Medical care data digital management method and system

Abstract

The invention provides a medical care data digital management method and system, and the method can obtain a patient information graph structures, and can obtain the adjacent correlation representation vector of an adjacent vertex of an s-th patient information vertex for the s-th patient information vertex in an e-th information dimension, the first vertex correlation classification represents that adjacent vertexes of the corresponding s patient information vertex have relatively great influence on the s patient information vertex; the first adjacent representation vectors of the adjacent vertexes of the s patient information vertex under different information dimensions can be integrated to obtain the adjacent correlation representation vectors of the adjacent vertexes of the s patient information vertex under the e information dimension. On the basis of the mode, the dimension patient characterization vectors of the b patient information vertexes in the a information dimensions can be obtained, so that characterization information of the patient information vertexes in the single information dimension is richer, and the accuracy of obtaining the multi-patient characterization vectors is improved.

Description

Technical Field

[0001] The present application relates to the field of data processing, and in particular, to a method and system for digital management of medical care data. Background Art

[0002] In the field of medical care, with the continuous advancement of medical technology and the deepening of medical information construction, medical institutions have accumulated a huge amount of patient data. These data cover multiple information dimensions such as patients' physiological indicators, imaging data, pathology reports, medication records, living habits, etc., providing valuable basic resources for precision medicine and personalized care. However, how to efficiently process and analyze these multi-source heterogeneous data and extract valuable information from them to support clinical decision-making has become a major challenge facing the current digital management of medical care data.

[0003] Traditional data processing methods often rely on manual analysis and expert experience, which is not only inefficient but also difficult to fully capture the complex relationships and hidden patterns in the data. With the rapid development of big data and artificial intelligence technologies, especially the successful application of deep learning and machine learning in image recognition, natural language processing and other fields, new ideas have been provided for the automated processing and analysis of medical care data.

[0004] At present, although some studies have attempted to apply machine learning models to the processing of medical care data, these methods are mostly limited to the analysis of a single information dimension, ignoring the correlation and complementarity between data of different dimensions. In practical applications, the health status of patients is often affected by a combination of multiple factors, and the analysis results of a single dimension are often difficult to fully reflect the true situation of the patient. Therefore, a method that can integrate medical care data across dimensions is needed to achieve a comprehensive and multi-level understanding of the patient's health status.

[0005] In addition, existing data processing methods often have problems such as high computational complexity and poor model generalization when processing large-scale medical care data. With the continuous increase and update of medical data, how to build an efficient and scalable data processing framework to adapt to the ever-changing medical care needs is also one of the current problems that need to be solved. Summary of the invention

[0006] The object of the present invention is to provide a method and system for digital management of medical care data. The embodiment of the present application is implemented as follows:

[0007] In a first aspect, an embodiment of the present application provides a method for digital management of medical care data, the method comprising: obtaining a patient information graph structures; each patient information graph structure has b patient information vertices; wherein a≥1, b≥1; the information dimensions corresponding to the a patient information graph structures are inconsistent; in the e-th information dimension, obtaining the vertex-related classification of the adjacent vertices of the s-th patient information vertex for the s-th patient information vertex; wherein 1≤e≤a, 1≤s≤b; based on the vertex-related classification, determining the adjacent-related characterization vector of the adjacent vertices of the s-th patient information vertex for the s-th patient information vertex; the vertex-related classification includes a first vertex-related classification and a second vertex-related classification; the first vertex-related classification is used to characterize the adjacent-related characterization vector as based on the The first adjacent characterization vectors corresponding to the adjacent vertices of the s-th patient information vertex in the a information dimensions are determined; the second vertex-related classification is used to characterize the adjacent related characterization vector as being based on the adjacent vertices of the s-th patient information vertex, and the first adjacent characterization vector corresponding to the adjacent vertices in the e-th information dimension is determined; in the e-th information dimension, the proposed enhanced patient characterization vector of the s-th patient information vertex is determined based on the adjacent vertices of the s-th patient information vertex for the adjacent related characterization vector of the s-th patient information vertex; based on the proposed enhanced patient characterization vector, a dimensional patient characterization vector of the s-th patient information vertex in the e-th information dimension is generated, and the dimensional patient characterization vector is used to integrate and construct a vertex patient characterization vector that can characterize the s-th patient information vertex.

[0008] In a second aspect, an embodiment of the present application provides a computer system, comprising: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method described above is implemented.

[0009] The beneficial effects of the present application include at least: in the embodiment of the present application, a patient information graph structures can be obtained, each of which has b patient information vertices, and the information dimensions corresponding to the a patient information graph structures are inconsistent. For example, for a patient information vertex in an information dimension, in the e-th information dimension, the adjacent related characterization vector of the adjacent vertices of the s-th patient information vertex for the s-th patient information vertex can be obtained, and the adjacent related characterization vector is obtained based on the vertex-related classification of the adjacent vertices of the s-th patient information vertex for the s-th patient information vertex. If the vertex-related classification is the first vertex-related classification, the adjacent related characterization vector is determined based on the first adjacent characterization vectors corresponding to the adjacent vertices of the s-th patient information vertex in the a-information dimensions. In other words, the first vertex-related classification represents that the adjacent vertices of the corresponding s-th patient information vertex have a greater influence on the s-th patient information vertex, and the first adjacent characterization vectors of the adjacent vertices of the s-th patient information vertex in different information dimensions can be integrated to obtain the adjacent related characterization vector of the adjacent vertices of the s-th patient information vertex in the e-th information dimension. In this way, the multi-dimensional representation information enhancement under the guidance of the interaction weight is completed. In one information dimension, the multi-dimensional representation information integration is performed on the adjacent vertices with high correlation (such as the main vertex), and the adjacent vertices of the s-th patient information vertex are targeted for the s-th patient information vertex. The dimensional patient representation vector of the s-th patient information vertex in the e-th information dimension is obtained. Among them, 1≤s≤b, that is, the s-th patient information vertex is any patient information vertex among the b patient information vertices. Based on the above method, the dimensional patient representation vectors of the b patient information vertices in the a information dimensions can be obtained. According to the correlation coefficient between the patient information vertices and the patient information vertices, the adjacent vertices of the patient information vertices are targeted for multi-dimensional representation information integration, so that the representation information of the patient information vertices in the single information dimension is richer. At the same time, the repeated representation information that may be caused in the process is reduced, thereby increasing the accuracy of obtaining multiple patient representation vectors. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in describing the embodiments of the present application are briefly introduced below.

[0011] Figure 1 It is a flow chart of a method for digital management of medical care data provided in an embodiment of the present application.

[0012] Figure 2 It is a schematic diagram of the composition of a computer system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0013] The execution subject of the medical care data digital management method in the embodiment of the present application is a computer system, including but not limited to a server, a personal computer, a laptop, a tablet computer, a smart phone, etc. Figure 1 As shown, the method includes:

[0014] Step S100: obtaining a patient information graph structures; each patient information graph structure has b patient information vertices; wherein a≥1, b≥1; and the information dimensions corresponding to the a patient information graph structures are inconsistent.

[0015] In step S100 of the present embodiment, the computer system retrieves a patient information graph structures of a specific patient from the hospital database. Here, "a" is an integer greater than or equal to 1, representing the diversity of patient data. For example, for a patient with long-term diabetes, the system can retrieve 3 (i.e., a=3) different patient information graph structures, namely:

[0016] Physiological indicator graph structure: This graph structure contains various physiological indicator data of patients, such as blood sugar level, blood pressure, heart rate, etc. Each physiological indicator exists as a patient information vertex, and the vertices may be connected by edges to indicate the mutual influence or temporal sequence between these indicators. For example, the blood sugar vertex may point to the blood pressure vertex through an edge, indicating that changes in blood sugar levels can affect blood pressure.

[0017] Medication record graph structure: This graph structure records the patient's medication history, including drug name, dosage, medication time, etc. Each drug or each medication record is a patient information vertex, and the edges between vertices may represent interactions between drugs or the order of medication.

[0018] Image examination graph structure: This graph structure is built based on the patient's medical imaging data, such as X-rays, MRI images, etc. Although these raw data are not directly stored in the form of a graph structure, the computer system can extract key features (such as lesion location, size, morphology, etc.) through image recognition technology and use these features as patient information vertices. The edges between vertices may represent changes in lesions at different examination time points.

[0019] The b patient information vertices (i.e., graph nodes) in each patient information graph structure represent various entities related to medical care, which together constitute a comprehensive description of the patient's health status. For example, in the physiological indicator graph structure, b may be equal to the number of physiological indicators currently monitored by the patient, such as blood sugar, blood pressure, heart rate, etc.; in the medication record graph structure, b may correspond to the different types of drugs or the number of medication records that the patient has taken.

[0020] Each patient information graph structure corresponds to a different information dimension (that is, different data modality), and together they provide a multi-angle view of the patient's health status. As mentioned earlier, although these graph structures are for the same patient, their data sources, presentation forms, and analysis purposes are different. By integrating this information from different dimensions, the computer system can more comprehensively understand the patient's health status and provide more accurate medical care recommendations to the patient.

[0021] Step S200: In the e-th information dimension, obtain the vertex-related classification of the adjacent vertices of the s-th patient information vertex for the s-th patient information vertex; wherein 1≤e≤a, 1≤s≤b.

[0022] In the embodiment of the present application, it is assumed that the computer system is processing the medical data of a heart patient, and these data have been organized into multiple patient information graph structures according to different information dimensions (modalities). The patient information vertices in each graph structure represent various entities related to medical care, such as electrocardiogram results, blood test results, medication records, etc. Now, the goal of the computer system is to analyze the relationship between these vertices in order to better understand and predict the patient's health status.

[0023] Specifically in step S200, it is assumed that the focus is on the e-th information dimension, that is, the electrocardiogram result (for the sake of simplicity, it is assumed that e=1), and the s-th patient information vertex is currently being analyzed, which represents the specific result of a certain electrocardiogram examination of the patient (for the sake of convenience, it is assumed that s=5, representing the fifth electrocardiogram result). In this step, the computer system identifies and analyzes the vertex-related classification of the adjacent vertices of the fifth electrocardiogram result (that is, other vertices directly related to the electrocardiogram result). The "vertex-related classification" here actually refers to the type of association or the strength of association between the adjacent vertices and the fifth electrocardiogram result. In actual operation, the computer system can use a series of algorithms or models to help determine these association types. For example, it can use graph embedding-based algorithms (such as DeepWalk, Node2Vec, etc.) to learn low-dimensional representations of vertices in the graph, which can capture the structural similarity and semantic association between vertices. Based on these low-dimensional representations, the system can then calculate the similarity or distance between neighboring vertices and the 5th EKG result vertex, thereby inferring the type of association between them.

[0024] Specifically, if an adjacent vertex (such as the vertex representing the patient's most recent blood test result) is close to the vertex of the fifth electrocardiogram result in the calculated low-dimensional space, the system can classify their association type as "strongly correlated", which means that there may be a direct mutual influence between the two test results or they may point to a certain health condition. On the contrary, if an adjacent vertex (such as the vertex representing the patient's medication record one year ago) is far away from the vertex of the fifth electrocardiogram result in the calculated low-dimensional space, the system can classify their association type as "weakly correlated" or "uncorrelated".

[0025] Step S300: Determine the adjacent related characterization vectors of the s-th patient information vertex for the adjacent vertices of the s-th patient information vertex based on vertex-related classification; the vertex-related classification includes a first vertex-related classification and a second vertex-related classification; the first vertex-related classification is used to characterize the adjacent related characterization vectors as the adjacent vertices based on the s-th patient information vertex, and the first adjacent characterization vectors corresponding to the adjacent vertices in a information dimensions are determined; the second vertex-related classification is used to characterize the adjacent related characterization vectors as the adjacent vertices based on the s-th patient information vertex, and the first adjacent characterization vectors corresponding to the adjacent vertices in the e-th information dimension are determined.

[0026] In the embodiment of the present application, it is assumed that the computer system is processing the medical data of a patient with a complex disease, and the data has been organized into multiple patient information graph structures, each of which represents a different dimension of the patient's health status (such as physiological indicators, imaging data, medication records, etc.). Now, the computer system focuses on the sth patient information vertex in the eth information dimension (for example, the heart rate measurement value representing a patient at a specific time point).

