dimensional relationship graph includes: S21. Mine the association weight between disease d and treatment plan t based on the graph attention network: , where is the learnable attention mechanism parameter vector, W is the weight matrix of the graph neural network, and N d is the set of neighbor nodes of disease node d; S22. Extract the symptom time series data from the electronic health record, and predict the disease evolution path through LSTM: , where h t is the LSTM hidden state at time step t, x t is the input symptom feature at time step t, h t-1 is the hidden state at time step t-1, is the predicted probability of the disease evolution path, Wy is the weight matrix of the output layer, and by is the bias term of the output layer.

[0028] In the above technology, a dynamic multi-dimensional relationship graph is constructed based on the Graph Attention Network (GAT). Nodes represent entities such as diseases, symptoms, and treatment plans, and edges represent the associations between entities (such as "causing", "taboo"). GAT calculates the weights between nodes through the attention mechanism. For example, the association strength between a disease and a treatment plan is determined by the following process: concatenating the embedding vectors of the disease node and the treatment plan node, calculating the attention coefficient through the LeakyReLU activation function, and generating the final association weight after normalization. At the same time, the system extracts symptom time series data from the electronic health record, uses the LSTM model to predict the disease evolution path (such as "cough → pneumonia → respiratory failure"), and injects the prediction result into the graph as a time series feature. During query, the relational database (storing structured diagnosis and treatment records) and the non-relational graph database (storing graph associations) are retrieved synchronously, and a candidate set is returned.

[0029] Furthermore, in the graph, the association weight between the disease node d and the treatment plan node t is calculated by GAT. First, linearly transform the node features h d ,h t to W g h d and W g h t . After concatenation, calculate the similarity with the attention vector a through the LeakyReLU activation function, and finally obtain the attention coefficient through Softmax normalization, which reflects the priority of the treatment plan.

[0030] S03. Adopt the federated learning framework to aggregate encrypted parameters and generate a global model; Specifically, as Figure 4 shown, the above federated learning framework includes: S31. Inject differential Gaussian noise into the local parameter update of each medical institution node M i : ; Among them, is the local parameter update amount of the i-th medical institution, is the gradient clipping function, is the gradient of the loss function of the i-th node, takes 1; S32. Encrypt and aggregate Δw i using the Paillier homomorphic encryption algorithm: , where ΔW enc is the aggregated result of the encrypted global parameters, and N is the total number of medical institutions participating in the federated learning.

[0031] In the above technology, when training the local models of each medical institution, differential privacy technology is used to add noise to the gradients, clip the gradients (limit the maximum norm C), and inject Gaussian noise , ensuring that a single piece of data cannot be inferred from the gradients. The encrypted parameters are uploaded to the central server through the Paillier homomorphic encryption algorithm, and the global model is aggregated and generated. The server dynamically allocates aggregation weights according to the performance of the validation sets of each node (such as the AUC value), and the nodes with higher performance contribute more to the global model. Finally, the global model is distributed to each node through secure multi-party computation (SMPC) to avoid privacy leakage caused by plaintext transmission.

[0032] S04. Generate a recommendation list based on the patient's real-time physiological indicators and diagnosis stage; Specifically, as Figure 5 shown, the generation of the above recommendation list includes: S41. Filter through the real-time physiological indicator P t and the diagnosis stage S filtering options, and the filtering conditions are: , where r i is the i-th candidate solution in the recommendation list, Match(r i , P t ) is the matching degree between the solution r i and the real-time physiological indicator P t , θ1 = 0.7, and Stage(r i ) is the diagnosis stage applicable to the solution r i .

[0033] S42. Perform cross-modal association on the structured diagnosis and treatment record D struct and the unstructured imaging report D image : , where Score(r j ) is the cross-modal association score, BERT(D struct ) is the BERT encoding vector of the structured diagnosis and treatment record, and ResNet(D image) is the ResNet feature vector of the unstructured imaging report.

[0034] As described above, the recommendation engine receives the patient's real-time physiological indicators (such as heart rate, blood oxygen saturation) and the diagnosis stage (such as "postoperative recovery period"), and filters the candidate set through the following rules: Calculate the Euclidean distance between the recommended item and the patient's indicators, and eliminate the items that exceed the threshold θ1. Only retain the recommended items labeled with the current diagnosis stage. For unstructured imaging reports (such as CT scan text descriptions), a cross-modal alignment model is used to map the image features and text features to the same space, and relevant treatment plans are matched through cosine similarity.

[0035] Further, the patient's real-time physiological indicators P t (such as blood pressure, heart rate) and the normalized feature vector of the recommended item r i are used to calculate the L2 distance and normalized into a matching score. Only retain the recommended items whose matching degree exceeds the threshold θ1 and conform to the current diagnosis stage S.

