An individualized disease AI auxiliary diagnosis system based on multi-modal physical examination data

Abstract

The application relates to the technical field of information communication, and particularly discloses an AI-assisted diagnosis system for individual diseases based on multi-modal physical examination data, which comprises the following modules: a feature processing module, which is used for converting multi-modal physical examination data into an initial feature vector; the initial feature vector is a feature representation of a node in Euclidean space; a feature analysis module, which is used for analyzing the initial feature vector to generate a hyperbolic node feature vector; a health state analysis module, which is used for determining a health state evaluation result based on the hyperbolic node feature vector; and a result output module, which is used for calculating and outputting a health deviation score of the multi-modal physical examination data based on the hyperbolic node feature vector and the hyperbolic geodesic distance of the corresponding node in a reference population, and taking the health deviation score and the health state evaluation result as a multi-dimensional diagnosis output result.

Description

Technical Field

[0001] This invention relates to the field of information and communication technology, and in particular to a personalized AI-assisted disease diagnosis system based on multimodal physical examination data. Background Technology

[0002] As health checkups evolve towards digitalization and intelligence, the process generates a large amount of data from diverse sources and with varying structures. Existing disease prediction or intelligent diagnostic methods mostly analyze single modalities or a few structured indicators, typically employing simple feature concatenation and ordinary machine learning models. This makes it difficult to fully uncover the potential complementary information and deep semantic relationships between different modalities of health checkup data. Especially in real-world health checkup scenarios, the occurrence and development of diseases are often not determined by a single indicator, but rather by a complex result of the combined effects of multiple factors, including physiological indicators, behavioral characteristics, psychological state, social environment, and past health information. Existing methods struggle to effectively characterize these high-order relationships involving multiple factors, multiple levels, and multiple granularities, limiting the accuracy, stability, and generalization ability of diagnostic results.

[0003] Therefore, this invention proposes a personalized AI-assisted disease diagnosis system based on multimodal physical examination data. Summary of the Invention

[0004] This invention provides a personalized AI-assisted disease diagnosis system based on multimodal physical examination data to address the technical problems in existing technologies, such as the difficulty in effectively integrating and accurately representing multi-source heterogeneous physical examination data within a unified framework; the difficulty in accurately modeling high-order correlations and potential hierarchical structures among physical examination indicators; and the insufficient robustness and accuracy of diagnostic results in real-world physical examination scenarios with missing data and noise interference.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A personalized AI-assisted disease diagnosis system based on multimodal physical examination data includes: The feature processing module is used to convert multimodal physical examination data into initial feature vectors. Multimodal physical examination data includes demographic information, questionnaire data, laboratory test indicators, vital sign data, physical measurement data, imaging examination results, and data from at least two different modalities in past medical history. The initial feature vectors are used to characterize the basic health information of at least two different modalities of data in Euclidean space and their differences in semantic hierarchy, statistical distribution, and feature dimensions. The initial feature vectors are the feature representations of nodes in Euclidean space. The feature analysis module is used to analyze the initial feature vectors and generate hyperbolic node feature vectors. The hyperbolic node feature vectors are used to characterize the comprehensive health status of multimodal physical examination data determined in a hyperbolic space with learnable curvature. The nodes are the examinees corresponding to the multimodal physical examination data. The health status analysis module is used to determine the health status assessment results based on hyperbolic node feature vectors; The results output module is used to calculate and output the health deviation score of multimodal physical examination data based on the hyperbolic feature vector of hyperbolic node and the hyperbolic geodesic distance of the corresponding node in the reference population. The health deviation score and the health status assessment results are used as multidimensional diagnostic output results.

[0006] Based on the aforementioned technical means, by projecting the initial feature vectors onto a hyperbolic space manifold with learnable curvature, the multi-scale hierarchical structure of multimodal physical examination data is presented with minimal geometric distortion; by dynamically constructing a hypergraph topology, high-order correlations between multiple physical examination indicators are effectively captured; global structural features and local high-frequency detail features are extracted by parallel spatial message passing stream and spectral transformation stream, respectively, and then geometrically gated adaptively fused to achieve a balance between hierarchical structure and local details; and by using learnable curvature parameters, each processing stage adaptively matches the inherent hierarchy of the data, maintaining robustness and accuracy even under conditions of missing data and noise interference.

[0007] Furthermore, the feature analysis module includes: Hyperbolic projection units are used to project the initial feature vectors onto a hyperbolic manifold with learnable curvature through exponential and logarithmic mappings, generating the initial representation of nodes in hyperbolic space; Hypergraph building unit is used to construct a hypergraph topology based on the initial representation of nodes in hyperbolic space, where each hyperedge in the hypergraph topology connects two or more nodes. The feature extraction unit is used to extract global structural features and local high-frequency detail features from the hypergraph topology through a parallel spatial message passing stream and a spectral transform stream based on graph wavelet transform, respectively. The geometric fusion unit is used to adaptively weight and fuse global structural features and local high-frequency detail features to generate hyperbolic node feature vectors.

[0008] Furthermore, the hyperbolic projection unit includes: Curvature parameter generation subunit is used to learn and generate curvature parameters based on semantic hierarchy differences; The logarithmic mapping subunit is used to project the initial feature vector from Euclidean space to hyperbolic space based on the curvature parameter, so as to obtain the initial representation of the node in hyperbolic space; The parameter adjustment subunit is used to determine the volume growth rate of the hyperbolic space based on the distribution density and curvature parameters of the initial representation in the hyperbolic space, and to perform geometric redistribution on the initial representation based on the volume growth rate to obtain the adjusted initial representation.

[0009] Furthermore, the hypergraph building unit includes: The neighborhood calculation subunit is used to calculate the hyperbolic geodesic distance between nodes based on the initial representation of the nodes in hyperbolic space, and to determine the preset number of nearest neighbor nodes for each node based on the hyperbolic geodesic distance. The hyperedge generation subunit is used to form hyperedges with each node and its corresponding neighboring nodes, thereby generating a hypergraph topology; the hypergraph topology evolves dynamically as the node representation is updated.

[0010] Furthermore, the feature extraction unit includes: Spatial message passing stream is used to aggregate node representations belonging to the same hyperedge along the node-to-hyperedge direction into a hyperedge representation using the hyperbolic Einstein midpoint algorithm, based on the hypergraph topology. Then, the hyperedge representation is propagated back to each node along the hyperedge-to-node direction to update the node representation and generate a node representation that encodes global structural features. The spectral transform stream is used to perform multi-scale spectral filtering on the initial representation of nodes in hyperbolic space through hyperbolic graph wavelet transform, and to extract node representations that encode local high-frequency detail features.

