Chronic patient screening and personalized processing method and system based on large model

Abstract

The invention provides a chronic patient screening and personalized processing method and system based on a large model, and relates to the technical field of information management, and the method comprises the steps: obtaining and integrating hierarchical diagnosis and treatment coding data of a patient and continuous physiological signal data collected by wearable equipment; respectively mapping the multi-modal data into embedded representations located in a non-Euclidean feature space through a specific model, and executing efficient feature fusion in the non-Euclidean feature space; and calling a large model trained by federated distillation to carry out deep analysis on the fused features to generate a patient recall priority index. According to the method and the device, accurate identification and recall of high-risk patients can be realized, and after the patients see a doctor, a personalized management path scheme is automatically matched or generated according to the diagnosis results of the patients. By making full use of complex internal association of patient data, on the premise that data privacy is protected, accuracy and efficiency of early screening of chronic diseases are remarkably improved, and intelligentization and individuation of subsequent management are achieved.

Description

Technical Field

[0001] This application relates to the field of information management technology, and more specifically, to a method and system for screening and personalized treatment of chronic disease patients based on large models. Background Technology

[0002] With the increasing prevalence of chronic diseases and the aging population, early screening and long-term management of chronic diseases have become core issues urgently needing to be addressed by the healthcare system. Currently, most medical institutions are still in a state of information fragmentation, with patient data scattered across different departments and systems, lacking unified, structured integration and intelligent processing methods. In chronic disease management practice, traditional methods rely heavily on manual screening rules and experience-based judgment, often based on a single data source, making it difficult to consider multimodal information such as diagnostic codes and continuous physiological signals. This affects the accuracy of screening and restricts the development of personalized intervention pathways. Furthermore, due to data heterogeneity between different medical institutions and privacy compliance requirements for model training, it is difficult to build a unified, high-quality screening system.

[0003] Therefore, how to integrate multi-source heterogeneous patient data in a multi-medical institution environment, build a unified and efficient intelligent screening process, and achieve accurate identification and personalized path configuration for patients with chronic diseases has become a key technical problem that urgently needs to be solved. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this application provides a method and system for screening and personalized treatment of chronic disease patients based on a large model.

[0005] In a first aspect, embodiments of this application provide a method for screening and personalized treatment of chronic disease patients based on a large model, including:

[0006] Acquire object data of the target entity and generate an object data set, the object data set including target encoded data with hierarchical relationship and continuous time-series signal data collected by wearable device;

[0007] The target encoded data is processed by a first preset model to generate a first embedded representation of the target encoded data, wherein the first embedded representation is located in a preset non-Euclidean feature space.

[0008] The continuous time-series signal data is processed by a second preset model to generate a second embedded representation of the continuous time-series signal data;

[0009] In the non-Euclidean feature space, a fusion operation is performed on the first embedding representation and the second embedding representation to generate an aligned feature set;

[0010] The large model trained by federated distillation is invoked to process the aligned feature set to generate a set of recall metrics that characterize the recall priority of target individuals.

[0011] If the recall priority in the recall indicator set meets the preset recall threshold, the corresponding individual is identified as the target recall individual and a recall notification is sent; after the target recall individual completes the target interaction, it is registered as a target management individual.

[0012] Based on the object type indicated by the analysis results of the target management individual, a personalized processing path scheme is matched or generated from the preset management path library, and the scheme is assigned to the target management individual.

[0013] Optionally, generating the first embedding representation of the target encoded data includes:

[0014] The target encoded data is converted into nodes, and edges are constructed based on the hierarchical relationship to generate an initial hierarchical graph;

[0015] A neighbor aggregation operation is performed on the initial hierarchical graph using a hyperbolic graph convolutional network to generate a hierarchical graph representation.

[0016] Perform a hyperbolic convolution operation on the hierarchical graph representation to generate a hierarchical embedding representation;

[0017] Based on the hierarchical embedding representation, the first embedding representation is generated, and the first embedding representation is located in the non-Euclidean feature space.

[0018] Optionally, generating a second embedded representation of the continuous time-series signal data includes:

[0019] The continuous temporal signal data is iteratively processed using a forward long short-term memory network to generate forward hidden states;

[0020] The reverse sequence of the continuous time-series signal data is iteratively processed by a backward long short-term memory network to generate a backward hidden state.

[0021] By fusing the forward and backward hidden states, a temporal embedding representation is generated;

[0022] The second embedding representation is generated based on the temporal embedding representation.

[0023] Optionally, the non-Euclidean feature space is a hyperbolic space; the generated aligned feature set includes:

[0024] Perform a logarithmic mapping with respect to the origin of the hyperbolic space on the first embedded representation, project the first embedded representation to the tangent space of the origin, and generate a first tangent space embedded representation;

[0025] In the slicing space, a combination operation is performed on the first slicing space embedding representation and the second embedding representation to generate a fused embedding representation;

[0026] An exponential mapping based on the origin of the hyperbolic space is performed on the fused embedding representation to map the fused embedding representation back to the hyperbolic space, generating the aligned feature set.

[0027] Optionally, the generation of the fused embedding representation includes:

[0028] Based on the first slicing space embedding representation, the corresponding target management path is extracted from the preset management path library, and a path embedding representation representing the target management path is generated;

[0029] The path embedding representation is used to generate a channel attention vector, which is used to indicate the relative importance of different signal channels in the continuous time-series signal data;

[0030] The first fusion sub-step is executed, in which the channel attention vector is applied to the forward hidden state for weighting, and the weighted forward hidden state is fused with the first tangent space embedding representation to generate a forward context representation;

[0031] The second fusion sub-step is executed, in which the channel attention vector is applied to weight the backward hidden state, and the weighted backward hidden state is fused with the first tangent space embedding representation to generate a backward context representation;

[0032] The forward context representation and the backward context representation are finally fused to generate the fused embedded representation.

[0033] Optionally, generating the alignment feature set includes:

[0034] Perform the exponential mapping on the fused embedding representation to generate an initial aligned embedding;

[0035] Based on the statistical time-series characteristics of the continuous time-series signal data, a risk adjustment factor characterizing the recall priority of the individual is calculated;

[0036] In the hyperbolic space, the position of the initial alignment embedding is adjusted according to the risk adjustment factor along the geodesic path connecting the origin and the initial alignment embedding to generate the alignment feature set.

[0037] Optionally, the set of recall metrics for generating the recall priority of target individuals includes:

[0038] The alignment feature set, the path embedding representation, and the risk adjustment factor are combined to generate enhanced input features;

[0039] The enhanced input features are input into the large model for processing to generate the recall index set.

[0040] Optionally, the generation of enhanced input features includes:

[0041] Obtain the identity information of the target entity corresponding to the current object data, and generate an organizational embedding representation that characterizes the data characteristics of the target entity;

[0042] Based on the mechanism embedding representation, generate one or more gating vectors;

[0043] The gating vector is applied to perform a gating fusion operation on the alignment feature set, the path embedding representation, and the risk adjustment factor to generate the enhanced input features.

[0044] Optionally, the large model trained by the federated distillation includes:

[0045] Deploy a local teacher model for each of the target entities;

[0046] The local teacher model processes the alignment feature set to generate an intermediate layer feature representation, which serves as local knowledge, characterizing the fusion state of the target encoded data and the continuous temporal signal data.

[0047] A global student model is deployed on the central server as the larger model;

[0048] The global student model receives the intermediate layer feature representation from one or more of the local teacher models, and updates the parameters of the global student model based on the principle of minimizing the difference between the feature representation generated by the global student model itself and the intermediate layer feature representation.

[0049] Secondly, embodiments of this disclosure also provide a system for screening and personalized treatment of chronic diseases based on a large model, including:

[0050] The acquisition module is used to acquire object data of the target entity and generate an object data set, wherein the object data set includes target encoded data with hierarchical relationship and continuous time-series signal data acquired by the wearable device;

[0051] The processing module is configured to process the target encoded data through a first preset model to generate a first embedded representation of the target encoded data, wherein the first embedded representation is located in a preset non-Euclidean feature space; process the continuous time-series signal data through a second preset model to generate a second embedded representation of the continuous time-series signal data; and perform a fusion operation on the first embedded representation and the second embedded representation in the non-Euclidean feature space to generate an aligned feature set.

