Hospital management scenario simulation and evaluation method and system based on generative adversarial network

Abstract

The application relates to a hospital management scene simulation and evaluation method and system based on a generative adversarial network, which comprises the following steps: collecting hospital full-dimension management data, and generating a management data matrix through time-space alignment and standardization. Unsupervised learning is performed through a restricted Boltzmann machine to output an index probability distribution model and a connection weight matrix. A hospital management graph is constructed based on this, and an improved PageRank algorithm is used to calculate a node management importance score. In combination with a management target, graph embedding and adaptive clustering are performed to obtain a node embedding vector and a classification label set. The connection weight matrix, the importance score and the classification label are comprehensively used to calculate a comprehensive management evaluation score. Finally, the probability distribution model, the evaluation score, the classification label and an external intervention strategy are input into a generative adversarial network, and after adversarial training, a dynamic management scene prediction report and an optimization strategy scheme are output, so that the scientificity and the fine-grained level of hospital management decision-making are improved.

Description

Technical Field

[0001] This application relates to the field of medical information technology, and in particular to a method and system for simulating and evaluating hospital management scenarios based on generative adversarial networks. Background Technology

[0002] With the deepening of medical informatization and the increasing demand for refined hospital management, modern hospitals are able to collect and accumulate massive amounts of multi-dimensional management data in their daily operations. This data covers various aspects such as departmental operations, staffing, financial revenue and expenditure, and medical quality evaluation. Hospital management decision-makers generally hope to use the analysis of this historical data to more scientifically assess the overall operational status of the hospital and, based on this, predict the potential impact of different management strategies, thereby optimizing resource allocation and improving management efficiency and service quality. Currently, common hospital management evaluation methods mostly rely on statistical analysis based on fixed indicators, balanced scorecards, or expert experience models. These methods can reflect certain static characteristics of hospital operations to a certain extent and provide a reference for management decisions.

[0003] However, due to the high complexity and dynamism of hospital management systems, numerous non-linear and implicit interrelationships exist between various management dimensions. Traditional methods struggle to comprehensively and accurately depict these complex relationships, leading to evaluation results that are often limited to superficial listings of indicators or simple weighted summaries, failing to reveal the deeper structural problems behind the data. Secondly, existing evaluation models lack the ability to simulate the chain reactions and long-term dynamic effects that management decisions may trigger. This makes it difficult for managers to effectively predict the overall effects of strategies before implementation, particularly in quantifying the differentiated impacts of a strategy adjustment on different departments and business dimensions within the hospital, thus increasing the uncertainty and potential risks of management decisions. Summary of the Invention

[0004] To address the aforementioned technical issues, this application provides a method and system for simulating and evaluating hospital management scenarios based on generative adversarial networks.

[0005] Firstly, this application provides a method for simulating and evaluating hospital management scenarios based on generative adversarial networks, employing the following technical solution:

[0006] Collect comprehensive hospital management data, including departmental operation indicators, staffing parameters, financial revenue and expenditure data, and medical quality evaluation data;

[0007] The management data is subjected to spatiotemporal alignment and standardization to generate a management data matrix;

[0008] The management data matrix is ​​subjected to unsupervised learning using a restricted Boltzmann machine, and the output index probability distribution model and the inter-index connection weight matrix are generated.

[0009] A hospital management graph structure is constructed based on the inter-indicator connection weight matrix, where nodes represent management entities, edge weights are assigned by the inter-indicator connection weight matrix, and the management importance score of each node is calculated using an improved PageRank algorithm.

[0010] Based on the preset management target configuration parameters and the management importance score, a node embedding vector is generated by a graph embedding algorithm, and an adaptive clustering analysis is used to classify and label the nodes, outputting a set of node classification labels;

[0011] Combining the inter-indicator connection weight matrix, management importance score, and node classification label set, the comprehensive management evaluation score is calculated by grouping by category;

[0012] The probability distribution model of the indicators, the comprehensive management evaluation score, the set of node classification labels and the parameters of external intervention strategies are input into a pre-configured generative adversarial network. After adversarial training, a dynamic management scenario prediction report and an optimization strategy scheme are output.

[0013] By adopting the above technical solution, scattered and static management data is gradually transformed into a probabilistic model of deeply correlated indicators, node scores reflecting structural importance, and a set of labels reflecting functional clustering. Finally, through generative adversarial networks, a digital twin system capable of predicting and optimizing complex management interventions is constructed. This technical solution not only provides a multi-dimensional and quantifiable comprehensive assessment score of the hospital's historical and current status, but also serves as a powerful decision support tool, simulating the potential chain reactions and comprehensive effects of various management strategies in the future. This enhances the predictability, scientific rigor, and precision of hospital management, effectively mitigates decision-making risks, and guides resources to achieve optimal allocation.

[0014] Secondly, this application provides a hospital management scenario simulation and evaluation system based on generative adversarial networks, employing the following technical solution:

[0015] The multi-source data acquisition module is used to collect hospital management data from all dimensions, including departmental operation indicators, staffing parameters, financial revenue and expenditure data, and medical quality evaluation data.