[0011] Furthermore, the health status analysis module includes: The distance calculation unit is used to calculate the hyperbolic geodesic distance between the feature vector of the hyperbolic node and the reference hyperbolic node with a preset known health category; the hyperbolic geodesic distance is the shortest path length of the hyperbolic node feature vector and the reference hyperbolic node along the geodesic in hyperbolic space with learnable curvature; The classification and determination unit is used to determine the health category to which the multimodal physical examination data belongs based on the hyperbolic geodesic distance, and generate health status assessment results.

[0012] Furthermore, the classification determination unit includes: The distance comparison subunit is used to compare the hyperbolic feature vector of the hyperbolic node with the hyperbolic geodesic distance between each reference hyperbolic node with a preset known health category, and to determine the health category corresponding to the reference hyperbolic node with the smallest distance. The category determination sub-unit is used to determine the health category corresponding to the reference hyperbolic node with the smallest distance as the health category to which the multimodal physical examination data belongs, and to generate health status assessment results.

[0013] Furthermore, it also includes a scoring calculation unit, which includes: The confidence assessment subunit is used to perform diagnostic classification on the hyperbolic node feature vector after multiple perturbations through multiple random inactivation forward propagation, and calculate the consistency ratio of each classification result as the confidence score of the health status assessment result. The deviation calculation subunit is used to normalize the distance into a health deviation score based on the hyperbolic geodesic distance between the hyperbolic node feature vector and the reference hyperbolic node; the multimodal physical examination data are positively correlated with the typical features and health deviation scores of known health categories.

[0014] Furthermore, the result output module includes: The deviation calculation unit is used to calculate the subject's group health deviation score based on the hyperbolic node feature vector and the hyperbolic geodesic distance between the corresponding node in the reference population; the group health deviation score represents the degree of deviation of the subject's overall health status from the normal state of the homogeneous reference population; The longitudinal deviation calculation unit is used to calculate the longitudinal health deviation score of the examinee based on the hyperbolic geodesic distance between the hyperbolic node feature vector and the examinee's historical hyperbolic node feature vector; the longitudinal health deviation score represents the extent of change of the examinee's current health status relative to their own historical baseline level; The multidimensional diagnostic output unit integrates health status assessment results, population health deviation scores, and longitudinal health deviation scores into multidimensional diagnostic output results, which are then presented in the form of a visual report.

[0015] The beneficial effects of this invention are as follows: (1) This invention projects the initial feature vector onto a hyperbolic space manifold with learnable curvature and utilizes the exponential volume growth characteristic of hyperbolic space to present a multi-scale hierarchical structure with minimal geometric distortion, effectively solving the problem of geometric feature distortion caused by traditional Euclidean space models when representing hierarchical physical examination data.

[0016] (2) This invention captures the high-order correlation between different modal health examination indicators by dynamically constructing a hypergraph topology structure, with each hyperedge connecting two or more nodes, effectively solving the problem that simple graph models cannot effectively capture the complex high-order relationship of multiple health examination indicators causing disease.

[0017] (3) This invention extracts global structural features and local high-frequency detail features by parallel running spatial message passing stream and spectral transformation stream, and performs adaptive weighted fusion through geometric fusion unit, thereby achieving a balance between global hierarchical structure and local geometric details, effectively solving the problems of "oversmoothing" of deep node features and insufficient extraction of local high-frequency boundary information in deep graph neural networks.

[0018] (4) The present invention assigns learnable curvature parameters to each information processing stage, enabling shallow layers to operate with a smaller curvature to capture local features precisely, while deep layers use a larger curvature to parse deeply nested hierarchical structures, effectively solving the problem that fixed network curvature and static graph topology cannot adapt to changes in the semantic scale of complex medical data.

[0019] (5) Even in real physical examination scenarios where physical examination data has missing values, noise interference and incomplete structure, the present invention can still achieve intelligent disease diagnosis with high robustness, high accuracy and high generalization ability.

[0020] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the structure of a personalized disease AI-assisted diagnosis system based on multimodal physical examination data in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a disease intelligent diagnostic device based on multimodal physical examination data in an embodiment of the present invention. Detailed Implementation

[0023] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0024] Example 1 Reference Figure 1 This invention provides a personalized disease AI-assisted diagnosis system based on multimodal physical examination data, including a feature processing module, a feature analysis module, a health status analysis module, and a result output module.

[0025] The feature processing module is used to convert multimodal physical examination data into an initial feature vector.

[0026] The multimodal physical examination data includes demographic information, questionnaire data, laboratory test indicators, vital sign data, anthropometric data, imaging examination results, and data from at least two different modalities in past medical history.

[0027] Specifically, multimodal health check data comes from various records generated by examinees during the health check process.

[0028] The aforementioned demographic information includes basic information such as age, gender, occupation, and place of residence; questionnaire scale data includes subjective or semi-subjective assessment results such as mental health assessment scales, lifestyle questionnaires, and family medical history questionnaires; laboratory test indicators include quantitative indicators obtained through laboratory testing such as complete blood count, comprehensive biochemistry, tumor markers, and hormone levels; vital sign data includes basic physiological parameters such as heart rate, blood pressure, body temperature, and respiratory rate; anthropometric data includes anthropometric indicators obtained through physical measurement such as height, weight, waist circumference, hip circumference, and body fat percentage; imaging examination results include findings and diagnostic opinions of medical imaging such as X-ray, CT, MRI, and ultrasound; and past medical history includes historical health information such as past health status assessment records, surgical history, medication history, and allergy history.

[0029] In one possible implementation, the feature processing module cleans, normalizes, encodes, and vectorizes the physical examination data from different modalities, with different data formats and semantic levels, and then uniformly encodes them into a numerical vector of a preset dimension to form the initial feature vector of the node. The initial feature vector is used to characterize the basic health information of the at least two different modalities of data in Euclidean space and the differences in semantic level, statistical distribution, and feature dimension.

[0030] The feature analysis module is used to analyze the initial feature vector and generate hyperbolic node feature vectors.

[0031] The hyperbolic node feature vector is used to characterize the comprehensive health status representation of the multimodal physical examination data determined in a hyperbolic space with learnable curvature.

[0032] Specifically, the initial feature vector is projected onto a hyperbolic space manifold with learnable curvature to obtain an initial representation; the hyperbolic geodesic distance between nodes is calculated based on the initial representation, and the hypergraph topology is dynamically constructed; global structural features and local high-frequency detail features are extracted by parallel spatial message passing stream and spectral transformation stream, respectively, and the hyperbolic node feature vector is generated after adaptive weighted fusion.

[0033] The hyperbolic manifold described above is a non-Euclidean geometric space with negative curvature. Unlike flat Euclidean space, the volume of hyperbolic space grows exponentially with the radius. This property allows it to embed data with tree-like or hierarchical structures with extremely low geometric distortion.