[0052] The recall module is used to call a large model trained by federated distillation to process the alignment feature set to generate a set of recall indicators that characterize the recall priority of the target individual; if the recall priority in the recall indicator set meets the preset recall threshold, the corresponding individual is identified as the target recall individual and a recall notification is sent; after the target recall individual completes the target interaction, it is registered as a target management individual.

[0053] The management module is used to match or generate personalized processing path schemes from a preset management path library based on the object type indicated by the analysis results of the target management individual, and to assign the schemes to the target management individual.

[0054] Compared with existing technologies, this application overcomes the limitations of existing technologies in terms of data integration capabilities, model generalization capabilities, and path adaptation capabilities. By introducing a bimodal embedding fusion mechanism of diagnostic codes and continuous physiological signals, and utilizing hyperbolic space to enhance feature representation capabilities, this method can more accurately capture the health risk characteristics of individuals with chronic diseases across different dimensions. Simultaneously, a federated distillation mechanism is used to train the large model, enabling medical institutions to participate in collaborative model optimization without sharing raw data, thus ensuring a balance between data privacy and model generalization capabilities. Furthermore, through deep integration with a chronic disease management path library, personalized path generation and precise allocation for different individual types are achieved, promoting the transformation of chronic disease management from passive response to proactive intervention, thereby improving overall medical efficiency and individual health levels. Attached Figure Description

[0055] Figure 1 A flowchart illustrating the method for screening and personalized treatment of chronic disease patients based on a large model, provided in the embodiments of this application;

[0056] Figure 2 A flowchart illustrating a first embedded representation method for generating the target encoded data, provided in an embodiment of this application;

[0057] Figure 3 A flowchart illustrating a second embedded representation method for generating the continuous time-series signal data, provided in an embodiment of this application;

[0058] Figure 4 This is a schematic diagram of a large-model-based chronic disease patient screening and personalized treatment system provided in an embodiment of this application. Detailed Implementation

[0059] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0060] See Figure 1 The diagram shows a flowchart of a method for screening and personalized treatment of chronic disease patients based on a large model, provided in an embodiment of this application. The apparatus includes steps S101 to S105, wherein:

[0061] S101: Obtain object data of the target entity and generate an object data set, wherein the object data set includes target encoded data with hierarchical relationship and continuous time-series signal data collected by wearable device;

[0062] S102: Process the target encoded data through a first preset model to generate a first embedded representation of the target encoded data, the first embedded representation being located in a preset non-Euclidean feature space; process the continuous time-series signal data through a second preset model to generate a second embedded representation of the continuous time-series signal data;

[0063] S103: In the non-Euclidean feature space, perform a fusion operation on the first embedding representation and the second embedding representation to generate an aligned feature set;

[0064] S104: Call the large model trained by federated distillation to process the alignment feature set to generate a set of recall indicators that characterize the recall priority of the target individual; if the recall priority in the recall indicator set meets the preset recall threshold, the corresponding individual is identified as the target recall individual and a recall notification is sent; after the target recall individual completes the target interaction, it is registered as a target management individual.

[0065] S105: Based on the object type indicated by the analysis results of the target management individual, match or generate a personalized processing path scheme from the preset management path library, and assign the scheme to the target management individual.

[0066] This application can be applied to medical information processing systems, such as computing devices deployed on servers or cloud platforms. The method includes multiple steps, and step S101 is described in detail below.

[0067] In practical implementation, the target entity is first identified. This target entity can be a single, pre-defined medical institution or a collection of multiple medical institutions, such as a hospital, clinic, or member institution of a medical alliance. In a multi-institutional environment, the system can obtain object data from the clinical data repository (CDR) or electronic health record (EHR) of each institution through secure interfaces, such as the application programming interface (API) based on the Health Level Seven Fast Healthcare Interoperability Resources (HL7 FHIR) standard. Specifically, the object data may include historical target interaction records, real-time monitoring data, etc. This data can be obtained through batch import of historical object data or by capturing newly generated data in real-time incrementally to ensure the timeliness and completeness of the data.

[0068] The acquired data needs to undergo preprocessing, such as data cleaning and standardization. This includes converting codes from different institutions into the standard International Classification of Diseases, 10th Revision (ICD-10) or Systematized Nomenclature of Medicine—Clinical Terms (SNOMED CT) format, before generating an object dataset. This object dataset is a structured dataset, which can be organized using database tables or vector storage, and includes two main types of data:

[0069] The first type can be target coding data with a hierarchical relationship. This data comes from an individual's medical records, such as International Classification of Diseases (ICD) codes or surgical codes. These codes have a tree-like hierarchical relationship, such as the hierarchical structure of chapter-category-subcategory-specific code in ICD-10.

[0070] For example, for individuals with chronic diseases, this might include codes such as E11 (type 2 diabetes) and its sub-code E11.2 (with kidney complications) to reflect the classification and association of the disease. This data can be extracted from the EHR system and constructed as a graph structure or nested list for easier subsequent processing.

[0071] The second type can be continuous time-series signal data collected by wearable devices: This data is collected through monitoring devices such as smart bracelets, heart rate monitors, or continuous glucose monitors, and is in the form of time-series signals, such as continuous value sequences of heart rate, blood pressure, blood glucose levels, or activity levels. This data can be transmitted to the medical institution's system via Bluetooth or cloud synchronization interfaces. The sampling frequency can be once per minute or once per hour, and the data length depends on the monitoring period, such as 24-hour continuous recording.

[0072] When generating the object dataset, a unique identifier (such as an individual ID) can be assigned to each individual, and the two types of data mentioned above can be stored together, for example, using JavaScript Object Notation (JSON) format or Pandas DataFrame structure, to facilitate subsequent analysis. This step ensures the integration of multimodal data, providing a comprehensive data foundation for the screening and management of individuals with chronic diseases.

[0073] In practice, this step can be performed by the acquisition module, which has the processing capacity to support large-scale object data, such as processing tens of thousands of records per day, and complies with data privacy regulations, such as the General Data Protection Regulation (GDPR) or the Health Insurance Portability and Accountability Act (HIPAA), protecting sensitive information through encrypted transmission and access control.

[0074] Regarding S102 above:

[0075] In specific implementation, step S102 includes two parallel processing sub-processes, which perform feature extraction and embedding representation generation for the target encoded data and continuous time-series signal data, respectively, in order to capture the inherent structure and dynamic characteristics of the data and provide a unified representation for subsequent fusion.

[0076] First, the target encoded data is processed by a first preset model to generate a first embedded representation of the target encoded data. Specifically, the first preset model can be a preset graph representation learning model, such as a variant based on a graph neural network, specifically designed for processing structured data with hierarchical relationships. This model first converts the target encoded data into a format suitable for graph processing, for example, treating the encoding as nodes and defining connections based on hierarchical relationships to form a graph structure. Then, through multi-level aggregation and transformation operations, the hierarchical features of the nodes are extracted, ultimately generating the first embedded representation. This first embedded representation resides in a preset non-Euclidean feature space, such as hyperbolic space or spherical space. This spatial design helps to better represent tree-like or hierarchical data, avoiding the dimensionality explosion problem in traditional Euclidean space.

[0077] For example, the dimension of the embedded representation can be set according to the application scenario, such as a 128-dimensional or 256-dimensional vector, and model training and inference can be implemented using libraries such as PyTorch Geometric. For the ICD-10 encoding E11 and its sub-encodings of diabetic individuals, the model can capture hierarchical associations from general disease categories to specific complications, improving the semantic richness of the representation.

