Diabetes risk prediction method based on TCN incremental learning, terminal equipment and storage medium

By adopting the TCN incremental learning method in the diabetes knowledge graph, combined with TransR and TCN, the problems of insufficient correlation and difficult to capture the timing dynamics of the diabetes knowledge graph are solved, and the accuracy and scalability of link prediction are improved, which is suitable for clinical practice. Domain timing knowledge graph.

CN119028580APending Publication Date: 2024-11-26CENT SOUTH UNIV

Patent Information

Application Number
CN202411083948.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-08
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

The prior art lacks correlation in the diabetes knowledge graph, which cannot fully reflect the relationship between real entities, and traditional methods are difficult to effectively cope with the dynamics of the timing knowledge graph, resulting in insufficient accuracy of link prediction.

Method used

Using a method based on TCN incremental learning, combining TransR conversion model and TCN time convolution network, the triplets are vectorized and incremental calculations are performed to capture timing information and contextual associations, and more accurate feature vectors are extracted for diabetes risk. predict.

Benefits of technology

It improves the accuracy of diabetes knowledge graph link prediction, highlights semantic and timing information in clinical facts, has strong scalability and robustness, and is suitable for clinical field timing knowledge graphs of different scales and complexities.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119028580A_ABST
    Figure CN119028580A_ABST
Patent Text Reader

Abstract

The invention discloses a diabetes risk prediction method based on TCN incremental learning, a terminal device and a storage medium, in combination with a TransR conversion model and a TCN time convolutional network, time sequence information and context association information in a map are effectively utilized, and the accuracy of link prediction is improved. According to the TransR model, entity projection is carried out in a relation space containing different semantics, so that effective vectorization of entities and relations in the atlas is realized. According to the method, the diversity of the entities under different relations can be captured, and a solid foundation is provided for subsequent time sequence analysis. The TCN is used for processing entities and relation sequences in the time sequence knowledge graph, so that more accurate feature vectors are extracted.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to intelligent medical technology, and in particular to a diabetes risk prediction method based on TCN incremental learning. Background Art

[0002] Diabetes is a metabolic disease characterized by hyperglycemia caused by insufficient insulin secretion or impaired insulin action. As the global incidence of diabetes increases year by year, prevention and treatment become increasingly important. The prevention and treatment of diabetes not only requires the support of medical expertise, but also requires the systematization and visualization of this knowledge so that doctors, patients and the general public can understand and apply it. As an emerging technology, knowledge graphs can effectively organize and manage a large amount of knowledge information, which is of great significance for the integration and dissemination of diabetes prevention and treatment knowledge.

[0003] Knowledge Graph (KG) is a knowledge base launched by Google in 2012 to enhance its intelligent search engine service. It describes entities and their relationships in the real world in a structured way. With its excellent semantic processing ability and open interconnection ability, the knowledge graph can standardize the expression of high-quality knowledge data and reveal the connection between knowledge. It uses nodes and edges to represent entities, attributes and the relationship between entities, forming a network knowledge structure. Through the knowledge graph, knowledge reasoning can be realized and query efficiency can be improved. The knowledge graph includes entities, classes, relationships and attributes, and is divided into general knowledge graphs and domain knowledge graphs. The general knowledge graph focuses on the breadth of knowledge, while the domain knowledge graph explores the entities and their relationships in a specific field in depth.

[0004] In the medical field, knowledge graphs can systematically integrate medical knowledge, realize rapid retrieval and reasoning of knowledge, and provide strong support for medical research and clinical applications. Liu Yan and others constructed a medical-related knowledge graph, and Nie Lili used NLP technology to construct a medical consultation knowledge graph based on disease triples. However, there are still some challenges in the development of knowledge graphs in the field of diabetes, such as: the existing diabetes knowledge graphs have too little correlation and knowledge missing, and cannot fully reflect the relationship between entities. Therefore, how to design an effective method to predict the relationship between entities has become an extremely important issue.

