Old-age chronic disease risk prediction method and system based on time sequence knowledge graph

Abstract

The application discloses a kind of old chronic disease risk prediction method and system based on time sequence knowledge graph, belong to medical health information technology field, wherein the method includes the following steps: step S1: obtaining the health data set of elderly population;Step S2: based on ESOA-XGBoost algorithm, the feature selection and sorting of each health data in health data set are carried out, to determine the correlation between variables, generate health data feature set;Step S3: build and train the time sequence knowledge graph prediction model including STKGR-PR model and T i PNN model, the relationship between target entity, other entity, time sequence information is predicted by time sequence knowledge graph prediction model;Step S4: according to the trained time sequence knowledge graph prediction model, the risk of old chronic disease of elderly population is predicted;By the method, the old chronic disease risk prediction result with good prediction effect and high accuracy can be obtained.

Description

Technical Field

[0001] This invention relates to the field of medical and health information technology, and in particular to a method and system for predicting the risk of chronic diseases in the elderly based on time-series knowledge graphs. Background Technology

[0002] Existing methods for predicting the risk of chronic diseases in the elderly can be mainly classified into the following categories: prediction methods based on statistical models, prediction methods based on machine learning, prediction methods based on time series analysis, and prediction methods based on knowledge graphs.

[0003] These methods can predict the risk of chronic diseases in the elderly to some extent, but they have certain limitations: statistical model-based prediction methods are simple to use and easy to interpret, but their effectiveness is limited when dealing with high-dimensional and nonlinear data, and they are difficult to capture complex nonlinear relationships and interaction effects; machine learning-based prediction methods can achieve good prediction results in some cases, but they are sensitive to parameter selection, have high computational complexity, and may perform poorly when dealing with large-scale datasets; time series analysis-based prediction methods can effectively capture the time dependence of data, but they have limitations in feature selection and complex relationship modeling, and cannot fully utilize multi-dimensional health data for comprehensive risk prediction; knowledge graph-based prediction methods capture complex relationships in data by constructing and reasoning knowledge graphs, but they still face challenges in dealing with sparse data and inference paths, and require a large amount of labeled data for training.

[0004] In summary, the above-mentioned methods for predicting the risk of chronic diseases in the elderly are not very accurate. Therefore, it is necessary to propose a method for predicting the risk of chronic diseases in the elderly with better predictive performance. Summary of the Invention

[0005] In view of the above-mentioned prior art, the present invention provides a method and system for predicting the risk of chronic diseases in the elderly based on time-series knowledge graphs, which mainly solves the technical problems existing in the above-mentioned background art.

[0006] To achieve the above objectives, the technical solution of this invention is implemented as follows:

[0007] In a first aspect, the present invention provides a method for predicting the risk of chronic diseases in the elderly based on time-series knowledge graphs, the method comprising the following steps:

[0008] Step S1: Obtain the health dataset of the elderly population;

[0009] Step S2: Based on the ESOA-XGBoost algorithm, perform feature selection and sorting on each health data in the health dataset to determine the correlation between variables and generate a health data feature set;

[0010] Step S3: Construct and train a temporal knowledge graph prediction model including the STKGR-PR model and the TiPNN model, and use the temporal knowledge graph prediction model to predict the relationship between the target entity, other entities, and temporal information;

[0011] Step S4: Based on the trained time-series knowledge graph prediction model, predict the risk of chronic diseases in the elderly population.

[0012] Optionally, step S2 specifically includes:

[0013] The health dataset is updated using the ESOA algorithm. After each update, the health dataset is filtered according to the discrimination criteria to obtain a subset of health data.

[0014] The health data subset is input into the XGBoost model for training, and the health data subset is adjusted according to the training results.

[0015] The contribution of each feature of the health data subset to the output of the XGBoost model is quantified using the SHAP method to obtain the health data feature set.

[0016] Optionally, the health dataset is updated using the ESOA algorithm, specifically including:

[0017] A waiting strategy is implemented to perform preliminary position updates on the health dataset, using the following formula:

[0018] x a,i =x i +exp(-t / (0.1×t max ))×0.1×hop×g i

[0019] Where, x a,i The updated position, x i This indicates the current position, exp represents the natural exponential function, and t is the current iteration number. max The maximum number of iterations is given, hop represents the feasible region of the independent variables, and g is the maximum number of iterations. i It is a random number;

[0020] Based on the results of the waiting strategy, a random walk strategy is further implemented to update the positions of the health dataset. The specific formula is as follows:

[0021]

[0022] Where, r b,i It is a random number in the range (π / 2, π / 2);

[0023] Based on the results of the random walk strategy, an encirclement strategy is further implemented to optimize the position adjustment; the specific formula is as follows:

[0024] x c,i =(1-r i -r j )×x i +r i ×D h +r j ×D g

[0025] Among them, D h D represents the difference matrix between the egret team's current position and its optimal position. g R represents the difference matrix between the current position and the optimal position of the egret population. i and r j It is a random number in the range [0, 0.5].

[0026] The specific formula for the discrimination condition is as follows:

[0027] x s,i =[x a,i ,x b,i ,x c,i ]

[0028] c i =argmin(y s,i )

[0029]

[0030] Where, x s,i Let x represent a vector containing candidate feature values ​​generated from three different sources. a,i ,x b,i ,x c,i ;c i This indicates the selection of an index from a specific source, choosing the feature source that generates the minimum error or the best result. s,i and y i is the fitness value, and r is a random number.

[0031] Optionally, the STKGR-PR model is used to map the relationship between the target entity and other entities into a multi-dimensional space through multi-hop path reasoning and reward mechanisms, thereby vectorizing the relationship between the target entity and other entities in the temporal knowledge graph prediction model and predicting the relationship between the target entity and other entities; the TiPNN model predicts the chronic disease risk of the elderly population based on temporal information and the dependency relationship between the target entities.

