Medical quality intelligent evaluation and early warning prediction method and system based on multi-dimensional indexes

Abstract

The invention provides a medical quality intelligent evaluation and early warning prediction method and system based on a multi-dimensional index, and relates to the technical field of medical quality evaluation, and the method comprises the steps: calculating the fluctuation feature value of a medical quality impact factor, constructing a quality sensitivity scoring model, and carrying out the dynamic sampling to extract a quality feature vector. And performing feature fusion by calculating the clinical association strength and the feature importance score of the feature group to obtain a medical quality score. And performing multi-scale wavelet transform on the medical quality score, extracting a quality evolution feature sequence, calculating a department quality state vector, constructing a quality propagation graph, and identifying a quality control node. And finally, a multi-objective constraint optimization equation is constructed based on the risk propagation prediction sequence, a quality intervention strategy set is solved, and an optimal intervention scheme is screened, so that intelligent assessment, early warning and prediction of medical quality are realized, and the medical quality management level is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technology of medical quality assessment, and in particular to an intelligent medical quality assessment and early warning prediction method and system based on multi-dimensional indicators. Background Art

[0002] Medical quality management is a key link to ensure the safety and effectiveness of medical services. Traditional medical quality assessment methods mainly rely on manual experience judgment, which has problems such as inconsistent assessment criteria, poor timeliness, and strong subjectivity, and it is difficult to detect potential quality risks in a timely manner. With the continuous improvement of medical informatization, various types of clinical data have grown rapidly, but the existing quality assessment models have limited analysis capabilities for multi-source heterogeneous data and cannot fully explore the quality evolution laws contained in the data.

[0003] Currently, the medical quality early warning prediction methods generally use single-index threshold judgment, lacking in-depth analysis of the quality fluctuation characteristics and propagation laws, resulting in insufficient early warning accuracy. At the same time, the existing early warning models do not sufficiently consider the relevance of the diagnosis and treatment paths between departments, and it is difficult to effectively identify the propagation paths and key control nodes of quality risks, restricting the effect of continuous improvement of medical quality.

[0004] Medical quality management faces challenges such as high data dimensions, complex quality characteristics, and dynamic changes in risk propagation. It is necessary to establish an assessment system integrating multi-dimensional indicators to achieve accurate characterization of the quality status and intelligent prediction of risk evolution. Therefore, there is an urgent need for an intelligent medical quality assessment and early warning prediction method based on multi-dimensional indicators to improve the scientificity and initiative of medical quality management. Summary of the Invention

[0005] The embodiments of the present invention provide an intelligent medical quality assessment and early warning prediction method and system based on multi-dimensional indicators, which can solve the problems in the prior art.

[0006] In the first aspect of the embodiments of the present invention, There is provided an intelligent medical quality assessment and early warning prediction method based on multi-dimensional indicators, including: Obtaining real-time medical data, calculating the fluctuation characteristic values of medical quality impact factors, constructing a quality sensitivity scoring model based on the fluctuation characteristic values, calculating the data sampling priority, establishing a hierarchical sampling frequency matrix, dynamically sampling the real-time medical data according to the hierarchical sampling frequency matrix to obtain sampling data, extracting the quality feature vectors of the sampling data and setting differential time windows, calculating the clinical association strength and feature importance scores of the feature groups within the differential time windows, and performing feature fusion to obtain the medical quality score; Perform multi-scale wavelet transform on the medical quality score to obtain the quality fluctuation component, calculate the local trend function of the quality fluctuation component at different time scales, construct the cross-correlation matrix between the local trend function and the quality fluctuation component, extract the quality evolution feature sequence, calculate the dynamic time warping distance to obtain the department quality state vector, construct the quality propagation graph based on the department quality state vector, fuse the diagnosis and treatment path information to construct a heterogeneous information network, calculate the propagation influence factor of the heterogeneous information network, and identify the quality control nodes; Calculate the network topological entropy of the quality propagation graph, combine it with the sample entropy of the quality fluctuation component to obtain the quality risk entropy, construct the risk accumulation function based on the quality risk entropy, input the centrality index of the quality control node to obtain the risk propagation prediction sequence, construct the multi-objective constraint optimization equation based on the risk propagation prediction sequence, solve the multi-objective constraint optimization equation to obtain the quality intervention strategy set, and screen the optimal intervention plan based on the quality intervention strategy set.

[0007] In an alternative embodiment, Obtain real-time medical data, calculate the fluctuation eigenvalue of the medical quality influence factor, construct a quality sensitivity score model based on the fluctuation eigenvalue, calculate the data sampling priority, establish a hierarchical sampling frequency matrix, and perform dynamic sampling on the real-time medical data according to the hierarchical sampling frequency matrix to obtain the sampling data including: Input the fluctuation eigenvalue into a deep neural network, the deep neural network includes a feature extraction layer, an attention mechanism layer and a fully connected layer, learn the combined weights of the short-term fluctuation intensity, medium-term fluctuation period and long-term fluctuation trend in the fluctuation eigenvalue through the deep neural network to construct a quality sensitivity score model, and calculate the quality sensitivity score based on the quality sensitivity score model; Perform non-linear mapping on the quality sensitivity score to obtain the initial sampling priority, construct a double exponential smoothing model, input the initial sampling priority into the double exponential smoothing model, calculate the smoothing parameter based on the historical trend and future prediction of the real-time medical data, and perform smoothing processing on the initial sampling priority through the smoothing parameter to obtain the final sampling priority; Construct an adaptive hierarchical threshold according to the final sampling priority, divide the real-time medical data into multiple priority levels based on the adaptive hierarchical threshold to obtain the priority hierarchical result, calculate the data mean, variance and distribution density of each level in the priority hierarchical result, and calculate the corresponding basic sampling frequency of each level based on the data mean, variance and distribution density; Calculate the timeliness decay value, quality correlation decay value, and business importance decay value of the calculation data, combine them to obtain a time decay factor, construct a three-dimensional hierarchical sampling frequency matrix based on the time decay factor, the basic sampling frequency, and the quality sensitivity score, query the sampling frequency from the three-dimensional hierarchical sampling frequency matrix based on the quality sensitivity score, and generate a sampling timestamp sequence that satisfies the sampling interval constraint according to the sampling frequency; Obtain data samples at the time points corresponding to the sampling timestamp sequence, determine the spline interpolation order based on the characteristics of the real-time medical data, interpolate and reconstruct the sampling interval of the data samples using the spline interpolation order to obtain reconstructed data, and detect and correct the outliers in the reconstructed data to obtain the final sampling data set.

[0008] In an alternative embodiment, Perform multi-scale wavelet transform on the medical quality score to obtain a quality fluctuation component, calculate the local trend function of the quality fluctuation component at different time scales, construct a cross-correlation matrix of the local trend function and the quality fluctuation component, extract the quality evolution feature sequence, and calculate the dynamic time warping distance to obtain the department quality state vector, including: Construct an asymmetric wavelet basis function with adaptive time-frequency resolution, perform continuous wavelet transform on the medical quality score based on the asymmetric wavelet basis function to obtain quality fluctuation components in different frequency domains; Decompose the quality fluctuation component into its morphological components, extract the basic morphological elements of the quality fluctuation component, and perform smooth fitting on the basic morphological elements using a sliding time window to obtain local trend functions at different time scales; Perform fractional-order differential operations on the local trend function and the quality fluctuation component respectively, construct a fractional-order differential operator, and obtain a phase evolution sequence; calculate the instantaneous phase and instantaneous amplitude of the phase evolution sequence, calculate the generalized complementary coefficient at multiple time scales based on the instantaneous phase and instantaneous amplitude, and construct a cross-correlation matrix reflecting multi-scale characteristics; Perform spectral decomposition of the cross-correlation matrix on a dynamic system to obtain an eigenvalue sequence and an eigenvector sequence, and construct a non-linear feature projection transformer based on the eigenvalue sequence and the eigenvector sequence; input the cross-correlation matrix into the feature projection transformer for non-linear dimensionality reduction to obtain a quality evolution feature sequence; Calculate the geodesic distance between adjacent eigenvectors in the quality evolution feature sequence, organize the geodesic distances into a local distance matrix; use a constrained dynamic programming algorithm to solve the optimal alignment path for the local distance matrix; calculate the quality trend prediction value based on the optimal alignment path and the quality evolution feature sequence; Extract the local deformation gradient of the optimal regularized path, input the local deformation gradient into a recurrent neural network, and calculate the mass mutation probability through a non-linear activation function and a temporal memory unit; Perform dynamic time warping on the mass evolution feature sequence and a standard sequence to obtain a warping distance sequence, and use a time weight function to perform weighted accumulation on the warping distance sequence to obtain a mass accumulation deviation; Normalize and perform dimensional transformation on the mass trend prediction value, the mass mutation probability, and the mass accumulation deviation in a feature space to construct a standardized department quality status vector.

[0009] In an optional embodiment, Construct a quality propagation graph based on the department quality status vector, fuse the diagnosis and treatment path information to construct a heterogeneous information network, calculate the propagation influence factor of the heterogeneous information network, and identify quality control nodes, including: Construct a kernel function mapping matrix, perform non-linear mapping on the department quality status vector to obtain a mapped feature vector; calculate the cosine similarity of the mapped feature vectors between department nodes to construct a similarity matrix; construct an initial quality propagation graph based on the similarity matrix, where a quality status edge is established between nodes when the similarity value corresponding to a department node exceeds a preset dynamic threshold, and the weight of the quality status edge is the corresponding similarity value; Construct a time sliding window based on patient transfer records, perform exponential temporal weighting on the patient transfer frequency within the window to obtain a temporally weighted transfer frequency, and perform normalization processing on the temporally weighted transfer frequency to obtain the weight value of the patient transfer edge; extract the consultation frequency and examination application frequency based on doctor's order execution records, and combine them after assigning weight coefficients according to the business type to obtain the weight value of the medical collaboration edge; Integrate the patient transfer edge and the medical collaboration edge and their corresponding weight values into the initial quality propagation graph to construct a heterogeneous information network; for each department node in the heterogeneous information network, construct a multi-layer attention network based on the features of its first-order neighborhood nodes on the quality status edge, the patient transfer edge, and the medical collaboration edge; use the multi-layer attention network to calculate the weight contribution coefficients of each edge respectively, and perform weighted summation of the weight contribution coefficients and the weight values of the corresponding edges to obtain the propagation influence factor of the department node; Construct a random walk iteration rule based on the topological structure of the heterogeneous information network, iteratively update the propagation influence factor according to the random walk iteration rule, calculate the Euclidean distance of the propagation influence factor sequences obtained from two adjacent iterations, and obtain the final propagation influence factor when the Euclidean distance is less than the convergence threshold. Sort the department nodes according to the final propagation influence factor, and combine the in-degree distribution and betweenness centrality of each department node on the three types of edges, and use the analytic hierarchy process to identify quality control nodes.

