A knowledge graph-based personal health assessment reasoning method

Abstract

The application relates to the field of knowledge graphs, and in particular to a personal health assessment reasoning method based on a knowledge graph, which comprises the following steps: embedding user physiological indexes, clinical characteristics and medical knowledge graph data into vectors, constructing a probabilistic hypercone with initial health entities as a starting point for reasoning; in each round of reasoning, a relationship conversion matrix is selected according to boundary uncertainty to transform the hypercone and screen candidate entities; if the entities are associated with physiological indexes, the entity and index vectors are fused by means of an attention mechanism to generate a context-aware entity state; the comprehensive confidence is calculated based on path fragments, and the path reward is calculated in combination with the radial distribution KL divergence before and after and the ontology information gain; after iteration until the termination condition, the effective reasoning path is screened according to the cumulative reward, and a health state inference graph is constructed. The application can quantify reasoning uncertainty, fuse heterogeneous health data, objectively judge reasoning links, and improve the accuracy and result interpretability of health assessment.

Description

Technical Field

[0001] This application relates to the field of knowledge graphs, and more particularly to a reasoning method for personal health assessment based on knowledge graphs. Background Technology

[0002] Personal health assessment is a core component of intelligent health management systems. Its core objective is to quantitatively analyze an individual's overall physiological state, uncover hidden health risks, and provide quantitative evidence for chronic disease intervention and clinical auxiliary diagnosis. Medical knowledge graphs possess the advantages of structured and relational knowledge representation, capable of comprehensively depicting the causal and relational logic between diseases, clinical symptoms, physiological indicators, interventional drugs, and risk factors. Integrating multidimensional personal health data with standardized medical knowledge graphs, and relying on artificial intelligence graph reasoning algorithms to achieve graded assessment of health status and prediction of disease progression trends, has become a mainstream approach for home-based health monitoring and primary care clinical decision support.

[0003] Individual health evolution and disease progression are coupled by multiple levels of physiological mechanisms. There is a highly complex nonlinear relationship between microscopic continuous physiological test indicators and macroscopic discrete clinical symptoms. Therefore, refined personal health assessment cannot be simply classified as a single data binary or multi-classification task. It must rely on medical knowledge graphs to carry out multi-step chain semantic reasoning and explore the pathological relationships hidden behind health data layer by layer.

[0004] Existing health reasoning assessment schemes based on medical knowledge graphs still suffer from several key technical deficiencies. A Chinese patent application with publication number CN113871003A discloses a disease-aided differential diagnosis system based on a causal medical knowledge graph. This application constructs a causal medical knowledge graph, imports positive and negative clinical data from patients to generate personalized sub-graphs, and uses these graphs to perform static causal reasoning, building a causal medical graph diagnostic framework capable of basic disease differential reasoning. However, this application only addresses static matching diagnosis for standardized confirmed cases, exhibiting significant shortcomings in dynamic, multi-step reasoning scenarios involving diverse and heterogeneous health data from ordinary individuals. It cannot simultaneously address the three core requirements of deep integration of heterogeneous health data, quantification of uncertainty throughout the reasoning process, and generation of a complete and interpretable reasoning chain, making it difficult to adapt to the personalized, refined, and highly interpretable modern intelligent health assessment scenarios.

[0005] Therefore, in view of the technical defects of existing technologies, such as insufficient heterogeneous data fusion, inability to quantify inference uncertainty, and lack of complete credibility evaluation and visualization of inference paths, this application proposes a personal health assessment inference method that significantly improves the individualized accuracy, uncertainty characterization ability and clinical interpretability of personal health assessment. Summary of the Invention

[0006] To address the technical problems of insufficient heterogeneous data fusion, inability to quantify inference uncertainty, and lack of complete credibility evaluation and visualization of inference paths in existing technologies, this application provides a knowledge graph-based personal health assessment inference method, including:

[0007] The system acquires physiological indicators, clinical characteristics, and medical knowledge graph data of the user to be evaluated, and embeds these data into a high-dimensional embedding space using unified vectorization. Starting with the initial healthy entity of the user, the inference state is initialized as a probabilistic hypercone containing a central axis vector and a radial boundary distribution. In each inference step, a relation transformation matrix is ​​selected based on the uncertainty of the radial boundary distribution and the reachable relation set of the current entity. The relation transformation matrix is ​​applied to the current probabilistic hypercone to generate the next region, and the graph entities within the region are extracted as candidate entities. If a candidate entity is associated with a physiological indicator, the candidate entity embedding is fused using an attention mechanism based on the inference path context. Vectors and physiological indicators are used to generate context-aware entity states and update the central axis vector of the next probabilistic hypercone. Path segments are composed of historical paths and candidate entities. Semantic consistency scores are obtained through a knowledge graph embedding model, and the comprehensive confidence of the path segments is calculated by combining prior probabilities. The KL divergence of the radial boundary distribution between the current and previous steps is calculated. The reward value of the current path segment is calculated by combining the information content gain of the candidate entities in the medical ontology. The reasoning calculation is repeated until the preset termination condition is met. The cumulative reward is the sum of the reward values ​​of all path segments on each reasoning path. Valid reasoning paths are selected based on the cumulative reward, and a health status inference graph is constructed and output.

[0008] This application constructs a probabilistic hypercone based on a medical knowledge graph to achieve multi-step iterative health reasoning. On the one hand, the radial boundary distribution of the probabilistic hypercone fully characterizes the uncertainty of the entire health deduction process, effectively representing the probabilistic fuzzy boundary brought about by the complex correlation between physiological indicators and clinical manifestations. On the other hand, it uses an attention mechanism to fuse two types of heterogeneous data, discrete graph entities and continuous physiological indicators, and achieves semantic alignment of the two types of data based on the context of the reasoning path, fully exploring individual-specific health characteristics. At the same time, it quantifies the credibility of the path through semantic confidence, KL divergence and medical ontology information gain, selects effective reasoning links based on cumulative rewards and generates a health status inference graph, forming a complete and traceable medical evidence chain. It takes into account the fusion of individualized health data, the quantification of reasoning uncertainty and the interpretability of the reasoning process, significantly improving the accuracy and clinical reference value of personal health assessment results.

[0009] In one embodiment, starting with the initial health entity of the user to be evaluated, the inference state is initialized as a probabilistic hypercone containing a central axis vector and a radial boundary distribution, including: Retrieve the user's medical text records and extract the confirmed disease entity names as the initial health entities; input the initial health entities into a pre-trained graph encoder and extract the feature vectors of the entities in the high-dimensional embedding space as the central axis vectors; count the total number of first-order neighboring nodes of the initial health entities in the medical knowledge graph, multiply the total number of neighboring nodes by a preset coefficient as the initial uncertainty adjustment parameter of the radial boundary distribution, and determine the shape parameters of the radial boundary distribution based on the initial uncertainty adjustment parameter to complete the initialization of the probabilistic hypercone.

[0010] In one embodiment, the relation transformation matrix is ​​selected based on the degree of uncertainty in the radial boundary distribution and the current set of reachable relations of the entity, including: Calculate the differential entropy of the current probability hypercone radial boundary distribution as the uncertainty quantification value; The uncertainty quantification value is compared with a preset entropy threshold; If the uncertainty quantification value is greater than the entropy threshold, then a low-rank generalization matrix trained with nuclear norm penalty is selected as the relation transformation matrix from the candidate relation transformation matrix set corresponding to the current entity reachable relation set; if the uncertainty quantification value is less than or equal to the entropy threshold, then a high-rank specialization matrix trained with L2 regularization is selected as the relation transformation matrix from the candidate relation transformation matrix set corresponding to the current entity reachable relation set.