[0032] Optionally, construct the STKGR-PR model, specifically including:

[0033] The state representation and action space are defined as follows:

[0034] s t =[r q ;e t h t ;tim]

[0035] Among them, s t Let e ​​be the state at time t. t For the current entity, r q To query the relationship, h t This represents historical path information, where tim is the timestamp.

[0036] The defined motion space is represented as follows:

[0037] A t ={(r,e′,tim)∣(e t ,r,e′,tim)∈TKG}∪{(self-loop:e t )};

[0038] Among them, A t For the action space, it represents the action space from the current entity e. t All possible quadruplets emitted (e t The set of (r, e', tim), where self-loop is a self-looping action that indicates staying at the current node;

[0039] Define the transformation strategy, reward function, policy network, and dynamic path completion strategy, wherein the transformation strategy is expressed as:

[0040] s t+1 =[r q ;e t+1 h t+1 ;tim]

[0041] Among them, s t+1 Let e ​​be the state at time t+1. t+1 For the entity at time t+1, r q To query the relationship, h t+1 This represents the historical path information at time t+1, where tim is the timestamp.

[0042] The reward function is expressed as follows:

[0043]

[0044] Among them, R p Let E(e) represent the reward function. o ,r q ,e TD(·) represents the expected score function of path embedding, and D(·) represents the cosine similarity function.

[0045] The policy network is represented as follows:

[0046] h t =LSTM(h t-1 ,a t-1 )

[0047] π θ (a t |s t )=σ(A t ×W1RELU(W2∣s t ))

[0048] Among them, h t The hidden state at the current moment represents the path reasoning history information up to this point. LSTM is a Long Short-Term Memory network. t-1 This is the hidden state from the previous moment, containing historical information about previous reasoning paths, a t-1 The action selected in the previous time step, σ represents the softmax operator, W1 and W2 are the parameters of the two policy networks, and π is the value of π. θ (a t |s t ) is the policy network in state s t Choose action a t The probability distribution of A t For the current state s t The corresponding action embedding matrix, where each row corresponds to a possible action, s t The current state includes the current entity, the relationships queried, and the historical path information inferred to the current step; ReLU is the activation function.

[0049] The dynamic path completion strategy is expressed as follows:

[0050]

[0051] A t =A t +A add

[0052] p((r p ,e p ,tim p )|s t )=p((r p ,tim p )|s t )p(e p |(r p ,tim p ),s t )

[0053] in, This represents the combination of the j-th relation and the timestamp, where w represents the state s in the current state. t Below, the probability of each action being selected is given. SoftMax is used to transform the input values ​​into a probability distribution, ensuring that the sum of all output values ​​is 1. Each value represents the probability of selecting the corresponding action in the current state. MLP(s) t ) is for the current state s t Processing, s t It refers to the current state, including the current entity's location and historical information; A t A represents the original action space. add Represents a high-quality set of potential actions; p((r p ,tim p )∣s t ) indicates that in state s t Lower selection relation r p and timestamp tim p The probability, p(e) p ∣(r p ,tim p ),s t ) indicates that given relation r p and timestamp tim p Select tail entity e p The probability of.

[0054] Optionally, a TiPNN model is constructed, specifically including:

[0055] Construct a historical timeline, define time paths, and initialize the time path representation. The constructed historical timeline is represented as follows:

[0056]

[0057] in, Represents a historical timeline, containing subgraph information for all timestamps from time t-m+1 to t; ← indicates a build or update operation used to generate the historical timeline, (s,r τ (,o) represents an edge in the historical timeline graph, where s is the main entity and r is the edge. τ G is a relation type with a time attribute τ, where o is the object entity; (s,r,o) represents a fact in a subgraph at timestamp τ, including the subject entity s, the relation r, and the object entity o; τ The subgraph represents the time stamp τ, which contains all the facts at that specific point in time; t represents the query timestamp; m represents the length of the historical information used for reasoning, i.e., how many historical timestamps are considered.

[0058] The defined time path is represented as follows:

[0059]

[0060] Where ⊕ represents the path aggregation operator, This represents the set of paths from s to o in the historical temporal graph. The cardinality of the path set;

[0061] The initialization time path is represented as follows:

[0062]

[0063] A time path aggregation layer is set up, which includes a time edge merging function and a path aggregation function, wherein the time edge merging function is expressed as follows:

[0064]

[0065] w r (z,p τ ,o)=g(Ψ r (p)||Y(Δτ))

[0066] Δτ=|τ-t q |

[0067]

[0068] Where TMSG(·) is the time edge merging function, w represents the temporal path representation of entity z in the (l-1)th iteration. r (z,p τ (,o) represents the time edge (z,p) τ The query-aware time representation of ,o), where r is the relation type, τ is the time attribute, z is the starting entity, and o is the ending entity; Ψ r (p) represents the query relation-aware basic static representation of edge type p, and Y(Δτ) represents the temporal edge type p. r The temporal embedding, g(·) is a feedforward neural network, t q This represents the query timestamp, Δτ represents the time interval between query timestamps, and w * and φ * It is a learnable parameter and the dimension of the vector representation; d: represents the dimension of the vector representation, that is, the dimension of the output temporal encoded vector;

[0069] The path aggregation function is expressed as follows:

[0070]

[0071] Where AGG(·) is the path aggregation function, w r Query-aware time representation indicating time edge type. It is the time relationship p in the historical timeline. r The time edge that occurs between entity z and candidate entity o at historical time t;

[0072] A scoring function and a loss function are set. The scoring function is used to calculate the conditional probability of future object candidates, and the loss function guides the iterative optimization of the time-series knowledge graph prediction model. The scoring function is expressed as follows:

[0073]

[0074] in, It is a feedforward neural network, where σ(·) is the sigmoid function;

[0075] By minimizing the negative log-likelihood loss of positive and negative samples, a regularization term is introduced to obtain the final loss function;

[0076] The loss function is expressed as:

[0077]

[0078] Among them, L TKG For negative log-likelihood loss, L REG Let be the regularization parameter, L be the final loss function, and p(s,r,o) represent the probability that the target object o exists given subject s and relation r. For the constant term used in normalization, negative sample triples The probability, For negative log-likelihood; R T R: Represents matrix R and its transpose R T The product of α and I, where α is a scalar hyperparameter and I is the identity matrix.

