Real-time whole course information acquisition method for orthopedic patient

Abstract

The invention relates to a real-time whole course information acquisition method for orthopedic patients, and belongs to the field of intelligent monitoring and risk management, and the core process comprises the following steps: personalized model construction: generating a personalized initial state vector based on patient physiological baseline data and a predefined biomechanical causal map; real-time state inference: inferring in-vivo bio-mechanical states which cannot be directly observed in real time by utilizing a causal graph neural network and combining external observable data; risk quantification and decision making: calculating a risk influence index of a real-time state, and comparing the risk influence index with a preset threshold value to generate an active information acquisition decision; and closed-loop information collection: pushing an intervention task to the patient according to the decision, and feeding back newly collected data to the network to continuously update the state. The problem of hysteresis of traditional rehabilitation monitoring is solved, and a passive and delayed management mode in the past is converted into active and real-time accurate monitoring.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent monitoring and risk management of orthopedic rehabilitation, in particular to a real-time full-course information collection method for orthopedic patients. BACKGROUND

[0002] Information collection for the rehabilitation process of orthopedic patients is a technology aimed at achieving precise and forward-looking management of patients; during the rehabilitation process of patients, the biomechanical state in the body is the core basis for judging rehabilitation progress, predicting future trends and developing personalized guidance programs;

[0003] However, in actual rehabilitation management, these key in-vivo biomechanical states cannot be directly observed. Existing rehabilitation monitoring methods mainly rely on the subjective feedback of patients and infrequent in-hospital examinations. This passive and delayed information acquisition mode leads to obvious delays in rehabilitation monitoring and makes it difficult to accurately predict the recovery trajectory of patients and provide truly personalized rehabilitation guidance, thereby affecting the accuracy and forward-looking nature of rehabilitation management.

[0004] The above information disclosed in the above background section is only used to strengthen the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0005] The purpose of the present application is to provide a real-time full-course information collection method for orthopedic patients to solve the problems raised in the above background technology; specifically, the technical solution of the present application includes the following steps:

[0006] S1, personalized model construction step: obtaining the static physiological baseline data of the patient and the predefined biomechanical causal graph; based on the static physiological baseline data, calculating the initial weight of each state variable node in the biomechanical causal graph; combining the initial weight with the predefined node basic state vector to generate a personalized initial state vector for each node;

[0007] S2, real-time state inference step: collecting external observable data of the patient; using a causal graph neural network to infer the real-time biomechanical state of all unobserved nodes, wherein the causal graph neural network uses the biomechanical causal graph as the topological structure and is initialized using the personalized initial state vector, and the external observable data is input as evidence nodes;

[0008] S3, risk quantification and decision step: based on the real-time biomechanical state, calculating the risk influence index of each node; comparing the risk influence index with the preset risk influence trigger threshold to generate an active information collection decision;

[0009] S4, closed-loop information collection step: in response to the active information collection decision, selecting and pushing an intervention task to the patient; feeding the newly collected data through the execution of the intervention task to the causal graph neural network to update the real-time biomechanical state.

[0010] Optionally, in the personalized model construction step, the static physiological baseline data includes age data, gender data, weight data, and surgery type data.

[0011] Optionally, in the real-time state inference step, the training of the causal graph neural network is optimized using a composite loss function, which is composed of a data fidelity loss and a causal consistency loss.

[0012] Optionally, the data fidelity loss is used to measure the difference between the predicted value of the evidence node and the true measured value; the causal consistency loss is used to punish the influence connection that violates the preset topological structure of the biomechanical causal graph.

[0013] Optionally, in the risk quantification and decision step, the step of calculating the risk influence index includes:

[0014] S31, taking the real-time biomechanical state as the initial condition for forward reasoning to generate a future rehabilitation trajectory;

[0015] S32, using Monte Carlo random inactivation technology to quantify the prediction uncertainty of each node;

[0016] S33, by backpropagation of the causal graph neural network, calculating the sensitivity of the long-term rehabilitation result to the node state;

[0017] S34, multiplying the prediction uncertainty and the sensitivity to obtain the risk influence index of the node.

[0018] Optionally, in the risk quantification and decision step, the generation of the active information collection decision includes:

[0019] If the risk influence index of any node exceeds the risk influence trigger threshold, a decision to trigger information collection is generated;

[0020] If the risk influence index of all nodes does not exceed the risk influence trigger threshold, a decision to continue executing the real-time state inference step is generated.

