Intelligent nursing management system based on artificial intelligence

Abstract

The invention relates to an intelligent nursing management system based on artificial intelligence, in particular to the field of intelligent nursing services, and realizes efficient and personalized nursing services by integrating multi-source heterogeneous data and an advanced artificial intelligence technology. According to the system, wearable equipment, physiological signals, environment data and emotion data are fused into low-dimensional spatiotemporal features by using a dynamic graph convolutional network, the data processing efficiency is greatly improved, an LSTM model is trained locally through a federated learning framework and a gradient distillation mechanism, the communication burden is reduced while privacy protection is optimized, and the user experience is improved. A dynamic path generation module analyzes user behaviors, generates personalized nursing paths and marks potential risks, and finally, a meta-reinforcement learning mechanism dynamically adjusts a feature weighting coefficient to ensure that a new user is adaptive within 15 minutes, so that the intelligence, accuracy and efficiency of nursing services are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of smart nursing services, and more specifically, to a smart nursing management system based on artificial intelligence. Background Technology

[0002] With the rapid development of intelligent health management devices, home care and community care are gradually becoming important health management methods for patients with chronic diseases and individuals requiring long-term care. Patients can collect their health data in real time through wearable devices, such as physiological parameters like heart rate, blood pressure, steps, and body temperature. At the same time, the devices can also record patients' daily activity patterns, sleep quality, and other behavioral information. This data provides an important basis for the development of personalized care plans, and care service providers can design specific health management programs for patients based on this data.

[0003] Existing nursing management systems primarily generate plans using static algorithms and rule engines based on historical data. These plans often fail to effectively address unforeseen events or fluctuations in users' personalized needs. For example, when the external environment changes, such as a sudden epidemic or severe weather, the fixed plans of existing systems are difficult to adjust in a timely manner, resulting in nursing plans that cannot adapt to changes in users' health under these circumstances. At the same time, existing systems fail to fully integrate the dynamic adjustments of users' emotional changes and lifestyle habits when processing user behavior data, resulting in health management plans that do not accurately meet actual needs. Current algorithms lack adaptive capabilities, fail to learn and adjust nursing plans in real time, and cannot effectively integrate data from multiple dimensions such as environment, user physiology, and behavior. This makes the adaptability of existing nursing management systems significantly insufficient in dynamic environments. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing an artificial intelligence-based intelligent nursing management system. Through technologies such as multi-source heterogeneous data fusion, privacy-optimized modeling, dynamic path generation, and strategy evolution, it provides efficient and personalized nursing services, thereby solving the problems mentioned in the background.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: it includes a multi-source heterogeneous processing module, a privacy optimization modeling module, a dynamic path generation module, and a policy evolution implementation module;

[0006] Multi-source heterogeneous processing module: When receiving real-time physiological signals from wearable devices Environmental sensor data and unstructured user sentiment data At that time, a dynamic graph convolutional network is used to perform manifold space mapping, fusing cross-modal data into a low-dimensional spatiotemporal feature tensor. ;

[0007] Privacy optimization modeling module: when the spatiotemporal feature tensor When inputting into the local terminal, the federated learning framework is driven by the gradient distillation mechanism to train a lightweight LSTM prediction model locally and generate gradient residual vectors, which are then uploaded to the cloud to aggregate and update the global model parameters.

[0008] Dynamic path generation module: When obtaining global model parameters, it applies nonlinear dynamic equations to analyze the hidden states of user behavior. By combining the Lyapunov index spectrum to detect behavioral inflection points, a personalized care pathway containing risk markers is output. ;

[0009] Strategy Evolution Implementation Module: When the nursing path When inputting to the policy optimizer, a meta-reinforcement learning mechanism is used to adjust the feature weighting coefficients. It generates adaptive closed-loop control commands and feeds them back to the multi-source heterogeneous processing module, ensuring that the adaptation delay for new users is ≤15 minutes.

[0010] In a preferred embodiment, the specific operation of manifold space mapping in the multi-source heterogeneous processing module is as follows:

[0011] First, the original data is taken as input and mapped to the Riemannian manifold space using differential geometric operators, whose metric tensor... The data is dynamically generated based on its inherent relationships. Then, homogeneous coordinate transformation is performed in space through exponential mapping to unify the data scale. Finally, a homogeneous feature matrix is ​​output.

[0012] The specific operation of the dynamic graph convolutional network to extract the spatiotemporal features of the original data is as follows:

[0013] After completing the manifold space mapping, graph convolution operations are used to extract the spatiotemporal features of the data. Specifically, a spatiotemporal graph is constructed, in which the adjacency matrix of the graph is dynamically updated according to the similarity between feature vectors. Then, spectral convolution kernels are defined in the manifold space, and features are enhanced through convolution operations, thereby extracting spatiotemporal correlation features between different data sources.

[0014] In a preferred embodiment, the specific operation of cross-modal data dimensionality reduction is as follows:

[0015] After enhancing the spatiotemporal features, an unsupervised learning method is used to reduce the dimensionality of the features. By solving the Riemann Laplacian operator, the important dimensions in the manifold space are found, and the dimension with the largest difference in eigenvalues ​​is selected for projection, thereby preserving the nonlinear structure of the data. Finally, the features are compressed into low-dimensional feature tensors through tensor reconstruction, and the final compression result is output.

