A gout staging prediction method and system based on federated learning and a knowledge graph, and a storage medium

Abstract

The application discloses a gout staging prediction method and system based on federated learning and a knowledge graph and a storage medium, relates to the field of intelligent medical treatment, and aims at solving the problems that the existing machine learning model is difficult to comprehensively and accurately capture key information in case data, privacy data is prone to be leaked, and data islands exist among hospitals. A client constructs a keyword-related knowledge graph subgraph based on patient medical record data sets; a center server constructs an R-GCN model; the client trains the R-GCN model by using the knowledge graph subgraph; model parameters are sent to the center server for aggregation after being encrypted; and the model is trained until the model meets the training condition. The trained model is used to encode the knowledge graph subgraph to obtain encoding information R1; text feature data in the data set is vectorized and encoded, and numerical features are transversely connected to obtain encoding information R2; and after the encoding information R1 and R2 are transversely connected, a deep learning network model is used to classify and predict different stages of gout.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical technology, and more specifically, to a method, system, and storage medium for predicting gout staging based on federated learning and knowledge graphs. Background Technology

[0002] With the development of medical informatization and digitalization, the medical field has accumulated a vast amount of data, including electronic medical records, medical research literature, and clinical trial results. The processing and analysis of this data are of great significance for improving medical quality, promoting medical research, and developing new therapies. Knowledge graphs and graph neural networks, as advanced artificial intelligence technologies, are increasingly widely used in the medical field. Knowledge graphs represent patients in a structured form, including patient information, symptoms, physical examination records, test results, and other entities and their interrelationships. Graph neural networks learn node representations from this graph structure, capturing the structural and semantic information of the nodes, and classifying and predicting gout stages.

[0003] Traditional machine learning models are centralized learning methods that require integrating patient data from various hospitals. This integration process increases the risk of privacy breaches. Furthermore, due to the high privacy requirements of medical data, some hospitals are unwilling to share data, creating data silos. Additionally, existing machine learning models struggle to comprehensively and accurately capture key information from case data. Summary of the Invention

[0004] The technical problem to be solved by this invention is:

[0005] Existing machine learning models struggle to capture key information from case data comprehensively and accurately, and pose risks of privacy breaches and data silos between hospitals.

[0006] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0007] This invention provides a method for predicting gout staging based on federated learning and knowledge graphs. The method is based on a federated learning model, which includes a central server and N clients, and includes the following steps:

[0008] S1. The client obtains patient medical records from the hospital, and after data preprocessing, obtains a medical dataset S. Based on the dataset S, feature extraction is performed to obtain a feature dataset F1. According to different feature types, the feature dataset F1 is divided into a numerical feature dataset F2 and a text feature dataset F3.

[0009] S2. Construct a graph structure based on dataset S to form a knowledge graph G; construct keyword groups from case feature data, query the entities of keywords in the knowledge graph, and obtain knowledge graph subgraphs G1, G2...Gn related to the keywords;

[0010] S3. The central server has built an R-GCN model and initialized the model;

[0011] S4. The central server distributes the model to each client. Each client trains the R-GCN model based on the knowledge graph subgraph. After training, noise is added to the model parameter data for encryption, and the data is compressed. The model parameters are then uploaded to the central server.

[0012] S5. The central server performs noise reduction processing on the model parameter data of each client, then aggregates the data, updates the model parameters, and obtains the global model.

[0013] S6. The central server verifies the effect of the global model. If the result does not meet the conditions, the operation of S4-S5 is repeated until the result meets the conditions and training is terminated.

[0014] S7. The trained R-GCN model is used to encode the knowledge graph subgraph to obtain the encoded information R1.

[0015] S8. Based on the improved MC-BERT model, the text feature data in F3 is vectorized and encoded. The encoded information is horizontally connected with the numerical feature data in the normalized F2 to obtain the encoded information R2. The improved MC-BERT model uses multiple masking layers in the encoding layer to mask phrases and professional medical texts. At the same time, a KeywordsEmbedding layer is introduced in the input layer to obtain a text representation that integrates keyword information.

[0016] S9. After horizontally connecting the encoded information R1 and the encoded information R2, the deep learning network model is used to classify and predict the different stages of gout.

[0017] Furthermore, S4 employs differential privacy to add random noise to the model parameter data, obfuscating the original dataset to achieve encryption of the model parameters, i.e.:

[0018] P r [M(x)∈S]-P r [M(x')∈S]≤ε

[0019] Where x and x' are adjacent datasets, M represents the algorithm for adding noise to the data, S represents the result of adding noise to the data, and P r Let ε represent the probability, and let ε represent the privacy parameter of differential privacy.

