An intelligent intervention strategy generation method for whole-cycle health management

Abstract

The application discloses an intelligent intervention strategy generation method for whole-cycle health management, and belongs to the technical field of health management, and specifically comprises the following steps: acquiring heterogeneous health original data, inputting the heterogeneous health original data into a modal alignment network, extracting context correlation information through a cross attention mechanism, and generating a fixed-dimension cross-modal feature representation; inputting the cross-modal feature representation into a hypergraph topology construction module according to a time sequence, taking time sequence mapping hyperedges and feature nodes as vertices, and constructing a time sequence health hypergraph structure containing a time dimension; outputting a health time evolution feature matrix expressing the evolution relationship of time nodes; inputting the health time evolution feature matrix into a hidden Markov model to generate a dynamic digital health portrait matching the above-mentioned hidden state; introducing the dynamic digital health portrait into a reinforcement learning decision framework to output an action sequence as a candidate intervention path; and decoding and mapping the action sequence in combination with the constraint rules of the dynamic digital health portrait to generate an intelligent intervention strategy.

Description

Technical Field

[0001] This invention relates to the field of health management technology, and more specifically to a method for generating intelligent intervention strategies for full-cycle health management. Background Technology

[0002] The rapid development of the Internet of Things (IoT), wearable devices, and healthcare information technology has made full-cycle health management a crucial direction in modern digital healthcare. Full-cycle health management aims to continuously collect and track multi-dimensional health data, including physiological indicators, medical records, and lifestyle behaviors, across various stages from prevention, daily care, disease diagnosis, to rehabilitation. This data is then used to comprehensively assess an individual's health status using artificial intelligence and big data analytics. Existing intelligent health management methods typically rely on basic vital sign data or interim medical test reports obtained from data collection terminals. They employ conventional machine learning models or pre-set expert rule bases to classify and predict current health risks, and then output corresponding health guidance suggestions, exercise and dietary plans, or medication intervention reminders. This digitally driven health management model provides the basic technical architecture and implementation path for personalized medicine and proactive health interventions.

[0003] However, existing technologies face technical bottlenecks when processing full-cycle health data, including the difficulty in deeply integrating multimodal heterogeneous data and accurately extracting long-term time-series features. Full-cycle health management involves multidimensional data generated by users at different life stages and health states, which are highly discrete, time-varying, and unstructured. Traditional health analysis methods typically only perform static modeling for local time points or single-modal data, lacking the ability to dynamically capture long-term evolution patterns and perform cross-modal correlation analysis. This limitation in data processing dimensions prevents the system from constructing accurate and coherent personal health digital profiles, resulting in a lack of comprehensive and accurate data support for subsequent intervention strategies, significantly limiting the scientific validity of these strategies. Summary of the Invention

[0004] The purpose of this invention is to provide a method for generating intelligent intervention strategies for full-cycle health management, thereby addressing the problems in the background technology:

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A method for generating intelligent intervention strategies for full-cycle health management includes the following steps:

[0007] S1: Obtain heterogeneous raw health data, input it into a modality alignment network, extract contextual information through a cross-attention mechanism, and generate a fixed-dimensional cross-modal feature representation;

[0008] S2: Input the cross-modal feature representation into the hypergraph topology construction module according to the time series, and construct a temporal health hypergraph structure containing the time dimension with the time series mapping hyperedge and feature nodes as vertices;

[0009] S3: Utilize a spatiotemporal graph network to perform node aggregation calculations on the temporal health hypergraph structure, update the state of feature nodes, and output a health temporal evolution feature matrix that expresses the evolutionary relationship of time nodes;

[0010] S4: Input the health temporal evolution feature matrix into the hidden Markov model, map the transition path of the implicit state, and generate a dynamic digital health profile that matches the above implicit state.

[0011] S5: Import the dynamic digital health profile into the reinforcement learning decision framework, perform decision process traversal calculations in the preset intervention knowledge nodes, and output action sequences as candidate intervention paths.

[0012] S6: Input the candidate intervention path into the autoregressive decoder, combine it with the constraint rules of the dynamic digital health profile to decode and map the action sequence, and generate an intelligent intervention strategy.

[0013] As a further aspect of the present invention: In step S1, the process of acquiring heterogeneous health raw data, inputting it into a modality alignment network, extracting contextual association information through a cross-attention mechanism, and generating a fixed-dimensional cross-modal feature representation is as follows:

[0014] Obtain heterogeneous health raw data, input the heterogeneous health raw data into a multi-branch feature encoder for independent mapping operation, and output the initial feature vectors corresponding to each modality;

[0015] The initial feature vectors are input into the modality alignment network, and the interaction weights between the initial feature vectors are calculated using the cross attention mechanism to extract contextual association information.

[0016] The initial feature vectors are weighted and fused based on contextual information, and the weighted fusion result is transformed through a linear mapping layer to generate a fixed-dimensional cross-modal feature representation.

[0017] As a further aspect of the present invention: In step S2, the process of inputting the cross-modal feature representation into the hypergraph topology construction module according to the time series, and constructing a temporal health hypergraph structure containing the time dimension with the time series mapping hyperedges and feature nodes as vertices, is as follows:

[0018] Receive cross-modal feature representations and corresponding time series, input the cross-modal feature representations into the hypergraph topology construction module, divide the cross-modal feature representations into timestamps according to the time series, and extract feature nodes corresponding to each timestamp;

[0019] Feature nodes are mapped to vertices of the topological data structure, an initial vertex set is established, the health status attribute vector contained in the feature nodes is extracted and assigned to the corresponding vertices, and a vertex association feature matrix with timestamp is generated.

[0020] Using the time interval span in the time series as the mapping condition, vertices belonging to the same time interval span are aggregated and divided, a hyperedge is constructed to enclose the vertices, and an association generation matrix connecting the corresponding hyperedge and the vertex is generated.