[0011] The overall uncertainty of the radial boundary distribution is quantified by differential entropy. Based on the entropy threshold, an adaptive matching mechanism between a low-rank generalization matrix and a high-rank specialized matrix is ​​used: when the inference uncertainty is high, a low-rank generalization matrix trained with a nuclear norm penalty is selected to broaden the semantic retrieval scope and avoid missing potential health risk associations; when the inference information is sufficient and the uncertainty is low, a high-rank specialized matrix trained with L2 regularization is selected to tighten the inference boundary and filter out irrelevant entity interference. Compared to the inference mode with a fixed single transformation matrix, this adaptive selection mechanism can adapt to the degree of ambiguity at different stages of disease development, balancing comprehensive risk screening with precise targeted inference, reducing invalid inference branches, and improving the computational efficiency and path selection accuracy of multi-step health inference.

[0012] In one embodiment, based on the inference path context, candidate entity embedding vectors and physiological indicators are fused through an attention mechanism to generate a context-aware entity state, including: The historical path sequence is input into a long short-term memory network to extract the hidden layer states, which serve as the inference path context. The candidate entity embedding vector and the physiological index quantization vector are mapped to the same dimension as the inference path context through fully connected layers. The inner product of the mapped candidate entity embedding vector and the mapped physiological index quantization vector with the inference path context is calculated. The two inner product values ​​are normalized by the Softmax function to obtain the fusion weights of each modality. The fusion weights are multiplied by the mapped candidate entity embedding vector and the mapped physiological index quantization vector and summed to obtain the context-aware entity state.

[0013] By capturing the temporal context information of the entire reasoning path using a long short-term memory network, and then dynamically allocating the fusion weights of two heterogeneous modalities, graph entities and physiological indicators, through dot product attention, the representation ratio of quantitative physiological indicators and qualitative medical entities can be adaptively adjusted in conjunction with the pre-reasoning logic, solving the problem of semantic alignment and fusion between discrete graph entities and continuous physiological indicators. The generated context-aware entity states carry complete historical reasoning information, making subsequent probabilistic hypercone updates fit the individual's real physiological characteristics, greatly improving the individualized accuracy of health reasoning.

[0014] In one embodiment, semantic consistency scores are obtained through a knowledge graph embedding model, and the comprehensive confidence of path segments is calculated by combining prior probabilities, including: The path segment is input into the knowledge graph embedding model, and the semantic consistency score of the path segment is output. The medical knowledge graph is queried and retrieved, and the frequency of occurrence of the corresponding relation type of the currently selected relation transformation matrix is ​​counted. The occurrence frequency and the total relation frequency are smoothed and normalized to obtain the prior probability that the value is in the probability interval of 0 to 1. The semantic consistency score is multiplied by the prior probability to obtain the comprehensive confidence of the path segment.

[0015] In one embodiment, the relationship type includes one or more of the following: symptom association, disease comorbidity, examination indicator association, drug treatment, or risk factor association.

[0016] In one embodiment, calculating the reward value for the current path segment includes: The Monte Carlo sampling method is used to extract samples from the radial boundary distributions of the current probabilistic hypercone and the previous probabilistic hypercone, respectively, and the KL divergence of the current radial boundary distribution relative to the previous radial boundary distribution is estimated. The depth of the candidate entity in the medical ontology tree structure is obtained, and the depth value represents the information content gain. The KL divergence and the information content gain are linearly weighted and summed to obtain the reward value of the current path segment.

[0017] By comprehensively evaluating the merits of a single reasoning path from two dimensions, we can accurately distinguish reasoning branches with good convergence and high diagnostic information value, providing an objective quantitative basis for subsequent screening of effective reasoning paths and optimizing the overall quality of healthy reasoning path screening.

[0018] In one embodiment, valid reasoning paths are selected based on cumulative rewards, including: When the maximum number of inference steps is reached or a terminal entity with no successor node is reached, a preset termination condition is triggered, inference calculation is stopped, and the cumulative reward value of all path segments on each inference path is calculated to obtain the cumulative reward value. Based on the average and standard deviation of the cumulative reward values ​​of all current inference paths, the sum of the average and standard deviation is calculated as the screening threshold. Inference paths with a cumulative reward value greater than or equal to the screening threshold are retained as valid inference paths, and the remaining inference paths are removed. When there is no inference path with a cumulative reward value greater than or equal to the screening threshold, at least one inference path with the highest cumulative reward value is retained as a valid inference path.

[0019] By adaptively generating screening thresholds based on the distribution characteristics of rewards for all reasoning paths, it can automatically filter out invalid paths with low diagnostic value and high reasoning uncertainty, while also stably retaining high-quality reasoning links with medical reference value, ensuring that subsequent health status inference graphs can generate complete and credible visual reasoning evidence chains.

[0020] In one embodiment, constructing and outputting a health status inference graph includes: Directed edges are established according to the order from the source entity to the target candidate entity in the effective inference path, and the inference relation corresponding to the relation transformation matrix is ​​used as the edge type of the directed edge. For each path segment in the effective inference path, the comprehensive confidence of the path segment is obtained, and the edge weight of the corresponding directed edge is calculated in combination with the cumulative reward value of the effective inference path to which the path segment belongs. When multiple effective inference paths contain the same directed edge, the multiple edge weights corresponding to the same directed edge are aggregated, and the aggregation result is used as the final weight of the directed edge.

[0021] In one embodiment, after extracting map entities within the region as candidate entities, the method further includes: After obtaining candidate entities, for each inference branch, the radial boundary distribution of the next step of the branch is updated according to the degree of dispersion of the candidate entity set corresponding to the branch relative to the distribution of the central axis quantity of the next step of the branch.

[0022] The inference uncertainty boundary is dynamically adjusted based on the degree of aggregation of entities in each round of retrieval. The more concentrated the distribution of inference entities, the tighter the retrieval scope becomes, while the more dispersed the entities are, the wider the boundary becomes. This enables the inference scope to be adaptively and dynamically adjusted according to individual health data. While reducing the amount of redundant and irrelevant entities retrieved, it also avoids the omission of potential health relationships and continuously optimizes the accuracy and coverage of each subsequent round of inference.