[0010] In an alternative embodiment, Calculating the network topology entropy of the quality propagation graph, combining it with the sample entropy of the quality fluctuation component to obtain the quality risk entropy, constructing a risk accumulation function based on the quality risk entropy, and inputting the centrality index of the quality control node to obtain a risk propagation prediction sequence, including: Calculating the connection probability between nodes and adjacent nodes in the quality propagation graph, performing a logarithmic operation on the connection probability and multiplying it by the original connection probability to obtain the local entropy of the node; calculating the degree distribution and betweenness centrality of the node based on the connection probability, constructing a node importance score; using the node importance score to calculate the weight coefficient, and performing a weighted sum of the local entropy of each node and the corresponding weight coefficient to obtain the network topology entropy of the quality propagation graph; Segmenting and normalizing the quality fluctuation component to obtain a standardized sequence, constructing a template sequence based on the standardized sequence according to a preset embedding dimension; counting the number of matches of the template sequence under a given similarity tolerance; calculating the natural logarithm ratio of the number of template matches in adjacent dimensions to obtain the sample entropy of the quality fluctuation component; Constructing a non-linear combination model including the network topology entropy and the sample entropy, the non-linear combination model including a linear term, a square term, and a cross term; constructing an optimization objective function based on historical risk data and node importance scores; solving the optimization objective function to obtain a combination coefficient, and substituting the combination coefficient, network topology entropy, and sample entropy into the non-linear combination model to obtain the quality risk entropy; Using the quality risk entropy to construct an exponential cumulative function, mapping the quality risk entropy to a cumulative rate coefficient, constructing a time weight function based on the time characteristics of the standardized sequence, and combining the time weight function with the cumulative function to obtain a risk accumulation function; Extracting the centrality index of the quality control node, performing dimensionality reduction and fusion on the centrality index using the principal component analysis method, smoothing the fused centrality index, and inputting the smoothed centrality index into the risk accumulation function to obtain a prediction sequence reflecting the risk propagation degree of the quality control node.

[0011] In an alternative embodiment, Constructing a multi-objective constrained optimization equation based on the risk propagation prediction sequence, solving the multi-objective constrained optimization equation to obtain a set of quality intervention strategies, and screening the optimal intervention plan based on the set of quality intervention strategies, including: Processing the risk propagation prediction sequence using a bidirectional time series network to obtain a hidden layer state sequence, calculating a corresponding risk weight sequence based on the hidden layer state sequence through a self-attention mechanism, and performing weighted fusion of the risk weight sequence and the hidden layer state sequence to obtain a risk feature; Construct an adaptive weight function and a dynamic volatility penalty function respectively based on the risk characteristics, and combine the product of the adaptive weight function and the risk value and the product of the dynamic volatility penalty function and the volatility amplitude to construct a risk control objective function; Construct a resource interaction graph according to the resource type and resource collaboration relationship, use a graph attention network to perform node representation learning on the resource interaction graph, construct resource constraint conditions through a gated graph network based on the obtained node representations, and at the same time construct a state space based on the risk characteristics, model the intervention process as a state-action sequence in the state space, calculate the intervention time sequence corresponding to the state-action sequence using a dual value network, and construct the weighted sum of the intervention time sequence and the time weight as a time response constraint condition; Combine the risk control objective function, resource constraint conditions and time response constraint conditions to construct a multi-objective constraint optimization equation; Construct a policy network and a value network for solving the multi-objective constraint optimization equation, use the gradient policy method to optimize the parameters of the policy network and the value network, and introduce a meta-learning mechanism to dynamically adjust the optimization strategy during the parameter optimization process to obtain a set of quality intervention strategies; Calculate the risk control effect of each strategy in the set of quality intervention strategies based on the risk control objective function, calculate the resource utilization efficiency based on the risk control objective function and resource constraint conditions, and calculate the time response performance based on the time response constraint conditions; Construct a metric network, input the risk control effect, resource utilization efficiency and time response performance of each strategy into the metric network, train the metric network through a contrastive learning method to obtain a strategy scoring model, use the strategy scoring model to score each strategy in the set of quality intervention strategies, and select the strategy with the highest score as the optimal intervention plan.

[0012] In an alternative embodiment, Construct a policy network and a value network for solving the multi-objective constraint optimization equation, use the gradient policy method to optimize the parameters of the policy network and the value network, and introduce a meta-learning mechanism to dynamically adjust the optimization strategy during the parameter optimization process to obtain a set of quality intervention strategies, including: Construct a policy network and a value network for solving the multi-objective constraint optimization equation. Among them, the policy network extracts features from the input state through a multi-layer attention structure to obtain a state representation, and generates an action probability distribution based on the state representation; the value network includes a first feature extraction sub-network and a second feature extraction sub-network, generates a first state value and a second state value respectively based on the state representation obtained in the policy network, and fuses the first state value and the second state value through an adaptive weight coefficient to obtain an estimated state value; Construct a composite reward function, where the composite reward function includes the negative term of the risk control objective function in the multi-objective constrained optimization equation, a resource constraint penalty term, and a time constraint penalty term. The resource constraint penalty term is constructed based on the resource constraint condition, and the time constraint penalty term is constructed based on the time response constraint condition. Dynamically update the Lagrange multiplier based on the constraint violation degrees of the resource constraint penalty term and the time constraint penalty term to obtain a constraint weighting coefficient; Calculate an immediate reward value based on the composite reward function, and calculate a temporal difference error based on the immediate reward value and state value estimation. Construct a temporal advantage estimation value using the temporal difference error, and construct a policy gradient loss using the temporal advantage estimation value and the action probability distribution. At the same time, construct a value network loss based on the mean square error between the immediate reward value and the state value estimation; Construct a meta-policy network, input the gradient of the policy gradient loss with respect to the policy parameters into the meta-policy network, adjust the policy parameters based on the gradient update direction output by the meta-policy network, calculate the information entropy of the action probability distribution, and construct the weighted sum of the information entropy and the policy gradient loss as an optimization objective with an exploration term. Calculate an adaptive exploration coefficient based on the gradient magnitude of the policy gradient loss with respect to the policy parameters; Based on the optimization objective with an exploration term and the value network loss, use the projected gradient method to iteratively update the policy parameters and the value network parameters respectively. Project the updated policy parameters into the feasible region that satisfies the resource constraint condition and the time response constraint condition by minimizing the parameter offset, and alternately optimize the parameters of the policy network, the dual-stream value network, and the meta-policy network to obtain a set of quality intervention strategies.

[0013] In the second aspect of the embodiments of the present invention, there is provided an intelligent medical quality evaluation and early warning prediction system based on multi-dimensional indicators, including: A first unit for obtaining real-time medical data, calculating the fluctuation characteristic values of medical quality impact factors, constructing a quality sensitivity scoring model based on the fluctuation characteristic values, calculating the data sampling priority, establishing a hierarchical sampling frequency matrix, dynamically sampling the real-time medical data according to the hierarchical sampling frequency matrix to obtain sampling data, extracting the quality feature vectors of the sampling data and setting differential time windows, calculating the clinical correlation intensity and feature importance scores of the feature groups within the differential time windows, and performing feature fusion to obtain a medical quality score; A second unit, configured to perform multi-scale wavelet transform on the medical quality score to obtain a quality fluctuation component, calculate a local trend function of the quality fluctuation component at different time scales, construct a cross-correlation matrix between the local trend function and the quality fluctuation component, extract a quality evolution feature sequence, calculate a dynamic time warping distance to obtain a department quality status vector, construct a quality propagation graph based on the department quality status vector, fuse diagnosis and treatment path information to construct a heterogeneous information network, calculate a propagation influence factor of the heterogeneous information network, and identify quality control nodes; A third unit, configured to calculate the network topology entropy of the quality propagation graph, combine it with the sample entropy of the quality fluctuation component to obtain a quality risk entropy, construct a risk accumulation function based on the quality risk entropy, input the centrality index of the quality control node to obtain a risk propagation prediction sequence, construct a multi-objective constrained optimization equation based on the risk propagation prediction sequence, solve the multi-objective constrained optimization equation to obtain a set of quality intervention strategies, and screen an optimal intervention plan based on the set of quality intervention strategies.

[0014] The third aspect of the embodiments of the present invention provides an electronic device, 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 described above.

[0015] The fourth aspect of the embodiments of the present invention, provides 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.

[0016] In this embodiment, a quality sensitivity scoring model is constructed based on multi-dimensional indicators and fluctuation eigenvalues, and feature fusion is performed in combination with clinical association strength and feature importance, which can more accurately evaluate medical quality, avoid the limitations of single indicators, and achieve a comprehensive evaluation of medical quality. By methods such as multi-scale wavelet transform and dynamic time warping distance, a quality propagation graph is constructed and quality control nodes are identified, the quality trend, mutation probability, and cumulative deviation are predicted, and based on the risk entropy and multi-objective constrained optimization equation, precise intervention strategies are formulated, thereby effectively preventing and controlling medical quality risks. Dynamic sampling is performed based on the hierarchical sampling frequency matrix, avoiding data redundancy and improving data processing efficiency. At the same time, through the multi-objective constrained optimization equation, a balance is sought among risk control, resource constraints, and time response, optimizing the allocation of medical resources and improving the intervention efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a schematic flowchart of the method for intelligent evaluation and early warning prediction of medical quality based on multi-dimensional indicators in the embodiments of the present invention; Figure 2 It is a comparison chart of the classification accuracy of the quality status of different departments in the embodiments of the present invention; Figure 3 It is an analysis chart of the correlation between the dissemination influence factor and the quality index in the embodiments of the present invention; Figure 4 It is a comparison chart of the convergence performance of the multi-objective constraint optimization method in this embodiment. Detailed implementation manners

[0018] 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. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0019] 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.

[0020] Figure 1 It is a schematic flowchart of the intelligent medical quality evaluation and early warning prediction method based on multi-dimensional indicators in the embodiments of the present invention, as Figure 1 shown, and the method includes: Obtain real-time medical data, calculate the fluctuation characteristic values of medical quality influence factors, construct a quality sensitivity scoring model based on the fluctuation characteristic values, calculate the data sampling priority, establish a hierarchical sampling frequency matrix, perform dynamic sampling on the real-time medical data according to the hierarchical sampling frequency matrix to obtain sampling data, extract the quality feature vectors of the sampling data and set differential time windows, calculate the clinical correlation intensity and feature importance scores of the feature groups within the differential time windows, and perform feature fusion to obtain the medical quality score; Perform multi-scale wavelet transform on the medical quality score to obtain the quality fluctuation component, calculate the local trend function of the quality fluctuation component at different time scales, construct a cross-correlation matrix of the local trend function and the quality fluctuation component, extract the quality evolution feature sequence, calculate the dynamic time warping distance to obtain the department quality status vector, construct a quality propagation graph based on the department quality status vector, fuse the diagnosis and treatment path information to construct a heterogeneous information network, calculate the propagation influence factor of the heterogeneous information network, and identify the quality control nodes; Calculate the network topology entropy of the mass propagation graph, combine it with the sample entropy of the mass fluctuation component to obtain the mass risk entropy, construct a risk accumulation function based on the mass risk entropy, input the centrality index of the quality control node to obtain a risk propagation prediction sequence, construct a multi-objective constraint optimization equation based on the risk propagation prediction sequence, solve the multi-objective constraint optimization equation to obtain a set of quality intervention strategies, and screen the optimal intervention plan based on the set of quality intervention strategies.