[0021] Extract the vertex association feature matrix and association occurrence matrix, and assemble the topology structure of vertices and hyperedges with timestamps inside the hypergraph topology construction module to construct a temporally healthy hypergraph structure containing the time dimension.

[0022] As a further aspect of the present invention: In step S3, the process of using a spatiotemporal graph network to perform node aggregation calculation on the temporal health hypergraph structure, updating the state of feature nodes, and outputting a health temporal evolution feature matrix expressing the evolutionary relationship of time nodes is as follows:

[0023] The temporal health hypergraph structure is input into the graph convolutional layer of the spatiotemporal graph network. Based on the hyperedge generation matrix, the feature nodes inside the same hyperedge are subjected to a feature weighted summation operation to generate the spatial dimension node aggregation calculation result.

[0024] Extract the node aggregation calculation results in the spatial dimension and input them into the nonlinear mapping layer to perform numerical transformation. Use the new feature vector generated by the mapping transformation to replace the original node parameters and update the feature node state.

[0025] All updated feature node states are input into the temporal memory layer of the spatiotemporal graph network in chronological order to calculate the historical state transfer weights between adjacent timestamps and extract the evolution relationship of time nodes.

[0026] Based on the evolution relationship of time nodes, cross-time step feature splicing operation is performed on the updated feature node states in different time series, and all time series feature vectors are summarized to output the healthy temporal evolution feature matrix.

[0027] As a further aspect of the present invention: the specific content of inputting all updated feature node states into the temporal memory layer of the spatiotemporal graph network in chronological order, calculating the historical state transfer weights between adjacent timestamps, and extracting the evolution relationship of time nodes includes:

[0028] Extract the corresponding time stamps, arrange the states of all updated feature nodes according to the time sequence, and input the arranged state sequence into the temporal memory layer of the spatiotemporal graph network.

[0029] The temporal memory layer reads the current time step input state and the hidden state retained by adjacent timestamps, combines the features of the two and inputs them into the internal door control unit to calculate the historical state transfer weight between adjacent timestamps.

[0030] Based on the weights passed from historical states, the hidden states are used to preserve features and update values, outputting a hidden feature matrix that includes step-time dependency attributes, and extracting the evolution relationship at time nodes.

[0031] As a further aspect of the present invention: in step S4, the process of inputting the health temporal evolution feature matrix into the hidden Markov model, mapping the transition path of the implicit state, and generating a dynamic digital health profile matching the aforementioned implicit state is as follows:

[0032] Extract the health temporal evolution feature matrix and input it into a fully connected mapping network. Perform dimensionality reduction projection on the continuous vectors inside the matrix, calculate the observation probability distribution sequence for each timestamp node, and input the observation probability distribution sequence into a hidden Markov model.

[0033] The Hidden Markov Model receives the observation probability distribution sequence, combines it with the state transition probability matrix to perform maximum likelihood path decoding calculation, and outputs the implicit state sequence corresponding to each time stamp node, thereby mapping the transition path of the implicit state.

[0034] Based on the transition path of the implicit state, the implicit state sequence is extracted. The health attribute tags corresponding to each implicit state are retrieved from the preset database. All health attribute tags are concatenated according to the timestamp order to generate a dynamic digital health profile that matches the above implicit state.

[0035] As a further aspect of the present invention: In step S5, the process of importing the dynamic digital health profile into the reinforcement learning decision framework, performing decision process traversal calculations in preset intervention knowledge nodes, and outputting action sequences as candidate intervention paths is as follows:

[0036] The dynamic digital health profile is imported into the reinforcement learning decision-making framework, and its health attribute labels are encoded to generate environmental state features. The operation instructions contained in the preset intervention knowledge nodes are read to construct the decision action space.

[0037] The reinforcement learning decision-making framework performs traversal calculations of the decision-making process in the decision-making action space based on the characteristics of the environmental state. It deduces the expected health state corresponding to each operation instruction through a pre-trained state transition model and calculates the single-step reward value based on the expected health state.

[0038] The cumulative expected return of each traversal branch is obtained by summing the single-step return values ​​over time. The traversal branch with the largest cumulative expected return is selected, and the operation instructions contained therein are concatenated in order. The output action sequence is used as a candidate intervention path.

[0039] As a further aspect of the present invention: In step S6, the process of inputting the candidate intervention path into the autoregressive decoder, combining the constraint rules of the dynamic digital health profile to decode and map the action sequence, and generating an intelligent intervention strategy is as follows:

[0040] Extract the physiological attributes contained in the dynamic digital health profile to construct constraint rules, input the action sequence inside the candidate intervention path into the autoregressive decoder to generate the initial semantic vector for decoding;

[0041] The autoregressive decoder predicts and outputs word units one by one based on the initial semantic vector. During the word unit generation stage, it calls the constraint rules to perform a masking filtering operation on the candidate words and removes illegal word units.

[0042] The tokens retained by the masking filtering operation are mapped to text characters with the highest conditional probability. The token prediction and character concatenation steps are executed cyclically to generate an intelligent intervention strategy for text format.

[0043] The beneficial effects of this invention are:

[0044] This invention breaks down the structural and semantic barriers between heterogeneous raw health data by introducing a modality alignment network and a cross-attention mechanism, achieving deep association and fusion of multimodal data in a fixed dimension. Building upon this, the invention employs a hypergraph topology construction module and a spatiotemporal graph network to construct and aggregate discrete cross-modal features into a temporal health hypergraph structure using time series as the mapping condition. This invention completes the node association calculation of multidimensional complex health indicators in the spatial dimension and accurately captures the long-term dynamic evolution patterns across different life stages of a user in the temporal dimension. It effectively overcomes the problem of incomplete feature extraction caused by traditional single-modality or static time-node modeling, laying a high-dimensional spatiotemporal data foundation for accurate extrapolation of full-cycle health status.