[0023] The technical solution of this application has the following beneficial technical effects: First, it adopts a probabilistic hypercone modeling inference state that includes a central axis vector and radial boundary distribution. It uses differential entropy to adaptively switch between low-rank and high-rank relation transformation matrices and dynamically updates the radial boundary parameters based on the dispersion of candidate entity distribution. This fully characterizes the fuzzy boundaries and probabilistic uncertainties in the disease evolution process, solving the shortcomings of traditional methods that only output deterministic conclusions and cannot distinguish risk confidence. It takes into account both comprehensive screening of potential health risks and targeted and accurate inference. Secondly, by relying on long short-term memory networks to extract the temporal context of reasoning, and combining attention mechanisms to adaptively fuse physiological indicator quantification vectors and graph entity embedding vectors, the semantic alignment barrier of heterogeneous health data is broken down, so that the reasoning process fully incorporates individualized physiological characteristics, significantly improving the personalized adaptability of health assessment. Third, it integrates the KL divergence of distribution differences with the information gain of medical ontology to construct a path reward index. Based on the overall path reward distribution, it adaptively generates a screening threshold and provides a fallback retention mechanism to achieve an objective quantitative evaluation of the quality of the reasoning link and automatically eliminate low-value invalid branches. Fourth, after screening highly credible and effective reasoning paths, it constructs a health status inference graph to form a complete and traceable medical logical evidence chain. This solves the problems of black box reasoning and lack of visual credibility basis in traditional methods, and comprehensively improves the accuracy of personal health assessment, uncertainty characterization ability and clinical interpretability. Attached Figure Description

[0024] Figure 1 This is a flowchart of a knowledge graph-based personal health assessment reasoning method according to an embodiment of this application.

[0025] Figure 2 This is a multimodal feature fusion weight distribution diagram according to an embodiment of this application. Detailed Implementation

[0026] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0027] Figure 1 This is a flowchart illustrating a knowledge graph-based personal health assessment reasoning method according to an embodiment of this application. Figure 1 As shown, the knowledge graph-based personal health assessment reasoning method includes steps S1 to S4, which are described in detail below.

[0028] S1: Obtain physiological indicators, clinical characteristics, and medical knowledge graph data of the user to be evaluated, and embed all types of data into a high-dimensional embedding space in a unified vectorized manner; Starting from the initial healthy entity of the user to be evaluated, initialize the inference state as a probabilistic hypercone containing a central axis vector and a radial boundary distribution.

[0029] The system reads electronic medical record data and wearable device data from the users to be evaluated. Continuous physiological indicators such as heart rate and blood pressure are extracted from the wearable device data, and clinical manifestations such as cough and fatigue are extracted from the chief complaint, diagnostic records, or symptom descriptions in the medical records. Natural language processing tools are then used to segment, remove stop words, and identify medical entities from the clinical manifestation text. Simultaneously, a medical terminology dictionary and medical knowledge graph are loaded using the Neo4j graph database to extract medical entities such as diseases, symptoms, examination indicators, and drugs, along with their relational triples.

[0030] The extracted medical entities are mapped to dense embedding vectors, which are then used as input features for the graph encoder nodes. The graph encoding is performed by combining the adjacency matrix of the medical knowledge graph, and the output is a high-dimensional entity vector containing the semantics of the graph structure.

[0031] For continuous physiological indicator data, outlier processing, missing value imputation, and normalization are first performed on the values ​​of indicators with different dimensions such as heart rate, blood pressure, blood glucose, and blood oxygen saturation. The values ​​are then spliced ​​together according to the preset indicator order to form a physiological indicator numerical feature vector. The physiological indicator numerical feature vector is then input into a feature extraction network composed of a multilayer perceptron. The feature extraction network outputs a physiological indicator quantification vector that matches the spatial dimension of the medical entity embedding.

[0032] The feature extraction network includes an input layer, at least one hidden fully connected layer, a nonlinear activation layer, a normalization layer, and an output mapping layer. The feature vector of physiological indicators is input into the feature extraction network composed of a multilayer perceptron. The fully connected layer completes feature combination and dimensionality enhancement, the nonlinear activation layer enhances the nonlinear correlation expression between indicators, the normalization layer stabilizes the numerical distribution, and finally the output mapping layer generates a physiological indicator quantization vector that matches the embedding space dimension of the medical entity.

[0033] Using medical entities corresponding to user complaints, confirmed diagnosis records, or initial abnormal indicators as initial health entities, the high-dimensional entity vector corresponding to this initial health entity is used as the initial value of the central axis vector of the probabilistic hypercone, and it is allowed to be updated as a learnable high-dimensional tensor during training. Simultaneously, a Beta distribution is used to model the radial boundary of the probabilistic hypercone, initializing α and β shape parameter vectors greater than 0. The shape parameters describe the probability density distribution of the radial boundary as it extends along the central axis, thus characterizing the uncertainty space range of the health state at the inference starting point.

[0034] As an optional implementation, starting with the initial health entity of the user to be evaluated, the inference state is initialized as a probabilistic hypercone, including: Retrieve the user's medical text records and extract the entity names of the diagnosed diseases as the initial health entities; The initial healthy entity is input into the pre-trained graph encoder, and the high-dimensional entity vector of the initial healthy entity in the high-dimensional embedding space is extracted as the central axis vector. The total number of first-order neighboring nodes of the initial healthy entity in the medical knowledge graph is counted. The total number of neighboring nodes is multiplied by a preset coefficient as the initial uncertainty adjustment parameter of the radial boundary distribution. Based on the initial uncertainty adjustment parameter, the shape parameter of the radial boundary distribution is determined, and the initialization of the probabilistic hypercone is completed.

[0035] In an optional embodiment, the user's electronic medical records and diagnostic reports are scanned using natural language processing technology, and the named entity recognition model is used to extract the diagnosed disease entity names, such as type 2 diabetes and hypertension, as the initial health entities.

[0036] The named entity recognition model comprises a network structure consisting of a bidirectional long short-term memory network and a conditional random field layer cascaded together. The bidirectional long short-term memory network is used to extract the forward and backward semantic features of each word in the input medical text sequence and concatenate them to form a contextual semantic representation; the conditional random field layer is used to learn the transition constraint relationship between adjacent entity labels to avoid combinations in the output label sequence that do not conform to the entity labeling rules.

[0037] The input to the Named Entity Recognition (NAME) model is a sequence of word vectors from a medical text record, processed through word segmentation and word vectorization. During runtime, the model first inputs the word vector sequence into a bidirectional long short-term memory (LSTM) network to obtain the contextual hidden state for each word. Then, a linear mapping layer converts the contextual hidden state into an emission score matrix, representing the degree of matching between each word and different entity category labels. Subsequently, a conditional random field (CRF) layer integrates the emission score matrix and the pre-learned label transition score matrix, employing maximum a posteriori (MAP) decoding or Viterbi decoding algorithms to select the label sequence with the highest overall score from all candidate label sequences as the output. Finally, the NAME model outputs the entity category label corresponding to each input word, such as disease, symptom, examination indicator, or non-entity label, and parses the initial healthy entity from continuous entity label fragments.

[0038] In this system, each vector in the word vector sequence corresponds to an input word segmentation, the emission score matrix is ​​used to represent the correspondence between the word segmentation and the entity category label, and the label transition score matrix is ​​used to represent the rationality of the transition between adjacent entity category labels.

[0039] The extracted initial health entity is input into a graph encoder pre-trained on a medical knowledge graph to generate a high-dimensional entity vector corresponding to the initial health entity.

[0040] The graph encoder employs a multi-layer graph attention network structure, including an entity embedding layer, a multi-head attention layer, a feature aggregation layer, and a feedforward neural network layer. Specifically, based on the entity vocabulary in the medical knowledge graph, initial healthy entities are converted into entity index identifiers and mapped to initial entity embedding vectors through the entity embedding layer. Subsequently, the node feature matrix and adjacency matrix from the medical knowledge graph are input into the graph encoder, where the node feature matrix contains the initial embedding vectors of the initial healthy entities and their neighboring entities, and the adjacency matrix represents the connection relationships between entities.