[0021] Among them, the fluctuation eigenvalues include short-term fluctuation intensity (measuring the change range of data in a short time), medium-term fluctuation period (describing the change law of data in a medium time range), and long-term fluctuation trend (reflecting the change direction of data in a long time range).

[0022] Through the comprehensive evaluation of these eigenvalues, construct a quality sensitivity scoring model to quantify the sensitivity of data to medical quality. Based on the results of the scoring model, prioritize the real-time medical data to determine the sampling order of key data. Then establish a hierarchical sampling frequency matrix to ensure that high-priority data is sampled at a higher frequency and low-priority data is sampled at a lower frequency, so as to achieve dynamic sampling. After the sampled data is cleaned and processed, extract its quality feature vector, and set different time windows for different feature vectors to capture important feature changes within a specific time range. Then, analyze the clinical association strength (reflecting the correlation between features and clinical indicators) and feature importance score (quantifying the key importance of features in the overall data) of the feature group within the time window, and fuse the two to generate a medical quality score.

[0023] In an optional implementation manner, obtain real-time medical data, calculate the fluctuation eigenvalues of the medical quality impact factor, construct a quality sensitivity scoring model based on the fluctuation eigenvalues, calculate the data sampling priority, establish a hierarchical sampling frequency matrix, and perform dynamic sampling on the real-time medical data according to the hierarchical sampling frequency matrix to obtain sampled data, including: Input the fluctuation eigenvalues into a deep neural network, the deep neural network includes a feature extraction layer, an attention mechanism layer and a fully connected layer, and learn the combined weights of the short-term fluctuation intensity, medium-term fluctuation period and long-term fluctuation trend in the fluctuation eigenvalues through the deep neural network to construct a quality sensitivity scoring model, and calculate the quality sensitivity score based on the quality sensitivity scoring model; Perform a non-linear mapping on the quality sensitivity score to obtain an initial sampling priority, construct a double exponential smoothing model, input the initial sampling priority into the double exponential smoothing model, calculate the smoothing parameter based on the historical trend and future prediction of the real-time medical data, and smooth the initial sampling priority through the smoothing parameter to obtain the final sampling priority; Construct an adaptive hierarchical threshold according to the final sampling priority, divide the real-time medical data into multiple priority levels based on the adaptive hierarchical threshold to obtain a priority hierarchical result, calculate the data mean, variance, and distribution density of each level in the priority hierarchical result, and calculate the corresponding basic sampling frequency of each level based on the data mean, variance, and distribution density; Calculate the data timeliness decay value, quality relevance decay value, and business importance decay value and combine them to obtain a time decay factor. Construct a three-dimensional hierarchical sampling frequency matrix based on the time decay factor, basic sampling frequency, and quality sensitivity score. Query the sampling frequency from the three-dimensional hierarchical sampling frequency matrix based on the quality sensitivity score, and generate a sampling timestamp sequence that satisfies the sampling interval constraint according to the sampling frequency; Obtain data samples at the time points corresponding to the sampling timestamp sequence, determine the spline interpolation order based on the characteristics of the real-time medical data, interpolate and reconstruct the sampling interval of the data samples using the spline interpolation order to obtain reconstructed data, and detect and correct the outliers in the reconstructed data to obtain the final sampling data set.

[0024] Exemplarily, construct a quality sensitivity scoring model. Arrange the collected real-time medical data in chronological order, calculate the differences between adjacent data points to obtain a sequence of fluctuation characteristic values. Input these fluctuation characteristic values into a deep neural network. This deep neural network includes a feature extraction layer, an attention mechanism layer, and a fully connected layer. The feature extraction layer is used to extract the short-term fluctuation intensity in the fluctuation characteristic values, such as calculating the maximum fluctuation amplitude within every 10 seconds. The attention mechanism layer focuses on the medium-term fluctuation period, such as identifying the regular changes in the heart rate per minute. The fully connected layer integrates the short-term fluctuation intensity, medium-term fluctuation period, and long-term fluctuation trend (such as the overall upward or downward trend of the patient's blood pressure), and learns the combined weights between them, and finally outputs a quality sensitivity score. For example, if the patient's heart rate suddenly accelerates, the short-term fluctuation intensity will be high, the attention mechanism will focus on this change, and the fully connected layer combines the long-term trend to judge the severity of the condition and finally gives a higher quality sensitivity score.

[0025] Calculate the data sampling priority based on the quality sensitivity score. Perform a non-linear mapping on the quality sensitivity score, for example, using the Sigmoid function, to map the score to a value between 0 and 1 to obtain the initial sampling priority. To avoid drastic changes in the sampling frequency caused by fluctuations in a single score, use a double exponential smoothing model to smooth the initial sampling priority. This model takes into account historical trends and future predictions. For example, if a patient's blood pressure continues to rise, even if the short-term fluctuations are small, the model will predict that future fluctuations may increase, thus increasing the sampling priority. Dynamically adjust the smoothing parameter according to the changes in real-time data. For example, when the data fluctuates greatly, reduce the smoothing parameter to respond to changes more quickly; when the data fluctuates little, increase the smoothing parameter to keep the sampling frequency stable. The result after smoothing is the final sampling priority.

[0026] Establish a hierarchical sampling frequency matrix. Set an adaptive hierarchical threshold according to the final sampling priority, and divide the real-time medical data into different priority levels. For example, divide the priorities from high to low into three levels: high, medium, and low. Calculate the mean, variance, and distribution density of the data within each level respectively. For example, the mean of the data in the high-priority level represents the average value of the data in this level, the variance represents the degree of data fluctuation, and the distribution density describes the concentration degree of the data within different value ranges. Set a basic sampling frequency for each level according to these statistical characteristics. For example, since the data in the high-priority level fluctuates greatly, set a higher basic sampling frequency; since the data in the low-priority level is relatively stable, set a lower basic sampling frequency.

[0027] To further optimize the sampling strategy, introduce a time decay factor. Calculate the data timeliness decay value. For example, the timeliness of data closer to the current time is higher and the decay value is smaller. Calculate the data quality correlation decay value. For example, the decay value of indicators related to the current condition is smaller. Calculate the business importance decay value. For example, the decay value of indicators related to key treatment links is smaller. Combine these decay values to obtain the final time decay factor.

[0028] Construct a three-dimensional hierarchical sampling frequency matrix based on the basic sampling frequency, quality sensitivity score, and time decay factor. Query the final sampling frequency from the matrix according to the quality sensitivity score. Generate a sampling timestamp sequence that meets the sampling interval constraint based on the sampling frequency. For example, if the sampling frequency is once per second, the timestamp sequence is 1 second, 2 seconds, 3 seconds, etc. Obtain data samples at the moments corresponding to these timestamps. Determine the order of spline interpolation according to the characteristics of the real-time medical data, such as the smoothness of data changes. Use spline interpolation to interpolate and reconstruct the sampling interval of the data samples to obtain denser data. Detect and correct possible outliers in the reconstructed data, such as values outside the normal physiological range, and finally obtain a high-quality sampling data set.

[0029] Suppose the blood pressure data of a patient rapidly increases within a short period, and its fluctuation eigenvalue will be relatively large, resulting in a high quality sensitivity score. The system will classify it into the high-priority level and assign a relatively high basic sampling frequency. At the same time, due to the high timeliness of the data, the time decay factor is relatively small. Finally, the system will query a relatively high sampling frequency from the three-dimensional matrix, such as sampling once per second, and generate a sampling timestamp sequence accordingly. The system will collect blood pressure data at these time points and use spline interpolation to reconstruct the data, finally obtaining a sampling data set containing dense blood pressure data.

[0030] In this embodiment, through the quality sensitivity score and the hierarchical sampling strategy, important data is preferentially collected, over-sampling of low-value data is avoided, and the data sampling efficiency is improved. Through spline interpolation and outlier correction, the integrity and accuracy of the sampling data are guaranteed, and the data quality is improved. Through the adaptive hierarchical threshold and the dynamic smoothing parameter, the system can dynamically adjust the sampling strategy according to the changes in real-time data, enhancing the adaptability of the system.

[0031] In an alternative embodiment, perform multi-scale wavelet transform on the medical quality score to obtain the quality fluctuation component, calculate the local trend function of the quality fluctuation component at different time scales, construct the cross-correlation matrix between the local trend function and the quality fluctuation component, extract the quality evolution feature sequence, and calculate the dynamic time warping distance to obtain the department quality status vector, including: Construct an asymmetric wavelet basis function with adaptive time-frequency resolution, and perform continuous wavelet transform on the medical quality score based on the asymmetric wavelet basis function to obtain the quality fluctuation components in different frequency domains; Decompose the quality fluctuation component into its morphological components, extract the basic morphological elements of the quality fluctuation component, and use a sliding time window to perform smooth fitting on the basic morphological elements to obtain the local trend functions at different time scales; Perform fractional-order differential operations on the local trend function and the quality fluctuation component respectively, construct a fractional-order differential operator, and obtain the phase evolution sequence; calculate the instantaneous phase and instantaneous amplitude of the phase evolution sequence, calculate the generalized complementary coefficient at multiple time scales based on the instantaneous phase and instantaneous amplitude, and construct a cross-correlation matrix reflecting multi-scale characteristics; Perform spectral decomposition of the dynamic system on the cross-correlation matrix to obtain the eigenvalue sequence and eigenvector sequence, and construct a non-linear feature projection transformer based on the eigenvalue sequence and eigenvector sequence; input the cross-correlation matrix into the feature projection transformer for non-linear dimensionality reduction to obtain the quality evolution feature sequence; Calculate the geodesic distance between adjacent feature vectors in the quality evolution feature sequence, and organize the geodesic distances into a local distance matrix; solve the optimal alignment path for the local distance matrix using a constrained dynamic programming algorithm; calculate the quality trend prediction value based on the optimal alignment path and the quality evolution feature sequence; Extract the local deformation gradient of the optimal alignment path, input the local deformation gradient into a recurrent neural network, and calculate the quality mutation probability through a non-linear activation function and a temporal memory unit; Perform dynamic time warping on the quality evolution feature sequence and a standard sequence to obtain a warping distance sequence, and use a time weight function to perform weighted accumulation on the warping distance sequence to obtain the quality cumulative deviation; Normalize and perform dimensional transformation on the quality trend prediction value, quality mutation probability, and quality cumulative deviation in the feature space to construct a standardized department quality status vector.