[0005] Link prediction technology of knowledge graphs is crucial for identifying and supplementing missing information in graphs, and has become a core issue in the current field of knowledge graph research. As a special type of knowledge graph, temporal knowledge graphs are widely used in various fields, especially in the field of clinical medicine, such as the course management of chronic diseases such as diabetes, where their importance is particularly prominent. As a typical chronic disease, the course of diabetes is a complex process that gradually develops and changes over time. However, traditional knowledge graph link prediction methods are often based on static data and are difficult to effectively deal with temporal knowledge graphs with dynamic characteristics. Therefore, link prediction on temporal knowledge graphs in the field of clinical medicine, such as predicting complications that may be caused by diabetes, requires special consideration of the dynamic nature of the course of the disease over time. This challenge requires researchers to not only have a deep understanding of the characteristics of temporal knowledge graphs, but also to develop new algorithms and models to more accurately capture and predict the dynamic patterns of disease development. This is not only of great significance to improving the application of knowledge graphs in the medical field, but also brings new development opportunities and challenges to traditional knowledge graph link prediction methods.

[0006] The network commonly used to process sequence problems is the RNN (Recurrent Neural Network) recurrent neural network, and one of the popular architectures is the LSTM (Long Short-Term Memory) long short-term memory network. However, networks based on the RNN structure generally have limitations. They can only read once. When processing a character, they must wait until the previous character is processed before processing the next character. This means that RNN cannot perform large-scale parallel processing like CNN (Convolution Network), which will result in too long a time consumption.

[0007] Entities and relationships in temporal knowledge graphs evolve over time. This dynamic nature requires link prediction methods to capture and model time dependencies and change patterns. Therefore, in order to more effectively perform link prediction on temporal knowledge graphs, it is necessary to explore and develop new algorithms and models to adapt to and utilize these dynamic change characteristics. The existing diabetes knowledge graphs have too few associations and missing knowledge, and cannot fully reflect the relationship between entities. In the research on link prediction of knowledge graphs, most of the existing work does focus on general knowledge graphs with timeliness, but lacks research on temporal features that are particularly prominent in clinical knowledge graphs. The RNN-based network structure has limitations, such as taking too long and not being able to process data in parallel. Summary of the invention

[0008] The technical problem to be solved by the present invention is to provide a diabetes risk prediction method, terminal device and storage medium based on TCN incremental learning to improve the prediction accuracy in view of the shortcomings of the existing technology.

[0009] In order to solve the above technical problems, the technical solution adopted by the present invention is: a diabetes risk prediction method based on TCN incremental learning, comprising the following steps:

[0010] S1. Construct a diabetes time series knowledge graph, wherein the diabetes time series knowledge graph is presented in the form of a triple <h, r, t>; wherein h is the head entity, r is the relationship, and t is the tail entity;

[0011] S2. Embed the triples <h, r, t> in the diabetes time series knowledge graph into a low-dimensional space to obtain vectorized triples; all vectorized triples form a matrix <l h ,l r .l t > L ; L is the number of triplets;

[0012] S3. Use the matrix as the input of TCN to obtain the feature vector They are l h ,l r .l t Output after TCN;

[0013] S4. Combine all feature vectors, use the combined feature vector as the input of the classifier to obtain a classification result vector, use the classification result vector as the input of the activation function, and obtain the probability distribution of the shared relationship r between the user entity and other entities.

[0014] The present invention combines the TransR conversion model and the TCN temporal convolutional network to effectively utilize the temporal information and contextual association information in the graph, thereby improving the accuracy of link prediction. The TransR model achieves effective vectorization of entities and relationships in the graph by projecting entities in a relational space containing different semantics. This method can capture the diversity of entities under different relationships and provides a solid foundation for subsequent temporal analysis. TCN is used to process entity and relationship sequences in temporal knowledge graphs to extract more accurate feature vectors.

[0015] In step S2, the calculation formula for the head entity and the tail entity in the vectorized triple is: r =M r ·h, t r =M r ·t; where h r and t r are the representations of the head entity and the tail entity projected into the subspace of the relation through the projection matrix, M r is the projection matrix corresponding to the relation r.

[0016] <l h ,l r .l t > L ={(h r1 ,r1,t r1 ),(h r2 ,r2,t r2 ),...,(h rL ,r L ,t rL )}; where h rL ,r L ,t rL They are the head entity, relation, and tail entity of the Lth vectorized triple respectively;

[0017] In step S4, the specific implementation process of combining all feature vectors includes:

[0018] Set the output V of TCN at the initial moment (0) The output V of TCN at the first time point (1) Equal, that is, V (0) =V (1) ;

[0019] The output V of TCN at the i-th time point is calculated using the following formula: (i) :V (i) '=V (i-1) +V (i) ; V (i) 'Assign the output V of TCN at the i-th time point (i) , get the updated output of TCN at the i-th time point;

[0020] And so on, until we get the output V of TCN at the nth time point: V = V (n) .