[0016] In a preferred embodiment, the specific operation of generating the gradient residual vector is as follows:

[0017] First, a spatiotemporal feature tensor is input and modeled using an LSTM model. The output at the current time is predicted based on data from the past time steps. To reduce the complexity of the high-dimensional space, the model uses the covariant derivative on the Riemannian manifold to calculate the gradient of the loss function. Finally, the calculated loss gradient is used as the local gradient tensor.

[0018] The gradient distillation mechanism is as follows:

[0019] Download the global model from the cloud. The virtual gradient of the global model is calculated, and then the Euclidean gradient of the global model is mapped to the manifold space where the local model is located using the projection operator. Then, a residual vector is calculated through logarithmic mapping, which represents the difference between the local model gradient and the global model gradient.

[0020] In a preferred embodiment, the specific operation of updating the global model parameters is as follows:

[0021] First, the update direction of the local model is restored to the correct direction in the manifold space using an exponential mapping. Next, the update values ​​of all local models are aggregated through Riemannian centroid calculation to obtain a new global model parameter. During the aggregation process, the parameters are weighted according to the curvature of the manifold in which each local model is located. Finally, the updated global model parameter is obtained through weighted aggregation.

[0022] In a preferred embodiment, the dynamic path generation module analyzes the hidden state of user behavior. The specific operation is as follows:

[0023] First, define the hidden state vector of user behavior. It resides in a multidimensional phase space, and the evolution of user behavior is described by an equation consisting of two parts: a conservative force component, representing the change in behavior over time; and a time-delayed chaotic term, used to simulate the memory effect of user behavior. The conservative force component is represented by the Hamiltonian. The description includes one linear term and one nonlinear term. A time-delay chaotic term is introduced to introduce nonlinear chaos through an antisymmetric matrix, simulating abrupt changes and complex dynamics of behavior. Finally, it outputs a sequence of user behavior trajectories in phase space. , representing the hidden state at different points in time;

[0024] The specific steps for detecting behavioral inflection points are as follows:

[0025] First, the tangent space evolution matrix is ​​generated by calculating the variational equation. Then, orthogonal QR decomposition is used to iterate the evolution matrix to calculate the maximum Lyapunov exponent. ,like If a point is reached, it indicates that the system behavior has reached an inflection point, which is then marked as a behavioral inflection point. The time series of these inflection points then constitute the output inflection point marking sequence.

[0026] In a preferred embodiment, a personalized care pathway incorporating risk markers is generated. The specific steps are as follows:

[0027] Based on the output inflection point marker sequence, the planning of the nursing pathway is optimized. First, a weighted directed graph is constructed, where each node represents a physiological state and a nursing action, and the edge weight represents the transition cost from one state to another, including risk and distance penalty. Then, the Monge-Kandrovich optimal transport method is used to solve for the optimal path.

[0028] This path minimizes the cost from the initial state to the target state and takes into account the deviation penalty at the inflection point. The final output is an optimal nursing path, which includes the time point, nursing action and corresponding risk value.

[0029] In a preferred embodiment, in the strategy evolution implementation module, the strategy optimizer optimizes the care path. The optimization operation is as follows:

[0030] The nursing pathway is modeled as a game system. By using the adversarial optimization of patient and environmental strategies, the patient's nursing pathway strategy is optimized through a minimax game process, and the Hamiltonian function is output.

[0031] In a preferred embodiment, the feature weighting coefficient is adjusted. The specific operation is as follows:

[0032] After game optimization, a meta-reinforcement learning mechanism is adopted. By updating the policy parameters, the system can adapt to different situations. Based on the Hamiltonian function, the system generates enhanced samples and updates the policy. The policy is adjusted by projecting the latent space policy onto the Riemannian manifold, enabling the system to quickly adapt to new changes.

[0033] In a preferred embodiment, the specific operation for generating adaptive closed-loop control commands is as follows:

[0034] After policy optimization, the system maps features to Lie algebras and adjusts the policy output by calculating feature weighting coefficients. This part combines the policy's sensitivity and risk factors to generate feature weighting coefficients to determine the policy's effectiveness.

[0035] Finally, the system calculates the curvature scalar and adjusts the dynamic regularization intensity, generates control commands by calculating the feedback vector, and feeds these control commands back to the multi-source heterogeneous processing module to ensure that the system's adaptation latency is less than 15 minutes and can quickly respond to new user needs.

[0036] The beneficial effects of this invention are as follows: This AI-based intelligent nursing management system provides efficient and personalized nursing services through technologies such as multi-source heterogeneous data fusion, privacy-optimized modeling, dynamic path generation, and strategy evolution. First, the system fuses physiological signals from wearable devices, environmental sensor data, and user emotional data using a dynamic graph convolutional network, compressing dimensionality and extracting low-dimensional spatiotemporal features to improve data processing efficiency. Second, it utilizes gradient distillation and a federated learning framework to train a lightweight LSTM model locally, effectively reducing the amount of communication data and ensuring privacy and security. The dynamic path generation module analyzes user behavior based on the global model, generates personalized nursing paths, and performs risk labeling based on behavioral inflection points. Finally, a meta-reinforcement learning mechanism is used to optimize feature weighting coefficients, ensuring a latency of less than 15 minutes when adapting to new users, achieving rapid and accurate personalized nursing. This system not only improves the intelligence and accuracy of nursing services but also reduces communication costs while ensuring data security. Attached Figure Description

