Risk early warning method and device, electronic equipment and storage medium

Abstract

The invention discloses a risk early warning method and device, electronic equipment and a storage medium, and relates to the technical field of risk early warning. According to the risk early warning method and device, personalized learning is carried out based on a medical knowledge base in combination with multi-modal semantic embedding, and a user exclusive causal map is generated to consider individual differences; the model is updated in real time, the health state posterior probability is accurately calculated through a filtering algorithm, and misjudgment is reduced; a target event probability and a high-probability causal path are deduced, a hierarchical intervention instruction and an active disturbance calibration mechanism are matched, a user feedback dynamic optimization map is combined, and medical rationality and personalized adaptation are guaranteed, so that the problems that individual differences are ignored, high false alarms and missing alarms are likely to be generated, and the accuracy and the reliability of medical analysis are poor in existing threshold judgment or statistical correlation analysis can be solved. According to the technical scheme of the invention, the technical problems in the prior art are solved, and the technical effects of reducing the false report and missing report rate of health risks, enhancing the individual suitability, maintaining the medical rationality, and improving the continuous effectiveness of the home health service and the user credibility are achieved.

Description

Technical Field

[0001] This disclosure relates to the field of risk warning technology, and in particular to a risk warning method and apparatus, electronic device and storage medium. Background Technology

[0002] With the increasing global trend of population aging, home-based elderly care, as an important technological direction in the field of smart health, is widely used in scenarios of health monitoring and risk prevention for the elderly. Among related technologies, a multimodal data acquisition and analysis system has been constructed through the collaborative operation of wearable devices, environmental sensors, and visual monitoring systems. Specifically, this technology covers the entire process from physiological indicator monitoring and behavioral pattern recognition to health risk assessment, including key aspects such as data fusion, statistical correlation analysis, and threshold-based early warning. Based on this, existing systems mainly rely on fixed threshold judgments or traditional machine learning models for risk identification. Threshold methods trigger alarms by setting preset physiological indicator thresholds, while machine learning models focus on mining statistical correlations between data, but have not yet broken through the modeling and deduction of the intrinsic causal mechanisms of health states.

[0003] This method, which uses threshold judgment or statistical correlation analysis, ignores individual differences and is prone to high false alarm or false negative rates. This makes it difficult for the system to maintain medical rationality and personalized adaptation capabilities in long-term operation, ultimately affecting the continued effectiveness of home-based health care services and user trust. Summary of the Invention

[0004] This disclosure provides a risk warning method, device, electronic device, and storage medium. Its main purpose is to address issues affecting the continued effectiveness and user trust in home-based elderly care services.

[0005] According to a first aspect of this disclosure, a risk warning method is provided, comprising: The model topology is initialized based on a medical knowledge base, and the model is then trained in a personalized manner using multimodal semantic embedding vector sequences to generate a user-specific causal graph. The tree model state is updated in real time. The current posterior probability distribution of each health state latent variable is recursively calculated by combining the current semantic embedding vector with the posterior probability distribution of the previous time step using a filtering algorithm. Based on the updated model state, a forward simulation is performed to extrapolate the probability of the target event occurring within a preset time window and to identify the high-probability causal path leading to the event. Based on risk probability and causal path, hierarchical intervention instructions are generated. Causal relationships are experimentally calibrated through an active perturbation verification mechanism, and the structure or parameters of the causal graph are adjusted based on user feedback.

[0006] Optionally, the step of initializing the model topology based on a medical knowledge base and performing personalized learning on the model through multimodal semantic embedding vector sequences to generate a user-specific causal graph includes: The model topology is initialized based on a medical knowledge base and formally defined using a dynamic Bayesian network. The set of nodes represents latent variables of health states that cannot be directly observed, and the set of directed edges represents the conditional dependencies between variables, including instantaneous dependencies at the same point in time and temporal dependencies across time points. Structure learning is performed using a fraction-based causal discovery algorithm, with the objective function being to maximize the log marginal likelihood of user data. The expected value maximization algorithm is then used for parameter learning to obtain the optimal parameters of the conditional probability distribution from the user data.

[0007] Optionally, the real-time update of the tree model state, using a filtering algorithm to combine the current semantic embedding vector with the posterior probability distribution of the previous time step, recursively calculates the current posterior probability distribution of each health state latent variable, including: The particle filter algorithm is used as the filtering algorithm, and the posterior probability distribution of each latent variable at the current time is calculated recursively. The semantic embedding vector of the new observation is weighted and fused with the posterior probability distribution of the previous time step to generate the health status estimate at the current time step.

[0008] Optionally, the step of performing forward simulation based on the updated model state to extrapolate the probability of the target event occurring within a preset time window and identifying high-probability causal paths leading to the event includes: Multiple future state evolution paths are generated using a forward sampling simulation method, and the proportion of sequences in all paths where the target event occurs at least once is used as the risk probability. Statistical analysis is used to identify high-probability causal paths leading to the target event, where the causal path is the state transition sequence that occurs most frequently.

[0009] Optionally, the method further includes: An online causal discovery algorithm is used to continuously update the dynamic causal digital twin model; wherein, the online causal discovery algorithm includes a time-series-based streaming version of causal testing or an incremental causal structure learning algorithm; The model's graph structure is dynamically adjusted through an incremental structure learning algorithm. When a change in the distribution of user health status is detected, the causal relationship that needs to be corrected is automatically identified and the connection relationship of the corresponding node is updated.

[0010] According to a second aspect of this disclosure, a risk warning device is provided, comprising: The generation unit is used to initialize the model topology based on the medical knowledge base and to perform personalized learning on the model through multimodal semantic embedding vector sequences to generate a user-specific causal graph. The computational unit is used to update the tree model state in real time. It uses a filtering algorithm to combine the current semantic embedding vector with the posterior probability distribution of the previous time step to recursively calculate the current posterior probability distribution of each health state latent variable. The simulation unit is used to perform forward simulation based on the updated model state, extrapolate the probability of the occurrence of the target event within a preset time window, and identify the high-probability causal path that leads to the event. The calibration unit is used to generate graded intervention instructions based on risk probability and causal path, conduct experimental calibration of causal relationships through an active perturbation verification mechanism, and adjust the structure or parameters of the causal graph based on user feedback.