[0032] Exemplarily, first obtain the department medical quality score data. The department medical quality score data can be obtained from multiple channels such as the hospital information system, electronic medical record system, patient satisfaction survey, etc. For example, the monthly average patient satisfaction scores of a certain department in the past year can be collected to form time series data. Construct an asymmetric wavelet basis function with adaptive time-frequency resolution. The reason for using the asymmetric wavelet basis function is that it can better capture the asymmetric fluctuation patterns in the medical quality score data. For example, the process of quality improvement may be relatively slow, while the process of quality decline may be relatively rapid. By adjusting the shape parameters of the wavelet basis function, it can be made to better match the fluctuation patterns of the actual data. For example, select an asymmetric wavelet basis function that can adapt to rapid decline and slow rise.

[0033] Perform continuous wavelet transform on the medical quality score data. Use the constructed asymmetric wavelet basis function to perform continuous wavelet transform on the medical quality score data to obtain the quality fluctuation components in different frequency domains. The continuous wavelet transform can decompose the original medical quality score data into fluctuation components of different frequencies. For example, high-frequency components can reflect short-term quality fluctuations, and low-frequency components can reflect long-term quality trends.

[0034] Perform morphological decomposition on the quality fluctuation components. Decompose the quality fluctuation components into several basic morphological elements, which can represent different patterns in the quality fluctuations, such as upward trends, downward trends, periodic fluctuations, etc. Assume that the basic morphological elements after decomposing the high-frequency components are upward trend, downward trend, and periodic fluctuation.

[0035] Smoothly fit the basic morphological elements. Use a sliding time window to smoothly fit the basic morphological elements to obtain local trend functions at different time scales. The size of the sliding time window can be adjusted according to the actual situation. For example, a month, a quarter, or a year can be used as the time window. Assume that a three-month time window is used to smoothly fit the upward trend in the high-frequency component to obtain the local trend function.

[0036] Perform fractional-order differential operations on the local trend function and the quality fluctuation component. Fractional-order differential operations can be used to describe the local change characteristics of a function. For example, they can be used to calculate the slope, curvature, etc. of a function. Assume that after performing fractional-order differential operations on the high-frequency component, a phase evolution sequence is obtained.

[0037] Calculate the instantaneous phase and instantaneous amplitude of the phase evolution sequence. The instantaneous phase and instantaneous amplitude can be used to describe the dynamic change characteristics of the phase evolution sequence. Assume that the calculated instantaneous phase sequence is: [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9], and the instantaneous amplitude sequence is: [1, 2, 1, 2, 1, 2, 1, 2, 1].

[0038] Calculate the generalized complementary coefficient based on the instantaneous phase and instantaneous amplitude, and construct a cross-correlation matrix. The generalized complementary coefficient can be used to measure the correlation between the local trend function and the quality fluctuation component at different time scales. Assume that the constructed cross-correlation matrix is a 3x3 matrix, where each element represents the correlation coefficient between the local trend function and the quality fluctuation component at different time scales.

[0039] Perform spectral decomposition of the cross-correlation matrix for the dynamic system. Through spectral decomposition, the cross-correlation matrix can be decomposed into an eigenvalue sequence and an eigenvector sequence, and these eigenvalues and eigenvectors can be used to describe the main patterns in the quality evolution process.

[0040] Construct a non-linear feature projection transformer based on the eigenvalue sequence and the eigenvector sequence. The non-linear feature projection transformer can project the high-dimensional cross-correlation matrix into a low-dimensional feature space, thereby extracting the most representative quality evolution features.

[0041] Input the cross-correlation matrix into the feature projection transformer for non-linear dimensionality reduction to obtain the quality evolution feature sequence. The quality evolution feature sequence can be used to describe the evolution trend of the department's medical quality.

[0042] Calculate the geodesic distance between adjacent eigenvectors in the quality evolution feature sequence. The geodesic distance can be used to measure the similarity between eigenvectors.

[0043] The optimal alignment path is solved for the local distance matrix using a constrained dynamic programming algorithm. The optimal alignment path can be used to describe the most likely evolution trajectory during the quality evolution process.

[0044] The quality trend prediction value is calculated based on the optimal alignment path and the quality evolution feature sequence. The quality trend prediction value can be used to predict the change trend of the medical quality of the department in a future period.

[0045] The local deformation gradient of the optimal alignment path is extracted. The local deformation gradient can be used to describe the possibility of mutation during the quality evolution process.

[0046] The local deformation gradient is input into a recurrent neural network to calculate the quality mutation probability. The quality mutation probability can be used to predict the possibility of a mutation in the medical quality of the department in a future period.

[0047] The quality evolution feature sequence and the standard sequence are subjected to dynamic time warping to obtain a warping distance sequence. The standard sequence can be the best quality level in history or the average quality level of departments of the same type.

[0048] The time weight function is used to perform weighted accumulation on the warping distance sequence to obtain the quality cumulative deviation. The quality cumulative deviation can be used to measure the gap between the medical quality of the department and the standard sequence.

[0049] The quality trend prediction value, the quality mutation probability, and the quality cumulative deviation are normalized and dimension-transformed in the feature space to construct a standardized department quality state vector. The department quality state vector can be used to comprehensively evaluate the medical quality of the department.

[0050] Figure 2 This is a comparison chart of the classification accuracy of the quality states of different departments in the embodiments of the present invention, as Figure 2As shown, the figure shows the comparison of the accuracy rates of different methods in the quality status classification of each medical department. It can be seen from the figure that the technical solution of the present invention shows excellent classification accuracy rates in all departments. In the internal medicine department, the accuracy rate of the technical solution of the present invention reaches 95.7%, while that of the SVM classifier is 87.2%, that of the random forest is 85.3%, and that of the KNN classifier is 82.1%. In the surgical field, the accuracy rate of the technical solution of the present invention is 93.8%, leading the 85.9%, 83.7% and 79.6% of other methods. The pediatric field is particularly prominent. The accuracy rate of the technical solution of the present invention is as high as 96.2%, which is 9.5 percentage points higher than that of the second-place SVM classifier. In complex departments such as the obstetrics and gynecology department and the emergency department, the technical solution of the present invention still maintains high accuracy rates of 88.9% and 87.3%, while other methods generally drop below 80%. Particularly noteworthy is that in departments with complex and variable data characteristics such as the medical technology department (such as the laboratory department and the imaging department), the accuracy rate of the technical solution of the present invention is 91.4%, far higher than the 78.5%, 76.3% and 72.8% of other methods. This result shows that the method for constructing the department quality status vector integrating multi-scale wavelet transform and dynamic time warping in the technical solution of the present invention can more comprehensively capture the unique characteristics of the quality evaluation of different departments, realize more accurate quality status classification, and provide strong support for the refined management of the hospital.

[0051] Through multi-scale wavelet transform and morphological decomposition, the fluctuation patterns and trend changes in the medical quality scoring data can be captured more precisely, thereby improving the accuracy of medical quality assessment. It is possible to monitor and predict the medical quality of departments in real time, timely discover potential quality problems, and issue early warnings to provide decision-making support for hospital managers. It can help hospital managers better understand the evolution law of the medical quality of departments, formulate more targeted improvement measures, and promote the continuous improvement of medical quality.

[0052] In an optional implementation manner, a quality propagation graph is constructed based on the department quality status vector, heterogeneous information networks are constructed by integrating diagnosis and treatment path information, the propagation influence factors of the heterogeneous information networks are calculated, and the quality control nodes are identified, including: Construct a kernel function mapping matrix, perform a non-linear mapping on the department quality status vector to obtain a mapped feature vector; calculate the cosine similarity of the mapped feature vectors between department nodes to construct a similarity matrix; construct an initial quality propagation graph based on the similarity matrix, where a quality status edge is established between nodes when the similarity value corresponding to the department node exceeds a preset dynamic threshold, and the weight of the quality status edge is the corresponding similarity value; Construct a time-sliding window based on the patient transfer records, perform exponential time series weighting on the patient transfer frequency within the window to obtain the time series weighted transfer frequency, and perform normalization processing on the time series weighted transfer frequency to obtain the weight value of the patient transfer edge; extract the consultation frequency and examination application frequency based on the doctor's order execution records, and combine them after assigning weight coefficients according to the business type to obtain the weight value of the medical collaboration edge; Integrate the patient transfer edge, medical collaboration edge and their corresponding weight values into the initial quality propagation graph to construct a heterogeneous information network; for each department node in the heterogeneous information network, construct a multi-layer attention network based on the features of its first-order neighbor nodes on the quality status edge, patient transfer edge and medical collaboration edge; use the multi-layer attention network to calculate the weight contribution coefficients of each edge respectively, and perform weighted summation of the weight contribution coefficients and the weight values of the corresponding edges to obtain the propagation influence factor of the department node; Construct a random walk iteration rule based on the topological structure of the heterogeneous information network, iteratively update the propagation influence factor according to the random walk iteration rule, calculate the Euclidean distance of the propagation influence factor sequence obtained from two adjacent iterations, and obtain the final propagation influence factor when the Euclidean distance is less than the convergence threshold. Rank the department nodes according to the final propagation influence factor, and combine the in-degree distribution and betweenness centrality of each department node on the three types of edges, and use the analytic hierarchy process to identify the quality control nodes.

[0053] Exemplarily, first, collect the quality status data of the departments, such as indicators like patient satisfaction, average length of stay, and postoperative complication rate, to construct the department quality status vector. Taking the Department of Cardiology as an example, assume that the patient satisfaction is 95 points, the average length of stay is 7 days, and the postoperative complication rate is 2%. Then the quality status vector of the Department of Cardiology is [95, 7, 2]. Similar vectors are constructed for each department.

[0054] Then, use a kernel function to perform non-linear mapping on the department quality status vector. For example, use the radial basis kernel function to map the vector [95, 7, 2] of the Department of Cardiology to a higher-dimensional feature space to obtain the mapped feature vector. The same mapping operation is performed on other departments.

[0055] Next, calculate the cosine similarity of the mapped feature vectors between department nodes to construct a similarity matrix. For example, calculate the cosine similarity between the mapped feature vectors of the Department of Cardiology and the Department of Respiratory Medicine as 0.85, and the similarity between the Department of Cardiology and the Department of Orthopedics as 0.6. Then, in the similarity matrix, the corresponding value between the Department of Cardiology and the Department of Respiratory Medicine is 0.85, and the corresponding value between the Department of Cardiology and the Department of Orthopedics is 0.6, and so on, to construct a complete similarity matrix.

[0056] Construct an initial quality propagation graph based on the similarity matrix. Set a dynamic threshold, such as 0.7. If the similarity value between two department nodes exceeds this threshold, establish a quality status edge between these two nodes, and the weight of the edge is the corresponding similarity value. For example, the similarity between the Department of Cardiology and the Department of Respiratory Medicine is 0.85, which exceeds the threshold of 0.7, so a quality status edge with a weight of 0.85 is established between the Department of Cardiology and the Department of Respiratory Medicine.