[0037] Figure 1 This is a flowchart of the method of the present invention;

[0038] Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0040] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0041] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0042] Example 1

[0043] This embodiment provides, for example Figure 1-2 The present invention relates to an artificial intelligence-based intelligent nursing management system, which specifically includes: a multi-source heterogeneous processing module, a privacy optimization modeling module, a dynamic path generation module, and a strategy evolution implementation module;

[0044] Multi-source heterogeneous processing module: When receiving real-time physiological signals from wearable devices Environmental sensor data and unstructured user sentiment data At that time, a dynamic graph convolutional network is used to perform manifold space mapping, fusing cross-modal data into a low-dimensional spatiotemporal feature tensor. This results in a feature dimension compression rate of ≥90%.

[0045] Privacy-optimized modeling module: when spatiotemporal feature tensor When inputting to the local terminal, a federated learning framework driven by a gradient distillation mechanism is used to train a lightweight LSTM prediction model locally and generate gradient residual vectors, which are then uploaded to the cloud to aggregate and update global model parameters, reducing the amount of communication data by 80%.

[0046] Dynamic path generation module: When obtaining global model parameters, it applies nonlinear dynamic equations to analyze the hidden states of user behavior. By combining the Lyapunov index spectrum to detect behavioral inflection points, a personalized care pathway containing risk markers is output. ;

[0047] Strategy Evolution Implementation Module: When the Care Path When inputting to the policy optimizer, a meta-reinforcement learning mechanism is used to adjust the feature weighting coefficients. It generates adaptive closed-loop control commands and feeds them back to the multi-source heterogeneous processing module, ensuring that the adaptation delay for new users is ≤15 minutes.

[0048] In this embodiment, the multi-source heterogeneous processing module needs to be specifically described. The specific operation of manifold space mapping is as follows:

[0049] First, the raw data (physiological signals) Environmental sensor data and unstructured user sentiment data ,in, Indicates a time step. Taking heterogeneous dimensions as input, the data is mapped to the Riemannian manifold space through differential geometric operators. Its metric tensor Dynamically generated from the inherent relationships within the data, the expression is:

[0050] ;

[0051] in, This represents the curvature adjustment factor, ranging from [0, 0.5]. This represents the gradient during feature alignment, controlling the degree of manifold curvature. The formula adaptively adjusts the manifold curvature to preserve cross-modal topology (e.g., ...). (The physiological-environment coupling relationship); secondly, homogeneous coordinate transformation is performed in space through exponential mapping to unify the data scale, the expression of which is:

[0052] ;

[0053] in, To represent the homogeneous features after mapping, i.e., the coordinates of the data in the manifold space, we use an exponential mapping. To standardize data scales This represents the origin of the manifold (initialized to a point of zero curvature). Cross-modal weights, used to describe the correlation between different data sources, are expressed using mutual information entropy. Calculated;

[0054] This approach ensures that different data sources have a uniform scale and can be represented and processed in the same manifold space. Finally, a homogeneous feature matrix is ​​output. ;

[0055] The specific operations of dynamic graph convolutional networks to extract spatiotemporal features from raw data are as follows:

[0056] After completing the manifold space mapping, graph convolution operations are used to extract the spatiotemporal features of the data. Specifically, a spatiotemporal graph is constructed, where the adjacency matrix of the graph is dynamically updated based on the similarity between feature vectors, and its expression is:

[0057] ;

[0058] in, Indicates at time step The dynamic adjacency matrix describes the spatiotemporal correlation between features, with each element... Representation of features At any moment The similarity between them Index representing features, This represents the manifold Sigmoid function, used to calculate the similarity between features. Representation of features The inner product on a manifold, i.e., calculating the cosine of their included angle in Riemannian space. Representation of features The norm in Riemannian space, calculated based on manifold metrics, serves the following purposes: it calculates the cosine of the angle between eigenvectors in Riemannian space to generate a time-varying correlation graph; then, it defines a spectral convolution kernel in the manifold space, with the following expression:

[0059] ;

[0060] in, The feature map after the convolution operation is the nth feature map. The output features of the layer Indicates the index of the current convolutional layer. Let represent the manifold Fourier transform, which is used to perform frequency domain transformation on a manifold, so that signals can be processed in the frequency domain of the manifold, rather than just in the flat Euclidean space. This represents the inverse Fourier transform of the manifold, which transforms the frequency domain data on the manifold back to the original space. This represents a frequency domain filter (constructed via Chebyshev polynomial approximation) that weights features in the frequency domain, in graph convolution. Its function is to control the intensity of different frequency components, and it is designed based on the graph Laplacian operator. This represents the Hadamard product, also known as element-wise multiplication. It represents the multiplication of corresponding elements between two matrices (or vectors). Indicates frequency domain filter and features after manifold Fourier transform Perform element-wise multiplication. This represents the curvature diagonal matrix, i.e., the curvature information of the manifold. It is a diagonal matrix whose elements reflect the geometric properties and curvature of the features. ,in, The metric tensor describes the geometry on the manifold. Indicates taking The diagonal elements form a curvature diagonal matrix. , Indicates the first The input feature map of the first layer is passed as input to the next layer for convolution operation; the convolution operation enhances the features, thereby extracting spatiotemporal correlation features between different data sources, and finally outputting spatiotemporally enhanced features. ( (represents the number of convolution channels).

