Dynamic Monitoring Method and System for the Compliance of Medical Service Projects Based on Artificial Intelligence

Abstract

The present invention provides a method and system for dynamically monitoring the compliance of medical service items based on artificial intelligence, which relates to the technical field of medical service management. The method includes establishing a data correlation relationship matrix by using the non-equilibrium entropy theory, extracting risk features through a chaotic neural network, combining deep generative adversarial diagnosis and treatment modeling and heterogeneous medical network neural evolution analysis to generate a risk prediction spectrum, and constructing a time-series evolution model by using a causal inference decision tree to locate abnormal risk points. The present invention can realize the real-time monitoring and risk warning of the compliance of medical service items, improve the supervision efficiency, and reduce the medical risk.

Description

Technical Field

[0001] The present invention relates to the technical field of medical service management, and particularly to a method and system for dynamically monitoring the compliance of medical service items based on artificial intelligence. Background Art

[0002] With the continuous improvement of the complexity of medical service items, the monitoring of medical service compliance has become a key link in medical quality management. The existing monitoring of medical service compliance mainly adopts the methods of manual review and rule matching, and judges compliance by setting fixed monitoring indicators and thresholds. At the same time, some medical institutions have begun to use artificial intelligence technologies such as machine learning to analyze medical service data to identify and warn against non-compliant behaviors.

[0003] However, the existing technologies have the following problems: traditional rule matching methods are difficult to cope with the dynamic changes of medical service items, and fixed monitoring indicators cannot accurately reflect the actual operation risks; existing machine learning methods mainly focus on static features and lack the ability to dynamically analyze the risk evolution in the medical service process; the correlation and temporal characteristics between medical service items are not adequately considered, making it difficult to achieve systematic risk prediction; there is a lack of an immediate intervention mechanism for abnormal risks, and the risk prevention and control effect is limited.

[0004] In summary, it is urgent to solve the technical problems in the dynamic monitoring of medical service compliance. Specifically, by establishing a data correlation relationship matrix to achieve dynamic knowledge mapping, using deep generative adversarial networks and heterogeneous medical networks to extract diagnosis and treatment features, and constructing a risk evolution model based on causal inference decision trees, a complete set of risk identification, prediction, and intervention mechanisms can be formed, thereby improving the accuracy and real-time performance of medical service compliance monitoring. The present invention can solve the problems in the existing technologies. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for dynamically monitoring the compliance of medical service items based on artificial intelligence, which can solve the problems in the existing technologies.

[0006] In the first aspect of the embodiments of the present invention,

[0007] A method for dynamically monitoring the compliance of medical service items based on artificial intelligence is provided, including:

[0008] Using the non-equilibrium entropy theory to perform dynamic association mapping on medical service item data, establishing a data correlation relationship matrix, and obtaining project knowledge association data; based on the data correlation relationship matrix, using a chaotic neural network to extract the risk feature distribution in the project knowledge association data; using a Lyapunov stability analyzer to perform long-term trend prediction on the risk feature distribution, and generating a risk benchmark model including risk assessment weights and dynamic monitoring thresholds;

[0009] Using a deep generative adversarial diagnosis and treatment modeling device to extract diagnosis and treatment feature sequences from project operation data and construct a knowledge representation model; based on the knowledge representation model, using a heterogeneous medical network neural evolver to analyze the association strength and temporal characteristics among various diagnosis and treatment activities, generating a risk prediction spectrum after comparing with historical compliance records, and combining with a risk benchmark model, dynamically evolving and analyzing the risk prediction spectrum through a medical value dynamics framework to output real-time compliance analysis results;

[0010] Using a causal inference decision tree to construct a temporal evolution model of real-time compliance evaluation results, calculating the risk evolution gradient, dynamically decomposing the risk state matrix based on the risk evolution gradient, locating abnormal risk points, and generating risk warning information and risk intervention strategies according to the spatio-temporal distribution characteristics of the abnormal risk points; feeding back the spatio-temporal distribution characteristics, risk warning information and risk intervention strategies to the data association relationship matrix to update the medical service project knowledge association data.

[0011] In an alternative embodiment,

[0012] Using the non-equilibrium state entropy theory to perform dynamic association mapping on medical service project data, establishing a data association relationship matrix, and obtaining project knowledge association data; based on the data association relationship matrix, using a chaotic neural network to extract the risk feature distribution in the project knowledge association data; using a Lyapunov stability analyzer to perform long-term trend prediction on the risk feature distribution, and generating a risk benchmark model including risk assessment weights and dynamic monitoring thresholds:

[0013] Constructing an initial data set for medical service project data according to service type, execution time, operator, service object, and operation records; obtaining the non-equilibrium state entropy value between any two data items according to the probability distribution of each data item in the corresponding dimension in the initial data set, and calculating the association strength between the data items according to the non-equilibrium state entropy value and the smoothing factor to construct a data association relationship matrix; introducing a time window mechanism into the data association relationship matrix, and weighted combining the current time window's data association relationship matrix with the historical data association relationship matrix through a forgetting factor to generate the project knowledge association data of the medical service project;

[0014] Taking the association strength in the project knowledge association data as the neuron connection weight to construct a chaotic neural network, using a non-linear iterative mapping to process the neuron input signal to keep the chaotic neural network in a chaotic critical state, and reducing the dimension of the output of the chaotic neural network through a locality-sensitive hashing algorithm to obtain the risk feature distribution of the medical service project;

[0015] Convert the risk feature distribution into a dynamic system equation, calculate the Lyapunov exponents of the dynamic system equation in each feature dimension, where the Lyapunov exponents characterize the degree of divergence of the trajectories in the feature dimensions; perform exponential transformation and normalization on the Lyapunov exponents to generate risk assessment weights; combine the historical data statistics of each feature dimension with the corresponding Lyapunov exponents to generate dynamic monitoring thresholds, and combine the risk assessment weights and the dynamic monitoring thresholds to construct a risk benchmark model for medical service items.

[0016] In an alternative embodiment,

[0017] Construct a chaotic neural network with the association strength in the project knowledge association data as the neuron connection weights, use a non - linear iterative mapping to process the neuron input signals to keep the chaotic neural network in a chaotic critical state, and perform dimensionality reduction on the output of the chaotic neural network through the locality - sensitive hashing algorithm. The obtained risk feature distribution of the medical service item includes:

[0018] Construct the association strength of the project knowledge association data as the neuron connection weights, combine the non - linear iterative mapping to control the chaotic critical state of the network, complete the dynamic update of the weights through the chaotic adaptive adjustment mechanism, monitor the network synchronization degree using the chaotic resonance detection mechanism, and obtain the risk feature distribution using the local similarity mapping. Specifically, it includes:

[0019] Construct the initial connection weight matrix of the chaotic neural network according to the association strength in the project knowledge association data; input the input signals of adjacent neurons and the initial connection weight matrix into the signal accumulation unit, calculate the cumulative state value at the current moment according to the attenuation coefficient, process the cumulative state value through non - linear iterative mapping and compare it with the activation threshold. When the cumulative state value is greater than the activation threshold, trigger neuron activation and enter the transient inhibition period; calculate the positive activation energy and negative inhibition energy and convert them into chaotic adjustment signals to keep the chaotic neural network in a chaotic critical state;

[0020] Construct a chaotic adaptive adjustment module, obtain the activation timing information of adjacent neurons, calculate the weight adjustment increment and dynamically fuse it with the initial connection weight matrix to update the connection weight matrix; adjust the corresponding sensitivity threshold according to the activation state of the neurons, and adjust the signal transmission intensity based on the overall synchronization degree of the chaotic neural network;

[0021] Set up a chaotic resonance detection module, calculate the synchronization degree based on the activation patterns of neurons in the chaotic neural network. When the synchronization degree is greater than the preset steady - state threshold, enter the feature extraction state and generate a feature vector;

[0022] Perform dimensionality reduction mapping on the feature vector using the locality - sensitive hashing algorithm to obtain the risk feature distribution of the medical service item.

[0023] In an alternative embodiment,

[0024] Utilize a deep generative adversarial diagnosis and treatment modeling device to extract a diagnosis and treatment feature sequence from project operation data and construct a knowledge representation model; based on the knowledge representation model, use a heterogeneous medical network neural evolver to analyze the association strength and temporal characteristics between various diagnosis and treatment activities, generate a risk prediction spectrum after comparing with historical compliance records, and combine with a risk benchmark model to perform dynamic evolution analysis on the risk prediction spectrum through a medical value dynamics framework, and output real-time compliance analysis results including:

[0025] Construct a deep generative adversarial diagnosis and treatment modeling device, set a residual connection structure in the generator network of the deep generative adversarial diagnosis and treatment modeling device, set a feature fusion gating unit in each connection layer of the residual connection structure, and adaptively fuse features at different levels through the feature fusion gating unit to obtain fused features; set multiple convolution kernels with different scales in the discriminator network of the deep generative adversarial diagnosis and treatment modeling device, and perform multi-scale decomposition on the fused features by using grouped convolution to obtain decomposed features; input medical service project operation data into the deep generative adversarial diagnosis and treatment modeling device, extract a diagnosis and treatment feature sequence based on the decomposed features, group the diagnosis and treatment feature sequence according to temporal characteristics, behavior characteristics, and resource characteristics, and construct a knowledge representation model;

[0026] Map the diagnosis and treatment feature sequence in the knowledge representation model to the feature nodes of a heterogeneous medical network, and construct association edges based on the temporal relationship, behavior association, and resource dependence of the feature nodes; calculate the node importance of the feature nodes as the starting probability of random walk, and perform temporal walk to obtain a multi-hop walk path; calculate the node transition probability matrix according to the multi-hop walk path, and construct a diagnosis and treatment activity association strength matrix;

[0027] Perform temporal alignment on the diagnosis and treatment activity association strength matrix and historical compliance records to obtain aligned data, input the aligned data into a temporal convolutional network with a dilated convolutional structure, extract multi-scale temporal features and calculate a risk propagation coefficient, and generate a risk prediction spectrum; construct a medical value dynamics equation set based on the risk prediction spectrum, set a risk propagation function and a value loss function in the medical value dynamics equation set, the risk propagation function describes the diffusion process of risk between the feature nodes according to the risk propagation coefficient, and the value loss function quantifies the impact of risk on the value of medical services; solve the medical value dynamics equation set to obtain the dynamic evolution trajectory of the risk state, and output real-time compliance analysis results based on the dynamic evolution trajectory.