[0034] In this embodiment, the hyperbolic space manifold is modeled using the Poincaré sphere model, which maps the entire infinitely extending hyperbolic space into an open sphere of finite radius. The hyperbolic distance between any two points within the sphere increases dramatically as it approaches the boundary of the sphere.

[0035] The aforementioned learnable curvature means that the curvature parameter of the hyperbolic space manifold is not a pre-set fixed value, but is an optimizable variable during system training, automatically adjusted according to the actual layer depth and distribution characteristics of the input data.

[0036] The curvature mentioned above determines the degree of curvature of the space: the closer the curvature is to zero, the closer the space is to a flat Euclidean space, which is suitable for capturing fine local features; the greater the curvature, the faster the volume of the space grows, which is suitable for parsing deeply nested global hierarchical structures.

[0037] The node is the individual being examined corresponding to the multimodal physical examination data.

[0038] The aforementioned individuals are natural persons who generate the multimodal physical examination data during the physical examination process. Each individual is mapped to at least one node in the system. The node carries the initial feature vector obtained by the feature processing module after the individual is transformed, which serves as the data basis for subsequent feature analysis.

[0039] Each of the above nodes corresponds to a set of information after uniformly encoding all the multimodal data of a subject.

[0040] It should be understood that the aforementioned multimodal physical examination data is not real-time collected physical examination data, but rather historical physical examination records that have already been collected, stored, and preprocessed. This system analyzes and processes the received historical physical examination data and outputs diagnostic classification results and health deviation scores as auxiliary references. These results do not constitute direct health status assessment conclusions, but rather intermediate processing results that provide information support for medical professionals.

[0041] In one possible implementation, the hyperbolic space manifold is modeled using the Poincaré sphere model, where an n-dimensional hyperbolic space is defined as satisfying... The set of all points x, where c is the curvature parameter.

[0042] The curvature parameters mentioned above can be learned and adjusted during system training, so that a smaller curvature is used in the shallow processing stage to capture local features precisely, and a larger curvature is used in the deep processing stage to analyze the global hierarchical structure.

[0043] The hyperbolic geodesic distance mentioned above is the shortest path length between two points along a geodesic line in hyperbolic space, calculated using Möbius addition and the inverse hyperbolic tangent function.

[0044] The hyperbolic Einstein midpoint algorithm ensures that the aggregation result lies precisely on the hyperbolic manifold by assigning a Lorentz factor that conforms to the geometry of hyperbolic space to each node for weighted aggregation.

[0045] The aforementioned health status analysis module is used to determine the health status assessment result based on the hyperbolic node feature vector.

[0046] The above health status assessment results are category determination information output after comprehensive analysis of multimodal physical examination data of multiple examinees.

[0047] For example, classification labels such as "high-risk diabetes", "high probability of malignancy of thyroid nodules" or "good health" are only intermediate outputs after data processing and are provided for medical professionals to refer to when making comprehensive judgments.

[0048] In one possible implementation, the health status analysis module calculates the hyperbolic geodesic distance between the feature vector of the hyperbolic node and a reference hyperbolic node with a pre-defined known health category, and determines the health category corresponding to the reference hyperbolic node with the smallest distance as the health status assessment result. The reference hyperbolic nodes are stored in the system after preprocessing historical physical examination data of a large number of diagnosed samples by the feature processing module and the feature analysis module.

[0049] The health status analysis module also includes a confidence assessment mechanism, which calculates the consistency ratio of each classification result through multiple random inactivation forward propagation, and uses it as the confidence score of the health status assessment result to provide medical professionals with information on the reliability of diagnostic references.

[0050] The result output module is used to calculate and output the health deviation score of the multimodal physical examination data based on the hyperbolic node feature vector and the hyperbolic geodesic distance of the corresponding node in the reference population, and to use the health deviation score and the health status assessment result as the multidimensional diagnostic output result.

[0051] In one possible implementation, the result output module calculates two types of health deviation scores: the first type is a population health deviation score, calculated based on the hyperbolic geodesic distance between the hyperbolic node feature vector and the corresponding node in the homogeneous reference population, representing the degree of deviation of the examinee's overall health status from the normal state of the homogeneous reference population; the second type is a longitudinal health deviation score, calculated based on the hyperbolic geodesic distance between the hyperbolic node feature vector and the examinee's historical hyperbolic node feature vector, representing the magnitude of change in the examinee's current health status relative to their historical baseline level.

[0052] The results output module integrates health status assessment results, confidence scores, population health deviation scores, and longitudinal health deviation scores into multidimensional diagnostic output results, which are presented in the form of a visual report.

[0053] The aforementioned visualization report includes a map showing the subject's health status in hyperbolic space, a map showing the degree of deviation from the reference population, and a map showing the trend of change compared to their own historical baseline, providing clinicians with multi-dimensional and interpretable information support.

[0054] Example 2 Based on Embodiment 1, this embodiment further defines the internal structure of the feature analysis module.

[0055] The feature analysis module includes a hyperbolic projection unit, a hypergraph construction unit, a feature extraction unit, and a geometric fusion unit.

[0056] The hyperbolic projection unit is used to project the initial feature vector onto a hyperbolic space manifold with learnable curvature through exponential and logarithmic mappings, thereby generating an initial representation of the node in the hyperbolic space.

[0057] The learning of the curvature parameters mentioned above is completed during the end-to-end training process of the system. The specific mechanism is as follows: the curvature parameters of each network layer are transformed into a set of learnable scalars and constrained to positive values ​​through the Softplus activation function to ensure that the hyperbolic space defined by the curvature parameters is geometrically valid; the curvature parameters, together with network parameters such as feature transformation matrix and gating weights, are part of the trainable parameters of the entire system. Instead of setting an independent loss function term, the unified optimization objective is to minimize the diagnostic classification loss (such as cross-entropy loss).

[0058] During training, the gradient of the diagnostic classification loss with respect to the curvature parameters of each layer is calculated synchronously using the gradient backpropagation algorithm, and the curvature parameters are iteratively updated using stochastic gradient descent or its variant optimizer. The curvature parameters of shallow networks naturally tend to smaller values ​​during optimization to maintain a relatively flat local space, which is beneficial for capturing fine local features in physical examination data; while the curvature parameters of deep networks naturally tend to larger values ​​to provide stronger spatial expansion capabilities.

[0059] Specifically, hyperbolic space is modeled using the Poincaré sphere model.