[0045] This invention utilizes Hidden Markov Models to perform deep mapping of implicit state transition paths on the temporal evolution characteristics of health, successfully constructing a dynamic digital health profile with high temporal coherence and evolution tracking capabilities, completely solving the technical shortcomings of traditional discrete and fragmented personal profiles. More importantly, this invention seamlessly integrates this dynamic profile into a reinforcement learning decision-making framework for traversing and optimizing global intervention actions, using an autoregressive decoder combined with physiological constraint rules to perform text decoding and safety mask filtering on candidate actions. This mechanism integrates long-term expected health returns with strict medical safety red lines, ensuring that the final intelligent intervention strategy not only has comprehensive and accurate data support, but also significantly improves the intervention effectiveness of full-cycle health management. Attached Figure Description

[0046] The invention will now be further described with reference to the accompanying drawings.

[0047] Figure 1 This is a flowchart illustrating a method for generating intelligent intervention strategies for full-cycle health management according to the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Please see Figure 1 As shown, this invention provides a method for generating intelligent intervention strategies for full-cycle health management, comprising the following steps:

[0050] S1: Obtain heterogeneous raw health data, input it into a modality alignment network, extract contextual information through a cross-attention mechanism, and generate a fixed-dimensional cross-modal feature representation;

[0051] S2: Input the cross-modal feature representation into the hypergraph topology construction module according to the time series, and construct a temporal health hypergraph structure containing the time dimension with the time series mapping hyperedge and feature nodes as vertices;

[0052] S3: Utilize a spatiotemporal graph network to perform node aggregation calculations on the temporal health hypergraph structure, update the state of feature nodes, and output a health temporal evolution feature matrix that expresses the evolutionary relationship of time nodes;

[0053] S4: Input the health temporal evolution feature matrix into the hidden Markov model, map the transition path of the implicit state, and generate a dynamic digital health profile that matches the above implicit state.

[0054] S5: Import the dynamic digital health profile into the reinforcement learning decision framework, perform decision process traversal calculations in the preset intervention knowledge nodes, and output action sequences as candidate intervention paths.

[0055] S6: Input the candidate intervention path into the autoregressive decoder, combine it with the constraint rules of the dynamic digital health profile to decode and map the action sequence, and generate an intelligent intervention strategy.

[0056] In one embodiment of the present invention, step S1 involves acquiring heterogeneous raw health data, inputting it into a modality alignment network, extracting contextual association information through a cross-attention mechanism, and generating a fixed-dimensional cross-modal feature representation.

[0057] Heterogeneous raw health data is acquired and input into a multi-branch feature encoder for independent mapping operations, outputting initial feature vectors corresponding to each modality. Specifically, the heterogeneous raw health data includes various data types from different monitoring devices and acquisition terminals, such as one-dimensional electrocardiogram (ECG) sequences reflecting heart rate and rhythm, two-dimensional medical images reflecting organ and tissue morphology, and electronic medical record text data recording patient history and diagnoses. These different types of data have completely different data structures and feature distribution dimensions. To effectively process this heterogeneous raw health data, a multi-branch feature encoder is pre-constructed. The multi-branch feature encoder contains multiple parallel processing branches, each specifically responsible for processing data of a particular modality. The extracted one-dimensional ECG sequence signals are input into the corresponding time-series feature extraction branch for convolution operations; the two-dimensional medical images are input into the corresponding spatial image feature extraction branch for pooling and dimensionality reduction operations; and the electronic medical record text data are input into the corresponding natural language processing branch for word vector embedding calculations. Each processing branch within the multi-branch feature encoder performs an independent mapping operation on the input data, transforming the original unstructured or semi-structured data into a high-dimensional feature matrix that the computer can understand, and finally outputting the initial feature vector corresponding to each modality.

[0058] Initial feature vectors are input into the modality alignment network, and the interaction weights between the initial feature vectors are calculated using a cross-attention mechanism to extract contextual information. The initial feature vectors obtained after processing by the multi-branch feature encoder reside in their own independent feature spaces, exhibiting dimensional differences and lacking direct semantic connections. To break down these data modality barriers, the initial feature vectors of all modalities are simultaneously input into the modality alignment network for deep interactive computation. The core component of the modality alignment network includes a multi-head cross-attention mechanism module. During the operation of this module, the initial feature vectors of the first modality are mapped to a query matrix, and the initial feature vectors of the second modality are mapped to a key matrix and a value matrix, respectively. The inner product multiplication of the query matrix and the key matrix is ​​performed to calculate the spatial and semantic similarity scores of the feature vectors from different modalities. A normalized activation function is used to non-linearly scale the similarity scores, generating an interaction weight matrix ranging from 0 to 1. The interaction weight matrix accurately reflects the interdependencies of the feature nodes of each modality in a specific health intervention scenario. Based on the generated interaction weight matrix, the modality alignment network can capture the common pathological patterns hidden behind different heterogeneous data sources, thereby accurately extracting multi-dimensional contextual information.

[0059] The initial feature vectors are weighted and fused based on contextual information. A linear mapping layer then transforms the weighted fusion result to generate a fixed-dimensional cross-modal feature representation. After extracting multi-level contextual information, it is used as a guiding factor in the original feature spaces. Specifically, the calculated interaction weight matrix is ​​multiplied with the initial feature vectors of the corresponding modalities to strengthen the expression of key health features with high relevance while suppressing irrelevant noise and redundant features. The weighted feature vectors of each modality are concatenated at the feature level to complete the weighted fusion step, forming a fused feature tensor containing comprehensive health attributes. Because the original feature dimensions of each modality differ significantly, the weighted fusion result often has an extremely large and irregular dimensional space, which is detrimental to the rapid convergence and computational optimization of deep learning models. Therefore, a linear mapping layer in a fully connected neural network is needed to reconstruct the dimensions of the weighted fusion result. The linear mapping layer contains a pre-defined weight matrix and bias vector, which projects the high-dimensional weighted fusion result into a unified and compact low-dimensional feature space through linear transformation operations. After transformation by the linear mapping layer, the final output is a vector representation with a uniform size, which successfully generates a fixed-dimensional cross-modal feature representation.