[0028] In an alternative embodiment,

[0029] Calculate the node importance of the feature nodes as the starting probability of random walk, and perform temporal walk to obtain multi-hop walk paths, including:

[0030] Calculate the comprehensive importance by fusing the community structure features and historical activity sequence features of the feature nodes as the starting probability of random walk; during the temporal walk, dynamically update the transition probability based on the local importance and temporal importance of the nodes, and introduce a path sampling and skeleton extraction mechanism to generate multi-hop walk paths that meet the temporal transition constraints, specifically including:

[0031] Obtain the connection relation matrix of the feature nodes, decompose the connection relation matrix using spectral clustering method to obtain the eigenvector matrix, and identify the node community structure based on the eigenvector matrix; calculate the structural density between node pairs within the node community structure to obtain the density matrix, and calculate the propagation ability by combining the in-degree and out-degree distributions of the nodes to obtain the propagation matrix; calculate the local importance of the nodes according to the density matrix and the propagation matrix;

[0032] Collect the historical activity sequences of the feature nodes, perform time series decomposition on the historical activity sequences to obtain periodic components, construct a temporal template based on the periodic components, match the current activity sequence of the feature nodes with the temporal template to obtain the matching degree distribution, and calculate the temporal importance of the feature nodes according to the matching degree distribution;

[0033] Input the local importance and the temporal importance into a feature fusion network, the feature fusion network dynamically adjusts the fusion weights based on an error feedback mechanism, outputs the comprehensive importance of the feature nodes, and normalizes it to obtain the starting probability distribution of random walk;

[0034] Calculate the forward transition probability and backward transition probability between nodes according to the starting probability distribution, and construct a temporal transition probability matrix based on the forward transition probability and the backward transition probability; initialize the non-homogeneous temporal walk strategy with the temporal transition probability matrix, and update the temporal transition probability matrix based on the local importance and temporal importance of the current node in each walk step; execute the non-homogeneous temporal walk strategy to obtain an initial walk path, perform segmented sampling on the initial walk path to obtain a sampling path set, extract temporal key points from the sampling path set to determine the path skeleton, and generate multi-hop walk paths based on the path skeleton, where the multi-hop walk paths contain node jump sequences that meet the temporal transition constraints.

[0035] In an alternative embodiment,

[0036] Construct a time-series evolution model of real-time compliance evaluation results using a causal inference decision tree, calculate the risk evolution gradient, dynamically decompose the risk state matrix based on the risk evolution gradient, locate abnormal risk points, and generate risk warning information and risk intervention strategies according to the spatio-temporal distribution characteristics of the abnormal risk points; feedback the spatio-temporal distribution characteristics, risk warning information and risk intervention strategies to the data association relationship matrix to update the knowledge association data of medical service items, including:

[0037] Obtain the risk state data in the medical service process and extract the risk state characteristics. According to the hierarchical relationship of the risk state characteristics, construct the nodes of the causal inference decision tree, establish the transfer connections between the nodes based on the time-series change relationship of the risk state data, and generate a multi-level causal inference decision tree; calculate the state transfer weights between adjacent nodes in the multi-level causal inference decision tree, normalize the state transfer weights to obtain the state transfer probability, and construct a risk state transfer matrix.

[0038] Obtain the risk state transfer matrices of adjacent time steps, calculate the difference between the risk state transfer matrices, and obtain the risk evolution gradient according to the ratio of the difference to the time interval; perform singular value decomposition on the risk evolution gradient to obtain the eigenmatrix, and identify the abnormal risk points of the risk situation mutation based on the singular values in the eigenmatrix.

[0039] Extract the time span and spatial distribution information of the abnormal risk points, construct a spatio-temporal feature vector in combination with the mutation degree of the abnormal risk points, determine the risk warning level according to the spatio-temporal feature vector, and generate a risk intervention strategy based on the risk warning level and the spatio-temporal feature vector.

[0040] Input the spatio-temporal feature vector, the risk warning level and the risk intervention strategy into the feature fusion function, update the data association relationship matrix according to the output result of the feature fusion function, and optimize the knowledge association data of medical service items based on the data association relationship matrix.

[0041] In an alternative embodiment,

[0042] Obtain the risk state data in the medical service process and extract the risk state characteristics. According to the hierarchical relationship of the risk state characteristics, construct the nodes of the causal inference decision tree, establish the transfer connections between the nodes based on the time-series change relationship of the risk state data, and generate a multi-level causal inference decision tree, including:

[0043] By performing standardization and dimensionality reduction processing on the medical service risk state data, using an autoencoder network to extract the latent feature vector and calculate the risk state similarity, clustering based on the similarity to form risk state nodes, establishing the transfer connections between the nodes according to the time-series change relationship, and optimizing the connection structure of the decision tree using the maximum information coefficient, specifically including:

[0044] Obtain risk status data during the medical service process and extract risk status features, perform standardization processing on the risk status features to obtain standardized features, and use the principal component analysis method to perform dimensionality reduction processing on the standardized features to obtain dimensionality-reduced risk status features;

[0045] Input the dimensionality-reduced risk status features into an autoencoder network. The encoder part of the autoencoder network maps the dimensionality-reduced risk status features to latent feature vectors, and the decoder part of the autoencoder network reconstructs the dimensionality-reduced risk status features based on the latent feature vectors. Calculate the Euclidean distance between risk statuses based on the latent feature vectors to obtain risk status similarity;

[0046] Perform clustering operations based on the risk status similarity. According to a preset similar similarity threshold, classify the corresponding risk statuses into a risk status cluster to obtain multiple risk status clusters at different levels, and map each risk status cluster to a risk status node of a causal inference decision tree; Based on the temporal changes of the risk statuses in the risk status cluster, calculate the temporal correlation strength between the risk status nodes. The temporal correlation strength is determined according to the jump frequency of the risk status in the time series;

[0047] Establish a directed connection relationship between the risk status nodes according to the temporal correlation strength, establish a transfer connection between the risk status node pairs with the temporal correlation strength greater than the preset strength threshold, and construct an initial structure of a multi-level causal inference decision tree based on the risk status nodes and the transfer connections;

[0048] Use the maximum information coefficient method to calculate the causal relationship strength between the risk status nodes, optimize the initial structure of the multi-level causal inference decision tree based on the causal relationship strength, delete the transfer connections with the causal relationship strength lower than the preset causal threshold, and retain the transfer connections higher than the preset causal threshold to obtain an optimized multi-level causal inference decision tree.

[0049] In the second aspect of the embodiments of the present invention,

[0050] Provide an artificial intelligence-based dynamic monitoring system for the compliance of medical service items, including:

[0051] A first unit for dynamically associating and mapping medical service item data using the non-equilibrium entropy theory, establishing a data association relationship matrix, and obtaining project knowledge association data; Based on the data association relationship matrix, use a chaotic neural network to extract the risk feature distribution in the project knowledge association data; Use a Lyapunov stability analyzer to perform long-term trend prediction on the risk feature distribution to generate a risk benchmark model including risk assessment weights and dynamic monitoring thresholds;

[0052] A second unit is configured to extract a sequence of diagnosis and treatment features from project operation data by using a deep generative adversarial diagnosis and treatment modeling device, and construct a knowledge representation model; based on the knowledge representation model, an heterogeneous medical network neural evolver is adopted to analyze the association strength and temporal characteristics among various diagnosis and treatment activities, generate a risk prediction spectrum after comparing with historical compliance records, and combine with a risk benchmark model to perform dynamic evolution analysis on the risk prediction spectrum through a medical value dynamics framework, and output a real-time compliance analysis result;

[0053] A third unit is configured to construct a temporal evolution model of the real-time compliance assessment result by using a causal inference decision tree, calculate a risk evolution gradient, dynamically decompose a risk state matrix based on the risk evolution gradient, locate abnormal risk points, and generate a risk warning message and a risk intervention strategy according to the spatio-temporal distribution characteristics of the abnormal risk points; feedback the spatio-temporal distribution characteristics, the risk warning message and the risk intervention strategy to a data association relationship matrix, and update the knowledge association data of medical service items.

[0054] In a third aspect of the embodiments of the present invention,

[0055] a kind of electronic device is provided, including:

[0056] a processor;

[0057] a memory for storing instructions executable by the processor;

[0058] wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0059] In a fourth aspect of the embodiments of the present invention,

[0060] a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0061] In the embodiments of the present invention, through dynamic association mapping and risk feature extraction, the compliance of medical service items can be monitored in real time, and the safety and effectiveness of medical services can be improved; by using deep generative adversarial modeling and heterogeneous medical network analysis, the correlation and risks among diagnosis and treatment activities can be accurately identified, the allocation of medical resources can be optimized, and the compliance risk can be reduced; through a temporal evolution model and a risk warning mechanism, abnormal risk points can be discovered in time and corresponding intervention strategies can be formulated, and the adaptability and continuous improvement ability of medical services can be enhanced. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 is a schematic flowchart of a method for dynamically monitoring the compliance of medical service items based on artificial intelligence according to an embodiment of the present invention;

[0063] Figure 2This is a schematic structural diagram of the compliance dynamic monitoring system for medical service items based on artificial intelligence in the embodiments of the present invention. Specific embodiments

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only a part rather than all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0065] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0066] Figure 1 This is a schematic flow diagram of the compliance dynamic monitoring method for medical service items based on artificial intelligence in the embodiments of the present invention. As Figure 1 shown, the method includes:

[0067] S101. Dynamically associate and map medical service item data using the non-equilibrium entropy theory, establish a data association relationship matrix, and obtain project knowledge association data; based on the data association relationship matrix, use a chaotic neural network to extract the risk feature distribution in the project knowledge association data; use a Lyapunov stability analyzer to predict the long-term trend of the risk feature distribution, and generate a risk benchmark model including risk assessment weights and dynamic monitoring thresholds;

[0068] In this embodiment, by constructing a data association matrix using the non-equilibrium entropy theory, the complex dynamic relationships hidden in medical service items can be captured, and the data utilization efficiency can be improved; using a chaotic neural network to accurately extract the risk feature distribution, enhancing the detection ability for non-linear and complex risk patterns; realizing long-term trend prediction through Lyapunov stability analysis, and constructing a benchmark model including risk weights and dynamic monitoring thresholds to provide timely early warnings and decision-making support for medical risk management.