[0060] n-dimensional hyperbolic space is defined as: ; Where D (Disk / Sphere): represents the space itself, i.e., the Poincaré sphere, which is an abstract spatial container where all data points exist within its boundary; c (Curvature parameter): this is the curvature parameter that defines the degree of curvature of the space (Curvature, c>0), and the actual curvature of the space is -c; the value of c determines the size of the Poincaré sphere and the curvature of the internal space; by adjusting c, the volume growth rate of the space can be changed to adapt to the hierarchical structure of different datasets; n (Dimension): it represents the dimension (n) of this space; x (Point / Vector): represents a point in the n-dimensional space, which is a specific n-dimensional real vector. The feature vector of each node corresponding to each physical examination data point, after transformation, will be represented as a point x in this space; (n-dimensional real space): indicates that x is a vector existing in n-dimensional real space; (norm / distance to origin): represents the Euclidean norm of vector x, which is the "length" of x, or the straight-line distance from the point to the center of the sphere; (Radius of curvature): is the radius of the sphere's boundary determined by the curvature c. The larger c is, the smaller the space of the sphere. (Constraints): The Euclidean distance from all data points x to the origin must be strictly less than half of this boundary; all points can only be "inside" the sphere and cannot touch the boundary. Riemannian metric tensor With Euclidean measurement It has a conformal relation, and its expression is: ; When x approaches At the boundary, the conformal factor diverges, therefore, the infinitesimal Euclidean displacement near the boundary corresponds to a large hyperbolic distance, resulting in exponential volume growth.

[0061] The mapping relationship between Euclidean space and hyperbolic manifold is realized by exponential and logarithmic mappings, ensuring that all transformations conform to hyperbolic geometric constraints.

[0062] For the tangent vector The exponential mapping is defined as: ; The exponential mapping is an operation that maps vectors in the tangent space back to a hyperbolic manifold. It is a tangent vector in the tangent space of the manifold at the origin o. The tangent space is a Euclidean space isomorphic to Rn and supports standard vector operations. Let v be the Euclidean norm of the tangent vector v, representing the straight-line distance from the origin to v in the tangent space; c is the curvature parameter, which controls the degree of curvature of the manifold. For the hyperbolic tangent function, the radial component is compressed to the range [0, 1 / c) to ensure that the mapping result always lies inside the Poincaré sphere; Let v represent the unit vector of direction. While keeping the direction unchanged, map the tangent vector back to the manifold along the geodesic. The geometric meaning of this mapping is: starting from the origin o, along the direction of the tangent vector v, with velocity v... The point on the manifold reached after moving along the geodesic line for a unit of time.

[0063] For point The logarithmic mapping is given by the following equation: ; The logarithmic mapping is the inverse operation of the exponential mapping, mapping points on a hyperbolic manifold back to the tangent space, where: It is a point in the Poincaré sphere model. Let y be the Euclidean norm of point y, which is the straight-line distance from that point to the origin. The inverse hyperbolic tangent function reduces the radial position of a point to a distance in the tangent space; as y approaches the boundary, → The inverse hyperbolic tangent function diverges, therefore the boundary points cannot be returned to the finite tangent space by a logarithmic mapping; Similarly, it is a unit vector of direction, maintaining the direction from the origin to y; the geometric meaning of the mapping is: given a point y on a manifold, calculate the initial velocity vector v required to reach that point per unit time, starting from the origin o and moving along a geodesic. This initial velocity vector is the result of the logarithmic mapping.

[0064] The hypergraph construction unit is used to construct a hypergraph topology based on the initial representation of the nodes in hyperbolic space, wherein each hyperedge in the hypergraph topology connects two or more nodes.

[0065] The hypergraph described above is a generalized graph structure. In a regular graph, an edge can only connect two nodes, while in a hypergraph, a hyperedge can connect two or more nodes at the same time. The group of nodes connected by the hyperedge represents a certain high-order correlation between the multimodal physical examination data corresponding to these nodes.

[0066] Higher-order associations refer to a joint effect pattern that is jointly presented among three or more data items and cannot be simply deduced from pairwise relationships.

[0067] The hypergraph is constructed using the hyperbolic K-nearest neighbor (H-KNN) strategy: Calculate the pairwise hyperbolic distance matrix between nodes, where the geodesic distance between two points is defined as: ; in This represents Mobius addition, an operation that yields the precise shortest path distance in curved space.

[0068] Mobius addition generalizes Euclidean vector addition to hyperbolic manifolds, and is defined as follows: ; Then, for each node, select the K nearest nodes to construct a hyperedge, thereby capturing the local higher-order associations between nodes.

[0069] The feature extraction unit is used to extract global structural features and local high-frequency detail features from the hypergraph topology through a parallel spatial message passing stream and a spectral transform stream based on graph wavelet transform, respectively.

[0070] Spatial message passing flow refers to the computational process of regularly passing and aggregating information between nodes and hyperedges in a hypergraph along specific directions. Based on the hypergraph topology, spatial message passing flow first aggregates node representations belonging to the same hyperedge into a hyperedge representation along the node-to-hyperedge direction using the Hyperbolic Einstein Midpoint Algorithm. The Hyperbolic Einstein Midpoint Algorithm is an aggregation operation that calculates the center positions of multiple points in hyperbolic space. This algorithm introduces the Lorentz factor as a weight and corrects the aggregation contribution based on the radial position of each node in hyperbolic space, ensuring that the aggregation result is strictly on the hyperbolic manifold and does not produce geometric distortion. Along the hyperedge-to-node direction, the aggregated hyperedge representation is backpropagated to each node connected by the hyperedge to update the node representation of each node. After the complete message passing loop from node to hyperedge and from hyperedge to node, a node representation encoding global structural features is finally generated. Global structural features refer to macroscopic information reflecting the position of a node in the entire hypergraph topology, its group, and the overall association pattern with other groups.

[0071] Specifically, information first flows from nodes to hyperedges, and is aggregated using the hyperbolic Einstein midpoint algorithm, for each node embedding set { The aggregated hyperedge representation is calculated as follows: ; in The Einstein weights, determined by the radial position of the node in hyperbolic space, are used. The enhanced hyperedge representation is then backpropagated to its constituent nodes, completing a full message-passing loop. In the Poincaré sphere model, the closer a point is to the boundary of the sphere, the higher its norm. The larger it is, the more it corresponds to The larger the value, the better; It is an element in the hypergraph incidence matrix, a scalar value representing the connection strength or dependency relationship between node i and hyperedge j; The embedding or feature vector of node i is the coordinate representation of node i in hyperbolic space (Poincaré sphere) at the current time.

[0072] The spectral transform stream extracts local high-frequency components through hyperbolic wavelet transform.

[0073] Specifically, we first define a hyperbolic adjacency matrix based on a heat kernel: Construct the normalized hyperbolic Laplace operator: ; To overcome the limitations of large-scale graph computation, a Chebyshev polynomial expansion approximation spectral filter is employed: This eliminates the need for explicit Laplace eigenvalue decomposition.