[0057] Next, construct a time-sliding window based on the patient transfer records between departments. For example, the window size is set to 7 days. Count the transfer frequency of patients between departments within each window, and perform exponential time-series weighting on the frequency so that recent data has a greater impact on the weight. For example, there are 5 patients transferred between the Department of Cardiology and the Department of Respiratory Medicine in the past 7 days, and the weighted transfer frequency after exponential time-series weighting is 6.2. Normalize the weighted transfer frequencies between all departments to obtain the weight values of the patient transfer edges. For example, the weight value of the patient transfer edge between the Department of Cardiology and the Department of Respiratory Medicine is 0.15.

[0058] Then, extract the consultation frequency and examination application frequency based on the doctor's order execution records. For example, the Department of Cardiology applies for examinations 10 times to the Department of Radiology and 5 times to the Department of Ultrasound. Assign weight coefficients according to the business type. For example, the consultation weight coefficient is 0.8, and the examination application weight coefficient is 0.5. Combine the weighted consultation frequency and examination application frequency to obtain the weight value of the medical collaboration edge. For example, the weight value of the medical collaboration edge between the Department of Cardiology and the Department of Radiology is 0.8×10 + 0.5×0 = 8.

[0059] Integrate the patient transfer edges, medical collaboration edges and their corresponding weight values into the initial quality propagation graph to construct a heterogeneous information network.

[0060] For each department node in the heterogeneous information network, construct a multi-layer attention network based on the features of its first-order neighborhood nodes on the quality status edge, patient transfer edge and medical collaboration edge. Use the multi-layer attention network to calculate the weight contribution coefficients of each edge respectively. For example, the weight contribution coefficient of the quality status edge of the Department of Cardiology node is 0.4, the weight contribution coefficient of the patient transfer edge is 0.3, and the weight contribution coefficient of the medical collaboration edge is 0.3. Perform weighted summation of the weight contribution coefficients and the weight values of the corresponding edges to obtain the propagation influence factor of the department node. For example, the propagation influence factor of the Department of Cardiology is 0.4×0.85 + 0.3×0.15 + 0.3×8 = 2.785.

[0061] Construct a random walk iteration rule based on the topological structure of the heterogeneous information network, and iteratively update the propagation influence factor according to the random walk iteration rule. Calculate the Euclidean distance of the propagation influence factor sequences obtained from two adjacent iterations. When the Euclidean distance is less than the convergence threshold, such as 0.001, obtain the final propagation influence factor.

[0062] Rank the department nodes according to the final propagation influence factor. Combining the in-degree and out-degree distributions and betweenness centrality of each department node on three types of edges, use the analytic hierarchy process to identify quality control nodes.

[0063] Figure 3 This is the correlation analysis diagram of the propagation influence factor and quality indicators in the embodiment of the present invention, as Figure 3 shown. This diagram analyzes the correlation between the propagation influence factors calculated by different methods and the hospital quality indicators. The horizontal axis represents 5 main medical quality indicators: the incidence of medical adverse events, the average length of stay, the antibiotic usage rate, the surgical complication rate, and patient satisfaction; the vertical axis represents the correlation coefficient (absolute value) with these indicators. The results show that the correlation coefficient between the propagation influence factor calculated by this technical solution and the incidence of medical adverse events reaches 0.87, which is significantly higher than that of the improved PageRank algorithm (0.76) and the traditional centrality index (0.65); the correlation coefficient with the average length of stay is 0.82, which is also higher than the other two methods (0.71 and 0.59 respectively). For the antibiotic usage rate, the surgical complication rate, and patient satisfaction, the correlation coefficients of this technical solution are 0.79, 0.83, and 0.81 respectively, all of which are better than the control methods. This indicates that by integrating multi-dimensional information such as quality status, patient flow, and medical collaboration, the propagation influence factor calculated by this technical solution can more accurately reflect the actual importance of departments in the hospital quality management system, providing a scientific basis for accurately identifying quality control nodes.

[0064] In this embodiment, by integrating multi-source heterogeneous information, departments that have the greatest impact on the overall medical quality can be more accurately identified, thereby achieving precise quality control and avoiding waste of resources. By focusing on and intervening in key departments for quality control, the overall medical quality can be more effectively improved, the allocation of medical resources can be optimized, and the efficiency of medical management can be enhanced. By analyzing the quality propagation relationship and collaboration mode between departments, learning and communication between departments can be promoted, and the overall collaborative development of medical services can be promoted.

[0065] In an alternative embodiment, calculate the network topological entropy of the quality propagation graph, combine it with the sample entropy of the quality fluctuation component to obtain the quality risk entropy, construct a risk accumulation function based on the quality risk entropy, and input the centrality index of the quality control node to obtain a risk propagation prediction sequence, including: Calculate the connection probability between a node and its adjacent nodes in the quality propagation graph, perform a logarithmic operation on the connection probability and multiply it by the original connection probability to obtain the local entropy of the node; calculate the degree distribution and betweenness centrality of the node based on the connection probability, and construct a node importance score; use the node importance score to calculate the weight coefficient, and perform a weighted sum of the local entropy of each node and the corresponding weight coefficient to obtain the network topological entropy of the quality propagation graph; Segment and normalize the quality fluctuation component to obtain a standardized sequence, and construct a template sequence based on the standardized sequence according to a preset embedding dimension; count the number of matches of the template sequence under a given similarity tolerance; calculate the natural logarithm ratio of the number of template matches in adjacent dimensions to obtain the sample entropy of the quality fluctuation component. Construct a non - linear combination model including the network topology entropy and the sample entropy, where the non - linear combination model includes linear terms, square terms, and cross terms; construct an optimization objective function based on historical risk data and node importance scores; solve the optimization objective function to obtain combination coefficients, and substitute the combination coefficients, network topology entropy, and sample entropy into the non - linear combination model to obtain the quality risk entropy. Use the quality risk entropy to construct an exponential cumulative function, map the quality risk entropy to a cumulative rate coefficient, construct a time - weight function based on the time characteristics of the standardized sequence, and combine the time - weight function with the cumulative function to obtain a risk cumulative function. Extract the centrality index of the quality control node, use the principal component analysis method to reduce the dimension and fuse the centrality index, smooth the fused centrality index, and input the smoothed centrality index into the risk cumulative function to obtain a prediction sequence reflecting the risk propagation degree of the quality control node.

[0066] Exemplarily, first, construct a quality propagation graph. According to the quality correlation relationship between the components in the system, construct a directed or undirected quality propagation graph. The nodes in the graph represent different components in the system, and the edges represent the quality influence relationship between them. For example, in the manufacturing process, raw materials, parts, production equipment, etc. can be used as nodes, and the processing, assembly, etc. relationships between them can be used as edges. The weight of the edge can be set according to the degree of quality influence. For example, the degree of influence of the quality defect of a certain part on the quality of the final product.

[0067] Calculate the network topological entropy of the quality propagation graph. This step aims to quantify the structural complexity and information transfer efficiency of the quality propagation graph. First, calculate the connection probability between nodes and their adjacent nodes. Then, perform a logarithmic operation on the connection probability and multiply it by the original connection probability to obtain the local entropy of the node, which is used to characterize the information transfer ability of the node itself. To measure the importance of a node in the network, calculate the degree distribution (number of connections) and betweenness centrality (the number of times a node appears in the shortest paths in the network). Combine the degree distribution and betweenness centrality to construct a node importance score. Finally, use the node importance score to calculate the weight coefficient, and perform a weighted sum of the local entropy of each node and the corresponding weight coefficient to obtain the network topological entropy of the quality propagation graph. For example, if the connection probability of a certain node is 0.8, then its local entropy is 0.8×log(0.8). Suppose the degree of this node is 3 and the betweenness centrality is 10, then its importance score can be calculated according to a pre-set rule. For example, importance score = 0.5×degree + 0.5×betweenness centrality.

[0068] Calculate the sample entropy of the quality fluctuation component. This step aims to quantify the complexity and uncertainty of the quality fluctuation sequence. First, segment and normalize the quality fluctuation component to obtain a standardized sequence. For example, divide the original quality data according to time periods and perform min-max normalization. Then, construct template sequences based on the standardized sequence according to a preset embedding dimension (for example, the dimension is 2). Given a similarity tolerance (for example, 0.2), count the number of matches of the template sequences. Finally, calculate the natural logarithm ratio of the number of template matches in adjacent dimensions to obtain the sample entropy of the quality fluctuation component.

[0069] Construct the quality risk entropy. Combine the network topological entropy and the sample entropy to obtain the quality risk entropy, which is used to comprehensively characterize the system quality risk. Construct a non-linear combination model that includes the network topological entropy and the sample entropy. This model includes linear terms, square terms, and cross terms. Based on historical risk data and node importance scores, construct an optimization objective function. For example, minimize the difference between the predicted risk and the actual risk. Solve the optimization objective function to obtain the combination coefficients. Substitute the combination coefficients, the network topological entropy, and the sample entropy into the non-linear combination model to obtain the quality risk entropy.

[0070] Construct the risk accumulation function. Use the quality risk entropy to construct an exponential cumulative function, which maps the quality risk entropy to a cumulative rate coefficient. Based on the time characteristics of the standardized sequence, construct a time weight function. For example, as time goes by, the weight gradually increases. Combine the time weight function and the cumulative function to obtain the risk accumulation function.

[0071] Finally, predict the risk propagation of quality control nodes. Extract the centrality indicators of quality control nodes, such as degree centrality, closeness centrality, betweenness centrality, etc. Use the principal component analysis method to reduce the dimension and fuse the centrality indicators. Smooth the fused centrality indicators, for example, using the moving average method. Input the smoothed centrality indicators into the risk accumulation function to obtain a prediction sequence reflecting the degree of risk propagation of quality control nodes.

[0072] Suppose there are two quality control nodes, and their fused centrality indicators are 0.8 and 0.6 respectively. Input these two values into the risk accumulation function to obtain the corresponding risk prediction values of 0.9 and 0.7. Then it can be considered that the node with a centrality indicator of 0.8 has a higher degree of risk propagation.

[0073] In this embodiment, the network topology structure and quality fluctuation characteristics are comprehensively considered, which can more comprehensively reflect the system quality risk, thereby improving the risk prediction accuracy. Decompose the quality risk into network topology entropy and sample entropy, and correlate them through the risk accumulation function, making the risk prediction result more interpretable. It is applicable to various complex systems and can be adjusted according to specific application scenarios, with strong practicality.