[0069] S102. Use a deep generative adversarial diagnosis and treatment modeling device to extract the diagnosis and treatment feature sequence from the project operation data, and construct a knowledge representation model; based on the knowledge representation model, use a heterogeneous medical network neural evolver to analyze the association strength and temporal characteristics between various diagnosis and treatment activities, generate a risk prediction spectrum after comparing with historical compliance records, and combine it with the risk benchmark model. Through the medical value dynamics framework, perform dynamic evolution analysis on the risk prediction spectrum, and output the real-time compliance analysis result;

[0070] In this embodiment, a deep generative adversarial diagnosis and treatment modeling device is used to extract diagnosis and treatment feature sequences from project data, and a unified knowledge representation model is constructed, improving the ability to capture and express key diagnosis and treatment information; through a heterogeneous medical network neural evolver, the association strength and temporal characteristics between various diagnosis and treatment activities are deeply analyzed, and the complex interaction relationship between medical activities can be more comprehensively reflected; a risk prediction spectrum is generated in combination with historical compliance records and docked with a risk benchmark model, realizing the dynamic evolution analysis of medical risks and providing forward-looking data support for risk early warning; finally, real-time compliance analysis results are output through a medical value dynamics framework, which helps to detect anomalies in a timely manner and ensure the standardization and safety of medical processes.

[0071] S103. Use a causal inference decision tree to construct a temporal evolution model of real-time compliance evaluation results, calculate the risk evolution gradient, dynamically decompose the risk state matrix based on the risk evolution gradient, locate abnormal risk points, and generate risk early warning information and risk intervention strategies according to the spatio-temporal distribution characteristics of the abnormal risk points; feedback the spatio-temporal distribution characteristics, risk early warning information and risk intervention strategies to the data association relationship matrix to update the knowledge association data of medical service items.

[0072] In this embodiment, through the temporal evolution model constructed by the causal inference decision tree, the real-time tracking of the compliance evaluation results and the capture of the risk evolution trend are realized, enhancing the risk dynamic monitoring ability; using the risk evolution gradient to dynamically decompose the risk state matrix can accurately locate abnormal risk points, providing a clear basis for early warning and intervention; automatically generating risk early warning information and intervention strategies according to the spatio-temporal distribution characteristics of abnormal risk points improves the response speed and accuracy of risk management; feedbacking the spatio-temporal distribution characteristics, warning information and intervention strategies to the data association relationship matrix realizes the dynamic update of the knowledge association data of medical service items, promoting the continuous optimization and upgrading of the risk management system.

[0073] In an alternative implementation manner, the non-equilibrium state entropy theory is used to perform dynamic association mapping on medical service item data, establish a data association relationship matrix, and obtain project knowledge association data; based on the data association relationship matrix, a chaotic neural network is used to extract the risk feature distribution in the project knowledge association data; the Lyapunov stability analyzer is used to predict the long-term trend of the risk feature distribution, and a risk benchmark model including risk assessment weights and dynamic monitoring thresholds is generated, including:

[0074] Construct an initial data set for medical service item data according to service type, execution time, operator, service object, and operation record; obtain the non-equilibrium entropy value between any two data items according to the probability distribution of each data item in the initial data set on the corresponding dimension, calculate the association strength between the data items according to the non-equilibrium entropy value and the smoothing factor, and construct a data association relationship matrix; introduce a time window mechanism into the data association relationship matrix, and weight and combine the data association relationship matrix of the current time window and the historical data association relationship matrix through a forgetting factor to generate item knowledge association data for medical service items;

[0075] Use the association strength in the item knowledge association data as the neuron connection weight to construct a chaotic neuron network, adopt a non-linear iterative mapping to process the neuron input signal to keep the chaotic neuron network in the chaotic critical state, and reduce the dimension of the output of the chaotic neuron network through the locality-sensitive hashing algorithm to obtain the risk feature distribution of medical service items;

[0076] Convert the risk feature distribution into a dynamic system equation, calculate the Lyapunov exponent of the dynamic system equation on each feature dimension, and the Lyapunov exponent characterizes the degree of divergence of the trajectory of the feature dimension; perform exponential transformation and normalization processing on the Lyapunov exponent to generate a risk assessment weight; combine the historical data statistics value of each feature dimension and the Lyapunov exponent of the corresponding dimension to generate a dynamic monitoring threshold, and combine the risk assessment weight and the dynamic monitoring threshold to construct a risk benchmark model for medical service items.

[0077] In a specific implementation, first, construct an initial data set for medical service items. Collect data related to medical service items, including information such as service type, execution time, operator, service object, and detailed operation records. For example, all surgical records of a hospital in one year can be collected, including surgical type, surgical time, surgeon, anesthesiologist, patient information, and various index records during the operation.

[0078] Next, calculate the association strength between the data items and construct a data association relationship matrix. Calculate the non-equilibrium entropy value between any two data items according to the probability distribution of each data item in the initial data set on different dimensions. The non-equilibrium entropy value reflects the degree of difference between the two data items. For example, the non-equilibrium entropy value between two surgical records of the same surgical type is relatively small, while the non-equilibrium entropy value between two surgical records of different surgical types is relatively large. Then, calculate the association strength between the data items according to the non-equilibrium entropy value and the preset smoothing factor. The smoothing factor is used to adjust the sensitivity of the association strength. The higher the association strength, the stronger the association between the two data items. Combine the association strengths between all data items into a matrix, that is, the data association relationship matrix.

[0079] Introduce a time window mechanism to generate project knowledge association data. To capture the dynamic association relationships between medical service project data, a time window mechanism is introduced. The data association relationship matrix of the current time window and the historical data association relationship matrix are weighted and combined. For example, the association relationship matrix of the most recent month can be weighted and combined with the association relationship matrix of the past three months, and the weights can be set according to the proximity in time, with the greater weight for the more recent time. The matrix after weighted combination is the project knowledge association data of medical service projects.

[0080] Construct a chaotic neural network to extract the risk feature distribution. Use the association strength in the project knowledge association data as the connection weights between neurons to construct a chaotic neural network. Input the medical service project data into the chaotic neural network, and use a non-linear iterative mapping to process the neuron input signals to keep the chaotic neural network at the chaotic critical state. The neural network at the chaotic critical state has higher sensitivity and information processing ability. Then, use the locality-sensitive hashing algorithm to reduce the dimension of the output of the chaotic neural network to obtain the risk feature distribution of medical service projects. For example, the high-dimensional output data can be reduced to two or three dimensions for subsequent analysis and processing.

[0081] Conduct Lyapunov stability analysis to generate risk assessment weights and dynamic monitoring thresholds. Transform the risk feature distribution into a dynamic system equation, and calculate the Lyapunov exponents of the dynamic system equation in each feature dimension. The Lyapunov exponents characterize the divergence degree of the trajectories in the feature dimensions and can be used to evaluate the long-term trend of risks. Perform exponential transformation and normalization on the Lyapunov exponents to generate risk assessment weights. For example, the Lyapunov exponents can be exponentiated and then normalized to obtain the risk assessment weights for each feature dimension. Combine the historical data statistical values of each feature dimension and the corresponding Lyapunov exponents to generate dynamic monitoring thresholds. For example, the dynamic monitoring thresholds can be set according to the mean and standard deviation of the historical data, combined with the Lyapunov exponents.

[0082] Finally, construct a risk benchmark model for medical service projects. Combine the risk assessment weights and dynamic monitoring thresholds to construct a risk benchmark model for medical service projects. This model can be used to evaluate the risk levels of medical service projects and monitor the risks in real time.

[0083] In this embodiment, through the non-equilibrium entropy theory and the chaotic neural network, the risk characteristics in the medical service item data can be captured more comprehensively, improving the accuracy of risk assessment; through the time window mechanism and the Lyapunov stability analysis, the dynamic monitoring of the medical service item risk can be realized, and potential risks can be detected in a timely manner; the constructed risk benchmark model can provide decision-making support for medical institutions in risk management and help them formulate effective risk prevention and control measures.

[0084] In an alternative implementation, the association strength in the project knowledge association data is used as the neuron connection weight to construct a chaotic neural network. The neuron input signal is processed by a non-linear iterative mapping to keep the chaotic neural network in the chaotic critical state. The output of the chaotic neural network is dimension-reduced by the locality-sensitive hashing algorithm, and the risk characteristic distribution of the medical service item is obtained, including:

[0085] The association strength of the project knowledge association data is constructed as the neuron connection weight, combined with the non-linear iterative mapping to control the chaotic critical state of the network. The weight dynamic update is completed through the chaotic adaptive regulation mechanism, the network synchronization degree is monitored by the chaotic resonance detection mechanism, and the risk characteristic distribution is obtained by using the local similarity mapping, specifically including:

[0086] Construct the initial connection weight matrix of the chaotic neural network according to the association strength in the project knowledge association data; input the input signals of adjacent neurons and the initial connection weight matrix into the signal accumulation unit, calculate the cumulative state value at the current moment according to the attenuation coefficient, process the cumulative state value through the non-linear iterative mapping and compare it with the activation threshold. When the cumulative state value is greater than the activation threshold, the neuron is activated and enters the transient inhibition period; calculate the positive activation energy and the negative inhibition energy and convert them into chaotic regulation signals to keep the chaotic neural network in the chaotic critical state;

[0087] Construct a chaotic adaptive regulation module, obtain the activation timing information of adjacent neurons, calculate the weight adjustment increment and dynamically fuse it with the initial connection weight matrix to update the connection weight matrix; adjust the corresponding sensitivity threshold according to the activation state of the neurons, and adjust the signal transmission intensity based on the overall synchronization degree of the chaotic neural network;

[0088] Set a chaotic resonance detection module, calculate the synchronization degree based on the activation patterns of the neurons in the chaotic neural network. When the synchronization degree is greater than the preset steady state threshold, enter the feature extraction state and generate a feature vector;

[0089] Use the locality-sensitive hashing algorithm to perform dimension reduction mapping on the feature vector to obtain the risk characteristic distribution of the medical service item.

