AI-based full life cycle health intervention method

Abstract

This invention relates to an AI-based full-lifecycle health intervention method, comprising: constructing a metaverse virtual space for health management; creating and maintaining a dynamic virtual health record for each patient, serving as their digital twin and core data hub; providing rehabilitation activities to patients in the space through virtual avatars configured for users, supporting real-time multi-party interaction, with multi-dimensional health data generated during the process being collected in real time and updated to the record; using a multimodal time-series prediction model to perform real-time analysis and trend prediction of the health data, generating risk prediction information; and triggering and executing personalized health intervention measures based on this information. This invention achieves the integration and continuous management of patients' full-dimensional health data, completing the transformation from a passive response to a proactive early warning disease management model, and providing Alzheimer's patients with an immersive, proactive, and full-cycle comprehensive health management solution by constructing an online and offline linked closed-loop intervention system.

Description

Technical Field

[0001] This invention relates to the field of medical and health information technology and artificial intelligence, and in particular to an AI-based method for full-life-cycle health intervention. Background Technology

[0002] Alzheimer's disease (AD) is a complex neurodegenerative disease, and its rehabilitation management involves multidimensional and dynamically changing physical and mental states and diverse interventions. Patient diagnostic information, rehabilitation records, daily behavior, and socio-psychological data are usually scattered across different medical institutions, family settings, and various independent applications, lacking a unified, dynamically updated data hub for panoramic aggregation and integration, making it difficult to form a continuous and complete assessment of the patient's condition.

[0003] Traditional health management relies heavily on periodic outpatient assessments and subjective reports from patients and their families, making it difficult to provide early warnings of risks. Interventions often lag behind actual changes in the condition, failing to prevent problems before they arise and resulting in low management efficiency.

[0004] Existing cognitive and motor rehabilitation training often relies on offline guidance or simple tablet applications, which are monotonous and result in poor patient compliance. At the same time, training programs are difficult to dynamically and accurately adjust in difficulty based on the patient's real-time performance, limiting their personalization.

[0005] In recent years, emerging technologies such as metaverse, digital twins, and artificial intelligence have offered new possibilities for health management. However, how to deeply integrate these technologies to build a systematic solution specifically designed for Alzheimer's disease (AD) patients, capable of connecting the entire chain of data collection, intelligent analysis, proactive intervention, and collaborative management, remains an unsolved technical challenge. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing an AI-based full-lifecycle health intervention method to solve the problems existing in the prior art.

[0007] To achieve the above objectives, this invention provides an AI-based method for full-lifecycle health intervention, the method comprising: Construct a metaverse virtual space for Alzheimer's disease health management; In the metaverse virtual space, a dynamically updated virtual health record is created for each patient, serving as the core data hub for the patient, and together with the patient's virtual avatar, constitutes the patient's digital twin; In the aforementioned metaverse virtual space, virtual avatars are created and configured for patients, their families, and medical staff. Through these virtual avatars, rehabilitation activities are provided to patients, and real-time interaction among multiple parties is supported. These rehabilitation activities include at least virtual cognitive training courses and virtual exercise rehabilitation games. The behavioral and social data generated during these rehabilitation activities and real-time interactions constitute multi-dimensional health data, which is collected in real time. This multi-dimensional health data is collected in real time and updated to the corresponding virtual health profile. Based on a preset multimodal time-series prediction model, the multidimensional health data of a preset duration is analyzed and trend predicted in real time to generate predictive information for potential complications or functional decline risks. Based on the predicted information, health intervention measures are triggered and implemented.

[0008] In one possible implementation, the construction of the metaverse virtual space for Alzheimer's disease health management specifically includes: Based on the full life cycle needs of Alzheimer's patients from diagnosis and rehabilitation to daily living support, multiple functionally independent but logically related virtual scene modules are set up; the virtual scene modules include at least: a virtual rehabilitation center for structured cognitive training, a virtual community square for social interaction and group support, a virtual home environment for simulating daily living skills training, and a virtual support center for family education and medical care collaboration. Based on scene planning, 3D modeling and game engine technology are used to construct 3D visual models of each virtual scene module; within the virtual rehabilitation center, interactive objects and triggers are preset and bound to cognitive training tasks; within the virtual community square, interaction rules supporting voice dialogue, facial expression communication, and preset group activities between virtual avatars are set; within the virtual home environment, interactive points that can simulate daily operations are set and linked to life skills assessment logic; within the virtual support center, document sharing, video conferencing, and planning dashboard tools are integrated to provide an interactive interface for collaborative management; A unified user account system is used to create unique identifiers for patients, their families, and medical staff. Based on this user account system, each user is provided with the ability to create and customize a virtual avatar, which can exist continuously and move freely in all virtual scene modules. Establish standardized data interfaces between the metaverse virtual space and external systems to ensure access to data streams from medical information systems and wearable devices; at the same time, integrate the various virtual scene modules, user systems and data interfaces through a unified central server or distributed network architecture to form a persistent, multi-user concurrent access coherent virtual world, namely the metaverse virtual space.

[0009] In one possible implementation, the creation of a dynamically updated virtual health record for each patient in the metaverse virtual space specifically includes: A multi-layered digital profile is defined for each patient, comprising: an identity and foundation layer, a temporal data layer, a document and report layer, and a 3D twin vital signs layer. The identity and foundation layer stores the patient's unique identifier, demographic information, and basic diagnostic information. The temporal data layer uses a database table structure with preset fields to store temporal data of accessed physiological signals, behavioral events, and environmental interactions. The document and report layer stores non-temporally structured documents and medical image indexes. The 3D twin vital signs layer is used to associate with and drive a 3D visualization model representing the patient in the metaverse. Deploy standardized data access interfaces connected to the digital archives, including: a medical information system interface, an IoT device interface, and a metaverse behavior log interface; the medical information system interface is used to pull electronic medical records, test results, and image reports from external medical systems to the document and report layer; the IoT device interface is used to receive physiological and environmental data uploaded by wearable devices and environmental sensors to the time-series data layer; the metaverse behavior log interface is used to capture patients' rehabilitation activities and social interaction data in the metaverse, and classify and store them in the time-series data layer and the document and report layer. Based on the identity and information of the base layer, a basic virtual avatar is generated or matched as the digital twin of the patient; a mapping relationship is established between the key real-time data in the time series data layer and the state parameters of the virtual avatar to realize the driving force of data to visualized state; The raw data imported through the interface is cleaned, de-identified, and time-aligned; and features are extracted from the processed data based on preset rules or models to update the data association graph within the virtual health record. The digital archive is configured with update conditions including timed triggering, data arrival triggering, and AI prediction-based triggering; when the conditions are met, data integration, feature updates, or digital twin state synchronization are automatically performed. Configure a role-based access control list for the digital archives and enable audit logging with operator, timestamp, and context for all data operations.

[0010] In one possible implementation, providing rehabilitation activities to patients and supporting real-time interaction through a metaverse virtual space specifically includes: The virtual cognitive training course adopts adaptive difficulty, which dynamically adjusts the task complexity and cognitive load type of the next training cycle based on the error rate and reaction time recorded by the patient in the virtual health record recently. The virtual motion rehabilitation game integrates an inertial measurement unit and a computer vision motion capture system to quantify and assess the patient's motion parameters in real time; the game provides real-time audiovisual biofeedback when the patient performs specific rehabilitation movements such as weight transfer and gait cycle. When patients participate in the real-time interaction, the spatial distance between their virtual avatars, the duration of the interaction, the emotional tendency of their verbal communication, and their non-verbal actions are all captured by the system and stored in their virtual health records as structured social interaction data.

[0011] In one possible implementation, the multimodal temporal prediction model is a multi-branch fusion network with a Transformer encoder at its core; including: A multimodal feature extraction module is used to extract deep features of different types of data in parallel from the virtual health record, including: a temporal feature extraction branch, a text feature extraction branch, and a visual feature extraction branch. The temporal feature extraction branch is composed of a long short-term memory network, whose input is a sequence of physiological signals and behavioral events extracted from the virtual health record and aligned by timestamps, and whose output is a temporal feature vector representing individual physiological and behavioral patterns. The text feature extraction branch is composed of a natural language processing model, whose input is medical text reports and cognitive assessment records in the virtual health record, and whose output is a text feature vector representing semantic information and clinical conditions. The visual feature extraction branch is composed of a three-dimensional convolutional neural network, whose input is patient interaction video data captured in the metaverse virtual space, and whose output is a visual feature vector representing social interaction and non-verbal behavior patterns. The feature fusion and encoding module is used to integrate multimodal features and generate a context-aware fusion representation, including: a cross-modal attention fusion submodule and a shared Transformer encoder; the cross-modal attention fusion submodule takes the temporal feature vector, text feature vector and visual feature vector as input, and its function is to align and weight the three feature vectors, and its output is a unified fusion feature representation; the shared Transformer encoder takes the fusion feature representation as input, and its function is to perform global context modeling and deep encoding on the fused features, and its output is an encoded feature vector containing long-term dependencies; The prediction output module is used to generate the final risk prediction and analysis results based on the encoded features, including a prediction head; the input of the prediction head is the encoded feature vector, and the output is the prediction result; the prediction result specifically includes: the risk probability value of a specific adverse event occurring at one or more preset time points in the future; the input feature attribution identifier that contributes the most to the risk probability value; wherein, the specific adverse event includes a fall, an episode of agitated behavior, or a significant decline in cognitive function score.