[0021] Optionally, in the closed-loop information collection step, the selection of the intervention task is based on the principle of maximizing risk reduction, and is selected from a pre-defined intervention task library.

[0022] Optionally, the step of selecting the intervention task includes:

[0023] S41, identifying a set of nodes that each selectable intervention task can provide information for;

[0024] S42, aggregate the risk influence indexes of all nodes in each node set to obtain a total risk influence;

[0025] S43, select an intervention task with the maximum total risk influence as an optimal intervention task for pushing.

[0026] Compared with the prior art, the present application has the following beneficial effects:

[0027] 1. Real-time and accurate inference of the state in the patient's body is achieved, solving the problem of hysteresis caused by the traditional rehabilitation monitoring relying on patient subjective feedback and low-frequency in-hospital examination; by constructing a biomechanical causal graph and using a causal graph neural network to process external observable data, the biomechanical state in the body that cannot be directly observed can be inferred in real time, and the passive and delayed management mode in the past is changed to active and real-time precise monitoring;

[0028] 2. An individualized rehabilitation mechanism model is constructed, improving the individualization and accuracy of rehabilitation guidance; the method integrates the patient's age, weight, surgery type and other static physiological baseline data in the initial stage, generates an initial state vector conforming to the individual characteristics of each patient, ensures that all subsequent inferences and predictions are based on a digital model starting point closer to the patient's real situation, and provides a reliable foundation for realizing truly personalized rehabilitation guidance;

[0029] 3. A forward-looking risk quantification and decision mechanism is introduced, realizing the transition from passive response to active risk management; the method not only infers the current state, but also predicts the future rehabilitation trajectory, and by comprehensively evaluating the uncertainty of node prediction and its sensitivity to long-term rehabilitation results, a risk influence index is calculated, so that the key risk points with the greatest impact on rehabilitation results can be actively identified, and forward-looking risk warning is realized;

[0030] 4. A closed-loop active information acquisition system is established, significantly improving the efficiency and pertinence of information acquisition; when a key risk is identified, the system will select and push a task that can most effectively eliminate the current core uncertainty from the pre-defined intervention task library based on the principle of maximizing risk reduction; this target-oriented closed-loop feedback mechanism ensures that every time the patient is requested for information is efficient and necessary, avoiding blind disturbance, and realizing intelligent and precise management of the rehabilitation process. BRIEF DESCRIPTION OF DRAWINGS

[0031] The present application will be further explained in conjunction with the accompanying drawings and embodiments:

[0032] Figure 1 is a flowchart of the method of the present application. DETAILED DESCRIPTION

[0033] In order to make the objects, technical solutions and advantages of the present application clearer, further detailed description will be given below in combination with specific embodiments.

[0034] Embodiment 1:

[0035] Please refer to Figure 1 , a real-time full-course information acquisition method for orthopedic patients, comprising the following steps:

[0036] S1, a personalized model construction step: obtaining static physiological baseline data of the patient and a predefined biomechanics causal graph; based on the static physiological baseline data, calculating the initial weight of each state variable node in the biomechanics causal graph; combining the initial weight with the predefined node basic state vector to generate a personalized initial state vector for each node;

[0037] S2, a real-time state inference step: collecting external observable data of the patient; using a causal graph neural network to infer the real-time biomechanics state of all unobserved nodes, wherein the causal graph neural network uses the biomechanics causal graph as the topological structure and uses the personalized initial state vector for initialization, and also inputs the external observable data as evidence nodes;

[0038] S3, a risk quantification and decision step: based on the real-time biomechanics state, calculating the risk influence index of each node; comparing the risk influence index with a preset risk influence trigger threshold to generate an active information acquisition decision;

[0039] S4, a closed-loop information acquisition step: in response to the active information acquisition decision, selecting and pushing an intervention task to the patient; feeding back the newly collected data through the execution of the intervention task to the causal graph neural network to update the real-time biomechanics state.