[0041] During graph encoding, the multi-head attention layer calculates the influence weights of different neighboring entities on the current entity based on the feature similarity and graph connectivity between the initial healthy entity and the first-order neighboring entities; the feature aggregation layer performs weighted aggregation of the features of the neighboring entities based on the influence weights to obtain an intermediate entity representation that integrates the semantics of the neighboring entities; the feedforward neural network layer performs nonlinear mapping and dimension adjustment on the intermediate entity representation to output a high-dimensional entity vector containing the semantics of the medical graph structure.

[0042] The graph encoder outputs a high-dimensional entity vector for each medical entity. These high-dimensional entity vectors are dense floating-point vectors ranging from 64 to 256 dimensions. All high-dimensional entity vectors are mapped to the same high-dimensional embedding space, and the high-dimensional entity vector corresponding to the initial healthy entity is retrieved as the central axis vector of the probabilistic hypercone. For example, the medical entity "type 2 diabetes" is encoded by the graph encoder to generate a 128-dimensional dense floating-point high-dimensional entity vector. When initializing the probabilistic hypercone, if "type 2 diabetes" is the initial healthy entity for the user to be evaluated, the corresponding 128-dimensional high-dimensional entity vector is retrieved and set as the central axis vector of the probabilistic hypercone.

[0043] The total number of first-order neighbor nodes of the initial healthy entity in the medical knowledge graph is queried in the graph database. A preset coefficient is introduced based on the average number of first-order neighbor nodes of all entities in the medical knowledge graph. This coefficient is used to compress or expand the numerical range of the number of first-order neighbor nodes to avoid the problem of saturation of concentration parameters and loss of distinguishability in subsequent calculations. The preset coefficient is the reciprocal of the average of the first-order neighbors of all entities.

[0044] The initial uncertainty adjustment parameter u is obtained by multiplying the total number of first-order adjacent nodes by a preset coefficient. Subsequently, a Beta distribution is used to model the radial boundary of the probabilistic hypercone, converting u into a concentration parameter τ of the Beta distribution. ,in and λ is the preset concentration parameter range, λ is the adjustment coefficient, and u is the initial uncertainty adjustment parameter.

[0045] An initial mean μ is set for the radial boundary distribution, for example, μ is set to 0.5, to make the radial probability distribution symmetrical, eliminate the retrieval bias in the high-dimensional embedding space, and ensure that there is no one-sided retrieval tilt in the healthy inference process. The α shape parameter and β shape parameter of the Beta distribution are calculated according to α=μτ and β=(1-μ)τ. When the number of first-order neighboring nodes is large, u increases and τ decreases, the Beta distribution is more dispersed, and the coverage of the probability hypercone is wider; when the number of first-order neighboring nodes is small, u decreases and τ increases, the Beta distribution is more concentrated, and the coverage of the probability hypercone is narrower, thus completing the initialization of the probability hypercone.

[0046] S2, in each inference step, select the relation transformation matrix according to the uncertainty of the radial boundary distribution and the reachability relation set of the current entity; apply the matrix to the current probabilistic hypercone to generate the next region, and extract the graph entities in the region as candidate entities; if the candidate entity is associated with physiological indicators, then according to the inference path context, fuse the candidate entity embedding vector and physiological indicators through the attention mechanism to generate the context-aware entity state, and update the central axis vector of the next probabilistic hypercone.

[0047] A pre-built relation transformation operator library is constructed according to the relation types in the medical knowledge graph. Each relation type corresponds to at least one set of relation transformation matrices, which are used to characterize the transition direction of the corresponding medical relation in the high-dimensional embedding space. Preferably, each relation type is configured with two types of matrices: a low-rank generalization matrix and a high-rank specialization matrix, which are used for inference states with high and low uncertainty, respectively. The operator library is pre-built offline and directly retrieved and used during the online inference stage.

[0048] During online reasoning, the reachable relation set is first determined based on the adjacency relationships of the current entity in the medical knowledge graph. Specifically, using the current entity as the starting point of the graph query, the first-order adjacency edges directly connected to the current entity are retrieved in the medical knowledge graph, and the relation types and adjacent entity nodes corresponding to the first-order adjacency edges are extracted. The relation types include at least one of the following: symptom association, disease concurrency, examination indicator association, drug treatment, or risk factor association. All retrieved first-order adjacency relation types are deduplicated to form the reachable relation set of the current entity.

[0049] Extract a set of candidate relation transformation matrices corresponding to the reachable relation set from the relation transformation operator library. Read the uncertainty of the radial boundary distribution of the current probabilistic hypercone. If the uncertainty of the current radial boundary distribution is greater than a preset threshold, select a low-rank generalization matrix from the candidate relation transformation matrix set; if the uncertainty of the current radial boundary distribution is less than or equal to the preset threshold, select a high-rank specialization matrix from the candidate relation transformation matrix set. The reachable relation set of the current entity limits the medical semantic direction, and the uncertainty of the radial boundary distribution limits the generalization or specialization form of the relation transformation matrix.

[0050] If the same current entity corresponds to multiple reachable relation types, the current central axis vector is mapped using the relation transformation matrix corresponding to each reachable relation type, resulting in multiple independent inference branches and corresponding intermediate central axis vectors. Each branch independently holds its own probabilistic hypercone parameter. Candidate entities are recalled based on each intermediate central axis vector, and a comprehensive score is calculated by combining semantic consistency score, prior probability, and candidate entity inclusion probability. One or more relation transformation results with higher comprehensive scores are retained for further inference.

[0051] The matrix is ​​applied to the current probabilistic hypercone to generate the next region and extract the graph entities falling into the region as candidate entities. Specifically, let the central axis vector of the current probabilistic hypercone be... The selected relational transformation matrix is Transform the relation into a matrix With the central axis vector Multiply to obtain the mapped intermediate central axis vector, and then perform L2 normalization on the intermediate central axis vector to maintain its unit direction vector form, which will serve as the center direction of the next region.

[0052] The Beta radial boundary distribution is obtained based on the α and β shape parameters of the current radial boundary distribution. A normalized radial value corresponding to a preset confidence quantile is selected, for example, the 95th quantile is chosen as the normalized radial boundary. When the α and β shape parameters are in vector form, the quantiles corresponding to each dimension are calculated separately, and the average, maximum, or weighted average is taken as the unified normalized radial boundary. This normalized radial boundary is multiplied by a preset maximum angle range to obtain the angle coverage threshold around the central axis vector.

[0053] L2 normalization is applied to each entity vector in the medical knowledge graph. The cosine similarity between each entity vector and the central axis vector is calculated, and this cosine similarity is converted into the angular deviation of the entity vector relative to the central axis vector using an inverse cosine function. If the angular deviation is less than or equal to the angular coverage threshold, the entity is determined to fall into the next region and added to the candidate entity set.