[0061] The specific operations for cross-modal data dimensionality reduction are as follows:

[0062] After enhancing the spatiotemporal features, an unsupervised learning method is used to reduce the dimensionality of the features by solving the Riemann Laplacian operator, the calculation formula of which is:

[0063]

[0064] in, Denotes the Laplace operator on a manifold. It is a characteristic function. These are the corresponding eigenvalues. Denotes the divergence operator, This represents the gradient operator on the manifold, which performs dimensionality reduction and feature extraction by calculating the rate of change of the signal on the manifold;

[0065] Find the important dimensions in the manifold space and select the dimension with the largest eigenvalue difference for projection, thereby preserving the nonlinear structure of the data. Its expression is:

[0066] ;

[0067] in, This represents the selected optimal projection matrix. Represents the space of orthogonal matrices. Represents the eigenvalues, selected from the previous ones. The eigenvectors corresponding to the eigenvalues ​​of the maximum gaps are selected. The method is as follows:

[0068] ;

[0069] in, The dimension representing the feature will Set as To ensure that the feature dimensions after dimensionality reduction are appropriate;

[0070] Finally, the features are compressed into a low-dimensional feature tensor through tensor reconstruction, the expression of which is:

[0071] ;

[0072] in, This indicates matrix multiplication along the third dimension of the tensor; this operation incorporates spatiotemporal features. Projecting onto the optimal low-dimensional space achieves feature compression;

[0073] Output the final compression result. ,in, This represents the compressed spacetime tensor. Indicates the time dimension. This represents the feature dimension after dimensionality reduction. This represents the number of channels (with the following constraints:) , The zero norm represents the number of non-zero elements. This constraint ensures that the number of non-zero elements in the compressed tensor is less than 10% of the original tensor.

[0074] In this embodiment, the privacy optimization modeling module specifically describes the following operation for generating the gradient residual vector:

[0075] First, input a spatiotemporal feature tensor. And it uses an LSTM model for modeling, defining locally learnable parameters. This is the set of learnable parameters for the local LSTM model, where The parameters of the LSTM model belong to the set. That is, the space of all possible LSTM parameters, construct the prediction equation:

[0076] ;

[0077] in, Indicates the model at time step The predicted output represents the result of the LSTM's prediction based on the input spatiotemporal feature tensor. This represents the output of the LSTM model, i.e., the output from... arrive Feature tensor of time step The predictions made This represents the backtracking window, i.e., the window used by LSTM in the past. Predict the output based on data from each time step; predict the output at the current time step. Based on the past To reduce the complexity of high-dimensional space, the model uses covariant derivatives on the Riemannian manifold to calculate the gradient of the loss function, based on data from each time step. The formula is as follows:

[0078] ;

[0079] in, Indicates parameters of the local LSTM model loss function The gradient, calculated during LSTM training, represents the sensitivity of the loss function to changes in the model parameters. Christoffel notation represents the geometric structure of a Riemannian manifold and describes the effect of changes in the coordinate system on the derivative. On a manifold, the calculation of the covariant derivative considers not only the magnitude of the gradient but also the curvature of the manifold. Represents the loss function For model parameters The partial derivatives, It is a parameter index, representing a specific parameter in the LSTM (such as the weights or biases of a certain layer). Representing manifold coordinates A tiny change in direction, used to describe a small displacement in space, is part of the differential form on a Riemannian manifold. The partial derivative on the manifold is used to describe the rate of change along a certain direction on the manifold. It is another parameter index used to represent a specific direction or coordinate on the manifold; the covariant derivative helps to avoid the dimensional distortion that may occur in traditional Euclidean space within the manifold space. Finally, the calculated loss gradient is used as the local gradient tensor. ,in, Representation and model parameters The relevant tangent space is the set of all vectors passing through a point in the manifold; here, it represents the model parameters in the manifold. The local gradient space, Indicates in tangent space The loss gradient in the LSTM model parameters The gradient vector on the manifold explains how to adjust the parameters to reduce the value of the loss function; it ensures that the gradient is updated along the tangent space of the manifold.

[0080] The gradient distillation mechanism is as follows:

[0081] local gradient tensor As input, where Indicates the model parameters gradient, Indicates the parameters of the local model. The loss function represents the error or loss between the model output and the actual labels, and the global model is downloaded from the cloud. And calculate the virtual gradient of the global model, the formula of which is:

[0082] ;

[0083] in, The projection operator represents the projection of a vector from Euclidean space onto a manifold. tangent space , The tangent space of a manifold, usually referring to a locally linearized space at a certain point, is here the space projected onto the parameters of the model. Represents the teacher model loss function relative to local model parameters The gradient;

[0084] Then, use the projection operator. The Euclidean gradient of the global model is mapped to the manifold space of the local model. Then, a residual vector is calculated through logarithmic mapping. ,in This represents the dimension of the model parameter space, which is the number of parameters. or The dimension of is calculated using the following formula:

[0085] ;

[0086] in, The logarithmic map represents the inverse operation of mapping a tangent vector in a manifold space to a Euclidean space. The logarithmic map is a common technique in manifold learning, transforming vectors in a nonlinear space into operable forms in a linear space. This represents the residual vector, which is the difference between the local gradient and the teacher model gradient. It helps the local model adjust its parameters to better learn the knowledge from the teacher model.