[0020] Furthermore, the noise added to the data is Gaussian noise, and the density function of Gaussian noise is:

[0021]

[0022] Where f(x) is the probability density function of Gaussian noise, σ is the standard deviation of the noise, μ is the mean of the noise, and x is the value of the noise.

[0023] Furthermore, S5 also includes weighted averaging of model parameter data from each client and then aggregating the data.

[0024] Furthermore, S6 introduces the range and step size of hyperparameters for cross-validation, and selects the best-performing parameter combination as the hyperparameters of the global model. Its objective function is:

[0025]

[0026] Where, N i r Let c represent the set of neighboring nodes of node i whose relation is r. i,r It is a regularization constant with a value of It is a linear transformation function.

[0027] Furthermore, the classification model described in S9 employs a deep feedforward neural network model.

[0028] Furthermore, the deep feedforward neural network model uses the ReLU function as the activation function for each hidden layer and constructs a softmax function to determine the stage of gout.

[0029] A gout staging prediction system based on federated learning and knowledge graphs, the system having program modules corresponding to the steps of the method described in any of the above technical solutions, and executing the steps in the gout staging prediction method based on federated learning and knowledge graphs at runtime.

[0030] A computer-readable storage medium storing a computer program configured to, when invoked by a processor, implement the steps of the gout staging prediction method based on federated learning and knowledge graphs as described in any of the above technical solutions.

[0031] Compared with the prior art, the beneficial effects of the present invention are:

[0032] This invention proposes a gout staging prediction method based on federated learning and knowledge graphs. By constructing a knowledge graph subgraph and integrating a federated learning framework on top of a graph neural network, the accuracy of gout staging diagnosis is improved, while simultaneously addressing the risks of privacy data leakage and the problem of data silos. This invention employs an improved MC-BERT model to vectorize textual feature data. The model uses multiple masking layers at the encoding level to mask phrases and professional medical text; simultaneously, a Keyword Embedding layer is introduced at the input layer, helping the model better understand and utilize important content in the text information, improving the model's sensitivity to key information and its ability to capture local information in text sequences, and enabling it to capture key information in the data more comprehensively and accurately. This invention also introduces numerical augmentation methods to fully mine the potential features of numerical information, allowing the model to more comprehensively consider the relationship between text and numerical information, improving the overall understanding of electronic medical records. Attached Figure Description

[0033] Figure 1 This is a flowchart of the gout staging prediction method based on federated learning and knowledge graph in an embodiment of the present invention.

[0034] Figure 2 This is a diagram of the federated learning architecture in an embodiment of the present invention;

[0035] Figure 3 These are the keywords and entities used in the embodiments of this invention;

[0036] Figure 4 This is a knowledge graph subgraph in an embodiment of the present invention.

[0037] To enable those skilled in the art to better understand the present invention, exemplary embodiments or examples of the present invention will be described below in conjunction with the accompanying drawings. Obviously, the described embodiments or examples are merely some, not all, of the embodiments or examples of the present invention. All other embodiments or examples obtained by those skilled in the art based on the embodiments or examples of the present invention without inventive effort should fall within the scope of protection of the present invention.

[0038] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0039] Specific Implementation Plan 1: (e.g.) Figure 1 and Figure 2 As shown, this invention provides a method for predicting gout staging based on federated learning and knowledge graphs. The method is based on a federated learning model, which includes a central server and N clients, and includes the following steps:

[0040] S1. The client obtains patient medical records from the hospital, and after data preprocessing, obtains a medical dataset S. Based on the dataset S, feature extraction is performed to obtain a feature dataset F1. According to different feature types, the feature dataset F1 is divided into a numerical feature dataset F2 and a text feature dataset F3.

[0041] S2. Construct a graph structure based on dataset S to form a knowledge graph G; construct keyword groups from case feature data, query the entities of keywords in the knowledge graph, and obtain knowledge graph subgraphs G1, G2...Gn related to the keywords;

[0042] S3. The central server has built an R-GCN model and initialized the model;

[0043] S4. The central server distributes the model to each client. Each client trains the R-GCN model based on the knowledge graph subgraph. After training, noise is added to the model parameter data for encryption, and the data is compressed. The model parameters are then uploaded to the central server.

[0044] S5. The central server performs noise reduction processing on the model parameter data of each client, then aggregates the data, updates the model parameters, and obtains the global model.

[0045] S6. The central server verifies the effect of the global model. If the result does not meet the conditions, the operation of S4-S5 is repeated until the result meets the conditions and training is terminated.

[0046] S7. The trained R-GCN model is used to encode the knowledge graph subgraph to obtain the encoded information R1.

