Risk assessment stabilization method for senile chronic disease detection

Abstract

The invention discloses a risk assessment stabilization method for senile chronic disease detection, and relates to the technical field of medical data analysis, and the method comprises the steps: carrying out federal dynamic time warping processing on encrypted patient data streams of a plurality of medical institutions, generating a joint feature space mapping matrix, and obtaining a standard feature tensor through homomorphic encryption; inputting the standard feature tensor into a dynamic medical knowledge graph construction module, and generating a knowledge graph embedding matrix through a space-time sensitivity enhanced cross-modal attention mechanism; and fusing the standard feature tensor and the knowledge graph embedding matrix, constructing a dynamic hypergraph structure, executing dual-channel hypergraph convolution calculation, and outputting a patient-knowledge joint embedding matrix. According to the method, in the construction of the dynamic knowledge graph, the time sensitivity weight of the entity relationship is quantified through the exponential decay function, and the time-space enhanced embedded matrix is generated in combination with the cross-modal attention mechanism, so that the time sequence discrimination of the concurrent disease association strength is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data analysis, and in particular to a stable risk assessment method for detecting chronic diseases in the elderly. Background Art

[0002] Currently, the application of federated learning technology in medical data analysis is primarily focused on the construction of multi-institutional data collaboration and privacy protection frameworks. Existing methods typically employ homomorphic encryption protocols (such as the Paillier algorithm) and differential privacy mechanisms (such as Laplace noise injection) to securely aggregate data across medical institutions, and align physiological indicator time series data through static time window interception strategies. In the construction of medical knowledge graphs, mainstream technologies build static association networks based on entity-relationship triplets (such as disease-treatment-drug) based on clinical guidelines, ICD codes, and drug instructions, and utilize graph attention mechanisms for feature embedding. Furthermore, hypergraph convolution methods are being used to capture the complex associations between patients and multiple diseases and symptoms, while techniques such as variational autoencoders enhance the robustness of risk assessment through probabilistic modeling.

[0003] However, these methods have significant limitations when processing multi-source, heterogeneous time series data: First, traditional time alignment strategies rely on fixed time windows, which cannot eliminate global phase shifts caused by sampling frequency differences between medical institutions, leading to the accumulation of feature space mapping errors. Second, static knowledge graphs lack the ability to model the dynamic evolution of a patient's disease course and cannot quantify the time-sensitive weights of entity relationships that decay over time, resulting in distorted associations between concurrent diseases and drug contraindications. For example, the concurrent relationship between diabetes and hypertension may exhibit different weights in the acute phase and the stable phase, but existing technologies do not introduce a time-sensitive decay function for dynamic correction, resulting in a decrease in the confidence level of risk assessment. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a stable risk assessment method for chronic disease detection in the elderly to solve the problem of poor stability of chronic disease risk assessment caused by insufficient accuracy of global phase alignment of multi-source medical time series under the federated learning framework and lack of timeliness modeling of dynamic medical knowledge graphs.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: In a first aspect, the present invention provides a stable risk assessment method for geriatric chronic disease detection, comprising: performing federated dynamic time warping on encrypted patient data streams from multiple medical institutions, generating a joint feature space mapping matrix, and obtaining a standard feature tensor through homomorphic encryption; Inputting the standard feature tensor into a dynamic medical knowledge graph construction module, and generating a knowledge graph embedding matrix through a cross-modal attention mechanism with enhanced spatiotemporal sensitivity; Fusing the standard feature tensor with the knowledge graph embedding matrix, constructing a dynamic hypergraph structure and performing a dual-path hypergraph convolution calculation to output a joint patient-knowledge embedding matrix; Calculating a credibility score based on the joint embedding matrix, optimizing a credibility weighted loss function in combination with variational inference, and generating a weighted risk prediction vector; The risk prediction vector is used as a teacher signal, and an optimized global model is output through contribution-weighted federated parameter aggregation and dynamic knowledge distillation.

[0007] As a preferred solution of the risk assessment stabilization method for elderly chronic disease detection according to the present invention, the generating of the joint feature space mapping matrix comprises the following steps: Configure a two-way authenticated SSL / TLS communication channel between the medical institution and the data processing center to encapsulate and encrypt data stream packets; Perform hash verification and decryption on the received encrypted data stream to generate a structured data table; Convert the structured data table into a two-dimensional numerical matrix, traverse the physiological indicator categories in the column order of the two-dimensional numerical matrix, and construct a three-dimensional time series matrix; The time warping operator is iteratively updated by the alternating direction multiplication method to perform phase shift compensation on the three-dimensional time series matrix and generate a joint feature space mapping matrix.

[0008] As a preferred solution of the risk assessment stabilization method for detecting chronic diseases in the elderly described in the present invention, the obtaining of the standard feature tensor refers to weighted summation of the joint feature space mapping matrix through the Paillier homomorphic encryption algorithm and superposition of Laplace noise.

[0009] As a preferred solution of the risk assessment stabilization method for elderly chronic disease detection according to the present invention, the construction of the dynamic medical knowledge graph includes the following steps: Based on the disease classification coding of the standard feature tensor, the mapping generates a set of composite entities; Construct triple sets based on clinical guidelines and drug instructions; Based on the time series data of patients with multiple diseases co-occurring in the standard feature tensor, the co-occurrence frequency between disease entities is counted to generate a dynamic medical knowledge graph.

[0010] As a preferred solution of the risk assessment stabilization method for elderly chronic disease detection according to the present invention, the generating of the knowledge graph embedding matrix comprises the following steps: Match the dynamic medical knowledge graph with the patient's physiological indicator time series fragments to generate a set of four-tuples; The spatiotemporal adjacency matrix between entities is calculated by using time-sensitivity weights and exponential decay functions, and the knowledge graph embedding matrix is ​​generated by combining the cross-modal attention mechanism.