[0012] In one possible implementation, the training method for the multimodal time-series prediction model includes: Training sample sets are extracted from the virtual health records of historical patients. Each training sample includes input data and label data. The input data consists of multimodal data within a preset historical time window that has undergone time alignment and normalization, specifically including physiological and behavioral time-series data, medical text reports, and corresponding interactive video clips. The label data consists of supervised learning labels constructed based on health events actually recorded within a preset future time period after the historical time window. The labels are multi-task labels, including at least a binary classification label indicating whether a specific adverse event occurred and a regression label indicating the time of event occurrence. A phased pre-training strategy is adopted to independently initialize each branch network constituting the multimodal temporal prediction model. The independent initialization of each branch network includes: pre-training the natural language processing model in the text feature extraction branch on a large general text dataset; pre-training the three-dimensional convolutional neural network in the visual feature extraction branch on a large general visual dataset; and pre-training the long short-term memory network in the temporal feature extraction branch using unlabeled medical time-series data through a self-supervised learning task. The pre-trained text feature extraction branch, visual feature extraction branch, and temporal feature extraction branch are assembled with the untrained cross-modal attention fusion module, Transformer encoder, and prediction head to form an initial prediction model. On the training sample set, with the goal of minimizing the multi-task loss function, the initial prediction model is jointly fine-tuned end-to-end to obtain the multimodal temporal prediction model; the multi-task loss function is composed of a weighted sum of cross-entropy classification loss, root mean square error regression loss, and contrastive loss used to promote multimodal feature alignment; After the multimodal time series prediction model is deployed, a federated learning framework is used for continuous optimization. The continuous optimization includes: each terminal using its private data to fine-tune its local model copy to generate model update parameters; the model update parameters of each terminal are encrypted and aggregated on a central server through a secure aggregation protocol to generate a global model update; and the global model update is distributed to each terminal to update its local model copy.

[0013] In one possible implementation, the triggering and execution of health intervention measures based on predictive information specifically includes: Receive prediction information generated by the multimodal time series prediction model, wherein the prediction information includes at least the risk event type, risk probability value, and feature attribution identifier; The risk probability value is compared with multiple preset risk thresholds to classify the risk level into three levels: low, medium, and high. Based on the risk level, risk event type, and characteristic attribution identifier, one or more initial intervention instructions are matched and generated from a predefined intervention strategy knowledge base; the intervention strategy knowledge base stores the mapping relationship between different risk scenarios and corresponding intervention measure types, execution objects, and execution platforms; Based on the patient's personal preferences in the virtual health record, the current metaverse virtual space scene, the compliance with historical intervention measures, and the feedback effect, the parameters of the initial intervention instruction are individually fine-tuned to form an executable personalized intervention plan; the personalized fine-tuning includes: adjusting the way the intervention information is expressed, selecting the virtual scene preferred by the patient as the intervention trigger point, or adapting to the real-world resources available to the patient; Based on the execution platform specified in the personalized intervention plan, the intervention instructions are distributed to the corresponding terminal or system and triggered for execution: If the execution platform is a metaverse virtual space, the system will present visual or auditory guidance information, adjust the difficulty of rehabilitation tasks, or dispatch virtual assistants to provide companionship and assistance in the environment where the patient's virtual avatar is located. If the execution platform is an associated mobile application or web platform, the system will push structured warning notifications, specific nursing suggestions, or rehabilitation plan adjustment to-do items to the client of the patient's family or medical staff. If the platform is an offline medical service system, the system generates and sends service work orders to community nursing centers, pharmacies or attending physician workstations through standardized interfaces to initiate offline follow-up, drug delivery or appointment adjustment processes. After the intervention measures are implemented, the system monitors and collects the patient's response data and status change data through the aforementioned metaverse virtual space, IoT devices, and manual reporting channels; The response data, state change data, and prediction information and intervention plan of this intervention are associated and stored in the patient's virtual health record as a complete intervention-feedback record. This is used to evaluate the effectiveness of the intervention measures and optimize the intervention strategy knowledge base and the multimodal time series prediction model. When the probability of a specific risk indicated by the predicted information exceeds a preset threshold, the system automatically executes a tiered intervention strategy. The tiered intervention strategy includes Level 1 intervention, Level 2 intervention, and Level 3 intervention. Level 1 intervention is guidance within the metaverse, including: popping up contextual reminders in the patient's virtual avatar, adjusting the complexity of the virtual environment, or assigning a guiding virtual assistant for companionship and task guidance. Level 2 intervention is cross-platform notification, including: sending a structured early warning report and nursing suggestion list to the bound family member's mobile terminal, and generating pending follow-up tasks or medication re-examination reminders to the responsible medical staff's collaborative management platform. Level 3 intervention is initiating offline linkage, including: when the risk level is high, automatically generating referral suggestions or emergency contact information and pushing it to the nearest community health service center.

[0014] In one possible implementation, the method further includes: For the raw data imported into the virtual health record, noise is added at the data source end or immediately after access, using differential privacy technology, or feature extraction is performed locally using federated learning. Only the encrypted feature vectors or model gradients are uploaded to the central server for aggregation analysis, making it impossible to identify specific patient individuals from the circulating data.

[0015] In one possible implementation, the method further includes: An edge computing architecture is adopted, and the core interactive logic and real-time data analysis module of the virtual cognitive training course and sports rehabilitation game are deployed on edge nodes close to the user. For latency-sensitive operations such as virtual avatar movement and environmental rendering, edge nodes process and respond in real time. Only data such as behavior logs and aggregated features that need to be stored and analyzed for a long time are asynchronously uploaded to the virtual health record in the central cloud.

[0016] In one possible implementation, the method further includes: The collaborative management platform integrates an insurance payment interface and a service subscription module. Once the health intervention measures are triggered and implemented, the system automatically generates structured service credentials and effect reports; According to the insurance terms pre-authorized by the patient or their family, the service voucher that meets the claim conditions is submitted to the insurance company's system through the insurance payment interface to initiate an automatic claim process; or, according to the subscription package, the corresponding service fee is deducted from the prepaid account.

[0017] By applying the AI-based full-lifecycle health intervention method provided in this invention, and through the construction of a metaverse space and virtual health records, it systematically integrates heterogeneous data from multiple sources, including disease diagnosis and treatment, rehabilitation training, daily life, and metaverse-specific behaviors and social interactions, for the first time. This forms a dynamically updated digital twin of the patient, addressing the pain points of fragmented and scattered health data and providing a complete and continuous data foundation for precise analysis. Furthermore, by utilizing a multimodal temporal prediction model to perform real-time analysis and trend prediction on the fused data, it can generate predictive information for risks such as falls and agitated behaviors in advance, achieving a leap from post-event treatment to pre-event warning and in-event intervention in proactive management. Furthermore, based on the predictive information, personalized intervention measures can be triggered simultaneously in the metaverse environment, mobile platforms, and offline medical services, forming a three-dimensional intervention network of virtual guidance, family reminders, and physical services, ensuring seamless connection from risk insight to physical action. Furthermore, immersive and gamified metaverse rehabilitation activities significantly increase patient participation; the system dynamically adjusts task difficulty based on real-time performance data and provides immediate biofeedback, keeping training within a personalized optimal challenge range, thereby optimizing rehabilitation outcomes. Furthermore, this application, based on a collaborative management platform within the same metaverse space, breaks through the limitations of time and space, enabling real-time information sharing and collaborative development of rehabilitation plans among family members and medical staff, thereby improving the collaboration efficiency and decision-making consistency of the care team. Attached Figure Description

[0018] Figure 1 A flowchart of the AI-based full-life-cycle health intervention method provided by this invention; Figure 2 for Figure 1 Flowchart for step 110; Figure 3 for Figure 1 Flowchart for step 120; Figure 4 for Figure 1 Flowchart for step 130; Figure 5 This is a structural diagram of a multimodal time series prediction model; Figure 6 for Figure 1 The flowchart for step 150. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Figure 1 A flowchart of an AI-based full-lifecycle health intervention method, such as... Figure 1 As shown, the method includes the following steps: Step 110: Construct a metaverse virtual space for Alzheimer's disease health management.

[0021] according to Figure 2 As shown, step 110 includes the following steps: Step 1101: Based on the full life cycle needs of Alzheimer's patients from diagnosis and rehabilitation to daily living support, multiple functionally independent but logically related virtual scene modules are set up; the virtual scene modules include at least: a virtual rehabilitation center for structured cognitive training, a virtual community square for social interaction and group support, a virtual home environment for simulating daily living skills training, and a virtual support center for family education and medical care collaboration.

[0022] First, based on the full life-cycle care needs of Alzheimer's disease patients, from mild cognitive impairment to severe dementia, a systematic scenario planning is conducted. This embodiment designs and sets up four core, functionally independent yet logically interconnected virtual scenarios, forming a complete virtual care ecosystem: The Virtual Rehabilitation Center is designed to address the core symptoms of cognitive decline in Alzheimer's disease (AD) patients. It is a structured training environment designed to provide targeted training for specific cognitive domains such as memory, attention, and executive function through repetitive, progressive tasks. Its design avoids complex and confusing decorations, employing soothing colors and clear visual guidance.

[0023] The virtual community square is designed to address the social withdrawal and loneliness often experienced by Alzheimer's disease (AD) patients. This module simulates a safe and friendly public space to promote non-stressful social interactions between patients and between patients and their families / healthcare staff, aiming to maintain patients' social skills and emotional connections.