[0078] Secondly, the continuous time-series signal data is processed by a second preset model to generate a second embedded representation of the continuous time-series signal data. Specifically, the second preset model can be a preset time-series processing model, such as a variant based on a recurrent neural network, specifically designed to capture long-term dependencies and patterns in time series. This model takes continuous time-series signal data as input, such as multi-channel time series, like heart rate and blood glucose sequences, extracts hidden features through iterative calculation, and fuses contextual information to generate the second embedded representation. This embedded representation can be a fixed-length vector, for example, by summarizing sequence features through pooling or attention mechanisms, with dimensions matching the first embedded representation for subsequent operations.

[0079] In practical implementation, for 24-hour heart rate monitoring data, the model can identify abnormal fluctuation patterns, such as elevated heart rate at night, and generate a compressed representation reflecting an individual's physiological state. The model can be implemented using TensorFlow or PyTorch frameworks, supporting GPU acceleration for processing large-scale sequence data.

[0080] This ensures low dimensionality and high information density in the embedded representation, laying the foundation for further integration of multimodal data while maintaining computational efficiency.

[0081] Regarding the above S103:

[0082] In specific implementation, step S103 performs a fusion operation on the first and second embedded representations to achieve alignment of multimodal features in the non-Euclidean feature space. Specifically, this fusion operation aims to integrate the hierarchical information of the target encoded data with the temporal dynamics of the continuous time-series signal data to generate a unified set of aligned features, which facilitates subsequent model processing.

[0083] First, the second embedding representation is projected or adapted into the same non-Euclidean feature space as the first embedding representation to ensure that the two are geometrically compatible.

[0084] The choice of this space, such as a hyperbolic or spherical space, helps to handle hierarchical and nonlinear relationships and avoids information loss as in traditional spaces. Then, a combination operation is performed within this space, such as through weighted fusion, concatenation, or geometric transformation, to integrate the two embedded representations into a single feature representation. This operation can take into account the relative importance of features, for example, by adjusting based on attention mechanisms or preset weights to highlight key patterns.

[0085] Finally, through mapping or optimization steps, an aligned feature set is generated. This set can be a sequence of vectors or a tensor, with dimensions set according to the input embedding; for example, after fusion, it may be a 512-dimensional feature set, used to characterize the individual's overall state. For example, for hierarchical encoding and heart rate sequences of diabetic individuals, this fusion can capture the correlation between disease classification and physiological fluctuations, forming a more comprehensive risk profile. The fusion operation can be implemented using geometry learning extension libraries such as PyTorch or TensorFlow, supporting parallel computation to process batches of object data.

[0086] This ensures the accuracy and computational efficiency of feature alignment, providing a robust foundation for subsequent analysis.

[0087] Regarding S104 above:

[0088] In its implementation, step S104 involves calling a large model trained through federated distillation for prediction and performing individual recall and registration operations based on the results. Specifically, this step aims to generate quantitative indicators using aligned feature sets to support intelligent screening and management decisions for individuals with chronic diseases.

[0089] First, a large model trained via federated distillation is invoked to process the aligned feature set. This large model can be a pre-defined deep learning model, such as a variant based on the Transformer architecture, trained in a multi-institutional healthcare environment via federated distillation to achieve knowledge transfer without directly sharing the original data, thereby protecting individual privacy. The model takes the aligned feature set as input, processes the features (e.g., through fully connected layers or attention mechanisms), and outputs a set of recall metrics. This set of recall metrics includes multiple dimensions, such as risk scores and probability values, where recall priority can be a comprehensive score used to characterize the degree to which an individual is recalled.

[0090] Then, if the recall priority in the recall indicator set meets a preset recall threshold, such as a score higher than 0.7, the corresponding individual is identified as a target recall individual, and a recall notification is sent. This notification can be delivered via SMS, email, or mobile app push notification, and includes target interaction suggestions and appointment links to improve individual response rates. After the target recall individual completes the target interaction, for example, by confirming the target interaction record through the system, they are registered as a target management individual. This registration process can be automatically integrated into the medical institution's management platform, for example, updating individual files and allocating tracking resources.

[0091] Understandably, the threshold can be dynamically adjusted based on clinical standards.

[0092] For example, for an individual with diabetes and a set of aligned features, the large model can generate a recall priority score of 0.85, send a notification when the threshold is met, and add the individual to the management queue after confirmation of the target interaction. This ensures the accuracy of individual screening and the continuity of the management process, supporting subsequent personalized interventions.

[0093] Regarding the above S105:

[0094] In practice, step S105 matches or generates pathway plans based on the analysis results of the target management individuals and assigns them to the individuals. Specifically, this step aims to provide targeted intervention plans for registered target management individuals, thereby achieving personalized and refined chronic disease management.

[0095] First, the analysis results for the target managed individual are obtained. These results can originate from target interaction records, such as standard disease codes or clinical reports extracted through the EHR system, indicating object types such as type 2 diabetes, hypertension, etc. The results may include auxiliary details such as disease severity and complication information to refine path selection.

[0096] Then, personalized processing path solutions are matched or generated from a pre-defined management path library. This management path library is a pre-defined database or knowledge base that stores various standardized path templates, such as paths built based on clinical guidelines, including modules for dietary guidance, drug treatment, exercise plans, regular monitoring, and follow-up arrangements. The matching process can employ rule-based or similarity calculation methods. For example, it queries the library for corresponding templates based on object type; if a direct match exists, it is selected; if no exact match is found or personalized adjustments are needed, a new solution is generated. For instance, it modifies template parameters using a rule engine combined with individual age, gender, physiological indicators, and other data to form a custom path. This generation can use pre-defined algorithms, such as decision trees or simple neural networks, to ensure the feasibility and safety of the path.

[0097] Finally, the treatment plan is assigned to the target individual. This assignment process can be completed through a digital platform, such as pushing the plan content to the individual's mobile health application (APP), doctor's workstation interface, or integrating it into the individual's electronic medical record system. The plan content is typically presented in the form of a structured document, which can be in PDF format or designed as an interactive web page. The plan will include detailed schedules, reminder mechanisms, and tracking modules for monitoring the treatment process, thereby improving individual adherence and management effectiveness.

[0098] For example, for an individual diagnosed with type 2 diabetes, the matching pathway could include daily blood glucose monitoring, weekly exercise goals, and dietary recommendations. If the individual has kidney complications, an adjusted pathway could be generated to increase the frequency of kidney function tests.

[0099] This ensures the targeted and effective management approach, improving the overall health of individuals with chronic diseases.

[0100] Optional, see Figure 2 The flowchart illustrates a first embedded representation method for generating the target encoded data, provided in an embodiment of this application, comprising steps S201 to S204, wherein:

[0101] S201: Convert the target encoded data into nodes, and construct edges based on the hierarchical relationship to generate an initial hierarchical graph;

[0102] S202: Perform a neighbor aggregation operation on the initial hierarchical graph using a hyperbolic graph convolutional network to generate a hierarchical graph representation;

[0103] S203: Perform a hyperbolic convolution operation on the hierarchical graph representation to generate a hierarchical embedding representation;

[0104] S204: Based on the hierarchical embedding representation, generate the first embedding representation, which is located in the non-Euclidean feature space.

[0105] This optional implementation aims to address how to effectively capture the hierarchical relationships of target encoded data to generate more accurate embedding representations, thereby improving the accuracy and robustness of multimodal data processing.

[0106] Specifically, this optional implementation refines the process of generating the first embedded representation of the target encoded data in step S102. First, the system converts the target encoded data into nodes. Specifically, each diagnostic code is considered an independent node. For example, the International Classification of Diseases, 10th Revision (ICD-10) code E11 represents type 2 diabetes and can be converted into one node, while its sub-code E11.2 represents type 2 diabetes with renal complications and can be converted into another node. Each node can carry corresponding attribute information, such as the descriptive text or statistical frequency corresponding to the code. The initial feature vector of a node can be initialized through a preset node encoder, which can be a simple embedding structure that maps the diagnostic code string into a fixed-dimensional low-dimensional vector using word embedding technology, for example, 64 dimensions, to preserve necessary semantic information.