[0011] As a preferred solution of the risk assessment stabilization method for elderly chronic disease detection according to the present invention, the output of the patient-knowledge joint embedding matrix includes the following steps: Perform triple interaction operations based on the patient feature vector and the entity vector of the knowledge graph embedding matrix to calculate the triple interaction score; Filter the effectively connected patient pairs according to the dynamic threshold to generate the original hyperedge set; Create a patient-hyperedge correlation matrix and a diagonal matrix with time weights, and combine principal component analysis to reduce the dimension of patient feature slices to form the initial hidden representation matrix; Generate the hypergraph convolution pathway output through patient degree matrix normalization and hypergraph convolution pathway calculation; Based on the graph attention mechanism, the entity relationship weights of the knowledge graph embedding matrix are integrated to generate the graph attention path output; The hypergraph convolution pathway is fused with the output of the graph attention pathway to generate a joint patient-knowledge embedding matrix.

[0012] As a preferred solution of the stable risk assessment method for elderly chronic disease detection according to the present invention, the generating of a weighted risk prediction vector comprises the following steps: Based on the patient-knowledge joint embedding matrix, the same set of patients is filtered by disease category label, and the mean vector of the embedding vectors of the same patients is calculated; Patient credibility scores were generated by the ratio of the squared Euclidean distance to the maximum value; Map the patient embedding vector to the mean vector and variance vector of the variational distribution, and define the standard normal distribution as the prior distribution; Combining the credibility score weighted binary cross entropy loss and KL divergence term, we construct a total loss function and optimize the variational parameters through backpropagation; Based on the updated patient-knowledge joint embedding matrix, a weighted risk prediction vector is generated.

[0013] As a preferred solution of the risk assessment stabilization method for elderly chronic disease detection according to the present invention, the output optimized global model includes the following steps: Extract the risk prediction vector as the teacher signal and calculate the AUC value of each local model as the contribution weight; Aggregate global model parameters by contribution weights to generate an initial global model; Based on the JS divergence optimization parameters of the student model and the teacher signal, the teacher model is updated through the sliding average strategy; The regularization coefficient and contribution aggregation weight are dynamically adjusted according to the AUC performance index to generate an optimized global model.

[0014] In a second aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the risk assessment stabilization method for detecting chronic diseases in the elderly as described in the first aspect of the present invention.

[0015] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the risk assessment stabilization method for detecting chronic diseases in the elderly as described in the first aspect of the present invention.

[0016] The beneficial effects of the present invention are as follows: through the federated dynamic time warping algorithm, a unified time axis benchmark is constructed in the encrypted domain, and the alternating direction multiplier method is used to optimize the time warping operator to achieve nonlinear phase compensation of cross-institutional time series, thereby reducing the global alignment error of the three-dimensional time series matrix; in the construction of the dynamic knowledge graph, the time sensitivity weight of the entity relationship is quantified by the exponential decay function, and the spatiotemporal enhanced embedding matrix is ​​generated in combination with the cross-modal attention mechanism, thereby improving the temporal discriminability of the association strength of concurrent diseases; based on the variational inference loss function of the credibility score, the deviation of the variational distribution from the standard normal prior is dynamically constrained by the KL divergence, thereby suppressing the prediction interference of low-confidence patients and improving the stability of the model AUC indicator. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 Flowchart of a stable risk assessment method for chronic disease detection in the elderly.

[0019] Figure 2 Generate a schematic diagram for the joint feature space mapping matrix.

[0020] Figure 3 Constructing a schematic for a dynamic knowledge graph.

[0021] Figure 4 Schematic diagram of dual-path hypergraph convolution. DETAILED DESCRIPTION

[0022] In order to make the above objectives, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0023] In the following description, a lot of specific details are set forth in order to provide a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the concept of the present application, therefore the present application is not limited to the specific embodiments disclosed below.

[0024] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an embodiment that is separate or alternative to other embodiments.

[0025] Reference Figure 1 For one embodiment of the present application, the embodiment provides a risk assessment and stabilization method for detection of chronic diseases in the elderly, comprising the following steps: S1, performing federated dynamic time warping processing on encrypted patient data streams of multiple medical institutions, generating a joint feature space mapping matrix and obtaining a standard feature tensor through homomorphic encryption, please refer to Figure 2 , specifically comprising the following steps: A bidirectional authentication SSL / TLS communication channel is configured between the medical institutions and the data processing center, the medical institutions perform identity authentication on the transmission channel through pre-assigned encryption certificates, and generate an encrypted transmission link; encrypted data stream data packets from each medical institution are received through the encrypted transmission link, the data packet encapsulation format conforms to the HL7 FHIR standard, and contains a timestamp field, a patient anonymous identifier and a physiological indicator value sequence.

[0026] SHA-256 hash check is performed on the received encrypted data stream data packets, the hash value of the data packet content is calculated and compared with the attached digital signature, and the encrypted data stream data packets with matching hash values and not tampered are screened; the homomorphic encryption private key is used to decrypt the encrypted data stream data packets that pass the verification, and the original data stream containing time sequence markers is restored, each record in the data stream contains a standard timestamp and corresponding physiological indicator measurement value.

[0027] The decrypted raw data stream is parsed according to the preset field delimiter, and the timestamp sequence and the corresponding physiological indicator values ​​such as blood pressure, blood sugar and blood oxygen saturation are extracted to generate a structured data table containing patient number, timestamp, indicator code and measurement value; the structured data table is arranged in ascending order by timestamp, and the multi-dimensional physiological indicator measurement values ​​under continuous timestamps are converted into a two-dimensional numerical matrix for each patient, where the matrix row index corresponds to the time point serial number and the column index corresponds to the physiological indicator category of the disease classification code.

[0028] The timestamp metadata defined in the HL7 FHIR standard is parsed from the column header information of the two-dimensional numerical matrix, and the precise time mark corresponding to each time point is extracted to generate a timestamp index list. The list length is consistent with the number of rows in the two-dimensional numerical matrix. According to the disease classification code corresponding to the column index of the two-dimensional numerical matrix, physiological indicator categories such as blood pressure, blood sugar, and blood oxygen saturation are mapped to independent coding fields to form a comparison table between indicator codes and matrix column numbers.

[0029] Physiological indicator categories are traversed in column order of a two-dimensional numerical matrix, and the entire column of data is extracted for each indicator category to generate an independent time series vector with the same length as the timestamp index list. The timestamp index lists in the data streams of different medical institutions are compared, and a unified time axis benchmark is established with the minimum time granularity as the unit. Linear interpolation is performed on independent time series vectors with sampling time deviations to generate a corrected sequence with strictly aligned timestamps.