[0024] The virtual home environment module is designed to assess and train patients' activities of daily living (ADL) skills. It highly replicates a typical home environment, such as the living room, kitchen, and bedroom, simulating daily activities like cooking, dressing, and tidying up. This safe and risk-free virtual setting assesses patients' instrumental ADL skills and provides opportunities for repeated training.

[0025] The Virtual Support Center module serves collaborative care teams. It provides an online collaborative workspace for family members, community doctors, specialists, rehabilitation therapists, and others located in different places, facilitating information sharing, solution discussion, and decision-making.

[0026] Step 1102: Based on the virtual scene, construct the 3D visual model of each virtual scene module using 3D modeling and game engine technology; within the virtual rehabilitation center, preset interactive objects and triggers bound to cognitive training tasks; within the virtual community square, set interaction rules supporting voice dialogue, facial expression communication, and preset group activities between virtual avatars; within the virtual home environment, set interaction points that can simulate daily operations and link them to life skills assessment logic; within the virtual support center, integrate document sharing, video conferencing, and planning dashboard tools to provide an interactive interface for collaborative management.

[0027] Specifically, based on the virtual scene planning in step 1101, high-precision, low-polygon 3D visual models of each scene module are created using professional 3D modeling software such as Blender and 3ds Max to ensure smooth rendering on mainstream hardware. Subsequently, game engines such as Unity or Unreal Engine are used for integration and interactive development.

[0028] Within the virtual rehabilitation center, scripts are developed to link specific cognitive training tasks, such as matching memory cards and creating virtual supermarket shopping lists, with interactive objects in the scene, such as cards and products. Triggers are set to detect user actions and calculate key performance indicators such as task completion and reaction time based on these actions.

[0029] Within the virtual community square, third-party voice communication SDKs, such as Agora and facial capture plugins, are integrated to enable real-time voice dialogue and synchronized basic facial expressions, such as smiling and nodding, between virtual avatars. Simultaneously, pre-set scripts for activities such as virtual concerts and group gardening guide users to participate in structured group interactions.

[0030] In a virtual home environment, "interaction points" are set up on objects such as kitchen kettles and wardrobes. When the patient's virtual avatar approaches and triggers these points, the system initiates a simulated task such as "boiling water" or "choosing an outer garment." Every step in the task process, such as whether the stove was turned off or whether appropriate seasonal clothing was selected, is recorded and linked to a pre-set set of daily living ability assessment logic to generate an assessment report.

[0031] Within the virtual support center, lightweight collaborative tools can be directly integrated or developed, such as a shared document viewer, a WebRTC-based video conferencing window, and a multi-user editable rehabilitation plan dashboard.

[0032] The activities of daily living (ADL) assessment algorithm is deployed within a virtual home environment. Its purpose is to automatically and objectively transform patients' natural interactions during simulated daily tasks, such as meal preparation, laundry, and medication administration, into structured clinical assessment data, enabling continuous monitoring and quantitative analysis of their instrumental activities of daily living.

[0033] The life skills assessment logic processes interactive stream data from the virtual environment in real time. It analyzes the completion rate and sequential correctness of task steps using a rule engine, and employs temporal pattern analysis to identify abnormal patterns in behavior, such as hesitation, indecisiveness, and repetitive errors. This enables refined and repeatable measurement of what was done and how it was done.

[0034] Quantified behavioral characteristics are mapped to clinically significant cognitive and functional dimensions, such as procedural memory, executive function, safety awareness, and tool usage ability. A configurable weighted scoring model or lightweight machine learning model generates a comprehensive ability score and risk item list. The results can be correlated with standardized clinical assessment scales, such as the IADL, providing healthcare professionals with directly referable digital reports.

[0035] The living ability assessment logic possesses adaptive and evolutionary capabilities. On the one hand, the difficulty of its assessment tasks can be dynamically adjusted according to the disease stage in the patient's file. On the other hand, all assessment results, as key multi-dimensional health data, flow back to the patient's virtual health file, not only to plot long-term ability change curves but also to provide crucial training and validation data for upstream AI risk prediction models, thereby driving the entire intervention system to form a continuous optimization closed loop of assessment-prediction-intervention-reassessment.

[0036] In summary, this daily living ability assessment logic, by creating safe and controllable virtual task scenarios and combining computational models to deeply interpret behavior, successfully transforms the traditionally subjective and difficult-to-perform assessment of daily living abilities into an objective, continuous, and integrable core technology component in the entire digital intervention ecosystem. This greatly enhances the ability to accurately manage and intervene in the functional status of Alzheimer's patients at an early stage.

[0037] Step 1103: Create unique identity identifiers for patients, their families, and medical staff through a unified user account system; based on the user account system, provide each user with virtual avatar creation and customization functions, and the virtual avatar can exist continuously and move freely in all virtual scene modules.

[0038] Specifically, to achieve a unique user identity and immersive experience within the metaverse, a unified user account system is constructed. This system employs a role-based access control model, creating accounts and assigning unique identifiers (UIDs) to different roles such as patients, family members, and medical staff. Based on this system, a virtual avatar editor is provided. In this editor, users, especially patients and their families, can choose a basic appearance from a pre-set, user-friendly avatar model library and perform limited personalization, such as hairstyle and clothing color. The user account system maintains an avatar state mapping table for each user account. Regardless of which virtual scene a user logs into, such as a rehabilitation center or community square, their virtual avatar's appearance, location, and basic state (as seen in the last logged-out scene) remain consistent and persistent, enabling seamless cross-scene roaming and enhancing the immersive experience and user sense of belonging.

[0039] The avatar model library is a collection of digital assets stored on the system server, containing multiple 3D character models that have been modeled, skeletally bound, and configured with basic animations. Based on the patient's identity and base layer data in their virtual health record, such as age and gender, the system can automatically recommend or directly match a base image from the library that best reflects the common characteristics of the patient's age group using a simple rule engine. The selected base image becomes the initial 3D mesh and texture data for the patient's digital twin in the metaverse. The core function of the avatar state mapping table is to calculate and map multi-dimensional health data imported into the record, such as physiological indicators, behavioral events, and AI risk predictions, into a set of visualized state parameters for the virtual avatar in real time using predefined rules or models. These parameters directly control the avatar's appearance, behavior, and interactive feedback in the metaverse, transforming invisible health data into a visible and perceptible virtual image state. The avatar state mapping table is organized and stored using a key-value pair structure. The key represents adjustable avatar state parameters, such as fatigue level, mood index, engagement level, and current risk indicator. The value corresponds to the current calculated value of the parameter, such as percentage, level, or label. This value is dynamically generated by the mapping rules based on the latest health data. How to obtain the avatar state parameters can be achieved using existing technologies, which will not be elaborated here. The digital twin achieves data-driven processing through a state mapping table.

[0040] Step 1104: Establish a standardized data interface between the metaverse virtual space and external systems to ensure access to data streams from medical information systems and wearable devices; at the same time, integrate the various virtual scene modules, user systems and data interfaces through a unified central server or distributed network architecture to form a persistent, multi-user concurrent access coherent virtual world, namely the metaverse virtual space.

[0041] Specifically, to make the metaverse virtual space a true data hub and real-time service platform, it is essential to achieve data connectivity with external real-world systems. This requires developing standardized data interfaces, adopting medical data standards such as HL7 FHIR (Fast Healthcare Interoperability Resources), and developing interfaces with Hospital Information Systems (HIS) and Laboratory Information Systems (LIS) for periodically retrieving patients' electronic medical records, test reports, and other data. Simultaneously, an IoT interface based on MQTT or WebSocket protocols should be designed to receive real-time physiological and activity data streams from devices such as smart bracelets and wearable fall detectors.

[0042] All the virtual scenario modules, user account system, and internal and external data interfaces constructed above are deployed in a unified software architecture. Depending on the expected user scale, either a central server architecture or a distributed microservice architecture can be chosen. In a central server architecture, all logic and data are centralized on a cloud server, facilitating maintenance and updates, and is suitable for initial deployment. In a distributed microservice architecture, different scenario modules, user services, data processing, and other services are broken down into independent microservices, which collaborate through an API gateway. This architecture is more flexible and better supports concurrent access from multiple users and high availability.

[0043] Ultimately, through the above integration, a persistent (data and state are continuously preserved), scalable, and coherent virtual world that can be synchronized with real-world medical data streams in real time is formed. This is the metaverse virtual space dedicated to AD health management, providing a core carrier for all subsequent health data collection, analysis, and intervention services.

[0044] Step 120: In the metaverse virtual space, create a dynamically updated virtual health profile for each patient, which serves as the core data hub for the patient and together with the patient's virtual avatar, constitutes the patient's digital twin. like Figure 3 As shown, step 120 includes the following: Step 1201: Define a multi-layered digital profile for each patient, comprising: an identity and foundation layer, a temporal data layer, a document and report layer, and a 3D twin vital signs layer. The identity and foundation layer stores the patient's unique identifier, demographic information, and basic diagnostic information. The temporal data layer uses a database table structure with preset fields to store temporal data of accessed physiological signals, behavioral events, and environmental interactions. The document and report layer stores non-temporally structured documents and medical image indexes. The 3D twin vital signs layer is used to associate with and drive a 3D visualization model representing the patient in the metaverse.

[0045] Specifically, to achieve systematic management of patients' comprehensive health information, a digital profile object with a clear hierarchical logic is instantiated for each patient. Each layer of the digital profile corresponds to a type of data lifecycle and purpose.