[0087] It represents the difference between the local model gradient and the global model gradient. This process helps the local model gradually approach the global model during training, thereby achieving better optimization results.

[0088] The specific steps to update the global model parameters are as follows:

[0089] First, the update direction of the local model is restored to the correct direction in the manifold space using an exponential mapping, the expression of which is:

[0090] ;

[0091] in, Represents the local model The parameter update amount is mapped exponentially. We obtain the parameters used to update the local model. This indicates that the residual vector By mapping back to the manifold space through an exponential mapping, the update direction of the local model can be obtained;

[0092] Next, all updated values ​​from the local model will be aggregated using the Riemann centroid calculation, which is performed as follows:

[0093] ;

[0094] in, Represents minimizing the objective function That is, finding the global model parameters by minimizing the distance. Represents the local model The weights are used to describe the importance of the model in global aggregation. In manifold The upper metric is the square of the distance between two points. This is the global model parameter. Local model The distance between them;

[0095] To obtain a new global model parameter, during the aggregation process, weights are calculated based on the curvature of the manifold in which each local model resides. The weight calculation formula is as follows:

[0096] ;

[0097] in, Represented for each local model The weights are calculated based on the curvature of the local manifold, with higher weights assigned to models in regions of high curvature (complex models). Represents a metric tensor, describing distances and angles on a manifold. The determinant of the metric tensor reflects the curvature of a local region on the manifold;

[0098] Local models with higher curvature (typically representing more complex models) will have a greater impact on the update of the global model. Ultimately, the updated global model parameters are obtained through weighted aggregation. .

[0099] In this embodiment, the dynamic path generation module, which analyzes the hidden state of user behavior, needs to be specifically described. The specific operation is as follows:

[0100] First, set the global model parameters As input, define the hidden state vector of user behavior. It resides in a multidimensional phase space (dimension 1). The evolution of user behavior is described by an equation, the expression of which is:

[0101] ;

[0102] in, Indicates the user's hidden state The time-varying derivative is used to describe the dynamic evolution of user behavior. Represents global parameters The relevant matrices or operators are used to describe the linear part of the behavior in the dynamic equations. Represents Hamiltonian For hidden states The gradient of the Hamiltonian describes the conservative force component of user behavior, and the Hamiltonian... A mathematical function, typically consisting of two parts: a linear term and a nonlinear term, describes the dynamics of user behavior, and takes the following form: ,in, From global model parameters The derived time-delay chaotic term simulates the memory effect of user behavior through time-delay states. Introducing the influence of past behavior; the time-delay chaotic term is passed through an antisymmetric matrix. Nonlinear chaos is introduced, and a sign function is used to simulate abrupt changes in behavior, whose antisymmetric matrix... The form is:

[0103] ;

[0104] in, Represents the chaos intensity factor, and ,in The Hessian matrix representing the Hamiltonian. Represents the identity matrix. Represents the tensor product. This represents a time lag window, used to define the memory of past actions;

[0105] This consists of two parts: first, a conservative force component, representing the change in behavior over time; and second, a time-delay chaotic term, used to simulate the memory effect of user behavior. The conservative force component is composed of the Hamiltonian. The description includes one linear term and one nonlinear term. A time-delay chaotic term is introduced to introduce nonlinear chaos through an antisymmetric matrix, simulating abrupt changes and complex dynamics of behavior. Finally, it outputs a sequence of user behavior trajectories in phase space. , representing the hidden state at different points in time;

[0106] The specific steps for detecting behavioral inflection points are as follows:

[0107] First, the phase space trajectory As input, the tangent space evolution matrix is ​​generated by calculating the variational equation, and the specific calculation formula is as follows:

[0108] ;

[0109] in, This represents the tangent space matrix, used to describe the evolution of phase space perturbations. The Jacobian matrix describes the points... The rate of change of the system at that point, and ;

[0110] Next, orthogonal QR decomposition is used to iterate the evolution matrix, with the following formula:

[0111] ;

[0112] in, Represents an orthogonal matrix. Represents an upper triangular matrix. Representation matrix The diagonal elements are used to update each feature value;

[0113] Thus, the maximum Lyapunov exponent can be calculated. The calculation formula is as follows:

[0114] ;

[0115] in, Represents an orthogonal matrix transpose, The maximum Lyapunov exponent is calculated from the selected eigenvector. Used to determine the stability of the system;

[0116] like This indicates that the system behavior has reached an inflection point, which is marked as a behavioral inflection point. The time series of these inflection points then constitute the output inflection point label sequence. ;

[0117] Generate personalized care pathways that include risk markers The specific steps are as follows:

[0118] Based on the output inflection point marker sequence, the planning of the nursing pathway is optimized. First, a weighted directed graph is constructed, and the weighted directed graph is defined. :node: Edge weight: ,in, This represents the set of nodes, i.e., the state of the system. This represents a set of edges, i.e., transitions between nodes. The weight of an edge represents the cost of moving from one node to another. Represents a node The corresponding risk value (such as the risk of physiological state). Represents a node and Distance metric between, typically This indicates differences in physiological state or differences in nursing actions. This represents a constant used to adjust the effect of distance on the weights. This indicates that the cost of transitioning decreases with state distance; transitioning to a more distant state or action will incur a higher cost.