[0054] For each candidate entity in the candidate entity set, it is determined whether it has a physiological indicator attribute association in the medical knowledge graph. If so, the corresponding physiological indicator quantization vector is extracted, and the historical inference node sequence is extracted. The path hidden state is extracted using a Long Short-Term Memory (LSTM) network as the inference path context. The candidate entity embedding vector and the corresponding physiological indicator quantization vector are mapped to vectors of the same dimension as the inference path context, and the fusion weights are calculated using a dot product attention mechanism. The fusion weights are applied to the candidate entity embedding vector and the physiological indicator quantization vector, respectively. After weighted summation and layer normalization, a context-aware entity state vector corresponding to the candidate entity is generated. Subsequently, the context-aware entity state vector is fused with the intermediate central axis vector of the corresponding relation transformation result to obtain the next central axis vector of the inference branch to which the candidate entity belongs. If the candidate entity is not associated with a physiological indicator value, the intermediate central axis vector is directly used as the next central axis vector of the inference branch.

[0055] For each inference branch, the α and β vectors of the next radial boundary distribution of that branch are updated based on the dispersion of the candidate entity set corresponding to that branch relative to the central axis vector of the next step of that branch. Specifically, the angular deviation between each entity vector in the candidate entity set corresponding to that branch and the central axis vector of the next step of that branch is calculated, and each angular deviation is divided by a preset maximum angular range to obtain a normalized angular deviation sequence; the average or variance of the normalized angular deviation sequence is taken to obtain the normalized dispersion; the normalized dispersion is used as the current uncertainty adjustment parameter, and converted into the α shape parameter and β shape parameter of the next radial boundary distribution according to the same Beta distribution parameter mapping method as in the initialization phase, thereby completing the update of the radial boundary distribution.

[0056] As an optional implementation, a relation transformation matrix is ​​selected from the operator library based on the degree of uncertainty of the radial boundary distribution and the current entity's reachability relation set, including: Calculate the differential entropy of the current probability hypercone radial boundary distribution as the uncertainty quantification value; The set of reachable relations for the current entity is determined based on the adjacency relationship of the current entity in the medical knowledge graph, and a set of candidate relation transformation matrices corresponding to the reachable relation set is extracted from the operator library; The uncertainty quantification value is compared with a preset entropy threshold; If the uncertainty quantification value is greater than the entropy threshold, then the low-rank generalization matrix trained with nuclear norm penalty is selected as the relation transformation matrix from the candidate relation transformation matrix set; If the uncertainty quantization value is less than or equal to the entropy threshold, then the high-rank specialization matrix trained with L2 regularization is selected as the relation transformation matrix from the candidate relation transformation matrix set.

[0057] More specifically, based on the α-shape and β-shape parameters of the current radial boundary distribution, the differential entropy function of the Beta distribution object in the statistical computing library is invoked to calculate the differential entropy of the current Beta distribution. This differential entropy is then normalized, and the normalized result is used as the current uncertainty quantification value. Subsequently, the current uncertainty quantification value is compared with a preset entropy threshold, where the entropy threshold can be determined based on the distribution characteristics of uncertainty quantification values ​​in the historical medical record validation set, such as selecting the median, upper quartile, or empirical quantile as the threshold.

[0058] When the current uncertainty quantification value is greater than the entropy threshold, a low-rank generalization matrix trained with nuclear norm penalty is selected as the current relation transformation matrix from the set of candidate relation transformation matrices corresponding to the current entity reachable relation set. For example, a 128×128-dimensional low-rank relation transformation matrix with a rank of 8 can be selected to guide the probabilistic hypercone to a wider candidate region. When the current uncertainty quantification value is less than or equal to the entropy threshold, a high-rank specialization matrix trained with L2 regularization is selected as the current relation transformation matrix from the set of candidate relation transformation matrices corresponding to the current entity reachable relation set. For example, a full-rank 128×128-dimensional high-rank relation transformation matrix can be selected to refine the mapping of the current central axis vector.

[0059] As an optional implementation, an attention mechanism is used to assign fusion weights to the candidate entity embedding vector and the physiological indicator quantization vector based on the inference path context, generating a context-aware entity state, including: The historical path sequence is input into the Long Short-Term Memory network to extract the hidden layer state, which serves as the context for inference paths; The candidate entity embedding vector and the physiological indicator quantization vector are respectively mapped to the same dimension as the inference path context through a fully connected layer; The inner product of the mapped candidate entity embedding vector and the mapped physiological index quantization vector with the inference path context is calculated respectively. The two inner product values ​​are then normalized using the Softmax function to obtain the fusion weights of each modality. The fusion weights are multiplied by the mapped candidate entity embedding vector and the mapped physiological index quantization vector, respectively, and then summed to obtain the context-aware entity state.

[0060] More specifically, the variable-length historical reasoning path sequence, which consists of alternating nodes and relationships, is sequentially input into a long short-term memory network with a multi-layer structure, and the hidden layer state vector output at the last time step is extracted, for example, with a dimension of 128, as the reasoning path context.

[0061] The Long Short-Term Memory (LSTM) network consists of an input layer, multiple repeating memory cell modules, and an output layer. Each memory cell contains a forget gate, an input gate, candidate memory cell states, and an output gate, which helps alleviate the vanishing gradient problem during long sequence feature extraction. The input to the LTM network is a temporal path feature matrix composed of historical nodes and relational feature vectors.

[0062] At each time step, the Long Short-Term Memory (LSTM) network receives the current input vector and recursively calculates it by combining the hidden layer state and cell state from the previous time step. The LTM network uses a forget gate to determine which information from the cell state in the previous time step needs to be retained or forgotten, controls the proportion of current input features written into the cell state through the input gate, and generates new memory content that can be written to the current time step using candidate cell states. The information retained by the forget gate is fused with the new memory content written by the input gate to obtain the cell state for the current time step; then, the output gate controls the amount of information output from the cell state to the hidden layer state to generate the hidden layer state for the current time step.

[0063] Extract the candidate entity embedding vector, with an example dimension of 256, and the physiological indicator quantification vector associated with the candidate entity, for example, the feature vector corresponding to fasting blood glucose of 8.5 mmol / L after min-max normalization encoding, with a dimension of 32. The two types of vectors are non-linearly mapped through independently configured fully connected layers and linear rectified activation functions, so that the output dimensions of the two are uniformly aligned to a 128-dimensional space, consistent with the inference path context dimension.

[0064] Based on the dot product attention mechanism, the inner product of the 128-dimensional candidate entity mapping vector and the 128-dimensional inference path context vector, as well as the inner product of the 128-dimensional physiological indicator mapping vector and the context vector, are calculated to obtain two real scalars. These two real scalars are then input into the Softmax function for normalization to obtain the fusion weights. The distribution of the fusion weights of candidate entity features and physiological indicator features under different inference samples is shown below. Figure 2 As shown. For example, if the inner products of the two sets are 1.2 and 2.5 respectively, and they are input into the binary classification Softmax function for normalization, the output will be two sets of fusion weights with a sum of 1. For example, the entity fusion weight is 0.21 and the physiological indicator fusion weight is 0.79.

[0065] The two obtained fusion weights are multiplied by their corresponding 128-dimensional mapping vectors, and then the two weighted vectors are added together to obtain a 128-dimensional context-aware entity state. This context-aware entity state does not replace the original entity embedding in the knowledge graph embedding model, but is used as an individualized state compensation quantity. It is added to the intermediate central axis vector of the correspondence transformation result by residual addition, and the added vector is L2 normalized to update the central axis vector of the next probabilistic hypercone.