[0030] Check whether the timestamp intervals in the corrected independent time series vectors meet the sampling frequency requirements defined by the HL7 FHIR standard, and remove abnormal sequences with consecutive missing timestamps exceeding the maximum allowed number of consecutive missing timestamps. Reassemble the aligned and verified independent time series vectors according to the original ICD-11 coding order to construct a three-dimensional time series matrix.

[0031] To further illustrate, the first dimension of the three-dimensional time series matrix corresponds to the medical institution number, the second dimension corresponds to the timestamp index, and the third dimension corresponds to the physiological indicator category code.

[0032] The time window length of federated dynamic time warping is set to twice the maximum allowed sampling interval defined by the HL7 FHIR standard, the Hilbert-Schmidt regularization coefficient is defined to be 0.1, and the time warping operator is initialized to the identity matrix.

[0033] Time series segments of each medical institution under a unified time axis benchmark were extracted from the three-dimensional time series matrix. The segment length was equal to the time window length set initially, and the segment interval was the minimum sampling interval specified by the HL7 FHIR standard. The Frobenius norm difference of physiological indicators at adjacent time points within each time series segment was calculated to generate a difference matrix set. The dimension of the difference matrix was consistent with the number of medical institutions in the three-dimensional time series matrix.

[0034] Based on the difference matrix set, the time warping operator is iteratively updated by the alternating direction multiplication method, so that the time warping operator can simultaneously satisfy the minimization of the difference between adjacent time points and the Hilbert-Schmidt norm constraint. The iteration termination condition is that the difference between two adjacent time warping operator updates is less than the convergence threshold, and the optimized time warping operator is obtained.

[0035] To further explain, the convergence threshold is used to determine whether the update of the time warping operator has reached a converged state. It is defined according to the typical machine precision defined by the IEEE 754 standard. It can effectively distinguish between computational errors and actual parameter changes, and avoid misjudgment of convergence due to floating-point rounding errors.

[0036] The optimized time warping operator is applied to the time axis dimension of the three-dimensional time series matrix to perform phase shift compensation on the time series of medical institutions with sampling delays. The shift amount is determined by the corresponding element value of the unit matrix of the time warping operator; the three-dimensional time series matrix that has undergone phase compensation is expanded according to the patient dimension and reconstructed into a two-dimensional feature tensor. The row dimension of the tensor corresponds to the combined index of the patient number and timestamp, and the column dimension corresponds to the physiological indicator category of the disease classification code.

[0037] The two-dimensional feature tensor is split into N sub-matrices according to the patient number dimension. Each sub-matrix corresponds to the feature data of a single medical institution. The number of rows in the sub-matrix is ​​the number of patients corresponding to the single medical institution multiplied by the number of timestamps, and the number of columns is the number of physiological indicator categories coded by the disease classification. The row-wise covariance matrix is ​​calculated for the feature sub-matrix of each medical institution. The dimension of the covariance matrix is ​​the number of physiological indicator categories multiplied by the number of physiological indicator categories. The matrix elements reflect the statistical correlation between different physiological indicators.

[0038] The covariance matrix was decomposed into eigenvalues, and the eigenvectors corresponding to the first d largest eigenvalues ​​were selected to form a projection matrix. The number of rows in the projection matrix was the number of physiological indicator categories, and the number of columns was d feature dimensions. The column vectors of the projection matrix were normalized using the L2 norm so that the module length of the projection vector of each feature dimension was 1, eliminating the scale deviation caused by the difference in the amount of data among medical institutions. The standardized projection matrices of N medical institutions were arranged in order of the medical institution numbers to form a three-dimensional initial mapping matrix set. The first dimension of the set was the medical institution number, the second dimension was the number of physiological indicator categories, and the third dimension was d feature dimensions.

[0039] The Paillier homomorphic encryption algorithm is used to generate a 2048-bit public-private key pair. The public key is distributed to each medical institution, and the private key is stored in the hardware security module of the central coordinator. Each medical institution uses the public key to encrypt its own projection matrix in the initial mapping matrix set element by element, and calculates the weight coefficient based on the proportion of patients in each medical institution in the two-dimensional feature tensor. In the encrypted state, a weighted sum operation is performed on each corresponding element of the N encrypted projection matrices. The set of weight coefficients corresponds one-to-one to the medical institution number dimension to generate an encrypted joint mapping matrix.

[0040] Laplace noise is superimposed on each element value of the encrypted joint mapping matrix. The amount of noise is controlled by a preset privacy budget parameter, and the noise amplitude is inversely proportional to the absolute value of the matrix element. The noisy encrypted joint mapping matrix is ​​rearranged according to the physiological indicator category dimension and the feature dimension to form a final encrypted joint mapping matrix with the number of rows equal to the number of physiological indicator categories and the number of columns equal to the feature dimensions.

[0041] It is further explained that the privacy budget parameter is preset according to the medical data privacy regulations and the requirements of the federated learning protocol, and represents the maximum privacy leakage risk allowed in the entire federated learning process.

[0042] Read the Paillier homomorphic encryption private key from the hardware security module, load the final encrypted joint mapping matrix, define the privacy protection threshold, and initialize the iteration step size of the proximal operator; iteratively update each element of the joint mapping matrix in the encrypted state as follows: If the absolute value of an element exceeds the privacy protection threshold, it shrinks toward zero; if the element change rate exceeds the stability condition threshold, it is truncated according to the iteration step size to complete the proximal optimization; the encrypted joint mapping matrix that completes the proximal optimization is decrypted element by element using the Paillier private key to generate a floating-point numerical matrix, and the privacy budget parameter of the decrypted floating-point numerical matrix is ​​calculated and compared with the preset privacy budget parameter. When the calculated privacy budget parameter is greater than the preset privacy budget parameter, the noise superposition process is triggered to be re-executed; when the calculated privacy budget parameter is less than the preset privacy budget parameter, it indicates that the verification is passed, and the floating-point numerical matrix is ​​stored in a structure where rows correspond to physiological indicator categories and columns correspond to feature dimensions to generate a joint feature space mapping matrix.

[0043] It is further explained that the privacy protection threshold is set based on the data sensitivity classification and privacy budget parameter requirements to control the leakage of sensitive information; the stability condition threshold is set based on the convergence of the proximal optimization algorithm, numerical stability and federated communication constraints to ensure the stability of training.