[0119] Each node represents a physiological state and a nursing action. The edge weights represent the transition cost from one state to another, including risk and distance penalties. Then, the Monge-Kandrovich optimal transport method is used to solve for the optimal path, and its calculation formula is as follows:

[0120] ;

[0121] in, Represents the distribution from the initial state To the target state distribution All possible transmission plans, This represents the cost of transfers between nodes, including a risk penalty term, i.e., the cost from the node... The cost, This represents the penalty coefficient for deviation from the inflection point, used to adjust the degree of deviation from the inflection point. These represent the actual and target physiological states, respectively. Indicates at time The distance between the actual physiological state and the target physiological state is usually measured using Euclidean distance.

[0122] This path minimizes the cost from the initial state to the target state and takes into account the deviation penalty at the inflection point, ultimately outputting an optimal nursing path. ,in, Indicates time points in the nursing process. Indicates time Corresponding nursing actions, Indicates time The corresponding risk value is used to describe the risk of physiological state and nursing action, including time point, nursing action and corresponding risk value.

[0123] In this embodiment, the specific module to be described is the strategy evolution implementation module, where the strategy optimizer optimizes the nursing pathway. The optimization operation is as follows:

[0124] The nursing pathway is modeled as a game system, utilizing adversarial optimization of patient and environmental strategies. The goal is to optimize the patient's nursing pathway strategy through a minimax game process, outputting a Hamiltonian function. ,in, Indicates time, This represents the system's state vector, a vector used to describe the system's current state. Indicating patient strategy, Let represent the costate vector; thus ensuring that the optimal strategy can be found in a changing environment. Its expression is:

[0125]

[0126] in, This refers to a patient's strategy, which is learnable; that is, the patient's course of action in different environments will adjust over time or as the environment changes. This refers to an environmental perturbation strategy, which is adversarial, meaning that changes in the environment or external factors interfere with the patient's behavior, and these changes are predictable or modelable. This represents a minimax game strategy, an adversarial game between the patient's strategy and the environment's perturbation strategy. The patient attempts to minimize their own loss, while the environment's perturbation strategy attempts to maximize the disturbance to the patient. This represents the integral symbol, indicating the time interval from 0 to... The entire time period, It is a termination time, representing the time span of the entire strategy optimization process. This represents the loss function, which is at time t. Given the patient's actions Feedback from the environment The loss or reward is used to measure the performance of the patient's strategy. Let KL divergence be the denoted KL divergence, a function that measures the difference between two probability distributions; here, it measures the patient's strategy. With global strategy The differences between them This indicates the strength of meta-knowledge transfer, which controls how patient strategies are affected by global knowledge (…). The initial value of 0.7 indicates that the initial patient strategy will rely more on the overall strategy of the environment. This represents the final value function, i.e., the function that performs a function in time. The results of the patient-time strategy Indicates in The system state at any given moment. This indicates the final reward or loss;

[0127] Adjusting feature weighting coefficients The specific operation is as follows:

[0128] After game optimization, the Hamiltonian function is... As input, a meta-reinforcement learning mechanism is adopted. By updating the policy parameters, the system can adapt to different situations. Based on the Hamiltonian function, the system generates augmented samples and updates the policy, the expression of which is:

[0129] ;

[0130] in, This refers to a dataset that has been augmented with new samples. The purpose of this augmentation is to increase the diversity of the training dataset by generating new samples. Indicates system state gradient, Indicates relative to the end of time Some changes, Indicates from the normal distribution The noise from the mid-sample is used to perturb the augmented samples, ensuring sample diversity. The covariance matrix, representing co-state variables, is used to describe the correlation between different samples. This means that enhanced samples are generated by making small adjustments to the system's costate variables (such as Lagrange multipliers);

[0131] By adjusting the policy through latent space policy projection onto the Riemannian manifold, the system can quickly adapt to new changes. Its expression is:

[0132] ;

[0133] in, The policy network, typically a function learned through a neural network, determines what action to take given a state. The parameters of the policy network are typically the weights and biases of the neural network. These parameters are updated through learning. The exponential mapping of a manifold is a mathematical operation used to map a given point to a new point along a direction on the manifold. A manifold is a space with a specific geometric structure, and the mapping operation here refers to a certain update performed on the manifold. Indicates the network parameters of the policy The gradient is used to update the policy by calculating the derivative of the loss function with respect to the parameters. Indicates in augmented samples The expected value, calculated by averaging all augmented samples, is used to optimize the strategy. This represents the meta-learning rate, the learning rate used to update the policy. It is dynamically adjusted based on the model's performance, helping the model adapt quickly during the learning process. Representing costate variables, in optimal control problems, it describes the strength of constraints; it is used in this process to adjust and update the policy.

[0134] The specific steps for generating adaptive closed-loop control commands are as follows:

[0135] After policy optimization, the system maps features to Lie algebras. :

[0136] ;

[0137] in, Lie algebras, which are the dynamic or characteristic state spaces of a system, are discussed here. Depend on This indicates a special type of Lie algebra structure used to describe the rotation and translation of rigid body motion. Characteristic function gradient, This represents a linear transformation matrix used to map features. The feature function represents the policy-related characteristics, describing the behavior of a feature over time or in the environment. This represents a mapping from a vector to an antisymmetric matrix.