[0060] In one embodiment of the present invention, step S2, in which the cross-modal feature representation is input into the hypergraph topology construction module according to the time series, and the process of constructing a temporal health hypergraph structure containing the time dimension with the time series mapping hyperedges and feature nodes as vertices, is as follows:

[0061] The system receives cross-modal feature representations and corresponding time-series data, and inputs the complete multi-dimensional cross-modal feature representations into a pre-configured hypergraph topology construction module. This module is capable of processing complex spatiotemporally correlated data. Upon receiving the input data, the parsing unit within the module immediately performs precise timestamp segmentation of the continuous cross-modal feature representations based on the accompanying time-series information. During the segmentation process, the algorithm strictly cuts the originally continuous or mixed feature stream into discrete time segments according to the set time granularity. Each time segment corresponds to a specific timestamp. For each timestamp, the extraction engine extracts the multi-modal fusion data unique to that moment, thereby accurately extracting the independent feature nodes corresponding to each timestamp. These feature nodes not only contain health and physiological indicators at specific time points but also retain the deep semantic associations after modality alignment processing. Through this fine-grained segmentation operation based on the time axis, the originally static feature representations are endowed with dynamic life-cycle attributes, enabling the complex health data to be arranged in an orderly manner in the time dimension, thus laying a solid data foundation for constructing complex dynamic topology networks.

[0062] The feature nodes obtained in the preceding steps are mapped one by one to basic vertices in a high-dimensional topological data structure, establishing a large-scale initial vertex set in the virtual data space. This set constitutes the basic skeleton of the entire hypergraph network. After the vertex mapping is completed, the computing unit deeply analyzes the internal structure of each feature node and extracts the health status attribute vectors contained within the feature nodes. These attribute vectors record the quantitative values ​​of various physiological parameters and vital signs. The computing unit accurately assigns these extracted high-value health status attribute vectors to the corresponding topological vertices, transforming each isolated vertex into an entity unit carrying rich health information. To facilitate matrix operations in the graph neural network, the algorithm structurally integrates the attribute vectors of all vertices, combines them with the timestamp information inherent in each vertex, and arranges them according to specific linear algebra rules to finally generate a vertex association feature matrix with clear timestamps. This matrix not only reflects the numerical distribution of various health indicators in the spatial dimension but also locks in the specific occurrence time of these indicators in the temporal dimension, providing a standardized computational foundation for mining temporal evolution patterns.

[0063] Based on the given time series parameters, the time interval span is extracted as the core mapping condition for constructing the hypergraph topology. The algorithm engine scans all vertices in the initial vertex set, focusing on comparing the timestamps of each vertex, and logically aggregating and dividing multiple vertices belonging to the same time interval span. In this process, changes in health indicators within the same specific time period often have a high degree of intrinsic correlation and co-evolution characteristics. Therefore, a data structure capable of expressing high-order relationships is needed to describe them. The construction module instantiates a hyperedge object that transcends traditional graph structures and uses this hyperedge to wrap the aforementioned aggregated vertex groups. A hyperedge can connect any number of vertices simultaneously, thus perfectly showcasing the many-to-many interaction relationships between complex health data. After all hyperedges are accurately established and wrap the corresponding vertices, the module records the membership relationship between vertices and hyperedges in binary or real number form, generating a detailed association occurrence matrix that records the connectivity status of nodes. The elements in this matrix precisely indicate which vertex exists in which hyperedge, thus completing the networked expression of complex relationships within the health data time segment.

[0064] The algorithm extracts the vertex association feature matrix and the association occurrence matrix, which record the network connectivity state in detail, from the memory storage area. These two high-dimensional matrices contain all the basic materials needed to construct a complex temporal health network. In the internal workspace of the hypergraph topology construction module, the computing engine performs precise topological assembly of all vertices with clear timestamps and the multiple hyperedges connecting them based on the data mapping relationship provided by these two matrices. In the assembly stage, the algorithm not only establishes the spatial positional relationships between vertices, but also assigns the network structure a dynamic time axis attribute based on the timestamps, so that the entire topology graph can show the dynamic evolution process of health indicators over time. By performing deep geometric reconstruction of the discrete feature matrix and connection matrix, the originally flat numerical information is successfully transformed into a three-dimensional graph computing architecture with temporal evolution characteristics, and finally a temporal health hypergraph structure containing rich temporal dimension information is constructed. This constructed temporal hypergraph structure can seamlessly connect to various advanced graph neural network models, providing a highly representative logical carrier for immediately carrying out deep node aggregation calculations and state inference.

[0065] In one embodiment of the present invention, step S3, which involves using a spatiotemporal graph network to perform node aggregation calculations on the temporal health hypergraph structure, updating the state of feature nodes, and outputting a health temporal evolution feature matrix expressing the evolutionary relationship of time nodes, is as follows:

[0066] The completed temporal health hypergraph structure is fully input into a pre-configured spatiotemporal graph network, specifically guided into the graph convolutional layer within this network structure for deep spatial feature extraction. Within the computational environment of the graph convolutional layer, the algorithm engine first precisely extracts the hyperedge generation matrix contained in the temporal health hypergraph structure. This matrix records the complex hierarchical connections between all feature nodes and multiple hyperedges in a precise mathematical expression. Based on the topological guidance provided by this crucial hyperedge generation matrix, the graph convolutional layer quickly locates all sets of feature nodes belonging to the same hyperedge. Since feature nodes enclosed by the same hyperedge have a high potential for synergy in physiological mechanisms or complications, the computational unit performs a weighted summation operation on the attribute vectors of these specific feature nodes. During the weighted summation calculation, different coefficient values ​​are assigned based on the node's centrality within the hyperedge or a preset medical prior weight. Through multiply-add operations of multidimensional matrices, scattered local physiological indicators are fused into a high-order structured feature. This computational method effectively breaks the limitation of connecting nodes in pairs in traditional simple graph structures, realizes the comprehensive evaluation of the joint influence of multiple nodes, and finally successfully generates node aggregation calculation results that reflect the interaction of complex health states in the spatial dimension.