[0079] Optionally, the STKGR-PR model is used to map the relationship between the target entity and other entities into a multi-dimensional space through multi-hop path reasoning and reward mechanisms, thereby vectorizing the relationship between the target entity and other entities in the temporal knowledge graph and predicting the relationship between the target entity and other entities. Specifically, this includes:

[0080] Actions are selected based on the health data feature set, and state transitions are performed based on the actions;

[0081] Calculate the reward based on the transformed state;

[0082] The policy network is optimized based on the reward to generate the probability of selecting the action;

[0083] A dynamic path completion strategy is implemented based on the probability of generating the selected action, selecting high-probability relation-timestamp combinations and tail entities to generate a high-quality action set.

[0084] Optionally, the TiPNN model predicts the risk of chronic diseases in the elderly population based on temporal information and the dependencies between target entities, specifically including:

[0085] Based on time series information, a historical time map is constructed, and time paths are defined and initialized to obtain time path characteristics;

[0086] Based on the time path features, time edge information is calculated using a time edge merging function in the time path aggregation layer, and the time edge information is integrated using a path aggregation function to obtain new time path features.

[0087] The new time path features are used to calculate the conditional probability of future object candidates through a scoring function, thereby predicting the risk of chronic diseases in the elderly population. In this process, the parameters in the TiPNN model are continuously adjusted iteratively through a loss function.

[0088] Optionally, step S3 further includes: evaluating the time-series knowledge graph prediction model to determine the reliability of the time-series knowledge graph prediction model, wherein the evaluation includes error analysis and accuracy analysis.

[0089] Secondly, the present invention also provides a risk prediction system for chronic diseases in the elderly based on time-series knowledge graphs. The system is used to implement the risk prediction method for chronic diseases in the elderly based on time-series knowledge graphs. The risk prediction system for chronic diseases in the elderly based on time-series knowledge graphs includes a data acquisition and processing module, a feature selection module, and a data prediction module.

[0090] The data acquisition and processing module is used to acquire health data of elderly patients and process the health data to obtain a health dataset. The health data comes from the patient's medical records, examination reports, and lifestyle habits. The processing of the health data includes removing outliers, filling in missing values, and standardizing the data.

[0091] The feature selection module, based on the ESOA-XGBoost algorithm, performs feature selection and sorting on the health dataset output by the data acquisition and processing module to generate a health data feature set.

[0092] The data prediction module is used to construct and train a time-series knowledge graph prediction model that includes the STKGR-PR model and the TiPNN model. The health data feature set generated by the feature selection module is input into the STKGR-PR model for training. Using multi-hop path reasoning and reward mechanisms, the relationship between the target entity and other entities is mapped into a multi-dimensional space to learn the complex relationship between the target entity and other entities. The output of the STKGR-PR model is input into the TiPNN model for deep learning and sequential learning. By predicting the relationship between time-series information and target entities, the chronic disease risk of the elderly population can be predicted.

[0093] The beneficial effects of this invention are as follows: First, a health dataset is acquired. Based on the ESOA algorithm combined with the XGBoost model, feature selection and sorting of health data from the elderly population are performed to generate a health data feature set, optimizing the feature selection process and improving the accuracy and robustness of the model. Based on the STKGR-PR model, multi-hop path reasoning and reward mechanisms are used to quantify the complex relationships between health data indicators. Through multi-hop path reasoning and reward mechanisms, the problems of path sparsity and reward sparsity in knowledge graphs can be solved, improving reasoning ability and prediction accuracy. Based on the TiPNN model combined with the temporal information in the trained temporal knowledge graph prediction model, deep learning and sequential learning are performed to capture dynamic health data features that change over time, further improving the model's learning ability and prediction accuracy, and reducing dependence on large-scale labeled data. Finally, a trained temporal knowledge graph prediction model is obtained. Based on the trained temporal knowledge graph prediction model, the risk of chronic diseases in the elderly population is predicted, thereby obtaining a prediction result of chronic disease risk in the elderly with good prediction effect and high accuracy. Attached Figure Description

[0094] Figure 1 This is a schematic diagram of the method for predicting the risk of chronic diseases in the elderly based on time-series knowledge graphs in an embodiment of the present invention;

[0095] Figure 2 This is a schematic diagram of an elderly chronic disease risk prediction system based on time-series knowledge graphs in an embodiment of the present invention. Detailed Implementation

[0096] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. In the following description, the expression "some embodiments" refers to a subset of all possible embodiments; however, it should be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments and can be combined with each other without conflict.

[0097] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.

[0098] It should be understood that the present invention can be embodied in various forms and should not be construed as being limited to the embodiments set forth herein. Rather, providing these embodiments will make the disclosure thorough and complete, and will fully convey the scope of the invention to those skilled in the art. Furthermore, the terminology used herein is intended only to describe particular embodiments and is not intended to limit the invention. When used herein, the singular forms “a,” “an,” and “the” are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “compose” and / or “comprising,” when used in this specification, identify the presence of the stated features, integers, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups. When used herein, the term “and / or” includes any and all combinations of the associated listed items.

[0099] It should also be noted that when an element is referred to as being "fixed to" another element, it can be directly attached to the other element or there may be an intervening element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "inner," "outer," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementation.

[0100] To fully understand this invention, a detailed structure will be presented in the following description to illustrate the technical solution proposed by this invention. Optional embodiments of the invention are described in detail below; however, in addition to these detailed descriptions, the invention may have other embodiments.