[0027] In step S200, the computer system has determined the vertex-related classifications of the adjacent vertices for the s-th patient information vertex. These classifications include the first vertex-related classification and the second vertex-related classification. The first classification indicates that the representation vector of the adjacent vertex has a strong correlation with the representation vector of the s-th vertex in all a information dimensions and needs to be considered comprehensively; while the second classification indicates that the representation vector of the adjacent vertex is significantly correlated with the s-th vertex mainly in the e-th information dimension.

[0028] Based on these classifications, the task of step S300 is to generate an adjacent correlation characterization vector (i.e., adjacent correlation feature) for the sth patient information vertex. This process involves how to integrate the information of adjacent vertices according to vertex correlation classification to form a richer and more accurate description of the sth vertex.

[0029] For the case of the first vertex-related classification, the computer system obtains the first adjacent characterization vectors corresponding to all adjacent vertices of the s-th patient information vertex in a information dimensions. These characterization vectors may contain rich contextual information, such as the results of adjacent examinations, relevant medical history, etc. Then, the computer system adopts a feature fusion strategy (such as splicing, weighted summation, attention mechanism, etc.) to integrate these characterization vectors from different dimensions. For example, if a weighted summation method is adopted, the computer system can assign different weights to each adjacent characterization vector according to the strength of association between the adjacent vertex and the s-th vertex (calculated by step S200), and then add them up to obtain the adjacent related characterization vector.

[0030] For the case of the second vertex-related classification, it is relatively simple. The computer system only needs to pay attention to the first adjacent representation vectors of the adjacent vertices in the e-th information dimension. These vectors are directly related to the representation vector of the s-th patient information vertex in the current dimension. The computer system can directly use these adjacent representation vectors, or obtain adjacent related representation vectors through some simple transformation (such as standardization).

[0031] In practice, computer systems can use machine learning models (such as graph attention networks, graph convolutional networks, etc.) to assist in this step. These models can automatically learn how to integrate the information of adjacent vertices based on the type and strength of the association between vertices, thereby generating more accurate and meaningful adjacent correlation representation vectors.

[0032] Finally, through the processing of step S300, the computer system generates one or more adjacent related representation vectors for the sth patient information vertex. These vectors not only capture the direct connection between the adjacent vertices and the sth vertex, but also integrate contextual information from multiple information dimensions, providing more comprehensive and in-depth data support for subsequent data analysis and health prediction.

[0033] Step S400: In the e-th information dimension, a proposed enhanced patient representation vector of the s-th patient information vertex is determined based on adjacent related representation vectors of adjacent vertices of the s-th patient information vertex with respect to the s-th patient information vertex.

[0034] In the embodiment of the present application, it is assumed that the computer system is processing the medical data of a diabetic patient, and the data has been organized into multiple patient information graph structures according to different information dimensions (such as blood sugar monitoring, medication records, living habits, etc.). Now, the computer system focuses on the sth patient information vertex in the eth information dimension (for example, blood sugar monitoring dimension), which represents a specific blood sugar measurement result of the patient.

[0035] In step S300, the computer system has determined the adjacent related characterization vectors of the adjacent vertices of the sth patient information vertex according to the vertex related classification. These vectors capture the characteristic information of other medical entities (such as the previous blood sugar measurement result, the diet record of the same period, etc.) associated with the sth vertex.

[0036] The goal of step S400 is to use these adjacent related representation vectors to strengthen the representation of the sth patient information vertex, thereby obtaining a proposed enhanced patient representation vector that can more comprehensively reflect the patient's current health status. In the processing of the computer system, this usually involves the following key steps:

[0037] First, the computer system analyzes the relationship between the adjacent related representation vectors and the initial representation vector of the sth patient information vertex. This relationship may be quantified by calculating the similarity, distance or correlation between them. For example, cosine similarity can be used as a metric to calculate the cosine value of the angle between each adjacent related representation vector and the initial representation vector to evaluate the similarity between them. Then, the computer system weights the adjacent related representation vectors according to these relationships. The basis for weighting can be their similarity, correlation strength or importance with the initial representation vector. The purpose of weighting is to highlight the adjacent vertex information that contributes more to the representation of the sth patient information vertex, while suppressing those that contribute less. For example, if the representation vector of a certain adjacent vertex is highly similar to the initial representation vector, it indicates that the adjacent vertex is highly consistent with the current vertex in terms of health status, so a higher weight should be given when weighting.

[0038] Finally, the computer system fuses the weighted adjacent related representation vector with the initial representation vector of the sth patient information vertex to generate a quasi-enhanced patient representation vector. The fusion method can be simple vector addition, weighted summation, or a more complex feature fusion algorithm. For example, a linear combination method can be used to add the weighted adjacent related representation vector and the initial representation vector according to a certain weight to obtain the final quasi-enhanced patient representation vector.

[0039] In practical applications, the computer system can also use machine learning models to assist in completing the processing of step S400. These models may include but are not limited to recurrent neural networks (RNNs), long short-term memory networks (LSTMs), graph neural networks (GNNs), etc., which can automatically learn how to strengthen the representation of the target vertex based on the information of the adjacent vertices. By training these models, the computer system can continuously improve the accuracy and efficiency of generating the patient representation vector to be strengthened.

[0040] Step S500: Based on the patient representation vector to be enhanced, a dimensional patient representation vector of the sth patient information vertex in the eth information dimension is generated, and the dimensional patient representation vector is used to integrate and construct a vertex patient representation vector that can represent the sth patient information vertex.

[0041] In the embodiment of the present application, step S500 aims to further refine and integrate the proposed enhanced patient representation vector obtained in step S400 to generate a dimensional patient representation vector that can comprehensively reflect the patient's health status in the e-th information dimension.

[0042] Assume that a computer system is processing the medical data of a chronic disease patient (such as a diabetic patient), and these data are organized into multiple patient information graph structures, each of which represents a different dimension of the patient's health status, such as blood sugar level, medication record, living habits, etc. Now, focus on the sth patient information vertex in the eth information dimension (for example, the blood sugar level dimension), which represents the specific value of the patient in a certain blood sugar test. In step S400, the computer system has generated a quasi-reinforced patient representation vector for the sth patient information vertex based on the information of adjacent vertices. This vector not only contains the information of the sth vertex itself, but also integrates the information of other medical entities associated with it, thereby providing a richer and more comprehensive description of the patient's health status.

[0043] The task of step S500 is to further process the proposed enhanced patient representation vector to generate a dimensional patient representation vector. This step may involve multiple sub-processes, including feature selection, feature transformation, dimensionality reduction, etc., to ensure that the generated dimensional patient representation vector retains key information and is convenient for subsequent processing and analysis. Specifically, the computer system can use one or more of the following methods to generate a dimensional patient representation vector:

[0044] Feature selection and transformation: First, the computer system can select the most representative features for the e-th information dimension from the patient representation vector to be enhanced based on domain knowledge or data characteristics. Then, these features are transformed, such as standardized and normalized, to eliminate the influence of different dimensions and scales.

[0045] Nonlinear mapping: In order to capture the nonlinear relationship in the patient representation vector to be enhanced, kernel methods (such as kernel techniques in support vector machines) or deep learning models (such as autoencoders and deep neural networks) can be used to map the original vector to a new high-dimensional or low-dimensional space. In this new space, the patient's health status may be expressed in a clearer and more separable form.

[0046] Graph embedding technology: If the patient information is represented in the form of a graph structure, graph embedding technology (such as Node2Vec, DeepWalk, etc.) can also be used to generate dimensional patient representation vectors. These technologies can learn low-dimensional representations of vertices in the graph while retaining the structural relationship and semantic information between vertices.

[0047] Ensemble learning methods: In order to further improve the accuracy and robustness of the dimensional patient representation vector, the computer system can also use ensemble learning methods (such as random forests, gradient boosting trees, etc.).

[0048] These methods combine the prediction results of multiple weak learners to generate the final strong learner prediction results, thereby effectively reducing overfitting and improving generalization ability.

[0049] Finally, after the above processing, the computer system generates a dimensional patient representation vector for the sth patient information vertex in the eth information dimension. This vector not only contains the patient's key health information in this dimension, but also has good interpretability and analyzability. With the generation of dimensional patient representation vectors of all relevant dimensions, the computer system can integrate them to construct a vertex patient representation vector that can fully represent the sth patient information vertex. This vertex patient representation vector will become the basis for subsequent health prediction, risk assessment, and formulation of personalized care plans.

[0050] As an implementation manner, step S200, in the e-th information dimension, obtaining vertex-related classifications of adjacent vertices of the s-th patient information vertex for the s-th patient information vertex, may specifically include:

[0051] Step S210: in the e-th information dimension, obtaining the first adjacent characterization vector of the adjacent vertices of the s-th patient information vertex, and obtaining the initial patient information vertex characterization vector of the s-th patient information vertex;

[0052] Step S220: In the e-th information dimension, according to the first adjacent characterization vector and the initial patient information vertex characterization vector, determining the initial characterization vector correlation coefficient of the adjacent vertices of the s-th patient information vertex with respect to the s-th patient information vertex;

[0053] Step S230: if the correlation coefficient of the initial characterization vector in the e-th information dimension is greater than a preset value, the first vertex correlation classification is performed as a vertex correlation classification of the adjacent vertices of the s-th patient information vertex for the s-th patient information vertex;

[0054] Step S240: If the correlation coefficient of the initial characterization vector in the e-th information dimension is not greater than a preset value, the second vertex correlation classification is performed as a vertex correlation classification of the adjacent vertices of the s-th patient information vertex for the s-th patient information vertex.

[0055] In an example scenario, suppose that the medical record data of a heart patient is being processed. The data is organized into one or more patient information graph structures, each of which represents a different information dimension, such as physiological indicators, medication records, imaging data, etc. Now, the computer system focuses on the e-th information dimension (for example, the electrocardiogram data dimension) and is ready to analyze the relationship between the s-th patient information vertex (for example, the vertex representing the result of a certain electrocardiogram examination) and its adjacent vertices in this dimension. The computer system first obtains the first adjacent representation vectors of all adjacent vertices of the s-th patient information vertex in the e-th information dimension. These vectors are feature representations of adjacent vertices in this dimension, and may contain rich diagnostic information, such as the time of adjacent electrocardiogram examinations, abnormal result indicators, heart rate variability, etc. For example, if the adjacent vertex represents the result of another electrocardiogram examination, then its first adjacent representation vector may contain feature values ​​such as ST segment changes, QRS complex morphology, QT interval, etc. of this examination. At the same time, the computer system obtains the initial representation vector of the s-th patient information vertex itself. This vector is the basic feature representation of vertex s in the ECG data dimension, and also contains the key information of the ECG examination. For example, this vector may include the specific values ​​and morphological characteristics of electrophysiological parameters such as heart rate, P wave, QRS wave, and T wave.

[0056] After obtaining the first neighbor characterization vectors of the neighboring vertices and the initial characterization vector of the s-th patient information vertex, the computer system calculates the correlation between them. This step is the basis for understanding how the neighboring vertices affect or relate to the s-th vertex.

[0057] Correlation calculation method: Correlation can be calculated by a variety of statistical measures, such as Pearson correlation coefficient, Spearman rank correlation coefficient or mutual information. In the context of healthcare data, Pearson correlation coefficient is often used because of its simplicity and wide applicability. The calculation formula of Pearson correlation coefficient is:

[0058]

[0059] Among them, xi and yi represent the observed values ​​of the corresponding elements in the two vectors, and are the means of the two vectors, and n is the length of the vector.

[0060] For example, suppose there are two vectors, one is the initial characterization vector of the sth patient information vertex (such as heart rate, ST segment deviation, etc.), and the other is the first adjacent characterization vector of a certain adjacent vertex (representing the result of another electrocardiogram examination). The computer system will use the above formula to calculate the Pearson correlation coefficient between the two vectors to quantify the strength and direction of the linear relationship between them.

[0061] The value of the correlation coefficient ranges from -1 to 1. Values ​​close to 1 or -1 indicate strong positive or negative correlation, while values ​​close to 0 indicate almost no linear relationship. The output of this step is the correlation coefficient of the initial representation vector between the neighboring vertices and the s-th patient information vertex.