[0044] According to the medical number dimension in the three-dimensional time series matrix of federated dynamic time warping, the patient time series data is decomposed into N independent time series sub-matrices, each of which contains all the patient time series data of a single institution. Matrix multiplication is performed on the time series sub-matrix of each medical institution and the joint feature space mapping matrix. The projection calculation formula is: time series sub-matrix (patient × time × indicator) multiplied by the mapping matrix (indicator × feature) to generate an intermediate projection matrix (patient × time × feature). The intermediate projection matrices of N medical institutions are vertically spliced ​​according to the patient number dimension to form a three-dimensional tensor structure. Check whether the number of patients in the three-dimensional tensor is consistent with the total number of original data records, whether the number of timestamp indexes matches the unified time axis reference length, and whether the number of feature dimensions is equal to the number of columns in the joint feature space mapping matrix. The verified three-dimensional tensors are sorted and stored in ascending order of patient number and timestamp index to generate a standard feature tensor, completing data federation alignment and feature space conversion.

[0045] S2, inputs the standard feature tensor into the dynamic medical knowledge graph construction module, and generates the knowledge graph embedding matrix through the cross-modal attention mechanism with enhanced spatiotemporal sensitivity, see Figure 3 , specifically including the following steps: Disease classification codes are extracted from the column dimension of the standard feature tensor, and each code is mapped to a clinical entity name, including disease names such as hypertension, diabetes, and chronic kidney disease, to generate a disease entity set. Based on the disease entity set, the associated symptoms and signs defined in the clinical guidelines are queried. For example, hypertension corresponds to headache and dizziness, and diabetes corresponds to polyuria and thirst. The symptoms and signs are added to the entity set to form a composite entity set containing diseases and symptoms.

[0046] Identify drug-related monitoring indicators from the physiological indicator categories of the standard feature tensor and map them to the names of active ingredients according to the drug instructions. For example, blood sugar indicators are associated with insulin and metformin, and blood pressure indicators are associated with amlodipine. Add the drug ingredients to the composite entity set.

[0047] Based on the treatment options recommended by clinical guidelines, triples in the form of (disease entity, treatment relationship, drug ingredient) are constructed, such as (diabetes, drug treatment, insulin) and (hypertension, drug treatment, amlodipine), to generate the initial treatment relationship.

[0048] Analyze the time series data of patients with multiple diseases co-occurring in the standard feature tensor, count the co-occurrence frequencies between disease entities, and generate (disease A, concurrent relationship, disease B) triples when the co-occurrence frequency exceeds the minimum disease co-occurrence frequency threshold, such as (diabetes, concurrent, hypertension), to form a concurrent relationship.

[0049] It is further explained that the minimum disease co-occurrence frequency threshold is a quantitative standard determined by clinical guidelines, statistical tests and data distribution. It is used to generate reliable disease co-occurrence relationships in the knowledge graph and is the core parameter for disease association mining in medical federated learning.

[0050] According to the contraindication description in the drug instructions, a (drug ingredient, contraindication relationship, disease / symptom) triple is constructed, such as (insulin, contraindication, hypoglycemia) and (amlodipine, contraindication, severe liver dysfunction), to generate a contraindication relationship.

[0051] Check whether the disease entities, symptoms and signs, and drug ingredients in the triple set exist in the entity set, remove triples containing undefined entities, and output a complete medical knowledge graph triple set.

[0052] The composite entity set is used as the node and the triple set as the labeled edge to form a directed graph structure. The node types in the graph include disease entities, symptom weight and drug ingredients, and the edge types include treatment relationships, concurrent relationships and contraindication relationships, completing the initial construction of the medical knowledge graph.

[0053] The time series of physiological indicators of each patient in the standard feature tensor is matched with the triples of the dynamic medical knowledge graph, and a timestamp is added to each edge to mark the time period when the corresponding relationship appears in the corresponding patient data, thereby generating a dynamic medical knowledge graph with time series attributes.

[0054] Read a set of triples with timestamps from a dynamic medical knowledge graph with time series attributes, extract the patient number and corresponding time period associated with each triple, intercept the physiological indicator time series fragments within the corresponding time period in the standard feature tensor, and generate a set of four-tuples (head entity, relationship, tail entity, time series fragment).

[0055] Check whether the length of the time series segment in the quadruple meets the minimum sampling number requirement defined by the HL7 FHIR standard, and remove triples with insufficient number of time points to generate a valid basic triple set.

[0056] A timestamp interval tag is attached to each triple in the valid basic triple set. The tag format is a string pair of a start timestamp and an end timestamp. The current processing timestamp is read. For each triple in the valid basic triple set with a timestamp tag, the number of days between the midpoint of the timestamp interval and the current timestamp is calculated. The time interval is input into an exponential decay function to calculate a time sensitivity weight value. The weight value range is constrained between 0 and 1. The shorter the time interval, the higher the weight value.

[0057] The time sensitivity weight value is appended as a new attribute to the valid basic triple set to generate a triple set with weight labeling. Each element in the set contains (head entity, relationship, tail entity, time series segment, time sensitivity weight).

[0058] All disease entities, symptom and sign entities, and drug ingredient entities are extracted from the composite entity set, classified by entity type, and a globally unique digital index is generated to form a mapping table from entity name to index number.

[0059] Create an all-zero matrix with the number of rows and columns equal to the total number of composite entity sets. The matrix data type is floating point and is used to store the spatiotemporal relationship weight values. Traverse the set of triples marked with weights, locate the matrix row and column positions according to the index numbers of the head entity and the tail entity, write the time-sensitive weight values ​​into the corresponding all-zero matrix elements, and take the maximum weight value if there are multiple relationships for the same entity pair. Check whether the all-zero matrix meets the symmetry requirements of the undirected graph, force the edges of the treatment relationship and concurrent relationship types to be symmetric, keep the taboo relationship unidirectional, and output the spatiotemporal adjacency matrix.

[0060] Determine the feature dimension of the standard feature tensor as d, preset the embedding dimension of the dynamic medical knowledge graph with time series attributes as k, and create two trainable parameter matrices: The query weight matrix has d rows and k columns, the key weight matrix has d rows and k columns, and the Xavier normal distribution initialization method is used to generate initial values.