[0011] Optionally, the generation unit is further configured to: The model topology is initialized based on a medical knowledge base and formally defined using a dynamic Bayesian network. The set of nodes represents latent variables of health states that cannot be directly observed, and the set of directed edges represents the conditional dependencies between variables, including instantaneous dependencies at the same point in time and temporal dependencies across time points. Structure learning is performed using a fraction-based causal discovery algorithm, with the objective function being to maximize the log marginal likelihood of user data. The expected value maximization algorithm is then used for parameter learning to obtain the optimal parameters of the conditional probability distribution from the user data.

[0012] Optionally, the computing unit is further configured to: The particle filter algorithm is used as the filtering algorithm, and the posterior probability distribution of each latent variable at the current time is calculated recursively. The semantic embedding vector of the new observation is weighted and fused with the posterior probability distribution of the previous time step to generate the health status estimate at the current time step.

[0013] Optionally, the simulation unit is further configured to: Multiple future state evolution paths are generated using a forward sampling simulation device, and the proportion of sequences in all paths where the target event occurs at least once is used as the risk probability. Statistical analysis is used to identify high-probability causal paths leading to the target event, where the causal path is the state transition sequence that occurs most frequently.

[0014] Optionally, the device further includes: An update unit is used to continuously update the dynamic causal digital twin model using an online causal discovery algorithm; wherein, the online causal discovery algorithm includes a time-series-based streaming version of causal testing or an incremental causal structure learning algorithm; The detection unit is used to dynamically adjust the graph structure of the model through an incremental structure learning algorithm. When a change in the distribution of user health status is detected, it automatically identifies the causal relationship that needs to be corrected and updates the connection relationship of the corresponding node.

[0015] According to a third aspect of this disclosure, an electronic device is provided, comprising: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method described in the first aspect above.

[0016] According to a fourth aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are configured to cause the computer to perform the method described in the first aspect above.

[0017] According to a fifth aspect of this disclosure, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the method described in the first aspect above.

[0018] The risk warning method, device, electronic device, and storage medium disclosed herein, through this application, can generate user-specific causal graphs by performing personalized learning based on a medical knowledge base and multimodal semantic embedding to take into account individual differences; update the model in real time and accurately calculate the posterior probability of health status through filtering algorithms to reduce misjudgments; infer the probability of target events and high-probability causal paths, and combine them with graded intervention instructions and active perturbation calibration mechanisms, dynamically optimizing the graph based on user feedback, ensuring medical rationality and personalized adaptation. Therefore, it can solve the technical problems of existing threshold judgment or statistical association analysis ignoring individual differences, easily generating high false positives and false negatives, and making it difficult to maintain medical rationality and personalized adaptation, thus affecting the continuous effectiveness of home-based health care services and user trust. It achieves the technical effects of reducing the false positive and false negative rates of health risks, enhancing individual adaptability, maintaining medical rationality, and improving the continuous effectiveness of home-based health care services and user trust.

[0019] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0020] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein: Figure 1 A schematic flowchart illustrating a risk warning method provided in an embodiment of this disclosure; Figure 2 A structural block diagram of a risk early warning system provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a multimodal semantic perception module provided in an embodiment of this application; Figure 4 A schematic diagram of the processing flow of a dynamic causal digital twin module provided in an embodiment of this application; Figure 5 A flowchart illustrating a forward-looking risk assessment module provided in an embodiment of this application; Figure 6 A schematic diagram of the processing flow of a closed-loop self-evolutionary intervention module provided in an embodiment of this application; Figure 7 A schematic diagram of the structure of a risk warning device provided in an embodiment of this disclosure; Figure 8 A schematic diagram of another risk warning device provided in an embodiment of this disclosure; Figure 9 A schematic block diagram of an example electronic device provided for embodiments of this disclosure. Detailed Implementation

[0021] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0022] The risk warning method, apparatus, electronic device, and storage medium of embodiments of this disclosure are described below with reference to the accompanying drawings.

[0023] Figure 1 This is a flowchart illustrating a risk warning method provided in an embodiment of the present disclosure.

[0024] like Figure 1 As shown, the method includes the following steps: Step 101: Initialize the model topology based on the medical knowledge base, and perform personalized learning on the model through multimodal semantic embedding vector sequences to generate a user-specific causal graph; The model topology was initialized based on a medical knowledge base, which includes general medical knowledge and causal relationships between diseases and health-influencing factors. The initialized model topology graphically presents health-related latent variables and their initial dependencies. Latent variables may include indicators that have a key impact on the user's health status but cannot be directly observed, such as sleep quality and heart rate fluctuations. Dependencies include both instantaneous associations between different latent variables at the same time dimension and cross-temporal effects of latent variables at different time dimensions. Subsequently, a multimodal semantic embedding vector sequence was introduced to conduct personalized model learning. This multimodal semantic embedding vector sequence was obtained by processing multi-source heterogeneous data from the home environment. This data comes from gait parameters and body posture collected by visual sensors, temperature, humidity, and light intensity collected by environmental sensors, and heart rate and heart rate variability collected by physiological sensors. These data are converted into vector sequences with unified semantic representations through a unified embedding function, which can comprehensively and accurately represent the user's real-time status in different dimensions.

[0025] In the personalized learning process, based on the initial model topology and multimodal semantic embedding vector sequence, the optimal graph structure that fits the current user is first searched in the initial structure space using a structure learning algorithm. Then, the conditional probability distribution parameters of the model are optimized from the user's exclusive data using a parameter learning algorithm. Finally, an exclusive causal graph that can accurately reflect the internal causal mechanism of the user's health status is generated, providing a core basis for the subsequent dynamic tracking of changes in the user's health status and the conduct of risk assessment.

[0026] Step 102: Update the tree model state in real time. Use a filtering algorithm to combine the current semantic embedding vector with the posterior probability distribution of the previous time step to recursively calculate the current posterior probability distribution of each health state latent variable. The core of defining the tree model state to be updated is the set of latent variables representing the user's health status and the relationships between these variables. These latent variables encompass indicators that are crucial for health and wellness risk assessment but cannot be directly observed, such as sleep quality, blood pressure fluctuations, and daytime activity levels. The real-time nature of the tree model state directly determines the reliability of subsequent risk analysis. During the update process, two core inputs are required: first, the current semantic embedding vector, generated by processing various sensor data in the home environment through multimodal perception, comprehensively reflecting the user's current physiological state, behavioral characteristics, and environmental influences; second, the posterior probability distribution of each latent health status variable from the previous time step, which is a quantitative estimate of the user's health status in the previous cycle, providing historical evidence for the current update.