[0062] After calculating the correlation coefficient of the initial characterization vector, the computer system compares it with a preset threshold to determine the type of association between the adjacent vertex and the sth patient information vertex (i.e., vertex correlation classification). The preset value is a critical value determined based on domain knowledge, data characteristics, or experiments, and is used to distinguish associations of different strengths. In the field of medical care, this value may be adjusted according to the specific application scenario and data set. For example, if a correlation coefficient greater than 0.7 is considered to indicate a strong correlation, then the preset value can be set to 0.7. If the calculated initial characterization vector correlation coefficient is greater than the preset value, the computer system believes that there is a strong association between the adjacent vertex and the sth patient information vertex, and therefore assigns the first vertex correlation classification (the characterization adjacent correlation characterization vector is determined based on the first adjacent characterization vector of all relevant adjacent vertices) to the pair of vertices. This means that in subsequent steps, the information of all strongly correlated adjacent vertices will be comprehensively considered to strengthen the characterization of the sth patient information vertex.

[0063] If the correlation coefficient of the initial characterization vector is not greater than a preset value, the computer system considers that the association between the adjacent vertex and the s-th patient information vertex is weak or does not exist directly.

[0064] In this case, the computer system assigns the second vertex correlation classification (characterizing the adjacent correlation characterization vector as determined based only on the first adjacent characterization vector of the adjacent vertices under the current information dimension) to the pair of vertices. This means that in subsequent steps, only the information of the adjacent vertices under the current information dimension will be considered, while other possible weakly related or irrelevant information will be ignored.

[0065] Suppose you are analyzing the ECG data graph structure of a heart disease patient. The sth patient information vertex represents the patient's most recent ECG result, while adjacent vertices may include the patient's previous ECG results, blood test results, medication records, etc.

[0066] In the e-th information dimension (electrocardiogram data), the computer system first obtains the initial characterization vector of the s-th vertex (such as heart rate, ST segment deviation, etc.) and the first adjacent characterization vector of the adjacent vertex. Next, the correlation coefficient of the initial characterization vector between the adjacent vertex and the s-th vertex is calculated. Assume that the correlation coefficient between a certain adjacent vertex (representing the patient's last electrocardiogram result) and the s-th vertex is as high as 0.8, which exceeds the preset value of 0.7, and is therefore judged to be strongly correlated. Based on this result, the computer system assigns the first vertex correlation classification to the pair of vertices, which means that in subsequent steps, the information of the adjacent vertex will be comprehensively considered to strengthen the characterization of the s-th patient information vertex.

[0067] On the contrary, if the correlation coefficient between another adjacent vertex (such as blood test results) and the sth vertex is only 0.3, which is lower than the preset value, then the pair of vertices will be assigned the second vertex correlation classification, and only the information of this adjacent vertex under the ECG data dimension will be considered in subsequent steps.

[0068] Through the above steps, the computer system can effectively distinguish the strength of association between vertices in the patient information graph structure, providing an important classification basis for subsequent data processing and analysis. This classification not only helps to reduce information redundancy and noise, but also improves the efficiency and accuracy of data processing, providing more reliable support for medical care decision-making.

[0069] As an implementation manner, the number of adjacent vertices of the s-th patient information vertex is X, X≥1; step S220, in the e-th information dimension, determining the correlation coefficient of the initial representation vector of the adjacent vertices of the s-th patient information vertex with respect to the s-th patient information vertex according to the first adjacent representation vector and the initial patient information vertex representation vector, may specifically include:

[0070] Step S221: In the e-th information dimension, a representation information integration operation is performed on the first adjacent representation vectors corresponding to the X adjacent vertices and the initial patient information vertex representation vector to obtain the initial vertex integration vector of the s-th patient information vertex;

[0071] Step S222: in the e-th information dimension, obtaining initial vertex integration subvectors between the X first adjacent representation vectors and the initial patient information vertex representation vector respectively;

[0072] Step S223: In the e-th information dimension, according to the ratio between the X initial vertex integration subvectors and the initial vertex integration vector, determine the initial representation vector correlation coefficients of the X adjacent vertices for the s-th patient information vertex.

[0073] The goal of step S220 is to calculate the correlation coefficient between the adjacent vertices of the s-th patient information vertex and its own initial representation vector to evaluate the similarity or correlation between them.

[0074] Suppose you are processing the medical records of a diabetic patient. These data are organized into multiple patient information graph structures, each of which represents a different medical information dimension, such as blood sugar levels, medication records, physical examination results, etc. Now, the computer system focuses on the e-th information dimension (for example, the blood sugar level dimension) and is ready to analyze the relationship between the s-th patient information vertex under this dimension (for example, representing a blood sugar measurement result at a specific time point) and its X adjacent vertices (other blood sugar measurement results or related medical events).

[0075] In step S221, the computer system integrates the first adjacent characterization vectors of the X adjacent vertices and the initial characterization vector of the sth patient information vertex. The characterization vector usually includes a series of values, which represent the key features or attributes of the vertex.

[0076] Each adjacent vertex has a corresponding first adjacent characterization vector, which captures the key information of the adjacent vertex. In the blood sugar level dimension, these vectors may include multiple features such as blood sugar concentration, measurement time, and diet before measurement. The initial characterization vector of the s-th patient information vertex also contains a series of eigenvalues, which describe the specific information of the vertex, such as a specific blood sugar measurement result. The specific method of integration depends on the characteristics of the data and the analysis requirements. In the field of medical care, a common method is to integrate information using methods such as weighted average or principal component analysis. However, due to the complexity and diversity of medical data, the integration operation may require more sophisticated processing. Assume that a simple weighted average method is used for integration. First, a weight needs to be assigned to the first adjacent characterization vector of each adjacent vertex. This weight can be determined based on the time distance between the adjacent vertex and the s-th vertex, the severity of the disease, or other relevant factors. Then, these weighted vectors are added together to obtain the initial vertex integration vector.

[0077] For example, suppose there are 3 adjacent vertices (X=3), and the first adjacent characterization vector of each vertex is The corresponding weights are w1, w2, and w3, and their sum is equal to 1. Then the initial vertex integration vector It can be calculated as:

[0078]

[0079] This integration vector Represents the comprehensive reflection of all adjacent vertex information.

[0080] After obtaining the initial vertex integration vector, the computer system further analyzes the relationship between the first adjacent characterization vector of each adjacent vertex and the initial patient information vertex characterization vector. This step is achieved by calculating the sub-vectors between them, which can be understood here as a certain form of interaction result between the adjacent vertex characterization vector and the initial characterization vector. One possible method is to calculate the difference vector, dot product, or more complex similarity measure (such as cosine similarity) between the two vectors. However, in the context of step S222, it is more likely to be looking for a way to reflect the influence of adjacent vertices on the initial vertex. Since directly calculating sub-vectors is not a standard mathematical operation, an alternative solution can be adopted: calculate a certain "difference" or "similarity" indicator between the first adjacent characterization vector of each adjacent vertex and the initial patient information vertex characterization vector, and use this indicator as a proxy for the "sub-vector". Here, cosine similarity can be selected as the similarity indicator.

[0081] Cosine similarity calculation formula:

[0082]

[0083] Among them, a and b represent two vectors, represents their dot product, denote their Euclidean lengths respectively.

[0084] For each adjacent vertex, the cosine similarity between its first adjacent representation vector and the initial patient information vertex representation vector can be calculated, and this similarity value can be used as a proxy for the initial vertex integration subvector (in this specific scenario, it is actually a scalar value).

[0085] For example, suppose the initial patient information vertex representation vector is The first adjacent characterization vector of the first adjacent vertex is Then the cosine similarity between them can be calculated as:

[0086]

[0087] Similarly, cosine similarity can be calculated for other adjacent vertices to obtain a series of similarity values, which can be regarded as a "scalarized" representation of the initial vertex integration sub-vector.

[0088] In step S223, the computer system determines the initial characterization vector correlation coefficient of each adjacent vertex for the sth patient information vertex based on the relationship between the initial vertex integration subvector (here, the cosine similarity value) obtained in step S222 and the initial vertex integration vector. However, since what is actually obtained in step S222 is a series of similarity values ​​rather than a true vector, the explanation of this step needs to be adjusted to conform to the actual situation.

[0089] In this scenario, the similarity value of each adjacent vertex can be regarded as a measure of its correlation with the initial patient information vertex representation vector. However, in order to obtain a more standardized correlation coefficient, these similarity values ​​may need to be further processed. A simple method is to directly use the similarity value as the correlation coefficient.

[0090] If the cosine similarity value calculated in step S222 is directly used as a proxy for the correlation coefficient, then for the first adjacent vertex, its initial characterization vector "correlation coefficient" for the s-th patient information vertex is sim1. Similarly, the "correlation coefficients" of other adjacent vertices can be obtained. If you want these values ​​to more directly reflect the "contribution" of the adjacent vertices to the initial vertex, you can consider normalizing the similarity values. For example, you can use the softmax function to convert the similarity value into a probability distribution form:

[0091]

[0092] Among them, corr i represents the normalized correlation coefficient (or contribution) of the i-th adjacent vertex, and simi is the similarity value calculated in step S222.

[0093] By implementing the above steps S221 to S223, the computer system can accurately calculate the correlation between the adjacent vertices of the sth patient information vertex and its own initial representation vector. This result not only helps to understand the complex relationship network between vertices, but also provides a solid foundation for subsequent data analysis and health prediction. In practical applications, these steps can be adjusted and optimized according to specific data characteristics and analysis requirements.

[0094] In one implementation, step S300, determining the adjacent related characterization vectors of the adjacent vertices of the s-th patient information vertex for the s-th patient information vertex according to the vertex related classification, may specifically include:

[0095] Step S310: If the vertex-related classification is the first vertex-related classification, the first adjacent representation vectors corresponding to the adjacent vertices of the s-th patient information vertex in a information dimensions are obtained, the representation information of the a first adjacent representation vectors is integrated to obtain an adjacent integrated representation vector, and the adjacent integrated representation vector is subjected to unified modal mapping to obtain an adjacent related representation vector of the adjacent vertices of the s-th patient information vertex for the s-th patient information vertex;

[0096] Step S320: If the vertex-related classification is the second vertex-related classification, then obtain the first adjacent representation vector corresponding to the adjacent vertices of the s-th patient information vertex in the e-th information dimension, and use the first adjacent representation vector in the e-th information dimension as the adjacent related representation vector of the adjacent vertices of the s-th patient information vertex for the s-th patient information vertex.

[0097] In the embodiment of the present application, step S300 constructs a deep correlation representation between patient information vertices according to vertex correlation classification. Specifically, the goal of step S300 is to generate an adjacent correlation representation vector of adjacent vertices for the sth patient information vertex according to the vertex correlation classification determined in step S200. This vector will be used for subsequent data processing and analysis to more comprehensively reflect the health status of the patient.

[0098] When the vertex correlation classification is the first vertex correlation classification, it means that the adjacent vertices of the sth patient information vertex have a significant impact on the sth vertex in multiple information dimensions (i.e., a dimensions). Therefore, the computer system comprehensively considers the information of these adjacent vertices in all relevant dimensions to generate the adjacent correlation representation vector.

[0099] First, the computer system retrieves the first adjacent representation vectors corresponding to all adjacent vertices of the sth patient information vertex in a information dimensions from the database. These vectors may contain rich medical information, such as physiological indicators, medication records, imaging data, etc. Next, the system needs to integrate these first adjacent representation vectors from different dimensions. There are many ways to integrate, depending on the characteristics of the data and the analysis requirements. The following are several possible integration methods and their specific implementations:

[0100] Feature averaging method: The system can calculate the average value of all first-neighbor characterization vectors and use it as the neighboring integrated characterization vector.

[0101]

[0102] Represents the first adjacent representation vector on the i-th information dimension.

[0103] Concatenation method: The system can also directly concatenate all first adjacent representation vectors into a longer vector. This method retains information in all dimensions, but may cause the vector dimension to be too high, increasing the complexity of subsequent processing.

[0104] Feature weighting method: The system can assign different weights to each first adjacent characterization vector according to the strength of association between the adjacent vertices and the sth vertex (such as the correlation coefficient calculated in step S220), and then perform weighted summation. This method is more flexible and can perform differentiated processing according to the importance of different dimensions.

[0105] Attention mechanism: In more advanced applications, the system can use the attention mechanism to dynamically adjust the weights of information of different dimensions. The attention mechanism can learn which dimensions of information are more important for generating adjacent related representation vectors, thereby achieving more accurate feature extraction.

[0106] Unified modal mapping: The integrated vector may contain multiple different types of data (such as numerical, text, image, etc.), and the representation and scale of these data may be different. Therefore, the system needs to perform unified modal mapping to convert the integrated vector into a unified representation space. This usually involves operations such as feature scaling and encoding conversion.

[0107] Feature scaling: Scale feature values ​​to the same scale using methods such as max-min normalization or Z-score normalization.