[0066] S3 consists of path segments composed of historical paths and candidate entities. Semantic consistency scores are obtained through a knowledge graph embedding model, and the comprehensive confidence of the path segments is calculated by combining prior probabilities. The KL divergence of the radial boundary distribution between the current step and the previous step is calculated, and the reward value of the current path segment is calculated by combining the information content gain of the candidate entities in the medical ontology.

[0067] The path segment from the starting entity to the current entity is input into the TransE knowledge graph distance scoring model. The pre-trained high-dimensional entity vectors and relation vectors corresponding to the head entity, relation, and tail entity within the path are extracted. The L1-norm distance or L2-norm distance of the head entity vector plus the relation vector minus the tail entity vector is calculated. The calculated distance d is substituted into the negative exponential function score=exp(-d), or transformed to the probability interval of 0 to 1 using the sigmoid negative mapping function score=sigmoid(-d), to obtain the semantic consistency score. The frequency of occurrence of the relation type corresponding to the currently selected relation transformation matrix in the medical knowledge graph is statistically analyzed. The occurrence frequency and the total relation frequency are smoothed and normalized to obtain the prior probability that the value falls within the probability interval of 0 to 1. The semantic consistency score is multiplied by the prior probability to obtain the comprehensive confidence of the path segment.

[0068] The radial boundary distribution parameters of the current step and the previous step are obtained respectively. The KL divergence of the radial boundary distribution of the current step relative to the radial boundary distribution of the previous step is calculated to obtain the divergence measure representing the degree of change in the model space exploration. Using the information theory information content algorithm, based on the directed acyclic graph of the medical ontology, the joint occurrence probability of the candidate entity node and its related subordinate nodes is statistically analyzed, and the negative logarithm of the joint occurrence probability is calculated as the information content gain. As an optional simplified implementation, the depth of the candidate entity in the medical ontology tree structure can be directly obtained, and the information content gain can be approximated by this depth value.

[0069] By setting constant weight coefficients, the KL divergence term and the information content gain term are weighted and summed to obtain the immediate reward value of the current path segment. When this reward value is used for reinforcement learning policy updates, the path segment, relation transformation matrix, overall confidence, and reward value can be stored in the trajectory sample cache for subsequent parameter update calculations of the policy network.

[0070] As an optional implementation, a path segment is composed of historical paths and candidate entities. A semantic consistency score is obtained through a knowledge graph embedding model, and the comprehensive confidence of the path segment is calculated by combining prior probabilities, including: Input path fragments into the knowledge graph embedding model and output the semantic consistency score of the path fragments; Search and retrieve data in the medical knowledge graph, count the frequency of occurrence of the corresponding relation type in the currently selected relation transformation matrix, and perform smoothing and normalization on the occurrence frequency and the total relation frequency to obtain the prior probability with a value in the probability interval of 0 to 1. The semantic consistency score is multiplied by the prior probability to obtain the overall confidence level of the path segment.

[0071] More specifically, the current triplet reasoning path segment, consisting of the source entity, reasoning relation, and target candidate entity, is parsed and input into a fine-tuned knowledge graph embedding model. The knowledge graph embedding model employs a translation distance scoring network structure, comprising an entity embedding layer and a relation embedding layer. The entity embedding layer maps discrete source entities and target candidate entities to continuous vectors, while the relation embedding layer maps reasoning relations to translation vectors within the same continuous vector space.

[0072] The input to the knowledge graph embedding model is a triplet index identifier consisting of a source entity, a reasoning relation, and a target candidate entity. The formula for calculating the intra-network distance is: Where h is the source entity vector, r is the inference relation vector, and t is the target candidate entity vector. This represents the distance between the L1 norm and the L2 norm.

[0073] The calculated distance values ​​undergo a reverse mapping process. For example, a negative exponential function, score=exp(-d), is used to convert the distance values ​​into likelihood scores within the probability interval of 0 to 1, and this likelihood score is used as the semantic consistency score. Simultaneously, the frequency of the currently selected relationship is counted in the medical knowledge graph. To avoid the zero-probability division problem caused by low-frequency relationships, a smoothing term with a value of 1 is introduced. This smoothing term of 1 is added to the counted frequency and divided by the sum of the total relationship frequency and the total smoothing term to calculate the prior probability. Subsequently, the semantic consistency score is multiplied by the prior probability to obtain the overall confidence of the path segment.

[0074] For example, the semantic consistency score for hypertension complicated with ventricular hypertrophy is 0.85, and the calculated prior probability value is 0.15. Multiplying the semantic consistency score of 0.85 with the prior probability of 0.15 yields a product of 0.1275, which is used as the overall confidence of the current path segment.

[0075] As an optional implementation, the KL divergence of the radial boundary distribution of the probabilistic hypercone between the current and previous steps is calculated. Combined with the information content gain of the candidate entities in the medical ontology, the reward value of the current path segment is calculated, including: The Monte Carlo sampling method is used to extract samples from the radial boundary distributions of the current probability hypercone and the previous probability hypercone, respectively, and the KL divergence of the current radial boundary distribution relative to the previous radial boundary distribution is estimated. The depth of candidate entities in the medical ontology tree structure is used to characterize the information content gain. The reward value of the current path segment is obtained by linearly weighting and summing the KL divergence and the information content gain.

[0076] More specifically, the Monte Carlo sampling method is used to extract radial sample points from the radial boundary Beta distributions of the current probability hypercone and the previous step's probability hypercone, respectively. The preferred range for the number of samples is 1000 to 5000. The radial sample sets obtained from the current step and the previous step are divided into the same discrete intervals, and the sample proportions corresponding to each discrete interval are statistically analyzed to obtain the sample probability distribution within each discrete interval. Based on the discrete probability distributions, the KL divergence between the two radial boundary distributions is calculated.

[0077] Simultaneously, the depth of the current candidate entity is obtained in the medical ontology tree structure, and this depth value represents the information content gain. For example, the root node has a depth of 0, and the depth of the proliferative diabetic retinopathy node is 4, with this depth value of 4 being used as the information content gain.

[0078] The KL divergence and information content gain are linearly weighted and summed to obtain the immediate reward value of the path segment. During the linear weighted summation, the ranges of the divergence weight and depth weight are determined through grid optimization using multiple sets of medical record datasets. The optimal range for the divergence weight is 0.5 to 2.0, and the optimal range for the depth weight is 0.1 to 0.5. Within this range, the model achieves the best combination of path selection accuracy and training convergence speed.

[0079] S4. Repeat the reasoning calculation until the preset termination condition is met. The sum of the reward values ​​of all path segments on each reasoning path is used as the cumulative reward. Based on the cumulative reward, the effective reasoning path is selected and the health status inference graph is constructed and output.

[0080] Determine whether the current inference jump count has reached the preset maximum inference step count, whether the current candidate entity belongs to the target disease diagnosis category, or whether the current entity has no successor node; if any of these conditions are met, stop the multi-step inference calculation.

[0081] After inference is complete, the generated complete inference paths are traversed, and the reward values ​​of each path segment in each path are accumulated to obtain the corresponding cumulative reward value. A screening threshold is determined based on the average and standard deviation of the cumulative reward values ​​of all inference paths. Inference paths with a cumulative reward value greater than or equal to the screening threshold are retained as valid inference paths. If no inference path meets the screening criteria, at least one inference path with the highest cumulative reward value is retained as a valid inference path.