[0027] Subsequently, a filtering algorithm is employed to fuse the two types of inputs. Based on the currently observed semantic embedding vector, the algorithm dynamically adjusts and recursively calculates the posterior probability distribution from the previous time step, eliminating noise interference and integrating new observation information to ultimately obtain the current posterior probability distribution of each latent health variable. This distribution accurately quantifies the possible values ​​and corresponding probabilities of each latent variable at the current time step, clearly presenting the intrinsic characteristics of the user's current health status. This provides accurate status input for the subsequent prospective risk assessment module, ensuring that risk prediction is based on the latest and most reliable health status data.

[0028] Step 103: Perform forward simulation based on the updated model state to extrapolate the probability of the target event occurring within a preset time window and identify the high-probability causal path leading to the event; Analysis is conducted based on the updated model state, which represents the current posterior probability distribution of each latent variable of health status. This state accurately quantifies the probability of key latent variables such as sleep quality, heart rate stability, and activity level at the current moment, providing a reliable initial state basis for subsequent extrapolations. During forward simulation, the system, based on the aforementioned current state distribution, uses forward sampling simulation to repeatedly extrapolate the evolution path of health status within a preset time window. The preset time window can be flexibly set according to the characteristics of the target event; for example, it can be set to 12 hours for the risk of falls or 72 hours for the risk of chronic disease exacerbation. In each extrapolation, the system extracts specific state instances from the current state distribution and, combined with the state transition rules in the model, successively generates a complete state sequence from the current moment to the end of the time window.

[0029] After obtaining a large number of simulated future state sequences, the system calculates the percentage of sequences in which the target event occurs at least once. This percentage determines the probability of the target event occurring within a preset time window. Target events include adverse health events closely related to user well-being, such as falls and abnormal cardiovascular indicators. Simultaneously, the system tracks and analyzes all state evolution paths leading to the target event. By statistically analyzing the frequency of different paths, it identifies the one or more paths that occur most frequently—the high-probability causal paths leading to the target event. These paths clearly reveal the key influencing relationships between latent variables; for example, environmental noise causing sleep disruption, leading to heart rate fluctuations, and ultimately increasing the risk of cardiovascular events. The probability of the target event and the high-probability causal paths output in this step provide direct decision-making basis for subsequent targeted intervention measures, ensuring that interventions accurately target the root causes of risk.

[0030] Step 104: Generate graded intervention instructions based on risk probability and causal path, conduct experimental calibration of causal relationship through active perturbation verification mechanism, and adjust the structure or parameters of causal graph based on user feedback information.

[0031] Based on risk probability and causal path, the system generates tiered intervention instructions. Risk probability is the probability of the occurrence of the target adverse health event within a preset time window, while causal path is the link of influence of high-probability latent variables that lead to the event. The system internally sets risk thresholds to classify different risk levels. Based on the risk probability level and the specific content of the causal path, corresponding instructions are generated. For low-risk levels, non-invasive environmental adjustment instructions are generated; for medium-risk levels, contextualized interactive instructions based on causal paths are generated; and for high-risk levels, alarm instructions containing the causal path of the monitored phenomenon and risk quantification information are generated.

[0032] Subsequently, the causal relationship is calibrated through an active perturbation verification mechanism. The system first identifies the causal associations in the model whose confidence level needs to be verified, generates an active perturbation command that can apply small and controllable changes to the relevant states of the corresponding cause variables in reality, executes the command through the smart home device interface, observes the state changes of the result variables through the multimodal semantic perception module, and integrates the perturbation operation and the observed state change results into feedback information, which is used to adjust the conditional probability parameters of the corresponding causal relationship in the model.

[0033] The system adjusts the causal graph based on user feedback. It receives user feedback on the intervention effect or causal explanation through an interactive interface. After parsing the feedback, if it finds that the feedback points to confounding factors of health impact not included in the model, or indicates that there is a bias in the existing causal association, it generates an adjustment instruction to update the number of nodes or the dependency parameters between nodes in the causal graph. This ensures that the causal graph can continuously reflect the user's actual health status and provide more accurate model support for subsequent risk assessment.

[0034] In some embodiments, the step of initializing the model topology based on a medical knowledge base and performing personalized learning on the model through multimodal semantic embedding vector sequences to generate a user-specific causal graph includes: The model topology is initialized based on a medical knowledge base and formally defined using a dynamic Bayesian network. The set of nodes represents latent variables of health states that cannot be directly observed, and the set of directed edges represents the conditional dependencies between variables, including instantaneous dependencies at the same point in time and temporal dependencies across time points. Structure learning is performed using a fraction-based causal discovery algorithm, with the objective function being to maximize the log marginal likelihood of user data. The expected value maximization algorithm is then used for parameter learning to obtain the optimal parameters of the conditional probability distribution from the user data.

[0035] This feature is used to refine the specific implementation process of model building and personalized adaptation, ensuring that the generated causal graph conforms to both medical logic and individual user characteristics. First, the model topology is initialized based on a medical knowledge base. This knowledge base covers general medical common sense, established causal relationships between diseases and health influencing factors, etc., providing a medically sound framework for initialization. The initialized model is formally defined using a dynamic Bayesian network. The node set represents latent health variables that play a key role in the user's health status and are not directly observable, such as sleep quality, blood pressure stability, and daytime activity level. The directed edge set is used to characterize the conditional dependencies between variables. These dependencies have two dimensions: one is the instantaneous dependency between different latent variables at the same point in time, such as the interaction between heart rate and blood oxygen saturation at a certain moment; the other is the temporal dependency between latent variables at different points in time, such as the impact of the previous day's sleep quality on the current day's mental state.

