Multi-point scheduling intelligent health management system based on cloud data

Abstract

The invention relates to the technical field of health data management, and discloses a cloud data-based multi-point scheduling intelligent health management system, which comprises a multi-modal data acquisition and twinborn body construction module, a cloud data management module, a cloud data management module, a cloud data management module and a cloud data management module, the federated causal inference cloud brain module trains a global causal model under the condition of guaranteeing data privacy through federated learning; the collaborative decision-making and intervention scheduling module generates an intelligent intervention scheme based on the model; the abnormal resonance and risk traceability module identifies individual abnormity by comparing twinborn body states and finds a group resonance mode to trace unknown risks; and the closed-loop self-evolution module performs model iteration updating by using the execution result and the traceability risk. According to the technical scheme, the federated learning framework and the cloud causal inference model are combined, and the technical effects of global knowledge sharing and causal relationship mining are achieved on the premise that original data of all nodes are not directly accessed.

Description

Technical Field

[0001] This invention relates to the field of health data management technology, specifically to a multi-point scheduling intelligent health management system based on cloud data. Background Technology

[0002] The managed individuals are distributed across different regions, and their health data originates from various institutions and devices, forming a naturally distributed structure. How to effectively utilize this dispersed data to provide dynamic, reliable, and predictive health guidance for each individual is a core technological challenge currently facing the field of smart health management.

[0003] In the existing technology, there are various systems for health data analysis. One type of system adopts a centralized data processing architecture, which aggregates data to a central server and uses powerful computing resources for batch analysis, enabling the discovery of valuable statistical correlations from large-scale data. Another type of system uses rule-based expert systems, which encode mature clinical medical guidelines and knowledge, and can provide stable and reliable monitoring and alerting services for known health risk indicators (such as specific thresholds for blood glucose and blood pressure).

[0004] However, existing technologies still face several unresolved issues in practical applications. First, centralized data processing requires raw data to leave its parent institution, which is difficult to achieve in real-world multi-party collaboration scenarios due to data privacy and security regulations and data ownership issues, leading to ineffective data flow and the formation of de facto data silos. Second, existing systems rely heavily on correlation analysis when generating intervention recommendations, and their decision-making process is opaque to users and managers. When system recommendations conflict with human expert experience, there is a lack of objective, evidence-based conflict resolution, limiting the adoption rate and credibility of system recommendations. Finally, the risk monitoring capabilities of existing systems are limited to known risk factors and fixed judgment rules, failing to proactively identify weak, collective abnormal patterns caused by unknown factors. Furthermore, once deployed, system models typically lack automated closed-loop feedback mechanisms and cannot iterate based on real-world intervention effects and newly discovered knowledge, gradually reducing their long-term effectiveness. Therefore, those skilled in the art propose a cloud-based multi-point scheduling intelligent health management system to address these issues. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a cloud-based multi-point scheduling intelligent health management system, which solves the technical problems in existing technologies, such as the difficulty in joint data analysis due to data privacy and security restrictions, the lack of transparency in intelligent decision-making processes and conflict resolution mechanisms, and the inability to proactively discover unknown group risks and carry out closed-loop self-evolution.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A cloud-based, multi-point scheduling intelligent health management system includes:

[0008] The module includes a multimodal data acquisition and twin construction module, a federated causal inference cloud brain module, a collaborative decision-making and intervention scheduling module, an anomaly resonance and risk tracing module, and a closed-loop self-evolution module.

[0009] The multimodal data acquisition and twin construction module is used to perform the following operations for each of the multiple management objects:

[0010] Collect its multimodal health data, which includes static data, quasi-static data, high-frequency dynamic data, and interactive data;

[0011] The collected multimodal health data are fused into a unified multidimensional health state vector;

[0012] Based on the health status vector, a dynamic state transition model is instantiated for each managed object as its personalized health digital twin. The dynamic state transition model is used to calculate and output the predicted health status vector for the next moment based on the current health status vector, an externally input intervention vector, and a set of personalized model parameters characterizing the individual characteristics of the managed object.

[0013] The federated causal inference cloud brain module is used to train a global causal inference model based on the health status vectors of multiple nodes and corresponding intervention data through a federated learning framework. Its training steps include:

[0014] The global causal inference model, either initial or from the previous iteration, is distributed from a cloud server to multiple local servers deployed on edge nodes.

[0015] The local server of the edge node uses its local private health status vector and intervention data to train the received global causal inference model locally.

[0016] The local server of the edge node updates and uploads the encrypted model parameters generated after local training to the cloud server;

[0017] The cloud server securely aggregates the encrypted model parameter updates uploaded from all edge nodes participating in the training to update the global causal inference model.

[0018] The updated global causal inference model is used to characterize the causal relationship between interventions and health outcomes.