[0061] Marking the initial query weight matrix and the initial key weight matrix as trainable parameters allows the matrix element values ​​to be updated by the gradient descent algorithm during backpropagation.

[0062] The initial query weight matrix and the initial key weight matrix are stored in the memory buffer. The matrix memory layout is arranged in row-major order, ready for cross-modal feature conversion.

[0063] Extract the feature vector of a single patient from the standard feature tensor according to the patient number dimension, perform matrix multiplication on the feature vector of a single patient and the query weight matrix in the memory buffer to generate a patient query vector; extract the entity vector from the dynamic medical knowledge graph, perform matrix multiplication on the entity vector and the key weight matrix in the memory buffer to generate an entity key vector.

[0064] Traverse all combinations of patient query vectors and entity key vectors, calculate the cosine similarity between each pair of vectors, and constrain the similarity value range to be between -1 and 1 to generate a patient-entity similarity matrix; expand the patient-entity similarity matrix according to the patient number dimension and reshape it into a matrix structure with the number of rows being the number of patients and the number of columns being the number of entities. Each element represents the strength of the association between a specific patient and an entity, and is mapped to the [0,1] interval using the Sigmoid function to eliminate the influence of negative values ​​and generate a standardized attention score matrix.

[0065] Load the spatiotemporal adjacency matrix, where the number of rows and columns in the matrix is ​​equal to the number of entities, and the element value is the time sensitivity weight. Multiply the spatiotemporal adjacency matrix and the standardized attention score matrix element by element at the same row and column positions. The position with zero weight in the spatiotemporal adjacency matrix will suppress the corresponding attention score, generating a spatiotemporal enhanced attention score matrix.

[0066] The original entity vectors are extracted from the dynamic medical knowledge graph, and the spatiotemporal enhanced attention score matrix is ​​multiplied with the original entity vectors to generate the knowledge graph embedding matrix.

[0067] To further illustrate, the knowledge graph embedding matrix is ​​converted into a binary data stream in the order of entity index, with the first dimension of the matrix being the entity index number and the second dimension being the embedding dimension value; the binary data stream is written into a distributed file, with each entity index corresponding to an independent storage block, supporting fast retrieval of embedding vectors by entity number, completing the collaborative embedding of dynamic medical knowledge graph and patient characteristics, and generating a knowledge graph embedding matrix.

[0068] S3, fuses the standard feature tensor with the knowledge graph embedding matrix, builds a dynamic hypergraph structure and performs two-pass hypergraph convolution calculations to output the patient-knowledge joint embedding matrix, see Figure 4 , specifically including the following steps: For each pair of patient feature vectors in the standard feature tensor, combined with the entity vectors in the knowledge graph embedding matrix, a triple interaction operation is performed to obtain the triple interaction score, which is expressed as: ; in, is the triple interaction score, and is the patient feature vector, corresponding to the patient number and The feature representation of and Patient index (such as patient 1, patient 2), the value range is 1≤ , ≤ ,in is the total number of patients, an entity vector in a knowledge graph embedding matrix, an index variable for embedding vector dimension, a transpose of a patient feature vector.

[0069] reading a current processing timestamp, selecting the 95th percentile as an initial threshold value based on a triple interaction score distribution of historical patient data, calculating a dynamic threshold value according to the initial threshold value and a decay rate, the expression being: ; wherein, the dynamic threshold value, the initial threshold value, the decay rate, a normalized time difference between the current processing timestamp and an initial timestamp.

[0070] comparing the triple interaction score with the dynamic threshold value, marking a patient pair with a score higher than the dynamic threshold value as an effective connection to generate an effective connection marker set; arranging in ascending order according to patient number, double-loop traversing all patient numbers, if a patient pair ( , ) exists in the effective connection marker set, creating a hyperedge record and recording the patient number and the index position of the associated entity in the knowledge graph embedding matrix as well as the current processing timestamp to generate an original hyperedge set.

[0071] checking whether there is a record with the same patient number and entity index in the original hyperedge set, retaining the hyperedge record with the latest timestamp and eliminating duplicates to generate a de-duplicated hyperedge set, and appending a timestamp interval marker to each hyperedge in the hyperedge set, the start timestamp being the current processing timestamp minus the minimum sampling interval defined in the HL7 FHIR standard, and the end timestamp being the current processing timestamp, to generate a dynamic hyperedge set with timestamp markers.

[0072] based on the number of hyperedges in the dynamic hyperedge set, creating a patient-hyperedge all-zero association matrix and an all-zero diagonal matrix simultaneously; traversing each hyperedge in the dynamic hyperedge set, if the patient participates in the hyperedge, marking the all-zero association matrix as 1, otherwise keeping it as 0; filling the timestamp marker of the hyperedge with the current processing timestamp into the all-zero diagonal matrix to obtain a diagonal matrix with time weight.

[0073] checking whether the number of columns of the patient-hyperedge association matrix is equal to the number of rows of the weight diagonal matrix, if equal, storing the patient-hyperedge association matrix and the diagonal matrix with time weight into the memory buffer; if not equal, triggering the regeneration of the dynamic hyperedge set until the matrix dimensions are consistent.

[0074] Read the standard feature tensor from the memory buffer and perform a slice operation along the patient number dimension. Each slice extracts all time points and physiological indicator features of a single patient to generate multiple patient feature slices. The dimension of each slice is the number of time points multiplied by the number of physiological indicator categories, forming the original patient feature slice set.

[0075] Principal component analysis is performed on each slice in the original patient feature slice set, and the top d principal components with the highest variance contribution are retained. The dimension of each slice is reduced from the number of time points multiplied by the number of physiological indicator categories to d dimensions. The reduced dimension patient feature vector set is generated and arranged in ascending order by patient number to form the initial hidden representation matrix.

[0076] Define the current layer counter l=0, the maximum layer number L=3, read the patient-hyperedge association matrix and hyperedge weight matrix from the memory buffer; set the termination condition for the hidden representation update to be that the layer counter reaches L or the cosine similarity of the adjacent layer representation exceeds 0.99, and prepare to start the hypergraph convolution calculation; write the initial hidden representation matrix into the memory buffer in row-first order, with each row corresponding to a patient number.

[0077] Read the patient-hyperedge association matrix from the memory buffer and calculate the patient degree matrix value. The expression is: ; in, is the patient degree matrix value, is the number index of the hyperedge, is the total number of hyperedges, is a binary indicator function, indicating that the patient With super edge The relationship between them.