[0040] The embodiment of the present application provides a real-time full-course information acquisition method for orthopedic patients, which aims to solve the problem that the existing technology cannot directly observe the biomechanics state in the patient's body, resulting in delayed rehabilitation monitoring and inability to accurately predict and individualize guidance; the embodiment constructs a personalized biomechanics model and combines a causal graph neural network for real-time inference and risk prediction, and finally realizes accurate and forward-looking management of the rehabilitation process through closed-loop active information acquisition;

[0041] The overall technical process of the embodiment, as a complete and self-consistent technical closed loop, comprises the following steps:

[0042] The personalized model construction step S1 aims to create a digital rehabilitation mechanism model for each patient that conforms to their individual characteristics, serving as the foundational framework for all subsequent inferences and predictions. This step acquires the patient's static physiological baseline data and a predefined biomechanical causal graph. The biomechanical causal graph is a directed acyclic graph predefined by orthopedic medical experts as the basis of domain knowledge. Its function is to provide inviolable medical logical constraints for subsequent algorithms. Nodes represent key state variables in the rehabilitation process, and directed edges represent medical causal relationships between variables. Based on the static physiological baseline data, the initial weight of each state variable node in the biomechanical causal graph is calculated. In this embodiment, this calculation is implemented through a linear weighted model, designed to map discrete patient static information into initial state parameters that have a real impact on the model.

[0043] To integrate individual patient information into a general medical model, this embodiment uses the following formula to calculate the initial weights. :

[0044]

[0045] in, The initial weight of the i-th graph node is a dimensionless scalar, calculated by this formula.

[0046] The global baseline weight of a node is a dimensionless scalar that represents the average level of the general population and is preset in the system.

[0047] The total number of physiological baseline parameters, an integer, depends on the types of parameters collected;

[0048] The j-th physiological baseline parameter value after normalization is a dimensionless scalar obtained from the normalized raw data.

[0049] The weighting coefficient of the j-th physiological parameter is dimensionless; this weighting coefficient It was obtained in advance through offline multivariate regression analysis on a calibrated dataset covering numerous historical rehabilitation cases, reflecting the statistical strength of the influence of different physiological parameters on specific state variables;

[0050] This formula uses the patient's individual physiological parameters And the coefficient reflecting its importance Multiply and add to the baseline weights. Above, this ensures the calculated Both population commonality and individual difference are contained; combining the initial weight and the predefined node base state vector, the personalized initial state vector of each node is generated; the node base state vector is a learnable vector associated with the node type, which represents the general feature embedding of the node in general cases, and its dimension is determined by the network structure, and its initial value can be obtained by pre-training;

[0051] The personalized initial state vector is generated as follows:

[0052]

[0053] Wherein, : the personalized initial state vector of the i-th node, which is a vector, calculated by the formula, as the initial input of the subsequent graph neural network;

[0054] : the individualized scalar weight calculated in the previous step, which is a dimensionless scalar, as described above;

[0055] : the base state vector associated with the type of node i, which is a vector, predefined in the system or learned through training;

[0056] The physical meaning of this operation is to use the individualized scalar weight of the patient to scale or modulate a general base feature vector , so as to generate an initial state vector containing general domain knowledge and reflecting individual differences; for example, a patient with a larger body weight, the initial weight of the soft tissue stress related node will be correspondingly higher, so that the norm of the initial state vector of the patient is larger, so that the model reflects the individualized risk characteristics of the patient from the beginning of the calculation;

[0057] Real-time state inference step S2, the core of which is to use a causal graph neural network to infer the complete biomechanical state in the patient's body in real time according to the sparse external data generated by the patient on a daily basis; this step collects the patient's external observable data, such as the joint flexion angle collected in real time through wearable sensors, or the subjective pain score recorded through a mobile phone App, etc.; a causal graph neural network is used to infer the real-time biomechanical state of all unobserved nodes; in this embodiment, the causal graph neural network uses the biomechanical causal graph constructed in the previous step as the topological structure, which means that the connection relationship of the network is constrained by medical prior knowledge, avoiding the reasoning of a black box model that violates medical logic; and the network is initialized using the personalized initial state vector generated in the previous step, ensuring that the starting point of the inference conforms to the individual characteristics of the patient; the external observable data is input into the network as evidence nodes, and the graph neural network reversely infers the real-time values of all unobserved nodes according to the values of a few evidence nodes through an information propagation mechanism;