[0036] S05. Grab the latest literature through the medical knowledge update engine and update the graph nodes; Specifically, as Figure 6 shown, the above-mentioned medical knowledge update includes: S51. Obtain the input literature materials and slice them to obtain entity relationships; S52. Extract the entity relationship triples (h, r, t), and update the confidence of the graph node u through entity linking: , where Conf(u) is the updated confidence of the graph node u, Confold(u) is the original confidence of the node u, is the TF-IDF weight of the node u extracted from the literature, and ; S53. Dynamically adjust the graph topology based on the node degree, and delete the isolated nodes with degrees lower than the threshold θ2.

[0037] As described above, the knowledge update engine regularly grabs the latest literature from platforms such as PubMed and ClinicalTrials, and associates new entities (such as the new drug "Drug-X") with the existing nodes in the graph through entity linking technology. A context-aware model is used to identify entity mentions in the literature. Calculate the semantic similarity (based on BioBERT embedding) and co-occurrence frequency (TF-IDF) of candidate nodes. Integrate the historical confidence and freshness weight to update the node confidence. For controversial conclusions (such as reports of opposite effects of a certain drug), the system automatically triggers an artificial review process.

[0038] S06. Adopt a dynamic weight algorithm to integrate the results and return them to the client.

[0039] Specifically, the above-mentioned dynamic weight algorithm includes: S61. Calculate the weight according to the timeliness of data Source credibility and the data of user feedback , and calculate the weight: , where = 0.5, = 0.3, = 0.2; S62. Dynamically adjust the exposure weight according to the user click-through rate: , where CTR j is the click-through rate of the recommended item c j , = 0.1.

[0040] In summary, through the pre-trained BioBERT model and the Transformer encoder, high-precision parsing of the user's fuzzy semantics is achieved. Combining with real-time PubMed clinical trial data, the ambiguity of medical terms is effectively eliminated, and the accuracy of the generated structured query logic is significantly improved. Secondly, the graph attention network and the time series LSTM model of the dynamic multi-dimensional relationship graph can mine the high-order associations between diseases and treatment plans and the evolution laws of symptoms, improving the data association accuracy, so as to support in-depth reasoning in complex medical scenarios. Furthermore, the differential Gaussian noise injection and Paillier homomorphic encryption technology in the federated learning framework, on the premise of protecting the data privacy of each medical institution, realize the efficient update of the global model through the AUC dynamic weight aggregation mechanism. Further, the dynamic weight algorithm fuses multi-dimensions of timeliness, credibility, and user feedback, introduces dynamic exposure adjustment based on click-through rate feedback, and automatically eliminates outdated knowledge based on the TF-IDF and node confidence decay mechanism of the medical knowledge update engine, making the system faster from literature scraping to graph update, so as to make the recommended reference data scientific and timely.

[0041] Embodiment 2:

[0042] The embodiment of the present invention provides a non-transitory computer-readable storage medium, in which at least one instruction or at least one program segment is stored, and at least one instruction or at least one program segment is loaded and executed by a processor to implement the steps: Parse the fuzzy semantics input by the user through the pre-trained BioBERT model to generate structured query logic including entity types and relationship paths; Construct a dynamic multi-dimensional relationship graph based on the obtained structured query logic, synchronously query the relational database and the non-relational graph database, and return the candidate set; Adopt the federated learning framework to aggregate encryption parameters to generate a global model; Generate a recommended list based on the patient's real-time physiological indicators and diagnosis stage; The latest literature is fetched by a medical knowledge update engine to update the graph nodes; The results are integrated using a dynamic weight algorithm and returned to the client.

[0043] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the various embodiments provided in this application can include non-volatile and / or volatile memories. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0044] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the above-mentioned division of each functional unit and module is used as an example. In actual applications, the above functions can be allocated to different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above.

[0045] Embodiment 3:

[0046] The embodiment of the present invention provides an electronic device, including a processor and a memory. At least one instruction or at least one program segment is stored in the memory, and the at least one instruction or the at least one program segment is loaded and executed by the processor to implement the steps: Parse the fuzzy semantics of the user input through a pre-trained BioBERT model to generate a structured query logic including entity types and relationship paths; Build a dynamic multi-dimensional relationship graph based on the obtained structured query logic, synchronously query a relational database and a non-relational graph database, and return a candidate set; Adopt the federated learning framework to aggregate encrypted parameters and generate a global model; Generate a recommendation list based on the patient's real-time physiological indicators and diagnosis stage; Use the medical knowledge update engine to crawl the latest literature and update the graph nodes; Adopt the dynamic weight algorithm to integrate the results and return them to the client.

[0047] The above are only the preferred embodiments of the present invention, and do not impose any form of limitation on the present invention. Although the present invention has been disclosed above with the preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to equivalent embodiments by using the disclosed technical content within the scope of the technical solution of the present invention.