[0108] Encoding conversion: For non-numeric data (such as text, images), it is necessary to convert them into numerical vectors. This can be achieved through methods such as word embedding and convolutional neural network (CNN).

[0109] Generate adjacent related representation vectors: After the above processing, the system obtains a unified adjacent integrated representation vector. This vector integrates the information of all adjacent vertices in multiple information dimensions and can more comprehensively reflect the health status of the sth patient information vertex.

[0110] When the vertex correlation classification is the second vertex correlation classification, it means that the neighboring vertices of the sth patient information vertex have a significant impact on the sth vertex mainly in the current information dimension (i.e., the eth dimension). Therefore, the computer system only needs to consider the information in this dimension to generate the neighbor correlation representation vector.

[0111] The system can directly retrieve from the database the first adjacent representation vector corresponding to the adjacent vertices of the sth patient information vertex in the eth information dimension. Alternatively, since the influence of adjacent vertices is mainly concentrated in the current dimension, the system can directly use the first adjacent representation vector in this dimension as the adjacent related representation vector. This is simple and efficient, and can accurately reflect the influence of adjacent vertices on the current vertex. For example, suppose that the medical data of a heart patient is being processed. These data are organized into multiple patient information graph structures, each of which represents a different medical information dimension (such as electrocardiogram, blood test, imaging data, etc.). Now, the computer system focuses on a patient information vertex under the electrocardiogram dimension (representing a specific electrocardiogram result).

[0112] First vertex related classification: If the system determines that the adjacent vertices of the vertex have significant influence in multiple dimensions (such as electrocardiogram, blood test, etc.), step S310 is executed. The system first obtains the first adjacent representation vectors of these adjacent vertices in each dimension (such as electrocardiogram feature vector, blood test index vector, etc.), and then integrates them using methods such as feature splicing or attention mechanism to obtain adjacent integrated representation vectors. Finally, unified modal mapping is performed through operations such as feature scaling and encoding conversion to generate adjacent related representation vectors.

[0113] Second vertex related classification: If the system determines that the influence of the adjacent vertices is mainly concentrated in the electrocardiogram dimension, step S320 is executed. The system directly obtains the first adjacent characterization vector of the adjacent vertex in the electrocardiogram dimension and uses it as the adjacent related characterization vector.

[0114] As an implementation manner, step S400, in the e-th information dimension, determining the proposed enhanced patient representation vector of the s-th patient information vertex according to the adjacent related representation vectors of the adjacent vertices of the s-th patient information vertex for the s-th patient information vertex, may specifically include:

[0115] Step S410: in the e-th information dimension, according to the adjacent related characterization vectors of the adjacent vertices of the s-th patient information vertex with respect to the s-th patient information vertex and the initial patient information vertex characterization vector of the s-th patient information vertex, determine the correlation coefficient of the 0th level characterization vector of the adjacent vertices of the s-th patient information vertex with respect to the s-th patient information vertex;

[0116] Step S420: According to the correlation coefficient of the 0th level characterization vector, the adjacent vertices of the sth patient information vertex are eccentrically adjusted with respect to the adjacent related characterization vectors of the sth patient information vertex to obtain the quasi-enhanced patient characterization vector of the sth patient information vertex.

[0117] In step S400 of the embodiment of the present application, the goal of the computer system is to generate an enhanced and more comprehensive patient representation vector, i.e., a proposed enhanced patient representation vector, based on the adjacent related representation vectors provided by the adjacent vertices of the sth patient information vertex and the initial representation vector of the patient itself. This process is divided into two main steps: calculating the correlation coefficient of the 0th level representation vector and eccentrically adjusting the adjacent related representation vectors.

[0118] In the embodiment of the present application, the correlation coefficient of the level 0 representation vector reflects the similarity or correlation between the patient information vertex (such as a specific medical examination result) and its adjacent vertices (such as other related medical records or examination results) at the feature level. This step is the basis for understanding the interaction between vertices and constructing a reinforced representation vector.

[0119] First, the computer system obtains two key information: one is the initial patient information vertex representation vector of the sth patient information vertex, which represents the basic health status of the patient under this information dimension; the other is the adjacent related representation vector of the adjacent vertices of the sth patient information vertex for this vertex, which integrates the information of the adjacent vertices and provides additional context or association features. Next, the system needs to calculate the correlation coefficient between these two sets of vectors, but it is specifically pointed out here that "level 0" means that this is a calculation performed on the original feature layer or the lowest level representation, and no advanced feature transformation or abstraction has been introduced.

[0120] For example, suppose the sth patient information vertex represents the electrocardiogram result of a patient, and its initial patient information vertex representation vector contains direct measurements such as heart rate and ST segment deviation. At the same time, its adjacent vertices may include electrocardiogram results at other time points, blood test results, and the patient's medication records, etc. The information of these adjacent vertices is integrated into adjacent related representation vectors.

[0121] To calculate the correlation coefficient of the level 0 representation vector, the system can use the Pearson correlation coefficient, which is an indicator of the degree of linear correlation between two variables. For vectors, it can be obtained by calculating the ratio of the covariance between their corresponding elements and the product of their respective standard deviations.

[0122] Substituting the initial characterization vector and the adjacent related characterization vector of the sth patient information vertex into the above formula, we can get the correlation coefficient of the level 0 characterization vector between them. This value will be used in the subsequent steps as the basis for adjusting the weight of the adjacent related characterization vector.

[0123] After obtaining the correlation coefficient of the 0th level representation vector, the computer system uses this coefficient to adjust the weights of the adjacent related representation vectors to generate a proposed enhanced patient representation vector. This process is similar to "weighted voting" for the information of adjacent vertices, where the weight is determined by the correlation coefficient to ensure that the vertex information that is more relevant to the patient's initial representation has a larger proportion in the final representation.

[0124] First, the system determines the weight of each adjacent vertex based on the correlation coefficient of the 0th-level representation vector. Generally speaking, the higher the correlation coefficient, the stronger the correlation between the adjacent vertex and the initial patient representation, and therefore a higher weight should be given. The system then uses these weights to perform a weighted sum (or other forms of weighted fusion) of the elements in the adjacent related representation vectors to generate a vector for the proposed enhanced patient representation. This process is similar to feature weighting or attention mechanisms in machine learning, which can highlight the features that contribute most to patient representation.

[0125] For example, let's continue with the example of electrocardiogram results. Assume that the system has calculated that the correlation coefficient of the 0th level representation vector between the sth patient information vertex and one of its adjacent vertices (such as the last electrocardiogram result) is 0.8 (indicating a high correlation). When generating the proposed enhanced patient representation vector, the system will give a higher weight to the information of this adjacent vertex.

[0126] If the adjacent correlation representation vector is a multidimensional vector containing multiple features (such as heart rate changes, ST segment offset changes, etc.), the system will perform a weighted summation of these features according to the weights. For example, if a feature occupies an important position in the adjacent correlation representation vector and has a high correlation coefficient, the contribution of this feature will be greater in the weighted summation process.

[0127] The final generated quasi-enhanced patient representation vector not only contains the initial representation information of the patient itself, but also integrates the weighted adjacent vertex information, thereby providing a more comprehensive and accurate description of the patient's health status.

[0128] Through the implementation of step S400, the computer system can make full use of the relationship between vertices in the patient information graph structure, calculate the correlation coefficient of the representation vector and adjust the eccentricity of the adjacent related representation vectors to generate a simulated enhanced patient representation vector. This process not only enhances the richness and accuracy of the patient representation, but also provides strong support for subsequent data analysis and health prediction.

[0129] In one embodiment, step S500, based on the patient representation vector to be enhanced, generates a dimensional patient representation vector of the sth patient information vertex in the eth information dimension, which may specifically include:

[0130] Step S510: in the e-th information dimension, obtain the u-th level adjacent vertex representation vector corresponding to the adjacent vertices of the s-th patient information vertex, and obtain the u-th level patient representation vector corresponding to the s-th patient information vertex, wherein u≥1, and the quasi-enhanced patient representation vector of the s-th patient information vertex represents the first-level patient representation vector corresponding to the s-th patient information vertex when u=1;

[0131] Step S520: In the e-th information dimension, based on the u-th level adjacent vertex representation vectors corresponding to the adjacent vertices of the s-th patient information vertex, the u-th level patient representation vector corresponding to the s-th patient information vertex is enhanced to obtain the u+1-th level patient representation vector of the s-th patient information vertex;

[0132] Step S530: If u+1 is equal to the preset number of enhancements, the u+1-th level patient representation vector is determined as the dimensional patient representation vector of the s-th patient information vertex in the e-th information dimension.

[0133] In the embodiment of the present application, step S500 generates a dimensional patient representation vector that can comprehensively reflect the patient's health status in the e-th information dimension through multi-level feature enhancement and integration.

[0134] Before executing step S500, the computer system has generated a proposed enhanced patient characterization vector of the sth patient information vertex through step S400, which serves as the initial patient characterization (i.e., the first-level patient characterization vector), providing a starting point for subsequent feature enhancement. In step S510, the system needs to further obtain the u-th level adjacent vertex characterization vectors of other vertices (i.e., adjacent vertices) associated with the sth patient information vertex, as well as the u-th level patient characterization vector of the sth patient information vertex itself. For example, suppose that the medical data of a diabetic patient is being processed, where the e-th information dimension represents the patient's blood glucose monitoring data. Under this dimension, the sth patient information vertex may represent a specific blood glucose measurement result of the patient. In order to comprehensively evaluate the patient's blood glucose management, it is necessary to consider multiple blood glucose measurement results and their changing trends, which can all be regarded as characterization vectors of adjacent vertices. For example, adjacent vertex characterization vector: Assume that the patient's 5 blood glucose measurement results in the past week are considered as adjacent vertices, and each measurement result is represented as a vector containing features such as timestamp, blood glucose value, and whether it is measured before / after a meal. These vectors are the u-th level adjacent vertex representation vectors of adjacent vertices (here, since the processing has just begun, it can be considered that u=1, that is, the original measurement data is used).

[0135] Patient characterization vector: Similarly, the first-level patient characterization vector of the s-th patient information vertex (i.e., the proposed enhanced patient characterization vector) has been generated through step S400, which may be a comprehensive vector containing more context information and enhanced features.

[0136] After obtaining the u-th level representation vector, the core task of step S520 is to strengthen (feature enhance) the u-th level patient representation vector of the s-th patient information vertex to generate a higher-level u+1-th level patient representation vector. This process usually involves multiple technologies such as feature fusion, feature transformation, and feature selection, aiming to extract more discriminative and representative features from the original representation.

[0137] For example, for feature fusion: the system can fuse the u-th patient representation vector of the s-th patient information vertex with the u-th adjacent vertex representation vector of its adjacent vertex. The fusion method can be simple concatenation, weighted summation, or more complex graph neural network (GNN) or attention mechanism. For example, using the Graph Attention Network (GAT), the importance of adjacent vertices to the current vertex can be dynamically calculated, and the representation vectors of adjacent vertices can be weighted fused according to these importance weights.

[0138] For feature transformation: Based on the fusion, the system may also need to further transform the fused vector to extract higher-level features. This can be achieved through models such as multi-layer perceptron (MLP), convolutional neural network (CNN) or recurrent neural network (RNN). For example, using MLP can perform nonlinear transformation on the fused vector, thereby learning more abstract and complex feature representations.

[0139] For feature selection: In some cases, in order to avoid overfitting and improve the generalization ability of the model, the system may also need to select the transformed features. This can be achieved through methods such as regularization, feature importance evaluation or sparse coding.

[0140] Let's continue to take the blood glucose monitoring data of diabetic patients as an example. In step S520, the system can fuse the patient's specific blood glucose measurement result (i.e., the u-th level patient characterization vector of the s-th patient information vertex) with other blood glucose measurement results in the past week (i.e., the u-th level adjacent vertex characterization vector of adjacent vertices). The fusion method may be to use the GAT model to dynamically calculate the importance weights of adjacent vertices based on factors such as the time interval between each measurement result and the trend of blood glucose value changes, and normalize these weights to ensure that their sum is 1. Then, the characterization vectors of adjacent vertices are weighted and summed according to these weights to obtain a fused vector. Next, the system can use an MLP to perform a nonlinear transformation on the fused vector to extract higher-level features. Finally, the system can remove some redundant or unimportant features through feature selection methods to obtain a more compact and effective u+1-th level patient characterization vector.