[0058] Risk quantification and decision step S3 aims to use the model to predict the future rehabilitation trend of the patient and actively identify the key risk points affecting the prediction results; based on the real-time biomechanical state obtained in the previous step, the risk influence index of each node is calculated; the risk influence index is a self-defined composite index, which quantifies the fluctuation range of the long-term rehabilitation results caused by the uncertainty of the current state of a node, so as to identify the high-risk variables that need the most attention; the risk influence index is compared with the preset risk influence trigger threshold to generate an active information collection decision; the risk influence trigger threshold is a value set by clinical experts according to the acceptable risk level, which is used to determine whether the risk predicted by the current model exceeds the acceptable range, so as to decide whether to start intervention;

[0059] Closed-loop information collection step S4 actively requests the patient for key information that can eliminate the current core risk to the greatest extent according to the decision of the previous stage, thereby forming a complete closed-loop system from discovering unknown to actively clarifying; in response to the active information collection decision, that is, when the of a certain node exceeds the risk influence trigger threshold, the system is activated and selects to push an intervention task to the patient; the data newly collected by executing the intervention task is fed back to the causal graph neural network to update the real-time biomechanical state; the newly collected data is input into the model as new evidence nodes, and since these information is collected with a high degree of pertinence for eliminating the highest risk, it can significantly reduce the uncertainty of the related nodes, thereby effectively reducing the overall prediction risk and completing an intelligent closed loop;

[0060] The present application constructs a complete system from personalized modeling, real-time state inference, prospective risk quantification to closed-loop active information collection, compared with the passive and lagging mode of the prior art relying on patient subjective reports and low-frequency in-hospital examinations, can infer the invisible biomechanical state in the patient's body in real time and accurately, predict the rehabilitation trajectory, and actively identify and eliminate the uncertainty risk in the prediction, thereby bringing significant technical effects; this greatly improves the accuracy and forward-looking of the rehabilitation process management, and can provide reliable data and decision support for realizing truly personalized rehabilitation guidance.

[0061] Embodiment 2:

[0062] In the personalized model construction step, the static physiological baseline data includes age data, gender data, weight data and surgery type data.

[0063] This embodiment is a specific embodiment of the personalized model construction step of embodiment 1; in this step, in order to ensure the accuracy and comprehensiveness of the personalized model, the static physiological baseline data includes age data, gender data, weight data and surgery type data; these data are clinically recognized key factors affecting the orthopedic rehabilitation results, and by taking them as the initial input of the model, the initial calibration accuracy of the personalized biomechanical causal graph can be significantly improved; for example, age and weight directly affect the carrying capacity and recovery speed of the tissue, while gender and surgery type are closely related to endocrine level, tissue structure and injury mode; by explicitly specifying these specific physiological baseline data, the effectiveness and reliability of the personalized calibration of the model are ensured; compared with using ambiguous or incomplete data, the present scheme makes the calculation of the initial weight more targeted, so that the subsequent inference and prediction of the entire model start from a starting point closer to the real situation of the patient, improving the accuracy of the overall scheme.

[0064] Embodiment 3:

[0065] In the real-time state inference step, the training of the causal graph neural network is optimized using a composite loss function, and the composite loss function is composed of a data fidelity loss and a causal consistency loss.

[0066] The data fidelity loss is used to measure the difference between the predicted value of the evidence node of the model and the true measured value; the causal consistency loss is used to punish the influence connection that violates the preset topological structure of the biomechanical causal graph.

[0067] ​This embodiment is a further optimization of the causal graph neural network training process in the real-time state inference step of Embodiment 1; to ensure that the inference result not only fits the observation data, but also strictly follows the preset biomechanical causal law, in this embodiment, the training of the causal graph neural network is optimized using a composite loss function, which is composed of a data fidelity loss and a causal consistency loss;

[0068] The purpose of this composite loss function is to drive the network parameters to converge to two targets at the same time when minimizing the total loss by gradient descent method: one is to fit the observation data, and the other is to follow the causal structure; its specific form is as follows:

[0069]

[0070] Among them, : the total loss function, which is a dimensionless scalar, is the final goal of network optimization;

[0071] : data fidelity loss, which is a dimensionless scalar, is calculated by the following formula;

[0072] : causal consistency loss, which is a dimensionless scalar, is calculated by the following formula;

[0073] : causal constraint weight hyperparameter, which is dimensionless, is determined by grid search on the validation set to achieve the best generalization performance;