[0074] In an alternative embodiment, based on the risk propagation prediction sequence, construct a multi-objective constraint optimization equation, solve the multi-objective constraint optimization equation to obtain a set of quality intervention strategies, and screen the optimal intervention plan based on the set of quality intervention strategies, including: Use a bidirectional temporal network to process the risk propagation prediction sequence to obtain a hidden layer state sequence, calculate the corresponding risk weight sequence based on the hidden layer state sequence through the self-attention mechanism, and perform weighted fusion of the risk weight sequence and the hidden layer state sequence to obtain risk features; Based on the risk features, construct an adaptive weight function and a dynamic fluctuation penalty function respectively, and combine the product of the adaptive weight function and the risk value and the product of the dynamic fluctuation penalty function and the fluctuation amplitude to construct a risk control objective function; Construct a resource interaction graph according to the resource type and resource collaboration relationship, use the graph attention network to perform node representation learning on the resource interaction graph, based on the obtained node representations, construct resource constraint conditions through the gated graph network, and at the same time construct a state space based on the risk features, model the intervention process as a state-action sequence in the state space, use the dual value network to calculate the intervention time sequence corresponding to the state-action sequence, and construct a time response constraint condition by the weighted sum of the intervention time sequence and the time weight; Combine the risk control objective function, resource constraint conditions, and time response constraint conditions to construct a multi-objective constraint optimization equation; Construct a policy network and a value network for solving the multi-objective constrained optimization equation, optimize the parameters of the policy network and the value network using the gradient policy method, introduce a meta-learning mechanism to dynamically adjust the optimization strategy during the parameter optimization process, and obtain a set of quality intervention strategies; Calculate the risk control effect of each strategy in the set of quality intervention strategies based on the risk control objective function, calculate the resource utilization efficiency based on the risk control objective function and the resource constraint conditions, and calculate the time response performance based on the time response constraint conditions; Construct a metric network, input the risk control effect, resource utilization efficiency, and time response performance of each strategy into the metric network, train the metric network through a contrastive learning method to obtain a strategy scoring model, and use the strategy scoring model to score each strategy in the set of quality intervention strategies, and select the strategy with the highest score as the optimal intervention plan.

[0075] Exemplarily, a bidirectional time series network is used to process the risk propagation prediction sequence. For example, a bidirectional long short-term memory network (Bi-LSTM) is used, and the risk propagation prediction sequence is used as the input. The hidden layer state of the network contains information from both the past and the future at each time step, forming a sequence of hidden layer states. Taking a product defect rate prediction sequence as an example, the predicted defect rates for the next five days are 0.1, 0.12, 0.15, 0.11, and 0.09 respectively. This sequence is input into the Bi-LSTM network to obtain the corresponding hidden layer state at each time step.

[0076] Based on the obtained sequence of hidden layer states, calculate the risk weight sequence through the self-attention mechanism. The self-attention mechanism can capture the dependencies between different time steps in the sequence, thereby assigning different weights to different time steps. For example, the sequence of hidden layer states can be input into a self-attention layer, which will output a weight sequence indicating the importance of each time step. Suppose the obtained weight sequence is 0.2, 0.25, 0.3, 0.15, 0.1.

[0077] Perform weighted fusion of the risk weight sequence and the hidden layer state sequence to obtain the risk feature. Specifically, multiply the hidden layer state at each time step by its corresponding weight, and then sum the results of all time steps to obtain the final risk feature vector. For example, multiply and sum the hidden layer state and weight in the above example to obtain the fused risk feature.

[0078] Based on risk characteristics, construct an adaptive weight function and a dynamic volatility penalty function. The adaptive weight function dynamically adjusts the weight of the risk value according to the magnitude of the risk characteristics. The dynamic volatility penalty function dynamically adjusts the penalty intensity of the volatility amplitude according to the change amplitude of the risk characteristics. For example, the larger the risk characteristic value, the larger the value of the adaptive weight function, and the greater the weight of the risk value in the objective function; the greater the change amplitude of the risk characteristics, the larger the value of the dynamic volatility penalty function, and the greater the weight of the volatility amplitude in the objective function. Combine the product of the adaptive weight function and the risk value and the product of the dynamic volatility penalty function and the volatility amplitude to construct a risk control objective function.

[0079] Construct a resource interaction graph based on resource types and resource collaboration relationships. For example, if the resource types include human resources, funds, and equipment, their collaboration relationships can be expressed as: human resources operate equipment, and funds are used to purchase equipment and pay labor costs. These relationships can be represented by a graph, where nodes represent resource types and edges represent collaboration relationships. Use a graph attention network to perform node representation learning on the resource interaction graph. The graph attention network can learn the importance of each resource node and its relationship with other nodes. Based on the obtained node representations, construct resource constraint conditions through a gated graph network. The gated graph network can dynamically adjust resource allocation according to node representations and resource consumption conditions, so as to meet resource constraint conditions.

[0080] At the same time, construct a state space based on risk characteristics, and model the intervention process as a state-action sequence in the state space. For example, the state can represent the current risk level and resource allocation situation, and the action can represent different intervention measures. Use a dual value network to calculate the intervention time sequence corresponding to the state-action sequence, that is, the execution time of each intervention action. Construct the weighted sum of the intervention time sequence and the time weight as the time response constraint condition.

[0081] Combine the risk control objective function, resource constraint conditions, and time response constraint conditions to construct a multi-objective constraint optimization equation.

[0082] Construct a policy network and a value network for solving the multi-objective constraint optimization equation. The policy network is used to generate intervention policies, and the value network is used to evaluate the quality of the policies. Use the gradient policy method to optimize the parameters of the policy network and the value network. Introduce a meta-learning mechanism to dynamically adjust the optimization policy during the parameter optimization process to obtain a set of quality intervention policies.

[0083] Calculate the risk control effect of each policy in the set of quality intervention policies based on the risk control objective function, calculate the resource utilization efficiency based on the risk control objective function and resource constraint conditions, and calculate the time response performance based on the time response constraint conditions.

[0084] Construct a metric network, input the risk control effect, resource utilization efficiency, and time response performance of each policy into the metric network, and train the metric network through a contrastive learning method to obtain a policy scoring model. Use the policy scoring model to score each policy in the quality intervention policy set, and select the policy with the highest score as the optimal intervention plan.

[0085] In this embodiment, the intervention policy can be dynamically adjusted according to the risk propagation prediction sequence and resource constraint conditions, so as to more effectively control risks. By constructing a resource interaction graph and using a graph attention network, this method can optimize resource allocation and improve resource utilization efficiency. Considering the time response constraint conditions and using a dual value network to calculate the intervention time sequence, the intervention can be completed in a shorter time and the time response speed can be improved.

[0086] In an alternative implementation, construct a policy network and a value network for solving the multi-objective constrained optimization equation, use the gradient policy method to optimize the parameters of the policy network and the value network, and introduce a meta-learning mechanism to dynamically adjust the optimization policy during the parameter optimization process. The obtained quality intervention policy set includes: Construct a policy network and a value network for solving the multi-objective constrained optimization equation. Among them, the policy network extracts features from the input state through a multi-layer attention structure to obtain a state representation, and generates an action probability distribution based on the state representation; the value network includes a first feature extraction sub-network and a second feature extraction sub-network, generates a first state value and a second state value respectively based on the state representation obtained in the policy network, and fuses the first state value and the second state value through an adaptive weight coefficient to obtain a state value estimate; Construct a composite reward function, which includes a negative term of the risk control objective function in the multi-objective constrained optimization equation, a resource constraint penalty term, and a time constraint penalty term. The resource constraint penalty term is constructed based on the resource constraint conditions, and the time constraint penalty term is constructed based on the time response constraint conditions. Dynamically update the Lagrange multiplier based on the constraint violation degrees of the resource constraint penalty term and the time constraint penalty term to obtain a constraint weighting coefficient; Calculate the immediate reward value based on the composite reward function, and calculate the temporal difference error based on the immediate reward value and the state value estimate. Use the temporal difference error to construct a temporal advantage estimate value, use the temporal advantage estimate value and the action probability distribution to construct a policy gradient loss, and at the same time construct a value network loss based on the mean square error between the immediate reward value and the state value estimate; Construct a meta-policy network, input the gradient of the policy gradient loss with respect to the policy parameters into the meta-policy network, adjust the policy parameters based on the gradient update direction output by the meta-policy network, calculate the information entropy of the action probability distribution, and construct the weighted sum of the information entropy and the policy gradient loss as an optimization objective with an exploration term. Calculate the adaptive exploration coefficient based on the magnitude of the gradient of the policy gradient loss with respect to the policy parameters. Based on the optimization objective with an exploration term and the value network loss, use the projected gradient method to iteratively update the policy parameters and the value network parameters respectively. Project the updated policy parameters into the feasible region that satisfies the resource constraint condition and the time response constraint condition by minimizing the parameter offset, and alternately optimize the parameters of the policy network, the dual-stream value network, and the meta-policy network to obtain a set of quality intervention policies.

[0087] Exemplarily, construct a policy network and a value network. The policy network uses a multi-layer attention structure to extract the features of the input state and generate a state representation. Based on this state representation, the policy network outputs an action probability distribution, that is, the probabilities of performing various actions in the current state. The value network contains two feature extraction sub-networks, which respectively evaluate the first state value and the second state value based on the state representation generated by the policy network. These two state values are fused through an adaptive weight coefficient to obtain the final state value estimate, which is used to evaluate the long-term value of the current state. For example, the state can be the parameter configuration of the current production line, the action can be the adjustment of the parameters, and the state value can be the product quality and production efficiency after the adjustment.

[0088] Construct a composite reward function. The composite reward function consists of three parts: the negative term of the risk control objective function, the resource constraint penalty term, and the time constraint penalty term. The negative term of the risk control objective function is used to measure the risk brought by the intervention action. The resource constraint penalty term is constructed according to the resource constraint condition and is used to penalize actions that violate the resource limit. The time constraint penalty term is constructed according to the time response constraint condition and is used to penalize actions that exceed the specified time. The weights of the resource constraint penalty term and the time constraint penalty term are dynamically adjusted by the Lagrange multiplier, which is updated according to the degree of constraint violation, so as to obtain the constraint weighting coefficient. For example, if an action causes excessive resource consumption, the value of the resource constraint penalty term will increase, and the Lagrange multiplier will also be adjusted accordingly, resulting in a decrease in the reward value of this action. Assuming that the risk control objective is the product defect rate, the resource constraint is the energy consumption, and the time constraint is the production cycle, the composite reward function can be defined as the negative value of the product defect rate plus the weighted sum of the energy consumption penalty and the production cycle penalty.

[0089] Calculate the immediate reward value and the temporal difference error. Calculate the immediate reward value after performing a certain action in the current state based on the composite reward function. Use the immediate reward value and the state value estimate to calculate the temporal difference error, which reflects the gap between the current state value estimate and the actual long-term value. For example, if a certain action leads to an improvement in product quality, the immediate reward value will be higher. The temporal difference error is used to evaluate the accuracy of the current state value estimate.