[0107] Subsequently, the system constructs connecting edges between nodes based on the hierarchical relationship between codes. Specifically, according to the tree structure of the diagnostic codes, directed edges are established between parent and child nodes to represent the hierarchical relationship from general categories to specific subcategories. Each edge can also be assigned a numerical weight, such as a score calculated based on the clinical relevance or co-occurrence frequency between codes, to reflect the strength of the edge association; this value is typically between 0 and 1.

[0108] The graph structure can be constructed using graph building toolkits, such as the NetworkX graph processing library in Python, which combines the nodes and edges to generate a complete initial hierarchical graph. This graph has the structural characteristics of a directed acyclic graph, and the number of nodes in the graph depends on the number of diagnostic codes involved for a specific individual. For example, in the process of building a graph for an individual with a chronic disease, a compact tree structure containing 10 to 50 nodes is usually formed for subsequent processing by the graph neural network.

[0109] Next, a neighbor aggregation operation is performed on the initial hierarchical graph using a hyperbolic graph convolutional network to generate a hierarchical graph representation. Specifically, the hyperbolic graph convolutional network is a predefined variant of a graph neural network designed specifically for non-Euclidean spaces. It first projects the node features of the initial hierarchical graph into hyperbolic space, for example, by using an exponential mapping function to convert Euclidean vectors into hyperbolic vectors.

[0110] Then, a neighbor aggregation operation is performed within the hyperbolic space: for each node, the system collects information about its neighboring nodes, including its child and parent nodes. Subsequently, the system fuses the neighbor features using hyperbolic addition and hyperbolic transformation. The aggregation process involves calculating weights based on edge attributes and then performing a weighted summation of the neighbor features based on these weights. To preserve the curvature of the space, activation functions in the hyperbolic space are also applied, such as the hyperbolic tangent function, as a nonlinear transformation method.

[0111] For example, the aggregation process can be performed in multiple layers, typically two to four layers of convolutional structures, with each layer extending the aggregation scope of a node to its higher-order neighbor nodes. Ultimately, the system generates an updated vector for each node, which constitutes a hierarchical graph representation that effectively captures its local hierarchical context information within the structure.

[0112] Subsequently, hyperbolic convolution operations are performed on the hierarchical graph representation to generate a hierarchical embedding representation. Specifically, hyperbolic convolutional layers are applied to the hierarchical graph representation generated in the previous step: this involves global or multi-hop aggregation, such as further propagating features through message passing mechanisms, aggregating information from the root node downwards or from the leaf nodes upwards. Each convolutional operation includes a linear transformation simulated by hyperbolic matrix multiplication, weighted aggregation of neighboring features, and bias adjustment; all computations are performed in hyperbolic geometry to avoid distortion of the hierarchical data in Euclidean space. After the operation, a hierarchical embedding representation is obtained, such as a fixed-dimensional vector matrix where each node has an independent embedding, reflecting the global hierarchical structure.

[0113] Finally, based on the above hierarchical embedding representation, the system generates a first embedding representation. Specifically, pooling is first performed on the hierarchical embedding representations of all nodes, using average pooling or max pooling methods to compress graph-level features into a single vector; alternatively, the system can introduce an attention mechanism to weight the embedding vectors of different nodes, thereby highlighting key encoded information, such as features related to major disease nodes. The above generation steps ensure that the final first embedding representation lies within the non-Euclidean feature space, typically preserved in hyperbolic space, and its curvature parameters are further adjusted through normalization. The entire process can be extended and implemented using the PyTorch graph neural network library and supports GPU-accelerated computation to meet the needs of large-scale graph data processing.

[0114] For example, for the ICD-10 coding tree of diabetic individuals, this approach can generate a first embedding representation that captures the association from E11 to E11.2, thereby improving sensitivity to complications.

[0115] This ensures that the hierarchical information of the target encoded data is fully utilized, improves the quality of the embedded representation, and provides a more reliable foundation for overall chronic disease screening.

[0116] Optional, see Figure 3 The flowchart illustrates a second embedded representation method for generating the continuous time-series signal data provided in this application embodiment, including steps S301 to S304, wherein:

[0117] S301: The continuous time-series signal data is iteratively processed through a forward long short-term memory network to generate a forward hidden state;

[0118] S302: Iteratively process the reverse sequence of the continuous time-series signal data through a backward long short-term memory network to generate a backward hidden state;

[0119] S303: Fuse the forward and backward hidden states to generate a temporal embedding representation;

[0120] S304: Generate the second embedding representation based on the time-series embedding representation.

[0121] This optional implementation aims to address how to effectively capture the temporal dynamic features in continuous time-series signal data, thereby generating a more accurate embedded representation to improve the accuracy and robustness of multimodal data processing.

[0122] Specifically, this optional implementation refines the process of generating the second embedded representation of continuous time-series signal data in step S102. First, the system iteratively processes the continuous time-series signal data using a Forward Long Short-Term Memory (LSTM) network to generate forward hidden states. The Forward LSTM is a pre-defined recurrent neural network structure designed specifically for modeling sequence data, capable of capturing long-range dependencies. The system organizes multi-channel continuous physiological signals into a time-series input, where each time step corresponds to a set of signal observations, such as heart rate and blood glucose levels. The processing unfolds progressively from the beginning of the sequence: at each time step, the network receives the current input and, combined with the hidden state from the previous time step, makes a comprehensive decision on updating the current state through input gates, forget gates, and output gates. The input gate controls the degree of introduction of new information, the forget gate determines the proportion of historical information retained, and the output gate determines the hidden state output at that time step. This process continues until the end of the sequence, generating a corresponding forward hidden state vector for each time step to represent the dynamic context features from the past to the present.

[0123] Subsequently, the system reverses the time sequence of the original sequence to construct a reversed sequence, and processes it using a Backward Long Short-Term Memory (LSTM) network to generate backward hidden states. A reversed sequence refers to arranging the original time series in reverse chronological order; for example, rearranging a 24-hour heart rate sequence from night to day. The backward LSTM starts with the first time point of the reversed sequence, i.e., the last data point of the original sequence, and iteratively processes the sequence in reverse chronological order, using the same structure and gating mechanism as the forward network to progressively update the hidden states. This process captures the temporal context from future time points back to the present, compensating for potential information blind spots in the forward network when processing later data, and ultimately generating a backward hidden state vector with the same dimensions as the forward hidden states, for example, 128 dimensions.

[0124] Next, the system aligns the forward and backward hidden states at the time step level and merges them into a unified temporal embedding representation. This fusion can be achieved through vector concatenation, merging the forward and backward vectors of each corresponding time step into a higher-dimensional representation, such as merging them into 256 dimensions; alternatively, a weighted summation strategy can be used, where the fusion weights are determined by preset coefficients or a mean strategy. This fusion process ensures the integration of bidirectional dependency information of the sequence, forming a complete temporal context representation. The output is a vector matrix that varies with the sequence length, used to describe the dynamic evolution pattern of the entire continuous signal.

[0125] Subsequently, the system generates a second embedding representation based on the temporal embedding representation. Specifically, an average pooling method can be used to average the embedding vectors of all time steps to generate a fixed-dimensional representation; alternatively, an attention mechanism can be introduced to model the importance of each time step to highlight key events or abnormal patterns, such as feature vectors during moments of drastic heart rate fluctuations. The final generated second embedding representation is a fixed-length vector and maintains dimensionality compatibility with the first embedding representation for subsequent fusion processing in non-Euclidean space.

[0126] For example, when modeling an individual's 24-hour heart rate monitoring data, the system can use this method to generate a second embedding representation that fuses forward and backward states, thereby more sensitively capturing key patterns such as abnormal fluctuations at night.