[0046] The identity and foundation layer, serving as the root node of the archive, is implemented using relational database tables. In addition to storing the patient's unique identifier, which is associated with the user account system and demographic information, it also records core diagnostic information for Alzheimer's disease, such as clinical dementia rating and main symptoms, forming the static foundation of the archive.

[0047] The time-series data layer is designed for processing high-frequency, streaming data. It employs time-series databases such as InfluxDB and TimescaleDB, and its table structure includes pre-defined fields such as timestamps, data types (e.g., heart rate, steps, task completion events in the metaverse), data values, and data quality identifiers. This layer is specifically designed for efficiently storing and querying continuous monitoring data from IoT devices and metaverse logs.

[0048] The document and reporting layer manages unstructured and semi-structured documents. It utilizes object storage services combined with a search engine such as Elasticsearch. Raw files, including discharge summaries in PDF format and MRI image index files in DICOM format, are stored. Simultaneously, natural language processing techniques are used to extract and index key information from text reports, such as changes in MMSE scores and medication records, supporting rapid retrieval and correlation analysis.

[0049] The 3D twin vital signs layer maintains a data structure used to map and drive the patient's virtual avatar in the metaverse. It not only contains references to the avatar's static 3D model but also defines a series of state parameters, such as fatigue level, mood index, and completion rate of today's activities. These parameters will be bound to the real-time data in the lower layer through a rules engine.

[0050] Step 1202: Deploy standardized data access interfaces connected to the digital archives, including: a medical information system interface, an IoT device interface, and a metaverse behavior log interface; the medical information system interface is used to pull electronic medical records, test results, and image reports from external medical systems to the document and report layer; the IoT device interface is used to receive physiological and environmental data uploaded by wearable devices and environmental sensors to the time-series data layer; the metaverse behavior log interface is used to capture the patient's rehabilitation activities and social interactions in the metaverse, and classify and store them in the time-series data layer and the document and report layer.

[0051] Specifically, to enable dynamic updates to digital archives, real-time channels with internal and external data sources need to be established. This is achieved by deploying a set of standardized microservice interfaces. The medical information system interface is an asynchronous data retrieval service built on the HL7 FHIR RESTful API. It periodically or triggered by events, such as a patient completing an offline follow-up visit, sends requests to authorized external hospital information systems to retrieve the latest medical records and laboratory test results. After parsing, the structured portion of the retrieved data is stored in the identity and foundation layer, while the full-text document is stored in the document and report layer.

[0052] The IoT device interface is a message broker service based on the MQTT protocol. Devices such as smart bracelets worn by patients and indoor environmental sensors act as publishers, pushing encrypted physiological data packets, such as heart rate, nighttime activity levels, and indoor temperature and humidity, to this interface in real time. After verifying and parsing the data, the interface service writes it into the time-series data layer in the form of a data stream.

[0053] The Metaverse Behavior Log Interface is a core service integrated within the Metaverse server. It listens for all user interaction events in the virtual environment, such as starting cognitive training, talking to user A for 10 minutes in the community square, or completing the task of boiling water in the kitchen. These events are formatted into a unified JSON log. Quantitative performance data, such as reaction time and error count, is stored in the time-series data layer, while metadata describing the interaction, such as session summaries and activity video clip indexes, is stored in the document and report layer.

[0054] Step 1203: Based on the identity and information of the base layer, generate or match a base virtual avatar as the digital twin of the patient; establish a mapping relationship between key real-time data in the time-series data layer and the state parameters of the virtual avatar to realize the driving force of data to visualized state.

[0055] Specifically, static archives are linked to dynamic metaverse images to achieve a more intuitive presentation of data: Digital twin initialization: Based on identity and information in the base layer, such as age and gender, the system automatically matches or generates a basic virtual avatar from a preset virtual avatar that conforms to the characteristics of the elderly, or after confirmation by family members. This basic virtual avatar appears for the first time in the metaverse as the patient's digital twin.

[0056] The data-state mapping rule configuration uses a rule configuration engine to establish a mapping between time-series data and state parameters in the twin's symptom layer. For example, a configuration rule could be: IF average heart rate variability over the past 2 hours is below a threshold AND metaverse activity logs show frequent interruptions, THEN set the twin's state parameter fatigue level to high. Parameter changes trigger visual feedback from the avatar; for example, when fatigue is high, the avatar's movement speed slightly slows down or is accompanied by specific visual effects.

[0057] Step 1204: Clean, de-identify, and time-align the raw data imported through the interface; and extract features from the processed data based on preset rules or models to update the data association graph within the virtual health record.

[0058] Specifically, when faced with imported raw, multi-source, and heterogeneous data, automated governance is performed to improve data quality and uncover deeper correlations: Data cleaning and alignment involve denoising the raw data, such as removing sensor outliers, de-identifying data (e.g., removing direct personal identifiers), and time alignment. Time alignment is crucial; the system uses a network time protocol to synchronize timestamps from all data sources and uses the patient's local time as a reference to unify data from different frequencies onto the same timeline, laying the foundation for subsequent fusion analysis.

[0059] Using pre-defined rules, such as calculating average daily steps or lightweight machine learning models, such as identifying wandering behavior patterns from activity sequences, meaningful features are extracted from the cleaned data. These features, along with existing information in the archives, are used to update a built-in data association graph. This data association graph, centered on the patient entity, dynamically links nodes related to their symptoms, signs, behavioral patterns, treatment responses, etc., visually revealing the potential relationships between data.

[0060] Step 1205: Set update conditions for the digital archive, including timed triggering, data arrival triggering, and AI prediction-based triggering; when the conditions are met, automatically perform data integration, feature update, or digital twin state synchronization.

[0061] It should be noted that the AI ​​prediction trigger refers to the process whereby, when the risk probability output by the multimodal time-series prediction model exceeds a preset threshold, the output result serves as an event signal to trigger the virtual health record to update relevant labels, such as the "current risk label," to facilitate the matching of subsequent intervention strategies. The model output used for this trigger is based on data prior to the triggering time and does not include data updated in this instance, thus avoiding circular dependencies within the same round of prediction.

[0062] Specifically, to ensure that the records are almost synchronized with the patient's actual status in real time, an update strategy is preset that is triggered by multiple conditions, including timed triggering, data arrival triggering, and AI prediction triggering.

[0063] Scheduled triggers, for example, performing batch data aggregation and feature calculations every morning at midnight, and updating the activity summary of the past 24 hours.

[0064] Data arrival triggers the real-time update and risk assessment process of the archive whenever the IoT interface or metaverse log interface receives a new high-priority data packet, such as detecting a sudden change in acceleration suspected of being a fall.

[0065] AI prediction triggers the process. When the multimodal time-series prediction model outputs a high-risk warning in subsequent steps, the warning signal will serve as a trigger condition to drive the archive to actively update the current risk label and may preload relevant intervention plans into the archive's associated data area.

[0066] Step 1206: Configure a role-based access control list for the digital archive and enable audit logging with operator, timestamp and context for all data operations.

[0067] Specifically, considering the sensitivity of medical data, the system incorporates enterprise-level security controls, including: Role-based access control defines a detailed permission matrix. For example, a patient may only be able to view a summary of their own activities; family members can view detailed reports and receive alerts; the attending physician has read and write permissions for all data; and researchers can only access de-identified aggregated data. Access control lists are implemented at both the application gateway and database levels.

[0068] The entire process is audited through a dedicated audit service, recording all Create, Read, Update, and Delete operations on virtual health records. Each log entry includes an immutable timestamp, the operator's identity, the precise operation performed (e.g., "queried heart rate data from October 2023"), and the IP address from which the operation originated. This not only meets medical data compliance requirements but also provides a complete basis for post-event traceability and accountability.

[0069] Step 130: In the metaverse virtual space, create and configure virtual avatars for patients, family members, and medical staff; through the virtual avatars, provide rehabilitation activities to patients and support real-time interaction among multiple parties; wherein, the rehabilitation activities include at least virtual cognitive training courses and virtual sports rehabilitation games; the behavioral performance data and social data generated during the rehabilitation activities and the real-time interaction constitute multi-dimensional health data and are collected in real time; the multi-dimensional health data is collected in real time and updated to the corresponding virtual health profile; Among them, such as Figure 4 As shown, step 130 includes the following: Step 1301: The virtual cognitive training course adopts adaptive difficulty, which dynamically adjusts the task complexity and cognitive load type of the next training cycle based on the error rate and reaction time recorded by the patient in the virtual health record recently.

[0070] Specifically, virtual cognitive training is not a static set of numerical exercises, but rather a dynamic adjustment based on the patient's real-time abilities. The system maintains a structured cognitive task library, with tasks covering different cognitive domains such as memory, attention, executive function, and visuospatial. Each task has multiple difficulty levels with parameterized definitions, such as the number of memory items, the complexity of distractors, and the strictness of time constraints.

[0071] After each virtual cognitive training course, key performance indicators (KPIs) are automatically extracted from the course log, primarily including task error rate and average reaction time. These KPIs are then calculated and used as new data points, updating the time-series data layer of the patient's virtual health record in real time to form a continuous curve of changes in their cognitive abilities.

[0072] When planning the next training cycle, such as the next login or the task for the following day, the system invokes a lightweight decision-making algorithm. This algorithm takes the patient's recent error rate and reaction time trend in their records, such as over the past week, as its primary input. If recent performance consistently exceeds the success threshold, the algorithm will appropriately increase the complexity of the task in the next cycle, for example, by increasing the number of items to be memorized in a memory task.