[0074] Data fidelity loss , used to measure the difference between the predicted value of the evidence node by the model and the true measured value; its form is derived from the mean square error in statistics, and the calculation method is as follows:

[0075]

[0076] Among them, : the number of observable nodes in each training, which is an integer, depends on the data input in this batch;

[0077] : the normalized true measured value of the observable node, which is dimensionless, is obtained by collecting and processing through sensors and other devices;

[0078] : the predicted value of the observable node by the model, which is dimensionless, is calculated by forward propagation of the graph neural network;

[0079] Causal consistency loss is a loss function defined by the application, which is used to punish the influence connection that violates the preset topology of the biomechanical causal graph; the purpose is to force the internal influence relationship learned by the network to be consistent with the medical priori knowledge; the calculation method is as follows:

[0080]

[0081] Wherein, : the adjacency matrix of the predefined causal graph, is a matrix, which is preset in the system, indicates that there is a causal edge from node j to i, otherwise 0;

[0082] : the weight matrix dynamically generated by the neural network in the reasoning process through the attention mechanism, is a matrix, which is calculated by the internal parameters of the network, indicates the influence of node j on node i;

[0083] : Sigmoid function, is a function, which is used to normalize the influence value to 0-1;

[0084] When there is no causal path from j to i in the graph , if the network mistakenly learns a strong influence is larger, the term is 1, which leads to increase, thereby punishing the network parameters and forcing them to correct this wrong connection in subsequent updates;

[0085] By introducing this compound loss function, the application realizes the deep integration of data-driven and knowledge-driven; compared with the traditional neural network training method which simply relies on data fitting, the present scheme, through the causal consistency loss , integrates the priori knowledge of orthopedic medicine as a strong constraint into model training, effectively avoiding the model from producing inference results that violate biological common sense; this not only significantly improves the inference accuracy and robustness of the model under sparse observation data, but also enhances the explainability of the model and the credibility of clinical application.

[0086] Embodiment 4:

[0087] In the risk quantification and decision-making step, the step of calculating the risk influence index includes:

[0088] S31, taking the real-time biomechanical state as the initial condition to perform forward reasoning to generate a future rehabilitation trajectory;

[0089] S32, using Monte Carlo random inactivation technology to quantify the prediction uncertainty of each node;

[0090] S33, calculate the sensitivity of the long-term rehabilitation result to the node state by backpropagation on the causal graph neural network;

[0091] S34, multiply the prediction uncertainty and the sensitivity to obtain the risk influence index of the node.

[0092] This embodiment is a detailed description of the process of calculating the risk influence index in the risk quantification and decision-making step of embodiment 1; this calculation process precisely combines the predicted uncertainty and its potential impact on the clinical result, including the following steps:

[0093] The real-time biomechanical state is used as the initial condition for forward reasoning to generate the future rehabilitation trajectory; that is, the trained graph neural network model is used to iterate in the time dimension to predict the evolution trend of all key state variables in the future period of time;

[0094] The Monte Carlo random inactivation technique is used to quantify the prediction uncertainty of each node; Monte Carlo random inactivation is to perform times of forward propagation with random inactivation on the same input to obtain different prediction results; the uncertainty is quantified by calculating the standard deviation of these results; the calculation formula of the node prediction uncertainty is as follows:

[0095]

[0096] wherein, : the prediction uncertainty of the i-th node, which has the same dimension as the predicted value , is calculated by the formula;

[0097] : the total number of random forward propagation, which is an integer, is determined according to the convergence in the experiment, and is usually 30-50;

[0098] : the predicted value of the i-th node at the j-th propagation, which is a numerical value, is obtained by a single forward propagation of the model; : the average value of the j-th prediction of the i-th node, which is a numerical value, is obtained by averaging the j-th prediction of the i-th node;

[0099] : the predicted value of the i-th node at the j-th propagation, which is a numerical value, is obtained by a single forward propagation of the model; : the average value of the j-th prediction of the i-th node, which is a numerical value, is obtained by averaging the j-th prediction of the i-th node;

[0100] The sensitivity of the long-term rehabilitation result to the node state is calculated by backpropagation on the causal graph neural network; the long-term rehabilitation result ​​​​​​is a pre-defined index by clinical experts, representing the long-term key rehabilitation outcome, such as the expected maximum joint range of motion at the 90th day; sensitivity representative outcome index of the node current state value , which is automatically calculated by standard gradient backpropagation on the trained network model;