[0019] The collaborative decision-making and intervention scheduling module is used to respond to the health status vector of a specific managed object and, based on the global causal inference model, generate an intelligent intervention plan for that managed object. Its generation steps include:

[0020] By querying the global causal inference model, an optimal intervention recommendation for the managed object is obtained;

[0021] The optimal intervention recommendation is packaged into a decision support report, which includes the content of the optimal intervention recommendation, the expected health trajectory generated after the personalized health digital twin of the specific management object is executed, and historical intervention cases of individuals with similar characteristics retrieved from the global knowledge base.

[0022] Furthermore, the collaborative decision-making and intervention scheduling module is also used for:

[0023] When the intelligent intervention plan conflicts with an externally input human expert plan, the intelligent intervention plan and the human expert plan are defined as a first intervention hypothesis and a second intervention hypothesis, respectively.

[0024] The personalized health digital twin of the specific management object is invoked to simulate and execute the first intervention hypothesis and the second intervention hypothesis in parallel in a virtual environment, so as to generate two independent future health status prediction trajectories respectively;

[0025] Generate a visual comparison report containing the two independent future health status prediction trajectories to assist in the final clinical decision-making.

[0026] The abnormal resonance and risk tracing module is used to identify individual abnormal states by continuously comparing the real health state vector and the predicted health state vector of the personalized health digital twin, and to scan the personalized health digital twins of multiple managed objects to identify group-wide abnormal resonance patterns, thereby tracing the source of unknown risk factors. The steps for identifying the group-wide abnormal resonance patterns include:

[0027] Calculate the prediction deviation value between the actual health status vector and the predicted health status vector, and when the prediction deviation value is continuously higher than the dynamic threshold set for the managed object, mark the personalized health digital twin of the managed object as abnormal;

[0028] When a personalized health digital twin is marked as abnormal, a scan of personalized health digital twins of all managed objects is initiated. By employing a time series similarity matching algorithm, the similarity between the prediction deviation time series of different personalized health digital twins is calculated to identify other personalized health digital twins with prediction deviation time series patterns similar to the abnormality. All identified personalized health digital twins with similar patterns are formed into an abnormal resonance cluster. The deviation time series pattern commonly manifested by the abnormal resonance cluster is then determined as the group-wide abnormal resonance pattern.

[0029] Furthermore, the steps of the abnormal resonance and risk tracing module in tracing the unknown risk factors include:

[0030] Extract the shared metadata of the management objects corresponding to all personalized health digital twins within the abnormal resonance cluster;

[0031] Cross-analysis of the shared metadata was performed to identify common influencing factors;

[0032] Based on the common influencing factors, a new causal hypothesis is generated regarding the unknown risk factors.

[0033] The closed-loop self-evolution module is used to iteratively update the global causal inference model and the personalized health digital twin based on the execution results of the intelligent intervention plan and the traced unknown risk factors. The update steps include:

[0034] The execution results of the intelligent intervention plan are collected as routine intervention feedback data;

[0035] Data from the collaborative decision-making and intervention scheduling module during the handling of human-machine conflict is collected as counterfactual decision feedback data. Specifically, the counterfactual decision feedback data includes the intervention plan that is ultimately adopted, the actual health results after the implementation of the plan, the intervention plan that is rejected, and the predicted health trajectory obtained after virtual simulation of the rejected plan.

[0036] Collect the unknown risk factors generated by the abnormal resonance and risk tracing module and verified as new knowledge injection data;

[0037] Based on the conventional intervention feedback data, the counterfactual decision feedback data, and the new knowledge injection data, the global causal inference model and the personalized health digital twin are updated respectively.

[0038] This invention provides a cloud-based, multi-point scheduling intelligent health management system. It offers the following advantages:

[0039] 1. This invention adopts a technical solution that combines a federated learning framework with a cloud-based causal inference model. This solution achieves the technical effect of global knowledge sharing and causal relationship mining without directly accessing the original data of each node. Compared with the existing technical solutions that rely on centralized data storage for model training, this invention solves the shortcomings of high data privacy leakage risk and difficulty in breaking down cross-institutional data barriers.

[0040] 2. This invention designs a personalized health digital twin for virtual simulation and incorporates a collaborative decision-making mechanism. This design achieves the ability to provide quantitative predictive comparisons of different intervention programs, significantly enhancing the interpretability of intelligent intervention programs. Existing technologies typically output a black-box recommendation result directly. Compared to these, the solution of this invention overcomes the shortcomings of opaque decision-making processes and difficulty in gaining the trust of clinicians.

[0041] 3. This invention constructs an abnormal resonance and risk tracing module, and links it with a closed-loop self-evolution module that collects feedback from multiple sources. This construction enables the proactive discovery of unknown, cross-node group risks. The system can also continuously optimize its own model using newly discovered knowledge and decision results. Compared with the existing technology, which can only passively monitor known risks, this invention solves the technical problems of being unable to cope with unknown risks and the model performance deteriorating over time. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the system structure of the present invention;

[0043] Figure 2 This is a functional diagram of the multimodal data acquisition and twin construction module of the present invention;

[0044] Figure 3 This is a schematic diagram of the federated learning interaction process of the federated causal inference cloud brain module of the present invention;

[0045] Figure 4 This is a functional diagram of the collaborative decision-making and intervention scheduling module of the present invention;

[0046] Figure 5 This is a functional diagram of the abnormal resonance and risk tracing module of the present invention;

[0047] Figure 6 This is a schematic diagram of the data flow and functions of the closed-loop self-evolution module of the present invention.