[0036] In the personalized learning phase of the model, structure learning is first performed using a score-based causal discovery algorithm. This algorithm uses maximizing the log-marginal likelihood of user data as its core objective function, while introducing a model complexity penalty term to avoid overfitting. It searches for the optimal graph structure that best suits the current user within a space with an initial topological structure as a priori. After the structure is determined, parameter learning is performed using an expectation-maximization algorithm. Effective information is extracted from the user's multimodal semantic embedding vector sequence, and the conditional probability distribution among variables in the model is continuously optimized through iterative calculations. Finally, the optimal parameters of the conditional probability distribution are obtained. Combining the optimized structure and parameters, a personalized causal graph that accurately reflects the intrinsic causal mechanism of the user's health status is generated.

[0037] In some embodiments, the real-time update of the tree model state, which utilizes a filtering algorithm to combine the current semantic embedding vector with the posterior probability distribution of the previous time step, and recursively calculates the current posterior probability distribution of each health state latent variable, includes: The particle filter algorithm is used as the filtering algorithm, and the posterior probability distribution of each latent variable at the current time is calculated recursively. The semantic embedding vector of the new observation is weighted and fused with the posterior probability distribution of the previous time step to generate the health status estimate at the current time step.

[0038] This paper refines the specific implementation method of real-time state updates in the tree model to ensure the accuracy and efficiency of calculating the current posterior probability distribution of latent variables for each health state. Firstly, the particle filter algorithm is chosen as the core filtering algorithm. This algorithm has the advantage of handling the state evolution of nonlinear and non-Gaussian systems, and can well adapt to the characteristics of health state latent variables changing dynamically over time and being affected by multiple factors, providing reliable algorithmic support for recursive calculations.

[0039] During real-time updates, the posterior probability distribution of the latent variables of each health status at the previous time step is used as the initial basis. This distribution is a quantitative estimate of the user's health status in the previous period, including the probability information of latent variables such as sleep quality and heart rate fluctuations. The particle filter algorithm first transforms the posterior probability distribution of the previous time step into a large set of particles, each particle representing a possible combination of latent variables. Then, combined with the state transition rules in the model, the state of the latent variables corresponding to each particle is predicted to obtain the prior state distribution at the current time step.

[0040] Subsequently, the current semantic embedding vector of the new observation is introduced. This vector is generated by the multimodal semantic perception module through processing real-time sensor data, covering multi-dimensional state information such as the user's current physiological behavior and environment.

[0041] The algorithm calculates the matching degree between the predicted state of each particle and the current semantic embedding vector through an importance sampling step, and assigns weights to the particles accordingly, achieving a weighted fusion of the new observed semantic embedding vector and the posterior probability distribution of the previous time step. Finally, the weighted particle set is resampled and statistically analyzed to remove invalid particles with low weights and retain effective particles that accurately reflect the current state. By integrating the state information of the effective particles, the current posterior probability distribution of each health state latent variable is recursively calculated. This distribution can accurately characterize the current probability of the latent variable, providing real-time and reliable state data for subsequent prospective risk assessment.

[0042] In some embodiments, the step of performing forward simulation based on the updated model state to extrapolate the probability of the target event occurring within a preset time window and identifying high-probability causal paths leading to the event includes: Multiple future state evolution paths are generated using a forward sampling simulation method, and the proportion of sequences in all paths where the target event occurs at least once is used as the risk probability. Statistical analysis is used to identify high-probability causal paths leading to the target event, where the causal path is the state transition sequence that occurs most frequently.

[0043] Forward sampling simulation is performed based on the updated model state, which represents the current posterior probability distribution of each latent variable in the health state. This distribution accurately quantifies the current probability of key latent variables such as sleep quality, heart rate stability, and activity level, providing a reliable initial basis for the forward simulation. During the forward sampling simulation, the system samples different latent variable state instances multiple times from the current posterior probability distribution. Combining this with the model's built-in state transition rules, it recursively generates a complete state evolution path extending from the current moment to the end of a preset time window. By repeating this sampling and recursion operation multiple times, multiple future state evolution paths covering different possibilities are obtained.

[0044] After acquiring a sufficient number of evolutionary paths, the system filters and statistically analyzes all paths to identify those where the target event occurs at least once. Target events include adverse health events closely related to home-based health care, such as falls and abnormal cardiovascular indicators. The ratio of the number of these types of paths to the total number of paths is used as the probability of the target event occurring within a preset time window, i.e., the risk probability. This statistical method effectively reflects the likelihood of the target event occurring. Simultaneously, the system performs statistical analysis on all state evolutionary paths leading to the target event, tracking the state transition process of latent variables in each path, recording the frequency of different transition sequences, and filtering out the state transition sequences with the highest frequency. These sequences are the high-probability causal paths leading to the target event; for example, environmental noise causing sleep disruption, leading to heart rate fluctuations, and ultimately increasing the risk of cardiovascular events. The risk probability and high-probability causal paths obtained through this process provide a precise decision-making basis for the generation of subsequent tiered intervention instructions, ensuring that intervention measures accurately target the root cause of risk.

[0045] In some embodiments, the method further includes: An online causal discovery algorithm is used to continuously update the dynamic causal digital twin model; wherein, the online causal discovery algorithm includes a time-series-based streaming version of causal testing or an incremental causal structure learning algorithm; The model's graph structure is dynamically adjusted through an incremental structure learning algorithm. When a change in the distribution of user health status is detected, the causal relationship that needs to be corrected is automatically identified and the connection relationship of the corresponding node is updated.

[0046] It is clearly stipulated that online causal discovery algorithms should be used to continuously update the dynamic causal digital twin model. The dynamic causal digital twin model is a core model previously used to characterize latent variables of user health status and personalized causal relationships between variables. Its continuous updates ensure the accuracy of subsequent risk assessments and interventions. The online causal discovery algorithms employed include two key types: one is a streaming version of time-series causal testing, which can process time-series data continuously collected by the multimodal semantic perception module in real time, dynamically capturing the changing trends of potential causal relationships in the data, such as real-time monitoring of the strength fluctuations of the causal association between heart rate time-series data and sleep stage time-series data; the other is an incremental causal structure learning algorithm, which can process newly generated user data in a progressively accumulative manner, avoiding repeated calculations of historical data and improving model update efficiency.