[0090] In a specific embodiment, the construction process of the chaotic neural network first requires extracting detailed association data between items from the medical service database. This data includes information such as the order of medical treatment operations, medication compatibility rules, treatment plan combinations, etc. The system calculates the association frequency between each pair of items and converts it into a standardized association strength value. These association strength values are filled into the initial connection weight matrix. After normalizing the weight matrix, all weight values are mapped to the interval from zero to one as the initial connection strength of the neural network.

[0091] In the signal processing stage, each neuron continuously receives input signals from adjacent neurons. These signals are multiplied by the corresponding connection weights to obtain weighted inputs. The signal accumulation unit performs time integration on these weighted inputs and introduces a decay factor to simulate the actual signal decay process. When the accumulated value exceeds the preset activation threshold, the neuron generates an output pulse and immediately enters the refractory period. The output pulse is transmitted to adjacent neurons through synaptic connections. The system will record in detail the activation time and pulse interval information of each neuron for subsequent analysis.

[0092] To maintain the dynamic balance of the network, the system calculates the average activation frequency of all neurons in the network and the synchronous activation degree of the neuron population in real time. Based on these metrics, the system adjusts the chaotic control parameters and uses positive and negative feedback mechanisms to keep the network always at the critical state of chaos. This state not only ensures the sensitivity of the network to input signals but also avoids the network falling into local optima.

[0093] In the weight dynamic adjustment link, the system continuously records the activation time sequences of each pair of adjacent neurons and extracts neuron pairs with significant time sequence associations by calculating the correlation between the activation time sequences. For these highly correlated neuron pairs, the system strengthens the connection weights between them. The weight update calculates the adjustment amount using the Hebbian learning rule. By setting appropriate adjustment steps and learning rates, the adjustment amount is weighted and fused with the original weights to achieve the dynamic update of the weight matrix.

[0094] At the same time, the system monitors the changes in the activation frequency of each neuron and dynamically adjusts the sensitivity thresholds of each neuron according to these changes. For those neurons that are frequently activated, the system will correspondingly increase their sensitivity; while for neurons that are in an inactive state for a long time, their sensitivity will be reduced. This adaptive mechanism ensures the network's ability to quickly respond to abnormal patterns.

[0095] In the feature extraction stage, the system evaluates the overall synchronization degree of the network by calculating the average phase difference of the neuron population and the temporal aggregation degree of activation. When the synchronization degree reaches the preset condition, the system starts to extract the feature of the activation pattern of neurons, including the weight distribution of synaptic connections and the spatial correlation feature of neuron groups, and combines these features to form a high-dimensional feature vector. Finally, by constructing a family of locality-sensitive hashing functions, the feature vector is reduced in dimension to generate a low-dimensional risk feature representation, and finally a risk distribution map is output.

[0096] Exemplarily, taking the antibiotic use monitoring in a certain Class-III Grade-A hospital as a specific application scenario: The system first collects all medical data related to antibiotic use, including doctor's order prescribing records, pharmacy dispensing records, nursing execution records, etc. By analyzing these data, the system constructs a neuron network with twenty key monitoring points and sets the initial connection weights based on the business correlation degree. When the system detects the situation of frequent modification of antibiotic doctor's orders at night, the activation frequency of the corresponding neurons increases significantly, triggering the weight adaptive adjustment mechanism. The system then increases the monitoring sensitivity of the doctor's order modification link and strengthens the connection weight with the doctor's order review link. By tracking the execution situation after the doctor's order is modified, the system can analyze the abnormal pattern of drug consumption and detect the temporal change of the nursing execution record. Finally, based on the time synchronization degree, spatial aggregation, and process connection continuity of each link operation, the system generates a detailed antibiotic use compliance risk distribution map and accurately marks the high-risk time periods and operation links.

[0097] In this embodiment, using the project knowledge association data to construct the initial weight matrix provides a real dynamic association basis for the neuron network, thus more accurately reflecting the complex interaction relationship between medical projects; through the signal accumulation unit and the non-linear iterative mapping, the real-time monitoring of the neuron activation state is realized, ensuring that when the accumulated signal exceeds the threshold, activation is quickly triggered and enters transient inhibition, improving the sensitivity and response speed of the system to risk features; calculating and converting the positive activation energy and negative inhibition energy into chaotic adjustment signals effectively keeps the network in the chaotic critical state and enhances the system's ability to process complex and non-linear risk information; the chaotic adaptive adjustment module automatically adjusts the connection weight and sensitivity threshold according to the activation timing information, enabling the model to have an adaptive adjustment function and improving the overall robustness and adaptability; through chaotic resonance detection and the locality-sensitive hashing algorithm, the system can extract key features when the neuron synchronization reaches a steady state and perform dimensionality reduction mapping on the high-dimensional feature vector, so as to obtain a clear risk feature distribution, providing a reliable basis for subsequent risk assessment.

[0098] In an alternative embodiment, a deep generative adversarial diagnosis and treatment modeling device is used to extract a diagnosis and treatment feature sequence from project operation data to construct a knowledge representation model; based on the knowledge representation model, a heterogeneous medical network neural evolver is used to analyze the association strength and temporal characteristics between various diagnosis and treatment activities, generate a risk prediction spectrum after comparing with historical compliance records, and combine with a risk benchmark model to perform dynamic evolution analysis on the risk prediction spectrum through a medical value dynamics framework, and the output real-time compliance analysis result includes:

[0099] Construct a deep generative adversarial diagnosis and treatment modeling device. A residual connection structure is set in the generator network of the deep generative adversarial diagnosis and treatment modeling device, and a feature fusion gating unit is set in each connection layer of the residual connection structure. Adaptive fusion of features at different levels is performed through the feature fusion gating unit to obtain fused features; multiple convolution kernels of different scales are set in the discriminator network of the deep generative adversarial diagnosis and treatment modeling device, and grouped convolution is used to perform multi-scale decomposition on the fused features to obtain decomposed features; input the operation data of medical service projects into the deep generative adversarial diagnosis and treatment modeling device, extract the diagnosis and treatment feature sequence based on the decomposed features, group the diagnosis and treatment feature sequence according to temporal characteristics, behavior characteristics and resource characteristics, and construct a knowledge representation model;

[0100] Map the diagnosis and treatment feature sequence in the knowledge representation model to the feature nodes of a heterogeneous medical network, and construct association edges based on the temporal relationship, behavior association and resource dependence of the feature nodes; calculate the node importance of the feature nodes as the starting probability of random walk, and perform temporal walk to obtain a multi-hop walk path; calculate the node transition probability matrix according to the multi-hop walk path, and construct a diagnosis and treatment activity association strength matrix;

[0101] Perform temporal alignment on the diagnosis and treatment activity association strength matrix and historical compliance records to obtain aligned data, input the aligned data into a temporal convolutional network with a dilated convolution structure, extract multi-scale temporal features and calculate the risk propagation coefficient to generate a risk prediction spectrum; construct a medical value dynamics equation set based on the risk prediction spectrum, and set a risk propagation function and a value loss function in the medical value dynamics equation set. The risk propagation function describes the diffusion process of risk between the feature nodes according to the risk propagation coefficient, and the value loss function quantifies the impact of risk on the value of medical services; solve the medical value dynamics equation set to obtain the dynamic evolution trajectory of the risk state, and output the real-time compliance analysis result based on the dynamic evolution trajectory.

[0102] In a specific implementation, first, a deep generative adversarial diagnosis and treatment modeling device is constructed. This modeling device consists of two parts: a generator and a discriminator. A residual connection structure is set in the generator network, and a feature fusion gating unit is set in each connection layer. The role of the generator is to convert a random noise vector into a simulated diagnosis and treatment feature sequence. The residual connection structure helps train deeper networks, while the feature fusion gating unit can adaptively fuse features at different levels, thereby improving the quality of the generated sequence. For example, the gating unit can learn a weight based on the features of the current layer and the features of the previous layer to control the fusion ratio between the two.

[0103] Multiple convolution kernels of different scales are set in the discriminator network, and group convolution is used to perform multi-scale decomposition on the fused features output by the generator. The role of the discriminator is to distinguish between real diagnosis and treatment feature sequences and the simulated sequences generated by the generator. The multi-scale convolution kernels can capture features at different granularities, while group convolution can reduce the computational amount and improve the generalization ability of the model. For example, convolution kernels of sizes 3×3, 5×5, and 7×7 can be used to extract features at different scales.

[0104] The real operation data of medical service items is input into the deep generative adversarial diagnosis and treatment modeling device for training. Through adversarial training, the generator and the discriminator continuously improve their respective capabilities, and finally the generator can generate diagnosis and treatment feature sequences that are indistinguishable from the real ones. After the training is completed, the discriminator is used to extract the decomposed features, and the diagnosis and treatment feature sequences are extracted based on the decomposed features.

[0105] The extracted diagnosis and treatment feature sequences are grouped according to temporal features, behavioral features, and resource features to construct a knowledge representation model. For example, temporal features can include the occurrence time and duration of diagnosis and treatment activities; behavioral features can include specific behaviors such as diagnosis, treatment, and examination; resource features can include drugs, devices, and human resources used.

[0106] The heterogeneous medical network neural evolver analyzes the associations of diagnosis and treatment activities:

[0107] The diagnosis and treatment feature sequences in the knowledge representation model are mapped to the feature nodes of the heterogeneous medical network. For example, each diagnosis and treatment activity can be used as a node, and the attributes of the node can include temporal features, behavioral features, and resource features.

[0108] Association edges are constructed based on the temporal relationships, behavioral associations, and resource dependencies of the feature nodes. For example, if two diagnosis and treatment activities are closely connected in time, an edge can be established between them; if two diagnosis and treatment activities use the same resources, an edge can also be established between them.

[0109] Calculate the node importance of feature nodes as the starting probability of random walk. The node importance can be calculated according to indicators such as node degree and centrality. Perform temporal walk to obtain multi-hop walk paths. For example, one can start from an important node and perform random walk along the edges in the network, recording the sequence of nodes passed through.