[0047] S8. Vectorize the text feature data in F3, normalize the numerical feature data in F2, and connect the two horizontally to obtain the encoded information R2.

[0048] S9. After horizontally connecting the encoded information R1 and the encoded information R2, the deep learning network model is used to classify and predict the different stages of gout.

[0049] In this implementation plan: the feature set S shall include at least: age, gender, body mass index, chief complaint, present medical history, and other medical history. The examination results data shall include at least: blood and urine biochemical test results.

[0050] The numerical feature data in F2 is normalized by subtracting the mean from the original value and then dividing by the standard deviation, so that the resulting encoded information satisfies a normal distribution with a mean of 0 and a standard deviation of 1.

[0051] In the process of constructing the graph structure, entity recognition, relation extraction, and attribute extraction are performed on the dataset S to construct a graph structure G consisting of nodes (representing entities) and edges (representing relations between entities).

[0052] Extract keywords such as painful joints from the feature set F1; for example... Figure 3 As shown, the professional terminology used by doctors when writing medical records is extracted as keywords. For example, when describing painful joints, terms like left knee joint, right knee joint, left wrist joint, and right wrist joint are used. First, disease-related medical professional terms are collected to construct a keyword thesaurus. Then, the data in the dataset is matched with the data in the keyword thesaurus to extract keyword groups from the medical record data, denoted as K = {k1, k2, ..., kn}.

[0053] like Figure 4 As shown, when querying for entities with the keyword K in the knowledge graph G, a one-hop query is performed to obtain knowledge graph subgraphs G1, G2...Gn related to K; adjacency matrices A1, A2...An are constructed and uploaded respectively, and encoded using the R-GCN model.

[0054] This implementation uses the FR-CGN model for training, combining graph convolutional networks with relation types. It trains a relation graph convolutional layer on a subgraph of the knowledge graph, then computes the embedding representation for each node. These embedding representations are used to construct the next relation graph convolutional layer. It iteratively trains the relation graph convolutional layers until the node embedding representations converge, ultimately building a neural network model based on relation graph convolutions.

[0055] The improved MC-BERT is trained on medical data to better handle issues such as the long-tail distribution of medical vocabulary. The model employs multiple masking layers in the encoding layer, adding masking layers for phrases and for specific medical text. In each masking layer, the model divides the input text sequence into multiple blocks according to fixed rules. A portion of the words in each block are masked, while the remaining words participate in model training. The output of each masking layer is passed to the next masking layer for further processing until the final layer outputs the representation of the entire sequence.

[0056] Because each masking layer can focus on different lexical chunks and learn their specific linguistic features, the multi-masking MC-BERT model has a strong ability to capture local information in text sequences. Furthermore, since each masking layer can be trained independently, the improved MC-BERT also has a certain degree of parallel training capability.

[0057] A Keyword Embedding layer was introduced at the input layer of the model to obtain a text representation incorporating keyword information. Based on this, a text representation incorporating numerical data was obtained by horizontally linking the expanded numerical data. This allows the model to more comprehensively and accurately capture key information and numerical data from patient medical records, improving its ability to express electronic medical records. Numerical augmentation was employed, integrating numerical information as enhancement terms into the text representation. Given that the units and orders of magnitude of numerical information may be inconsistent, the input numerical terms were normalized to maintain consistency.

[0058] Specific Implementation Plan Two: In S4, differential privacy is used to add random noise to the model parameter data, obfuscating the original dataset to achieve encryption of the model parameters, that is:

[0059] P r [M(x)∈S]-P r [M(x')∈S]≤ε

[0060] Where x and x' are adjacent datasets, M represents the algorithm for adding noise to the data, S represents the result of adding noise to the data, and P r Let ε represent the probability, and let ε represent the privacy parameter of differential privacy. This implementation scheme is otherwise the same as specific implementation scheme one.

[0061] After training on the client side, differential privacy technology is used to add random noise to the model parameters, obfuscating the original dataset and thus encrypting the original model parameters. The formula is defined as follows:

[0062] P r [M(x)∈S]≤exp(ε)P r [M(y)∈S]+δ

[0063] Where P r [M(x)∈S] represents the probability of obtaining result S by executing algorithm M on dataset x containing sensitive data. ε is the privacy parameter of differential privacy, representing the strength of privacy protection. x and y are datasets that differ by only one element (also called adjacent datasets). δ is an additional parameter, also called the distortion parameter, used to constrain the trade-off between privacy protection and data availability. In differential privacy, the goal of privacy protection is to ensure that P... r [M(x)∈S] and P r [M(y)∈S] serves as a deviation constraint; if the deviation is small, it ensures that private data will not be leaked.