[0127] This approach enables the full modeling of temporal dependency features in continuous time-series signal data, enhancing the expressive power of the generated embedded representations and providing a more reliable and accurate data foundation for subsequent chronic disease screening and personalized treatment.

[0128] Optionally, the non-Euclidean feature space is a hyperbolic space; the generated aligned feature set includes:

[0129] Perform a logarithmic mapping with respect to the origin of the hyperbolic space on the first embedded representation, project the first embedded representation to the tangent space of the origin, and generate a first tangent space embedded representation;

[0130] In the slicing space, a combination operation is performed on the first slicing space embedding representation and the second embedding representation to generate a fused embedding representation;

[0131] An exponential mapping based on the origin of the hyperbolic space is performed on the fused embedding representation to map the fused embedding representation back to the hyperbolic space, generating the aligned feature set.

[0132] This optional implementation aims to address the problem of effectively fusing multimodal embeddings in hyperbolic space to generate aligned feature sets, thereby improving the accuracy and efficiency of feature integration.

[0133] Specifically, this optional implementation refines the process of generating the aligned feature set in step S103, wherein the non-Euclidean feature space specifically adopts a hyperbolic space model with negative curvature, which is naturally suitable for modeling multimodal data with hierarchical structure and scalability.

[0134] First, the system performs a logarithmic mapping operation on the first embedding representation, using the origin of hyperbolic space as a reference, projecting it onto the tangent space corresponding to the origin to generate a first tangent space embedding representation. In the implementation, the origin of hyperbolic space is set as a predefined zero vector, such as the center of a Poincaré sphere model. The system receives the first embedding representation as a vector input in hyperbolic space and transforms it to the Euclidean space corresponding to the tangent plane of the origin through a logarithmic mapping function. This transformation is equivalent to expanding a local neighborhood on the surface, thus facilitating subsequent linear operations. The generated first tangent space embedding representation retains a feature dimension similar to the original embedding, for example, 128 dimensions, and has a more simplified linear expressive capability.

[0135] Subsequently, within the aforementioned tangent space, the system combines the first tangent space embedding representation with the second embedding representation to generate a fused embedding representation. Before this, the system must ensure that the second embedding representation has been adapted to the same tangent space. Adaptation methods may include linearly projecting the temporal embedding through a predefined mapping module or feature transformation layer to achieve spatial and scale compatibility with graph-structured embeddings. The combination method may include vector concatenation or weighted summation, where the fusion weights can be set by static parameters or automatically learned from training data.

[0136] This operation can simultaneously preserve the disease-level information in the diagnostic coding and the dynamic patterns in the physiological signals, thus forming a fusion embedding representation with higher information density. Its dimension can be set to 256 dimensions, or it can be reduced to 128 dimensions through linear dimensionality reduction.

[0137] Next, the system performs an exponential mapping operation on the fused embedding representation, inversely transforming it from the tangent space back to the hyperbolic space to obtain the final aligned feature set. This mapping, with the origin of the hyperbolic space as a reference, maps the linear embedding representation in the tangent space back to the original non-Euclidean geometry.

[0138] In its implementation, the system performs coordinate transformation using an exponential mapping function, re-adapting the fused vectors to the hyperbolic metric, thereby generating a multimodal embedding result containing complete semantic and dynamic information. The generated aligned feature set can be a set of vectors or a tensor structure, representing the overall state of an individual within a specific time window.

[0139] In terms of technical implementation, the system can build the above hyperbolic mapping and tangent space operation process based on the GeoOpt geometry optimization learning library, and combine it with the PyTorch framework to realize automatic differentiation and batch processing capabilities, thereby adapting to real-time fusion computing tasks of multiple individuals.

[0140] For example, when analyzing individuals with diabetes, the system can fuse the hierarchical graph embedding formed by diagnostic codes with the temporal embedding of continuous heart rate data in the tangent space, and map the fusion result back to hyperbolic space to form an aligned feature set. This feature set can more sensitively capture the potential progression trend and physiological fluctuation characteristics of the disease, improving the comprehensive assessment ability of chronic disease risk.

[0141] This enables structured alignment and fusion of multimodal features in hyperbolic space, enhancing the expressiveness and usability of embedded expressions and providing a higher-quality feature foundation for subsequent chronic disease screening, classification, and personalized treatment.

[0142] Optionally, the generation of the fused embedding representation includes:

[0143] Based on the first slicing space embedding representation, the corresponding target management path is extracted from the preset management path library, and a path embedding representation representing the target management path is generated;

[0144] The path embedding representation is used to generate a channel attention vector, which is used to indicate the relative importance of different signal channels in the continuous time-series signal data;

[0145] The first fusion sub-step is executed, in which the channel attention vector is applied to the forward hidden state for weighting, and the weighted forward hidden state is fused with the first tangent space embedding representation to generate a forward context representation;

[0146] The second fusion sub-step is executed, in which the channel attention vector is applied to weight the backward hidden state, and the weighted backward hidden state is fused with the first tangent space embedding representation to generate a backward context representation;

[0147] The forward context representation and the backward context representation are finally fused to generate the fused embedded representation.

[0148] This optional implementation aims to address how to effectively integrate management path information to enhance the relevance of the fusion embedding representation, thereby improving the adaptability and fusion efficiency of chronic disease management paths.

[0149] Specifically, this optional implementation refines the processing flow for generating the fused embedding representation in step S103. First, based on the first scissor space embedding representation, the system retrieves and extracts the corresponding target management path from a pre-defined management path library, and generates a path embedding representation to characterize that path. The management path library stores various predefined path templates, which can be derived from standard intervention sequences in clinical practice guidelines, covering modules such as diet control, medication use, and physiological monitoring. The extraction process is implemented through a similarity matching mechanism: the system uses the first scissor space embedding representation as a query vector, compares it with the vector representations of each template in the path library, and selects the most similar template as the target management path through similarity calculation, such as using cosine similarity.

[0150] Subsequently, the system calls a preset embedding generator to process the structured description information of the selected path template, such as text description or modular sequence. This generator can adopt a multi-layer perceptron (MLP) structure and finally generate a fixed-length path embedding representation, such as 128 dimensions, to capture the semantic and intervention sequence features contained in the path.

[0151] Next, the system generates channel attention vectors based on the path embedding representation to indicate the relative importance of each signal channel in continuous time-series signal data.

[0152] In this process, the system inputs the path embedding representation into the attention generation module. First, a fully connected neural network is used for feature transformation. Then, a normalization function is applied to process the output, resulting in a vector-like set of attention weights. Each element of this vector corresponds to the attention coefficient of a physiological signal channel, for example, used to distinguish the importance of the heart rate channel from the blood glucose channel. The sum of the weights of all channels equals 1. The dimension of the attention vector depends on the number of signal channels, typically 2 to 5 channels, highlighting the core physiological indicators guided by the management path, such as emphasizing the salience of the blood glucose channel in a diabetes pathway.

[0153] Subsequently, the system executes the first fusion sub-step, which applies the aforementioned channel attention vector to the forward hidden state and fuses it with the first tangent space embedding representation to generate the forward context representation.

[0154] Specifically, the system first multiplies each channel dimension of the forward hidden state by the corresponding attention weight to achieve weighted enhancement of key signals. The weighted result is then fused with the first tangent space embedding representation. This fusion can be achieved through vector concatenation or element-wise weighted summation, for example, by expanding the two vectors to a unified dimension before performing element-wise addition. The final generated forward context representation is the fused representation vector, which can have a dimension of, for example, 256, to integrate forward temporal dynamics and hierarchical information.

[0155] The system executes the second fusion sub-step in the same manner, applying channel attention vectors to weight the backward hidden state and fusing the result with the first tangent space embedding representation to generate the backward context representation. The specific process is consistent with the aforementioned steps, ensuring that the backward sequence context is also included in the fused representation, thus forming a complete bidirectional embedding representation structure. The generated backward context representation can also be set to 256 dimensions to ensure consistency in the information structure.