[0141] After performing step S520 for feature enhancement multiple times, the system needs to determine whether the current number of enhancements (i.e., the value of u+1) has reached the preset number of enhancements. If the preset number has been reached, the currently generated patient representation vector is determined as the dimensional patient representation vector of the sth patient information vertex in the eth information dimension; if the preset number has not been reached, step S520 is continued to be performed for the next round of feature enhancement.

[0142] Setting the preset number of enhancements: The setting of the preset number of enhancements usually depends on the characteristics of the data, the analysis objectives, and the limitations of computing resources. In practical applications, the optimal number of enhancements can be determined by methods such as cross-validation. Generally speaking, as the number of enhancements increases, the discrimination and representativeness of the patient representation vector will gradually increase, but at the same time, the complexity of the calculation and the risk of overfitting will also increase. Therefore, it is necessary to find a balance between the enhancement effect and the computational cost.

[0143] For example, in the diabetes management case, it is assumed that after experimental verification, it is found that setting the number of enhancements to 3 times can achieve better enhancement effects while ensuring computational efficiency. Therefore, in step S530, the system checks the current number of enhancements after each execution of step S520. If it is the first enhancement (i.e., u+1=2), continue to execute step S520 for the second enhancement; if it is the second enhancement (i.e., u+1=3), the generated patient characterization vector is determined as a dimensional patient characterization vector; if it is the third or more enhancements, stop the enhancement process and output the final dimensional patient characterization vector.

[0144] By implementing the above step S500, the computer system can generate a dimensional patient representation vector that fully reflects the patient's health status in the e-th information dimension. This vector not only contains the initial representation information of the patient itself, but also integrates the context information and high-level features in multiple feature enhancement processes. This process not only deepens the computer system's understanding of individual differences in patients, but also provides strong data support for the subsequent formulation of precision medicine, health prediction and personalized care strategies.

[0145] As an implementation manner, step S520, in the e-th information dimension, based on the u-th level adjacent vertex representation vectors corresponding to the adjacent vertices of the s-th patient information vertex, performs a representation vector enhancement on the u-th level patient representation vector corresponding to the s-th patient information vertex to obtain the u+1-th level patient representation vector of the s-th patient information vertex, which may specifically include:

[0146] Step S521: in the e-th information dimension, obtain the u-th level representation vector correlation coefficient of the adjacent vertices of the s-th patient information vertex with respect to the s-th patient information vertex, and perform eccentric adjustment on the u-th level adjacent vertex representation vector according to the u-th level representation vector correlation coefficient to obtain the u-th level enhanced representation vector;

[0147] Step S522: In the e-th information dimension, the u-th level enhanced representation vector is used to perform representation vector enhancement on the u-th level patient representation vector corresponding to the s-th patient information vertex to obtain the u+1-th level patient representation vector of the s-th patient information vertex.

[0148] In the field of medical care, the health status of patients is often affected by many factors, including but not limited to physiological indicators, living habits, genetic background, environmental factors, etc. Therefore, when constructing the patient representation vector, it is necessary to comprehensively consider these factors and use feature enhancement technology to highlight important features and suppress irrelevant or noise information. Step S520 is designed to achieve this goal.

[0149] In step S521, the primary task of the computer system is to calculate the correlation coefficients of the adjacent vertices of the sth patient information vertex with respect to the uth level characterization vector of the vertex. These correlation coefficients reflect the similarity or relevance of the adjacent vertices with the central patient information vertex in the uth level feature space, and are an important basis for subsequent weighted processing.

[0150] For example, suppose that medical data of a heart patient is being processed, where the e-th information dimension represents the electrocardiogram result. The s-th patient information vertex represents a specific electrocardiogram result of the patient, and the adjacent vertices may include the patient's previous electrocardiogram record, blood test results, medication record, etc. Prior to step S520, the u-th level patient representation vector of the s-th patient information vertex and the u-th level adjacent vertex representation vector of the adjacent vertex have been generated through the previous steps.

[0151] Correlation coefficient calculation, in order to calculate the correlation coefficient of the u-th level representation vector, the computer system can use a variety of methods, such as Pearson correlation coefficient, Spearman rank correlation coefficient or cosine similarity, etc. The above content has been introduced and will not be repeated here.

[0152] For each adjacent vertex, the computer system calculates the cosine similarity between its u-th level adjacent vertex representation vector and the u-th level patient representation vector of the s-th patient information vertex to obtain the u-th level representation vector correlation coefficient.

[0153] After obtaining the correlation coefficients, the computer system performs eccentric adjustment (i.e., weighting) on ​​the representation vectors of adjacent vertices according to these coefficients. There are many ways to weight, which can be simple linear weighting or more complex nonlinear transformation. Here, linear weighting is used for illustration.

[0154] Let the u-th level adjacent vertex representation vector of the i-th adjacent vertex be Its correlation coefficient with the u-th level representation vector of the s-th patient information vertex is wi, then the weighted u-th level enhanced representation vector It can be calculated as:

[0155]

[0156] Where X is the number of adjacent vertices.

[0157] In step S522, the computer system will use the u-th level enhanced representation vector obtained in step S521 to enhance the u-th level patient representation vector of the s-th patient information vertex, thereby generating the u+1-th level patient representation vector. This process usually involves multiple technologies such as feature fusion and feature transformation.

[0158] The most direct way to fuse features is to concatenate or weighted sum the u-th level reinforcement representation vector and the u-th level patient representation vector. In the field of medical care, due to the high dimensionality and complexity of the data, simple concatenation or summation may not fully utilize the information in the reinforcement representation vector. Therefore, it is more common to use more complex fusion methods, such as graph neural networks (GNNs), attention mechanisms, etc.

[0159] Taking the attention mechanism as an example, the computer system can assign an attention weight to the enhanced representation vector of each adjacent vertex, which reflects the importance of the vertex to the current patient information vertex. Then, the enhanced representation vectors are weighted summed according to these weights to obtain the fused representation vector. This method can dynamically adjust the contribution of different adjacent vertices to the current patient representation, thereby improving the accuracy and robustness of the representation.

[0160] In order to simplify the description, a simple weighted summation method is still used for feature fusion. Suppose the u-th level patient representation vector of the s-th patient information vertex is Then the u+1th level patient representation vector It can be calculated as:

[0161]

[0162] Among them, α is a hyperparameter used to balance the contribution between the original patient representation vector and the enhanced representation vector. This parameter can be optimized through cross-validation and other methods.

[0163] After feature fusion, the computer system needs to further transform the fused representation vector to extract higher-level feature representations. This transformation can be achieved through models such as multi-layer perceptron (MLP), convolutional neural network (CNN) or recurrent neural network (RNN). Here, in order to simplify the explanation, it is assumed that no additional feature transformation steps are performed, and the fused representation vector is directly used as the u+1-th level patient representation vector.

[0164] Continuing with the medical data of heart disease patients as an example, it is assumed that the u-th level representation vector correlation coefficient between each adjacent vertex and the s-th patient information vertex has been calculated through step S521, and the representation vectors of the adjacent vertices have been eccentrically adjusted and weighted summed to obtain the u-th level enhanced representation vector.

[0165] In step S522, this enhanced characterization vector is weightedly summed with the u-th level patient characterization vector of the s-th patient information vertex (assuming α = 0.5) to obtain a fused characterization vector. This vector not only contains the initial information of the patient itself (reflected by the u-th level patient characterization vector), but also integrates the enhanced information of adjacent vertices (reflected by the u-th level enhanced characterization vector), thereby providing a more comprehensive and accurate description of the patient's health status. Finally, after multiple iterations of step S520 (i.e., multiple feature enhancements), a sequence of patient characterization vectors with gradually deepening levels and gradually richer features can be obtained. When the number of enhancements reaches a preset value, the final patient characterization vector is output as a dimensional patient characterization vector for subsequent tasks such as health prediction, disease diagnosis, and formulation of personalized nursing strategies.

[0166] In one embodiment, the method may further include:

[0167] Step S500A: Obtain the dimensional patient representation vectors corresponding to the s-th patient information vertex in a information dimensions, perform a representation information integration operation on the a-dimensional patient representation vectors corresponding to the s-th patient information vertex, and obtain the vertex patient representation vector of the s-th patient information vertex.

[0168] In order to more comprehensively understand and analyze the patient's overall health status, the computer system not only needs to pay attention to the patient's performance in each individual information dimension (such as physiological indicators, medication records, imaging data, etc.), but also needs to integrate the information in these different dimensions to form a unified patient representation vector.

[0169] Suppose you are processing the medical data of a patient with a chronic disease (such as a diabetic patient). The patient's medical records are scattered and stored in multiple different information dimensions, including but not limited to blood sugar monitoring, blood pressure measurement, blood lipid testing, medication records, and lifestyle surveys. Through the steps mentioned above, the computer system has generated corresponding dimensional patient representation vectors for this patient in each information dimension. Now, the goal is to integrate these scattered vectors into a unified vertex patient representation vector to more comprehensively reflect the patient's overall health status.

[0170] In step S500A, the first task of the computer system is to obtain the dimensional patient representation vectors corresponding to the s-th patient information vertex in the a information dimensions. These vectors are the results of the previous steps, and each vector captures the health characteristics of the patient in a specific dimension. Next, the system performs a representation information integration operation on these dimensional patient representation vectors. The specific method of the integration operation depends on the characteristics of the data and the analysis requirements, but generally includes the following steps:

[0171] Feature alignment: Since representation vectors of different dimensions may contain different numbers and types of features, feature alignment is performed first. This includes unifying the length, type, and possible scaling or normalization of the vectors to ensure that all vectors are compared and integrated under the same standard.

[0172] Feature fusion: After feature alignment, the system can use a variety of methods to fuse features. For example, a simple concatenation operation can directly connect features of all dimensions into a longer vector; weighted summation can assign different weights to different vectors according to the importance of each dimension; and more advanced methods such as using deep learning models (such as multi-layer perceptron MLP, long short-term memory network LSTM or graph neural network GNN) for feature extraction and fusion can capture the complex interactive relationship between different dimensions.

[0173] Dimensionality reduction and compression: When the dimensionality of the fused vector is too high, dimensionality reduction may be required to reduce computational complexity and avoid overfitting. Common dimensionality reduction methods include principal component analysis (PCA), linear discriminant analysis (LDA), and autoencoder.

[0174] Taking diabetic patients as an example, assume that three dimensional patient representation vectors are generated under the three dimensions of blood sugar monitoring, blood pressure measurement, and blood lipid examination. and These vectors may contain key health indicators under their respective dimensions, such as blood sugar concentration, blood pressure value, cholesterol and triglyceride levels, etc.

[0175] In step S500A, feature alignment is first performed to ensure that the lengths and types of the three vectors are consistent. Then, a concatenation operation can be selected to perform feature fusion to obtain a fused vector:

[0176]

[0177] However, since the concatenated vector may be too high in dimension, further dimensionality reduction may be required. Here, PCA and other methods can be used to project the fused vector into a low-dimensional space to obtain the final vertex patient representation vector

[0178] Finally, this vertex patient representation vector not only contains the patient's health characteristics in each individual information dimension, but also captures the complex relationship between these characteristics through feature fusion and dimensionality reduction. This vector will serve as the basis for subsequent health prediction, disease diagnosis and personalized care strategy formulation, providing strong support for the deep mining and intelligent analysis of medical care data.

[0179] In one embodiment, the method may further include:

[0180] Step S500B: Obtain the dimensional patient representation vectors corresponding to the s-th patient information vertex in the a information dimensions, obtain the patient identification code representation vector of the s-th patient information vertex, integrate the representation information of the a-dimensional patient representation vectors and the patient identification code representation vector corresponding to the s-th patient information vertex, and obtain the vertex patient representation vector of the s-th patient information vertex.

[0181] In order to better understand and analyze the overall health status of the patient, the computer system not only needs to integrate the specific data of the patient in different information dimensions, but also needs to consider the individual identity characteristics of the patient. To this end, step S500B proposes a new implementation method, that is, in the process of generating the vertex patient representation vector, the patient identification code representation vector is introduced and integrated with the dimensional patient representation vector of the patient in multiple information dimensions.

[0182] Suppose you are processing patient data from a large hospital, which covers a wide range of medical information dimensions, including but not limited to physiological indicators (such as heart rate, blood pressure, blood sugar, etc.), imaging data (such as X-rays, CT scans, etc.), medication records, surgical records, and lifestyle habits. Each patient has a unique patient identification code to identify their identity and medical records. Now, the goal of the computer system is to generate a comprehensive vertex patient representation vector for each patient for subsequent health prediction, disease diagnosis, and the formulation of personalized care strategies.