[0078] The diagonal elements are the sum of the elements in each row of the patient degree matrix, and the off-diagonal elements are 0, indicating the number of hyperedges in which each patient participates. The inverse square root of each diagonal element of the patient degree matrix is ​​taken to generate a normalized matrix. Chain matrix multiplication is performed on the normalized matrix, the patient-hyperedge association matrix, and the inverse matrix of the hyperedge weight matrix to obtain the intermediate representation matrix after diffusion. The intermediate representation matrix is ​​multiplied with the initial hidden representation matrix to generate a normalized diffusion result matrix.

[0079] Read the trainable parameter matrix from the parameter storage area, perform matrix multiplication on the normalized diffusion result matrix and the trainable parameter matrix, apply the ReLU activation function to the product result, set the negative values ​​to zero, and obtain the hypergraph convolution path output.

[0080] The relationship weights between entities are extracted from the knowledge graph embedding matrix and an adjacency matrix is ​​constructed, where the elements of the adjacency matrix represent the relationship weights between entities. If there is no direct relationship between two entities, the weight is 0.

[0081] Each patient hidden representation vector in the initial hidden representation matrix is ​​concatenated with the entity vector in the knowledge graph embedding matrix. The dot product operation is performed on the concatenated vector and the vector in the trainable parameter matrix. The vector is activated by the LeakyReLU function and normalized by the neighboring nodes using the Softmax function to generate an attention coefficient, which represents the patient's attention weight on the entity.

[0082] The trainable parameter matrix is ​​divided into four independent heads, and the attention aggregation results of each head are calculated separately; the ELU activation function is applied to the results of each head, and the outputs of the four heads are spliced ​​in the feature dimension to obtain the graph attention path output.

[0083] The output of the hypergraph convolutional path and the output of the graph attention path are linearly added to generate a fused representation matrix; layer normalization is performed on the fused representation matrix to preserve the stability of the data distribution; the normalized result is residually connected with the current hidden representation matrix to generate an updated hidden representation matrix.

[0084] Check whether the number of rows of the updated hidden representation matrix is ​​equal to the number of patients and whether the number of columns is equal to the embedding dimension; if the dimensions are inconsistent, trigger the recalculation process; if they are consistent, generate the patient-knowledge joint embedding matrix and store it in the memory buffer.

[0085] S4 calculates the credibility score based on the joint embedding matrix, optimizes the credibility weighted loss function in combination with variational inference, and generates a weighted risk prediction vector. Specifically, it includes the following steps: Read the patient-knowledge joint embedding matrix from the memory buffer. The number of rows in the matrix is ​​the number of patients, the number of columns is the embedding dimension, and each row corresponds to the embedding vector of a patient. Query the patient's disease category label (such as diabetes, hypertension) from the dynamic medical knowledge graph based on the patient number to form a patient-disease category mapping table. Traverse the patient-disease category mapping table and, for each patient, filter all patient embedding vectors with the same disease category label in the patient-knowledge joint embedding matrix to form a set of patients of the same type. For all embedding vectors in the set of patients of the same type, calculate the mean vector by column. The expression is: ; in, is the mean vector of the same type of patient set, indicating that the disease category is The average of all patient embedding vectors, For patients A collection of similar patients of the same disease category (such as diabetes, hypertension), including all patients Patient numbers with the same disease label, For patients The embedding vector of .

[0086] For the patient's embedding vector, the square of the Euclidean distance to the mean vector is calculated, and the expression is: ; in, For patients The embedding vector value of is the embedding vector dimension, For patients The embedding vector of The value of the dimension, is the mean vector In the The value of the dimension.

[0087] Traverse all patients in the same patient set and calculate the maximum square of the Euclidean distance between each patient's embedding vector and the mean vector. The expression is: ; The credibility score is generated by the ratio of the square of the Euclidean distance to the maximum value, and the expression is: ; in, For patients The credibility score of , the value range is [0, 1].

[0088] Extract the patient's embedding vector from the patient-knowledge joint embedding matrix, and directly copy the embedding value of the patient's embedding vector as the mean vector of the variational distribution; input the patient's embedding vector into a mapping network consisting of a fully connected layer and a nonlinear activation function to generate an intermediate vector; input the intermediate vector into another fully connected layer without an activation function, output the variance vector of the variational distribution, and store the mean vector and variance vector of the patient's variational distribution in the parameter storage area in order of patient number; Define the prior distribution and assign fixed parameters to the standard normal distribution, including the mean vector and covariance matrix. The standard normal distribution does not depend on patient data and only serves as the optimization target of the variational distribution. Each medical institution extracts the patient's true risk label from the local database to form a true risk label vector. Each element of the true risk label vector corresponds to a row in the patient-knowledge joint embedding matrix. The patient-knowledge joint embedding matrix is ​​input into the Sigmoid function, and the predicted risk value of each patient is calculated row by row. All predicted values ​​constitute a predicted value vector, which is expressed as: ; in, For patients The predicted risk value ranges from [0, 1].

[0089] Based on the true risk label vector and the predicted value vector, the binary cross entropy loss term for each patient is calculated as: ; in, is the binary cross entropy loss between the true label vector and the predicted value vector, For patients The true risk label vector.

[0090] Multiply the binary cross entropy loss term by the credibility score element-wise to obtain the weighted cross entropy loss term; Extract the mean vector and variance vector of the patient's variational distribution from the parameter matrix, and calculate the KL divergence between the variational distribution and the prior distribution. The expression is: ; in, is the variational distribution With prior distribution The KL divergence of is the variational distribution In the embedding dimension The variance vector of is the variational distribution In the embedding dimension The mean vector of .

[0091] Further explanation: The KL divergence term is used to constrain the variational distribution to be close to the standard normal distribution.

[0092] It should be noted that the expression of variational distribution is: ; in, For patients The variational distribution of the embedding vector, is the mean vector, is the variance vector, is the symbol of the normal distribution.

[0093] It is further explained that because the variational distribution is used to approximate the posterior distribution of the patient's embedding vector, by assuming a normal distribution, the mean vector and variance vector can be used for efficient optimization, and the KL divergence with the prior distribution can be conveniently calculated.