[0021] The present invention can more accurately capture the dynamic changes in the disease development process through incremental calculation of time series information, thereby providing more powerful support for link prediction.

[0022] In step S4, the classifier is a TCN temporal convolutional neural network trained with a triplet sequence.

[0023] As an inventive concept, the present invention also provides a computer-readable storage medium having a computer program / instruction stored thereon; the computer program / instruction implements the steps of the above method when executed by a processor.

[0024] As an inventive concept, the present invention also provides a computer program product, including a computer program / instruction; when the computer program / instruction is executed by a processor, the steps of the above method are implemented.

[0025] Compared with the prior art, the present invention has the following beneficial effects: the present invention combines the TransR conversion model and the TCN temporal convolutional network, effectively utilizes the temporal information and contextual association information in the graph, and improves the accuracy of link prediction. The present invention can highlight the implicit semantics and temporal information in clinical facts, and effectively utilize serialization learning to mine the previous and next dependency information, thereby making up for the shortcomings of the traditional link prediction model in terms of the accuracy of the prediction of the time-sensitive knowledge graph. In addition, the model also has strong scalability and robustness, and can adapt to clinical field temporal knowledge graphs of different scales and complexities. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a framework diagram of a temporal knowledge graph link prediction model based on TCN in an embodiment of the present invention;

[0027] Figure 2 This is a temporal link prediction model based on TCN sequence incremental learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0029] Example 1

[0030] This embodiment provides a diabetes risk prediction method based on TCN incremental learning. Incremental learning is added on the basis of the temporal convolutional network, and more abundant and effective relationship features are captured, ultimately improving the accuracy and efficiency of entity link prediction.

[0031] In order to accurately represent the entities and complex semantic relationships in the graph, we selected the TransR transformation model suitable for large-scale data. The TransR model achieves effective vectorization of entities and relationships in the graph by projecting entities in a relational space containing different semantics. This method can capture the diversity of entities under different relationships and provide a solid foundation for subsequent time series analysis.

[0032] On the basis of vectorization, we introduced TCN (Temporal Convolutional Network). TCN is a neural network specially designed for processing sequence data. It captures the temporal dependency and contextual information in the data through convolution operations. In our model, TCN is used to process entity and relationship sequences in the temporal knowledge graph to extract more accurate feature vectors.

[0033] In order to further improve the prediction accuracy, we adopted an incremental calculation method. Through the incremental calculation of time series information, we can more accurately capture the dynamic changes in the disease development process, thereby providing stronger support for link prediction. At the same time, combined with the recursive network structure of TCN, we can conduct deep learning, continuously optimize model parameters, and improve prediction performance.

[0034] The incremental TCN model we proposed has achieved remarkable results in the link prediction task of clinical temporal knowledge graphs. The model can highlight the implicit semantic and temporal information in clinical facts, and effectively use serialization learning to mine the previous and next dependency information, thus making up for the shortcomings of traditional link prediction models in terms of the accuracy of temporal knowledge graph prediction. In addition, the model is also highly scalable and robust, and can adapt to clinical temporal knowledge graphs of different scales and complexities.

[0035] As the core carrier of medical information, EMR electronic medical records contain a large amount of valuable clinical fact data. Based on the diabetes-related EMR data provided by a well-known tertiary hospital, the embodiment of the present invention successfully constructed a temporal knowledge graph focusing on the clinical field. The graph is centered on the triple structure of (entity, relationship, entity), deeply depicts the key entities and concepts in clinical data, and incorporates detailed semantic and temporal information. In order to further explore the deep connections between entities, the embodiment of the present invention also introduces a link prediction model based on deep learning, which aims to accurately reveal these connections and effectively deal with the semantic heterogeneity challenges that may be brought about by different data sources.