[0183] In step S500B, the computer system first obtains two key information: one is the dimensional patient representation vector corresponding to the patient at the sth information vertex (i.e., a specific medical record point of a specific patient) in the a information dimensions; the other is the patient identification code representation vector of the patient. Then, the system will integrate the representation information of these two types of vectors to generate the final vertex patient representation vector.

[0184] 1. Obtain the dimensional patient representation vector.

[0185] As mentioned above, dimensional patient representation vectors are generated through a series of complex data processing and analysis steps, which capture the health characteristics of patients in specific information dimensions. In step S500B, the computer system retrieves the dimensional patient representation vectors corresponding to the s-th patient information vertex in the a information dimensions from the database. These vectors may have different lengths and feature dimensions, so appropriate preprocessing and feature alignment are required before integration.

[0186] 2. Obtain the patient identification code representation vector.

[0187] The patient identification code representation vector is a digital representation of the patient's individual identity characteristics. In practical applications, the patient identification code may be a simple digital code, but the computer system usually converts it into a higher-dimensional vector form for subsequent numerical calculations and feature fusion. This conversion can be achieved through a variety of methods, such as hash functions, embedding layers, or graph embedding techniques (such as Node2Vec, DeepWalk, etc.).

[0188] Hash function: The patient identification code can be mapped to a fixed-length vector, but this method usually only retains the hash value of the identification code and cannot capture its underlying semantic information.

[0189] Embedding layer: In deep learning, the embedding layer can map discrete identifiers (such as patient identification codes) into a dense, low-dimensional continuous vector space. This approach can capture the potential relationship between identifiers and facilitate subsequent feature fusion and model training.

[0190] Graph embedding technology: If the patient identification code is associated with other entities in the hospital (such as doctors, departments, equipment, etc.), graph embedding technology can be used to generate a representation vector of the patient identification code. This method can use the node and edge information in the graph structure to enrich the individual feature representation of the patient.

[0191] 3. Representation information integration.

[0192] After obtaining the dimensional patient representation vector and the patient identification code representation vector, the computer system integrates the representation information of these two types of vectors. The purpose of the integration is to combine the individual identity characteristics of the patient with their health characteristics in different information dimensions to generate a more comprehensive and accurate patient representation vector.

[0193] There are many different ways to integrate data, depending on the characteristics of the data and the analysis requirements. Here are some common integration strategies:

[0194] Concatenation: directly concatenate the patient identification code representation vector and all dimensional patient representation vectors into a longer vector. This method is simple and intuitive, but it may cause the vector dimension to be too high, increasing the complexity of subsequent processing.

[0195] Weighted summation: Assign different weights to the patient identification code representation vector and each dimension patient representation vector, and then perform weighted summation. The weight assignment can be determined based on domain knowledge, data importance, or model learning results.

[0196] Attention mechanism: Use the attention mechanism to dynamically adjust the contribution of different vectors in the integration process. The attention mechanism can learn which vectors are more important to the current task and increase their weights accordingly.

[0197] Deep learning models: Use deep learning models such as multi-layer perceptron (MLP), convolutional neural network (CNN), recurrent neural network (RNN) or graph neural network (GNN) for feature extraction and fusion. These models can automatically learn complex relationships between vectors and generate higher-level feature representations.

[0198] The weight distribution process is learned through the attention model, which relies on a large amount of training data and iterative optimization. In practical applications, it may be necessary to use the patient's historical medical records as a training set to train the attention model and evaluate the performance of the model through cross-validation and other methods.

[0199] Step S500B introduces the patient identification code representation vector and integrates it with the dimensional patient representation vectors under multiple information dimensions, providing a new idea and method for generating a comprehensive and accurate patient representation vector. This method not only takes into account the specific health status of the patient under different information dimensions, but also takes into account the individual identity characteristics of the patient, which helps to improve the accuracy and personalization level of digital management of medical care data.

[0200] As an implementation manner, in step S500B, obtaining the patient identification code representation vector of the s-th patient information vertex may specifically include:

[0201] Step S500B1: obtaining an identification code graph structure; the identification code graph structure is consistent with a patient information graph structure, and the identification code graph structure includes b patient information vertices;

[0202] Step S500B2: In the identification code graph structure, filter and merge the identification code representation vectors of the adjacent vertices of the s-th patient information vertex in the l-th filtering layer to obtain the layer identification code representation vector of the s-th patient information vertex in the v+1-th filtering layer, where v≥1;

[0203] Step S500B3: If v+1 is equal to the preset conduction number, then obtain the layer identification code representation vectors of the s-th patient information vertex in the G filter layers respectively, integrate the representation information of the layer identification code representation vectors of the s-th patient information vertex in the G filter layers respectively, and obtain the patient identification code representation vector of the s-th patient information vertex; wherein, G≥1, G is the preset conduction number.

[0204] For example, imagine a large general hospital whose medical information system stores tens of thousands of patient records. Each patient has a unique patient identification code that uniquely identifies him or her in the system. At the same time, these records are organized into multiple patient information graph structures, each of which represents a different information dimension (such as physiological indicators, medication records, imaging data, etc.) to fully reflect the patient's health status. In order to generate a patient identification code representation vector, the computer system first constructs an identification code graph structure corresponding to the patient information graph structure.

[0205] In step S500B1, the computer system retrieves the patient identification codes of all patients from the medical information system and constructs an identification code graph structure based on these identification codes. This graph structure is consistent with the a patient information graph structures mentioned above in terms of topology, that is, they share the same vertex set (i.e., patient information vertices, the number of which is b), but the attributes of the edges and vertices are different. In the identification code graph structure, the vertices represent the patient's identity (i.e., the patient identification code), and the edges may represent some kind of association between patients (such as family members, common doctors, etc.), but this association is usually not necessary in the process of generating patient identification code representation vectors.

[0206] For example, suppose there are three patients in a hospital, and their patient identification codes are PID1, PID2, and PID3. When constructing the identification code graph structure, the computer system will add these three identification codes as vertices to the graph. Since this step mainly focuses on the representation of vertices (i.e., patient identification codes), the edge information is ignored at this stage.

[0207] In step S500B2, for example, in the identification code graph structure, each vertex (i.e., the patient identification code) needs to be converted into a high-dimensional representation vector for subsequent numerical calculations and feature fusion. In order to capture the potential relationship between vertices and enrich the information content of the representation vector, the computer system can adopt the filtering and merging strategy in the graph neural network (GNN) to gradually update the representation vector of each vertex through multi-layer graph convolution operations.

[0208] First, the computer system assigns an initial characterization vector to each vertex. This vector can be random or generated by some heuristic method (such as a hash value based on the patient identification code). It is important that the initial characterization vectors of all vertices have the same dimension.

[0209] Next, the system applies a graph convolutional network (GCN) or other types of graph neural networks to filter and merge the vertices in the identification code graph structure. In each filtering layer (or convolutional layer), the representation vector of each vertex is updated according to the representation vectors of its neighboring vertices. Specifically, the system calculates a new representation vector through aggregation functions (such as summation, averaging, maximum pooling, etc.) based on the structural information of the graph (i.e., the adjacency relationship of the vertices) and the feature information of the vertices (which may be random or heuristic in the initial stage). During the filtering and merging process, the system can use convolution kernels (or filters) to extract and transform features of the local neighborhood of the vertex. The parameters of these convolution kernels are obtained through model training, and they can learn how to effectively combine the information of adjacent vertices to update the representation vector of the current vertex. Through multiple layers of filtering and merging operations, the system can gradually capture the complex relationships between vertices and generate richer representation vectors.

[0210] For example, suppose that the adjacent vertices of PID1 (assuming PID2 and PID3) are filtered and merged at the lth filtering layer. The system will first obtain the representation vectors of PID2 and PID3 at this layer, and then calculate the intermediate representation vector of PID1 at this layer through an aggregation function (such as summation) based on the adjacency relationship between PID1 and them (in this example, assuming that PID1 is directly connected to PID2 and PID3):

[0211] The system can then apply a convolution kernel to transform the intermediate representation vector to generate the representation vector of PID1 at layer l+1. This process is repeated multiple times (i.e., multiple filter layers) until the preset number of transmissions v+1 is reached.

[0212] In step S500B3, the layer identification code representation vectors are integrated and the patient identification code representation vector is generated. For example, after the multi-layer filtering and merging operation, each patient identification code (i.e., the vertex in the identification code graph structure) will generate a corresponding layer identification code representation vector in multiple filtering layers. These vectors capture the characteristic information of the patient identification code at different abstract levels. In order to generate a comprehensive patient identification code representation vector, the computer system integrates these layer identification code representation vectors.

[0213] First, the system checks whether the current number of filter layers v+1 is equal to the preset number of transmission times G. If so, it means that all the predetermined filter merging operations have been completed; if not, continue to execute step S500B2 to perform filter merging of the next layer. When the preset number of transmission times is reached, the system collects the layer identification code representation vectors of each patient identification code in the G filter layers. Then, some integration strategy (such as splicing, weighted summation, attention mechanism, etc.) is used to merge these vectors into a comprehensive patient identification code representation vector.

[0214] Finally, the system uses the generated patient identification code representation vector as the output result of step S500B for subsequent data integration and analysis processes.

[0215] For example, assuming that the preset number of conduction times G is 3, the system plans to generate a patient identification code representation vector through a three-layer filter merging operation. After completing the filter merging of all three layers, the system collects the layer identification code representation vectors of PID1 in the three filter layers. Then, the system can use a splicing strategy to integrate these vectors.

[0216] By implementing step S500B, the computer system can generate a patient identification code representation vector using the filtering and merging strategy in the graph neural network. This process not only takes into account the information of the patient identification code itself, but also captures the potential relationship between patients through multi-layer graph convolution operations. The generated patient identification code representation vector not only enriches the dimension of patient information, but also provides strong support for subsequent data integration and analysis.

[0217] In one implementation, the b patient information vertices include a first vertex and a second vertex; based on this, the method may further include:

[0218] Step S610: if vertex patient characterization vectors corresponding to b patient information vertices are generated, then a first vertex patient characterization vector corresponding to a first vertex is obtained, and a second vertex patient characterization vector corresponding to a second vertex is obtained;

[0219] Step S620: taking the dot product result of the first vertex patient characterization vector and the second vertex patient characterization vector as the vertex correlation coefficient between the first vertex and the second vertex.

[0220] In the embodiment of the present application, it is assumed that the medical data of a patient with a complex disease is being processed. The patient suffers from multiple complications, and his medical records cover a wide range of medical information dimensions, such as physiological indicators (such as heart rate, blood pressure, blood sugar, etc.), imaging data (such as X-rays, MRI scans, etc.), medication records, surgical records, and genetic information, etc. These data are organized into multiple patient information graph structures, each of which represents an information dimension, and the vertices in the graph represent specific medical entities related to the dimension (such as a specific examination result, medication record, etc.).

[0221] In this example, we are interested in two vertices: the first vertex represents the patient's most recent echocardiogram results, and the second vertex represents the patient's most recent blood test results. The goal is to quantify the correlation between these two vertices in order to better understand how they jointly affect the patient's health status.

[0222] In step S610, the computer system first checks whether corresponding vertex patient characterization vectors have been generated for all b patient information vertices. These vectors are calculated step by step through the steps mentioned above (such as S400 and S500), which capture the health characteristics of each vertex in different information dimensions and perform multi-level feature enhancement and integration.

[0223] For the first vertex (cardiac ultrasound test results) and the second vertex (blood test results) in this example, the computer system will retrieve their corresponding vertex patient representation vectors from the repository. Assume that the patient representation vector of the first vertex is The patient characterization vector of the second vertex is

[0224]

[0225] In step S620, the computer system uses a dot product operation to calculate the vertex correlation coefficient between the first vertex and the second vertex. The dot product (also called the inner product) is a basic concept in linear algebra, which measures the similarity of two vectors in direction. When the two vectors are in the same direction, their dot product result is positive and large; when the directions are opposite, the dot product result is negative; when they are perpendicular, the dot product result is 0.

[0226] In this example, the first vertex patient representation vector and the second vertex patient characterization vector Perform dot multiplication and the result will be used as the vertex correlation coefficient. The calculation formula is as follows:

[0227]

[0228] Where n is the dimension of the vector (i.e. the number of features), v heart,i ,v blood,i They are the values ​​of the two vectors in the i-th dimension (feature).