[0067] The spatial dimension node aggregation calculation results generated by the above graph convolution operation are extracted and seamlessly guided from the output of the graph convolution layer to the closely connected nonlinear mapping layer for depthwise numerical transformation. Although the spatial dimension node aggregation calculation results integrate multiple node information within the same hyperedge, they may still remain in a simple linear superposition state, making it difficult to express the intricate nonlinear pathological relationships between human health physiological indicators. To enhance the feature representation ability and generalization performance of the network model, activation functions with nonlinear mapping mechanisms are deployed within the nonlinear mapping layer. These functions perform comprehensive numerical transformations on the input aggregated feature matrix. After completing this series of high-dimensional space to high-dimensional space mapping transformations, the nonlinear mapping layer calculates and outputs a set of new feature vectors with higher-level semantic information. The computation unit then accurately maps these new feature vectors generated by the mapping transformation back to their corresponding positions in the hypergraph topology, using them to directly overwrite and replace the original node parameter matrix. Through this parameter iteration mechanism, each feature node absorbs its global structural information in the hyperedge network, thereby completely completing the feature node state update operation.

[0068] All updated feature node states are input into the temporal memory layer of the spatiotemporal graph network in chronological order to calculate the historical state transfer weights between adjacent timestamps and extract the evolution relationships of time nodes; specific content includes:

[0069] After completing the fusion of node features in the spatial dimension, the computation engine accurately extracts the corresponding time stamps inherent in the state of each updated feature node. These time stamps record the absolute moment or relative time offset of each health indicator data point in a precise numerical format. Based on this time sequence, the computation engine initiates a built-in sorting algorithm to rigorously arrange all updated feature node states. During the arrangement process, all feature matrices that have acquired global topology awareness in the spatial graph convolutional network are reorganized into a continuous timeline data stream. This timeline data stream fully demonstrates the dynamic evolution of the patient's bodily functions from the past to the present. The arranged state sequence is then sequentially input into the temporal memory layer within the spatiotemporal graph network, which is dedicated to processing dynamic temporal data. The temporal memory layer possesses powerful sequential data throughput and caching capabilities, enabling it to receive and process these high-dimensional feature matrices strictly according to the chronological order of events, thus providing a solid data input foundation for uncovering the disease development patterns and physiological state evolution trends hidden within the long health monitoring period.

[0070] As the processing flow progresses into the temporal network, the temporal memory layer accurately reads the input state of the current time step and the hidden state preserved by the immediately adjacent timestamps. This hidden state is essentially a compressed feature vector containing historical health evolution information, recording the accumulated physiological characteristic changes across all previous time points. To fully explore the complex nonlinear relationship between the current input and historical accumulation, the computation module concatenates and combines these features along the channel dimension, jointly inputting them into the temporal memory layer's dedicated internal gating unit. This internal gating unit contains multiple activation function-based computational channels, capable of autonomously learning and determining which historical information needs to be forgotten and which current information needs to be prioritized. Through complex matrix multiplication and addition operations and nonlinear mapping, the gating unit accurately calculates the historical state transfer weights between adjacent timestamps. This weight precisely quantifies the direct intervention and potential influence of the previous time step's health physiological state on the current time step's health trajectory, providing core computational parameters for accurately predicting disease deterioration risk or recovery trends.

[0071] After obtaining precise quantitative parameters, the computation engine rigorously performs refined feature preservation and numerical update calculations on the old hidden states based on the weights transferred from the historical states calculated earlier. In the specific calculation steps, the weight matrix is ​​multiplied element-wise with the historical feature matrix to filter out redundant physiological information that no longer significantly impacts the current health assessment, while proportionally incorporating high-value, fresh pathological features extracted at the current time step. This dynamic update mechanism allows the hidden states to continuously absorb and refine the most useful health evolution features as time progresses. After a series of rigorous algebraic operations and state updates, the computation module finally outputs a stable hidden feature matrix containing step-time dependent attributes. This output feature matrix is ​​no longer an isolated time slice but a high-dimensional data set integrating long and short-term health memories. The computation module deeply analyzes the internal structure and distribution of this feature matrix, successfully extracting the evolutionary relationships of time nodes that determine the direction of health status, thus providing highly valuable dynamic evolutionary evidence for the entire intelligent assisted intervention process.

[0072] Based on the evolutionary relationships of time nodes extracted by the temporal memory layer, the processing engine begins the final fusion calculation of multi-time period data in the overall temporal network. Specifically, the processing engine accurately locates the state of updated feature nodes in different time series, using the evolutionary relationships of time nodes as a bridge and guide for cross-step connections, performing cross-time step feature splicing operations on these node features scattered across various time points. This splicing is not a simple physical arrangement and connection, but rather a weighted selection and structural reorganization of features at different time steps based on historical state transmission. This ensures that the spliced ​​comprehensive features retain the core physiological mutation information of each independent time node while incorporating the long-term health evolution trend throughout the entire monitoring period. After all cross-time step feature splicing operations are successfully completed, the computation module globally integrates and collects the splicing results distributed across all time dimensions, comprehensively summarizing all temporal feature vectors. These successfully summarized temporal feature vectors are reconstructed in memory into a tightly structured and dimensionally uniform high-dimensional array, ultimately outputting a stable health temporal evolution feature matrix that expresses rich evolutionary relationships of time nodes.