[0074] The geometric fusion unit is used to adaptively weight and fuse the global structural features and local high-frequency detail features to generate the hyperbolic node feature vector.

[0075] In one possible implementation, the fusion process is as follows: Spatial flow embedding and spectral flow embedding are projected onto the tangent space at the origin via a logarithmic map; geometric gating units are introduced within the tangent space, and gating vectors are generated by concatenating feature vectors and applying linear transformations and the sigmoid activation function. ; in, The feature vector obtained by projecting the spatial flow embedding of the i-th node onto the origin tangent space via a logarithmic mapping is an example of a vector that encodes the global structural information and hierarchical clustering features of the node in the hypergraph topology. This vector has dimension n and resides in a tangent space isomorphic to Rn. middle; The feature vector obtained by projecting the spectral flow embedding of the i-th node onto the origin tangent space through a logarithmic mapping is the vector that encodes the local high-frequency details and boundary information extracted by the hyperbolic wavelet transform. The dimension is n. Symbols (concatenation operators): This means concatenating the two n-dimensional vectors along the feature dimension to produce a 2n-dimensional joint vector; The learnable gated weight matrix has an n×2n dimension and belongs to the Euclidean parameters. Its function is to linearly map the concatenated 2n-dimensional vector to an n-dimensional vector, so that the output dimension is consistent with the input feature dimension; : Learnable gated bias vector, with dimension n, belonging to Euclidean parameters; Sigmoid(⋅): Logistic activation function, which compresses the output value to the (0,1) interval; The generated gated vector has a dimension of n, and each element... ,k∈(0,1), represents the importance weight of spatial flow features relative to spectral flow features in the k-th feature dimension.

[0076] Dimension-weighted fusion: ; Where ⊙: Hadamard product, i.e., element-wise multiplication, here implements a soft selection mechanism based on dimensions; : Dimensionally weighted gating vectors for spatial flow features; when When k→1, this dimension tends to preserve spatial flow information; The complementary parts of the gate vector are used for dimension-wise weighting of the spectral flow features. When As k→0, this dimension tends to preserve spectral flow information, and the sum of the two forms a convex combination of each dimension; The fused tangent space feature vector, with dimension n, still resides in the origin tangent space; the fused embedding is projected back onto the Poincaré sphere through exponential mapping to obtain the final hyperbolic node feature vector.

[0077] ; in, The final hyperbolic node feature vector of the i-th node effectively integrates global structural information and local spectral details while preserving geometric fidelity, serving as input for subsequent health status analysis.

[0078] Example 3 Based on Embodiment 2, this embodiment further defines the internal structure of the hyperbolic projection unit.

[0079] The hyperbolic projection unit includes a curvature parameter generation subunit, a logarithmic mapping subunit, and a parameter adjustment subunit.

[0080] The curvature parameter generation subunit mentioned above is used to automatically adjust the curvature based on the semantic hierarchy differences of multimodal data represented by the initial feature vector through learnable parameters.

[0081] Specifically, the differences in semantic levels among different modalities of data reflect the inherent hierarchical structure of the data.

[0082] For example, there are hierarchical differences in the data between different modalities, from basic physiological indicators to system function assessments and comprehensive health status, from micro to macro levels.

[0083] The curvature parameter generation subunit determines the curvature size required for each information processing stage based on these differences, so that subsequent stages can use a larger curvature to parse the deeply nested hierarchical structure.

[0084] The logarithmic mapping subunit is used to project the initial feature vector from Euclidean space to the hyperbolic space based on the curvature parameter, thereby obtaining the initial representation of the node in the hyperbolic space. The logarithmic mapping ensures that the feature vector in Euclidean space can be accurately mapped to a hyperbolic manifold with a specified curvature, preserving the relative position information of the vector in Euclidean space.

[0085] The parameter adjustment subunit is used to determine the volume growth rate of the hyperbolic space based on the distribution density of the initial representation in the hyperbolic space and the curvature parameter, and to perform geometric redistribution on the initial representation based on the volume growth rate to obtain the adjusted initial representation.

[0086] It should be understood that since the volume growth characteristics of hyperbolic space differ under different curvatures, the node distribution after initial mapping may not fully utilize the expressive capacity of hyperbolic space. By analyzing the distribution density of nodes, the volume growth rate can be adaptively adjusted to enable the nodes to achieve the optimal geometric distribution in hyperbolic space, thereby improving space utilization and enhancing the separability between categories.

[0087] Example 4 Based on Embodiment 2, this embodiment further defines the internal structure of the hypergraph building unit.

[0088] The hypergraph construction unit includes a neighborhood computation subunit and a hyperedge generation subunit.

[0089] The neighborhood calculation subunit is used to calculate the hyperbolic geodesic distance between nodes based on the initial representation of the nodes in hyperbolic space, and to determine a preset number of neighboring nodes that are closest to each node based on the hyperbolic geodesic distance.

[0090] Compared to Euclidean distance, the hyperbolic geodesic distance described above can more accurately reflect the true semantic relationship of nodes in the hierarchical structure. In particular, near the boundary of hyperbolic space, seemingly adjacent points in Euclidean space may be far apart in hyperbolic distance. This characteristic enables hyperbolic K-nearest neighbors to effectively distinguish nodes that are similar on the surface but have different meanings.

[0091] Alternatively, a hyperbolic density peak clustering strategy can be used to determine neighboring nodes.

[0092] Specifically, the hyperbolic density peak clustering strategy includes: calculating the hyperbolic local density and relative distance of each node; the hyperbolic local density is calculated using a Gaussian kernel function based on the local crowding of the node within the Poincaré sphere, with a truncation parameter that ensures only approximate neighbors contribute significantly; the hyperbolic relative distance represents the minimum hyperbolic geodesic distance from a node to any node with higher density; and identifying nodes with both high density and large relative distance as cluster centers, with each center and its followers forming a hyperedge.

[0093] It should be understood that the hyperbolic density peak clustering strategy has a natural ability to identify outliers for nodes located in low-density regions. Because it has both low density and large relative distance, it can effectively prevent the construction of hyperedges due to noise pollution.

[0094] When constructing the hyperedge, the hyperedge generation subunit also introduces a saturation gating mechanism to adaptively sparsify the connection probability between nodes and hyperedges.

[0095] Specifically, the hyperbolic geodesic distance between the node and the hyperedge centroid is used as the original geometric affinity, and the unnormalized association strength is calculated through the hyperbolic distance-aware attention mechanism. If the association strength is lower than the adaptive threshold, the gating is turned off by hard truncation, and low-quality connections are removed, thereby sparsifying the hypergraph topology and reducing structural noise.