[0229] The calculated vertex correlation coefficient can be used in a variety of application scenarios:

[0230] Health prediction: By analyzing the correlation between different vertices, a more accurate health prediction model can be built. For example, if it is found that the results of cardiac ultrasound examinations are highly correlated with blood test results, then the blood test results can be given a higher weight when predicting heart health.

[0231] Personalized care strategies: Understanding the correlation between vertices can help develop more personalized care strategies. If two test results are highly correlated and both point to a certain health risk, then joint care measures can be taken to target both risks.

[0232] Data mining and analysis: Vertex correlation coefficient can be used as one of the important features of data mining and analysis to discover potential medical patterns, association rules or outliers.

[0233] Graph structure optimization: When constructing or optimizing the patient information graph structure, the vertex correlation coefficient can be used to adjust the connection weights between vertices or add new edges to enhance the expressiveness and analytical value of the graph structure.

[0234] Through the implementation of steps S610 and S620, the computer system can effectively calculate the correlation between the patient information vertices, and this result is of great significance for in-depth understanding of the patient's health status, formulating personalized care strategies, and optimizing medical data analysis.

[0235] In one embodiment, the method may further include a training process of a representation vector reasoning neural network, which may specifically include:

[0236] Step S10: obtaining a graph structure samples; each graph structure sample has b patient information vertex samples; a≥1, b≥1; the information dimensions corresponding to the a graph structure samples are inconsistent;

[0237] Step S20: loading a graph structure samples into the initial representation vector inference neural network, obtaining the sample vertex-related classification of the adjacent vertex samples of the sth patient information vertex sample for the sth patient information vertex sample in the eth information dimension of the initial representation vector inference neural network; 1≤e≤a; 1≤s≤b;

[0238] Step S30: according to the sample vertex related classification, determine the sample adjacent related characterization vector of the sth patient information vertex sample for the adjacent vertex sample of the sth patient information vertex sample; the sample vertex related classification includes the first sample vertex related classification and the second sample vertex related classification; the first sample vertex related classification is used to characterize that the sample adjacent related characterization vector is based on the adjacent vertex sample of the sth patient information vertex sample, and the first sample adjacent characterization vectors corresponding to each other in a information dimensions are determined; the second sample vertex related classification is used to characterize that the sample adjacent related characterization vector is based on the adjacent vertex sample of the sth patient information vertex sample, and the first sample adjacent characterization vector corresponding to each other in the e information dimension is determined;

[0239] Step S40: in the e-th information dimension, according to the sample adjacent related representation vectors of the adjacent vertex samples of the s-th patient information vertex sample for the s-th patient information vertex sample, determine the proposed enhanced sample patient representation vector of the s-th patient information vertex sample;

[0240] Step S50: Based on the sample patient representation vector to be enhanced, generate a sample dimension patient representation vector of the sth patient information vertex sample in the eth information dimension; the sample dimension patient representation vector is used to integrate and construct a sample vertex patient representation vector that can represent the sth patient information vertex sample.

[0241] The principles of the above steps S10 to S50 can refer to the description of the above steps S100 to S500, which will not be repeated here.

[0242] Step S60: if the sample vertex patient characterization vectors corresponding to the b patient information vertex samples are generated, then according to the sample vertex patient characterization vectors corresponding to the b patient information vertex samples, the sample reasoning correlation coefficient of the sample vertex tuple in the b patient information vertex samples is determined; the sample vertex tuple includes two different patient information vertex samples in the b patient information vertex samples;

[0243] Step S70: updating the parameters of the initial representation vector inference neural network according to the sample inference correlation coefficient of the sample vertex binary to obtain the representation vector inference neural network.

[0244] After generating the sample vertex patient representation vectors corresponding to all b patient information vertex samples, the reasoning network will determine the sample reasoning correlation coefficient between the vertex samples based on these representation vectors. This correlation coefficient is an important indicator to measure the association or similarity between vertex samples.

[0245] For example, for each of the b patient information vertex samples in each sample, the inference network calculates the sample inference correlation coefficient between them. This can be achieved by calculating the cosine similarity, Euclidean distance or other similarity measures between the representation vectors. For example, for vertex samples i and j, their sample inference correlation coefficient can be calculated by the following formula:

[0246]

[0247] in, and are the sample vertex patient representation vectors of vertex samples i and j respectively.

[0248] In step S70, the inference network updates its parameters according to the sample inference correlation coefficient of the sample vertex binary. This step is implemented by the back propagation algorithm, which aims to minimize a loss function (such as cross entropy loss, mean square error loss, etc.) that measures the difference between the sample inference correlation coefficient predicted by the network and the actual correlation coefficient.

[0249] Suppose there is a training set containing multiple sample vertex pairs. For each sample vertex pair (i, j), there is a true correlation coefficient label yij and a correlation coefficient predicted by the network. A mean square error loss function can be defined to quantify the prediction error:

[0250]

[0251] Where N is the number of sample vertex pairs in the training set. By minimizing this loss function, we can use optimization algorithms such as gradient descent to update the parameters of the inference network so that it can better predict the correlation between vertex samples.

[0252] Through the above steps, the computer system can train a representation vector reasoning neural network that can process graph structured data. This network can automatically learn the associations between patient information vertices and generate a vector that can fully represent the patient's health status.

[0253] As an implementation method, b patient information vertex samples correspond to r sample vertex tuples; wherein r≥1; step S70, based on the sample reasoning correlation coefficient of the sample vertex tuple, the initial representation vector reasoning neural network is updated with parameters to obtain the representation vector reasoning neural network, which may specifically include:

[0254] Step S71: Among the r sample vertex pairs, the sample vertex pairs with connection lines in the a graph structure samples are used as positive sample vertex pairs, and the sample vertex pairs without connection lines in the a graph structure samples are used as negative sample vertex pairs;

[0255] Step S72: constructing an error calculation function based on the sample reasoning correlation coefficient of the positive sample vertex binary group and the sample reasoning correlation coefficient of the negative sample vertex binary group;

[0256] Step S73: updating the parameters of the initial representation vector inference neural network according to the error calculation function to obtain the representation vector inference neural network.

[0257] In the field of healthcare, patients' medical records are often stored in the form of a graph structure. The vertices in the graph represent different medical entities (such as diagnoses, test results, medication records, etc.), and the edges represent the relationships between these entities. In order to train the representation vector inference neural network, it is necessary to extract graph structure samples from a large number of medical records and build a training set based on these samples. In this process, how to distinguish between positive sample vertex binary groups and negative sample vertex binary groups becomes a key step.

[0258] In step S71, the computer system first traverses all sample vertex pairs in the training set. Each sample vertex pair consists of two patient information vertex samples, representing two different vertices in the graph structure. Next, the system will distinguish whether they are positive samples or negative samples based on the connection between the two vertices in the graph structure sample.

[0259] Positive example vertex tuple: If two vertices have a direct connection line in the graph structure sample (that is, there is a clear relationship between them, such as causal relationship, temporal order relationship, or structural adjacent relationship), then the vertex tuple they form is considered a positive example. For example, during a medical examination, there is a clear causal relationship between the patient's physiological indicator measurement result (vertex A) and the subsequent diagnosis result (vertex B), so the tuple formed by vertex A and vertex B is a positive example.

[0260] Negative example vertex pairs: On the contrary, if two vertices do not have a direct connection line in the graph structure example (that is, there is no obvious direct relationship between them), the vertex pair they form is considered a negative example. For example, in a patient's medical record, a medication record (vertex C) may not have a direct association with an unrelated test result (vertex D), so the pair consisting of vertex C and vertex D is a negative example.

[0261] For example, suppose there is a training set containing multiple graph structure examples, each of which represents a patient's medical records under different information dimensions. In these examples, multiple vertex pairs can be found. For example, in a graph structure example, there is a connection line between the vertex "abnormal electrocardiogram test results" and the vertex "heart disease diagnosis", so the pair they form is considered a positive example. However, there is no connection line between the vertex "normal blood sugar level" and the vertex "fracture treatment record", so the pair they form is considered a negative example.

[0262] In step S72, after distinguishing the positive sample vertex binary and the negative sample vertex binary, the computer system will then construct an error calculation function (i.e., loss function) based on the sample reasoning correlation coefficients of these binary groups. The loss function is an important indicator to measure the difference between the model prediction results and the actual results. By minimizing the loss function, the parameters of the neural network can be optimized to better fit the training data.

[0263] In the embodiments of the present application, a variety of loss functions can be used to calculate the difference between the correlation coefficient predicted by the model and the actual correlation coefficient. A commonly used method is the binary cross-entropy loss, which is particularly suitable for processing binary classification problems (such as determining whether a vertex binary is a positive sample or a negative sample).

[0264] However, in the scenario, it is not only necessary to judge whether the vertex binary is a positive sample or a negative sample, but also to consider the inference correlation coefficient between them. Therefore, the binary cross entropy loss can be modified to consider both the classification accuracy and the prediction accuracy of the correlation coefficient.

[0265] A possible way to construct the loss function is as follows:

[0266] Classification loss: For each vertex binary, we first calculate the binary cross entropy loss between the probability of it being a positive example (predicted by the model) and the actual label (0 for negative examples and 1 for positive examples).

[0267] Correlation coefficient loss: For the positive sample vertex pair, it is also necessary to calculate the difference between the inference correlation coefficient predicted by the model and the actual correlation coefficient. This can be achieved through the mean squared error (MSE) or other similar regression loss functions.

[0268] Comprehensive loss: Finally, the classification loss and correlation coefficient loss are weighted and summed according to certain weights to obtain the final loss function. The choice of weights can be determined according to the requirements of the specific task and the characteristics of the data set.

[0269] For example, suppose there is a training set of 100 sample vertex pairs, 50 of which are positive and 50 are negative. For each positive sample vertex pair, there is an actual correlation coefficient value (e.g., obtained by domain expert evaluation). During training, the model will make a classification prediction for each vertex pair (determine whether it is a positive sample or a negative sample) and predict the correlation coefficient value of the positive sample.

[0270] For the classification prediction part, the binary cross entropy loss can be used to calculate the difference between the predicted probability and the actual label. For the correlation coefficient prediction part, the mean square error loss can be used to calculate the difference between the predicted correlation coefficient and the actual correlation coefficient. Finally, the two parts of the loss are weighted and summed according to a certain ratio (such as 0.5:0.5) to obtain the comprehensive loss of each vertex binary. The loss of the entire training set is the average of the comprehensive losses of all vertex binary groups.

[0271] In step S73, after constructing the loss function, the computer system can update the parameters of the initial representation vector inference neural network according to the function. This process is usually implemented by a back propagation algorithm.

[0272] The back propagation algorithm is an effective method for training neural networks. It first calculates the model's prediction results and loss values ​​through forward propagation, then passes the loss values ​​backward along each layer of the network through back propagation and calculates the gradient of each layer's parameters. Based on these gradient values, gradient descent or other optimization algorithms can be used to update the network's parameters to reduce the loss value and improve the model's prediction accuracy.

[0273] In the embodiment of the present application, a variety of optimization algorithms can be selected to update the parameters of the characterization vector inference neural network. For example, Stochastic Gradient Descent (SGD), Adam optimizer (Adaptive Moment Estimation) or other more advanced optimization algorithms can be used. These algorithms can effectively adjust the network parameters to better fit the training data and generalize to unseen data.

[0274] For example, in step S73, the computer system uses the previously constructed loss function and the selected optimization algorithm to update the parameters of the representation vector inference neural network. Specifically, the system will first perform forward propagation calculations based on the current network parameters and input data (i.e., graph structure samples) to obtain prediction results and loss values; then calculate the gradient of each layer of parameters based on the loss value; and finally use the optimization algorithm to update the network parameters based on the gradient value. This process will be repeated for multiple iterations (i.e., training rounds) until the loss value is reduced to a certain level or reaches a preset number of iterations. In each iteration, the system will randomly select a batch of samples from the training set (i.e., small batch gradient descent) to update the network parameters to improve training efficiency and stability.

[0275] Through continuous parameter updating and optimization, the performance of the representation vector reasoning neural network will gradually improve and stabilize. The final model will be able to accurately represent the patient information vertices and understand the relationship between them, thus providing strong support for the analysis and application of medical care data.

[0276] The present application embodiment provides a computer system, such as Figure 2 As shown, the computer system 100 includes: a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, such as through a bus 102. Optionally, the computer system 100 may also include a transceiver 104. It should be noted that in actual applications, the transceiver 104 is not limited to one, and the structure of the computer system 100 does not constitute a limitation on the embodiments of the present application.