[0048] Among them, 100 is the multimodal data acquisition and twin construction module; 200 is the federated causal inference cloud brain module; 300 is the collaborative decision-making and intervention scheduling module; 400 is the abnormal resonance and risk tracing module; and 500 is the closed-loop self-evolution module. Detailed Implementation

[0049] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] See attached document Figure 1 , Figure 1 This is a structural block diagram of a cloud-based multi-point scheduling intelligent health management system according to an embodiment of the present invention. The system provided by the present invention may include: a multimodal data acquisition and twin construction module 100, a federated causal inference cloud brain module 200, a collaborative decision-making and intervention scheduling module 300, an anomaly resonance and risk tracing module 400, and a closed-loop self-evolution module 500.

[0051] In one embodiment of the present invention, the system's workflow is as follows:

[0052] The multimodal data acquisition and twin construction module 100 serves as the system's data input and model foundation unit. It connects to data acquisition devices of multiple managed objects, continuously acquiring multimodal health data for each object and fusing this heterogeneous data into a standardized, real-time updated health status vector. Simultaneously, based on the health status vector, this multimodal data acquisition and twin construction module 100 constructs and maintains a dynamically evolving, personalized health digital twin for each managed object. This personalized health digital twin is the fundamental computational model for all subsequent simulation and prediction operations.

[0053] The Federated Causal Inference Cloud Brain Module 200 connects to data processing units on multiple nodes, but it is designed not to directly access the raw health data stored on each node. Instead, the module receives standardized health state vectors and corresponding intervention data from each node. Under the scheduling of a federated learning framework, it trains a global causal inference model by performing steps such as model distribution, local training, encrypted update upload, and cloud aggregation. This global model represents the causal relationship between interventions and health outcomes and is stored on a cloud server.

[0054] The collaborative decision-making and intervention scheduling module 300 is the system's decision output unit. When an intervention plan needs to be generated for a specific managed object, the collaborative decision-making and intervention scheduling module 300 first obtains the object's latest health status vector from the multimodal data acquisition and twin construction module 100. Subsequently, the collaborative decision-making and intervention scheduling module 300 calls the global causal inference model in the federated causal inference cloud brain module 200 to perform calculations based on the input health status vector, generate an intelligent intervention plan, and output it.

[0055] The Anomaly Resonance and Risk Source Tracing Module 400 is the system's risk monitoring unit. It continuously acquires the true health status vector of each managed object from the multimodal data acquisition and twin construction module 100 and compares it with the predicted health status vector output by its personalized health digital twin. By calculating the deviation between the two, the Anomaly Resonance and Risk Source Tracing Module 400 identifies the individual's abnormal state. Once an anomaly is identified, the module scans the personalized health digital twins of all managed objects to discover group-wide anomaly resonance patterns and analyzes these patterns to trace the source of unknown risk factors.

[0056] The closed-loop self-evolution module 500 is the system's model iteration unit. This module receives actual execution result data of the intelligent intervention plan from the collaborative decision-making and intervention scheduling module 300, as well as data on unknown risk factors traced by the anomaly resonance and risk tracing module 400. The closed-loop self-evolution module 500 processes this multi-source feedback data and generates update instructions.

[0057] Ultimately, the update command is sent to the Federated Causal Inference Cloud Brain Module 200 and the Multimodal Data Acquisition and Twin Construction Module 100, respectively. The Federated Causal Inference Cloud Brain Module 200 updates its global causal inference model based on the command, while the Multimodal Data Acquisition and Twin Construction Module 100 updates the affected personalized health digital twins accordingly. This process constitutes a closed-loop operating mode for the system, enabling the system model to iterate based on newly generated data and information.

[0058] To further clarify the specific functions and working principles of each module in the embodiments of the present invention, the following will explain... Figure 1 The internal structure and specific implementation of the multimodal data acquisition and twin construction module 100, the federated causal inference cloud brain module 200, the collaborative decision-making and intervention scheduling module 300, the abnormal resonance and risk tracing module 400, and the closed-loop self-evolution module 500 shown are described in detail.

[0059] See attached document Figure 2 , Figure 2 This is a functional block diagram of a multimodal data acquisition and twin construction module 100 according to an embodiment of the present invention. In a specific embodiment of the present invention, the function of the multimodal data acquisition and twin construction module 100 is implemented through the following steps.