[0101] ESOA Algorithm: Egret Flock Optimization Algorithm;

[0102] XGBoost algorithm: Distributed gradient boosting algorithm;

[0103] STKGR-PR model: A sparse temporal knowledge graph multi-hop path reasoning model based on path completion and reward shaping;

[0104] TiPNN model: Based on a time-inductive path neural network;

[0105] Example 1

[0106] Please refer to the attached document. Figure 1 This application provides a method for predicting the risk of chronic diseases in the elderly based on time-series knowledge graphs, the method comprising the following steps:

[0107] Step S1: Obtain the health dataset of the elderly population;

[0108] First, health data of the elderly population is acquired, including physical examination data, medical history data, and lifestyle data. Then, the health data is cleaned, including removing outliers, missing values, or duplicate values, to ensure the accuracy and reliability of the data, resulting in a health dataset of the elderly population.

[0109] Step S2: Based on the ESOA-XGBoost algorithm, perform feature selection and sorting on each health data in the health dataset to determine the correlation between variables and generate a health data feature set;

[0110] Step S3: Construct and train a temporal knowledge graph prediction model including the STKGR-PR model and the TiPNN model, and use the temporal knowledge graph prediction model to predict the relationship between the target entity, other entities, and temporal information;

[0111] Step S4: Based on the trained time-series knowledge graph prediction model, predict the risk of chronic diseases in the elderly population.

[0112] By first processing health data to obtain a health dataset, and then inputting the health dataset into ESOA-XGBoost for feature filtering to obtain a health data feature set, the health data features are optimized. This health data feature set is then input into the STKGR-PR model for training. By capturing the relationships and temporal relationships between the target entity and other entities, the model captures the changes of each health data feature over time, thereby improving the accuracy of model predictions. The output of the STKGR-PR model is then input into the TiPNN model for optimization. The TiPNN model further captures dynamic health data features that change over time, enabling the model to learn complex temporal dependencies. Finally, a time-series knowledge graph prediction model with high learning ability and prediction accuracy is obtained, which can accurately predict the risk of chronic diseases in the elderly population.

[0113] As a further explanation of the above embodiments, step S2 specifically includes:

[0114] The settings include the egret population size n and the maximum number of iterations t. max The basic parameters of the ESOA algorithm, including the discrimination criteria;

[0115] The health dataset is updated using the ESOA algorithm. After each update, the health dataset is filtered according to the discrimination criteria to obtain a subset of health data.

[0116] Specifically, a waiting strategy is implemented to perform preliminary position updates on the health dataset, using the following formula:

[0117] x a,i =x i +exp(-t / (0.1×t max ))×0.1×hop×g i

[0118] Where, x a,i The updated position, x i This indicates the current position, exp represents the natural exponential function, and t is the current iteration number. max The maximum number of iterations is given, hop represents the feasible region of the independent variables, and g is the maximum number of iterations. i It is a random number;

[0119] Based on the results of the waiting strategy, a random walk strategy is further implemented to update the positions of the health dataset. The specific formula is as follows:

[0120]

[0121] Where, r b,i It is a random number in the range (π / 2, π / 2);

[0122] Based on the results of the random walk strategy, an encirclement strategy is further implemented to optimize the position adjustment; the specific formula is as follows:

[0123] x c,i =(1-r i -r j )×x i +r i ×D h +r j ×D g

[0124] Among them, D h D represents the difference matrix between the egret team's current position and its optimal position. g R represents the difference matrix between the current position and the optimal position of the egret population. i and r j It is a random number in the range [0, 0.5].

[0125] It should be noted that the waiting strategy involves making an initial position update based on the current state in the early stages of the prediction process, using random numbers to simulate random fluctuations of variables and adjust the model's state; the random walk strategy involves performing further random walks after the initial position update, allowing the model to explore different possible paths or states and get closer to the optimal solution; and the encirclement strategy further optimizes and adjusts the position, simulating a more precise search near the target area.

[0126] The specific formula for the discrimination condition is as follows:

[0127] x s,i =[x a,i ,x b,i ,x c,i ]

[0128] c i =argmin(y s,i )

[0129]

[0130] Where, x s,i Let x represent a vector containing candidate feature values ​​generated from three different sources (different models or methods). a,i ,x b,i ,x c,i ;c i This indicates the selection of an index from a specific source, choosing the feature source that generates the minimum error or the best result. s, i and y i This is the fitness value, and r is a random number;

[0131] The health data subset is input into the XGBoost model for training, and the health data subset is adjusted according to the training results.

[0132] Specifically, an objective function is defined, and training is performed based on XGBoost; the specific formula is as follows:

[0133]

[0134] Where l is the loss function and Ω is the regularization term. φ is the predicted value; φ is the predicted score function; and F is the function space of the decision tree. Let φ(|xi|) represent the prediction result, K represent the total number of trees, φ(|xi|) represent the prediction score of sample xi, F represent the function space associated with the decision tree, fk represent the Kth decision tree, and Ω(fj) represent the regularization term of the jth tree.

[0135] The contribution of each feature of the health data subset to the output of the XGBoost model is quantified using the SHAP method to obtain the health data feature set;

[0136] Specifically, the SHAP method is used to quantify the contribution of each feature to the model output, and the specific formula is as follows:

[0137]

[0138] Where, φ i This is the SHAP value, where z′ is the feature subset, M is the total number of features, |z′| represents the number of non-zero entries in z′, and f x(z') This represents the output of model f when making predictions using only features from a subset z′; f x (z′\i) represents the output of model f after removing feature i from subset z′, which measures the impact of feature i on the model output when it is added to the feature subset; |z′|!(M-|z′|-1)! / M! is a combination coefficient that represents how to weight subsets of different sizes when averaging the contribution of feature i across all subsets;

[0139] It should be noted that by enumerating subsets of all features, the impact of each feature on the model output under different combinations is quantified, and then the average side contribution of the features is calculated to determine which health indicator is more important. Based on this, the features that play a key role in the model are identified, and then important features are selected to be retained and unimportant features are removed, thereby optimizing the computational efficiency and prediction accuracy of the model and determining the final health data feature set.

[0140] By inputting the health dataset into ESOA and updating its location, the parameters can be continuously adjusted after each location update to filter features and obtain a subset of health data. This subset of health data is then input into the XGBoost model for training, continuously optimizing the input to the XGBoost model, thereby improving the predictive performance of the XGBoost model and obtaining a more relevant set of health data features.