Claims

1. A multi-channel data search fuzzy recommendation method, characterized in that The method includes the following steps: S01. Parse the fuzzy semantics of the user input through a pre-trained BioBERT model to generate a structured query logic including entity types and relationship paths; S02. Build a dynamic multi-dimensional relationship graph based on the obtained structured query logic, query the relational database and the non-relational graph database synchronously, and return a candidate set; S03. Adopt a federated learning framework to aggregate encrypted parameters to generate a global model; S04. Generate a recommendation list based on the patient's real-time physiological indicators and the diagnosis stage; S05. Crawl the latest literature through a medical knowledge update engine to update the graph nodes; S06. Integrate the results using a dynamic weight algorithm and return them to the client.

2. The multi-channel data search and fuzzy recommendation method according to claim 1, characterized in that The generation of the structured query logic in step S01 includes: S11. Use the BioBERT model to map the input statement s i to the medical entity space and output a probability distribution: ， Among them, BioBERT(s i ) is the input statement s i The feature vector encoded by the BioBERT model, W q is the query classification matrix, b q is the bias term, s i is the magic term semantic statement input by the user; S12. Obtain clinical trial data in real time, and use a Transformer encoder to extract the feature vectors of treatment methods.

3. A multi-channel data search fuzzy recommendation method according to claim 1, characterized in that The construction of the dynamic multi-dimensional relationship graph in step S02 includes: S21. Mine the association weight between disease d and treatment plan t based on a graph attention network; ， Among them, is the learnable attention mechanism parameter vector, W is the weight matrix of the graph neural network, and N d is the set of neighbor nodes of the disease node d; S22. Extract symptom time series data from the electronic health record, and predict the disease evolution path through LSTM; ， h t is the LSTM hidden state at time step t, x t is the input symptom feature at time step t, h t-1 is the hidden state at time step t - 1, is the predicted probability of the disease evolution path, Wy is the weight matrix of the output layer, and by is the bias term of the output layer.

4. A multi-channel data search fuzzy recommendation method according to claim 1, characterized in that The federated learning framework in step S03 includes: S31. Inject differential Gaussian noise into the local parameter update of each medical institution node M i : ； wherein, is the local parameter update amount of the i-th medical institution, is the gradient clipping function, is the gradient of the loss function of the i-th node, takes 1; S32. Aggregate after encrypting Δw using the Paillier homomorphic encryption algorithm i Encryption: ， Among them, ΔW enc is the aggregated result of the encrypted global parameters, and N is the total number of medical institutions participating in federated learning.

5. A multi-channel data search fuzzy recommendation method according to claim 1, characterized in that The generation of the recommendation list in S04 includes: S41. Through the real-time physiological index P t and the diagnosis stage S filtering option, the filtering condition is: ， where r i is the i-th candidate solution in the recommendation list, and Match(r i , P t ) is the matching degree between the solution r i and the real-time physiological index P t , θ1 = 0.7, and Stage(r i ) is the applicable diagnostic stage of the solution r i ; S42. Cross-modal association is performed on the structured diagnosis and treatment record D struct and the unstructured imaging report D image : ， Among them, Score(r j ) is the cross-modal correlation score, BERT(D struct ) is the BERT encoded vector of the structured diagnosis and treatment record, and ResNet(D image) ) is the ResNet feature vector of the unstructured imaging report.

6. A multi-channel data search fuzzy recommendation method according to claim 1, characterized in that The medical knowledge update in step S05 includes: S51. Obtain the input literature materials and slice them to obtain entity relationships; S52. Extract the entity relationship triples (h, r, t), and update the confidence of the graph node u through entity linking; ， Among them, Conf(u) is the updated confidence of graph node u, and Confold(u) is the original confidence of node u. is the TF-IDF weight of node u extracted from the literature, and ; S53. Dynamically adjust the graph topology based on the node degree, and delete the isolated nodes with degrees lower than the threshold θ2.

7. A multi-channel data search fuzzy recommendation method according to claim 1, characterized in that The dynamic weight algorithm in step S06 includes: S61. According to the timeliness of data Source credibility and the data of user feedback , calculate the weight: , where = 0.5, = 0.3, = 0.2; S62. The exposure weight is dynamically adjusted according to the user click-through rate: , where CTR j is the click-through rate of the recommended item c j , = 0.

1.

8. A non-transitory computer-readable storage medium storing at least one instruction or at least one program segment therein, characterized in that, The at least one instruction or the at least one program is loaded and executed by a processor to implement the steps of the multi-channel data search fuzzy recommendation method as described in any one of claims 1-7.

9. An electronic device, characterized in that, It includes a processor and a memory, and at least one instruction or at least one program is stored in the memory, and the at least one instruction or the at least one program is loaded and executed by a processor to implement the steps of the multi-channel data search fuzzy recommendation method as described in any one of claims 1-7.

Images