[0096] As the network training progresses, the hyperbolic representation of the nodes is continuously updated and optimized. Based on the updated node representation, the connection probabilities are recalculated and the hyperedges are reconstructed, enabling the hypergraph topology to adaptively reflect the evolution of the inherent relationships in the data without relying on a pre-set static correlation matrix. This dynamic evolution mechanism allows the model to proactively reshape information paths to fill topological gaps when there are missing values ​​or noise interference in the physical examination data, ensuring the robustness and accuracy of the diagnostic results under incomplete data conditions.

[0097] The hyperedge generation subunit is used to form a hyperedge with each node and its corresponding neighboring nodes, thereby generating the hypergraph topology.

[0098] The hypergraph topology evolves dynamically as the node representations are updated.

[0099] Specifically, the dynamic evolution of the hypergraph topology is carried out synchronously during the forward propagation of each training iteration step of the system, without setting an independent update cycle.

[0100] The specific update timing is as follows: In each iteration, after the node features are calculated by the hyperbolic projection unit to obtain the node representation in the current hyperbolic space, the hypergraph construction unit re-executes the neighborhood calculation and hyperedge generation operations based on the current node representation to generate a new hypergraph topology that matches the current node feature distribution.

[0101] The update method is as follows: Calculate the hyperbolic geodesic distance matrix between all node pairs based on the node representation; for each node, select the K nodes with the smallest distance to form its neighbor set; construct a hyperedge with the node as the center and its neighbor set as members; after traversing all nodes, the hypergraph inclination matrix of this iteration step is obtained.

[0102] The hypergraph correlation matrix described above is a continuously differentiable probability matrix, rather than a discrete binary matrix, to ensure that gradients can propagate backward. The number of iterations is consistent with the number of iterations in the overall system training, meaning that the hypergraph topology is updated once in each training step until training converges. After training converges, during the inference phase, for newly input subject data, the hypergraph construction unit also dynamically constructs the hypergraph based on the common node representations of the subject and the reference sample, without relying on a pre-set fixed graph structure.

[0103] Based on the above technical means, as the network training progresses, the hyperbolic representation of the nodes is continuously updated and optimized. The hyperedge is reconstructed based on the updated node representation, so that the hypergraph topology can adaptively reflect the evolution of the inherent relationship between the data, without relying on a preset static correlation matrix.

[0104] Example 5 Based on Embodiment 2, this embodiment further defines the internal structure of the feature extraction unit.

[0105] The feature extraction unit includes a spatial message passing stream and a spectral transformation stream, which run in parallel.

[0106] The spatial message passing stream is used to aggregate node representations belonging to the same hyperedge into a hyperedge representation along the hyperbola-to-hyperedge direction based on the hypergraph topology, and then propagate the hyperedge representation back to each node along the hyperedge-to-node direction to update the node representation and generate a node representation that encodes global structural features.

[0107] The aforementioned closed-loop message passing mechanism fully utilizes the metric properties of hyperbolic geometry. Unlike the smoothing effect in Euclidean space, message passing in hyperbolic space maintains or even amplifies the relative distance between embedded nodes as the iteration deepens, effectively alleviating the oversmoothing problem commonly found in deep graph neural networks.

[0108] The spectral transform stream is used to perform multi-scale spectral filtering on the initial representation of the node in hyperbolic space through hyperbolic graph wavelet transform, and to extract the node representation that encodes local high-frequency detail features.

[0109] The hyperbolic graph wavelet transform is based on the normalized hyperbolic Laplacian operator, which is generated by hyperbolic heat kernels and simultaneously encodes the graph's topological connectivity and hyperbolic geometric metric information.

[0110] Through multi-scale wavelet decomposition, spectral transform flow extracts node features from different frequency levels, paying particular attention to high-frequency signals that reflect early signs of disease, boundary abnormal indicators, and intermodal inconsistencies, thus compensating for the inherent low-pass bias of spatial message passing flow.

[0111] Example 6 Based on Embodiment 1, this embodiment further defines the internal structure of the health status analysis module.

[0112] The health status analysis module includes a distance calculation unit and a classification determination unit.

[0113] The distance calculation unit is used to calculate the hyperbolic geodesic distance between the hyperbolic node feature vector and a reference hyperbolic node with a preset known health category.

[0114] The hyperbolic geodesic distance is the shortest path length along the geodesic line between the feature vector of the hyperbolic node and the reference hyperbolic node in the hyperbolic space with learnable curvature. The reference hyperbolic node is obtained by processing a large amount of confirmed sample data through the feature processing module and the feature analysis module, and represents typical feature patterns of various diseases.

[0115] The classification and determination unit is used to determine the health category to which the multimodal physical examination data belongs based on the hyperbolic geodesic distance, and generate the health status assessment result.

[0116] Furthermore, the classification determination unit includes a distance comparison subunit and a category determination subunit.

[0117] The distance comparison subunit is used to compare the hyperbolic node feature vector with the hyperbolic geodesic distance between each preset known health category reference hyperbolic node, and determine the health category corresponding to the reference hyperbolic node with the smallest distance.

[0118] The category determination subunit is used to determine the health category corresponding to the reference hyperbolic node with the smallest distance as the health category to which the multimodal physical examination data belongs, and to generate the health status assessment result.

[0119] In addition, the health status analysis module also includes a scoring calculation unit for assessing the reliability and deviation of the diagnostic results.

[0120] The scoring calculation unit includes a confidence assessment subunit and a deviation calculation subunit.

[0121] The confidence assessment subunit is used to perform multiple perturbations on the hyperbolic node feature vectors through multiple random inactivation forward propagations, and calculate the consistency ratio of each classification result as the confidence score of the health status assessment result. This mechanism simulates the diagnostic performance stability of the model under different conditions and provides a reliable reference for clinical decision-making.

[0122] The deviation calculation subunit is used to normalize the distance between the hyperbolic node feature vector and the reference hyperbolic node into a health deviation score. The distance between the multimodal physical examination data and the typical features of the known health category is positively correlated with the health deviation score, that is, the higher the deviation score, the farther the examinee's health status is from the typical disease features.

[0123] Example 7 Based on Embodiment 1, this embodiment further defines the internal structure of the result output module.

[0124] The result output module includes a deviation calculation unit, a longitudinal deviation calculation unit, and a multidimensional diagnostic output unit.

[0125] The deviation calculation unit is used to calculate the group health deviation score of the examinee based on the hyperbolic node feature vector and the hyperbolic geodesic distance between the corresponding node in the reference population.