[0036] The temporal knowledge graph in the clinical field can be defined as a directed label graph G t =(t0,t e ,E,R,τ), where E represents the set of vertices in the graph, used to identify various entities; R is the set of edges, representing different factual relationships; τ is a function that maps any two entity pairs (E1, E2) to the relationship between them in a specific time period [t0,t e ] is the kth relation R arranged in chronological order. For example, consider a clinical time series knowledge graph G covering 2015-2016 t=(2015-1-1,2016-1-1,E,R,τ), for patient A, his blood sugar test can form a triple sequence L(t0,t e ,τ), which contains elements such as “{(A, blood sugar test|first normal)}, {(A, blood sugar test|second abnormally high)}, {(A, blood sugar test|third abnormally high)}...” and so on.

[0037] The task of temporal link prediction is to predict the possible clinical relationship R between two given entities E1 and E2 in such a temporal knowledge graph G. Considering that diabetic patients undergo multiple indicator tests during treatment, we can map the relevant triple sequence X in the diabetes temporal knowledge graph to (i) As input, through the link prediction model, it is inferred whether there is a potential disease relationship Y between the patient entity and the eye disease entity, so as to obtain the prediction result, which is the output. The entire prediction process can be expressed as

[0038] P={X,Y}={X (1) ,X (2) ,...,X (n) ,Y}

[0039] Where X (i) Represents a sequence of triples corresponding to a specific patient entity in multiple aspects (such as examination, treatment, disease progression, etc.), which describe the various attributes of the patient in detail. When our model makes predictions, the input is these triple sequences X in the clinical field time series knowledge graph. These sequences not only contain the patient's historical data, but also reflect the continuity in time. By analyzing these sequences, the model digs out the patterns and associations hidden in them, and then predicts the relationship between different entities (such as patients and possible complications). The output of the model is the prediction result of the possible clinical relationship between these entities.

[0040] The ultimate goal of the embodiment of the present invention is to establish a diabetes time series knowledge graph based on the actual EMR data of the hospital and the experience and knowledge of clinicians, and propose a diabetes time series knowledge graph link prediction model based on TCN incremental learning. The specific process is as follows:

[0041] Step 1: Based on the diabetes electronic medical record (EMR) data of a tertiary hospital, a diabetes time series knowledge graph was constructed. This graph is presented in the form of (entity, relationship, entity), which can intuitively display medical entities such as patients, drugs, examinations, and their relationships.

[0042] An entity is a node in a graph, representing a specific, unique thing or concept. In the embodiment of the present invention, each entity can be represented by a set of attributes and connected to other entities through relationships. The following is an example of an entity:

[0043] Person: Name, date of birth.

[0044] Step 2: Based on the TransR transformation model, the triples <h, r, t> in the diabetes time series knowledge graph G are transformed into

[0045] (where h is the head entity, r is the relationship, and t is the tail entity) is embedded into a low-dimensional space. The core idea of ​​the TransR model is to use different vector spaces to represent entities and relationships. In this model, entities are first embedded into the entity space and then projected into the relationship space through a relationship-specific projection matrix. This can better represent the different effects of different relationships on entities. For each relationship, TransR defines a projection matrix M r , projecting the entity vector from the entity space to the subspace of relation r, h r and t r It is expressed as:

[0046] h r =M r ·h,t r =M r ·t

[0047] h r and t r They are the representations of the head entity and the tail entity projected into the subspace of the relation through the projection matrix.

[0048] The loss function is trained by minimizing the following distance:

[0049] d r (h,t)=||h r +rt r || 2

[0050] In a series of triples <h,r,t> L After being input into the conversion model TransR for vectorization, a series of triples are obtained as vectorized representations <l h ,l r .l t > L ={(h r1 ,r1,t r1 ),(h r2 ,r2,t r2 ),...,(h rL ,r L ,t rL )}.

[0051] In the embodiment of the present invention, the projection matrix is ​​a matrix used to map entities from entity space to relational space.