[0073] If recent performance fluctuates or declines, the algorithm will maintain or slightly reduce the task difficulty, or even switch the type of cognitive load. For example, it may temporarily switch from a high-load working memory task to a lower-load sustained attention task to prevent the patient from feeling frustrated and maintain training motivation. This assessment-adjustment closed loop ensures the personalization and effectiveness of training.

[0074] Step 1302: The virtual motion rehabilitation game integrates an inertial measurement unit and a computer vision motion capture system to quantify and assess the patient's motion parameters in real time; the game provides real-time biofeedback in audiovisual form when the patient performs specific rehabilitation movements such as weight transfer and gait cycle.

[0075] Specifically, virtual exercise rehabilitation games aim to improve patients' adherence to tedious and repetitive physical training through fun and interactive activities, and provide accurate quantitative assessments of movement.

[0076] In multimodal motion capture and data fusion, patients wear devices with built-in IMU sensors, such as smart bracelets or straps, while playing games. The IMU provides high-frequency acceleration and angular velocity data for accurately calculating parameters such as joint angles and movement speed, and is particularly adept at capturing fast or small-amplitude movements. Simultaneously, motion videos are captured through patient terminals, such as depth-sensing cameras or RGB cameras. Using pre-trained pose estimation algorithms, such as OpenPose and MediaPipe, key points on the human body, such as the two-dimensional or three-dimensional coordinates of joints, are extracted in real time.

[0077] By combining IMU data with visual data through sensor fusion algorithms, such as Kalman filtering, the defects of visual perception being susceptible to occlusion and IMU drift can be compensated for. The final output is more stable and accurate real-time motion parameters, such as stride length, gait phase, and the amplitude of lateral and forward swing of the center of gravity in the gait cycle.

[0078] The game scenarios are designed to require patients to complete specific rehabilitation movements. For example, in the virtual river-crossing game, patients need to control their avatar to step on floating ice by shifting their center of gravity left and right.

[0079] When the aforementioned fusion technology detects that a patient has successfully completed a standard weight transfer movement, positive audiovisual feedback is immediately triggered, such as pleasant success sound effects, cheers from the virtual avatar, and an increase in in-game points. If the movement is detected as non-standard, such as insufficient transfer range or excessive forward leaning, guiding feedback is triggered: such as directional arrow prompts appearing at the edge of the screen, or gentle voice prompts from the virtual coach in the game, such as "Please move a little more to the left."

[0080] All these interactive events, quantified motion parameters, and feedback logs are captured in real time and written into the virtual health profile as behavioral performance data.

[0081] Step 1303: When the patient participates in the real-time interaction, the spatial distance between their virtual avatars, the duration of the interaction, the emotional tendency of their verbal communication, and their non-verbal actions are all captured by the system and stored in their virtual health record as structured social interaction data.

[0082] Specifically, social interactions in the metaverse are not only about emotional support, but their patterns are also an important window for assessing patients' social functioning and emotional state.

[0083] Multi-dimensional social signal capture includes spatial and temporal aspects, and language communication. Spatial and temporal analysis: The system records the spatial distance between the patient's virtual avatar and other avatars, which reflects social closeness and effective interaction time, excluding meaningless coexistence.

[0084] Language communication analysis, after obtaining the necessary authorization, performs real-time speech recognition on the speech in the interaction, and uses a sentiment analysis model to analyze the converted text to assess the emotional tendencies contained in the communication, such as positive, negative, neutral and emotional intensity.

[0085] Nonverbal action analysis records nonverbal social signals initiated by patients through preset actions of virtual avatars, such as waving, nodding, and clapping. Simultaneously, by analyzing the avatar's orientation and movement trajectory during interactions, the level of engagement can be indirectly inferred.

[0086] Data structuring and archiving: The raw signals captured above are processed in real time and transformed into structured social interaction data. For example, a 10-minute group chat might be recorded as an event containing the following fields: {Event ID, Participant list, Start time, End time, Average interpersonal distance, Patient's voice emotion score, Number of proactive nonverbal actions, Topic keywords}.

[0087] These structured event data are categorized and stored. Aggregated metrics, such as daily social interaction time and average sentiment score, are stored in the time-series data layer of the virtual health profile for trend observation; detailed event records are stored as log file indexes in the document and report layer for in-depth review and analysis.

[0088] All behavioral data generated and collected in steps 1301 to 1303, such as cognitive error rate, motor parameters, and social data, such as interaction indicators and emotional tendencies, collectively constitute multi-dimensional health data. This data is collected and processed in real time through the established metaverse behavior log interface (see step 1202), and immediately updated to the corresponding virtual health profile. This forms a real-time closed loop from environmental interaction to data accumulation, providing a fresh and continuous data stream for the AI ​​prediction in step 140, and is a key foundation for intelligent intervention throughout the entire life cycle.

[0089] Step 140: Based on the preset multimodal time-series prediction model, perform real-time analysis and trend prediction on the multidimensional health data for a preset duration to generate prediction information for potential complications or functional decline risks. Among them, such as Figure 5 As shown, the multimodal temporal prediction model is a multi-branch fusion network with a Transformer encoder at its core; it includes: a multimodal feature extraction module, a feature fusion and encoding module, and a prediction output module.

[0090] A multimodal feature extraction module is used to extract deep features of different types of data in parallel from the virtual health record, including: a temporal feature extraction branch, a text feature extraction branch, and a visual feature extraction branch. The temporal feature extraction branch is composed of a long short-term memory network, whose input is a sequence of physiological signals and behavioral events extracted from the virtual health record and aligned by timestamps, and whose output is a temporal feature vector representing individual physiological and behavioral patterns. The text feature extraction branch is composed of a natural language processing model, whose input is medical text reports and cognitive assessment records in the virtual health record, and whose output is a text feature vector representing semantic information and clinical conditions. The visual feature extraction branch is composed of a three-dimensional convolutional neural network, whose input is patient interaction video data captured in the metaverse virtual space, and whose output is a visual feature vector representing social interaction and non-verbal behavior patterns. The physiological signal sequence may include heart rate, blood pressure, and blood oxygen saturation; the behavioral event sequence may include login, logout, task start, task completion, and item interaction; and the medical text report may include discharge summary, medical record, and neuropsychological assessment report.

[0091] The feature fusion and encoding module integrates multimodal features and generates a context-aware fusion representation. It includes a cross-modal attention fusion submodule and a shared Transformer encoder. The cross-modal attention fusion submodule takes the temporal feature vector, textual feature vector, and visual feature vector as input, aligns and weights these three feature vectors, and outputs a unified fusion feature representation. The shared Transformer encoder takes the fusion feature representation as input and performs global context modeling and deep encoding on the fused features, outputting an encoded feature vector containing long-term dependencies. The Long Short-Term Memory (LSTM) network captures short-term local temporal patterns, such as physiological signal fluctuations over the past few hours, while the shared Transformer encoder performs global context modeling on the fused features, capturing dependencies across long time scales, such as days to weeks. The clear division of labor between the two avoids functional overlap.

[0092] The prediction output module is used to generate the final risk prediction and analysis results based on encoded features, including a prediction head. The prediction head takes the encoded feature vector as input and outputs the prediction result. The prediction result specifically includes: the risk probability value of a specific adverse event occurring at one or more preset time points in the future; and the attribution identifier of the input feature that contributes the most to the risk probability value. The specific adverse event includes a fall, an episode of agitated behavior, or a significant decline in cognitive function score. Here, a significant decline refers to a score decrease greater than a preset threshold, which can be set based on practical experience.

[0093] The feature attribution identifiers, for example, are calculated by integrating gradient or SHAP methods to determine the contribution of each input feature and map it to clinical terms. In one example, it can be generated as follows: During model prediction, a gradient-based method, such as integrating gradient or Gradient-Shap, is used to calculate the contribution of each dimension of the input feature vector to the final risk probability output; the top K features with the highest contribution, where K is a preset value, such as 5, are converted into clinically readable attribution descriptions, such as "a sudden drop in nighttime activity frequency over the past 3 days" or "increased amplitude of center of gravity swaying during virtual balance training," as the output feature attribution identifiers.

[0094] Furthermore, the training method for the multimodal time series prediction model includes: Training sample sets are extracted from the virtual health records of historical patients. Each training sample includes input data and label data. The input data consists of multimodal data within a preset historical time window that has undergone time alignment and normalization, specifically including physiological and behavioral time-series data, medical text reports, and corresponding interactive video clips. The label data consists of supervised learning labels constructed based on health events actually recorded within a preset future time period after the historical time window. The labels are multi-task labels, including at least a binary classification label indicating whether a specific adverse event occurred and a regression label indicating the time of event occurrence. A phased pre-training strategy is adopted to independently initialize each branch network constituting the multimodal temporal prediction model. The independent initialization of each branch network includes: pre-training the natural language processing model in the text feature extraction branch on a large general text dataset; pre-training the three-dimensional convolutional neural network in the visual feature extraction branch on a large general visual dataset; and pre-training the long short-term memory network in the temporal feature extraction branch using unlabeled medical time-series data through a self-supervised learning task. The pre-trained text feature extraction branch, visual feature extraction branch, and temporal feature extraction branch are assembled with the untrained cross-modal attention fusion module, Transformer encoder, and prediction head to form an initial prediction model. On the training sample set, the initial prediction model is jointly fine-tuned end-to-end with the objective of minimizing the multi-task loss function to obtain the multimodal temporal prediction model. The multi-task loss function consists of a weighted sum of cross-entropy classification loss, root mean square error regression loss, and contrastive loss to promote multimodal feature alignment. Before model deployment, the multimodal temporal prediction model is evaluated using an independent test set. Evaluation metrics include: AUC (ability to distinguish whether a risky event has occurred), F1 score (harmonic average of precision and recall), and BrierScore (calibration of probability prediction and stability of feature attribution, obtained through multiple perturbation tests). A model is considered to meet deployment requirements when AUC ≥ 0.85, F1 ≥ 0.80, and BrierScore ≤ 0.15.