[0073] In one embodiment of the present invention, step S4, which involves inputting the health temporal evolution feature matrix into a hidden Markov model to map the transition path of the implicit state and generate a dynamic digital health profile matching the implicit state, is as follows:

[0074] The health temporal evolution feature matrix output by the spatiotemporal graph network deep processing is extracted and completely input into a pre-defined fully connected mapping network. This feature matrix contains an extremely complex and high-dimensional set of continuous vectors. While these high-dimensional continuous vectors contain rich patterns of pathological evolution, their excessive dimensionality makes them unsuitable as direct input data for traditional probabilistic graphical models. Therefore, the fully connected mapping network utilizes its dense neural network layers to perform dimensionality reduction projection on the continuous vectors within the matrix. In the dimensionality reduction projection calculation step, high-dimensional features are smoothly mapped to a lower-dimensional and fixed-size feature space. Then, a normalized exponential function is used to probabilistically transform the dimensionality-reduced feature values. This transformation calculates the observation probability distribution sequence for each timestamp node, transforming the originally difficult-to-understand continuous feature variables into probability distribution values ​​representing the likelihood of different health manifestations. Finally, the calculated sequence of all observation probability distributions is stably input into a hidden Markov model in chronological order for decoding processing.

[0075] After receiving a series of continuous observation probability distribution sequences generated by the above calculations, the Hidden Markov Model (HMM) immediately activates its internal probabilistic graphical inference engine for in-depth analysis. The core advantage of this model lies in its ability to handle Markov processes containing invisible internal states. To uncover the true health evolution patterns hidden behind surface observation data, the model combines a state transition probability matrix pre-trained with a large amount of historical medical data to perform maximum likelihood path decoding calculations. During the execution phase of the maximum likelihood path decoding calculation, the algorithm globally searches all possible combinations of internal states and uses dynamic programming to calculate the optimal state sequence with the highest probability of occurrence. After completing the optimal path search, the model stably outputs the implicit state sequences corresponding to each timestamp node. In the medical field, these implicit states often represent a patient being in a specific hidden disease stage or a sub-healthy stage with a specific risk level. This accurately maps the transition paths of implicit states, thus clearly transforming the continuous and complex pathological changes into discrete internal state transition trajectories with a clear evolutionary direction.

[0076] Based on the transition paths of the implicit states determined in the previous step, the optimal sequence of implicit states is extracted. Since the extracted implicit states are merely a set of numerical codes without semantic information, to ensure they have real medical reference value and can be directly read by the computation module, the computing engine retrieves the corresponding health attribute tags for each implicit state from a pre-set database. This pre-set database stores a massive dictionary mapping state codes to specific medical descriptions. Through precise index lookup, each abstract mathematical state is assigned an objective tag with clear medical meaning. After completing the textual mapping of all states, the algorithm engine concatenates all extracted health attribute tags strictly according to their corresponding timestamps. This concatenation process not only preserves the content description of the health condition but also maintains the timeline of the physiological state's occurrence and development. Through this dynamically updated tag sequence combination method, a dynamic digital health profile that perfectly matches the aforementioned implicit states is successfully generated, providing intuitive and detailed digital medical evidence for making personalized intervention decisions.

[0077] In one embodiment of the present invention, step S5, which involves importing the dynamic digital health profile into a reinforcement learning decision framework, performing decision process traversal calculations at preset intervention knowledge nodes, and outputting action sequences as candidate intervention paths, is as follows:

[0078] The generated dynamic digital health profile is extracted and fully imported into a pre-built reinforcement learning decision-making framework. Within this framework, the computational engine first parses the input profile, extracting health attribute labels that record various physiological indicators and disease risks. To enable the machine to understand these labels in natural language or discrete numbered form, the encoding module vectorizes these health attribute labels, encoding them into environmental state features suitable for the reinforcement learning network. These features, presented as high-dimensional continuous vectors, accurately describe the overall health level of the intervened object. After establishing the environmental state features, the algorithm engine accesses the underlying medical knowledge graph to comprehensively read the operational instructions contained within the pre-defined intervention knowledge nodes. These instructions cover various medical measures, including medication adjustments, dietary interventions, and exercise recommendations. These collected operational instructions are discretized and structured, constructing a decision action space within the reinforcement learning decision-making framework's computational memory. This provides a complete and compliant action option for the agent to optimize actions in complex environments.

[0079] After acquiring the current environmental state characteristics and all available operational instructions, the reinforcement learning decision-making framework immediately initiates an intelligent optimization algorithm. Based on the environmental state characteristics, the framework iterates through the decision-making action space, simulating and attempting to execute every candidate operational instruction. To accurately assess the future impact of each operational instruction, the computation module calls an internally pre-trained state transition model using massive amounts of real medical record data. This pre-trained state transition model is essentially a complex deep neural network simulator capable of accurately predicting the expected health state corresponding to each operational instruction. The pre-trained state transition model receives the current environmental state characteristics and the hypothetical operational instructions to be executed, outputting a representation of the potential change in health state that the patient may achieve after intervention. After obtaining this expected health state, the reward function calculation module begins its work. The algorithm sets evaluation criteria biased towards disease improvement and the return of indicators to normal, calculating a single-step reward value based on the expected health state. This value objectively quantifies the immediate contribution of executing a specific operational instruction to improving health status.

[0080] Reinforcement learning algorithms not only focus on short-term gains but also on long-term health management goals. Therefore, the computational engine needs to conduct a long-term evaluation of all possible development paths generated by the aforementioned deductions. The computation module performs a time-series accumulation operation on the single-step reward values ​​along the simulated timeline. A discount factor can be introduced in this accumulation calculation to balance the weighting of short-term effects and long-term benefits. Through this deep accumulation calculation along different decision tree branches, the cumulative expected returns of each traversed branch are finally obtained. These cumulative expected returns represent the overall health benefits that can be obtained after taking a series of consecutive interventions. After completing the benefit evaluation of all potential branches, the decision engine performs a rigorous comparative screening operation, selecting the traversed branch with the largest cumulative expected return as the optimal solution. The algorithm module extracts this optimal branch and concatenates the operation instructions contained within it strictly according to the order of occurrence and execution. This concatenation operation combines the originally independent medical suggestions into a highly coherent and logically sound complete plan, stably outputting the action sequence as candidate intervention paths.