[0060] First, the multimodal data acquisition and twin construction module 100 collects multimodal health data for each managed object. This multimodal health data can be categorized into four types: static feature data, including genotype data, historical medical records, and drug allergy history data; quasi-static feature data, including various biochemical indicators and medical imaging diagnostic results from recent physical examination reports; high-frequency dynamic feature data, including continuous heart rate, blood oxygen saturation, activity level, and sleep structure collected through wearable devices; and interactive feature data, including daily dietary records and symptom complaints obtained through text or voice interaction.

[0061] After data acquisition, the multimodal data acquisition and twin construction module 100 performs data fusion processing on the heterogeneous data. This processing includes data cleaning, data format standardization, and timestamp alignment. The output of the fusion processing is a unified multidimensional health state vector, denoted as S. i,t This vector comprehensively represents the health snapshot of the managed object i at time point t.

[0062] After generating the health state vector S i,t Subsequently, the multimodal data acquisition and digital twin construction module 100 instantiates a dynamic state transition model for each managed object i. This model serves as the object's personalized health digital twin. The function of the dynamic state transition model is to receive the current state and a specified intervention plan, and calculate and output a predicted future health state. Its calculation process can be represented by the following formula:

[0063] S i,t+1 =f(S) i,t A i,t ,θ i )+∈ i,t ;

[0064] Wherein: S i,t S is a multidimensional health status vector of managed object i at time point t; i,t+1 It is the predicted health status vector of managed object i at the next time point t+1, calculated by the model; A i,t θ is an intervention vector that represents the intervention plan applied to the managed object i at time point t. The intervention vector may include parameters such as medication type and dosage, duration and intensity of physical activities; iis a set of personalized model parameters specific to the managed object i. This parameter set can be initialized using the static feature data of the object and can be subsequently updated by the closed-loop self-evolution module 500; f is a nonlinear state transition function used to calculate the process of health status changing with time and intervention measures. In one embodiment of the present invention, this function can be implemented by a Long Short-Term Memory (LSTM) network or a Gated Recurrent Unit (GRU) network to effectively handle the time-series characteristics of the health status vector; ∈ i,t It is a random error term used to characterize random physiological fluctuations or external environmental factors that the model fails to account for.

[0065] By performing the above instantiation process for each managed object, the multimodal data acquisition and twin construction module 100 establishes a unique, computable, personalized health digital twin for all managed objects in the system.

[0066] See attached document Figure 3 , Figure 3 This is a flowchart illustrating the federated learning interaction of a federated causal inference cloud brain module 200 according to an embodiment of the present invention. In a specific embodiment of the present invention, the functionality of the federated causal inference cloud brain module 200 is achieved through the following steps.

[0067] The Federated Causal Inference Cloud Brain Module 200 operates within a collaborative computing architecture of a cloud server and edge nodes. The cloud server is responsible for maintaining and iterating a global causal inference model, while multiple edge nodes deployed in different physical locations (e.g., different medical or health management institutions) are responsible for distributed computing using their local data. The Federated Causal Inference Cloud Brain Module 200 trains the global causal inference model by executing a federated learning protocol without transmitting the private raw data of each edge node.

[0068] One specific implementation of the federated learning protocol includes the following steps:

[0069] Model distribution: The cloud server distributes the model parameters of the global causal inference model for the current training round. Distribute to all edge nodes k participating in this round of training.

[0070] Local training: After receiving the model parameters, each edge node k uses its locally stored private dataset, which is not transmitted externally. (Including health status vectors and corresponding intervention data), the model is trained locally for a specified number of rounds to obtain a set of updated local model parameters.

[0071] Update Upload: Each edge node k encrypts its locally computed model parameter updates and uploads the encrypted results to the cloud server. Model parameter updates can be complete local model parameter updates. It can also be the gradient of parameters calculated locally.

[0072] Secure aggregation: After receiving encrypted model parameter updates from all participating nodes, the cloud server performs a secure aggregation operation to calculate the new global causal inference model parameters.

[0073] In one embodiment, the secure aggregation operation can be implemented by weighted averaging of the model parameters of all nodes, and the calculation process can be characterized by the following formula:

[0074]

[0075] in: and These are the model parameters of the global causal inference model for round t+1 and round t, respectively; is the local model parameter of edge node k after it completes local training in round t+1; N is the total number of edge nodes participating in this round of training; The dataset used for local training by edge node k The number of samples in the sample; It is the total number of samples in the dataset of all nodes participating in the training, i.e.

[0076] The training objective of the global causal inference model is to learn the causal relationship between interventions and health outcomes, rather than traditional statistical correlation. The model is trained to estimate a conditionally averaged intervention effect (CATE), which predicts the potential health outcome after implementing an intervention in a given health state. Therefore, the goal of model learning is to obtain a function G that takes the current state and the specified intervention as input and outputs a predicted potential outcome. The model training process employs a loss function designed based on causal inference theory to eliminate or mitigate the effects of confounding biases in the observed data during training.