[0138] The strategy output is adjusted by calculating the feature weighting coefficients. This part combines the strategy's sensitivity and risk factors to generate feature-weighted coefficients to determine the strategy's effectiveness. Its expression is:

[0139]

[0140] in, Indicates the first One characteristic in time The weighting coefficients are used to adjust the degree of influence of features in the strategy on the decision. This represents the determinant of the matrix product generated through the exponential mapping, combined with the risk parameter. To adjust the strategy, the exponential mapping is a mathematical operation that represents mapping Lie algebra elements to Lie groups. This represents the temperature parameter, used to adjust the model's response to high-risk scenarios. The value varies with risk An increase in risk indicates that the higher the risk, the greater the sensitivity to strategy adjustments. The policy sensitivity tensor describes the policy's sensitivity to changes in features; it is typically the derivative of the policy parameters with respect to the state and the Hamiltonian function.

[0141] Finally, the system calculates the curvature scalar and adjusts the dynamic regularization intensity. It then generates control commands by calculating the feedback vector, which are fed back to the multi-source heterogeneous processing module to ensure the system's adaptation latency is less than 15 minutes and to quickly respond to new user needs. The formula for calculating the curvature scalar is:

[0142] ;

[0143] in, The Riemann curvature scalar is a quantity that describes the curvature of space. It is typically used to represent the geometric properties of a region. The curvature can be obtained by calculating the product of the Riemann tensor and the metric tensor. This represents the Ricci tensor, a contracted form of the Riemann curvature tensor used to describe the curvature of a manifold. The metric tensor, used to describe the distance relationships between different points in a manifold, is a core object in Riemannian geometry. The Ricci tensor, which describes the curvature of space, is obtained by performing some calculations on the Ricci curvature tensor.

[0144] The formula for adjusting the dynamic regularization strength is:

[0145] ;

[0146] in, This indicates the regularization strength, used to adjust the feedback control strength of the system, helping the controller adjust its decisions within a specific time step. Indicates the initial regularization strength. This represents the hyperbolic tangent function, used to control the variation of regularization intensity with curvature. It is a smooth nonlinear function that can smoothly transition within a certain range. Scale factor representing curvature With curvature The product of these is used to adjust the change in regularization strength. The reference curvature is a baseline value obtained from historical data, used to compare the currently calculated curvature. Indicates the first Each feature weighting coefficient The magnitude of the gradient is used to measure the rate of change of the weighting coefficients. The sensitivity of a feature to the system's decision is assessed by calculating the magnitude of the gradient. This represents the gradient response coefficient, used to adjust for the effect of gradient changes. It is closely related to changes in the regularization strength, and ;

[0147] The expression for generating closed-loop control commands is:

[0148] ;

[0149] in, This indicates a feedback control command used to adjust the output of the control system to regulate the system's behavior based on the current characteristic weighting coefficients and regularization strength. It represents the rate of change of curvature with respect to time, describing the degree to which curvature changes over time.

[0150] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0151] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0152] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0153] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0154] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0155] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0156] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An intelligent nursing management system based on artificial intelligence, characterized in that, Specifically, it includes: Multi-source heterogeneous processing module, privacy optimization modeling module, dynamic path generation module, and policy evolution implementation module; Multi-source heterogeneous processing module: When receiving real-time physiological signals from wearable devices Environmental sensor data and unstructured data of user sentiment At that time, a dynamic graph convolutional network is used to perform manifold space mapping, fusing cross-modal data into a low-dimensional spatiotemporal feature tensor. ; Privacy optimization modeling module: when the spatiotemporal feature tensor When inputting into the local terminal, the federated learning framework is driven by the gradient distillation mechanism to train a lightweight LSTM prediction model locally and generate gradient residual vectors, which are then uploaded to the cloud to aggregate and update the global model parameters. Dynamic path generation module: When obtaining global model parameters, it applies nonlinear dynamic equations to analyze the hidden states of user behavior. By combining the Lyapunov index spectrum to detect behavioral inflection points, a personalized care pathway containing risk markers is output. ; Strategy Evolution Implementation Module: When the nursing path When inputting to the policy optimizer, a meta-reinforcement learning mechanism is used to adjust the feature weighting coefficients. It generates adaptive closed-loop control commands and feeds them back to the multi-source heterogeneous processing module, ensuring that the adaptation delay for new users is ≤15 minutes.

2. The intelligent nursing management system based on artificial intelligence according to claim 1, characterized in that: In the multi-source heterogeneous processing module, the specific operation of manifold space mapping is as follows: First, the original data is taken as input and mapped to the Riemannian manifold space using differential geometric operators, whose metric tensor... The data is dynamically generated based on its inherent correlations. Then, homogeneous coordinate transformation is performed in space through exponential mapping to unify the data scale. Finally, a homogeneous feature matrix is ​​output. The specific operation of the dynamic graph convolutional network to extract the spatiotemporal features of the original data is as follows: After completing the manifold space mapping, graph convolution operations are used to extract the spatiotemporal features of the data. Specifically, a spatiotemporal graph is constructed, in which the adjacency matrix of the graph is dynamically updated according to the similarity between feature vectors. Then, spectral convolution kernels are defined in the manifold space, and features are enhanced through convolution operations, thereby extracting spatiotemporal correlation features between different data sources.