[0047] During model updates, the focus is on dynamically adjusting the model's graph structure using an incremental structural learning algorithm. The system continuously monitors changes in the distribution of users' health status, specifically by comparing the distribution characteristics of multimodal semantic embedding vector sequences across different time periods. When a significant difference is detected, it indicates a change in the distribution of users' health status, such as a noticeable shift in the distribution of physiological indicators after a user starts taking a new medication. At this point, the incremental structural learning algorithm automatically analyzes the changes in causal relationships behind the data changes, identifying causal relationships in the model that need correction. This could be due to a weakening or disappearance of existing causal relationships, or the emergence of new causal relationships. Subsequently, the algorithm updates the connection relationships of corresponding nodes in the model's graph structure based on the identification results. For example, for a newly emerging causal relationship of "drug effect - blood pressure fluctuation," a new node representing "drug effect" is added to the graph structure, and a directed connection is established between it and the "blood pressure fluctuation" node. For a weakened "sleep quality - daytime activity" relationship, the weights or connection states of the edges between corresponding nodes are adjusted. This achieves continuous optimization of the dynamic causal digital twin model, ensuring that the model always accurately reflects the causal mechanism of the user's current health status.

[0048] The following example illustrates a risk warning method provided by an embodiment of this application.

[0049] Please see Figure 2 , Figure 2 This is a structural block diagram of a risk early warning system provided in an embodiment of this application. The system includes: a multimodal semantic perception module, a dynamic causal digital twin module, a forward-looking risk assessment module, and a closed-loop self-evolutionary intervention module.

[0050] The output of the multimodal semantic perception module is connected to the first input of the dynamic causal digital twin module. The output of the dynamic causal digital twin module is connected to the input of the prospective risk assessment module. The output of the prospective risk assessment module is connected to the input of the closed-loop self-evolutionary intervention module. The first output of the closed-loop self-evolutionary intervention module is used to output intervention instructions, and its second output is connected to the second input of the dynamic causal digital twin module, forming a closed-loop path for model adjustment.

[0051] In a specific embodiment, the workflow of the home-based elderly care risk early warning system provided by the present invention may include the following steps: The multimodal semantic awareness module continuously collects multi-source heterogeneous data from the user's environment and transmits the data stream through a unified multimodal embedding function. Processing is performed to transform it into a unified semantic space. semantic embedding vectors in This process can be formally represented as:

[0052] in, The raw data collected at time point t, Let k be the modality type of the data, and k be the dimension of the semantic space. This semantic embedding vector... It is used as real-time evidence representing the user's current state and is transmitted to the dynamic causal digital twin module.

[0053] The dynamic causal digital twin module receives real-time evidence from the multimodal semantic perception module. Based on this evidence, the state of its internal dynamic causal digital twin model G is updated using a filtering algorithm. The model G is used to characterize latent variables representing the user's health status. A probabilistic graphical model of the personalized causal relationships between the latent variables. This state update process aims to compute the posterior probability distribution of each latent variable, given all evidence up to the present. .

[0054] The forward-looking risk assessment module is based on the updated model status output by the dynamic causal digital twin module. Preset time window for the future The probability of occurrence of a specific adverse health event A within the body The module then performs a deduction. Simultaneously, it identifies high-probability causal paths in model G that lead from the current state to the adverse health event A. Ultimately, this module generates a value containing risk probabilities. With causal path The risk assessment results are then transmitted to the closed-loop self-evolutionary intervention module.

[0055] The closed-loop self-evolutionary intervention module receives the risk assessment results and, based on the risk probability... Quantitative values ​​and causal pathways The system generates and executes corresponding tiered intervention instructions based on the specific content of the instructions. These intervention instructions include, but are not limited to, environmental adjustment instructions, voice interaction instructions, or causal narrative alarms sent to designated receiving terminals.

[0056] The closed-loop self-evolutionary intervention module further generates feedback instructions for adjusting the dynamic causal digital twin model G based on the execution results of the hierarchical intervention instructions or user feedback information received from the external interaction interface. And the feedback instruction The data is transmitted to the second input terminal of the dynamic causal digital twin module. Based on this feedback instruction, the dynamic causal digital twin module adjusts the structure or parameters of its internal model G, thereby completing one cycle of closed-loop self-evolution of the model.

[0057] Through the cyclical execution of the above steps, this system constructs a complete closed loop from data perception, causal modeling, risk prediction to intervention feedback. This structure enables the causal model within the system to continuously self-correct and optimize based on actual conditions, adapting to the dynamic changes in the user's health status.

[0058] Please see Figure 3 , Figure 3 This is a schematic diagram of the structure of a multimodal semantic perception module provided in an embodiment of this application. The function of this module is to process multi-source, heterogeneous raw data collected by the system from the external environment into structured information with a unified representation format usable by subsequent modules. The working process of this module may include the following steps: The data acquisition module continuously collects time-series data reflecting user status and environmental information through a series of sensor devices deployed in the user's home environment. The data sources include: visual sensors for collecting gait parameters, body posture, body movement information during sleep, and respiratory rate; environmental sensors for collecting ambient temperature, humidity, light intensity, and air quality index; physiological indicator sensors for collecting heart rate, heart rate variability, and sleep stage data; and behavioral interaction sensors for collecting usage records of smart home appliances, door and window opening / closing status, and voice interaction data recorded by smart voice devices. All collected raw data... All include a timestamp and its corresponding modal type .

[0059] Unified semantic embedding: This module integrates data from different modalities collected during the data acquisition process. Through a unified multimodal embedding function Mapped to In a shared semantic space S of dimension , semantic embedding vectors are generated. .

[0060]

[0061] in, The goal of this mapping is to ensure that different modalities of information that are intrinsically related in the sense of health assessment have similar vector representations in the semantic space S. For example, the semantic embedding vectors generated by "slow gait" data collected through visual sensors and "negative emotion vocabulary" data collected through voice devices have relatively small Euclidean distances or other metric distances in the semantic space S.

[0062] In a specific embodiment, the multimodal embedding function This is achieved through a pre-trained neural network model. The model contains a set of modality-specific encoders and a shared projection head. For the input data... The system first determines its modal type The corresponding encoder is selected for feature extraction to generate an intermediate feature representation. This intermediate feature representation is then input into a shared projection head, ultimately outputting a K-dimensional semantic embedding vector. .