[0277] An embodiment of the present application provides a computer system. The computer system in the embodiment of the present application includes: one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors. When the one or more programs are executed by the processor, the method provided by the present application is implemented.

[0278] The above description is only a partial implementation method of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for digital management of medical care data, characterized in that: The method comprises: Obtain a patient information graph structures; each patient information graph structure has b patient information vertices; wherein a≥1, b≥1; and the information dimensions corresponding to the a patient information graph structures are inconsistent; In the e-th information dimension, obtain the vertex-related classification of the adjacent vertices of the s-th patient information vertex for the s-th patient information vertex; wherein 1≤e≤a, 1≤s≤b; According to the vertex-related classification, the adjacent vertices of the s-th patient information vertex and the adjacent related characterization vectors of the s-th patient information vertex are determined; the vertex-related classification includes a first vertex-related classification and a second vertex-related classification; the first vertex-related classification is used to characterize that the adjacent related characterization vector is determined based on the adjacent vertices of the s-th patient information vertex and the first adjacent characterization vectors corresponding to the adjacent vertices in a information dimensions are respectively determined; the second vertex-related classification is used to characterize that the adjacent related characterization vector is determined based on the adjacent vertices of the s-th patient information vertex and the first adjacent characterization vector corresponding to the adjacent vertices in the e-th information dimension is respectively determined; In the e-th information dimension, determining a proposed enhanced patient representation vector of the s-th patient information vertex according to adjacent related representation vectors of adjacent vertices of the s-th patient information vertex with respect to the s-th patient information vertex; Based on the proposed enhanced patient representation vector, a dimensional patient representation vector of the sth patient information vertex in the eth information dimension is generated, and the dimensional patient representation vector is used to integrate and construct a vertex patient representation vector that can represent the sth patient information vertex.

2. The method according to claim 1, characterized in that In the e-th information dimension, obtaining the vertex-related classification of the adjacent vertices of the s-th patient information vertex for the s-th patient information vertex includes: In the e-th information dimension, obtaining a first adjacent characterization vector of adjacent vertices of the s-th patient information vertex, and obtaining an initial patient information vertex characterization vector of the s-th patient information vertex; In the e-th information dimension, determining a correlation coefficient of an initial characterization vector of an adjacent vertex of the s-th patient information vertex with respect to the s-th patient information vertex according to the first adjacent characterization vector and the initial patient information vertex characterization vector; If the correlation coefficient of the initial characterization vector in the e-th information dimension is greater than a preset value, the first vertex correlation classification is performed as a vertex correlation classification of the adjacent vertices of the s-th patient information vertex for the s-th patient information vertex; If the correlation coefficient of the initial characterization vector in the e-th information dimension is not greater than the preset value, the second vertex correlation classification is performed as a vertex correlation classification of the adjacent vertices of the s-th patient information vertex for the s-th patient information vertex.

3. The method according to claim 2, characterized in that The number of adjacent vertices of the s-th patient information vertex is X, where X≥1; in the e-th information dimension, determining the correlation coefficient of the initial representation vector of the adjacent vertices of the s-th patient information vertex with respect to the s-th patient information vertex according to the first adjacent representation vector and the initial patient information vertex representation vector, including: In the e-th information dimension, a representation information integration operation is performed on the first adjacent representation vectors corresponding to the x adjacent vertices and the initial patient information vertex representation vector to obtain an initial vertex integration vector of the s-th patient information vertex; In the e-th information dimension, obtaining initial vertex integration subvectors between X first adjacent representation vectors and the initial patient information vertex representation vector respectively; In the e-th information dimension, the initial representation vector correlation coefficients of the X adjacent vertices for the s-th patient information vertex are determined according to the ratio between the X initial vertex integration subvectors and the initial vertex integration vector.

4. The method according to claim 1, characterized in that The step of determining the adjacent related characterization vector of the adjacent vertices of the s-th patient information vertex for the s-th patient information vertex according to the vertex related classification includes: If the vertex-related classification is the first vertex-related classification, then obtaining the first adjacent representation vectors corresponding to the adjacent vertices of the s-th patient information vertex in the a information dimensions, integrating the representation information of the a first adjacent representation vectors to obtain an adjacent integrated representation vector, performing unified modal mapping on the adjacent integrated representation vector to obtain an adjacent related representation vector of the adjacent vertices of the s-th patient information vertex for the s-th patient information vertex; If the vertex-related classification is the second vertex-related classification, obtaining the first adjacent characterization vector corresponding to the adjacent vertices of the s-th patient information vertex in the e-th information dimension, and using the first adjacent characterization vector in the e-th information dimension as the adjacent related characterization vector of the adjacent vertices of the s-th patient information vertex for the s-th patient information vertex; The method of determining the proposed enhanced patient representation vector of the sth patient information vertex in the eth information dimension according to the adjacent related representation vectors of the adjacent vertices of the sth patient information vertex with respect to the sth patient information vertex comprises: In the e-th information dimension, according to the adjacent related characterization vectors of the adjacent vertices of the s-th patient information vertex for the s-th patient information vertex and the initial patient information vertex characterization vector of the s-th patient information vertex, determine the correlation coefficient of the 0th level characterization vector of the adjacent vertices of the s-th patient information vertex for the s-th patient information vertex; According to the 0th level characterization vector correlation coefficient, the adjacent vertices of the sth patient information vertex are eccentrically adjusted with respect to the adjacent related characterization vectors of the sth patient information vertex to obtain a quasi-enhanced patient characterization vector of the sth patient information vertex; The step of generating a dimensional patient representation vector of the s-th patient information vertex in the e-th information dimension based on the proposed enhanced patient representation vector includes: In the e-th information dimension, obtain the u-th level adjacent vertex representation vector corresponding to the adjacent vertex of the s-th patient information vertex, and obtain the u-th level patient representation vector corresponding to the s-th patient information vertex, wherein u≥1, and the proposed enhanced patient representation vector of the s-th patient information vertex represents the first-level patient representation vector corresponding to the s-th patient information vertex when u=1; In the e-th information dimension, based on the u-th level adjacent vertex representation vectors corresponding to the adjacent vertices of the s-th patient information vertex, the u-th level patient representation vector corresponding to the s-th patient information vertex is enhanced to obtain the u+1-th level patient representation vector of the s-th patient information vertex; If u+1 is equal to the preset number of enhancements, the u+1-th level patient representation vector is determined as the dimensional patient representation vector of the s-th patient information vertex in the e-th information dimension.

5. The method according to claim 4, characterized in that In the e-th information dimension, according to the u-th level adjacent vertex representation vector corresponding to the adjacent vertex of the s-th patient information vertex, the u-th level patient representation vector corresponding to the s-th patient information vertex is enhanced to obtain the u+1-th level patient representation vector of the s-th patient information vertex, including: In the e-th information dimension, the correlation coefficient of the u-th level representation vector of the adjacent vertices of the s-th patient information vertex with respect to the s-th patient information vertex is obtained, and the u-th level adjacent vertex representation vector is eccentrically adjusted according to the u-th level representation vector correlation coefficient to obtain the u-th level enhanced representation vector; In the e-th information dimension, the u-th level enhanced representation vector is used to perform representation vector enhancement on the u-th level patient representation vector corresponding to the s-th patient information vertex to obtain the u+1-th level patient representation vector of the s-th patient information vertex.

6. The method according to claim 1, characterized in that The method further comprises: Obtaining the dimensional patient representation vectors corresponding to the s-th patient information vertex in the a information dimensions, and performing a representation information integration operation on the a-dimensional patient representation vectors corresponding to the s-th patient information vertex to obtain the vertex patient representation vector of the s-th patient information vertex; Alternatively, the method further comprises: Obtain the dimensional patient representation vectors corresponding to the s-th patient information vertex in the a information dimensions, obtain the patient identification code representation vector of the s-th patient information vertex, integrate the a-dimensional patient representation vectors corresponding to the s-th patient information vertex with the patient identification code representation vector, and obtain the vertex patient representation vector of the s-th patient information vertex.

7. The method according to claim 6, characterized in that The step of obtaining the patient identification code representation vector of the s-th patient information vertex includes: Acquire an identification code graph structure; the identification code graph structure is consistent with the a patient information graph structure, and the identification code graph structure includes the b patient information vertices; In the identification code graph structure, the adjacent vertex identification code representation vectors of the adjacent vertices of the s-th patient information vertex in the l-th filtering layer are filtered and merged to obtain the layer identification code representation vector of the s-th patient information vertex in the v+1-th filtering layer, where v≥1; If v+1 is equal to the preset conduction number, then the layer identification code representation vectors of the s-th patient information vertex in the G filter layers are obtained, and the layer identification code representation vectors of the s-th patient information vertex in the G filter layers are integrated to obtain the patient identification code representation vector of the s-th patient information vertex; wherein G≥1, G is the preset conduction number; The b patient information vertices include a first vertex and a second vertex; the method further includes: If vertex patient characterization vectors corresponding to the b patient information vertices are generated, then a first vertex patient characterization vector corresponding to the first vertex is obtained, and a second vertex patient characterization vector corresponding to the second vertex is obtained; The dot product result of the first vertex patient characterization vector and the second vertex patient characterization vector is used as the vertex correlation coefficient between the first vertex and the second vertex.

8. The method according to claim 1, characterized in that The method further comprises: Obtain a graph structure samples; each graph structure sample has b patient information vertex samples; a≥1, b≥1; the information dimensions corresponding to the a graph structure samples are inconsistent; The a graph structure samples are loaded into the initial representation vector inference neural network, and in the e-th information dimension of the initial representation vector inference neural network, the sample vertex-related classification of the adjacent vertex samples of the s-th patient information vertex sample is obtained for the s-th patient information vertex sample; 1≤e≤a; 1≤s≤b; According to the sample vertex related classification, determine the sample adjacent related characterization vector of the adjacent vertex samples of the s-th patient information vertex sample for the s-th patient information vertex sample; the sample vertex related classification includes a first sample vertex related classification and a second sample vertex related classification; the first sample vertex related classification is used to characterize that the sample adjacent related characterization vector is determined based on the adjacent vertex samples of the s-th patient information vertex sample, and the first sample adjacent characterization vectors corresponding to each other in a information dimensions are determined; the second sample vertex related classification is used to characterize that the sample adjacent related characterization vector is determined based on the adjacent vertex samples of the s-th patient information vertex sample, and the first sample adjacent characterization vector corresponding to each other in the e-th information dimension is determined; In the e-th information dimension, according to the sample adjacent related representation vector of the adjacent vertex samples of the s-th patient information vertex sample for the s-th patient information vertex sample, determine the proposed enhanced sample patient representation vector of the s-th patient information vertex sample; Based on the proposed enhanced sample patient representation vector, a sample dimension patient representation vector of the sth patient information vertex sample in the eth information dimension is generated; the sample dimension patient representation vector is used to integrate and construct a sample vertex patient representation vector that can represent the sth patient information vertex sample; If the sample vertex patient characterization vectors corresponding to the b patient information vertex samples are generated, then according to the sample vertex patient characterization vectors corresponding to the b patient information vertex samples, the sample reasoning correlation coefficient of the sample vertex tuple in the b patient information vertex samples is determined; the sample vertex tuple includes two different patient information vertex samples in the b patient information vertex samples; According to the sample reasoning correlation coefficient of the sample vertex binary, the parameters of the initial representation vector reasoning neural network are updated to obtain the representation vector reasoning neural network.

9. The method according to claim 8, characterized in that The b patient information vertex samples correspond to r sample vertex pairs; wherein r≥1; the initial representation vector inference neural network is parameter updated based on the sample inference correlation coefficient of the sample vertex pairs to obtain the representation vector inference neural network, including: Among the r sample vertex pairs, sample vertex pairs with connection lines in the a graph structure samples are used as positive sample vertex pairs, and sample vertex pairs without connection lines in the a graph structure samples are used as negative sample vertex pairs; Constructing an error calculation function based on the sample reasoning correlation coefficient of the positive sample vertex binary and the sample reasoning correlation coefficient of the negative sample vertex binary; The parameters of the initial representation vector inference neural network are updated according to the error calculation function to obtain the representation vector inference neural network.

10. A computer system, characterized in that: include: one or more processors; Memory; one or more computer programs; The one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method according to any one of claims 1 to 9 is implemented.

Images