[0156] Finally, the system performs a final fusion of the forward and backward context representations to generate a fused embedding representation. The fusion method can employ a vector concatenation followed by dimensionality reduction, or a weighted averaging strategy for feature compression, such as compressing the concatenated 512-dimensional vector to 128 dimensions using a linear layer. This fusion operation is performed in the tangent space, avoiding complex geometric calculations in hyperbolic space and ensuring operational efficiency and numerical stability.

[0157] For example, when processing data on diabetic subjects, the system can extract the corresponding blood glucose management path based on the diagnostic coding embedding, and enhance the representation strength of blood glucose channels in the forward and backward hidden states by generating channel attention vectors. Subsequently, it is fused with hierarchical information to form the final fused embedding representation, thereby improving the ability to identify abnormal physiological patterns and context adaptability.

[0158] In this way, path-guided fusion embedding construction is achieved, enhancing the structural consistency and adaptability of multimodal features, and providing a high-quality feature foundation for subsequent chronic disease screening and personalized treatment path development.

[0159] Optionally, generating the alignment feature set includes:

[0160] Perform the exponential mapping on the fused embedding representation to generate an initial aligned embedding;

[0161] Based on the statistical time-series characteristics of the continuous time-series signal data, a risk adjustment factor characterizing the recall priority of the individual is calculated;

[0162] In the hyperbolic space, the position of the initial alignment embedding is adjusted according to the risk adjustment factor along the geodesic path connecting the origin and the initial alignment embedding to generate the alignment feature set.

[0163] In this optional implementation, the aim is to address how to dynamically adjust the position of the fused embedding in hyperbolic space based on temporal features to generate a more risk-sensitive aligned feature set, thereby improving the accuracy of individual recall priority determination. The core principle lies in dynamically adjusting the embedding position along a geodesic path in hyperbolic space by introducing a risk adjustment factor. Leveraging the nonlinear amplification effect of hyperbolic geometry, the influence of high-risk signals is more effectively emphasized while reducing interference from low-risk signals, thus improving the match between the feature set and actual clinical recall needs, and enhancing the reliability and personalized adaptability of the screening process.

[0164] Specifically, this embodiment refines the process of generating the alignment feature set in step S103, including the following key steps:

[0165] First, an exponential mapping operation is performed on the fused multimodal embedding representation to generate an initial aligned embedding. Specifically, the system takes the fused embedding representation in the tangent space as input and performs an inverse projection using the exponential mapping function with the origin of hyperbolic space as the reference. To avoid vectors exceeding the model's defined boundaries, the system scales vectors exceeding the length limit before projection, keeping them within the range allowed by the hyperbolic space model, such as a Poincaré sphere. This mapping process maintains the original embedding dimension unchanged and achieves a smooth transition from Euclidean tangent space to hyperbolic space, ensuring that the embedding vectors structurally conform to hyperbolic metric constraints, providing a foundation for subsequent geometric adjustments.

[0166] Next, the system calculates statistical time-series characteristics representing an individual's state based on continuous time-series signal data, such as mean, standard deviation, trend slope, peak frequency, or coefficient of variation, and extracts representative indicators reflecting the individual's risk level. Based on this, the system inputs the statistical characteristics to a risk assessment module for processing. This module can be implemented using a multilayer perceptron structure, which may include a nested network composed of a set of fully connected layers and nonlinear activation functions. This module transforms the input characteristics layer by layer and finally outputs an adjustment factor representing the individual's risk level. This adjustment factor is a single scalar used to indicate whether the importance of the corresponding individual characteristics should be strengthened or weakened in subsequent embedding operations. The value range of this risk adjustment factor is usually limited to a preset positive interval to maintain the stability and controllability of feature adjustment.

[0167] After obtaining the risk adjustment factor, the system identifies the geodesic path connecting the origin and the initial embedding in hyperbolic space. This path corresponds to a shortest curve trajectory in hyperbolic space, typically represented by an arc in a Poincaré model. The system scales the distance of this path according to the specific value of the risk adjustment factor: if the factor is greater than 1, the embedding position is expanded away from the origin along the path direction to enhance the embedding's representation of anomalous signals; if the factor is less than 1, the embedding position is moved closer to the origin along the path direction to suppress the influence of low-risk or background signals in the embedding.

[0168] To ensure the accuracy and stability of the embedded adjustment, the system employs a coordinate-by-coordinate iterative approach for position updates during geodesic adjustment. Specifically, the system first calculates the distance between the initial embedded point and the origin, then scales this distance according to a risk adjustment factor, and repositions the vector in hyperbolic space accordingly. After each update, it is ensured that the adjustment result remains within the legal boundaries of the model. This adjustment process typically converges within a finite number of steps to guarantee computational efficiency and geometric consistency.

[0169] Finally, the system generates a set of risk-adjusted aligned embedding vectors, forming a risk-sensitive aligned feature set. This set can more prominently express the key information of high-risk individuals, helping the model identify key targets when performing recall priority judgment, and improving the clinical relevance and practicality of the recall task.

[0170] At the engineering implementation level, the system can be deployed based on existing geometric optimization libraries, such as GeoOpt (Geometric Optimization Library). GeoOpt has full support for hyperbolic space modeling and optimization, can handle batch adjustment operations with constrained embeddings, and provides the necessary optimizer support, making it suitable for large-scale clinical data processing scenarios.

[0171] For example, when dealing with individuals with diabetes among those with chronic diseases, the system can analyze the temporal characteristics of blood glucose fluctuations to generate risk adjustment factors for embedding expansion, and adjust the embedding position in hyperbolic space accordingly, so that key abnormal features are more clearly expressed in the feature set, thereby improving the ability of subsequent models to identify high-risk individuals.

[0172] Through the above methods, the final generated alignment feature set possesses good risk adaptability and expression accuracy in its geometric structure, which can effectively improve the rationality of recall priority ranking and provide a more reliable and practical data foundation for key population screening, dynamic assessment, and personalized intervention in chronic disease management scenarios. Optionally, the set of recall indicators representing the recall priority of target individuals includes:

[0173] The alignment feature set, the path embedding representation, and the risk adjustment factor are combined to generate enhanced input features;

[0174] The enhanced input features are input into the large model for processing to generate the recall index set.

[0175] This optional implementation aims to address how to effectively integrate multi-source features to generate a more comprehensive and adaptive set of recall indicators, thereby improving the accuracy and targeting of individual recall assessments. Its core principle lies in combining aligned features, path embeddings, and risk adjustment factors into enhanced input features, providing a richer and more structured information foundation for subsequent model inference. This enhances the comprehensive assessment capability of individual status, enabling the generated recall indicators to more accurately reflect individual priority, improving screening efficiency and the reliability of clinical decisions.

[0176] Specifically, this optional implementation refines the process of generating a set of recall metrics representing the recall priority of target individuals in step S104. First, the system integrates the aforementioned aligned feature set, path embedding representation, and risk adjustment factor to generate enhanced input features. In implementation, the system first ensures the dimensionality consistency of each feature module, for example, by converting the path embedding representation into a vector dimension consistent with the aligned feature set through a linear projection layer. Simultaneously, the risk adjustment factor, as a scalar parameter, is expanded into a repeating vector that matches other features to meet the requirements of the concatenated dimension.

[0177] After unifying the dimensions, the system performs a combination operation on the three types of features. The combination method can be a vector concatenation strategy, which concatenates the three features in a fixed order into a whole high-dimensional input vector; or a weighted fusion method, which superimposes and fuses the three types of features in a set ratio to achieve a weight balance of different information sources in the model input.

[0178] Next, the system enhances the input features and feeds them into the large model for processing to generate the final recall metric set. The large model typically employs a multi-layer Transformer structure for feature extraction and modeling. In this structure, each Transformer block first uses a self-attention mechanism to capture internal dependencies from the input, then extracts higher-level abstract representations through a feedforward network, and finally generates multi-dimensional recall metric outputs at the output layer using activation functions such as softmax or sigmoid. This metric set may include multiple sub-items such as risk score, recall probability, and disease priority, and the output is a fixed-length vector, for example, 5 dimensions, to facilitate downstream risk assessment and intervention screening. To ensure privacy compliance in a multi-institutional collaborative environment, a federated distillation mechanism is used during the large model training process to achieve model capability transfer and optimization without sharing the original data.