[0063] To achieve the aforementioned mapping objective, the neural network model can be trained using a contrastive learning method. During training, the system constructs positive and negative sample pairs. Positive sample pairs consist of data from different modalities that are temporally synchronized or logically strongly correlated, while negative sample pairs consist of unrelated data. The training objective of the model is to minimize the loss function, which drives the model to bring positive sample pairs closer together in the semantic space S, while simultaneously pushing negative sample pairs further apart.

[0064] Through the above steps, the multimodal semantic perception module transforms the raw, heterogeneous sensor signal stream into a standardized, high-information-density semantic embedding vector sequence. This vector sequence is then used as real-time evidence and fed into the dynamic causal digital twin module for subsequent state updates and causal inferences.

[0065] Please participate Figure 4 , Figure 4 This is a schematic diagram illustrating the processing flow of a dynamic causal digital twin module provided in an embodiment of this application. The module's function is to construct, personalize, and maintain in real-time a probabilistic graphical model that characterizes the inherent causal relationships of a user's health status. The module's operation may include the following steps: In this embodiment, the dynamic causal digital twin model G is formally defined and initialized using a dynamic Bass network. This network G consists of a node set V and a directed edge set E. The node set V represents a set of unobservable latent health state variables that have a decisive influence on the user's health status at a given time point. The state vector is represented as Where N is the total number of latent variables, For the i-th latent variable at time point The state. The directed edge set E represents the conditional dependencies between variables, including the same time point. The instantaneous dependencies between internal variables, and the intertemporal dependencies between the state at the previous time point t-1 and the state at the current time point t. The joint probability distribution of this dynamic Bass network follows the formula:

[0066] in, This represents the state sequence from time point 1 to T. This represents the corresponding semantic embedding vector sequence. It is a state transition model, and its specific form is as follows: ,in Represents a node The set of parent nodes. It is an observational model, representing the state of health given the current health status. Under these conditions, semantic embeddings were observed. The probability of this. The initial topology of the model. The prior distribution of parameters is pre-constructed based on a general medical knowledge base, providing a medically rational foundation for subsequent personalized learning.

[0067] In the personalized learning process of the model, to adapt the general model to the individual characteristics of a specific user, this module utilizes the time series of semantic embedding vectors belonging to that user, received from the multimodal semantic perception module. The initial model is then customized to generate a causal digital twin model tailored to each user. The process first performs structure learning, using a score-based causal discovery algorithm, in order to... The search is performed in the prior structure space to determine the optimal graph structure. The objective function for this optimization is:

[0068] in, It is the log-marginal likelihood of user data given a graph structure. This is a term used to penalize model complexity. It is the regularization coefficient. After determining the user-specific graph structure... Then, the process further performs parameter learning. The Expectation-Maximization (EM) algorithm is used to learn parameters from user data. Optimal parameters of each conditional probability distribution (CPD) in the learning model .

[0069] The model's real-time state update mechanism ensures that this module receives the latest semantic embedding vector at every time point during continuous system operation. This module utilizes filtering algorithms, such as particle filtering, combined with new observational evidence. The posterior probability distribution of each latent variable at the current time is recursively calculated using the posterior distribution of the state at the previous time step. This posterior probability distribution is considered the most accurate estimate of the user's current health status and is output to the prospective risk assessment module for subsequent risk extrapolation.

[0070] See attached document Figure 5 , Figure 5 This is a flowchart illustrating a forward-looking risk assessment module provided in an embodiment of this application. The module's function is to quantitatively predict potential future risks based on an estimate of the user's current health status and identify the key causes leading to those risks. The module's operation may include the following steps: The module receives status input, and its input is connected to the output of the dynamic causal digital twin module for use at each time point. Receive the posterior probability distribution representing the user's current health status. This probability distribution is a quantitative estimate of the current state of all latent variables in the dynamic causal digital twin model.

[0071] This module performs forward simulation and risk quantification. Based on the received current state distribution, it uses forward sampling simulation to repeatedly predict the state evolution path within a preset time window. In each prediction, the system samples specific state instances from the current state distribution and utilizes the state transition model in the model. Generate a path from... The future state sequence.

[0072] This module calculates a specific adverse health event based on a large number of extrapolated future state sequences. The probability of it occurring within this time window is defined as the risk value. The calculation method is as follows:

[0073] in, Indicates an event at a future point in time. It occurs. In practical implementation, this risk value... The approximation is obtained by statistically analyzing the proportion of sequences in which event A occurs at least once across all simulated future state sequences.

[0074] By identifying key causal paths during forward simulation, this module not only calculates risk probabilities but also records and analyzes the state evolution paths that ultimately lead to adverse health events. Through statistical analysis, the module identifies one or more state transition sequences with the highest frequency from the current state to the future event; these sequences are defined as key causal paths. These paths clarify the changes in the main latent variables that lead to increased predicted risk.

[0075] This module generates and outputs risk assessment results. It integrates the risk values ​​calculated in the forward simulation and risk quantification steps with the key causal paths identified in the key causal path identification step, into a structured risk assessment result. This risk assessment result is then output to the closed-loop self-evolutionary intervention module as the basis for its decision-making in generating intervention instructions.

[0076] Please see Figure 6 , Figure 6 This is a schematic diagram illustrating the processing flow of a closed-loop self-evolutionary intervention module provided in an embodiment of this application. The function of this module is to make decisions and interventions based on the evaluation results of upstream modules, and to drive the core model of the entire system to self-optimize through active verification and interactive feedback mechanisms. The working process of this module may include the following steps: This module receives and parses risk assessment results. Its input is connected to the output of the forward-looking risk assessment module, and it is used to receive risk assessment results that include risk probabilities and key causal paths.

[0077] This module generates and executes tiered intervention instructions, and internally sets at least two risk thresholds, including a low-risk threshold. and high risk threshold This module compares the received risk probability with the threshold and generates different levels of intervention instructions based on the comparison results: If the risk probability is less than the low-risk threshold, a level 1 intervention instruction is generated, which is a non-intrusive environmental fine-tuning instruction, such as adjusting indoor lighting or temperature through a smart home interface. If the risk probability is greater than or equal to the low-risk threshold but less than the high-risk threshold, a level 2 intervention instruction is generated, which is a contextualized interactive instruction based on the causal path, such as asking the user through a smart speaker to obtain more information, and the user's response is collected by the multimodal semantic perception module and used as new evidence. If the risk probability is greater than the high-risk threshold, a level 3 intervention instruction is generated, which is a structured causal narrative alert. This alert explicitly includes the monitored phenomenon, the causal path inferred by the model, and the quantified future risk, and is sent to the terminal device of the preset guardian or service platform through a network interface.