[0110] Calculate the node transition probability matrix based on the multi-hop walk paths and construct the correlation strength matrix of medical treatment activities. For example, if two nodes often appear together in the multi-hop walk paths, the correlation strength between them is high.

[0111] Risk prediction spectrum generation and analysis of the medical value dynamics framework:

[0112] Align the correlation strength matrix of medical treatment activities with historical compliance records in time series. For example, one can match the current medical treatment activity sequence with the sequences in the historical compliance records to find similar sequences.

[0113] Input the aligned data into a temporal convolutional network with a dilated convolutional structure. Dilated convolution can expand the receptive field, thereby capturing longer-term temporal dependencies. Extract multi-scale temporal features and calculate the risk propagation coefficient. The risk propagation coefficient can represent the possibility of risk propagation between different medical treatment activities.

[0114] Generate a risk prediction spectrum. The risk prediction spectrum can represent the risk levels of different medical treatment activities.

[0115] Construct a medical value dynamics equation set based on the risk prediction spectrum. Set a risk propagation function and a value loss function in the equation set. The risk propagation function describes the diffusion process of risk between feature nodes according to the risk propagation coefficient. The value loss function quantifies the impact of risk on the value of medical services. For example, if the risk of a medical treatment activity is very high, it may reduce the value of the entire medical service.

[0116] Solve the medical value dynamics equation set to obtain the dynamic evolution trajectory of the risk state. For example, one can simulate the process of risk propagation in the network and predict the risk levels of different medical treatment activities in the future for a period of time.

[0117] Output the real-time compliance analysis results based on the dynamic evolution trajectory. For example, one can judge whether the current medical treatment plan is compliant according to the risk state and give corresponding suggestions.

[0118] Data Case: Suppose a patient undergoes three medical treatment activities: A, B, and C. A and B are closely connected in time and both use the same resources; C has no direct association with A and B. The medical treatment feature sequences are extracted by a deep generative adversarial medical treatment modeling device, and a knowledge representation model is constructed. Then, A, B, and C are mapped to the nodes of a heterogeneous medical network, and an edge is established between A and B. The node transition probability matrix is calculated by random walk, and a medical treatment activity association intensity matrix is constructed. Finally, the association intensity matrix is compared with the historical compliance records to generate a risk prediction spectrum.

[0119] In this embodiment, by extracting medical treatment features through a deep learning model and combining heterogeneous network analysis and a medical value dynamics framework, the risks of medical treatment activities can be predicted more accurately, improving the accuracy of compliance analysis; the compliance of medical treatment activities can be analyzed in real time, and risk warnings can be given in a timely manner, which helps medical institutions take measures in a timely manner to avoid compliance risks; the propagation path and influence degree of risks among medical treatment activities can be clearly shown, improving the interpretability of compliance analysis and helping medical institutions better understand the causes of risks.

[0120] In an alternative embodiment, calculating the node importance of the feature nodes as the starting probability of random walk, and performing a temporal walk to obtain a multi-hop walk path includes:

[0121] Calculating the comprehensive importance by fusing the community structure features and historical activity sequence features of the feature nodes as the starting probability of random walk; in the process of performing temporal walk, dynamically updating the transition probability based on the local importance and temporal importance of the nodes, and introducing a path sampling and skeleton extraction mechanism to generate a multi-hop walk path that satisfies the temporal transition constraints, specifically including:

[0122] Obtaining the connection relationship matrix of the feature nodes, decomposing the connection relationship matrix by using a spectral clustering method to obtain a feature vector matrix, and identifying the node community structure based on the feature vector matrix; calculating the structural density between node pairs within the node community structure to obtain a density matrix, and calculating the propagation ability by combining the in-degree and out-degree distributions of the nodes to obtain a propagation matrix; calculating the local importance of the nodes according to the density matrix and the propagation matrix;

[0123] Collecting the historical activity sequences of the feature nodes, performing time series decomposition on the historical activity sequences to obtain periodic components, constructing a temporal template based on the periodic components, matching the current activity sequence of the feature nodes with the temporal template to obtain a matching degree distribution, and calculating the temporal importance of the feature nodes according to the matching degree distribution;

[0124] Input the local importance and the temporal importance into a feature fusion network. The feature fusion network dynamically adjusts the fusion weights based on an error feedback mechanism, outputs the comprehensive importance of the feature nodes, and normalizes it to obtain the initial probability distribution for random walk.

[0125] Calculate the forward transition probability and the backward transition probability between nodes according to the initial probability distribution, and construct a temporal transition probability matrix based on the forward transition probability and the backward transition probability; Initialize a non-homogeneous temporal random walk strategy with the temporal transition probability matrix. In each random walk step, update the temporal transition probability matrix based on the local importance and the temporal importance of the current node; Execute the non-homogeneous temporal random walk strategy to obtain an initial random walk path, perform segmented sampling on the initial random walk path to obtain a set of sampled paths, extract temporal key points from the set of sampled paths to determine the path skeleton, and generate a multi-hop random walk path based on the path skeleton. The multi-hop random walk path contains a node jump sequence that satisfies the temporal transition constraint.

[0126] In a specific embodiment, a method for temporal random walk of feature nodes is used to mine the dynamic association patterns of nodes in a complex network. This method fuses the static structural features and the dynamic temporal behavior features of nodes to generate a multi-hop random walk path containing time constraints.

[0127] First, obtain the connection relationship of feature nodes. For example, the adjacency matrix can be used to represent the connection relationship between nodes. Suppose there are 5 feature nodes, and the adjacency matrix can be represented as a 5×5 matrix, where the value of the matrix element is 1 indicating the existence of a connection between two nodes, and the value of 0 indicating the non-existence of a connection.

[0128] Then, use the spectral clustering method to decompose the connection relationship matrix to obtain a feature vector matrix. The spectral clustering method can divide nodes into different communities. Suppose the above 5 nodes are divided into two communities, {1, 2, 3} and {4, 5}, respectively, through spectral clustering.

[0129] Within each community, calculate the structural density between node pairs to obtain a density matrix. The structural density can be used to measure the tightness of the connection between nodes. For example, it can be represented by the number of common neighbors between node pairs. At the same time, calculate the propagation ability according to the in-degree and out-degree distribution of nodes to obtain a propagation matrix. The propagation ability can be used to measure the ability of nodes to spread information in the network. For example, it can be represented by the degree centrality of nodes. Combine the density matrix and the propagation matrix to calculate the local importance of nodes.

[0130] Next, collect the historical activity sequences of the feature nodes. For example, the activity data of each node over a period of time can be collected, such as the number of visits per hour. Perform time series decomposition on the historical activity sequences. For example, the time series can be decomposed into a trend component, a seasonal component, and a residual component. Based on the periodic component, such as the seasonal component, construct a time series template. Match the current activity sequence of the feature node with the time series template to obtain a matching degree distribution. For example, the dynamic time warping algorithm can be used to calculate the similarity between the current activity sequence and the time series template. Calculate the temporal importance of the feature node according to the matching degree distribution.

[0131] Input the local importance and the temporal importance into a feature fusion network. This network can be a multi-layer perceptron that dynamically adjusts the fusion weights through an error feedback mechanism and outputs the comprehensive importance of the feature nodes. For example, the backpropagation algorithm can be used to adjust the weights of the network. Normalize the comprehensive importance to obtain the initial probability distribution for random walks.

[0132] According to the initial probability distribution, calculate the forward transition probability and the backward transition probability between nodes, and construct a temporal transition probability matrix. The forward transition probability represents the probability of transferring from one node to another node, and the backward transition probability represents the probability of returning from one node to the previous node. Initialize the non-homogeneous temporal random walk strategy using the temporal transition probability matrix. In each random walk step, update the temporal transition probability matrix based on the local importance and the temporal importance of the current node. For example, the transition probability can be weighted and adjusted according to the local importance and the temporal importance of the current node. Execute the non-homogeneous temporal random walk strategy to obtain an initial random walk path.

[0133] Perform piecewise sampling on the initial random walk path to obtain a set of sampled paths. For example, the path can be sampled at fixed time intervals. Extract temporal key points from the set of sampled paths, such as nodes with higher local importance or temporal importance. Determine the path skeleton based on the temporal key points. Generate a multi-hop random walk path based on the path skeleton, which contains a sequence of node jumps that satisfy the temporal transition constraints.

[0134] In this embodiment, the community structure features and the historical activity sequence features of the nodes are fused, which can more accurately characterize the importance of the nodes, thereby improving the accuracy of the generated random walk paths and more effectively capturing the dynamic association patterns between the nodes; the transition probability is dynamically updated during the random walk process, and a path sampling and skeleton extraction mechanism is introduced, which can generate multi-hop random walk paths that satisfy the temporal transition constraints and better reflect the dynamic evolution process of the nodes in the network; by extracting the temporal key points, the evolution process of the path can be presented more clearly, improving the interpretability of the path and helping users understand the dynamic association patterns between the nodes in the network.

[0135] In an alternative embodiment, a temporal evolution model of real-time compliance evaluation results is constructed using a causal inference decision tree, the risk evolution gradient is calculated, the risk state matrix is dynamically decomposed based on the risk evolution gradient, abnormal risk points are located, and risk warning information and risk intervention strategies are generated according to the spatio-temporal distribution characteristics of the abnormal risk points; the spatio-temporal distribution characteristics, risk warning information and risk intervention strategies are fed back to the data association relationship matrix to update the knowledge association data of medical service items, including:

[0136] Obtain the risk state data in the medical service process and extract the risk state characteristics. According to the hierarchical relationship of the risk state characteristics, construct the nodes of the causal inference decision tree, and establish the transfer connections between the nodes based on the temporal change relationship of the risk state data to generate a multi-level causal inference decision tree; calculate the state transition weights between adjacent nodes in the multi-level causal inference decision tree, normalize the state transition weights to obtain the state transition probability, and construct the risk state transition matrix.

[0137] Obtain the risk state transition matrices at adjacent time steps, calculate the difference between the risk state transition matrices, and obtain the risk evolution gradient according to the ratio of the difference to the time interval; perform singular value decomposition on the risk evolution gradient to obtain the eigenmatrix, and identify the abnormal risk points of the risk situation mutation based on the singular values in the eigenmatrix.