[0064] Specific Implementation Plan 3: Since the data being processed is vector data, Gaussian noise is used to add noise to the data. The density function of Gaussian noise is:

[0065]

[0066] Where f(x) is the probability density function of Gaussian noise, σ is the standard deviation of the noise, which determines the intensity and width of the noise distribution, μ is the mean of the noise, which determines the center location of the noise, and x is the value of the noise. This implementation scheme is otherwise the same as specific implementation scheme two.

[0067] The central server acquires data from the client after differential privacy processing. This data typically includes encrypted model parameters, node features with added differential privacy noise, or other sensitive information. The central server then removes the differential privacy noise from the encrypted data to enable further processing and analysis. This usually involves inverse operations on the noise or the use of specific decryption techniques. Next, the central server aggregates the denoised data to construct a global model.

[0068] Specific Implementation Plan Four: S5 also includes weighted averaging of model parameter data from each client, followed by data aggregation. This implementation plan is otherwise identical to Specific Implementation Plan One.

[0069] Specific implementation plan five: In S6, the range and step size of hyperparameters are introduced for cross-validation, and the best-performing parameter combination is selected as the hyperparameters of the global model. The objective function is:

[0070]

[0071] in, Let c represent the set of neighboring nodes of node i whose relation is r. i,r It is a regularization constant with a value of It is a linear transformation function. This implementation scheme is otherwise the same as specific implementation scheme one.

[0072] When classifying datasets containing both large and small class samples, accuracy is typically used as a performance metric for the model. Since large class samples are the dominant samples in the dataset, a wider range of data features can be captured during training. Therefore, the accuracy for large class identification is often high, while the accuracy for small class identification is lower. Although traditional evaluation metrics may yield good results in this situation, the training results are clearly unsatisfactory. Therefore, it is necessary to find evaluation methods suitable for handling "imbalanced data".

[0073] For the actual classification of imbalanced datasets and the classification results determined by the algorithm, they are divided into 4 categories:

[0074] (1) True Positive (TP): It is judged as a positive sample and is actually a positive sample.

[0075] (2) False Positive (FP): The sample is judged as positive but is actually a negative sample.

[0076] (3) True Negative (TN): It is judged as a negative sample and is actually a negative sample.

[0077] (4) False Negative (FN): The sample is judged as negative when it is actually a positive sample.

[0078] This implementation plan uses evaluation metrics, primarily precision (P), recall (R), and F1-score, to comprehensively evaluate the results. The formulas are shown below:

[0079]

[0080] Specific Implementation Scheme Six: The deep learning network model described in S9 adopts a deep feedforward neural network model. All other aspects of this implementation scheme are the same as Specific Implementation Scheme One.

[0081] Specific implementation plan seven: Determine the label set D based on the number of gout disease stages, and perform multi-classification of the medical dataset based on a deep feedforward neural network model;

[0082] The number of neurons in the input layer is equal to the number of codes encoded in the encoding information R1 and R2;

[0083] The number of neurons in the output layer corresponds to the number of disease stages, i.e., the number of numerical values ​​in the label set D. The ReLU function is used as the activation function for each hidden layer, and a softmax function is constructed to determine the gout stage. This implementation scheme is otherwise identical to specific implementation scheme six.

[0084] The gout staging prediction method (algorithm) based on federated learning and knowledge graph proposed in this invention is the underlying technical core of this invention, and various products can be derived based on the algorithm.

[0085] Based on the method proposed in this invention, a gout staging prediction system based on federated learning and knowledge graphs is developed using a programming language. This system has program modules corresponding to the steps of the above-mentioned technical solution, and executes the steps in the above-mentioned gout staging prediction method based on federated learning and knowledge graphs when running.

[0086] The developed system (software) computer program is stored on a computer-readable storage medium, and the computer program is configured to implement the steps of the above-described gout staging prediction method based on federated learning and knowledge graphs when called by a processor. In other words, the invention is materialized on a carrier, becoming a computer program product.

[0087] Various implementations of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, application-specific integrated circuits (ASICs), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include: implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0088] The computational programs (also referred to as programs, software, software applications, or code) of this invention include machine instructions of a programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or apparatus (e.g., disk, optical disk, memory, programmable logic device PLD) for providing machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions as machine-readable signals. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0089] Example 1

[0090] The dataset used in this embodiment was collected from the gout clinic of a hospital. Since the original electronic medical records of the hospital contain a wealth of information, including a large amount of irrelevant information such as patient personal information, medication information, and hospitalization information, information irrelevant to the experimental purpose was removed from the original electronic medical records to reduce data dimensionality and protect patient privacy.