[0179] For example, when making recall judgments for individuals with diabetes, the system can integrate the individual's embedded features, path guidance, and dynamic risk adjustment factors as inputs. After inference, the large model outputs a recall score, such as 0.85, which corresponds to a higher priority processing suggestion.

[0180] This ensures that the generated recall index set is structurally comprehensive and dynamically adaptable, significantly improving the reliability of identifying high-risk individuals in chronic disease screening tasks and providing efficient support for the initiation of personalized treatment pathways and clinical intervention.

[0181] Optionally, the generation of enhanced input features includes:

[0182] Obtain the identity information of the target entity corresponding to the current object data, and generate an organizational embedding representation that characterizes the data characteristics of the target entity;

[0183] Based on the mechanism embedding representation, generate one or more gating vectors;

[0184] The gating vector is applied to perform a gating fusion operation on the alignment feature set, the path embedding representation, and the risk adjustment factor to generate the enhanced input features.

[0185] This optional implementation aims to address the feature mismatch problem caused by the heterogeneity of data sources in a multi-institutional environment, thereby generating more robust and adaptive enhanced input features and improving the model's generalization ability and recall accuracy in cross-institutional deployment scenarios. This application further introduces institutional embedding representation and a gating mechanism. By dynamically adjusting the feature fusion weights, the system can adjust features according to the data characteristics of different medical institutions, reducing heterogeneity interference, achieving adaptive construction of enhanced input features, and improving the overall model performance and stability.

[0186] Specifically, the process extracts identity information related to the organization from the target entity processing the current object data, and generates an organization embedding representation reflecting the characteristics of the organization's data. In implementation, organization information may include metadata such as organization number, organization level, geographical location, and the scale and quality of the collected data. This information is typically extracted from object data records or system logs and obtained through database indexing. After acquisition, the discrete or categorical attributes are vectorized: this information can be mapped to continuous vectors, for example, through one-hot encoding followed by dimensionality reduction processing in a fully connected network, or by using a word embedding mechanism to generate a fixed-length representation vector to express the organization's unique characteristics in terms of data source, collection mode, and quality.

[0187] Furthermore, one or more gated vectors are generated based on the mechanism embedding representation for subsequent fusion control. The gated vectors are output by a gate generation module, which can be constructed using a multilayer perceptron structure: first, the mechanism embedding representation undergoes feature transformation through hidden layers; then, the output values ​​are normalized to between 0 and 1 using a sigmoid activation function, ultimately forming multiple gated vectors. Each gated vector corresponds to a feature component to be fused, such as an alignment feature set, a path embedding representation, and a risk adjustment factor. The dimension of each gated vector is consistent with its corresponding feature vector.

[0188] The generation process of this gating vector reflects the importance of specific attributes of medical institutions to different types of features. For example, in primary care institutions with large data fluctuations, higher gating weights are assigned to risk adjustment factors to guide the model to pay more attention to individual abnormal patterns.

[0189] Furthermore, the system uses the aforementioned gate vectors to perform a weighted fusion operation on each feature vector, generating the final enhanced input features. Specifically, the system performs element-wise multiplication operations on the alignment feature set, path embedding representation, and risk adjustment factor with their respective gate vectors to achieve weight control. Subsequently, the weighted features of the three classes are fused, and a high-dimensional representation can be generated by vector concatenation.

[0190] Alternatively, element-wise weighted averaging can be used to integrate feature content. The above fusion process ensures that the fusion result fully reflects the characteristics of the institutional data and maintains format consistency with the large model input interface.

[0191] For example, when processing individuals from a community clinic, the system-generated institutional embedding representation can reflect the institution's limitations in terms of equipment configuration and data scale, thereby assigning higher fusion weights to risk adjustment factors, so that the final input features pay more attention to fluctuation anomalies in continuous physiological signals.

[0192] This enables the enhanced input features to adapt structurally to multi-institutional environments, improving the model's recall accuracy and stability in different medical scenarios, and providing cross-institutional consistent technical support for the task of screening and managing individuals with chronic diseases.

[0193] Optionally, the large model trained by the federated distillation includes:

[0194] Deploy a local teacher model for each of the target entities;

[0195] The local teacher model processes the alignment feature set to generate an intermediate layer feature representation, which serves as local knowledge, characterizing the fusion state of the target encoded data and the continuous temporal signal data.

[0196] A global student model is deployed on the central server as the larger model;

[0197] The global student model receives the intermediate layer feature representation from one or more of the local teacher models, and updates the parameters of the global student model based on the principle of minimizing the difference between the feature representation generated by the global student model itself and the intermediate layer feature representation.

[0198] This optional implementation aims to address the challenges of data privacy protection and model training efficiency in collaborative modeling across multiple healthcare institutions, thereby achieving efficient knowledge transfer and optimized global model training. This implementation employs a federated distillation mechanism as its core strategy. Without exchanging original object data, it updates model parameters by transmitting intermediate layer feature representations, enhancing privacy protection while improving the generalization performance and training efficiency of cross-institutional models.

[0199] Specifically, this optional implementation refines the federated distillation training process for the large model involved in step S104. First, an independent local teacher model is deployed within each target entity. A dedicated model instance can be configured for each institution, and this model structure can employ a deep neural network framework with a multi-layered Transformer structure, running on each institution's local server environment.

[0200] The teacher model can be initialized based on pre-trained parameters or historical data within the institution, ensuring that it always uses only local data during the training process, thereby avoiding the flow of raw data between institutions and complying with data sovereignty and privacy requirements.

[0201] Next, the local teacher model receives the alignment feature set corresponding to the internal object data of the institution and generates a set of intermediate layer feature representations, which are used to reflect the abstract knowledge information in the fusion state of the target encoded data and continuous time-series signal data.

[0202] During model execution, the system can select activation results generated by intermediate layers of the model and extract them to form intermediate vector representations, which are used to capture the multimodal semantics of fused features. This intermediate representation does not contain any original data fields, but only reflects the encoded and compressed model knowledge, and therefore can be used as a shareable and secure representation for subsequent uploading.

[0203] Subsequently, a global student model is deployed on a central server to aggregate and integrate knowledge information from multiple local models. The student model structure can be consistent with the local teacher model or a lightweight structure can be adopted. Its initial parameters can be random values ​​or derived from a third-party pre-trained model. This model is deployed on a central control node or cloud server to centrally complete parameter optimization tasks.

[0204] During training, the central server periodically receives intermediate feature representations from various medical institutions and inputs them into the global student model for alignment training. The system calculates the difference between the feature representations generated by the student model and the intermediate feature representations uploaded by the local teacher model under the same input conditions, and uses this difference as input to the loss function to guide the update of the model parameters. The entire training process is conducted through a secure channel, transmitting anonymized identifiers and vector data without transmitting any original medical records.

[0205] In practice, the model parameters can be updated using the standard backpropagation algorithm, and the model can be trained iteratively round by round using optimizers such as gradient descent until the feature fit of the global student model reaches the preset convergence criterion.

[0206] For example, in a medical consortium comprising multiple hospitals, each institution can run a teacher model locally to generate multimodal fusion feature representations of the target data and upload them to a central system periodically. The student models, on the other hand, learn local knowledge by minimizing the difference between their predictions and the uploaded features. In practice, this process can run on an hourly basis, ensuring a reasonable balance between training efficiency and model update frequency.

[0207] This enables knowledge integration and cross-institutional modeling capabilities while ensuring privacy and security, effectively improving the robustness and recall accuracy of large models in multi-source environments, and providing a safe, reliable, and efficient model support system for chronic disease screening and personalized treatment.