[0090] Construct the temporal advantage estimate and the policy gradient loss. Use the temporal difference error to construct the temporal advantage estimate, which is used to measure the advantage of performing a certain action in the current state. Use the temporal advantage estimate and the action probability distribution to construct the policy gradient loss, which is used to guide the update of the policy network parameters so that the policy network can generate a better action probability distribution. At the same time, construct the value network loss based on the mean square error between the immediate reward value and the state value estimate, which is used to guide the update of the value network parameters so that the value network can more accurately estimate the state value.

[0091] Construct the meta-policy network and the optimization objective with an exploration term. Input the gradient of the policy gradient loss with respect to the policy parameters into the meta-policy network, and adjust the policy parameters based on the gradient update direction output by the meta-policy network. Calculate the information entropy of the action probability distribution, and construct the weighted sum of the information entropy and the policy gradient loss as the optimization objective with an exploration term. The information entropy is used to encourage the policy network to explore different actions and avoid falling into local optima. Calculate the adaptive exploration coefficient based on the gradient norm of the policy gradient loss with respect to the policy parameters, which is used to dynamically adjust the degree of exploration.

[0092] Use the projected gradient method to iteratively update the policy parameters and the value network parameters. Project the updated policy parameters into the feasible region that satisfies the resource constraint condition and the time response constraint condition by minimizing the parameter offset. Alternately optimize the parameters of the policy network, the dual-stream value network, and the meta-policy network, and finally obtain the quality intervention policy set. For example, through continuous iterative updates, finally obtain a policy set that contains strategies on how to adjust parameters to achieve the best product quality and production efficiency under different production line states.

[0093] Figure 4This is a comparison chart of the convergence performance of the multi-objective constraint optimization method in this embodiment. This chart shows the performance comparison of different optimization methods in solving multi-objective constraint problems. Through constructing an adaptive weight function and a dynamic fluctuation penalty function, this technical solution achieves a better balance in three dimensions: risk control, resource utilization, and time response. In terms of the number of convergence iterations, this technical solution only requires 78 iterations to reach a stable solution, while the traditional linear weighted method (such as the weighted sum multi-objective optimization method proposed by Marler and Arora) requires 143 iterations, and the genetic algorithm based on random search (such as the NSGA-II algorithm proposed by Deb et al.) requires 217 iterations. In terms of the quality of the solution, the Pareto solution set obtained by this technical solution has an average improvement of 18.3% in the risk control effect, a 23.7% increase in resource utilization efficiency, and a 15.9% improvement in time response performance. Especially under strict resource constraint conditions (resource utilization rate > 85%), this technical solution can still maintain a high risk control effect (control rate 87.6%), while the risk control rates of the linear weighted method and the genetic algorithm under the same conditions are 71.4% and 68.9% respectively. By introducing the gradient strategy method and the meta-learning mechanism, this technical solution can dynamically adjust the optimization strategy, and its stability during the iteration process is significantly better than other methods, with the fluctuation amplitude reduced by 43.2%, which is of great significance for the stability of the solution in practical engineering applications.

[0094] In this embodiment, the meta-learning mechanism can dynamically adjust the optimization strategy, accelerate the model convergence speed, and thus improve the optimization efficiency. The introduction of the composite reward function and the constraint conditions enables the generated strategy to better meet the actual needs and have stronger robustness. The optimization objective with an exploration term can achieve a balance between exploring new strategies and exploiting existing strategies, avoiding falling into local optima, and thus finding the global optimal strategy.

[0095] In the second aspect of the present invention, there is provided an intelligent medical quality evaluation and early warning prediction system based on multi-dimensional indicators, and the system includes: A first unit for obtaining real-time medical data, calculating the fluctuation characteristic values of medical quality impact factors, where the fluctuation characteristic values include short-term fluctuation intensity, medium-term fluctuation period, and long-term fluctuation trend, obtaining real-time medical data, calculating the fluctuation characteristic values of medical quality impact factors, constructing a quality sensitivity scoring model based on the fluctuation characteristic values, calculating the data sampling priority, establishing a hierarchical sampling frequency matrix, dynamically sampling the real-time medical data according to the hierarchical sampling frequency matrix to obtain sampling data, extracting the quality feature vector of the sampling data and setting a differential time window, calculating the clinical correlation intensity and feature importance score of the feature group within the differential time window, and performing feature fusion to obtain the medical quality score; A second unit, configured to perform multi-scale wavelet transform on the medical quality score to obtain a quality fluctuation component, calculate local trend functions of the quality fluctuation component at different time scales, construct a cross-correlation matrix between the local trend functions and the quality fluctuation component, extract a quality evolution feature sequence, calculate the dynamic time warping distance to obtain a department quality status vector, construct a quality propagation graph based on the department quality status vector, fuse diagnosis and treatment path information to construct a heterogeneous information network, calculate the propagation influence factor of the heterogeneous information network, and identify quality control nodes; A third unit, configured to calculate the network topology entropy of the quality propagation graph, combine it with the sample entropy of the quality fluctuation component to obtain a quality risk entropy, construct a risk accumulation function based on the quality risk entropy, input the centrality index of the quality control node to obtain a risk propagation prediction sequence, construct a multi-objective constrained optimization equation based on the risk propagation prediction sequence, solve the multi-objective constrained optimization equation to obtain a set of quality intervention strategies, and screen the optimal intervention plan based on the set of quality intervention strategies.

[0096] In a third aspect of the embodiments of the present invention, there is provided an electronic device, 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 described above.

[0097] In a fourth aspect of the embodiments of the present invention, there is provided 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.

[0098] The present invention may be a method, apparatus, system, and / or 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.

[0099] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended 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. An intelligent medical quality evaluation and early warning prediction method based on multi-dimensional indicators, characterized in that Including: Obtain real-time medical data, calculate the fluctuation eigenvalue of the medical quality impact factor, construct a quality sensitivity scoring model based on the fluctuation eigenvalue, calculate the data sampling priority, establish a hierarchical sampling frequency matrix, perform dynamic sampling on the real-time medical data according to the hierarchical sampling frequency matrix to obtain sampled data, extract the quality feature vector of the sampled data and set a differential time window, calculate the clinical correlation strength and feature importance score of the feature group within the differential time window, and perform feature fusion to obtain the medical quality score; Perform multi-scale wavelet transform on the medical quality score to obtain a quality fluctuation component, calculate the local trend function of the quality fluctuation component at different time scales, construct a cross-correlation matrix between the local trend function and the quality fluctuation component, extract the quality evolution feature sequence, calculate the dynamic time warping distance to obtain the department quality state vector, construct a quality propagation graph based on the department quality state vector, fuse the diagnosis and treatment path information to construct a heterogeneous information network, calculate the propagation impact factor of the heterogeneous information network, and identify the quality control nodes; Calculate the network topological entropy of the quality propagation graph, combine it with the sample entropy of the quality fluctuation component to obtain the quality risk entropy, construct a risk accumulation function based on the quality risk entropy, input the centrality index of the quality control node to obtain the risk propagation prediction sequence, construct a multi-objective constraint optimization equation based on the risk propagation prediction sequence, solve the multi-objective constraint optimization equation to obtain a set of quality intervention strategies, and screen the optimal intervention plan based on the set of quality intervention strategies.

2. The method according to claim 1, wherein Obtain real-time medical data, calculate the fluctuation eigenvalue of the medical quality impact factor, construct a quality sensitivity scoring model based on the fluctuation eigenvalue, calculate the data sampling priority, establish a hierarchical sampling frequency matrix, and perform dynamic sampling on the real-time medical data according to the hierarchical sampling frequency matrix to obtain sampled data including: Input the fluctuation eigenvalue into a deep neural network, the deep neural network includes a feature extraction layer, an attention mechanism layer and a fully connected layer, learn the combined weight of the short-term fluctuation intensity, medium-term fluctuation period and long-term fluctuation trend in the fluctuation eigenvalue through the deep neural network to construct a quality sensitivity scoring model, and calculate the quality sensitivity score based on the quality sensitivity scoring model; Perform non-linear mapping on the quality sensitivity score to obtain an initial sampling priority, construct a double exponential smoothing model, input the initial sampling priority into the double exponential smoothing model, calculate the smoothing parameter based on the historical trend and future prediction of the real-time medical data, and perform smoothing processing on the initial sampling priority through the smoothing parameter to obtain the final sampling priority; Construct an adaptive hierarchical threshold according to the final sampling priority, divide the real-time medical data into multiple priority levels based on the adaptive hierarchical threshold to obtain a priority hierarchical result, calculate the data mean, variance and distribution density of each level in the priority hierarchical result, and calculate the basic sampling frequency corresponding to each level based on the data mean, variance and distribution density; Calculate the decay value of data timeliness, the decay value of quality relevance, and the decay value of business importance, and combine them to obtain a time decay factor. Based on the time decay factor, the basic sampling frequency, and the quality sensitivity score, construct a three-dimensional hierarchical sampling frequency matrix. Query the sampling frequency from the three-dimensional hierarchical sampling frequency matrix based on the quality sensitivity score, and generate a sampling timestamp sequence that satisfies the sampling interval constraint according to the sampling frequency; Obtain data samples at the time points corresponding to the sampling timestamp sequence, determine the spline interpolation order based on the characteristics of real-time medical data, and use the spline interpolation order to interpolate and reconstruct the sampling interval of the data samples to obtain reconstructed data. Detect and correct the outliers in the reconstructed data to obtain the final sampling data set.

3. The method according to claim 1, wherein Perform multi-scale wavelet transform on the medical quality score to obtain the quality fluctuation component. Calculate the local trend function of the quality fluctuation component at different time scales, construct a cross-correlation matrix between the local trend function and the quality fluctuation component, extract the quality evolution feature sequence, and calculate the dynamic time warping distance to obtain the department quality status vector, including: Construct an asymmetric wavelet basis function with adaptive time-frequency resolution, and perform continuous wavelet transform on the medical quality score based on the asymmetric wavelet basis function to obtain the quality fluctuation component in different frequency domains; Decompose the quality fluctuation component morphologically, extract the basic morphological elements of the quality fluctuation component, and use a sliding time window to smooth and fit the basic morphological elements to obtain the local trend function at different time scales; Perform fractional-order differential operations on the local trend function and the quality fluctuation component respectively, construct a fractional-order differential operator, and obtain the phase evolution sequence; calculate the instantaneous phase and instantaneous amplitude of the phase evolution sequence, calculate the generalized complementary coefficient at multiple time scales based on the instantaneous phase and instantaneous amplitude, and construct a cross-correlation matrix reflecting multi-scale characteristics; Perform dynamic system spectral decomposition on the cross-correlation matrix to obtain the eigenvalue sequence and eigenvector sequence, and construct a non-linear feature projection transformer based on the eigenvalue sequence and eigenvector sequence; input the cross-correlation matrix into the feature projection transformer for non-linear dimensionality reduction to obtain the quality evolution feature sequence; Calculate the geodesic distance between adjacent eigenvectors in the quality evolution feature sequence, organize the geodesic distance into a local distance matrix; use a constrained dynamic programming algorithm to solve the optimal warping path for the local distance matrix; calculate the quality trend prediction value according to the optimal warping path and the quality evolution feature sequence; Extract the local deformation gradient of the optimal warping path, input the local deformation gradient into a recurrent neural network, and calculate the quality mutation probability through a non-linear activation function and a temporal memory unit; Perform dynamic time warping on the quality evolution feature sequence and a standard sequence to obtain a warping distance sequence, and use a time weight function to perform weighted accumulation on the warping distance sequence to obtain the quality cumulative deviation; Normalize and perform dimensional transformation on the quality trend prediction value, the quality mutation probability, and the quality cumulative deviation in the feature space to construct a standardized department quality status vector.