[0082] A health status inference graph is constructed based on effective inference paths. The specific construction rules are as follows: the initial healthy entity, intermediate inference entities, and terminal disease entities are used as graph nodes, and directed edges are established according to the inference order from the source entity to the target candidate entity in the path segment. For each directed edge, the edge weight is calculated based on the comprehensive confidence of the corresponding path segment and the cumulative reward value of the path. When multiple effective inference paths contain the same directed edge, the weights of multiple edges are aggregated by mean, maximum value, or weighted average to obtain the final weight of the directed edge.

[0083] Finally, the output is a health status inference graph consisting of nodes, edges, and edge weights, and a directed graph data structure is constructed. For example, Matplotlib or other visualization tools can be called to generate a two-dimensional inference image, thereby providing a quantitative health status inference result with path interpretability to the medical clinical decision-making system.

[0084] As an optional implementation, the inference calculation is repeated until a preset termination condition is met. The inference paths are then filtered and aggregated based on the accumulated reward value to obtain valid inference paths, including: When the maximum number of inference steps is reached or a terminal entity with no successor node is reached, a preset termination condition is triggered to stop the inference calculation and the cumulative reward value of all path segments on each inference path is calculated. Based on the average and standard deviation of the cumulative reward values ​​of all current inference paths, calculate the sum of the average and standard deviation, and use it as the screening threshold; Retain inference paths whose cumulative reward value is greater than or equal to the screening threshold as valid inference paths, and remove the rest of the inference paths; When there is no reasoning path with a cumulative reward value greater than or equal to the screening threshold, at least one reasoning path with the highest cumulative reward value is retained as a valid reasoning path.

[0085] More specifically, a reasoning step counter is maintained. When the set maximum reasoning step count is reached, or when the current entity node is detected to have no successor node (i.e., out-degree is 0), a preset termination condition is triggered to stop the reasoning calculation.

[0086] Iterate through all the generated inference paths and sum the reward values ​​of all path segments contained in each path, for example, the sequence 1.12, 0.85, 1.30, to obtain the cumulative reward value of that path, for example, 3.27.

[0087] After calculating the cumulative reward value of all inference paths, assuming a total of 50 paths, calculate the average and standard deviation of the cumulative reward value, for example, the average is 2.40 and the standard deviation is 0.65.

[0088] The sum of the mean and standard deviation, 3.05, is used as the filtering threshold. The 50 inference paths are re-examined, retaining those with a cumulative reward value greater than or equal to the filtering threshold of 3.05 (e.g., paths with a cumulative reward value greater than 3.05 are considered valid). The remaining paths with cumulative reward values ​​below the filtering threshold are removed. If the cumulative reward value of all inference paths does not reach the filtering threshold, at least one inference path with the highest cumulative reward value is retained as a valid inference path to avoid empty path sets during the inference graph construction phase.

[0089] As an optional implementation, a health status inference graph is constructed and output, including: Directed edges are established according to the order from the source entity to the target candidate entity in the effective reasoning path, and the reasoning relation corresponding to the relation transformation matrix is ​​used as the edge type of the directed edge. For each path segment in the valid inference path, obtain the comprehensive confidence of the path segment, and calculate the edge weight of the corresponding directed edge by combining the cumulative reward value of the valid inference path to which the path segment belongs. When multiple valid inference paths contain the same directed edge, the weights of multiple edges corresponding to the same directed edge are aggregated, and the aggregation result is used as the final weight of the directed edge. Output a health status inference graph consisting of nodes, edges, and edge weights, and construct a directed graph data structure.

[0090] More specifically, the terminal node of the effective inference path is set as the terminal entity node of the health status inference graph, such as the acute myocardial infarction node. At the same time, the initial health entity (e.g., hypertension) is set as the starting node, and the intermediate inference entities in the inference path (e.g., vascular endothelial injury, coronary artery sclerosis) are set as intermediate nodes.

[0091] Based on this, directed edges are established according to the order from the source entity to the target candidate entity in the effective inference path, and the inference relation corresponding to the relation transformation matrix is ​​used as the edge type of the directed edge. For each path segment, the comprehensive confidence of the path segment is extracted, for example, 0.15, and the cumulative reward value of the effective inference path to which the path segment belongs is extracted, for example, 3.27. The cumulative reward value is subjected to max-min normalization to obtain the normalized cumulative reward value, for example, 0.82. Fixed hyperparameters are obtained by offline grid optimization as preset weight parameters. For example, the comprehensive confidence weight is set to 0.7 and the normalized cumulative reward value weight is set to 0.3. The sum of the two weights is fixed to 1. The comprehensive confidence of the path segment (0.15) and the normalized cumulative reward value (0.82) are weighted and summed to obtain the edge weight of the directed edge corresponding to the path segment. If multiple valid inference paths share the same directed edge, then the weights of the directed edge obtained in different paths are aggregated by mean, maximum, or weighted average. The aggregated result is used as the final weight of the directed edge, and the health status inference graph is output.

[0092] The ablation experiment used a large-scale Chinese medical knowledge graph and 50,000 electronic medical record texts as the test dataset, divided into training, validation, and test sets according to a preset ratio. Evaluation metrics included the top three hit rate, average reciprocal ranking, and end-to-end health status inference accuracy. The experimental hardware environment consisted of four large-memory graphics processors, with an initial learning rate set to 0.1% and a maximum inference step limit of 4 steps. Four sets of models were set up for comparison: the first set was the basic model, using a fixed full-rank relation transformation matrix and fusing entity vectors and physiological indicator vectors through concatenation; the second set introduced a mechanism based on uncertainty-based relation transformation matrix selection on top of the basic model; the third set introduced a multimodal feature fusion mechanism based on attention on top of the basic model; and the fourth set was a complete solution model incorporating the above technical features.

[0093] Experimental results show that the first group of basic models achieved a top-three hit rate of 68.5%, an average regression rank of 0.512, and an inference accuracy of 71.4%. The second group, after introducing a matrix selection mechanism, saw its top-three hit rate increase to 74.2%, its average regression rank reach 0.589, and its inference accuracy improve to 78.6%. The third group, after introducing an attention-based multimodal fusion mechanism, achieved a top-three hit rate of 76.5%, an average regression rank of 0.615, and an inference accuracy of 80.2%. The fourth group, with its complete solution, achieved a top-three hit rate of 83.7%, an average regression rank of 0.694, and an inference accuracy of 88.5%.

[0094] The comparison between the second and first groups shows that switching between the low-rank generalization matrix and the high-rank specialization matrix using uncertainty quantization results helps reduce the impact of long-tail interference relationships in the knowledge graph on the inference path selection, and preserves local feature details when the inference state tends to converge. The comparison between the third and first groups shows that assigning fusion weights to the entity embedding vector and the physiological indicator quantization vector based on the inference path context, and using the fused context-aware entity state to update the next probabilistic hypercone central axis vector, helps alleviate the problem of insufficient multimodal feature representation under static splicing. The experimental results of the fourth group show that the probabilistic boundary constraint mechanism and the multimodal feature weighting mechanism can work synergistically to improve the accuracy of health status inference and the reliability of inference results on the test dataset.

[0095] In a second embodiment, this application also discloses a knowledge graph-based personal health assessment reasoning system, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement a knowledge graph-based personal health assessment reasoning method according to this application.