[0141] As a further explanation of the above embodiments, the STKGR-PR model is used to map the relationship between the target entity and other entities into a multi-dimensional space through multi-hop path reasoning and reward mechanism, thereby realizing the vectorization of the relationship between the target entity and other entities in the time-series knowledge graph prediction model, and thus predicting the relationship between the target entity and other entities.

[0142] The construction of the STKGR-PR model specifically includes:

[0143] The state representation and action space are defined as follows:

[0144] s t =[r q ;e t h t ;tim]

[0145] Among them, s t Let e ​​be the state at time t. t For the current entity, r q To query the relationship, h t This represents historical path information, where tim is the timestamp.

[0146] The defined motion space is represented as follows:

[0147] A t ={(r,e′,tim)∣(e t ,r,e′,tim)∈TKG}∪{(self-loop:e t )};

[0148] Among them, A t For the action space, it represents the action space from the current entity e. t All possible quadruplets emitted (e t The set of (r, e', tim), where self-loop is a self-looping action that indicates staying at the current node;

[0149] Define the transformation strategy, reward function, policy network, and dynamic path completion strategy, wherein the transformation strategy is expressed as:

[0150] s t+1 =[r q ;et+1 h t+1 ;tim]

[0151] Among them, s t+1 Let e ​​be the state at time t+1. t+1 For the entity at time t+1, r q To query the relationship, h t+1 This represents the historical path information at time t+1, where tim is the timestamp.

[0152] The reward function is expressed as follows:

[0153]

[0154] Among them, R p Let E(e) represent the reward function. o ,r q ,e T D(·) represents the expected score function of path embedding, and D(·) represents the cosine similarity function.

[0155] The policy network is represented as follows:

[0156] h t =LSTM(h t-1 ,a t-1 )

[0157] π θ (a t |s t )=σ(A t ×W1RELU(W2∣s t ))

[0158] Among them, h t The hidden state at the current moment represents the path reasoning history information up to this point. LSTM is a Long Short-Term Memory network. t-1 This is the hidden state from the previous moment, containing historical information about previous reasoning paths, a t-1 The action selected in the previous time step, σ represents the softmax operator, W1 and W2 are the parameters of the two policy networks, and π is the value of π. θ (a t |s t ) is the policy network in state s t Choose action a t The probability distribution of A t For the current state s t The corresponding action embedding matrix, where each row corresponds to a possible action (e.g., the relationship between the current entity and other entities that can be reached), s t The current state includes the current entity, the relationships queried, and the historical path information inferred to the current step; ReLU is the activation function.

[0159] The dynamic path completion strategy is expressed as follows:

[0160]

[0161] A t =A t +A add

[0162] p((r p ,e p ,tim p )|s t )=p((r p ,tim p )|s t )p(e p |(r p ,tim p ),s t )

[0163] in, This represents the combination of the j-th relation and the timestamp, where w represents the state s in the current state. t Below, the probability of each action (relationship) being selected is given. SoftMax is used to transform the input values ​​into a probability distribution, ensuring that the sum of all output values ​​is 1. Each value represents the probability of selecting the corresponding action in the current state. MLP(s) t ) is for the current state s t Processing, s t It refers to the current state, including the current entity's location and historical information; A t A represents the original action space. add Represents a high-quality set of potential actions; p((r p ,tim p )∣s t ) indicates that in state s t Lower selection relation r p and timestamp tim p The probability, p(e) p ∣(r p ,tim p ),s t ) indicates that given relation r p and timestamp tim p Select tail entity e p The probability of;

[0164] The STKGR-PR model is used to map the relationship between a target entity and other entities into a multi-dimensional space through multi-hop path reasoning and a reward mechanism. This vectorizes the relationship between the target entity and other entities in the temporal knowledge graph, thereby predicting the relationship between the target entity and other entities. Specifically, it includes:

[0165] Actions are selected based on the health data feature set, and state transitions are performed based on the actions;

[0166] Calculate the reward based on the transformed state;

[0167] The policy network is optimized based on the reward to generate the probability of selecting the action;

[0168] A dynamic path completion strategy is implemented based on the probability of generating the selected action, selecting high-probability relation-timestamp combinations and tail entities to generate a high-quality action set;

[0169] It should be noted that an action refers to a quadruple of relation, entity, path information, and timestamp in a temporal knowledge graph. These are used for reasoning and transitions in the temporal knowledge graph prediction model. The temporal knowledge graph prediction model needs to select an appropriate action (i.e., associated health data features and time information) from the current state (i.e., a specific health data point in the health data feature set) for the next step of reasoning. The selection of these actions is based on the health data feature set. The temporal knowledge graph prediction model uses the health data feature set to determine which features are most important and selects the next action based on their time series information. The content of the health data feature set directly affects the decision-making of the temporal knowledge graph prediction model in multi-hop path reasoning, and finally obtains the relevant health data features.

[0170] A high-quality action set refers to the optimal action (the most relevant health data feature in the health data feature set) selected through multi-hop path reasoning and reward mechanism. By using time series information, the model can capture the most relevant health data features within a specific time period, thereby associating these health data features with chronic disease risk. These action sets not only include disease-related features, but also the time dependence of complex relationships with health indicators. In this way, the model can accurately predict the risk of chronic diseases in the elderly.

[0171] The STKGR-PR model is responsible for capturing the complex dependencies between target entities and other entities in a time-series knowledge graph prediction model, specifically, capturing the complex dependencies between different health data features. Since the STKGR-PR model needs to select an appropriate action (i.e., associated health features and time information) from the current state (i.e., a specific health data point) for the next inference step, and these action selections are based on the health data feature set, the STKGR-PR model determines which features are most important and selects the next action based on their time-series information. Therefore, during the continuous training of the STKGR-PR model, the time factor is considered, enabling the STKGR-PR model to track and understand the changes of various health indicators over time. This allows the model to evaluate the quality of the inference path through a reward mechanism, correlate the time changes between different health indicators, and improve the model's predictive ability.