[0077] See attached document Figure 4 , Figure 4 This is a functional flowchart of a collaborative decision-making and intervention scheduling module 300 according to an embodiment of the present invention. In a specific embodiment of the present invention, the function of the collaborative decision-making and intervention scheduling module 300 is implemented through the following steps.

[0078] After receiving an instruction to generate an intervention plan for a specific managed object i, the collaborative decision-making and intervention scheduling module 300 first obtains the latest health status vector S of the object from the multimodal data acquisition and twin construction module 100. i,t .

[0079] Subsequently, the collaborative decision-making and intervention scheduling module 300, based on this health state vector S i,t By querying the global causal inference model G trained by the Federated Causal Inference Cloud Brain module 200, an optimization problem is solved to generate optimal intervention recommendations. The solution process can be characterized by the following formula:

[0080]

[0081] in: G(S) is the calculated optimal intervention recommendation for managed object i at time point t; A is a vector of candidate intervention measures traversed; G(S) is the optimal intervention recommendation for managed object i at time point t. i,t A) is a global causal inference model for input state S. i,t The expected health outcome output by candidate intervention A; U(·) is a predefined utility function used to quantify an expected health outcome, such as weighting a set of health indicators to maximize a comprehensive health score or minimize the risk prediction value of a specific disease.

[0082] After generating the optimal intervention recommendations Subsequently, the collaborative decision-making and intervention scheduling module 300 encapsulates it into a decision support report. This report specifically includes three components: optimal intervention recommendations. The specific content; calling the personalized health digital twin model of management object i, and... The projected future health trajectory is generated after calculation using the input; and a data source with similar characteristics to S is retrieved from a global knowledge base. i,t Historical intervention cases of individuals with similar feature vectors.

[0083] In another embodiment of the present invention, the collaborative decision-making and intervention scheduling module 300 further includes a mechanism for handling conflicting plans. When the collaborative decision-making and intervention scheduling module 300 generates an intelligent intervention plan... Compared to a human expert protocol A′ input from an external source (e.g., a clinician). i,t When inconsistency occurs, this mechanism is triggered and the following steps are executed:

[0084] The intelligent intervention plan and the human expert plan are defined as a first intervention hypothesis and a second intervention hypothesis, respectively.

[0085] The personalized health digital twin of a specific managed object i, i.e. its dynamic state transition model f, is invoked and simulated in parallel in a virtual environment.

[0086] Under the first intervention hypothesis, through continuous iterative calculations, the first independent future health status prediction trajectory is generated, denoted as .

[0087] Under the second intervention hypothesis, a second independent future health status prediction trajectory is generated through continuous iterative calculations, denoted as...

[0088] Where T is the length of the preset simulation time window.

[0089] The collaborative decision-making and intervention scheduling module 300 ultimately generates a visual comparison report containing two independent future health status prediction trajectories, and outputs the report to human decision-makers to assist them in making the final intervention plan selection.

[0090] See attached document Figure 5 , Figure 5 This is a functional flowchart of an abnormal resonance and risk tracing module 400 according to an embodiment of the present invention. In a specific embodiment of the present invention, the function of the abnormal resonance and risk tracing module 400 is implemented through the following steps.

[0091] The abnormal resonance and risk tracing module 400 continuously monitors the health status of each managed object i. This monitoring is achieved by comparing the object's true health status vector obtained from the multimodal data acquisition and twin construction module 100 with the predicted health status vector output by the object's personalized health digital twin.

[0092] First, the anomaly resonance and risk tracing module 400 calculates the prediction deviation between the actual health state vector and the predicted health state vector using a distance or divergence metric function. This calculation process can be represented by the following formula:

[0093]

[0094] Where: d i,t It is the prediction deviation value of managed object i at time point t; It is the vector of the actual health status of managed object i at time point t; D(·,·) is the predicted health status vector output by the personalized health digital twin of managed object i at time point t; D(·,·) is a metric function used to calculate the difference between the two vectors.

[0095] In one embodiment, the function can be Mahalanobis distance or KL divergence.

[0096] After calculating the prediction deviation value d i,t The post-abnormal resonance and risk tracing module 400 compares it with a dynamic threshold τ set for the managed object i. i Compare the prediction deviation value d. i,t The value remains above the dynamic threshold τ for a preset time window. i At that time, the abnormal resonance and risk tracing module 400 marks the personalized health digital twin of the managed object i as abnormal.

[0097] When any personalized health digital twin in the system is marked as abnormal, the anomaly resonance and risk tracing module 400 initiates a scan of the personalized health digital twins of all managed objects to identify group-wide anomaly resonance patterns. This scan employs a time-series similarity matching algorithm to calculate the prediction deviation time series {d} of different personalized health digital twins. j,t The similarity between individuals is used to identify personalized health digital twins that exhibit similar predictive bias patterns across different physiological dimensions to those of individuals with labeled abnormalities. In one embodiment, the time series similarity matching algorithm is the Dynamic Time Warping (DTW) algorithm.