[0095] After the multimodal time series prediction model is deployed, a federated learning framework is used for continuous optimization. The continuous optimization includes: each terminal using its private data to fine-tune its local model copy to generate model update parameters; the model update parameters of each terminal are encrypted and aggregated on a central server through a secure aggregation protocol to generate a global model update; and the global model update is distributed to each terminal to update its local model copy.

[0096] The terminals mentioned refer to the medical institutions or edge computing nodes authorized by the patients that participate in the model training. Each terminal has its own virtual health record data of the local patients, and the data does not leave the local area.

[0097] The contrast loss, used to facilitate the alignment of temporal features, textual features, and visual features before fusion, is defined as follows: L_contrastive=-log(exp(sim(z_t,z_t') / τ) / Σexp(sim(z_t,z_k) / τ)) Where z_t is the temporal feature vector, z_t' is the text feature vector or visual feature vector belonging to the same patient and time window as z_t, z_k is other feature vectors (negative samples) within the batch, sim(·) is the cosine similarity function, and τ is the temperature coefficient.

[0098] Step 150: Based on the predicted information, trigger and execute health intervention measures.

[0099] Specifically, such as Figure 6 As shown, step 150 includes the following: Step 1501: Receive prediction information generated by the multimodal time series prediction model, wherein the prediction information includes at least the risk event type, risk probability value, and feature attribution identifier.

[0100] Specifically, the system receives the output from the multimodal time series prediction model, which is a structured prediction information packet containing at least: Risk event types, clearly defined risk classifications, such as the risk of falling within the next 72 hours, the risk of agitated behavior within the next week, and the risk of a decline in cognitive function (MMSE equivalent score) of more than 3 points within the next month.

[0101] Risk probability value is a quantified probability value, such as 0.85, which represents the confidence level that the event will occur.

[0102] Feature attribution labels indicate which input features, such as a sharp drop in nighttime activity frequency over the past three days, an increase in the proportion of negative emotional words in recent social interactions, and increased center of gravity swaying during balance training games, contribute most to the prediction of current high risk. This provides a direct basis for subsequent targeted interventions.

[0103] Step 1502: Compare the risk probability value with multiple preset risk thresholds to classify the risk level into three levels: low, medium, and high.

[0104] Specifically, the system maintains a configurable set of risk threshold rules. This set of rules is differentiated based on different types of risk events and the patient's current disease stage, and is obtained from the virtual health record.

[0105] For example, the threshold for the risk of falling can be set as follows: low risk, probability < 0.3; medium risk, 0.3 ≤ probability < 0.7; high risk, probability ≥ 0.7.

[0106] The system compares the received risk probability values ​​with the corresponding event type thresholds in real time, automatically classifying the risk level into three levels. This level is the primary decision-making basis for triggering intervention strategies of different intensities.

[0107] Step 1503: Based on the risk level, risk event type, and characteristic attribution identifier, match and generate one or more initial intervention instructions from a predefined intervention strategy knowledge base; the intervention strategy knowledge base stores the mapping relationship between different risk scenarios and corresponding intervention measure types, execution objects, and execution platforms.

[0108] Specifically, the system has a pre-built knowledge base for intervention strategies, which is constructed from medical expert rules and historical data on successful intervention cases. Essentially, it is a decision support system that combines IF-THEN rules with case-based reasoning.

[0109] The knowledge base structure maps each record to a risk scenario and a set of interventions. A risk scenario is defined by three dimensions: risk level, risk event type, and key attribution characteristics.

[0110] Table 1 shows an example of an intervention strategy knowledge base. Taking fall risk as an example, matching can be achieved using a multi-level rule engine.

[0111] Table 1 The system uses the risk level output in step 1502, the risk event type in step 1501, and the characteristic attribution identifier as joint query keys to search the knowledge base. For example, if the scenario "{high risk, fall risk, attribution: decreased balance ability}" is matched, an initial set of intervention instructions may be generated: [Instruction A: Strengthen balance training in the metaverse; Instruction B: Send a fall prevention environment checklist to family members; Instruction C: Suggest that doctors assess whether the medication causes dizziness].

[0112] Step 1504: Based on the patient's personal preferences in the virtual health record, the current metaverse virtual space scene, the compliance with historical intervention measures, and the feedback effect, the parameters of the initial intervention instruction are individually fine-tuned to form an executable personalized intervention plan; the personalized fine-tuning includes: adjusting the way the intervention information is expressed, selecting a virtual scene preferred by the patient as the intervention trigger point, or adapting to real-world resources available to the patient.

[0113] Specifically, the initial intervention instructions are a general template, but they need to be adapted to the patient's specific situation to form the final plan.

[0114] The system queries the patient's virtual health record to obtain personal preferences, such as a preference for gentle reminders, the current metaverse scene (e.g., triggering environmental safety reminders more naturally in a virtual living room), historical intervention compliance (changing the format if the response rate to a certain type of reminder is low), and feedback from similar past interventions. For patients who prefer simplicity, lengthy nursing suggestions are condensed into a list of key points; for patients who need encouragement, positive tone is added to the messages. If a patient is about to log in for cognitive training, a balance ability reminder is set before the training begins; if the patient is frequently active in the virtual community square, a virtual nurse can be arranged to proactively care for and assess them in that scenario.

[0115] If the knowledge base suggests arranging in-person physical therapy, but the patient's records show that this service is not available in their community, the system will automatically adjust and recommend online video guidance courses or delivery of home rehabilitation equipment.

[0116] Step 1505: Based on the execution platform specified in the personalized intervention plan, distribute the intervention instructions to the corresponding terminal or system and trigger execution: If the execution platform is a metaverse virtual space, the system will present visual or auditory guidance information, adjust the difficulty of rehabilitation tasks, or dispatch virtual assistants to provide companionship and assistance in the environment where the patient's virtual avatar is located. If the execution platform is an associated mobile application or web platform, the system will push structured warning notifications, specific nursing suggestions, or rehabilitation plan adjustment to-do items to the client of the patient's family or medical staff. If the platform is an offline medical service system, the system generates and sends service work orders to community nursing centers, pharmacies, or attending physician workstations through standardized interfaces to initiate offline follow-up, medication delivery, or appointment adjustment processes.

[0117] Specifically, the personalized intervention plan specifies the execution platform for each instruction. The system distributes instructions to the corresponding execution terminal or system through dedicated interfaces on each platform.

[0118] When the metaverse virtual space is executed, instructions are sent to the target patient's client via the event bus of the metaverse server. This execution manifests as: non-invasive, contextualized visual cues appearing within the avatar's field of vision, such as highlighted safety paths on the ground; automatically adjusting the difficulty parameters of specific tasks in the day's rehabilitation game; or dispatching a pre-set virtual assistant avatar to approach and provide voice guidance and companionship.

[0119] When executed by related applications / web platforms, structured notifications are sent to dedicated clients of family members or healthcare professionals via push notification services or internal APIs. The notifications include clear risk statements, specific and actionable care recommendations (such as suggesting accompanying the patient for two short walks that day), and to-do items requiring healthcare confirmation or processing, such as generating a task card for checking a specific medication dosage in the collaborative management platform.

[0120] When offline healthcare service systems are implemented, intervention instructions are converted into standard service orders through interfaces that conform to medical information exchange standards, such as HL7 FHIR, and automatically sent to integrated external systems. For example, a service order for a community nurse to conduct a fall risk assessment at home is generated and sent to the community health information system; or a request for medication adherence support package delivery is generated and sent to the partner pharmacy platform.

[0121] Step 1506: After the intervention measures are implemented, the system monitors and collects the patient's response data and status change data through the metaverse virtual space, IoT devices and manual reporting channels.

[0122] Specifically, after the intervention is performed, the system immediately starts multi-channel monitoring to capture the patient's immediate response and short-term changes in condition.

[0123] Through monitoring within the metaverse, the system records whether the patient followed the guidance in the virtual environment, their subsequent activity trajectory, and changes in the emotional avatar's facial expressions. When monitoring via IoT devices, wearable devices can continuously monitor physiological parameters after intervention, such as whether the heart rate stabilizes and whether activity patterns change. When using manual reporting channels, a convenient feedback portal is provided through a client application for family members or medical staff to upload subjective observation records, such as whether the patient's mood has significantly improved or they still refuse to go out.

[0124] Step 1507: Associate the response data, state change data, and prediction information and intervention plan of this intervention, and store them as a complete intervention-feedback record in the patient's virtual health record for evaluation of the effectiveness of the intervention measures and optimization of the intervention strategy knowledge base and the multimodal time series prediction model.

[0125] Specifically, the system packages the entire data from this intervention into an intervention-feedback record. The record stores the original prediction information, the personalized intervention plan implemented, the response and status change data collected in step 1506, and the result label indicating whether the predicted risk event was avoided.