[0172] As a further illustration of the above embodiments, the TiPNN model predicts the risk of chronic diseases in the elderly population based on temporal information and the dependency relationship between target entities.

[0173] The construction of the TiPNN model specifically includes:

[0174] Construct a historical timeline, define time paths, and initialize the time path representation. The constructed historical timeline is represented as follows:

[0175]

[0176] in, Represents a historical timeline, containing subgraph information for all timestamps from time t-m+1 to t; ← indicates a build or update operation used to generate the historical timeline, (s,r τ (,o) represents an edge (or time edge) in a historical timeline graph, where s is the main entity and r is the timeline. τ G is a relation type with a time attribute τ, where o is the object entity; (s,r,o) represents a fact (or triple) in the subgraph at timestamp τ, consisting of the subject entity s, the relation r, and the object entity o; τ The subgraph represents the time stamp τ, which contains all the facts at that specific point in time; t: represents the query timestamp; m: represents the length of the historical information used for reasoning, i.e., how many historical timestamps are considered.

[0177] The defined time path is represented as follows:

[0178]

[0179] Where ⊕ represents the path aggregation operator, This represents the set of paths from s to o in the historical temporal graph. The cardinality of the path set;

[0180] The initialization time path is represented as follows:

[0181]

[0182] A time path aggregation layer is set up, which includes a time edge merging function and a path aggregation function, wherein the time edge merging function is expressed as follows:

[0183]

[0184] w r (z,p τ ,o)=g(Ψ r (p)||Y(Δτ))

[0185] Δτ=|τ-t q |

[0186]

[0187] Where TMSG(·) is the time edge merging function, w represents the temporal path representation of entity z in the (l-1)th iteration. r (z,p τ (,o) represents the time edge (z,p) τ The query-aware time representation of ,o), where r is the relation type, τ is the time attribute, z is the starting entity, and o is the ending entity; Ψ r (p) represents the query relation-aware basic static representation of edge type p, and Y(Δτ) represents the temporal edge type p. r The temporal embedding, g(·) is a feedforward neural network, t q This represents the query timestamp, Δτ represents the time interval between query timestamps, and w * and φ * It is a learnable parameter and the dimension of the vector representation; d: represents the dimension of the vector representation, that is, the dimension of the output temporal encoded vector;

[0188] The path aggregation function is represented as:

[0189]

[0190] Where AGG(·) is the path aggregation function, w r Query-aware time representation indicating time edge type. It is the time relationship p in the historical timeline. r The time edge that occurs between entity z and candidate entity o at historical time t;

[0191] A scoring function and a loss function are set. The scoring function is used to calculate the conditional probability of future object candidates, and the loss function guides the iterative optimization of the time-series knowledge graph prediction model. The scoring function is expressed as follows:

[0192]

[0193] in, It is a feedforward neural network, where σ(·) is the sigmoid function;

[0194] By minimizing the negative log-likelihood loss of positive and negative samples, a regularization term is introduced to obtain the final loss function;

[0195] The loss function is expressed as:

[0196]

[0197] Among them, L TKG For negative log-likelihood loss, L REG Let be the regularization parameter, L be the final loss function, and p(s,r,o) represent the probability that the target object o exists given subject s and relation r. For the constant term used in normalization, negative sample triples The probability, For negative log-likelihood; R T R: Represents matrix R and its transpose R T The product of α and I, where α is a scalar hyperparameter and I is the identity matrix.

[0198] The TiPNN model predicts the risk of chronic diseases in the elderly population based on temporal information and the dependencies between target entities, specifically including:

[0199] Based on time series information, a historical time map is constructed, and time paths are defined and initialized to obtain time path characteristics;

[0200] Based on the time path features, time edge information is calculated using a time edge merging function in the time path aggregation layer, and the time edge information is integrated using a path aggregation function to obtain new time path features.

[0201] The new time path features are used to calculate the conditional probability of future object candidates through a scoring function, thereby predicting the risk of chronic diseases in the elderly population. In this process, the parameters in the TiPNN model are continuously adjusted iteratively through a loss function.

[0202] It should be noted that the time path aggregation layer in the TiPNN model is responsible for learning the characteristics of each health indicator changing over time through the layer message passing mechanism. The aggregation of time paths combines time information and path information, iteratively expands the path length, captures the time path characteristics between the query target entity and other entities, and further optimizes the TiPNN model through this serialized learning method, enabling the TiPNN model to learn complex time dependencies.

[0203] The temporal edge merging function in the Ti PNN model is used to process the edges in the time path, merging the temporal information and the features of the edges. This ensures that the model has high expressive power when capturing time-related dynamics and can reflect the dynamic changes of various health indicators at different time points.

[0204] As a further explanation of the above embodiments, step S3 further includes: evaluating the time-series knowledge graph prediction model to determine the reliability of the time-series knowledge graph prediction model, wherein the evaluation includes error analysis and accuracy analysis.

[0205] After acquiring health data, data cleaning is performed to remove outliers, imput missing values, and standardize the data to obtain a health dataset. This ensures the quality and consistency of the health data. Then, based on ESOA combined with the XGBoost algorithm, feature selection is performed on each health data point, eliminating redundant features and retaining key features to generate a health data feature set. ESOA effectively selects important features, reduces the impact of noisy features, and improves the accuracy and robustness of the model. In the step of building the temporal knowledge graph model, a temporal knowledge graph prediction model is constructed and trained based on the health data feature set output by ESOA and XGBoost. The STKGR-PR model, combined with multi-hop path reasoning and a reward mechanism, captures the complex dependencies between features to predict the relationship between the target entity (elderly chronic diseases) and other entities (health features). The TiPNN model is then used to predict the risk of elderly chronic diseases according to the predicted dependencies between the target entity and other entities. Finally, model evaluation is performed, and error and accuracy analyses are conducted on the prediction results to determine the reliability of the prediction model. Furthermore, this invention employs the TiPNN model for deep optimization and sequential learning, which highlights the potential semantic and multivariate information between health indicators and effectively mines their dependencies using sequential learning. The TiPNN model overcomes the limitation of low accuracy in traditional link prediction models, improving the accuracy of risk prediction by combining feature selection and temporal knowledge graphs. Therefore, the prediction method provided by this invention can effectively improve the accuracy and reliability of risk prediction for chronic diseases in the elderly, reduce the impact of noisy data, capture complex temporal dependencies, and, through deep optimization, reduce the dependence on large-scale labeled data and computational resources, thereby improving the model's learning ability and prediction accuracy.