[0098] All identified personalized health digital twins with similar predictive bias time series patterns collectively constitute an anomalous resonance cluster C. Furthermore, the deviation time series pattern collectively exhibited by this anomalous resonance cluster C is identified by the anomalous resonance and risk tracing module 400 as a group-wide anomalous resonance pattern.

[0099] After identifying the collective abnormal resonance pattern and forming an abnormal resonance cluster C, the abnormal resonance and risk tracing module 400 performs the step of tracing the unknown risk factors. This step includes: first, extracting shared metadata of the management objects corresponding to all personalized health digital twins within the abnormal resonance cluster C. The metadata may include geographical location, affiliated institution, batch number of specific drugs, and source information of specific foods; then, performing cross-analysis on the extracted shared metadata to identify common influencing factors; finally, based on the identified common influencing factors, generating a new causal hypothesis about the unknown risk factors and outputting the hypothesis.

[0100] See attached document Figure 6 , Figure 6 This is a data flow and functional block diagram of a closed-loop self-evolution module 500 according to an embodiment of the present invention. In a specific embodiment of the present invention, the function of the closed-loop self-evolution module 500 is implemented through the following steps.

[0101] The closed-loop self-evolution module 500 is the unit responsible for model iteration in the system. It continuously updates the key models within the system by collecting, processing, and distributing multi-source feedback data. The feedback data collected by the closed-loop self-evolution module 500 specifically includes three types.

[0102] The first type is routine intervention feedback data. This data originates from the results of the intelligent intervention plan output by the collaborative decision-making and intervention scheduling module 300 after its execution. For each intervention, the closed-loop self-evolution module 500 collects the actual health status before the intervention, the intervention measures implemented, and the actual health status after the intervention, forming a standard training data tuple. This tuple can be represented as:

[0103]

[0104] Wherein: S t It is the state vector before intervention; A t It is the vector of the intervention measures being implemented; It is the true state vector after intervention.

[0105] The second type is counterfactual decision feedback data. This data originates from the complete decision-making process of the collaborative decision-making and intervention scheduling module 300 when handling human-machine conflict. The closed-loop self-evolution module 500 collects all relevant information from this process, forming a counterfactual training data tuple. This tuple can be represented as:

[0106]

[0107] Wherein: S i,t It is the health status vector of a specific managed object i before the decision is made; It is the intervention plan that is ultimately adopted and implemented; These are the actual health outcomes resulting from the implementation of the adopted plan; It is an intervention plan that was rejected in the conflict; It is a predicted health trajectory obtained by performing a virtual simulation on the rejected proposal.

[0108] The third type is new knowledge-injected data. This data originates from unknown risk factors output by the anomaly resonance and risk tracing module 400 and has been externally verified and confirmed. This data represents a new, confirmed causal relationship.

[0109] After collecting the three types of data mentioned above, the closed-loop self-evolution module 500 processes and distributes them. Routine intervention feedback data and counterfactual decision feedback data are formatted as training samples and transmitted to the federated causal inference cloud brain module 200 to update the global causal inference model in the next round of federated learning. Simultaneously, the portion of these data relevant to a specific individual is also transmitted to the multimodal data acquisition and twin construction module 100 to update and calibrate the model parameters of that individual's personalized health digital twin.

[0110] For newly injected knowledge data, the closed-loop self-evolution module 500 interprets it as instructions to modify the model structure. For example, a newly identified risk factor may be added as a new input feature to the global model in the federated causal inference cloud brain module 200, or a specific constraint term may be added to the model structure to directly reflect the causal relationship of the newly discovered data. Through this process, the system achieves closed-loop iteration and self-evolution based on multi-source data feedback.

[0111] To further illustrate the collaborative working process of the technical solution of this invention, a specific working scenario example will be used below.

[0112] Scenario Description: This embodiment is applied to the unified smart health management of multiple elderly patients with chronic diseases (hereinafter referred to as "managed objects") located in three health management centers (A, B, and C) at different geographical locations. Centers A, B, and C serve as edge nodes of the system, each responsible for local data processing of its managed objects. The core modules of this invention are deployed on a cloud server.

[0113] First, the initialization of the personalized health digital twin: Taking subject A (suffering from hypertension) at center A as an example. After the system starts, the multimodal data acquisition and twin construction module 100 collects multimodal health data of subject A, including his / her historical medical records, recent physical examination reports, daily blood pressure readings uploaded through a smart blood pressure monitor, and medication records and dietary information entered through the human-computer interaction interface. The multimodal data acquisition and twin construction module 100 integrates these data into a health state vector S of subject A at the current moment. 甲,t Based on this vector and A's static feature data, a personalized health digital twin model was initialized for A.