[0052] Step 3: When constructing a clinical domain knowledge graph with time series, the embodiment of the present invention introduces an improved temporal convolutional network called "incremental TCN". This method is particularly suitable for processing time series features in clinical data and can more accurately extract and represent key information in triples. Specifically, the matrix composed of batch vectors obtained in the second step is h ,l r .l t > L As input, it is sent to the incremental TCN for processing. The incremental TCN extracts features of entities and relations in the triples by carefully analyzing the context information, especially focusing on the association and changes between the previous and next vectors. It outputs a more accurate feature vector

[0053] Step 4: Incremental combination of TCN sequence feature values. When processing the upper and lower triplet sequences, if the latter vector is found to have non-zero values ​​at the same position as the previous vector, we will incrementally add these values ​​to strengthen the features of these positions. The strengthened vector will be used as the input of the next vector and continue to participate in the subsequent incremental combination process, such as Figure 2 As shown. Each grid corresponds to each bit in the vector, white represents no data, slash represents data, and double slash represents data after incremental addition. For specific steps, see Algorithm 1;

[0054] Algorithm 1: Incremental combination of sequence features

[0055] Input: TCN output vector V (m) ;

[0056] Output: feature vector V after sequence increment combination;

[0057] V: vector after feature combination layer, V (m) : The vector produced by the mth TCN in the sequence increment layer.

[0058] ①V (0) =V (1)

[0059] ②for i=0to n:

[0060] ③

[0061] V (i) ′=V (i-1) +V (i)

[0062] V(i) =V (i) '

[0063] ④end for

[0064] ⑤V=V (n)

[0065] Step 5: For each class of triples, the above feature combination extraction is performed through incremental calculation until the final feature vector s is obtained as the input of the classifier TCN to classify the results. Finally, through the Sigmoid layer calculation, the probability distribution of the relationship Y between the patient entity and other entities is obtained, that is, the probability that the patient has other symptoms.

[0066] The entire training process uses back propagation to adjust parameters. The specific model training algorithm is shown in Algorithm 2.

[0067] Algorithm 2: Model training process

[0068] Input: a sequence of triples sorted from front to back in time;

[0069] Output: 0 or 1, 1 means the triple is correct, 0 means otherwise.

[0070] Epoch: number of training times;

[0071] D-Train / D-Test: training set / test set;

[0072] M: the number of TCNs in the sequence increment layer;

[0073] Loss: loss value in the classification layer;

[0074] Loss(m): the loss value of the mth TCN in the sequence increment layer;

[0075] S: the number of batch data sets;

[0076] In the i-th training: parameters of the m-th TCN in the sequence increment layer / parameters of the TCN in the classification layer.

[0077]

[0078]

[0079] Experimental results:

[0080] The data used in this example comes from the actual EMR data of a medical school. The knowledge graph is extracted from the EMR data of the endocrinology department of the hospital in the past 10 years. There are 61,000 entities, 53 different relationship types, and a training set size of 345,549 triples.

[0081] Compare models:

[0082] NTN, tensor neural network; DNN, deep neural network model; LSTM, long short-term memory network.

[0083] Evaluation indicators:

[0084] The present invention uses accuracy (AUC), recall (R) and F1-Score (F1-Score) as indicators for model evaluation in real time.

[0085] Accuracy is the ratio of the number of samples correctly predicted by the model to the total number of samples. It is one of the most common evaluation indicators and is used to measure the overall accuracy of the model. The calculation formula is:

[0086]

[0087] TP (True Positive): True positive, the number of positive samples predicted correctly.

[0088] TN (True Negative): True negative examples, the number of correctly predicted negative samples.

[0089] FP (False Positive): False positive examples, the number of negative samples that are mistakenly predicted as positive examples.

[0090] FN (False Negative): False negative examples, the number of positive samples that are mistakenly predicted as negative examples.

[0091] Recall, also known as sensitivity or true positive rate, is used to measure the model's ability to identify positive samples. The calculation formula is:

[0092]

[0093] In addition, you can also use the F1 score. The F1 score is the harmonic mean of precision and recall, which is used to find a balance between the two.

[0094] Table 1 Comparison between the method of the embodiment of the present invention and different models

[0095]

[0096] It can be seen from the experimental results that the incremental TCN model proposed in the embodiment of the present invention achieves the highest accuracy, recall rate and F1-Score, and the classification performance is significantly improved.

[0097] Example 2

[0098] Embodiment 2 of the present invention provides a terminal device corresponding to the above-mentioned embodiment 1. The terminal device may be a processing device for a client, such as a mobile phone, a laptop computer, a tablet computer, a desktop computer, etc., to execute the method of the above-mentioned embodiment.