[0206] Example 2

[0207] Please refer to the attached document. Figure 2 The present invention also provides a risk prediction system for chronic diseases in the elderly based on time-series knowledge graphs. The system is used to implement a method for predicting the risk of chronic diseases in the elderly based on time-series knowledge graphs. The risk prediction system for chronic diseases in the elderly based on time-series knowledge graphs includes a data acquisition and processing module, a feature selection module, and a data prediction module.

[0208] The data acquisition and processing module is used to acquire health data of elderly patients and process the health data to obtain a health dataset. The health data comes from the patient's medical records, examination reports, and lifestyle habits. The processing of the health data includes removing outliers, filling in missing values, and standardizing the data.

[0209] The feature selection module, based on the ESOA-XGBoost algorithm, performs feature selection and sorting on the health dataset output by the data acquisition and processing module to generate a health data feature set.

[0210] The data prediction module is used to construct and train a time-series knowledge graph prediction model that includes the STKGR-PR model and the TiPNN model. The health data feature set generated by the feature selection module is input into the STKGR-PR model for training. Using multi-hop path reasoning and reward mechanisms, the relationship between the target entity and other entities is mapped into a multi-dimensional space to learn the complex relationship between the target entity and other entities. The output of the STKGR-PR model is input into the TiPNN model for deep learning and sequential learning. By predicting the relationship between time-series information and target entities, the chronic disease risk of the elderly population can be predicted.

[0211] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting the risk of chronic diseases in the elderly based on time-series knowledge graphs, characterized in that, The method includes the following steps: Step S1: Obtain the health dataset of the elderly population; Step S2: Based on the ESOA-XGBoost algorithm, perform feature selection and sorting on each health data in the health dataset to determine the correlation between variables and generate a health data feature set; Specifically, step S2 includes: The health dataset is updated using the ESOA algorithm. After each update, the health dataset is filtered according to the discrimination criteria to obtain a subset of health data. The health data subset is input into the XGBoost model for training, and the health data subset is adjusted according to the training results. The contribution of each feature of the health data subset to the output of the XGBoost model is quantified using the SHAP method to obtain the health data feature set; Step S3: Construct and train a temporal knowledge graph prediction model including the STKGR-PR model and the TiPNN model, and use the temporal knowledge graph prediction model to predict the relationship between the target entity, other entities and temporal information; The STKGR-PR model is used to map the relationship between the target entity and other entities into a multi-dimensional space through multi-hop path reasoning and reward mechanisms, thereby vectorizing the relationship between the target entity and other entities in the temporal knowledge graph prediction model and predicting the relationship between the target entity and other entities. The TiPNN model predicts the chronic disease risk of the elderly population based on temporal information and the dependency relationship between the target entity. The output of the STKGR-PR model is input into the TiPNN model for deep learning and sequential learning, and the prediction of the chronic disease risk of the elderly population is achieved by predicting the relationship between temporal information and the target entity. Step S4: Based on the trained time-series knowledge graph prediction model, predict the risk of chronic diseases in the elderly population.

2. The method for predicting the risk of chronic diseases in the elderly based on time-series knowledge graphs according to claim 1, characterized in that, The construction of the STKGR-PR model specifically includes: The state representation and action space are defined as follows: in, For the present t The state at any given moment, For the current entity, To query the relationship, For historical path information, For timestamps; The defined motion space is represented as follows: ； in, For action space, it represents the action space from the current entity. All possible quadruplets starting from The set of self-loops, where self-loop represents a self-circulating action, indicating that the current node is being stayed at. Define the transformation strategy, reward function, policy network, and dynamic path completion strategy, wherein the transformation strategy is expressed as: in, for t+1 The current state e t+1 for t+1 The entity at any given moment To query the relationship, h t+1 for t+1 Historical path information at any given moment For timestamps; The reward function is expressed as follows: in, Represents the reward function, This represents the expected score function for path embedding. Represents the cosine similarity function; The policy network is represented as follows: in, h t The hidden state at the current moment represents the path reasoning history information up to this point. LSTM is a Long Short-Term Memory network. h t-1 This is the hidden state from the previous moment, containing historical information about the previous reasoning path. a t-1 The action selected in the previous time step, σ represents the softmax operator. W 1 and W 2 represents the parameters of the two policy networks. For the policy network in state s t Select action a t The probability distribution, A t Current state s t The corresponding action embedding matrix, where each row corresponds to a possible action. s t The state at time t includes the current entity, the queried relationship, the historical path information from the reasoning to the current step, and the timestamp; ReLU is the activation function. The dynamic path completion strategy is expressed as follows: in, This represents the combination of the j-th relation and the timestamp, where j represents the range 1 to R. w In the current state s t The probability of each action being selected. SoftMax MLP (Multi-Level Processing) is used to transform input values ​​into a probability distribution, ensuring that the sum of all output values ​​is 1. Each value represents the probability of choosing the corresponding action in the current state. s t ) is for the current state s t To process, s t It represents the current state, including the current entity's location and historical information; Represents the original action space. This represents a high-quality set of potential actions. Represents the updated set of entities; Indicates the state Lower selection relationship and timestamp The probability, Indicates that in a given relation and timestamp Select tail entity The probability, Indicates at a given time Under these conditions, the relationship ,entity and timestamp The probability of them happening simultaneously Indicates a specific point in time. Identified relationships or features Indicates at a point in time The entity with the highest probability.