[0094] The expression of the prior distribution is: ; in, For patients The prior distribution of the embedding vector is the standard normal distribution.

[0095] The weighted cross-entropy loss term and the KL divergence term are added by the total loss function by the patient number to obtain the total loss value, and the expression is: ; wherein, is the total loss value, and is a scalar, is a regularization coefficient, indicating a scalar hyperparameter, used to balance the weights of cross-entropy loss and KL divergence, and prevent overfitting.

[0096] The trainable parameters of the variational inference layer are extracted from the parameter storage area, including the mean vector and variance vector of the patient variational distribution and the patient-knowledge joint embedding matrix.

[0097] Based on the total loss value, the node starts, and the local gradient is calculated layer by layer along the reverse path, as follows: The gradient of the weighted cross-entropy loss term with respect to the predicted risk value is calculated, and the gradient amplitude is adjusted in combination with the credibility score. The gradient of the KL divergence term with respect to the variational mean vector and the variational variance vector is calculated, and the gradient is passed to the patient-knowledge joint embedding matrix by the chain rule according to the KL divergence formula. The gradient matrix of the patient-knowledge joint embedding matrix is generated, and the gradient storage area is written according to the parameter dimension.

[0098] The Adam optimizer is used to update the parameters based on the gradient matrix in the gradient storage area, as follows: The first moment estimate and second moment estimate variables of the Adam optimizer are initialized to all-zero vectors with the same dimension as the parameters to be updated; The first moment estimate and second moment estimate are corrected according to the current gradient value; The learning rate is 0.001, and the moment estimate is corrected by the bias. The parameter update amount is calculated, and the patient-knowledge joint embedding matrix and the mean vector and variance vector in the variational parameter are updated.

[0099] After each round of parameter update, the value of the KL divergence term is extracted from the total loss value, and compared with the KL divergence convergence threshold: If the KL divergence term exceeds the KL divergence convergence threshold, the regularization coefficient is attenuated in proportion (such as the new regularization coefficient ), reducing the weight of KL divergence in the total loss; If the KL divergence term does not exceed the KL divergence convergence threshold, the remains unchanged.

[0100] Further explanation, the KL divergence convergence threshold is a hyperparameter set based on the variational inference theory, experimental calibration and field consensus, used to dynamically control the attenuation logic of the regularization coefficient, to ensure that the federated learning model balances between privacy protection and prediction performance.

[0101] The updated patient-knowledge joint embedding matrix is ​​read from the parameter storage area and input into the Sigmoid function row by row to generate an updated prediction value vector. The credibility scores are filled into a vector in numerical order to generate a credibility score vector, and the credibility score vector is multiplied element-by-element with the updated prediction value vector to generate a weighted risk prediction vector, which is used to suppress the prediction value of low-credibility patients and enhance the prediction contribution of high-credibility patients.

[0102] S5 uses the risk prediction vector as the teacher signal and outputs the optimized global model through contribution-weighted federated parameter aggregation and dynamic knowledge distillation. The specific steps include: The patient's predicted value is extracted from the weighted risk prediction vector as the teacher signal to form a teacher signal vector. The local model parameters of each medical institution are read from the parameter storage area. Based on the patient's true risk label vector and the teacher signal vector, the AUC performance index of each local model is calculated. The AUC value reflects the discriminative ability of the local model in patient risk assessment. The higher the value, the greater the contribution. The expression is: ; in, AUC is a core performance indicator for evaluating the ability of the model to distinguish between positive and negative samples. The higher the AUC value, the stronger the model's ability to distinguish between positive samples (such as high-risk patients) and negative samples (such as low-risk patients). is the total number of true high-risk patients (positive samples), is the total number of true low-risk patients (negative samples), is the index of the positive sample, is the index of the negative sample, For the The teacher signal vector of positive samples (high-risk patients), For the The teacher signal vector of negative samples (low-risk patients), Positive samples for local models The predicted value of Negative samples for local models The predicted value of is an indicator function, which has a value of 1 if the condition in the brackets is true, and 0 otherwise.

[0103] Based on the AUC values ​​of the local models of each medical institution, the global model parameters are aggregated according to the following rules: The AUC value Power (such as =2) as contribution weight; The parameters of each local model are weighted and summed, with the weight being the ratio of the contribution weight of the corresponding medical institution to the total contribution. The weighted summation result is written into the global model parameter storage area of ​​each medical institution, overwriting the original parameters.

[0104] Further explanation: is an exponential coefficient, which is used to adjust the degree of magnification of the AUC value on the contribution weight. The second power strengthens the weight of high AUC institutions and suppresses the influence of low AUC institutions.

[0105] The constructed variational inference layer parameters and hypergraph convolution parameters are used as the initial parameters of the student model to ensure that the student model structure is consistent with the teacher model. The student model parameters are stored in an independent storage area and isolated from the teacher model parameters.

[0106] Using the teacher signal vector as the soft label, the student model generates the student model prediction value based on the joint patient-knowledge embedding matrix.

[0107] Calculate the JS divergence between the teacher signal vector and the student model prediction value to measure the difference in their distribution. The expression is: ; in, is the probability distribution of the teacher signal vector Probability distribution of the predicted values ​​with the student model JS divergence, is the probability distribution of the teacher signal vector (soft label), is the probability distribution of the student model’s predicted values, is the middle distribution between teachers and students, is the KL divergence between the probability distribution of the teacher signal vector and the intermediate distribution between the teacher and student distributions, It is the KL divergence between the probability distribution of the student model prediction value and the intermediate distribution.

[0108] Based on the calculated JS divergence, the student model parameters are optimized through back propagation to minimize the JS divergence loss.

[0109] The sliding average strategy is used to update the teacher model parameters, as follows: Read the updated parameters from the student model parameter storage area and update the teacher model parameters. For example: Teacher model parameters = 0.9 × original teacher parameters + 0.1 × student parameters; Overwrite the updated teacher model parameters into the teacher model parameter storage area.

[0110] Based on the patient's true risk label vector and the teacher model's prediction value, recalculate the AUC value of the global model: If the AUC value is improved compared with the previous round, keep the current parameters; If the AUC value decreases or remains the same, roll back to the previous round of parameters and increase the contribution aggregation value; The final updated teacher model parameters are merged with the global model parameters to generate an optimized global model, which is written into the federated model storage area of ​​each medical institution for call by each medical institution.