[0126] Specifically, hyperbolic node feature vectors generated by preprocessing each sample in the reference population using the feature processing module and feature analysis module are obtained to form a reference node set; the hyperbolic geodesic distance between the hyperbolic node feature vector of the examinee and each reference node in the reference node set is calculated, and the mean or median of the distance is taken as the baseline distance; the baseline distance is normalized and mapped to a preset scoring interval to obtain the population health deviation score.

[0127] It should be understood that the higher the health deviation score of the above-mentioned groups, the greater the degree of deviation of the examinee's overall health status from the normal state of the homogeneous reference group.

[0128] The reference population refers to healthy individuals who have the same or similar demographic characteristics as the subject of the examination.

[0129] The construction and storage of reference hyperbolic nodes are completed during the system training phase. The specific process is as follows: Stratified sampling of diagnosed subjects is performed from a large-scale physical examination database, categorized by health type. The sample size for each disease category is no less than a preset threshold to ensure statistical representativeness. The health category label for each sample is confirmed by clinicians according to diagnostic guidelines, and the label includes the disease classification name and corresponding diagnostic standard code. During training, all labeled sample data is input into the system, sequentially processed by the feature processing module to generate initial feature vectors, and then by the feature analysis module to generate hyperbolic node feature vectors. For each health category, the intra-class mean of the hyperbolic node feature vectors of all samples within that category is taken. This mean vector is the reference hyperbolic node for that category and is stored in the system memory for subsequent inference. In subsequent actual diagnosis, the hyperbolic geodesic distance between the hyperbolic node feature vector of a new subject and the reference hyperbolic nodes of each known health category stored in memory is calculated one by one. The health category corresponding to the nearest reference hyperbolic node is taken as the diagnostic classification result.

[0130] The population health deviation score characterizes the degree of deviation of the examinee's overall health status from the norm of a homogeneous reference population. The homogeneous reference population consists of healthy individuals with similar demographic characteristics (such as age and gender) to the examinee, and their corresponding hyperbolic node feature vectors are pre-stored in the system database.

[0131] The longitudinal deviation calculation unit is used to calculate the longitudinal health deviation score of the examinee based on the hyperbolic geodesic distance between the hyperbolic node feature vector and the examinee's historical hyperbolic node feature vector.

[0132] Specifically, the system obtains the historical hyperbolic node feature vectors of the subject after processing by the system in at least one previous physical examination; calculates the hyperbolic geodesic distance between the hyperbolic node feature vector generated in the current physical examination and the historical hyperbolic node feature vectors; and maps the hyperbolic geodesic distance to a preset scoring interval after normalization to obtain the longitudinal health deviation score.

[0133] The preset scoring range is a continuous range of values ​​from 0 to 100. A score close to 0 indicates that the examinee's current health status is basically consistent with their historical baseline level and has not changed significantly. A score close to 100 indicates that the examinee's current health status has deviated greatly from their historical baseline level, suggesting that there is significant fluctuation in health status.

[0134] Normalization can be achieved using max-min normalization or Z-score normalization methods to linearly map the hyperbolic geodesic distance to the scoring interval.

[0135] Optionally, the above scoring range can also be adjusted to other ranges according to actual clinical needs, such as a range of 0 to 10 or 0 to 1.

[0136] It should be understood that the larger the longitudinal health deviation score, the greater the change in the examinee's current health status relative to their historical baseline level, indicating that attention should be paid to the longitudinal development trend of the examinee's health status.

[0137] Optionally, when there are multiple historical physical examination records, the feature vector of the most recent historical hyperbolic node can be selected for calculation, or the average representation of multiple historical records can be selected as the individual baseline reference.

[0138] The longitudinal health deviation score characterizes the extent of change in the examinee's current health status relative to their historical baseline. The historical hyperbolic node feature vector is a feature vector stored after processing by this system from the examinee's previous physical examinations, reflecting the examinee's individual health baseline.

[0139] The multidimensional diagnostic output unit integrates the health status assessment results, the population health deviation score, and the longitudinal health deviation score into a multidimensional diagnostic output result, which is then presented in the form of a visual report. The visual report graphically displays the examinee's health status position in hyperbolic space, the degree of deviation from the reference population, and the trend of change compared to their own historical baseline, providing clinicians with intuitive, multidimensional, and interpretable auxiliary diagnostic information.

[0140] Example 8 like Figure 2 As shown in the figure, an intelligent disease diagnosis device based on multimodal physical examination data in an embodiment of the present invention includes: a processor and a memory.

[0141] A processor is an integrated circuit chip with signal processing capabilities, used to execute computer program instructions stored in memory to implement the functions of the feature processing module, feature analysis module, health status analysis module, and result output module in the above embodiments of the present invention. The processor can be one or more combinations of a central processing unit, a graphics processing unit, a tensor processor, or a neural network processor.

[0142] The memory is connected to the processor via a communication bus and is used to store computer program instructions, multimodal physical examination data, intermediate processing results, and final diagnostic output results. The memory includes random access memory and non-volatile memory. Random access memory is used to temporarily store data and programs being processed during system operation; non-volatile memory is used to persistently store system parameters, data of corresponding nodes in a preset reference population, and feature vectors of reference hyperbolic nodes for each known health category.

[0143] The input interface is used to receive multimodal physical examination data from externally input subjects. This multimodal physical examination data includes demographic information, questionnaire data, laboratory test results, vital sign data, anthropometric data, imaging examination results, and data from at least two different modalities of past medical history. The input interface can be connected to a physical examination information system, electronic medical record system, medical image archiving and communication system, or a manual data entry terminal.

[0144] The output interface is used to output multidimensional diagnostic results. These results include health status assessments, population health deviation scores, and longitudinal health deviation scores. The output interface can connect to display devices, printing devices, or hospital information systems to present clinical users with visual reports.

[0145] The communication bus is used to enable data exchange between the processor, memory, input interface, and output interface.

[0146] The device operates as follows: The processor acquires the multimodal physical examination data of the subject through the input interface and temporarily stores it in the memory; the processor executes a feature processing program to convert the multimodal physical examination data into initial feature vectors; the processor executes a feature analysis program to project the initial feature vectors onto a hyperbolic space manifold with learnable curvature, dynamically constructs a hypergraph topology, and extracts global structural features and local high-frequency detail features using spatial message passing flow and spectral transformation flow, respectively, and then adaptively fuses them through a geometric gating unit to generate hyperbolic node feature vectors; the processor executes a health status analysis program to calculate the hyperbolic geodesic distance between the hyperbolic node feature vectors and the reference hyperbolic nodes stored in the memory to determine the health status assessment results; the processor executes a result output program to calculate the group health deviation score and the longitudinal health deviation score, and integrates the above results into a multidimensional diagnostic output result, which is output through the output interface.