[0208] Based on the same inventive concept, this application also provides a system for screening and personalized treatment of chronic diseases based on a large model, which corresponds to the method for screening and personalized treatment of chronic diseases based on a large model. Since the principle of the system in this application is similar to the method for screening and personalized treatment of chronic diseases based on a large model in this application, the implementation of the system can refer to the implementation of the method, and the repeated parts will not be described again.

[0209] Reference Figure 4 The diagram shown is a schematic of a large-model-based chronic disease patient screening and personalized treatment system provided in an embodiment of this application. The system includes:

[0210] The acquisition module 10 is used to acquire object data of the target entity and generate an object data set, wherein the object data set includes target encoded data with hierarchical relationship and continuous time-series signal data acquired by the wearable device;

[0211] Processing module 20 is configured to process the target encoded data through a first preset model to generate a first embedded representation of the target encoded data, the first embedded representation being located in a preset non-Euclidean feature space; process the continuous time-series signal data through a second preset model to generate a second embedded representation of the continuous time-series signal data; and perform a fusion operation on the first embedded representation and the second embedded representation in the non-Euclidean feature space to generate an aligned feature set.

[0212] The recall module 30 is used to call a large model trained by federated distillation to process the alignment feature set to generate a set of recall indicators that characterize the recall priority of the target individual; if the recall priority in the recall indicator set meets the preset recall threshold, the corresponding individual is identified as the target recall individual and a recall notification is sent; after the target recall individual completes the target interaction, it is registered as a target management individual.

[0213] The management module 40 is used to match or generate a personalized processing path scheme from a preset management path library based on the object type indicated by the analysis results of the target management individual, and to assign the scheme to the target management individual.

[0214] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for screening and personalizing chronic disease patients based on large models, characterized in that, The method comprises the following steps: acquiring object data of a target entity and generating an object data set, the object data set comprising target coding data having a hierarchical relationship and continuous time series signal data collected by a wearable device; processing the target coding data through a first preset model to generate a first embedding representation of the target coding data, the first embedding representation being located in a preset non-Euclidean feature space; processing the continuous time series signal data through a second preset model to generate a second embedding representation of the continuous time series signal data; performing a fusion operation on the first embedding representation and the second embedding representation in the non-Euclidean feature space to generate an aligned feature set; calling a large model trained through federated distillation to process the aligned feature set to generate a recall indicator set representing the recall priority of a target individual; if the recall priority in the recall indicator set meets a preset recall threshold, determining the corresponding individual as a target recall individual and sending a recall notification; after the target recall individual completes a target interaction, registering the individual as a target management individual; according to the object type indicated by the analysis result of the target management individual, matching or generating a personalized processing path scheme from a preset management path library and assigning the scheme to the target management individual. 2.The large model-based chronic disease patient screening and individualized treatment method according to claim 1, wherein, The first embedding representation of the target coding data is generated by: converting the target coding data into nodes and constructing edges based on the hierarchical relationship to generate an initial hierarchical graph; performing a neighbor aggregation operation on the initial hierarchical graph through a hyperbolic graph convolutional network to generate a hierarchical graph representation; performing a hyperbolic convolution operation on the hierarchical graph representation to generate a hierarchical embedding representation; based on the hierarchical embedding representation, generating the first embedding representation, which is located in the non-Euclidean feature space. 3.The large model-based chronic disease patient screening and individualized treatment method of claim 2, wherein, The second embedding representation of the continuous time series signal data is generated by: iteratively processing the continuous time series signal data through a forward long short-term memory network to generate a forward hidden state; iteratively processing the reverse sequence of the continuous time series signal data through a backward long short-term memory network to generate a backward hidden state; fusing the forward and backward hidden states to generate a time series embedding representation; based on the time series embedding representation, generating the second embedding representation. 4.The method according to claim 3, wherein, The non-Euclidean feature space is a hyperbolic space; the generation of the aligned feature set comprises: performing a logarithmic mapping of the first embedding representation with reference to the origin of the hyperbolic space, projecting the first embedding representation to the tangent space of the origin to generate a first tangent space embedding representation; performing a combination operation on the first tangent space embedding representation and the second embedding representation in the tangent space to generate a fusion embedding representation; performing an exponential mapping of the fusion embedding representation with reference to the origin of the hyperbolic space to map the fusion embedding representation back to the hyperbolic space to generate the aligned feature set. 5.The method according to claim 4, wherein, The generation of the fusion embedding representation comprises: based on the first tangent space embedding representation, extracting a corresponding target management path from the preset management path library and generating a path embedding representation representing the target management path; generate a channel attention vector, which is used to indicate relative importance of different signal channels in the continuous time-series signal data, using the path embedding representation; perform a first fusion sub-step, apply the channel attention vector to weight the forward hidden state, and fuse the weighted forward hidden state with the first tangent space embedding representation to generate a forward context representation; perform a second fusion sub-step, apply the channel attention vector to weight the backward hidden state, and fuse the weighted backward hidden state with the first tangent space embedding representation to generate a backward context representation; perform a final fusion on the forward context representation and the backward context representation to generate the fused embedding representation. 6.The method according to claim 5, wherein, The generating the alignment feature set comprises: performing the exponential mapping on the fused embedding representation to generate an initial alignment embedding; calculating a risk adjustment factor representing the individual recall priority based on statistical time-series features of the continuous time-series signal data; adjusting a position of the initial alignment embedding according to the risk adjustment factor along a geodesic path connecting an origin point and the initial alignment embedding in the hyperbolic space to generate the alignment feature set. 7.The large model-based chronic disease patient screening and individualized treatment method according to claim 6, wherein, The generating the recall indicator set representing the target individual recall priority comprises: combining the alignment feature set, the path embedding representation, and the risk adjustment factor to generate an enhanced input feature; inputting the enhanced input feature into the large model for processing to generate the recall indicator set. 8.The large model-based chronic disease patient screening and individualized treatment method according to claim 7, wherein, The generating the enhanced input feature comprises: obtaining identity information of the target entity corresponding to the current object data, and generating an organization embedding representation representing data characteristics of the target entity; generating one or more gating vectors based on the organization embedding representation; applying the gating vectors to perform a gated fusion operation on the alignment feature set, the path embedding representation, and the risk adjustment factor to generate the enhanced input feature. 9.The large model-based chronic disease patient screening and individualized treatment method of claim 1, wherein, The large model trained by federated distillation comprises: deploying a local teacher model at each target entity; processing the alignment feature set by the local teacher model to generate an intermediate layer feature representation representing a fusion state of the target encoded data and the continuous time-series signal data as local knowledge; deploying a global student model as the large model at a central server; the global student model receives the intermediate layer feature representation from one or more local teacher models, and updates parameters of the global student model based on a principle of minimizing differences between a feature representation generated by the global student model itself and the intermediate layer feature representation.

10. A chronic disease patient screening and individualized treatment system based on a large model, characterized in that, comprise: a collection module configured to obtain object data of a target entity and generate an object data set, the object data set comprising target encoded data having a hierarchical relationship and continuous time-series signal data collected by a wearable device; a processing module configured to process the target encoded data by a first preset model to generate a first embedding representation of the target encoded data, the first embedding representation being located in a preset non-Euclidean feature space; generate a second embedding representation of the continuous time-series signal data by processing the continuous time-series signal data through a second preset model; perform a fusion operation on the first embedding representation and the second embedding representation in the non-Euclidean feature space to generate an aligned feature set; a recall module configured to call a large model trained through federated distillation to process the aligned feature set to generate a recall indicator set representing a recall priority of a target individual; if the recall priority in the recall indicator set meets a preset recall threshold, the corresponding individual is determined as a target recall individual, and a recall notification is sent; after the target recall individual completes a target interaction, the target recall individual is registered as a target management individual; a management module configured to match or generate an individualized processing path scheme from a preset management path library according to an object type indicated by an analysis result of the target management individual, and assign the scheme to the target management individual.

Images