[0114] Secondly, the routine intervention plan is generated based on federal causal inference: Health management personnel at Center A need to develop a blood pressure management plan for subject A for the next cycle. Upon receiving this instruction, the collaborative decision-making and intervention scheduling module 300 first obtains A's latest health status vector S. 甲,tSubsequently, the collaborative decision-making and intervention scheduling module 300 invokes the global causal inference model in the federated causal inference cloud brain module 200 for query calculation. This global model has been federated and trained by aggregating desensitized data from all objects in centers A, B, and C, thus containing broader intervention-effect causal knowledge. Based on the model's calculation results, the collaborative decision-making and intervention scheduling module 300 generates optimal intervention recommendations, such as "increasing the dosage of a certain antihypertensive drug by 5 mg / day," and encapsulates this into an intelligent intervention plan containing the expected blood pressure change trajectory before outputting it.

[0115] Then, collaborative decision-making to handle the conflict between human and machine intervention plans: When reviewing the intelligent intervention plan generated by the system, the clinician at Center A tended to maintain the current medication dosage but added a low-sodium diet recommendation. At this point, a conflict arose between the intelligent intervention plan and the human expert's plan. The conflict handling mechanism of the collaborative decision-making and intervention scheduling module 300 was triggered, defining the system plan as the first intervention hypothesis and the doctor's plan as the second intervention hypothesis. The collaborative decision-making and intervention scheduling module 300 invoked the personalized health digital twin of managed object A and conducted parallel simulations of the two hypotheses over a period of 7 days in a virtual environment, generating two independent expected blood pressure change trajectories. The collaborative decision-making and intervention scheduling module 300 generated a comparison report of these two trajectories in the form of a visual chart and presented it to the doctor. Through the comparison report, the doctor observed that the expected blood pressure control under the system plan was more stable and ultimately adopted the intelligent intervention plan generated by the system.

[0116] Finally: Proactive discovery and tracing of unknown group risks: Simultaneously, the Anomaly Resonance and Risk Tracing Module 400, running continuously in the background, discovered that the actual blood glucose levels of managed object B in Center B and managed object C in Center C had consistently exceeded the predicted values ​​given by their respective personalized health digital twins over the past three days, with prediction deviations continuously exceeding the dynamic thresholds set for them. The Anomaly Resonance and Risk Tracing Module 400 first marked B and C as anomalies. Subsequently, the module initiated a full-network scan and, through dynamic time warping algorithms, found that the blood glucose prediction deviation time series of B and C exhibited a high degree of pattern similarity. The module identified B and C as an anomaly resonance cluster and determined that this cluster collectively exhibited the group-wide anomaly resonance pattern of "unexpectedly elevated blood glucose."

[0117] Next, the Anomaly Resonance and Risk Source Tracing Module 400 initiated the source tracing procedure, automatically extracting and cross-analyzing the shared metadata of B and C. The analysis results showed that the only common influencing factor was that they had both recently started taking the same batch of a certain brand of nutritional supplements. Based on this, the Anomaly Resonance and Risk Source Tracing Module 400 generated a new causal hypothesis: "Taking this batch of nutritional supplements is causally associated with unexpected increases in blood sugar levels," and reported this hypothesis to the system administrator for offline verification.

[0118] Finally, based on multi-source feedback, the system undergoes self-evolution: the closed-loop self-evolution module 500 collects the feedback data generated in the above steps. First, the closed-loop self-evolution module 500 packages the complete decision-making process (including the adopted system solution, the rejected doctor's solution, the actual blood pressure results, and the simulated trajectory of the doctor's solution) into a counterfactual decision feedback data tuple. Second, after the generated causal hypothesis is verified offline, the closed-loop self-evolution module 500 injects it as new knowledge into the data.

[0119] After processing this data, the closed-loop self-evolution module 500 uses it to update the system model: counterfactual decision feedback data is used in the next round of federated learning to optimize the global causal inference model; while the newly confirmed "supplement-blood sugar" causal knowledge is directly used to modify the global model and the personalized health digital twin model of all subjects who take this batch of supplements, adding the supplement as a new risk feature to the model.

[0120] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A cloud-based multi-point scheduling intelligent health management system, characterized in that: include: The multimodal data acquisition and twin construction module is used to collect multimodal health data of multiple managed objects, and based on the multimodal health data of the managed objects, construct personalized health digital twins that represent the health status of the managed objects, and generate a real-time updated health status vector of the managed object. The federated causal inference cloud brain module is used to train a global causal inference model based on the health status vectors and corresponding intervention data of multiple nodes, while ensuring the data privacy of each node; The collaborative decision-making and intervention scheduling module is used to respond to the health status vector of a specific management object and generate an intelligent intervention plan for the management object based on the global causal inference model. The abnormal resonance and risk tracing module is used to identify the abnormal state of an individual by continuously comparing the real health status vector and the predicted health status vector of the personalized health digital twin, and to scan the personalized health digital twins of multiple managed objects to identify abnormal resonance patterns in groups and trace the source of unknown risk factors. The closed-loop self-evolution module is used to iteratively update the global causal inference model and the personalized health digital twin based on the execution results of the intelligent intervention plan and the traced unknown risk factors.