[0101] The risk influence index of the node is obtained by multiplying the prediction uncertainty and the sensitivity; the final risk influence index The calculation logic of the risk influence index of the node is to multiply the uncertainty of the node and its sensitivity to the long-term outcome, and the formula is as follows:

[0102]

[0103] wherein, : the risk influence index of the node, which has the same dimension as the long-term outcome , and is calculated by the formula;

[0104] : the node prediction uncertainty calculated in the previous step, which has the same dimension as , and is from the same source;

[0105] : the sensitivity calculated in the previous step, which has the dimension of , and is from the same source;

[0106] The logic of this formula is that the risk of a node depends not only on its own uncertainty , but also on its influence on the final rehabilitation outcome (sensitivity); only when both are high does the node constitute a real core risk; therefore, this embodiment defines an innovative risk quantification method; compared with the traditional method which only focuses on the prediction value or simple uncertainty, this scheme can more deeply reveal the key risk points in the rehabilitation process by calculating the risk influence index , which not only tells us which state is uncertain, but more importantly, which uncertain state has the greatest impact on the final rehabilitation outcome, so that the subsequent active information acquisition is more targeted and efficient, realizing the qualitative change from passive response to active risk management.

[0107] Embodiment 5:

[0108] In the risk quantification and decision step, the generation of the active information acquisition decision includes:

[0109] If the risk influence index of any node exceeds the risk influence trigger threshold, a decision to trigger information acquisition is generated;​

[0110] If the risk influence index of all nodes does not exceed the risk influence trigger threshold, a decision is generated to continue performing the real-time state inference step.

[0111] This embodiment is a specific implementation of the active information collection decision logic generated in the risk quantification and decision step of embodiment 1. This decision mechanism is simple and effective, ensuring that the system only intervenes when necessary. Specifically, the generation of the active information collection decision includes:

[0112] If the risk influence index of any node exceeds the risk influence trigger threshold a decision to trigger information collection is generated; this indicates that the prediction uncertainty of at least one node in the system and its potential negative impact on the long-term outcome has reached a clinically unacceptable level, and immediate action must be taken to eliminate this risk;

[0113] If the risk influence index of all nodes does not exceed the risk influence trigger threshold a decision is generated to continue performing the real-time state inference step; this indicates that the current model's prediction is within an acceptable risk range, and the system will continue to monitor silently in the background without disturbing the patient until a new risk appears;

[0114] By setting clear trigger thresholds and decision logic, this embodiment realizes an automated, non-invasive monitoring system; it avoids unnecessary and frequent information requests from patients, and only initiates intervention when a truly critical risk is identified, thereby achieving a good balance between ensuring rehabilitation safety and improving user experience.

[0115] Embodiment 6:

[0116] In the closed-loop information collection step, the selection of intervention tasks is based on the principle of maximizing risk reduction, and is selected from a pre-defined intervention task library.

[0117] The step of selecting an intervention task includes:

[0118] S41, identifying a set of nodes that each selectable intervention task can provide information for;

[0119] S42, aggregating and calculating the risk influence index of all nodes in each node set to obtain the total risk influence;

[0120] S43, selecting the intervention task with the maximum total risk influence as the optimal intervention task for pushing.

[0121] ​​This embodiment is a further refinement and optimization of how to select the optimal intervention task in the closed-loop information collection step of Embodiment 1; to ensure that the collected information can most efficiently eliminate the currently identified core risks, the selection of the intervention task is based on the principle of maximizing risk reduction, and is selected from a predefined intervention task library; the intervention task library is a pre-established database that stores a series of specific rehabilitation actions or questionnaires, and each task is pre-calibrated by a physical therapist and biomechanics expert, and the strong correlation between it and one or more unobserved nodes in the causal graph is explicitly stated;

[0122] The decision model for selecting the optimal intervention task is as follows:

[0123] Identify the node set that each selectable intervention task can provide information; that is, for each candidate task in the task library , find the node set that can provide observation data directly or indirectly by executing the task ; this correspondence is pre-calibrated and stored in the system;

[0124] Sum up the risk influence index of all nodes in each node set to obtain the total risk influence; that is, for each candidate task , calculate the sum of the risk influence indexes of all nodes in the corresponding node set ;