[0078] Performing active perturbation verification, this module periodically or under specific conditions identifies causal relationships with unverified confidence levels in dynamic causal digital twin models, such as a relationship arising from latent variables. point to To verify this relationship, the module performs the following sub-steps: First, it generates an active perturbation instruction capable of applying small, controllable changes to the state corresponding to the causal variable in the real world; this instruction is executed through the smart home device interface. Second, after the instruction is executed, the changes in the user's state related to the outcome variable are observed through the multimodal semantic perception module. Finally, the result of this perturbation-observation is encapsulated into a feedback instruction and transmitted to the dynamic causal digital twin module. Based on this feedback instruction, the dynamic causal digital twin module adjusts the conditional probability parameters in the model corresponding to this causal relationship using a Bayesian update method.

[0079] During the execution of interactive model corrections and secondary or tertiary intervention instructions, this module can provide users or guardians with counterfactual explanations based on key causal paths. The module further includes an interactive interface to receive feedback from users or guardians regarding this explanation. This module parses the received feedback. When the feedback identifies a confounding factor not included in the current model, the module generates a structural correction instruction. This instruction is transmitted to the dynamic causal digital twin module, guiding it to modify the model's graph structure, such as adding new nodes representing the confounding factor and adjusting the connections between relevant nodes based on the feedback.

[0080] Through the above steps, the closed-loop self-evolutionary intervention module not only performs the output function from early warning to intervention, but also constructs two parallel feedback paths (active perturbation verification and interactive model correction), continuously providing high-quality correction information to the dynamic causal digital twin module, thereby realizing the closed-loop self-evolution of the core model of the entire system.

[0081] Corresponding to the aforementioned risk warning method, this invention also proposes a risk warning device. Since the device embodiments of this invention correspond to the method embodiments described above, details not disclosed in the device embodiments can be referred to in the method embodiments, and will not be repeated here.

[0082] Figure 7 This is a schematic diagram of the structure of a risk warning device provided in an embodiment of this disclosure, as shown below. Figure 7 As shown, it includes: The generation unit 21 is used to initialize the model topology based on the medical knowledge base and to perform personalized learning on the model through multimodal semantic embedding vector sequences to generate a user-specific causal graph. The computing unit 22 is used to update the tree model state in real time. It uses a filtering algorithm to combine the current semantic embedding vector with the posterior probability distribution of the previous time step to recursively calculate the current posterior probability distribution of each health state latent variable. Simulation unit 23 is used to perform forward simulation based on the updated model state, extrapolate the probability of the occurrence of the target event within a preset time window, and identify the high-probability causal path that leads to the event. The calibration unit 24 is used to generate graded intervention instructions based on risk probability and causal path, conduct experimental calibration of causal relationship through active perturbation verification mechanism, and adjust the structure or parameters of causal graph based on user feedback information.

[0083] Furthermore, in one possible implementation of this disclosure embodiment, the generation unit 21 is further configured to: The model topology is initialized based on a medical knowledge base and formally defined using a dynamic Bayesian network. The set of nodes represents latent variables of health states that cannot be directly observed, and the set of directed edges represents the conditional dependencies between variables, including instantaneous dependencies at the same point in time and temporal dependencies across time points. Structure learning is performed using a fraction-based causal discovery algorithm, with the objective function being to maximize the log marginal likelihood of user data. The expected value maximization algorithm is then used for parameter learning to obtain the optimal parameters of the conditional probability distribution from the user data.

[0084] Furthermore, in one possible implementation of this disclosure, the computing unit 22 is further configured to: The particle filter algorithm is used as the filtering algorithm, and the posterior probability distribution of each latent variable at the current time is calculated recursively. The semantic embedding vector of the new observation is weighted and fused with the posterior probability distribution of the previous time step to generate the health status estimate at the current time step.

[0085] Furthermore, in one possible implementation of this disclosure, the simulation unit 23 is further configured to: Multiple future state evolution paths are generated using a forward sampling simulation device, and the proportion of sequences in all paths where the target event occurs at least once is used as the risk probability. Statistical analysis is used to identify high-probability causal paths leading to the target event, where the causal path is the state transition sequence that occurs most frequently.

[0086] Furthermore, in one possible implementation of the embodiments of this disclosure, such as Figure 8 As shown, the device further includes: The update unit 25 is used to continuously update the dynamic causal digital twin model using an online causal discovery algorithm; wherein, the online causal discovery algorithm includes a time-series-based causal test streaming version or an incremental causal structure learning algorithm. The detection unit 26 is used to dynamically adjust the graph structure of the model through an incremental structure learning algorithm. When a change in the distribution of user health status is detected, it automatically identifies the causal relationship that needs to be corrected and updates the connection relationship of the corresponding node.

[0087] It should be noted that the foregoing explanation of the method embodiments also applies to the apparatus of the embodiments of this disclosure, and the principle is the same. Therefore, the embodiments of this disclosure are not limited thereto.

[0088] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0089] Figure 9A schematic block diagram of an example electronic device 400 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0090] like Figure 9 As shown, device 400 includes a computing unit 401, which can perform various appropriate actions and processes based on a computer program stored in ROM (Read-Only Memory) 402 or a computer program loaded from storage unit 408 into RAM (Random Access Memory) 403. RAM 403 may also store various programs and data required for the operation of device 400. The computing unit 401, ROM 402, and RAM 403 are interconnected via bus 404. I / O (Input / Output) interface 405 is also connected to bus 404.