[0099] The terminal device of this embodiment includes a memory, a processor, and a computer program stored in the memory; the processor executes the computer program in the memory to implement the steps of the method in the above-mentioned embodiment 1.

[0100] In some implementations, the memory may be a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk memory.

[0101] In some other implementations, the processor may be a central processing unit (CPU), a digital signal processor (DSP), or other general-purpose processors of various types, which are not limited herein.

[0102] Example 3

[0103] Embodiment 3 of the present invention provides a computer-readable storage medium corresponding to the above embodiment 1, on which a computer program / instruction is stored. When the computer program / instruction is executed by a processor, the steps of the method of the above embodiment 1 are implemented.

[0104] Computer readable storage media can be tangible devices that hold and store instructions used by instruction execution devices. Computer readable storage media can be, for example, but not limited to, electronic storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any combination thereof.

[0105] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0106] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0107] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0108] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0109] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A diabetes risk prediction method based on TCN incremental learning, characterized in that: The following steps are involved: S1. Construct a diabetes time series knowledge graph, wherein the diabetes time series knowledge graph is presented in the form of a triple <h, r, t>; wherein h is the head entity, r is the relationship, and t is the tail entity; S2. Embed the triples <h, r, t> in the diabetes time series knowledge graph into a low-dimensional space to obtain vectorized triples; all vectorized triples form a matrix <l h ,l r .l t > L ; L is the number of triplets; S3. Use the matrix as the input of TCN to obtain the feature vector They are l h ,l r .l t Output after TCN; S4. Combine all feature vectors, use the combined feature vector as the input of the classifier to obtain a classification result vector, use the classification result vector as the input of the activation function, and obtain the probability distribution of the shared relationship r between the user entity and other entities.

2. The diabetes risk prediction method based on TCN incremental learning according to claim 1, characterized in that: In step S2, the calculation formula for the head entity and the tail entity in the vectorized triple is: r =M r ·h, t r =M r ·t; where h r and t r are the representations of the head entity and the tail entity projected into the subspace of the relation through the projection matrix, M r is the projection matrix corresponding to the relation r.

3. The diabetes risk prediction method based on TCN incremental learning according to claim 1, characterized in that: <l h ,l r .l t > L ={(h r1 ,r1,t r1 ),(h r2 ,r2,t r2 ),...,(h rL ,r L ,t rL )}; where h rL ,r L ,t rL They are respectively the head entity, relation, and tail entity of the Lth vectorized triple.

4. The diabetes risk prediction method based on TCN incremental learning according to claim 1, characterized in that: In step S4, the specific implementation process of combining all feature vectors includes: Set the output V of TCN at the initial moment (0) The output V of TCN at the first time point (1) Equal, that is, V (0) =V (1) ; The output V of TCN at the i-th time point is calculated using the following formula: (i) :V (i)‘ =V (i-1) +V (i) ; V (i)‘ Assign to the output V of TCN at the i-th time point (i) , get the updated output of TCN at the i-th time point; And so on, until we get the output V of TCN at the nth time point: V = V (n) .

5. The diabetes risk prediction method based on TCN incremental learning according to claim 1, characterized in that: In step S4, the classifier is a TCN temporal convolutional neural network trained with a triplet sequence.

6. A computer-readable storage medium having a computer program / instruction stored thereon; characterized in that: When the computer program / instructions are executed by a processor, the steps of the method described in any one of claims 1 to 5 are implemented.

7. A computer program product comprising a computer program / instructions; characterized in that When the computer program / instructions are executed by a processor, the steps of the method described in any one of claims 1 to 5 are implemented.

Citation Information

Patent Citations

  • Link prediction method and system for clinical temporal knowledge graph

    CN109036546A

  • Social network link prediction method adopting knowledge graph embedding and time convolution network

    CN111784081A

  • Panoramic health dynamic monitoring and prediction system based on time sequence knowledge graph

    CN111951965A

  • Diabetes prediction method and system fusing knowledge graph and deep learning

    CN115719625A

  • Diabetes question-answering method based on deep learning and knowledge graph

    CN116304085A

Cited By

  • Computing task scheduling method and system based on time sequence diagram network resource state prediction

    CN120492131A

  • Computing task scheduling method and system based on network resource status prediction of time sequence diagram

    CN120492131B