[0091] To further investigate the roles of text representation and knowledge representation in the model, and to verify the importance of numerical data and keyword information, a series of ablation experiments were conducted, including the removal of numerical data, keyword information, and knowledge representation (i.e., representation learning of knowledge graph subgraphs). The experimental results are shown in Table 1.

[0092] Table 1

[0093]

[0094] Therefore, it can be concluded that knowledge representation contributes most significantly to model performance, followed by the numerical data module, and finally keyword information. This embodiment verifies the effectiveness of knowledge representation in intelligent diagnosis, while highlighting the importance of numerical data.

[0095] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A method for predicting gout staging based on federated learning and knowledge graphs, characterized in that, The method is based on a federated learning model, which includes a central server and N clients, and includes the following steps: S1. The client obtains patient medical records from the hospital, and after data preprocessing, obtains a medical dataset S. Based on the dataset S, feature extraction is performed to obtain a feature dataset F1. According to different feature types, the feature dataset F1 is divided into a numerical feature dataset F2 and a text feature dataset F3. S2. Construct a graph structure based on dataset S to form a knowledge graph G; construct keyword groups from case feature data, query the entities of keywords in the knowledge graph, and obtain knowledge graph subgraphs G1, G2...Gn related to the keywords; S3. The central server has built an R-GCN model and initialized the model; S4. The central server distributes the model to each client. Each client trains the R-GCN model based on the knowledge graph subgraph. After training, noise is added to the model parameter data for encryption, and the data is compressed. The model parameters are then uploaded to the central server. S5. The central server performs noise reduction processing on the model parameter data of each client, then aggregates the data, updates the model parameters, and obtains the global model. S6. The central server verifies the effect of the global model. If the result does not meet the conditions, the operation of S4-S5 is repeated until the result meets the conditions and training is terminated. S7. The trained R-GCN model is used to encode the knowledge graph subgraph to obtain the encoded information R1. S8. Based on the improved MC-BERT model, the text feature data in F3 is vectorized and encoded. The encoded information is horizontally connected with the numerical feature data in the normalized F2 to obtain the encoded information R2. The improved MC-BERT model uses multiple masking layers in the encoding layer to mask phrases and professional medical texts. At the same time, a KeywordsEmbedding layer is introduced in the input layer to obtain a text representation that integrates keyword information. S9. After horizontally connecting the encoded information R1 and the encoded information R2, the deep learning network model is used to classify and predict the different stages of gout.

2. The gout staging prediction method based on federated learning and knowledge graph as described in claim 1, characterized in that, S4 employs differential privacy to add random noise to the model parameter data, obfuscating the original dataset and thus encrypting the model parameters. P r [M(x)∈S]-P r [M(x')∈S]≤ε Where x and x' are adjacent datasets, M represents the algorithm for adding noise to the data, S represents the result of adding noise to the data, and P r Let ε represent the probability, and let ε represent the privacy parameter of differential privacy.

3. The gout staging prediction method based on federated learning and knowledge graph as described in claim 2, characterized in that, The noise added to the data is Gaussian noise, and the density function of Gaussian noise is: Where f(x) is the probability density function of Gaussian noise, σ is the standard deviation of the noise, μ is the mean of the noise, and x is the value of the noise.

4. The gout staging prediction method based on federated learning and knowledge graph as described in claim 1, characterized in that, S5 also includes weighted averaging of model parameter data from each client and then aggregating the data.

5. The gout staging prediction method based on federated learning and knowledge graph as described in claim 1, characterized in that, S6 introduces the range and step size of hyperparameters for cross-validation, and selects the best-performing parameter combination as the hyperparameters of the global model. Its objective function is: in, Let c represent the set of neighboring nodes of node i whose relation is r. i,r It is a regularization constant with a value of It is a linear transformation function.

6. The gout staging prediction method based on federated learning and knowledge graph as described in claim 1, characterized in that, The classification model described in S9 uses a deep feedforward neural network model.

7. The gout staging prediction method based on federated learning and knowledge graphs according to claim 6, characterized in that, The deep feedforward neural network model uses the ReLU function as the activation function for each hidden layer and constructs a softmax function to determine the stage of gout.

8. A gout staging prediction system based on federated learning and knowledge graphs, characterized in that, The system has a program module corresponding to the steps of the method described in any one of claims 1 to 7, and executes the steps in the above-described gout staging prediction method based on federated learning and knowledge graph when it is run.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program configured to, when invoked by a processor, implement the steps of the gout staging prediction method based on federated learning and knowledge graphs as described in any one of claims 1 to 7.

Images