[0091] Multiple components in device 400 are connected to I / O interface 405, including: input unit 406, such as keyboard, mouse, etc.; output unit 407, such as various types of monitors, speakers, etc.; storage unit 408, such as disk, optical disk, etc.; and communication unit 409, such as network card, modem, wireless transceiver, etc. Communication unit 409 allows device 400 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0092] The computing unit 401 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 401 include, but are not limited to, CPUs (Central Processing Units), GPUs (Graphics Processing Units), various special-purpose AI (Artificial Intelligence) computing chips, various computing units running machine learning model algorithms, DSPs (Digital Signal Processors), and any suitable processor, controller, microcontroller, etc. The computing unit 401 performs the various methods and processes described above, such as risk warning methods. For example, in some embodiments, the risk warning method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 408. In some embodiments, part or all of the computer program may be loaded and / or installed on device 400 via ROM 402 and / or communication unit 409. When the computer program is loaded into RAM 403 and executed by the computing unit 401, one or more steps of the methods described above may be performed. Alternatively, in other embodiments, the computing unit 401 may be configured to perform the aforementioned risk warning method by any other suitable means (e.g., by means of firmware).

[0093] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, FPGAs (Field Programmable Gate Arrays), ASICs (Application-Specific Integrated Circuits), ASSPs (Application-Specific Standard Products), SOCs (System-on-Chips), CPLDs (Complex Programmable Logic Devices), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0094] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0095] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, RAM, ROM, EPROM (Electrically Programmable Read-Only Memory) or flash memory, optical fiber, CD-ROM (Compact Disc Read-Only Memory), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0096] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (Cathode-Ray Tube) or LCD (Liquid Crystal Display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0097] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or computing systems that include middleware components (e.g., application servers), or computing systems that include frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include LANs (Local Area Networks), WANs (Wide Area Networks), the Internet, and blockchain networks.

[0098] Computer systems can include clients and servers. Clients and servers are generally geographically separated and typically interact via communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. A server can be a cloud server, also known as a cloud computing server or cloud host, a hosting product within the cloud computing service system that addresses the shortcomings of traditional physical hosts and VPS (Virtual Private Server) services, such as high management difficulty and weak business scalability. Servers can also be servers for distributed systems or servers incorporating blockchain technology.

[0099] It's important to note that artificial intelligence (AI) is the study of enabling computers to simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, and planning). It encompasses both hardware and software technologies. AI hardware technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, and big data processing. AI software technologies primarily include computer vision, speech recognition, natural language processing, machine learning / deep learning, big data processing, and knowledge graph technologies.

[0100] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0101] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method of early warning of a risk, characterized in that, The method comprises the following steps: Initializing the model topology based on a medical knowledge base, and performing personalized learning on the model through a multi-modal semantic embedding vector sequence to generate a user-specific causal graph; Real-time updating of the tree model state, using a filtering algorithm to combine the current semantic embedding vector with the posterior probability distribution at the previous time, and recursively calculating the current posterior probability distribution of each health state latent variable; Based on the updated model state, forward simulation is performed to deduce the probability of occurrence of the target event within a preset time window, and to identify the high-probability causal path leading to the event; According to the risk probability and the causal path, a hierarchical intervention instruction is generated, the causal relationship is experimentally calibrated through an active disturbance verification mechanism, and the structure or parameters of the causal graph are adjusted based on user feedback information.

2. The method of claim 1, wherein, The method comprises the following steps: Initializing the model topology based on a medical knowledge base, and performing personalized learning on the model through a multi-modal semantic embedding vector sequence to generate a user-specific causal graph; Initializing the model topology based on a medical knowledge base, and performing personalized learning on the model through a multi-modal semantic embedding vector sequence to generate a user-specific causal graph; 3. The method of claim 1, wherein, The method comprises the following steps: Using a particle filtering algorithm as the filtering algorithm, the posterior probability distribution of each latent variable at the current time is recursively calculated; The newly observed semantic embedding vector and the posterior probability distribution at the previous time are weighted and fused to generate the health state estimate at the current time.

4. The method of claim 1, wherein, The method comprises the following steps: Generating multiple future state evolution paths through forward sampling simulation method, and calculating the sequence proportion of the target event occurring at least once in all paths as the risk probability; Using statistical analysis to identify the high-probability causal path leading to the target event, the causal path being the state transition sequence with the highest frequency.

5. The method of claim 1, wherein, The method further comprises the following steps: Using an online causal discovery algorithm to continuously update the dynamic causal digital twin model; wherein the online causal discovery algorithm includes a time series-based causal test streaming version or an incremental causal structure learning algorithm; Using an incremental structure learning algorithm to dynamically adjust the graph structure of the model, when detecting changes in the user's health state distribution, automatically identifying the causal relationship that needs to be corrected and updating the connection relationship of the corresponding nodes.

6. A risk early warning device, characterized by, The method comprises the following steps: Initializing the model topology based on a medical knowledge base, and performing personalized learning on the model through a multi-modal semantic embedding vector sequence to generate a user-specific causal graph; The computing unit is configured to update the tree model state in real time, combine the current semantic embedding vector with the posterior probability distribution at the previous time by using a filtering algorithm, and recursively calculate the current posterior probability distribution of each health state latent variable; The simulation unit is configured to perform forward simulation based on the updated model state, deduce the probability of occurrence of a target event within a preset time window, and identify a high-probability causal path leading to the event; The calibration unit is configured to generate a hierarchical intervention instruction according to the risk probability and the causal path, perform experimental calibration on the causal relationship through an active disturbance verification mechanism, and adjust the structure or parameters of the causal graph based on user feedback information.

7. The apparatus of claim 6, wherein, The generating unit is further configured to: initialize the model topology based on a medical knowledge base, and define the topology in the form of a dynamic Bayesian network; wherein a node set represents a health state latent variable that cannot be directly observed, and a directed edge set represents a conditional dependency relationship between variables, including an instantaneous dependency relationship at the same time point and a time series dependency relationship across time points; perform structure learning based on a score-based causal discovery algorithm, maximize the log marginal likelihood of user data as an objective function, and perform parameter learning by using an expectation maximization algorithm to obtain optimal parameters of the conditional probability distribution from the user data.

8. An electronic device, comprising: comprise: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-5.

9. A non-transitory computer-readable storage medium having stored thereon computer instructions, wherein, The computer instructions are used to enable the computer to perform the method of any one of claims 1-5.

10. A computer program product, characterised in that, comprise a computer program that, when executed by a processor, implements the method of any one of claims 1-5.

Images