[0096] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0097] It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the scope of protection of this application. Therefore, the scope of protection of this patent application shall be determined by the appended claims.

Claims

1. A knowledge graph-based reasoning method for personal health assessment, characterized in that, include: Acquire physiological indicators, clinical characteristics, and medical knowledge graph data of the users to be evaluated, and embed all types of data into a high-dimensional embedding space in a unified vectorized manner. Starting with the initial health entity of the user to be evaluated, the inference state is initialized as a probabilistic hypercone containing a central axis vector and a radial boundary distribution; In each step of reasoning, a relation transformation matrix is ​​selected based on the degree of uncertainty of the radial boundary distribution and the set of reachable relations of the current entity; The relation transformation matrix is ​​applied to the current probabilistic hypercone to generate the next region. Graph entities within the region are extracted as candidate entities. If a candidate entity is associated with a physiological indicator, the candidate entity embedding vector and the physiological indicator are fused through an attention mechanism based on the inference path context to generate a context-aware entity state and update the central axis vector of the next probabilistic hypercone. The path segment is composed of historical paths and candidate entities. The semantic consistency score is obtained through the knowledge graph embedding model. The comprehensive confidence of the path segment is calculated by combining the prior probability. The KL divergence of the radial boundary distribution between the current step and the previous step is calculated. The reward value of the current path segment is calculated by combining the information content gain of the candidate entity in the medical ontology. Repeated reasoning calculations until the preset termination condition is met, the cumulative reward is the sum of the reward values ​​of all path segments on each reasoning path, and the effective reasoning path is selected based on the cumulative reward to construct and output the health status inference graph.

2. The knowledge graph-based personal health assessment reasoning method according to claim 1, characterized in that, The initialization of the inference state, starting with the initial health entity of the user to be evaluated, as a probabilistic hypercone containing a central axis vector and a radial boundary distribution, includes: Retrieve the user's medical text records and extract the entity names of the diagnosed diseases as the initial health entities; The initial healthy entity is input into a pre-trained graph encoder, and the feature vector of the entity in the high-dimensional embedding space is extracted as the central axis vector. The total number of first-order adjacent nodes of the initial healthy entity in the medical knowledge graph is counted. The total number of adjacent nodes is multiplied by a preset coefficient as the initial uncertainty adjustment parameter of the radial boundary distribution. Based on the initial uncertainty adjustment parameter, the shape parameter of the radial boundary distribution is determined to complete the initialization of the probabilistic hypercone.

3. The personal health assessment reasoning method based on knowledge graphs according to claim 1, characterized in that, The step of selecting the relation transformation matrix based on the uncertainty of the radial boundary distribution and the current entity's reachability relation set includes: Calculate the differential entropy of the current probability hypercone radial boundary distribution as the uncertainty quantification value; The uncertainty quantification value is compared with a preset entropy threshold; If the uncertainty quantification value is greater than the entropy threshold, then a low-rank generalization matrix trained with nuclear norm penalty is selected as the relation transformation matrix from the candidate relation transformation matrix set corresponding to the current entity reachable relation set; if the uncertainty quantification value is less than or equal to the entropy threshold, then a high-rank specialization matrix trained with L2 regularization is selected as the relation transformation matrix from the candidate relation transformation matrix set corresponding to the current entity reachable relation set.

4. The knowledge graph-based personal health assessment reasoning method according to claim 3, characterized in that, The step of generating a context-aware entity state by fusing candidate entity embedding vectors and physiological indicators through an attention mechanism based on the inference path context includes: The historical path sequence is input into a long short-term memory network to extract the hidden layer state, which serves as the inference path context. The candidate entity embedding vector and the physiological indicator quantization vector are respectively mapped to the same dimension as the inference path context through a fully connected layer. The inner product of the mapped candidate entity embedding vector and the mapped physiological index quantization vector with the inference path context is calculated respectively. The two inner product values ​​are normalized by the Softmax function to obtain the fusion weight of each modality. The fusion weight is multiplied by the mapped candidate entity embedding vector and the mapped physiological index quantization vector respectively and summed to obtain the context-aware entity state.

5. The knowledge graph-based personal health assessment reasoning method according to claim 1, characterized in that, The step of obtaining semantic consistency scores through a knowledge graph embedding model and calculating the comprehensive confidence of path segments by combining prior probabilities includes: Input the path segment into the knowledge graph embedding model and output the semantic consistency score of the path segment; Search and retrieve data in the medical knowledge graph, count the frequency of occurrence of the corresponding relation type in the currently selected relation transformation matrix, and perform smoothing and normalization on the occurrence frequency and the total relation frequency to obtain the prior probability with a value in the probability interval of 0 to 1. The semantic consistency score is multiplied by the prior probability to obtain the overall confidence level of the path segment.

6. The knowledge graph-based personal health assessment reasoning method according to claim 5, characterized in that, The relationship types include one or more of the following: symptom association, disease comorbidity, examination indicator association, drug treatment, or risk factor association.

7. A knowledge graph-based personal health assessment reasoning method according to claim 1 or 2, characterized in that, The calculation of the reward value for the current path segment includes: The Monte Carlo sampling method is used to extract samples from the radial boundary distributions of the current probability hypercone and the previous probability hypercone, respectively, and the KL divergence of the current radial boundary distribution relative to the previous radial boundary distribution is estimated. Obtain the depth of the candidate entity in the medical ontology tree structure, and use the depth value to represent the information content gain; The reward value of the current path segment is obtained by linearly weighting and summing the KL divergence and the information content gain.

8. The personal health assessment reasoning method based on knowledge graphs according to claim 1, characterized in that, The selection of valid reasoning paths based on cumulative rewards includes: When the maximum number of inference steps is reached or a terminal entity with no successor node is reached, a preset termination condition is triggered, the inference calculation is stopped, and the cumulative reward value of all path segments on each inference path is calculated to obtain the cumulative reward value. Calculate the sum of the average and standard deviation of the cumulative reward values ​​of all current inference paths, and use it as the screening threshold; retain inference paths whose cumulative reward values ​​are greater than or equal to the screening threshold as valid inference paths, and remove the rest of the inference paths; When there is no reasoning path with a cumulative reward value greater than or equal to the screening threshold, at least one reasoning path with the highest cumulative reward value is retained as a valid reasoning path.

9. The knowledge graph-based personal health assessment reasoning method according to claim 1, characterized in that, The construction and output of the health status inference graph includes: Directed edges are established according to the order from the source entity to the target candidate entity in the effective reasoning path, and the reasoning relation corresponding to the relation transformation matrix is ​​used as the edge type of the directed edges. For each path segment in the valid inference path, obtain the comprehensive confidence of the path segment, and calculate the edge weight of the corresponding directed edge in combination with the cumulative reward value of the valid inference path to which the path segment belongs; When multiple valid inference paths contain the same directed edge, the weights of the multiple edges corresponding to the same directed edge are aggregated, and the aggregation result is used as the final weight of the directed edge.

10. The knowledge graph-based personal health assessment reasoning method according to claim 1, characterized in that, After extracting the map entities within the region as candidate entities, the process also includes: For each inference branch, the radial boundary distribution of the next step of the branch is updated based on the degree of dispersion of the candidate entity set corresponding to that branch relative to the central axis quantity of the next step of the branch.

Images