[0081] In one embodiment of the present invention, step S6, in which candidate intervention paths are input into an autoregressive decoder, and the action sequence is decoded and mapped in combination with the constraint rules of the dynamic digital health profile to generate an intelligent intervention strategy, is as follows:

[0082] The process involves extracting physiological attributes from dynamic digital health profiles to construct constraint rules. Specifically, the computational unit deeply analyzes various objective physiological indicators recorded in the health profile, such as the degree of liver and kidney function decline, known drug allergy history, and basal metabolic rate. These extracted physiological attributes are then transformed into machine-readable logical judgment conditions to construct constraint rules for ensuring medical safety. These constraint rules define which intervention actions are absolutely prohibited; for example, for feature vectors of severe renal insufficiency, recommending specific highly toxic drugs is prohibited. After establishing strict constraint rule boundaries, the data processing module inputs the complete action sequence contained within the candidate intervention path into a pre-trained autoregressive decoder. Upon receiving these discretized action instruction sequences output by reinforcement learning, the autoregressive decoder's internal input embedding layer performs high-dimensional word vector mapping on each action node in the sequence. Through a complex multi-layer self-attention mechanism for global feature extraction and contextual semantic fusion, the autoregressive decoder transforms the originally rigid intervention action sequence into a continuous feature representation containing rich medical logical connections, ultimately generating a stable initial semantic vector to guide text generation.

[0083] The autoregressive decoder predicts output words one by one based on the decoded initial semantic vector. During word generation, it invokes constraint rules to perform masking filtering on candidate words, eliminating those that violate the rules. After obtaining the decoded initial semantic vector, which incorporates complete action sequence information, the autoregressive decoder formally initiates the text translation and generation process. The autoregressive decoder strictly follows the global context provided by the decoded initial semantic vector, predicting output words one by one according to preset natural language grammar logic. In each time step calculation, the decoder's internal linear classification layer and normalized exponential function calculate the generation probability distribution of all candidate words for the entire pre-set medical vocabulary. To ensure that the generated medical text meets absolute clinical safety standards, in the critical word generation stage, the computation engine immediately invokes the constraint rules constructed in previous steps. The algorithm module uses these constraint rules to perform strict masking filtering on candidate words with a high generation probability at the current time step. Specifically, the algorithm logically compares the semantic features of candidate words with the taboo conditions set by the constraint rules. If a candidate word, such as the name of a specific drug, is found to conflict with a patient's allergy history, the masking mechanism will force the generation probability of that candidate word to be reduced to 0, thereby completely eliminating the illegal word at the physical computation level. This safety mechanism effectively prevents the language model from generating hallucinatory content with potential medical risks during the generation process.

[0084] The lexical units retained by the masking filtering operation are mapped to text characters with the highest conditional probability. The lexical prediction and character concatenation stages are iteratively executed to generate a text-formatted intelligent intervention strategy. After rigorous masking filtering, dangerous terms in the candidate vocabulary are completely eliminated, and the computation engine then selects the optimal term from the remaining safe vocabulary. The algorithm module sorts the terms according to their normalized probability values, accurately mapping the lexical units safely retained by the masking filtering operation to text characters with the highest conditional probability. This selected text character not only meets the semantic coherence requirements of the current context but also fully satisfies the personalized safety constraints of medical intervention. After generating a single character at the current time step, the output character is fed back into the end of the autoregressive decoder sequence as a known condition for the prediction of the next time step. The computation module iteratively executes the lexical prediction and character concatenation stages in this manner, continuously combining newly generated medical terms, medication instructions, and healthy living suggestions. The entire iterative generation process continues until the decoder predicts and outputs a pre-set end marker. After the termination marker triggers the stop mechanism, all the concatenated discrete characters are integrated and formatted, ultimately generating a stable and structurally complete text format that conforms to human reading habits—an intelligent intervention strategy.

[0085] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for generating intelligent intervention strategies for full-cycle health management, characterized in that, Includes the following steps: S1: Obtain heterogeneous raw health data, input it into a modality alignment network, extract contextual information through a cross-attention mechanism, and generate a fixed-dimensional cross-modal feature representation; S2: Input the cross-modal feature representation into the hypergraph topology construction module according to the time series, and construct a temporal health hypergraph structure containing the time dimension with the time series mapping hyperedge and feature nodes as vertices; S3: Utilize a spatiotemporal graph network to perform node aggregation calculations on the temporal health hypergraph structure, update the state of feature nodes, and output a health temporal evolution feature matrix that expresses the evolutionary relationship of time nodes; S4: Input the health temporal evolution feature matrix into the hidden Markov model, map the transition path of the implicit state, and generate a dynamic digital health profile that matches the above implicit state. S5: Import the dynamic digital health profile into the reinforcement learning decision framework, perform decision process traversal calculations in the preset intervention knowledge nodes, and output action sequences as candidate intervention paths. S6: Input the candidate intervention path into the autoregressive decoder, combine it with the constraint rules of the dynamic digital health profile to decode and map the action sequence, and generate an intelligent intervention strategy.

2. The method for generating intelligent intervention strategies for full-cycle health management according to claim 1, characterized in that, In step S1, the process of acquiring heterogeneous raw health data, inputting it into a modality alignment network, extracting contextual information through a cross-attention mechanism, and generating a fixed-dimensional cross-modal feature representation is as follows: Obtain heterogeneous health raw data, input the heterogeneous health raw data into a multi-branch feature encoder for independent mapping operation, and output the initial feature vectors corresponding to each modality; The initial feature vectors are input into the modality alignment network, and the interaction weights between the initial feature vectors are calculated using the cross attention mechanism to extract contextual association information. The initial feature vectors are weighted and fused based on contextual information, and the weighted fusion result is transformed through a linear mapping layer to generate a fixed-dimensional cross-modal feature representation.