[0138] Extract the time span and spatial distribution information of the abnormal risk points, construct a spatio-temporal feature vector in combination with the mutation degree of the abnormal risk points, determine the risk warning level according to the spatio-temporal feature vector, and generate a risk intervention strategy based on the risk warning level and the spatio-temporal feature vector.

[0139] Input the spatio-temporal feature vector, the risk warning level and the risk intervention strategy into the feature fusion function, update the data association relationship matrix according to the output result of the feature fusion function, and optimize the knowledge association data of medical service items based on the data association relationship matrix.

[0140] In a specific embodiment, first, the system collects risk state data from all aspects of medical services, including comprehensive information such as the patient diagnosis and treatment process, doctor's order execution, drug use, and nursing records. After data cleaning and standardization processing, the risk state characteristics in these data are extracted, such as abnormal operation time, dosage deviation, execution delay, etc. The system analyzes the hierarchical dependence relationship between these risk state characteristics and constructs the initial nodes of the causal inference decision tree. Subsequently, based on the evolution law of the risk state data in the time dimension, the transfer connection relationship between the nodes is established, and finally a multi-level causal inference decision tree structure is formed.

[0141] After the decision tree is constructed, the system calculates in detail the state transition weights between each pair of adjacent nodes. This weight value reflects the tendency of one risk state to transition to another. By normalizing all the transition weights, the standardized state transition probabilities are obtained, and these probability values are organized into a risk state transition matrix. This matrix completely records all possible risk state transition paths and their occurrence probabilities in the system.

[0142] The system continuously monitors the risk state transition matrix within adjacent time windows. By calculating the difference between the matrices at two time points and dividing this difference by the corresponding time interval, the risk evolution gradient reflecting the risk change rate is obtained. Performing singular value decomposition on this gradient matrix yields a feature matrix containing the main characteristic patterns. By analyzing the magnitudes and distributions of the singular values in the feature matrix, the system can accurately identify the abnormal risk points where the risk situation undergoes sudden changes.

[0143] For each identified abnormal risk point, the system extracts in detail the time span information of its occurrence, including the start time, duration, periodic characteristics, etc., and at the same time analyzes its spatial distribution in different medical service links. Combining the information in these time and space dimensions with the degree index of risk mutation, a complete spatio-temporal feature vector is constructed. Based on the performance of each dimension of this feature vector, the system determines the corresponding risk warning level and formulates targeted risk intervention strategies according to the warning level and specific spatio-temporal characteristics.

[0144] Finally, the system inputs these three types of information, namely the spatio-temporal feature vector, the risk warning level, and the risk intervention strategy, into the feature fusion function for integrated processing. The fused result is used to update the data association relationship matrix, and further optimize the knowledge association data of medical service items, forming a closed-loop risk management mechanism.

[0145] Specific application example: Taking the management of anesthetic drug use in the operating room of a certain hospital as an example:

[0146] The system first collects all data related to surgical anesthesia, including medication orders issued by anesthesiologists, nurse execution records, patient vital sign monitoring data, etc. By analyzing these data, the system identifies multiple risk state characteristics, such as abnormal anesthetic drug doses, inappropriate drug administration time intervals, intraoperative sign fluctuations, etc. Based on these characteristics, the system constructs a multi-level causal reasoning decision tree, where the bottom-layer nodes represent specific risk manifestations, the middle-layer nodes represent risk aggregation states, and the top-layer nodes represent the overall risk level.

[0147] The system monitors the changes in various indicators during the anesthesia process in real time and calculates the transition probabilities between risk states. For example, when it is found that the drug dosage is adjusted continuously multiple times during the operation, the system will calculate the probability of transitioning from the normal drug administration state to the abnormal dosage state. By continuously monitoring the data for a week, the system discovers that the transition probability of abnormal dosage suddenly increases during a certain period. Through singular value decomposition analysis, it is confirmed that this is an abnormal risk point.

[0148] Further analysis by the system reveals that this abnormal risk point mainly appears during the night shift period, and is concentrated in several operating rooms, involving specific medical staff. Based on these spatio-temporal characteristics, the system sets the risk warning level to a higher level and generates specific intervention strategies, including suggesting strengthening the training of night shift staff and adjusting the operating room scheduling mechanism, etc.

[0149] These findings and suggestions are input into the feature fusion function for integration, which is used to update the knowledge association data for anesthesia drug management. The updated association data shows that it is necessary to strengthen the supervision of anesthesia operations during the night shift and optimize the personnel allocation plan, so as to reduce the occurrence probability of similar risks. The whole process forms a complete closed loop from risk discovery to intervention and improvement, continuously enhancing the safety of anesthesia drug management.

[0150] In this embodiment, it is possible to obtain risk state data from the medical service process in real time, extract key risk characteristics, and realize the instant monitoring of risk states and change trends; by constructing a multi-level causal inference decision tree and a state transition matrix, the hierarchical relationship and temporal evolution between risk states are systematically revealed, providing an accurate quantitative description for risk changes; using the risk evolution gradient and singular value decomposition, the mutation points of the risk situation are effectively captured, abnormal risk states are discovered in a timely manner, and the sensitivity of the warning response is improved; extracting the time span and spatial distribution information of abnormal risk points, generating spatio-temporal feature vectors, and then determining the warning level and formulating targeted intervention strategies, enhancing the pertinence and effectiveness of risk disposal; through feature fusion feedback to update the data association relationship matrix, continuously optimizing the knowledge association data of medical service items, and realizing the dynamic self-adaptation and continuous improvement of the risk assessment model.

[0151] In an alternative implementation manner, obtaining risk state data from the medical service process and extracting risk state characteristics, constructing causal inference decision tree nodes according to the hierarchical relationship of the risk state characteristics, and establishing transfer connections between the nodes based on the temporal change relationship of the risk state data, generating a multi-level causal inference decision tree includes:

[0152] By standardizing and reducing the dimension of the medical service risk status data, using an autoencoder network to extract potential feature vectors and calculate the risk status similarity, forming risk status nodes based on similarity clustering, establishing transfer connections between nodes according to the temporal change relationship, and optimizing the connection structure of the decision tree using the maximum information coefficient, specifically including:

[0153] Obtain the risk status data in the medical service process and extract the risk status features, perform standardization processing on the risk status features to obtain standardized features, and use the principal component analysis method to reduce the dimension of the standardized features to obtain the reduced-dimension risk status features;

[0154] Input the reduced-dimension risk status features into the autoencoder network. The encoder part of the autoencoder network maps the reduced-dimension risk status features to potential feature vectors, and the decoder part of the autoencoder network reconstructs the reduced-dimension risk status features based on the potential feature vectors. Calculate the Euclidean distance between risk states according to the potential feature vectors to obtain the risk status similarity;

[0155] Perform clustering operations based on the risk status similarity. According to a preset similar similarity threshold, classify the corresponding risk states into a risk status cluster to obtain multiple different-level risk status clusters, and map each risk status cluster to a risk status node of the causal inference decision tree; Based on the temporal changes of the risk states in the risk status cluster, calculate the temporal correlation strength between the risk status nodes. The temporal correlation strength is determined according to the jump frequency of the risk status in the time series;

[0156] Establish a directed connection relationship between the risk status nodes according to the temporal correlation strength, establish a transfer connection between the risk status node pairs with the temporal correlation strength greater than the preset strength threshold, and construct the initial structure of the multi-level causal inference decision tree based on the risk status nodes and the transfer connections;

[0157] Use the maximum information coefficient method to calculate the causal relationship strength between the risk status nodes, optimize the initial structure of the multi-level causal inference decision tree based on the causal relationship strength, delete the transfer connections with the causal relationship strength lower than the preset causal threshold, and retain the transfer connections higher than the preset causal threshold to obtain the optimized multi-level causal inference decision tree.

[0158] In a specific implementation, the implementation process first comprehensively collects risk status data from the medical service system. This data covers risk information in various aspects such as the diagnosis and treatment process, medical order execution, equipment use, and personnel operation. The system extracts features from the collected raw data to obtain risk status features with multiple dimensions including time features, space features, operation features, etc. These features are standardized to eliminate the dimensional differences between different dimensions, so that all features are mapped into the same numerical interval. Then, the principal component analysis method is used to reduce the dimension of the standardized features, retaining the main feature information and reducing the data redundancy.

[0159] The risk status features after dimensionality reduction are input into a pre-trained autoencoder network. In the encoder part, the high-dimensional risk status features are compressed and mapped into low-dimensional latent feature vectors through a multi-layer neural network. This vector contains the core feature information of the original risk status. The decoder part then attempts to reconstruct the original risk status features through this latent feature vector, ensuring that the latent feature vector retains sufficient information. Based on these latent feature vectors, the system calculates the Euclidean distance between different risk statuses to obtain a similarity matrix reflecting the similarity degree of risk statuses.

[0160] The system uses a clustering algorithm to group the risk statuses. According to a preset similarity threshold, the risk statuses with a similarity higher than the threshold are grouped into the same risk status cluster. By adjusting the similarity threshold, different levels of risk status clusters are formed to reflect the hierarchical structure relationship of risk statuses. Each risk status cluster is mapped to a node in the causal inference decision tree. The system analyzes the change law of risk statuses over time in each risk status cluster, counts the jump frequency between risk statuses, and calculates the temporal association strength between different nodes based on this.

[0161] Based on the calculated temporal association strength, the system starts to construct the connection relationship between nodes. When the temporal association strength between two nodes exceeds the preset strength threshold, the system establishes a directed transfer connection between these two nodes. Transfer connections are established between all pairs of nodes that meet the conditions, thus forming the initial structure of the multi-level causal inference decision tree.

[0162] Finally, the system uses the maximum information coefficient method to deeply analyze the causal relationship between nodes. By calculating the causal relationship strength between each pair of connected nodes, the initial structure of the decision tree is optimized. The system deletes the transfer connections with a causal relationship strength lower than the preset threshold, only retaining the connections with significant causal relationships, and finally obtains the optimized multi-level causal inference decision tree.