4. The method according to claim 1, wherein Construct a quality propagation graph based on the department quality status vector, fuse the diagnosis and treatment path information to construct a heterogeneous information network, calculate the propagation influence factor of the heterogeneous information network, and identify quality control nodes, including: Construct a kernel function mapping matrix, perform a non-linear mapping on the department quality status vector to obtain a mapped feature vector; calculate the cosine similarity of the mapped feature vectors between department nodes to construct a similarity matrix; construct an initial quality propagation graph based on the similarity matrix, where a quality status edge is established between nodes when the similarity value corresponding to the department node exceeds a preset dynamic threshold, and the weight of the quality status edge is the corresponding similarity value; Construct a time sliding window based on the patient transfer records, perform exponential time series weighting on the patient transfer frequency within the window to obtain a time series weighted transfer frequency, and perform normalization processing on the time series weighted transfer frequency to obtain the weight value of the patient transfer edge; extract the consultation frequency and examination application frequency based on the doctor's order execution records, and combine them with weight coefficients according to the business type to obtain the weight value of the medical collaboration edge; Integrate the patient transfer edge and the medical collaboration edge and their corresponding weight values into the initial quality propagation graph to construct a heterogeneous information network; for each department node in the heterogeneous information network, construct a multi-layer attention network based on the characteristics of its first-order neighbor nodes on the quality status edge, the patient transfer edge, and the medical collaboration edge; use the multi-layer attention network to calculate the weight contribution coefficients of each edge respectively, and perform weighted summation of the weight contribution coefficients and the weight values of the corresponding edges to obtain the propagation influence factor of the department node; Construct a random walk iteration rule based on the topological structure of the heterogeneous information network, perform iterative update on the propagation influence factor according to the random walk iteration rule, calculate the Euclidean distance of the propagation influence factor sequences obtained from two adjacent iterations, obtain the final propagation influence factor when the Euclidean distance is less than the convergence threshold, sort the department nodes according to the final propagation influence factor, and combine the in-degree distribution and betweenness centrality of each department node on the three types of edges, and use the analytic hierarchy process to identify quality control nodes.

5. The method according to claim 1, wherein Calculate the network topological entropy of the quality propagation graph, combine it with the sample entropy of the quality fluctuation component to obtain the quality risk entropy, construct a risk accumulation function based on the quality risk entropy, and input the centrality index of the quality control node to obtain a risk propagation prediction sequence, including: Calculate the connection probability between the nodes and their adjacent nodes in the quality propagation graph, perform logarithmic operation on the connection probability and multiply it by the original connection probability to obtain the node local entropy; calculate the degree distribution and betweenness centrality of the node based on the connection probability, and construct a node importance score; use the node importance score to calculate the weight coefficient, and perform weighted summation of the local entropy of each node and the corresponding weight coefficient to obtain the network topological entropy of the quality propagation graph; Perform segmentation and normalization processing on the quality fluctuation component to obtain a standardized sequence, construct a template sequence based on the standardized sequence according to the preset embedding dimension; count the number of matches of the template sequence under a given similarity tolerance; calculate the natural logarithm ratio of the number of template matches in adjacent dimensions to obtain the sample entropy of the quality fluctuation component; Construct a non - linear combination model that includes the network topological entropy and sample entropy. The non - linear combination model includes linear terms, square terms, and cross terms. Construct an optimization objective function based on historical risk data and node importance scores. Solve the optimization objective function to obtain combination coefficients, and substitute the combination coefficients, network topological entropy, and sample entropy into the non - linear combination model to obtain the quality risk entropy. Construct an exponential cumulative function using the quality risk entropy, map the quality risk entropy to a cumulative rate coefficient, construct a time weight function based on the time characteristics of the standardized sequence, and combine the time weight function and the cumulative function to obtain a risk cumulative function. Extract the centrality index of quality control nodes, use the principal component analysis method to reduce the dimension and fuse the centrality index, smooth the fused centrality index, and input the smoothed centrality index into the risk cumulative function to obtain a prediction sequence reflecting the risk propagation degree of quality control nodes.

6. The method according to claim 1, characterized in that, Construct a multi - objective constrained optimization equation based on the risk propagation prediction sequence, solve the multi - objective constrained optimization equation to obtain a set of quality intervention strategies, and screen the optimal intervention plan based on the set of quality intervention strategies, including: Use a bidirectional time - series network to process the risk propagation prediction sequence to obtain a hidden - layer state sequence, calculate the corresponding risk weight sequence based on the hidden - layer state sequence through the self - attention mechanism, and perform weighted fusion on the risk weight sequence and the hidden - layer state sequence to obtain risk features. Construct an adaptive weight function and a dynamic fluctuation penalty function based on the risk features, and combine the product of the adaptive weight function and the risk value and the product of the dynamic fluctuation penalty function and the fluctuation amplitude to construct a risk control objective function. Construct a resource interaction graph according to the resource type and resource collaboration relationship, use the graph attention network to perform node representation learning on the resource interaction graph, construct resource constraint conditions based on the obtained node representations through a gated graph network, construct a state space based on the risk features at the same time, model the intervention process as a state - action sequence in the state space, calculate the intervention time series corresponding to the state - action sequence using a dual - value network, and construct a time response constraint condition by the weighted sum of the intervention time series and the time weight. Combine the risk control objective function, resource constraint conditions, and time response constraint conditions to construct a multi - objective constrained optimization equation. Construct a policy network and a value network for solving the multi - objective constrained optimization equation, optimize the parameters of the policy network and the value network using the gradient policy method, and introduce a meta - learning mechanism to dynamically adjust the optimization strategy during the parameter optimization process to obtain a set of quality intervention strategies. Calculate the risk control effect of each strategy in the set of quality intervention strategies based on the risk control objective function, calculate the resource utilization efficiency based on the risk control objective function and resource constraint conditions, and calculate the time response performance based on the time response constraint conditions. Construct a metric network, input the risk control effect, resource utilization efficiency, and time response performance of each policy into the metric network, train the metric network through a contrastive learning method to obtain a policy scoring model, use the policy scoring model to score each policy in the quality intervention policy set, and select the policy with the highest score as the optimal intervention plan.

7. The method according to claim 6, wherein Construct a policy network and a value network for solving the multi-objective constrained optimization equation, use the gradient policy method to optimize the parameters of the policy network and the value network, and introduce a meta-learning mechanism to dynamically adjust the optimization policy during the parameter optimization process, and obtain a quality intervention policy set including: Construct a policy network and a value network for solving the multi-objective constrained optimization equation. Among them, the policy network extracts features from the input state through a multi-layer attention structure to obtain a state representation, and generates an action probability distribution based on the state representation; the value network includes a first feature extraction sub-network and a second feature extraction sub-network, generates a first state value and a second state value respectively based on the state representation obtained in the policy network, and fuses the first state value and the second state value through an adaptive weight coefficient to obtain a state value estimate; Construct a composite reward function, which includes a negative term of the risk control objective function in the multi-objective constrained optimization equation, a resource constraint penalty term, and a time constraint penalty term. Among them, the resource constraint penalty term is constructed based on the resource constraint condition, and the time constraint penalty term is constructed based on the time response constraint condition. Dynamically update the Lagrange multiplier based on the constraint violation degrees of the resource constraint penalty term and the time constraint penalty term to obtain a constraint weighting coefficient; Calculate the immediate reward value based on the composite reward function, and calculate the temporal difference error based on the immediate reward value and the state value estimate. Use the temporal difference error to construct a temporal advantage estimate value, use the temporal advantage estimate value and the action probability distribution to construct a policy gradient loss, and at the same time construct a value network loss based on the mean square error between the immediate reward value and the state value estimate; Construct a meta-policy network, input the gradient of the policy gradient loss with respect to the policy parameters into the meta-policy network, adjust the policy parameters based on the gradient update direction output by the meta-policy network, calculate the information entropy of the action probability distribution, and construct the weighted sum of the information entropy and the policy gradient loss as an optimization objective with an exploration term. Calculate an adaptive exploration coefficient based on the gradient norm of the policy gradient loss with respect to the policy parameters; Based on the optimization objective with an exploration term and the value network loss, use the projected gradient method to iteratively update the policy parameters and the value network parameters respectively. Project the updated policy parameters into the feasible region that satisfies the resource constraint condition and the time response constraint condition by minimizing the parameter offset, and alternately optimize the parameters of the policy network, the dual-stream value network, and the meta-policy network to obtain a quality intervention policy set.

8. An intelligent medical quality assessment and early warning prediction system based on multi-dimensional indicators, which is used to implement the method described in any one of the foregoing claims 1-7, and is characterized in that, Including: A first unit, configured to obtain real-time medical data, calculate the fluctuation eigenvalue of the medical quality impact factor, construct a quality sensitivity scoring model based on the fluctuation eigenvalue, calculate the data sampling priority, establish a hierarchical sampling frequency matrix, dynamically sample the real-time medical data according to the hierarchical sampling frequency matrix to obtain sampled data, extract the quality feature vector of the sampled data and set a differential time window, calculate the clinical correlation strength and the feature importance score of the feature group within the differential time window, and perform feature fusion to obtain a medical quality score; A second unit, configured to perform multi-scale wavelet transform on the medical quality score to obtain a quality fluctuation component, calculate the local trend function of the quality fluctuation component at different time scales, construct a cross-correlation matrix of the local trend function and the quality fluctuation component, extract the quality evolution feature sequence, calculate the dynamic time warping distance to obtain a department quality status vector, construct a quality propagation graph based on the department quality status vector, fuse the diagnosis and treatment path information to construct a heterogeneous information network, calculate the propagation influence factor of the heterogeneous information network, and identify quality control nodes; A third unit, configured to calculate the network topological entropy of the quality propagation graph, combine it with the sample entropy of the quality fluctuation component to obtain a quality risk entropy, construct a risk accumulation function based on the quality risk entropy, input the centrality index of the quality control node to obtain a risk propagation prediction sequence, construct a multi-objective constrained optimization equation based on the risk propagation prediction sequence, solve the multi-objective constrained optimization equation to obtain a set of quality intervention strategies, and screen the optimal intervention plan based on the set of quality intervention strategies.

9. An electronic device, characterized in that, Comprising: 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