3. The method for predicting the risk of chronic diseases in the elderly based on time-series knowledge graphs according to claim 2, characterized in that, The construction of the TiPNN model specifically includes: Construct a historical timeline, define time paths, and initialize the time path representation. The constructed historical timeline is represented as follows: in, This represents a historical timeline, including data from different times. t-m+1 arrive t Subgraph information for all timestamps; This indicates a build or update operation used to generate a historical timeline. Represents an edge in a historical timeline graph, where It is the main entity. It has time attributes Relationship types, It is an objective entity; Indicates the timestamp A fact in a subgraph includes subject entity s, relation r, and object entity o; Indicates timestamp as A subgraph of time, which contains all the facts at that specific point in time; t Indicates the timestamp of the query; m This indicates the length of the historical information used for reasoning, i.e., how many historical timestamps are considered; The defined time path is represented as follows: in, Represents the path aggregation operator. In a historical temporal graph, from arrive The set of paths, The cardinality of the path set; The initialization time path is represented as follows: A time path aggregation layer is set up, which includes a time edge merging function and a path aggregation function, wherein the time edge merging function is expressed as follows: Where TMSG(·) is the time edge merging function, This represents the temporal path representation of entity z in the (l-1)th level iteration. Indicates time edge The query-aware time representation, where r is the relation type, It is a time attribute, z is the starting entity, and o is the ending entity; Representing edge type The query relationship awareness is a basic static representation. Indicates time edge type Time embedding, It is a feedforward neural network. Indicates the query timestamp. Indicates the time interval between query timestamps; and These are learnable parameters. d This indicates the dimension of the vector representation, i.e., the dimension of the output time-coded vector; The path aggregation function is expressed as follows: in, This is a path aggregation function. Query-aware time representation indicating time edge type. It is a time relationship in a historical timeline. In historical time Occurring in the entity and candidate entities The time boundary between them; A scoring function and a loss function are set. The scoring function is used to calculate the conditional probability of future object candidates, and the loss function is used to guide the iterative optimization of the time-series knowledge graph prediction model. The scoring function is expressed as follows: in, It is a feedforward neural network. It is the sigmoid function; By minimizing the negative log-likelihood loss of positive and negative samples, a regularization term is introduced to obtain the final loss function; The loss function is expressed as: in, For negative log-likelihood loss, For regularization parameters, Let p(s,r,o) be the final loss function, representing the probability that the target object o exists given subject s and relation r. For the constant term used in normalization, negative sample triples The probability, Treat it as a negative log-likelihood; : Represents a matrix With its transpose The product of For a scalar hyperparameter, It is an identity matrix.

4. The method for predicting the risk of chronic diseases in the elderly based on time-series knowledge graphs according to claim 3, characterized in that, The STKGR-PR model is used to map the relationship between a target entity and other entities into a multi-dimensional space through multi-hop path reasoning and a reward mechanism. This vectorizes the relationship between the target entity and other entities in the temporal knowledge graph, thereby predicting the relationship between the target entity and other entities. Specifically, it includes: Actions are selected based on the health data feature set, and state transitions are performed based on the actions; Calculate the reward based on the transformed state; The policy network is optimized based on the reward to generate the probability of selecting the action; A dynamic path completion strategy is implemented based on the probability of generating the selected action, selecting high-probability relation-timestamp combinations and tail entities to generate a high-quality action set.

5. The method for predicting the risk of chronic diseases in the elderly based on time-series knowledge graphs according to claim 4, characterized in that, The TiPNN model predicts the risk of chronic diseases in the elderly population based on temporal information and the dependencies between target entities, specifically including: Based on time series information, a historical time map is constructed, and time paths are defined and initialized to obtain time path characteristics; Based on the time path features, time edge information is calculated using a time edge merging function in the time path aggregation layer, and the time edge information is integrated using a path aggregation function to obtain new time path features. The new time path features are used to calculate the conditional probability of future object candidates through a scoring function, thereby predicting the risk of chronic diseases in the elderly population. In this process, the parameters in the TiPNN model are continuously adjusted iteratively through a loss function.

6. The method for predicting the risk of chronic diseases in the elderly based on time-series knowledge graphs according to claim 1, characterized in that, Step S3 further includes: evaluating the time-series knowledge graph prediction model to determine the reliability of the time-series knowledge graph prediction model, wherein the evaluation includes error analysis and accuracy analysis.

7. A system for predicting the risk of chronic diseases in the elderly based on a time-series knowledge graph, the system being used to implement the method for predicting the risk of chronic diseases in the elderly based on a time-series knowledge graph as described in any one of claims 1-6, characterized in that, The system includes a data acquisition and processing module, a feature selection module, and a data prediction module; The data acquisition and processing module is used to acquire health data of elderly patients and process the health data to obtain a health dataset. The health data comes from the patient's medical records, examination reports, and lifestyle habits. The processing of the health data includes removing outliers, filling in missing values, and standardizing the data. The feature selection module, based on the ESOA-XGBoost algorithm, performs feature selection and sorting on the health dataset output by the data acquisition and processing module to generate a health data feature set. Specifically, based on the ESOA-XGBoost algorithm, feature selection and sorting are performed on the health dataset output by the data acquisition and processing module to generate a health data feature set, including: The health dataset is updated using the ESOA algorithm. After each update, the health dataset is filtered according to the discrimination criteria to obtain a subset of health data. The health data subset is input into the XGBoost model for training, and the health data subset is adjusted according to the training results. The contribution of each feature of the health data subset to the output of the XGBoost model is quantified using the SHAP method to obtain the health data feature set; The data prediction module is used to construct and train a time-series knowledge graph prediction model that includes the STKGR-PR model and the TiPNN model. The health data feature set generated by the feature selection module is input into the STKGR-PR model for training. Using multi-hop path reasoning and reward mechanisms, the relationship between the target entity and other entities is mapped to a multi-dimensional space, and the complex relationship between the target entity and other entities is learned. The output of the STKGR-PR model is input into the TiPNN model for deep learning and sequential learning. By predicting the relationship between time-series information and target entities, the chronic disease risk of the elderly population can be predicted.

Images