[0111] This embodiment also provides a computer device, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the risk assessment stabilization method for chronic disease detection in the elderly proposed in the above embodiment.

[0112] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0113] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the risk assessment stabilization method for detecting chronic diseases in the elderly proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0114] In summary, the present invention constructs a unified time axis benchmark in the encrypted domain through a federated dynamic time warping algorithm, optimizes the time warping operator using the alternating direction multiplier method, realizes nonlinear phase compensation of cross-institutional time series, and reduces the global alignment error of the three-dimensional time series matrix; in the construction of the dynamic knowledge graph, the time sensitivity weight of the entity relationship is quantified by an exponential decay function, and a spatiotemporal enhanced embedding matrix is ​​generated in combination with the cross-modal attention mechanism, so that the temporal discriminability of the association strength of concurrent diseases is improved; the variational inference loss function based on the credibility score dynamically constrains the deviation of the variational distribution from the standard normal prior through the KL divergence, suppresses the prediction interference of low-confidence patients, and improves the stability of the model AUC indicator.

[0115] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A stable risk assessment method for elderly chronic disease detection, characterized by: include: Perform federated dynamic time warping on encrypted patient data streams from multiple medical institutions, generate a joint feature space mapping matrix, and obtain the standard feature tensor through homomorphic encryption; Inputting the standard feature tensor into a dynamic medical knowledge graph construction module, and generating a knowledge graph embedding matrix through a cross-modal attention mechanism with enhanced spatiotemporal sensitivity; Fusing the standard feature tensor with the knowledge graph embedding matrix, constructing a dynamic hypergraph structure and performing a dual-path hypergraph convolution calculation to output a joint patient-knowledge embedding matrix; Calculating a credibility score based on the joint embedding matrix, optimizing a credibility weighted loss function in combination with variational inference, and generating a weighted risk prediction vector; The risk prediction vector is used as a teacher signal, and an optimized global model is output through contribution-weighted federated parameter aggregation and dynamic knowledge distillation.

2. The stable risk assessment method for detecting chronic diseases in the elderly according to claim 1, characterized in that: Generating the joint feature space mapping matrix comprises the following steps: Configure a two-way authenticated SSL / TLS communication channel between the medical institution and the data processing center to encapsulate and encrypt data stream packets; Perform hash verification and decryption on the received encrypted data stream to generate a structured data table; Convert the structured data table into a two-dimensional numerical matrix, traverse the physiological indicator categories in the order of the columns of the two-dimensional numerical matrix, and construct a three-dimensional time series matrix; The time warping operator is iteratively updated by the alternating direction multiplication method to perform phase shift compensation on the three-dimensional time series matrix and generate a joint feature space mapping matrix.

3. The stable risk assessment method for detecting chronic diseases in the elderly according to claim 1, characterized in that: Obtaining the standard feature tensor refers to performing weighted summation on the joint feature space mapping matrix and superimposing Laplace noise through the Paillier homomorphic encryption algorithm.

4. The stable risk assessment method for elderly chronic disease detection according to claim 1, characterized in that: The construction of the dynamic medical knowledge graph includes the following steps: Based on the disease classification coding of the standard feature tensor, the mapping generates a set of composite entities; Construct triple sets based on clinical guidelines and drug instructions; Based on the time series data of patients with multiple diseases co-occurring in the standard feature tensor, the co-occurrence frequency between disease entities is counted to generate a dynamic medical knowledge graph.

5. The stable risk assessment method for detecting chronic diseases in the elderly according to claim 1, characterized in that: Generating the knowledge graph embedding matrix includes the following steps: Match the dynamic medical knowledge graph with the patient's physiological indicator time series fragments to generate a set of four-tuples; The spatiotemporal adjacency matrix between entities is calculated by using time-sensitivity weights and exponential decay functions, and the knowledge graph embedding matrix is ​​generated by combining the cross-modal attention mechanism.

6. The stable risk assessment method for elderly chronic disease detection according to claim 1, characterized in that: The output of the patient-knowledge joint embedding matrix includes the following steps: Perform triple interaction operations based on the patient feature vector and the entity vector of the knowledge graph embedding matrix to calculate the triple interaction score; Filter the effectively connected patient pairs according to the dynamic threshold to generate the original hyperedge set; Create a patient-hyperedge correlation matrix and a diagonal matrix with time weights, and combine principal component analysis to reduce the dimension of patient feature slices to form the initial hidden representation matrix; Generate the hypergraph convolution pathway output through patient degree matrix normalization and hypergraph convolution pathway calculation; Based on the graph attention mechanism, the entity relationship weights of the knowledge graph embedding matrix are integrated to generate the graph attention path output; The hypergraph convolution pathway is fused with the output of the graph attention pathway to generate a joint patient-knowledge embedding matrix.

7. The stable risk assessment method for detecting chronic diseases in the elderly according to claim 1, characterized in that: Generating a weighted risk prediction vector comprises the following steps: Based on the patient-knowledge joint embedding matrix, the same set of patients is filtered by disease category label, and the mean vector of the embedding vectors of the same patients is calculated; Patient credibility scores were generated by the ratio of the squared Euclidean distance to the maximum value; Map the patient embedding vector to the mean vector and variance vector of the variational distribution, and define the standard normal distribution as the prior distribution; Combining the credibility score weighted binary cross entropy loss and KL divergence term, we construct a total loss function and optimize the variational parameters through backpropagation; Based on the updated patient-knowledge joint embedding matrix, a weighted risk prediction vector is generated.

8. The stable risk assessment method for elderly chronic disease detection according to claim 1, characterized in that: The output of the optimized global model comprises the following steps: Extract the risk prediction vector as the teacher signal and calculate the AUC value of each local model as the contribution weight; Aggregate global model parameters by contribution weights to generate an initial global model; Based on the JS divergence optimization parameters of the student model and the teacher signal, the teacher model is updated through the sliding average strategy; The regularization coefficient and contribution aggregation weight are dynamically adjusted according to the AUC performance index to generate an optimized global model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the risk assessment stabilization method for detecting chronic diseases in the elderly according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the risk assessment stabilization method for detecting chronic diseases in the elderly according to any one of claims 1 to 8 are implemented.

Images