2. The cloud-based multi-point scheduling intelligent health management system according to claim 1, characterized in that, The steps in the multimodal data acquisition and twin construction module for constructing a personalized health digital twin representing the health status of the managed object include: The collected multimodal health data are fused into a unified multidimensional health state vector; Based on the health status vector, a dynamic state transition model is instantiated for the managed object as a personalized health digital twin of the managed object. The dynamic state transition model is used to calculate and output the predicted health status vector for the next moment based on the current health status vector, the externally input intervention vector, and the personalized model parameters that characterize the individual characteristics of the managed object.

3. The cloud-based multi-point scheduling intelligent health management system according to claim 1, characterized in that, The steps in training the global causal inference model using the federated learning framework in the federated causal inference cloud brain module include: The global causal inference model, either initial or from the previous iteration, is distributed from the cloud server to multiple local servers deployed on edge nodes. The local server of the edge node uses its local private health status vector and intervention data to train the received global causal inference model locally. The local server of the edge node updates and uploads the encrypted model parameters generated after local training to the cloud server; The cloud server securely aggregates the encrypted model parameter updates uploaded from all edge nodes participating in the training to update the global causal inference model, wherein the updated global causal inference model is used to characterize the causal relationship between intervention measures and health outcomes.

4. The cloud-based multi-point scheduling intelligent health management system according to claim 1, characterized in that, The steps in the collaborative decision-making and intervention scheduling module for generating an intelligent intervention plan for the managed object include: Based on the health status vector of a specific managed object, an optimal intervention recommendation for that managed object is obtained by querying the global causal inference model. The optimal intervention recommendation is packaged into a decision support report, which includes the content of the optimal intervention recommendation, the expected health trajectory generated after the personalized health digital twin of the specific management object is executed, and historical intervention cases of individuals with similar characteristics retrieved from the global knowledge base.

5. The cloud-based multi-point scheduling intelligent health management system according to claim 4, characterized in that, The collaborative decision-making and intervention scheduling module is also used for: When the intelligent intervention plan conflicts with the human expert plan input from the outside, the intelligent intervention plan and the human expert plan are defined as the first intervention hypothesis and the second intervention hypothesis, respectively. The personalized health digital twin of the specific management object is invoked to simulate and execute the first intervention hypothesis and the second intervention hypothesis in parallel in a virtual environment, so as to generate two independent future health status prediction trajectories respectively; Generates a visual comparison report containing two independent predictive trajectories of future health status to assist in final clinical decision-making.

6. The cloud-based multi-point scheduling intelligent health management system according to claim 1, characterized in that, The steps for identifying collective abnormal resonance patterns in the abnormal resonance and risk tracing module include: Calculate the prediction deviation value between the actual health status vector and the predicted health status vector, and when the prediction deviation value is continuously higher than the dynamic threshold set for the managed object, mark the personalized health digital twin of the managed object as abnormal; When a personalized health digital twin is marked as abnormal, a scan of the personalized health digital twins of all managed objects is initiated to identify other personalized health digital twins that have similar prediction bias time series patterns in different physiological dimensions to the abnormality. All identified personalized health digital twins are then formed into an abnormal resonance cluster, and the deviation time series pattern commonly manifested by the abnormal resonance cluster is determined as the group-wide abnormal resonance pattern.

7. The cloud-based multi-point scheduling intelligent health management system according to claim 6, characterized in that, The abnormal resonance and risk tracing module uses a time series similarity matching algorithm to calculate the similarity between the prediction deviation time series of different personalized health digital twins and to identify the similar prediction deviation time series patterns.

8. The cloud-based multi-point scheduling intelligent health management system according to claim 6, characterized in that, The steps for tracing unknown risk factors in the abnormal resonance and risk tracing module include: Extract the shared metadata of the management objects corresponding to all personalized health digital twins within the abnormal resonance cluster; Cross-analysis of the shared metadata was performed to identify common influencing factors; Based on the common influencing factors, a new causal hypothesis is generated regarding the unknown risk factors.

9. The cloud-based multi-point scheduling intelligent health management system according to claim 1, characterized in that, The steps for iteratively updating the global causal inference model and the personalized health digital twin in the closed-loop self-evolution module include: The execution results of the intelligent intervention plan are collected as routine intervention feedback data; Collect decision-making process data from the collaborative decision-making and intervention scheduling module when handling human-machine conflict as counterfactual decision feedback data; Collect the unknown risk factors generated by the abnormal resonance and risk tracing module and verified as new knowledge injection data; Based on the conventional intervention feedback data, the counterfactual decision feedback data, and the new knowledge injection data, the global causal inference model and the personalized health digital twin are updated respectively.

10. The cloud-based multi-point scheduling intelligent health management system according to claim 9, characterized in that, The counterfactual decision feedback data specifically includes: the intervention plan that was ultimately adopted, the actual health results after the implementation of the plan, the intervention plan that was rejected, and the predicted health trajectory obtained after virtual simulation of the rejected plan.

Images