[0126] Optimize the intervention strategy knowledge base by analyzing a large number of records and using machine learning, such as reinforcement learning, to discover more effective combinations of intervention strategies or to modify the applicable conditions of existing strategies, so that the knowledge base evolves from expert rule-driven to data intelligence-driven.

[0127] Optimize multimodal time series prediction models: These complete records with outcome labels constitute post-intervention outcome data, which can be used for incremental training or feedback fine-tuning of multimodal time series prediction models, allowing the models to learn which interventions are most likely to be effective under what risk characteristics, thereby making more accurate and actionable predictions in the future.

[0128] Step 1508: When the probability of a specific risk indicated by the predicted information exceeds a preset threshold, the system automatically executes a tiered intervention strategy. Level 1 intervention is guidance within the metaverse, including: popping up contextual reminders through the patient's virtual avatar, adjusting the complexity of the virtual environment they are in, or assigning a guiding virtual assistant to accompany and guide tasks; Level 2 intervention involves cross-platform notifications, including: sending structured early warning reports and nursing advice lists to the linked family members' mobile terminals, and generating follow-up tasks or medication re-examination reminders to be processed on the collaborative management platform of the responsible medical staff. The three-tiered intervention involves initiating offline collaboration, including automatically generating referral suggestions or emergency contact information when the risk level is high, and pushing it to the nearest community health service center.

[0129] Specifically, when the predicted probability of a specific risk exceeds a preset extremely high threshold (e.g., ≥0.9), the system will trigger a mandatory tiered intervention strategy. This strategy can run in parallel with or take precedence over the regular processing flow described in steps 1502-1505 above. Tiered intervention aims to provide a rapid and standardized emergency response to the highest-risk scenarios, without requiring personalized fine-tuning, to ensure timely response.

[0130] Specifically, to address high-risk scenarios requiring rapid response, such as when the probability value exceeds an extremely high threshold, the system has pre-set a simplified, mandatory tiered emergency response procedure. This procedure can be implemented in parallel with or prioritized over the aforementioned personalized procedures. Level 1 intervention provides immediate guidance within the metaverse. For example, a security alert that cannot be easily closed may pop up in the center of the user's field of vision; or the visual elements of the user's current virtual environment may be simplified to reduce cognitive load and prevent anxiety caused by confusion; at the same time, a guiding virtual assistant may be immediately assigned to follow along and provide step-by-step instructions.

[0131] Secondary intervention can involve cross-platform emergency notifications, sending alert reports with high-priority indicators to all linked family members' mobile devices and automatically contacting key contacts. It can also push red alerts to the responsible healthcare provider's collaborative platform and automatically generate urgent tasks that must be handled within a specified timeframe.

[0132] Level 3 intervention can be linked offline. When the system determines that the risk level is the highest, such as life-threatening, it can automatically generate emergency contact information or electronic referral forms containing the patient's location and risk type, provided that authorization is obtained or the preset emergency protocol is met. These forms can be directly pushed to the nearest emergency center, community health service center, or the patient's attending physician's mobile phone through a standardized interface, thus initiating an offline emergency response.

[0133] Furthermore, the method also includes: For the raw data imported into the virtual health record, noise is added at the data source end or immediately after access, using differential privacy technology, or feature extraction is performed locally using federated learning. Only the encrypted feature vectors or model gradients are uploaded to the central server for aggregation analysis, making it impossible to identify specific patient individuals from the circulating data.

[0134] Furthermore, the method also includes: An edge computing architecture is adopted, and the core interactive logic and real-time data analysis module of the virtual cognitive training course and sports rehabilitation game are deployed on edge nodes close to the user. For latency-sensitive operations such as virtual avatar movement and environmental rendering, edge nodes process and respond in real time. Only data such as behavior logs and aggregated features that need to be stored and analyzed for a long time are asynchronously uploaded to the virtual health record in the central cloud.

[0135] Furthermore, the method also includes: The collaborative management platform integrates an insurance payment interface and a service subscription module. Once the health intervention measures are triggered and implemented, the system automatically generates structured service credentials and effect reports; According to the insurance terms pre-authorized by the patient or their family, the service voucher that meets the claim conditions is submitted to the insurance company's system through the insurance payment interface to initiate an automatic claim process; or, according to the subscription package, the corresponding service fee is deducted from the prepaid account.

[0136] By applying the AI-based full-lifecycle health intervention method provided in this invention, and through the construction of a metaverse space and virtual health records, it systematically integrates heterogeneous data from multiple sources, including disease diagnosis and treatment, rehabilitation training, daily life, and metaverse-specific behaviors and social interactions, for the first time. This forms a dynamically updated digital twin of the patient, addressing the pain points of fragmented and scattered health data and providing a complete and continuous data foundation for precise analysis. Furthermore, by utilizing a multimodal temporal prediction model to perform real-time analysis and trend prediction on the fused data, it can generate predictive information for risks such as falls and agitated behaviors in advance, achieving a leap from post-event treatment to pre-event warning and in-event intervention in proactive management. Furthermore, based on the predictive information, personalized intervention measures can be triggered simultaneously in the metaverse environment, mobile platforms, and offline medical services, forming a three-dimensional intervention network of virtual guidance, family reminders, and physical services, ensuring seamless connection from risk insight to physical action. Furthermore, immersive and gamified metaverse rehabilitation activities significantly increase patient participation; the system dynamically adjusts task difficulty based on real-time performance data and provides immediate biofeedback, keeping training within a personalized optimal challenge range, thereby optimizing rehabilitation outcomes. Furthermore, this application, based on a collaborative management platform within the same metaverse space, breaks through the limitations of time and space, enabling real-time information sharing and collaborative development of rehabilitation plans among family members and medical staff, thereby improving the collaboration efficiency and decision-making consistency of the care team.

[0137] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0138] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented in hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0139] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An AI-based full life cycle health intervention method, characterized by, The method includes: Construct a metaverse virtual space for Alzheimer's disease health management; In the metaverse virtual space, a dynamically updated virtual health record is created for each patient, serving as the core data hub for the patient, and together with the patient's virtual avatar, constitutes the patient's digital twin; In the aforementioned metaverse virtual space, virtual avatars are created and configured for patients, their families, and medical staff. Through these virtual avatars, rehabilitation activities are provided to patients, and real-time interaction among multiple parties is supported. These rehabilitation activities include at least virtual cognitive training courses and virtual exercise rehabilitation games. The behavioral and social data generated during these rehabilitation activities and real-time interactions constitute multi-dimensional health data, which is collected in real time. This multi-dimensional health data is collected in real time and updated to the corresponding virtual health profile. Based on a preset multimodal time-series prediction model, the multidimensional health data of a preset duration is analyzed and trend predicted in real time to generate predictive information for potential complications or functional decline risks. Based on the predicted information, health intervention measures are triggered and implemented.

2. The method of claim 1, wherein, The constructed metaverse virtual space for Alzheimer's disease health management specifically includes: Based on the full life cycle needs of Alzheimer's patients from diagnosis and rehabilitation to daily living support, multiple functionally independent but logically related virtual scene modules are set up; the virtual scene modules include at least: a virtual rehabilitation center for structured cognitive training, a virtual community square for social interaction and group support, a virtual home environment for simulating daily living skills training, and a virtual support center for family education and medical care collaboration. Based on scene planning, 3D modeling and game engine technology are used to construct 3D visual models of each virtual scene module; within the virtual rehabilitation center, interactive objects and triggers are preset and bound to cognitive training tasks; within the virtual community square, interaction rules supporting voice dialogue, facial expression communication, and preset group activities between virtual avatars are set; within the virtual home environment, interactive points that can simulate daily operations are set and linked to life skills assessment logic; within the virtual support center, document sharing, video conferencing, and planning dashboard tools are integrated to provide an interactive interface for collaborative management; A unified user account system is used to create unique identifiers for patients, their families, and medical staff. Based on this user account system, each user is provided with the ability to create and customize a virtual avatar, which can exist continuously and move freely in all virtual scene modules. Establish standardized data interfaces between the metaverse virtual space and external systems to ensure access to data streams from medical information systems and wearable devices; at the same time, integrate the various virtual scene modules, user systems and data interfaces through a unified central server or distributed network architecture to form a persistent, multi-user concurrent access coherent virtual world, namely the metaverse virtual space.

3. The method of claim 1, wherein, The creation of dynamically updated virtual health records for each patient in the metaverse virtual space specifically includes: A multi-layered digital profile is defined for each patient, comprising: an identity and foundation layer, a temporal data layer, a document and report layer, and a 3D twin vital signs layer. The identity and foundation layer stores the patient's unique identifier, demographic information, and basic diagnostic information. The temporal data layer uses a database table structure with preset fields to store temporal data of accessed physiological signals, behavioral events, and environmental interactions. The document and report layer stores non-temporally structured documents and medical image indexes. The 3D twin vital signs layer is used to associate with and drive a 3D visualization model representing the patient in the metaverse. Deploy standardized data access interfaces connected to the digital archives, including: a medical information system interface, an IoT device interface, and a metaverse behavior log interface; the medical information system interface is used to pull electronic medical records, test results, and image reports from external medical systems to the document and report layer; the IoT device interface is used to receive physiological and environmental data uploaded by wearable devices and environmental sensors to the time-series data layer; the metaverse behavior log interface is used to capture patients' rehabilitation activities and social interaction data in the metaverse, and classify and store them in the time-series data layer and the document and report layer. Based on the identity and information of the base layer, a basic virtual avatar is generated or matched as the digital twin of the patient; a mapping relationship is established between the key real-time data in the time series data layer and the state parameters of the virtual avatar to realize the driving force of data to visualized state; The raw data imported through the interface is cleaned, de-identified, and time-aligned; and features are extracted from the processed data based on preset rules or models to update the data association graph within the virtual health record. The digital archive is configured with update conditions including timed triggering, data arrival triggering, and AI prediction-based triggering; when the conditions are met, data integration, feature updates, or digital twin state synchronization are automatically performed. Configure a role-based access control list for the digital archives and enable audit logging with operator, timestamp, and context for all data operations.