[0163] Exemplarily, taking the management of the emergency infusion room in a certain hospital as an example:

[0164] The system first collects all business data in the infusion room, including doctor's order information, medication preparation records, infusion execution records, nursing records, etc. Multiple risk status features are extracted from these data, such as delayed medication preparation time, abnormal infusion rate, drug compatibility risk, etc. After standardization processing, these feature data are input into the principal component analysis model, and the main feature dimensions that can explain 85% of the variance are retained after dimensionality reduction.

[0165] The autoencoder network receives these feature data after dimensionality reduction and compresses the feature data originally containing more than a dozen dimensions into three-dimensional latent feature vectors. The decoder successfully reconstructs the original features with these latent feature vectors, and the reconstruction error is within an acceptable range. The system calculates the Euclidean distance between the latent feature vectors corresponding to all infusion events and constructs a complete similarity matrix.

[0166] By setting different similarity thresholds, the system classifies the infusion risk status into three levels. The bottom level contains specific risk manifestations, such as abnormal single-dose administration time; the middle level represents the concentrated occurrence of a certain type of risk, such as abnormal consecutive multiple-dose administration times; the top level represents systemic risks, such as overall infusion process disorders. Each risk status cluster corresponds to a node in the decision tree.

[0167] The system analysis finds that when there is an abnormal single-dose administration time, there is a probability of more than 60% that consecutive administration abnormalities will occur within the next two hours. This temporal correlation strength exceeds the preset threshold, so a transfer connection is established between the corresponding nodes. Similarly, the system establishes connections between all nodes with significant temporal correlations, forming an initial decision tree structure.

[0168] After analysis using the maximum information coefficient method, it is found that although some transfer connections have temporal correlations, the causal relationships are weak. For example, the causal relationship strength between abnormal single-dose administration time and subsequent drug compatibility risk is low, and such connections are deleted by the system. The final decision tree clearly shows the true causal associations between various risk statuses in the infusion room, providing a reliable decision basis for risk management.

[0169] In this embodiment, through standardization processing, principal component analysis, and autoencoder network, the original risk status data is dimensionally reduced and key potential features are extracted, effectively reducing noise and redundant information, and improving the accuracy of feature expression. The Euclidean distance is used to calculate the risk status similarity, and similar risk statuses are grouped into the same cluster through clustering to form risk status nodes at multiple levels, laying a fine foundation for subsequent causal analysis. Based on the temporal changes and transfer connections of risk status nodes, a multi-level causal inference decision tree is constructed to capture the temporal correlation and dynamic evolution law between risk statuses, enhancing the understanding of the risk change trend. The maximum information coefficient method is used to quantify the strength of the causal relationship between nodes, and the decision tree structure is optimized to ensure that only high-strength transfer connections are retained, improving the robustness and prediction ability of the model. The overall process helps to accurately identify the changes in risk status and its causal mechanism, providing reliable and fine data support and decision-making basis for risk early warning and intervention strategy formulation in medical services.

[0170] Figure 2 FIG. is a schematic structural diagram of a compliance dynamic monitoring system for medical service items based on artificial intelligence according to an embodiment of the present invention, as Figure 2 shown, the system includes:

[0171] A first unit for dynamically associating and mapping medical service item data using non-equilibrium entropy theory, establishing a data association relationship matrix, and obtaining item knowledge association data; based on the data association relationship matrix, using a chaotic neuron network to extract the risk feature distribution in the item knowledge association data; using a Lyapunov stability analyzer to perform long-term trend prediction on the risk feature distribution, and generating a risk benchmark model including risk assessment weights and dynamic monitoring thresholds;

[0172] A second unit for extracting a diagnosis and treatment feature sequence from item operation data using a deep generative adversarial diagnosis and treatment modeler, and constructing a knowledge representation model; based on the knowledge representation model, using a heterogeneous medical network neural evolver to analyze the association strength and temporal features between various diagnosis and treatment activities, generating a risk prediction spectrum after comparison with historical compliance records, and combining with the risk benchmark model, performing dynamic evolution analysis on the risk prediction spectrum through a medical value dynamics framework, and outputting a real-time compliance analysis result;

[0173] A third unit for constructing a temporal evolution model of the real-time compliance assessment result using a causal inference decision tree, calculating the risk evolution gradient, dynamically decomposing the risk status matrix based on the risk evolution gradient, locating abnormal risk points, and generating risk early warning information and risk intervention strategies according to the spatio-temporal distribution characteristics of the abnormal risk points; feeding back the spatio-temporal distribution characteristics, risk early warning information, and risk intervention strategies to the data association relationship matrix to update the item knowledge association data of medical services.

[0174] In the third aspect of the embodiment of the present invention,

[0175] Provided is an electronic device, comprising:

[0176] a processor;

[0177] a memory for storing instructions executable by the processor;

[0178] wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0179] In a fourth aspect of the embodiments of the present invention,

[0180] provided is a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0181] The present invention may be a method, an apparatus, a system, and / or a computer program product. The computer program product may include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0182] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for dynamically monitoring the compliance of medical service items based on artificial intelligence, characterized in that, Including: Dynamically associating and mapping medical service item data using the non-equilibrium entropy theory, establishing a data association relationship matrix, and obtaining item knowledge association data; Based on the data association relationship matrix, using a chaotic neural network to extract the risk feature distribution in the item knowledge association data; Using a Lyapunov stability analyzer to conduct long-term trend prediction on the risk feature distribution, generating a risk benchmark model including risk assessment weights and dynamic monitoring thresholds; Using a deep generative adversarial diagnosis and treatment modeling device to extract diagnosis and treatment feature sequences from item operation data and construct a knowledge representation model; Based on the knowledge representation model, using a heterogeneous medical network neural evolver to analyze the association strength and temporal characteristics between various diagnosis and treatment activities, generating a risk prediction spectrum after comparing with historical compliance records, and combining with the risk benchmark model, dynamically evolving and analyzing the risk prediction spectrum through a medical value dynamics framework to output real-time compliance analysis results; Using a causal inference decision tree to construct a temporal evolution model of the real-time compliance assessment results, calculating the risk evolution gradient, dynamically decomposing the risk state matrix based on the risk evolution gradient, locating abnormal risk points, and generating risk warning information and risk intervention strategies according to the spatio-temporal distribution characteristics of the abnormal risk points; Feeding back the spatio-temporal distribution characteristics, risk warning information, and risk intervention strategies to the data association relationship matrix to update the medical service item knowledge association data.

2. The method according to claim 1, wherein Dynamically associating and mapping medical service item data using the non-equilibrium entropy theory, establishing a data association relationship matrix, and obtaining item knowledge association data; Based on the data association relationship matrix, using a chaotic neural network to extract the risk feature distribution in the item knowledge association data; Using a Lyapunov stability analyzer to conduct long-term trend prediction on the risk feature distribution, generating a risk benchmark model including risk assessment weights and dynamic monitoring thresholds including: Constructing an initial data set for medical service item data according to service type, execution time, operator, service object, and operation record; obtaining the non-equilibrium entropy value between any two data items according to the probability distribution of each data item in the corresponding dimension in the initial data set, calculating the association strength between the data items according to the non-equilibrium entropy value and the smoothing factor, and constructing a data association relationship matrix; introducing a time window mechanism in the data association relationship matrix, and weighted combining the data association relationship matrix of the current time window and the historical data association relationship matrix through a forgetting factor to generate the item knowledge association data of the medical service item; Using the association strength in the item knowledge association data as the neuron connection weight to construct a chaotic neural network, using a non-linear iterative mapping to process the neuron input signal to keep the chaotic neural network in a chaotic critical state, and reducing the dimension of the output of the chaotic neural network through a locality-sensitive hashing algorithm to obtain the risk feature distribution of the medical service item; Convert the risk feature distribution into a dynamic system equation, calculate the Lyapunov exponents of the dynamic system equation in each feature dimension, where the Lyapunov exponents characterize the degree of divergence of the trajectories in the feature dimensions; perform exponential transformation and normalization on the Lyapunov exponents to generate risk assessment weights; combine the historical data statistics of each feature dimension with the Lyapunov exponents of the corresponding dimension to generate dynamic monitoring thresholds, and combine the risk assessment weights and the dynamic monitoring thresholds to construct a risk benchmark model for medical service items.

3. The method according to claim 2, wherein Use the association strength in the project knowledge association data as the neuron connection weights to construct a chaotic neuron network. Employ non-linear iterative mapping to process the neuron input signals to keep the chaotic neuron network in the chaotic critical state. Reduce the dimension of the output of the chaotic neuron network through the locality-sensitive hashing algorithm to obtain the risk feature distribution of medical service items, including: Construct the association strength of the project knowledge association data as the neuron connection weights, combine non-linear iterative mapping to control the chaotic critical state of the network, complete the dynamic update of the weights through the chaotic adaptive adjustment mechanism, monitor the network synchronization degree using the chaotic resonance detection mechanism, and obtain the risk feature distribution by using the local similarity mapping, specifically including: Construct the initial connection weight matrix of the chaotic neuron network according to the association strength in the project knowledge association data; input the input signals of adjacent neurons and the initial connection weight matrix into the signal accumulation unit, calculate the cumulative state value at the current moment according to the attenuation coefficient, process the cumulative state value through non-linear iterative mapping and compare it with the activation threshold. When the cumulative state value is greater than the activation threshold, trigger neuron activation and enter the transient inhibition period; calculate the positive activation energy and the negative inhibition energy and convert them into chaotic adjustment signals to keep the chaotic neuron network in the chaotic critical state; Construct a chaotic adaptive adjustment module, obtain the activation timing information of adjacent neurons, calculate the weight adjustment increment and dynamically fuse it with the initial connection weight matrix to update the connection weight matrix; adjust the corresponding sensitivity threshold according to the activation state of the neurons, and adjust the signal transmission strength based on the overall synchronization degree of the chaotic neuron network; Set up a chaotic resonance detection module, calculate the synchronization degree based on the activation patterns of neurons in the chaotic neuron network. When the synchronization degree is greater than the preset steady-state threshold, enter the feature extraction state and generate a feature vector; Perform dimensionality reduction mapping on the feature vector using the locality-sensitive hashing algorithm to obtain the risk feature distribution of medical service items.