3. The intelligent nursing management system based on artificial intelligence according to claim 2, characterized in that: The specific operation of cross-modal data dimensionality reduction is as follows: After enhancing the spatiotemporal features, an unsupervised learning method is used to reduce the dimensionality of the features. By solving the Riemann Laplacian operator, the important dimensions in the manifold space are found, and the dimension with the largest difference in eigenvalues ​​is selected for projection, thereby preserving the nonlinear structure of the data. Finally, the features are compressed into low-dimensional feature tensors through tensor reconstruction, and the final compression result is output.

4. The intelligent nursing management system based on artificial intelligence according to claim 3, characterized in that: In the privacy optimization modeling module, the specific operation for generating the gradient residual vector is as follows: First, a spatiotemporal feature tensor is input and modeled using an LSTM model. The output at the current time is predicted based on data from the past time steps. To reduce the complexity of the high-dimensional space, the model uses the covariant derivative on the Riemannian manifold to calculate the gradient of the loss function. Finally, the calculated loss gradient is used as the local gradient tensor. The gradient distillation mechanism is as follows: Download the global model from the cloud. The virtual gradient of the global model is calculated, and then the Euclidean gradient of the global model is mapped to the manifold space where the local model is located using the projection operator. Then, a residual vector is calculated through logarithmic mapping, which represents the difference between the local model gradient and the global model gradient.

5. The intelligent nursing management system based on artificial intelligence according to claim 4, characterized in that: The specific steps to update the global model parameters are as follows: First, the update direction of the local model is restored to the correct direction in the manifold space using an exponential mapping. Next, the update values ​​of all local models are aggregated through Riemannian centroid calculation to obtain a new global model parameter. During the aggregation process, the parameters are weighted according to the curvature of the manifold in which each local model is located. Finally, the updated global model parameter is obtained through weighted aggregation.

6. The intelligent nursing management system based on artificial intelligence according to claim 5, characterized in that: In the dynamic path generation module, the hidden state of user behavior is analyzed. The specific operation is as follows: First, define the hidden state vector of user behavior. It resides in a multidimensional phase space, and the evolution of user behavior is described by an equation consisting of two parts: a conservative force component, representing the change in behavior over time; and a time-delayed chaotic term, used to simulate the memory effect of user behavior. The conservative force component is represented by the Hamiltonian. The description includes one linear term and one nonlinear term. A time-delay chaotic term is introduced to introduce nonlinear chaos through an antisymmetric matrix, simulating abrupt changes and complex dynamics of behavior. Finally, it outputs a sequence of user behavior trajectories in phase space. , representing the hidden state at different points in time; The specific steps for detecting behavioral inflection points are as follows: First, the tangent space evolution matrix is ​​generated by calculating the variational equation. Then, orthogonal QR decomposition is used to iterate the evolution matrix to calculate the maximum Lyapunov exponent. ,like If a point is reached, it indicates that the system behavior has reached an inflection point, which is then marked as a behavioral inflection point. The time series of these inflection points then constitute the output inflection point marking sequence.

7. The intelligent nursing management system based on artificial intelligence according to claim 6, characterized in that: Generate personalized care pathways that include risk markers The specific steps are as follows: Based on the output inflection point marker sequence, the planning of the nursing pathway is optimized. First, a weighted directed graph is constructed, where each node represents a physiological state and a nursing action, and the edge weight represents the transition cost from one state to another, including risk and distance penalty. Then, the Monge-Kandrovich optimal transport method is used to solve for the optimal path. This path minimizes the cost from the initial state to the target state and takes into account the deviation penalty at the inflection point. The final output is an optimal nursing path, which includes the time point, nursing action and corresponding risk value.

8. The intelligent nursing management system based on artificial intelligence according to claim 7, characterized in that: In the strategy evolution implementation module, the strategy optimizer optimizes the nursing pathway. The optimization operation is as follows: The nursing pathway is modeled as a game system. By using the adversarial optimization of patient and environmental strategies, the patient's nursing pathway strategy is optimized through a minimax game process, and the Hamiltonian function is output.

9. The intelligent nursing management system based on artificial intelligence according to claim 8, characterized in that: Adjusting feature weighting coefficients The specific operation is as follows: After game optimization, a meta-reinforcement learning mechanism is adopted. By updating the policy parameters, the system can adapt to different situations. Based on the Hamiltonian function, the system generates enhanced samples and updates the policy. The policy is adjusted by projecting the latent space policy onto the Riemannian manifold, enabling the system to quickly adapt to new changes.

10. The intelligent nursing management system based on artificial intelligence according to claim 9, characterized in that: The specific steps for generating adaptive closed-loop control commands are as follows: After policy optimization, the system maps features to Lie algebras and adjusts the policy output by calculating feature weighting coefficients. This part combines the policy's sensitivity and risk factors to generate feature weighting coefficients to determine the policy's effectiveness. Finally, the system calculates the curvature scalar and adjusts the dynamic regularization intensity, generates control commands by calculating the feedback vector, and feeds these control commands back to the multi-source heterogeneous processing module to ensure that the system's adaptation latency is less than 15 minutes and to quickly respond to new user needs.

Images