[0147] The device may also include a network communication module for data interaction with remote servers or cloud computing platforms to support regular updates of large-scale reference population data and remote training and deployment of models.

[0148] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0149] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A personalized AI-assisted disease diagnosis system based on multimodal physical examination data, characterized in that, The system includes: A feature processing module is used to convert multimodal physical examination data into an initial feature vector. The multimodal physical examination data includes demographic information, questionnaire scale data, laboratory test indicators, vital sign data, physical measurement data, imaging examination results, and data from at least two different modalities in past medical history. The initial feature vector is used to characterize the basic health information of the at least two different modalities of data in Euclidean space and their differences in semantic level, statistical distribution, and feature dimensions. The initial feature vector is the feature representation of the node in Euclidean space. The feature analysis module is used to analyze the initial feature vector to generate hyperbolic node feature vectors; the hyperbolic node feature vectors are used to characterize the comprehensive health status representation of the multimodal physical examination data in a hyperbolic space with learnable curvature; the node is the examinee corresponding to the multimodal physical examination data; The health status analysis module is used to determine the health status assessment result based on the hyperbolic node feature vector; The results output module is used to calculate and output the health deviation score of the multimodal physical examination data based on the hyperbolic node feature vector and the hyperbolic geodesic distance of the corresponding node in the reference population, and to use the health deviation score and the health status assessment result as the multidimensional diagnostic output result.

2. The personalized disease AI-assisted diagnosis system based on multimodal physical examination data according to claim 1, characterized in that, The feature analysis module includes: The hyperbolic projection unit is used to project the initial feature vector onto a hyperbolic space manifold with learnable curvature through exponential and logarithmic mappings, generating the initial representation of the node in the hyperbolic space; A hypergraph construction unit is used to construct a hypergraph topology based on the initial representation of the nodes in hyperbolic space, wherein each hyperedge in the hypergraph topology connects two or more nodes. The feature extraction unit is used to extract global structural features and local high-frequency detail features from the hypergraph topology through a parallel spatial message passing stream and a spectral transform stream based on graph wavelet transform, respectively. The geometric fusion unit is used to adaptively weight and fuse the global structural features and local high-frequency detail features to generate the hyperbolic node feature vector.

3. The personalized disease AI-assisted diagnosis system based on multimodal physical examination data according to claim 2, characterized in that, The hyperbolic projection unit includes: A curvature parameter generation subunit is used to learn and generate curvature parameters based on the semantic hierarchy differences; The logarithmic mapping subunit is used to project the initial feature vector from Euclidean space to the hyperbolic space based on the curvature parameter, thereby obtaining the initial representation of the node in the hyperbolic space. The parameter adjustment subunit is used to determine the volume growth rate of the hyperbolic space based on the distribution density of the initial representation in the hyperbolic space and the curvature parameter, and to perform geometric redistribution on the initial representation based on the volume growth rate to obtain the adjusted initial representation.

4. The personalized disease AI-assisted diagnosis system based on multimodal physical examination data according to claim 2, characterized in that, The hypergraph construction unit includes: The neighborhood calculation subunit is used to calculate the hyperbolic geodesic distance between nodes based on the initial representation of the nodes in hyperbolic space, and to determine a preset number of neighboring nodes that are closest to each node based on the hyperbolic geodesic distance. The hyperedge generation subunit is used to form a hyperedge with each node and its corresponding neighboring nodes, thereby generating the hypergraph topology; the hypergraph topology evolves dynamically as the node representation is updated.

5. A personalized disease AI-assisted diagnostic system based on multimodal physical examination data according to claim 2, characterized in that, The feature extraction unit includes: The spatial message passing subunit is used to aggregate the node representations belonging to the same hyperedge into a hyperedge representation along the hyperbolic Einstein midpoint algorithm based on the hypergraph topology, and then propagate the hyperedge representation back to each node along the hyperedge to update the node representation, thereby generating a node representation that encodes global structural features. The spectral transform subunit is used to perform multi-scale spectral filtering on the initial representation of the node in hyperbolic space through hyperbolic graph wavelet transform, and extract the node representation that encodes local high-frequency detail features.

6. A personalized disease AI-assisted diagnostic system based on multimodal physical examination data according to claim 1, characterized in that, The health status analysis module includes: The distance calculation unit is used to calculate the hyperbolic geodesic distance between the hyperbolic node feature vector and a reference hyperbolic node with a preset known health category; the hyperbolic geodesic distance is the shortest path length of the hyperbolic node feature vector and the reference hyperbolic node along the geodesic in the hyperbolic space with learnable curvature; The classification and determination unit is used to determine the health category to which the multimodal physical examination data belongs based on the hyperbolic geodesic distance, and generate the health status assessment result.

7. A personalized disease AI-assisted diagnostic system based on multimodal physical examination data according to claim 6, characterized in that, The classification determination unit includes: The distance comparison subunit is used to compare the hyperbolic node feature vector with the hyperbolic geodesic distance between each preset known health category reference hyperbolic node, and determine the health category corresponding to the reference hyperbolic node with the smallest distance; The category determination subunit is used to determine the health category corresponding to the reference hyperbolic node with the smallest distance as the health category to which the multimodal physical examination data belongs, and to generate the health status assessment result.

8. A personalized disease AI-assisted diagnostic system based on multimodal physical examination data according to claim 6, characterized in that, The scoring calculation unit includes: The confidence assessment subunit is used to perform diagnostic classification on the hyperbolic node feature vector after multiple perturbations through multiple random inactivation forward propagation, and calculate the consistency ratio of each classification result as the confidence score of the health status assessment result. The deviation calculation subunit is used to normalize the distance between the hyperbolic node feature vector and the reference hyperbolic node into a health deviation score; the distance between the multimodal physical examination data and the typical features of the known health category is positively correlated with the health deviation score.

9. A personalized disease AI-assisted diagnostic system based on multimodal physical examination data according to claim 1, characterized in that, The result output module includes: The deviation calculation unit is used to calculate the group health deviation score of the examinee based on the hyperbolic node feature vector and the hyperbolic geodesic distance between the corresponding node in the reference population; the group health deviation score represents the degree of deviation of the examinee's overall health status from the normal state of the homogeneous reference population. The longitudinal deviation calculation unit is used to calculate the longitudinal health deviation score of the examinee based on the hyperbolic geodesic distance between the hyperbolic node feature vector and the examinee's historical hyperbolic node feature vector; the longitudinal health deviation score represents the magnitude of change in the examinee's current health status relative to its own historical baseline level; The multidimensional diagnostic output unit is used to integrate the health status assessment results, the population health deviation score, and the longitudinal health deviation score into a multidimensional diagnostic output result, and present it in the form of a visual report.

Images