4. The method according to claim 1, characterized in that, The provision of rehabilitation activities to patients through the metaverse virtual space and support for real-time interaction specifically includes: The virtual cognitive training course adopts adaptive difficulty, which dynamically adjusts the task complexity and cognitive load type of the next training cycle based on the error rate and reaction time recorded by the patient in the virtual health record recently. The virtual motion rehabilitation game integrates an inertial measurement unit and a computer vision motion capture system to quantify and assess the patient's motion parameters in real time; the game provides real-time audiovisual biofeedback when the patient performs specific rehabilitation movements such as weight transfer and gait cycle. When patients participate in the real-time interaction, the spatial distance between their virtual avatars, the duration of the interaction, the emotional tendency of their verbal communication, and their non-verbal actions are all captured by the system and stored in their virtual health records as structured social interaction data.

5. The method according to claim 1, characterized in that, The multimodal temporal prediction model is a multi-branch fusion network with a Transformer encoder at its core; including: A multimodal feature extraction module is used to extract deep features of different types of data in parallel from the virtual health record, including: a temporal feature extraction branch, a text feature extraction branch, and a visual feature extraction branch. The temporal feature extraction branch is composed of a long short-term memory network, whose input is a sequence of physiological signals and behavioral events extracted from the virtual health record and aligned by timestamps, and whose output is a temporal feature vector representing individual physiological and behavioral patterns. The text feature extraction branch is composed of a natural language processing model, whose input is medical text reports and cognitive assessment records in the virtual health record, and whose output is a text feature vector representing semantic information and clinical conditions. The visual feature extraction branch is composed of a three-dimensional convolutional neural network, whose input is patient interaction video data captured in the metaverse virtual space, and whose output is a visual feature vector representing social interaction and non-verbal behavior patterns. The feature fusion and encoding module is used to integrate multimodal features and generate a context-aware fusion representation, including: a cross-modal attention fusion submodule and a shared Transformer encoder; the cross-modal attention fusion submodule takes the temporal feature vector, text feature vector and visual feature vector as input, and its function is to align and weight the three feature vectors, and its output is a unified fusion feature representation; the shared Transformer encoder takes the fusion feature representation as input, and its function is to perform global context modeling and deep encoding on the fused features, and its output is an encoded feature vector containing long-term dependencies; The prediction output module is used to generate the final risk prediction and analysis results based on the encoded features, including a prediction head; the input of the prediction head is the encoded feature vector, and the output is the prediction result; the prediction result specifically includes: the risk probability value of a specific adverse event occurring at one or more preset time points in the future; the input feature attribution identifier that contributes the most to the risk probability value; wherein, the specific adverse event includes a fall, an episode of agitated behavior, or a significant decline in cognitive function score.

6. The method according to claim 5, characterized in that, The training method for the multimodal time-series prediction model includes: Training sample sets are extracted from the virtual health records of historical patients. Each training sample includes input data and label data. The input data consists of multimodal data within a preset historical time window that has undergone time alignment and normalization, specifically including physiological and behavioral time-series data, medical text reports, and corresponding interactive video clips. The label data consists of supervised learning labels constructed based on health events actually recorded within a preset future time period after the historical time window. The labels are multi-task labels, including at least a binary classification label indicating whether a specific adverse event occurred and a regression label indicating the time of event occurrence. A phased pre-training strategy is adopted to independently initialize each branch network constituting the multimodal temporal prediction model. The independent initialization of each branch network includes: pre-training the natural language processing model in the text feature extraction branch on a large general text dataset; pre-training the three-dimensional convolutional neural network in the visual feature extraction branch on a large general visual dataset; and pre-training the long short-term memory network in the temporal feature extraction branch using unlabeled medical time-series data through a self-supervised learning task. The pre-trained text feature extraction branch, visual feature extraction branch, and temporal feature extraction branch are assembled with the untrained cross-modal attention fusion module, Transformer encoder, and prediction head to form an initial prediction model. On the training sample set, with the goal of minimizing the multi-task loss function, the initial prediction model is jointly fine-tuned end-to-end to obtain the multimodal temporal prediction model; the multi-task loss function is composed of a weighted sum of cross-entropy classification loss, root mean square error regression loss, and contrastive loss used to promote multimodal feature alignment; After the multimodal time series prediction model is deployed, a federated learning framework is used for continuous optimization. The continuous optimization includes: each terminal using its private data to fine-tune its local model copy to generate model update parameters; the model update parameters of each terminal are encrypted and aggregated on a central server through a secure aggregation protocol to generate a global model update; and the global model update is distributed to each terminal to update its local model copy.

7. The method according to claim 1 or 5, characterized in that, The health intervention measures triggered and executed based on predictive information specifically include: Receive prediction information generated by the multimodal time series prediction model, wherein the prediction information includes at least the risk event type, risk probability value, and feature attribution identifier; The risk probability value is compared with multiple preset risk thresholds to classify the risk level into three levels: low, medium, and high. Based on the risk level, risk event type, and characteristic attribution identifier, one or more initial intervention instructions are matched and generated from a predefined intervention strategy knowledge base; the intervention strategy knowledge base stores the mapping relationship between different risk scenarios and corresponding intervention measure types, execution objects, and execution platforms; Based on the patient's personal preferences in the virtual health record, the current metaverse virtual space scene, the compliance with historical intervention measures, and the feedback effect, the parameters of the initial intervention instruction are individually fine-tuned to form an executable personalized intervention plan; the personalized fine-tuning includes: adjusting the way the intervention information is expressed, selecting the virtual scene preferred by the patient as the intervention trigger point, or adapting to the real-world resources available to the patient; Based on the execution platform specified in the personalized intervention plan, the intervention instructions are distributed to the corresponding terminal or system and triggered for execution: If the execution platform is a metaverse virtual space, the system will present visual or auditory guidance information, adjust the difficulty of rehabilitation tasks, or dispatch virtual assistants to provide companionship and assistance in the environment where the patient's virtual avatar is located. If the execution platform is an associated mobile application or web platform, the system will push structured warning notifications, specific nursing suggestions, or rehabilitation plan adjustment to-do items to the client of the patient's family or medical staff. If the platform is an offline medical service system, the system generates and sends service work orders to community nursing centers, pharmacies or attending physician workstations through standardized interfaces to initiate offline follow-up, drug delivery or appointment adjustment processes. After the intervention measures are implemented, the system monitors and collects the patient's response data and status change data through the aforementioned metaverse virtual space, IoT devices, and manual reporting channels; The response data, state change data, and prediction information and intervention plan of this intervention are associated and stored in the patient's virtual health record as a complete intervention-feedback record. This is used to evaluate the effectiveness of the intervention measures and optimize the intervention strategy knowledge base and the multimodal time series prediction model. When the probability of a specific risk indicated by the predicted information exceeds a preset threshold, the system automatically executes a tiered intervention strategy. The tiered intervention strategy includes Level 1 intervention, Level 2 intervention, and Level 3 intervention. Level 1 intervention is guidance within the metaverse, including: popping up contextual reminders in the patient's virtual avatar, adjusting the complexity of the virtual environment, or assigning a guiding virtual assistant for companionship and task guidance. Level 2 intervention is cross-platform notification, including: sending a structured early warning report and nursing suggestion list to the bound family member's mobile terminal, and generating pending follow-up tasks or medication re-examination reminders to the responsible medical staff's collaborative management platform. Level 3 intervention is initiating offline linkage, including: when the risk level is high, automatically generating referral suggestions or emergency contact information and pushing it to the nearest community health service center.

8. The method according to any one of claims 1 or 3, characterized in that, The method further includes: For the raw data imported into the virtual health record, noise is added at the data source end or immediately after access, using differential privacy technology, or feature extraction is performed locally using federated learning. Only the encrypted feature vectors or model gradients are uploaded to the central server for aggregation analysis, making it impossible to identify specific patient individuals from the circulating data.

9. The method according to any one of claims 1 or 2, characterized in that, The method further includes: An edge computing architecture is adopted, and the core interactive logic and real-time data analysis module of the virtual cognitive training course and sports rehabilitation game are deployed on edge nodes close to the user. For latency-sensitive operations such as virtual avatar movement and environmental rendering, edge nodes process and respond in real time. Only data such as behavior logs and aggregated features that need to be stored and analyzed for a long time are asynchronously uploaded to the virtual health record in the central cloud.

10. The method according to claim 1, characterized in that, The method further includes: The collaborative management platform integrates an insurance payment interface and a service subscription module. Once the health intervention measures are triggered and implemented, the system automatically generates structured service credentials and effect reports; According to the insurance terms pre-authorized by the patient or their family, the service voucher that meets the claim conditions is submitted to the insurance company's system through the insurance payment interface to initiate an automatic claim process; or, according to the subscription package, the corresponding service fee is deducted from the prepaid account.

Images