4. The method according to claim 1, characterized in that Use a deep generative adversarial diagnosis and treatment modeling device to extract diagnosis and treatment feature sequences from project operation data and construct a knowledge representation model; based on the knowledge representation model, use a heterogeneous medical network neural evolver to analyze the association strength and timing features between various diagnosis and treatment activities, generate a risk prediction spectrum after comparing with historical compliance records, and combine with the risk benchmark model to perform dynamic evolution analysis on the risk prediction spectrum through the medical value dynamics framework, and output real-time compliance analysis results, including: Construct a deep generative adversarial diagnosis and treatment modeling device. A residual connection structure is set in the generator network of the deep generative adversarial diagnosis and treatment modeling device, and a feature fusion gating unit is set in each connection layer of the residual connection structure. The feature fusion gating unit adaptively fuses features at different levels to obtain fused features. In the discriminator network of the deep generative adversarial diagnosis and treatment modeling device, multiple convolution kernels of different scales are set, and the fused features are decomposed at multiple scales by using grouped convolution to obtain decomposed features. Input the operation data of medical service items into the deep generative adversarial diagnosis and treatment modeling device, extract the diagnosis and treatment feature sequence based on the decomposed features, group the diagnosis and treatment feature sequence according to time series features, behavior features and resource features, and construct a knowledge representation model. Map the diagnosis and treatment feature sequence in the knowledge representation model to the feature nodes of the heterogeneous medical network, and construct association edges based on the time series relationship, behavior association and resource dependence of the feature nodes. Calculate the node importance of the feature nodes as the starting probability of random walk, and execute time series walk to obtain a multi-hop walk path. Calculate the node transition probability matrix according to the multi-hop walk path, and construct a diagnosis and treatment activity association intensity matrix. Perform time series alignment on the diagnosis and treatment activity association intensity matrix and historical compliance records to obtain aligned data. Input the aligned data into a time series convolutional network with a dilated convolution structure, extract multi-scale time series features and calculate the risk propagation coefficient to generate a risk prediction spectrum. Construct a medical value dynamics equation set based on the risk prediction spectrum. A risk propagation function and a value loss function are set in the medical value dynamics equation set. The risk propagation function describes the diffusion process of risk between the feature nodes according to the risk propagation coefficient, and the value loss function quantifies the impact of risk on the value of medical services. Solve the medical value dynamics equation set to obtain the dynamic evolution trajectory of the risk state, and output the real-time compliance analysis result based on the dynamic evolution trajectory.

5. The method according to claim 4, characterized in that, Calculating the node importance of the feature nodes as the starting probability of random walk and executing time series walk to obtain a multi-hop walk path includes: Calculate the comprehensive importance by fusing the community structure features and historical activity sequence features of the feature nodes as the starting probability of random walk. During the execution of time series walk, dynamically update the transition probability based on the local importance and time series importance of the nodes, and introduce a path sampling and skeleton extraction mechanism to generate a multi-hop walk path that meets the time series transition constraints, specifically including: Obtain the connection relationship matrix of the feature nodes, decompose the connection relationship matrix by using spectral clustering method to obtain the feature vector matrix, and identify the node community structure based on the feature vector matrix. Calculate the structural density between node pairs in the node community structure to obtain the density matrix, and calculate the propagation ability in combination with the in-degree and out-degree distributions of the nodes to obtain the propagation matrix. Calculate the local importance of the nodes according to the density matrix and the propagation matrix. Collect the historical activity sequences of the feature nodes, perform time series decomposition on the historical activity sequences to obtain periodic components, construct a time series template based on the periodic components, match the current activity sequences of the feature nodes with the time series template to obtain a matching degree distribution, and calculate the time series importance of the feature nodes according to the matching degree distribution; Input the local importance and the time series importance into a feature fusion network. The feature fusion network dynamically adjusts the fusion weights based on an error feedback mechanism, outputs the comprehensive importance of the feature nodes, and normalizes it to obtain the initial probability distribution for random walk; Calculate the forward transition probability and the backward transition probability between nodes according to the initial probability distribution, and construct a time series transition probability matrix based on the forward transition probability and the backward transition probability; Initialize a non-homogeneous time series random walk strategy with the time series transition probability matrix. In each random walk step, update the time series transition probability matrix based on the local importance and the time series importance of the current node; Execute the non-homogeneous time series random walk strategy to obtain an initial random walk path, perform segment sampling on the initial random walk path to obtain a set of sampled paths, extract time series key points from the set of sampled paths to determine the path skeleton, and generate a multi-hop random walk path based on the path skeleton. The multi-hop random walk path contains a node jump sequence that satisfies the time series transition constraint; 6. The method according to claim 1, wherein Use a causal inference decision tree to construct a time series evolution model for real-time compliance evaluation results, calculate the risk evolution gradient, dynamically decompose the risk state matrix based on the risk evolution gradient, locate abnormal risk points, and generate risk warning information and risk intervention strategies according to the spatio-temporal distribution characteristics of the abnormal risk points; Feed back the spatio-temporal distribution characteristics, risk warning information, and risk intervention strategies to the data association relationship matrix to update the medical service item knowledge association data, including: Obtain the risk state data in the medical service process and extract the risk state characteristics. According to the hierarchical relationship of the risk state characteristics, construct the nodes of a causal inference decision tree, establish transfer connections between the nodes based on the time series change relationship of the risk state data, and generate a multi-level causal inference decision tree; Calculate the state transfer weights between adjacent nodes in the multi-level causal inference decision tree, normalize the state transfer weights to obtain the state transfer probability, and construct a risk state transfer matrix; Obtain the risk state transfer matrices at adjacent time steps, calculate the difference between the risk state transfer matrices, and obtain the risk evolution gradient according to the ratio of the difference to the time interval; Perform singular value decomposition on the risk evolution gradient to obtain a feature matrix, and identify abnormal risk points with sudden changes in the risk situation based on the singular values in the feature matrix; Extract the time span and spatial distribution information of the abnormal risk points, construct a spatio-temporal feature vector by combining the mutation degree of the abnormal risk points, determine the risk warning level according to the spatio-temporal feature vector, and generate a risk intervention strategy based on the risk warning level and the spatio-temporal feature vector; Input the spatio-temporal feature vector, the risk warning level, and the risk intervention strategy into the feature fusion function, update the data association relationship matrix according to the output result of the feature fusion function, and optimize the knowledge association data of medical service items based on the data association relationship matrix.

7. The method according to claim 6, wherein Obtain the risk status data in the medical service process and extract the risk status features. According to the hierarchical relationship of the risk status features, construct the causal inference decision tree nodes, and establish the transfer connections between the nodes based on the temporal change relationship of the risk status data. The generated multi-level causal inference decision tree includes: By performing standardization and dimensionality reduction processing on the medical service risk status data, using an autoencoder network to extract the latent feature vector and calculate the risk status similarity, forming risk status nodes based on similarity clustering, establishing transfer connections between nodes according to the temporal change relationship, and optimizing the connection structure of the decision tree using the maximum information coefficient. Specifically, it includes: Obtain the risk status data in the medical service process and extract the risk status features. Perform standardization processing on the risk status features to obtain standardized features, and use the principal component analysis method to perform dimensionality reduction processing on the standardized features to obtain the dimensionality-reduced risk status features; Input the dimensionality-reduced risk status features into the autoencoder network. The encoder part of the autoencoder network maps the dimensionality-reduced risk status features to the latent feature vector, and the decoder part of the autoencoder network reconstructs the dimensionality-reduced risk status features based on the latent feature vector. Calculate the Euclidean distance between the risk statuses according to the latent feature vector to obtain the risk status similarity; Perform clustering operations based on the risk status similarity. According to the preset similar similarity threshold, classify the corresponding risk statuses into a risk status cluster to obtain multiple risk status clusters at different levels. Map each risk status cluster to a risk status node of the causal inference decision tree; Based on the temporal changes of the risk statuses in the risk status cluster, calculate the temporal association strength between the risk status nodes. The temporal association strength is determined according to the jump frequency of the risk status in the time series; Establish a directed connection relationship between the risk status nodes according to the temporal association strength. Establish transfer connections between the risk status node pairs with the temporal association strength greater than the preset strength threshold, and construct the initial structure of the multi-level causal inference decision tree based on the risk status nodes and the transfer connections; Use the maximum information coefficient method to calculate the causal relationship strength between the risk status nodes, optimize the initial structure of the multi-level causal inference decision tree based on the causal relationship strength, delete the transfer connections with the causal relationship strength lower than the preset causal threshold, and retain the transfer connections higher than the preset causal threshold to obtain the optimized multi-level causal inference decision tree.

8. An artificial intelligence-based dynamic monitoring system for the compliance of medical service items, which is used to implement the method described in any one of the foregoing claims 1-7, and is characterized in that, Include: The first unit is used to perform dynamic association mapping on the medical service item data using the non-equilibrium state entropy theory, establish a data association relationship matrix, and obtain the item knowledge association data; Based on the data association relationship matrix, use a chaotic neural network to extract the risk feature distribution in the item knowledge association data; Using a Lyapunov stability analyzer to conduct long-term trend prediction on the risk characteristic distribution, and generating a risk benchmark model that includes risk assessment weights and dynamic monitoring thresholds; A second unit, configured to extract a diagnosis and treatment feature sequence from project operation data by using a deep generative adversarial diagnosis and treatment modeling device, and construct a knowledge representation model; Based on the knowledge representation model, using a heterogeneous medical network neural evolver to analyze the association strength and temporal characteristics among various diagnosis and treatment activities, generating a risk prediction spectrum after comparing with historical compliance records, and combining with the risk benchmark model, dynamically evolving and analyzing the risk prediction spectrum through a medical value dynamics framework, and outputting a real-time compliance analysis result; A third unit, configured to use a causal inference decision tree to construct a temporal evolution model of the real-time compliance assessment result, calculate a risk evolution gradient, dynamically decompose a risk state matrix based on the risk evolution gradient, locate abnormal risk points, and generate risk warning information and risk intervention strategies according to the spatio-temporal distribution characteristics of the abnormal risk points; Feeding back the spatio-temporal distribution characteristics, risk warning information and risk intervention strategies to a data association relationship matrix, and updating medical service project knowledge association data.

9. An electronic device, characterized in that, Including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 7 is implemented.

Images