3. The method for generating intelligent intervention strategies for full-cycle health management according to claim 1, characterized in that, In step S2, the process of inputting cross-modal feature representations into the hypergraph topology construction module according to time series, and constructing a temporal health hypergraph structure containing the time dimension using time series mapping hyperedges and feature nodes as vertices, is as follows: Receive cross-modal feature representations and corresponding time series, input the cross-modal feature representations into the hypergraph topology construction module, divide the cross-modal feature representations into timestamps according to the time series, and extract feature nodes corresponding to each timestamp; Feature nodes are mapped to vertices of the topological data structure, an initial vertex set is established, the health status attribute vector contained in the feature nodes is extracted and assigned to the corresponding vertices, and a vertex association feature matrix with timestamp is generated. Using the time interval span in the time series as the mapping condition, vertices belonging to the same time interval span are aggregated and divided, a hyperedge is constructed to enclose the vertices, and an association generation matrix connecting the corresponding hyperedge and the vertex is generated. Extract the vertex association feature matrix and association occurrence matrix, and assemble the topology structure of vertices and hyperedges with timestamps inside the hypergraph topology construction module to construct a temporally healthy hypergraph structure containing the time dimension.

4. The method for generating intelligent intervention strategies for full-cycle health management according to claim 1, characterized in that, In step S3, the process of using a spatiotemporal graph network to perform node aggregation calculations on the temporal health hypergraph structure, updating the state of feature nodes, and outputting a health temporal evolution feature matrix expressing the evolutionary relationship of time nodes is as follows: The temporal health hypergraph structure is input into the graph convolutional layer of the spatiotemporal graph network. Based on the hyperedge generation matrix, the feature nodes inside the same hyperedge are subjected to a feature weighted summation operation to generate the spatial dimension node aggregation calculation result. Extract the node aggregation calculation results in the spatial dimension and input them into the nonlinear mapping layer to perform numerical transformation. Use the new feature vector generated by the mapping transformation to replace the original node parameters and update the feature node state. All updated feature node states are input into the temporal memory layer of the spatiotemporal graph network in chronological order to calculate the historical state transfer weights between adjacent timestamps and extract the evolution relationship of time nodes. Based on the evolution relationship of time nodes, cross-time step feature splicing operation is performed on the updated feature node states in different time series, and all time series feature vectors are summarized to output the healthy temporal evolution feature matrix.

5. The method for generating intelligent intervention strategies for full-cycle health management according to claim 4, characterized in that, The specific steps of inputting all updated feature node states into the temporal memory layer of the spatiotemporal graph network in chronological order, calculating the historical state transfer weights between adjacent timestamps, and extracting the evolutionary relationships of time nodes include: Extract the corresponding time stamps, arrange the states of all updated feature nodes according to the time sequence, and input the arranged state sequence into the temporal memory layer of the spatiotemporal graph network. The temporal memory layer reads the current time step input state and the hidden state retained by adjacent timestamps, combines the features of the two and inputs them into the internal door control unit to calculate the historical state transfer weight between adjacent timestamps. Based on the weights passed from historical states, the hidden states are used to preserve features and update values, outputting a hidden feature matrix that includes step-time dependency attributes, and extracting the evolution relationship at time nodes.

6. The method for generating intelligent intervention strategies for full-cycle health management according to claim 1, characterized in that, In step S4, the process of inputting the health temporal evolution feature matrix into the hidden Markov model, mapping the transition path of the implicit state, and generating a dynamic digital health profile matching the aforementioned implicit state is as follows: Extract the health temporal evolution feature matrix and input it into a fully connected mapping network. Perform dimensionality reduction projection on the continuous vectors inside the matrix, calculate the observation probability distribution sequence for each timestamp node, and input the observation probability distribution sequence into a hidden Markov model. The Hidden Markov Model receives the observation probability distribution sequence, combines it with the state transition probability matrix to perform maximum likelihood path decoding calculation, and outputs the implicit state sequence corresponding to each time stamp node, thereby mapping the transition path of the implicit state. Based on the transition path of the implicit state, the implicit state sequence is extracted. The health attribute tags corresponding to each implicit state are retrieved from the preset database. All health attribute tags are concatenated according to the timestamp order to generate a dynamic digital health profile that matches the above implicit state.

7. The method for generating intelligent intervention strategies for full-cycle health management according to claim 1, characterized in that, In step S5, the process of importing the dynamic digital health profile into the reinforcement learning decision framework, performing decision process traversal calculations at preset intervention knowledge nodes, and outputting action sequences as candidate intervention paths is as follows: The dynamic digital health profile is imported into the reinforcement learning decision-making framework, and its health attribute labels are encoded to generate environmental state features. The operation instructions contained in the preset intervention knowledge nodes are read to construct the decision action space. The reinforcement learning decision-making framework performs traversal calculations of the decision-making process in the decision-making action space based on the characteristics of the environmental state. It deduces the expected health state corresponding to each operation instruction through a pre-trained state transition model and calculates the single-step reward value based on the expected health state. The cumulative expected return of each traversal branch is obtained by summing the single-step return values ​​over time. The traversal branch with the largest cumulative expected return is selected, and the operation instructions contained therein are concatenated in order. The output action sequence is used as a candidate intervention path.

8. The method for generating intelligent intervention strategies for full-cycle health management according to claim 1, characterized in that, In step S6, the process of inputting candidate intervention paths into an autoregressive decoder, combining the constraint rules of the dynamic digital health profile to decode and map the action sequence, and generating an intelligent intervention strategy is as follows: Extract the physiological attributes contained in the dynamic digital health profile to construct constraint rules, input the action sequence inside the candidate intervention path into the autoregressive decoder to generate the initial semantic vector for decoding; The autoregressive decoder predicts and outputs word units one by one based on the initial semantic vector. During the word unit generation stage, it calls the constraint rules to perform a masking filtering operation on the candidate words and removes illegal word units. The tokens retained by the masking filtering operation are mapped to text characters with the highest conditional probability. The token prediction and character concatenation steps are executed cyclically to generate an intelligent intervention strategy for text format.

Images