[0125] Select the intervention task with the largest total risk influence as the optimal intervention task for pushing; this selection process is expressed by the following formula:

[0126]

[0127] where, : the selected optimal intervention task, is the task identifier, obtained by this formula decision;

[0128] : the set of all selectable intervention tasks, is the set, from the predefined intervention task library;

[0129] : the node set that executing task a can provide information, is the set, pre-calibrated;

[0130] : the node risk influence index calculated in the previous step, is a numerical value, as above;

[0131] The logic of this formula is to select the intervention task that can solve the node set with the largest total risk influence in the current system at one time;

[0132] The embodiment ensures that each active information collection is the most efficient by maximizing the risk reduction principle; it avoids blindly asking patients for information, but accurately selects an intervention task that can solve the current source of maximum uncertainty; this not only maximizes the efficiency of the closed-loop information collection process, but also maximizes the reduction of the disturbance to the patient, and improves the intelligent level and practicability of the whole system.

[0133] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application rather than limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. A method for real-time, full-episode information acquisition for orthopedic patients, characterized in that, The method comprises the following steps: S1, a personalized model construction step: obtaining static physiological baseline data of a patient and a predefined biomechanical causal graph; based on the static physiological baseline data, calculating an initial weight of each state variable node in the biomechanical causal graph; combining the initial weight with a predefined node basic state vector to generate a personalized initial state vector of each node; S2, a real-time state inference step: collecting external observable data of the patient; using a causal graph neural network to infer real-time biomechanical states of all unobserved nodes, wherein the causal graph neural network is initialized using the personalized initial state vector and the external observable data is input as an evidence node; S3, a risk quantification and decision-making step: based on the real-time biomechanical states, calculating a risk influence index of each node; comparing the risk influence index with a preset risk influence trigger threshold to generate an active information collection decision; S4, a closed-loop information collection step: in response to the active information collection decision, selecting and pushing an intervention task to the patient; feeding the newly collected data obtained by executing the intervention task back to the causal graph neural network to update the real-time biomechanical states.

2. The real-time, full-length information collection method for orthopedic patients according to claim 1, characterized in that, In the personalized model construction step, the static physiological baseline data includes age data, gender data, weight data and surgery type data.

3. The real-time, full-length information collection method for orthopedic patients according to claim 1, wherein, In the real-time state inference step, the training of the causal graph neural network is optimized using a composite loss function, which is composed of a data fidelity loss and a causal consistency loss.

4. The real-time, full-length information collection method for orthopedic patients according to claim 3, characterized in that, The data fidelity loss is used to measure the difference between the predicted value of the model for the evidence node and the true measured value; the causal consistency loss is used to punish the influence connection that violates the preset topological structure of the biomechanical causal graph.

5. The real time complete information capturing method for orthopedic patients as claimed in claim 1 wherein, In the risk quantification and decision-making step, the step of calculating the risk influence index comprises: S31, using the real-time biomechanical states as initial conditions for forward reasoning to generate a future rehabilitation trajectory; S32, using Monte Carlo random inactivation technology to quantify the prediction uncertainty of each node; S33, calculating the sensitivity of the long-term rehabilitation result to the node state by backpropagation of the causal graph neural network; S34, multiplying the prediction uncertainty and the sensitivity to obtain the risk influence index of the node.

6. The real time complete information capturing method for orthopedic patients as claimed in claim 1 wherein, In the risk quantification and decision-making step, the generation of the active information collection decision comprises: if the risk influence index of any node exceeds the risk influence trigger threshold, a decision to trigger information collection is generated; if the risk influence index of all nodes does not exceed the risk influence trigger threshold, a decision to continue executing the real-time state inference step is generated.

7. The real time complete information capturing method for orthopedic patients as claimed in claim 1 wherein, In the closed-loop information collection step, the selection of the intervention task is based on the principle of maximizing risk reduction and is selected from a predefined intervention task library.

8. The method for real-time, full episode information gathering for orthopedic patients of claim 7, wherein, The step of selecting the intervention task comprises: S41, identifying a node set that each selectable intervention task can provide information for; S42, aggregating and calculating the risk influence index of all nodes in each node set to obtain a total risk influence; S43, selecting the intervention task with the maximum total risk influence as the optimal intervention task for pushing.

Images