[0016] The data processing module is used to perform spatiotemporal alignment and standardization on the management data to generate a management data matrix;

[0017] The unsupervised learning module is used to perform unsupervised learning on the management data matrix through a restricted Boltzmann machine, and outputs an index probability distribution model and an inter-index connection weight matrix.

[0018] The node importance quantification module is used to construct a hospital management graph structure based on the connection weight matrix between the indicators, and to calculate the management importance score of each node using an improved PageRank algorithm.

[0019] The dynamic clustering module is used to generate node embedding vectors through graph embedding algorithm based on preset management target configuration parameters and management importance scores, and to classify and label nodes using adaptive clustering analysis, outputting a set of node classification labels.

[0020] The comprehensive management assessment module is used to combine the weight matrix between the indicators, the management importance score, and the set of node classification labels to calculate the comprehensive management assessment score by category.

[0021] The adversarial training output module is used to input the index probability distribution model, comprehensive management evaluation score, node classification label set and external intervention strategy parameters into a pre-configured generative adversarial network, and output a dynamic management scenario prediction report and optimization strategy scheme after adversarial training.

[0022] Thirdly, this application provides a computer-readable storage medium, which adopts the following technical solution:

[0023] A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as in any of the methods in the first aspect.

[0024] In summary, this application includes at least one of the following beneficial technical effects: By constructing a digital twin system for hospital management, it transforms dispersed and static management data into dynamic and predictable decision support tools, achieving a fundamental leap from retrospective assessments based on historical data to forward-looking simulations of future possibilities. In practical applications, this application not only provides a comprehensive quantitative assessment of the overall and local operational management level of the hospital, integrating structural importance, functional clustering, and mutual influence, but more importantly, through the adversarial training mechanism of generative adversarial networks, it simulates the chain reactions and comprehensive effects that different management intervention strategies may trigger in the complex hospital operation network. This enables managers to scientifically foresee potential risks and benefits before implementing decisions, thereby greatly improving the predictability, scientific rigor, and precision of management decisions, effectively avoiding decision-making blind spots, and guiding limited medical and management resources to achieve optimal system allocation. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the first process of a hospital management scenario simulation and evaluation method based on generative adversarial networks, which is one embodiment of this application.

[0026] Figure 2This is a schematic diagram of the second process of a hospital management scenario simulation and evaluation method based on generative adversarial networks, which is one embodiment of this application.

[0027] Figure 3 This is a schematic diagram of the third process of a hospital management scenario simulation and evaluation method based on generative adversarial networks, which is one embodiment of this application.

[0028] Figure 4 This is a schematic diagram of the fourth process of a hospital management scenario simulation and evaluation method based on generative adversarial networks, according to one embodiment of this application.

[0029] Figure 5 This is a schematic diagram of the fifth process of a hospital management scenario simulation and evaluation method based on generative adversarial networks, which is one embodiment of this application.

[0030] Figure 6 This is a schematic diagram of the sixth process of a hospital management scenario simulation and evaluation method based on generative adversarial networks, which is one embodiment of this application.

[0031] Figure 7 This is a schematic diagram of the seventh process of a hospital management scenario simulation and evaluation method based on generative adversarial networks, which is one embodiment of this application.

[0032] Figure 8 This is a schematic diagram of the eighth process of a hospital management scenario simulation and evaluation method based on generative adversarial networks, one embodiment of this application. Detailed Implementation

[0033] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figures 1-8 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0034] This application discloses a method for simulating and evaluating hospital management scenarios based on generative adversarial networks.

[0035] Reference Figure 1 A method for simulating and evaluating hospital management scenarios based on generative adversarial networks, the method comprising:

[0036] Step S101: Collect hospital management data across all dimensions, including departmental operation indicators, staffing parameters, financial revenue and expenditure data, and medical quality evaluation data;

[0037] Among these, departmental operational indicators (such as outpatient volume and bed turnover rate) reflect the vitality and efficiency of business units; personnel allocation parameters (such as doctor-nurse ratio and professional title structure) are a quantitative representation of human resource investment; financial revenue and expenditure data (such as revenue composition and cost details) directly characterize the economic operation status; and medical quality evaluation data (such as cure rate and nosocomial infection rate) are the final output of the hospital's core value. By collecting these multi-dimensional management data, we can comprehensively depict the hospital's management landscape as much as possible, avoiding biases in model evaluation due to missing or incomplete data, and laying the foundation for building a high-fidelity digital twin.

[0038] Step S102: Perform spatiotemporal alignment and standardization on the management data to generate a management data matrix;

[0039] The management data matrix is ​​a well-organized, machine-readable two-dimensional data table, where each row represents an observation sample (such as monthly data of a department) and each column represents a management indicator.

[0040] Specifically, spatiotemporal alignment refers to unifying data from different sources and with different time granularities (such as daily, monthly, and yearly) into the same analytical time window and spatial dimension (such as uniformly summarizing by department), solving the problems of inconsistent data timestamps and misaligned statistical units. Standardization processing (such as normalization or Z-score standardization) transforms various indicators with vastly different dimensions and numerical ranges (such as monetary values ​​and percentage scores) into a unified order of magnitude, preventing some large-value indicators from "drowning" smaller but crucial indicators (such as medical quality indicators) during model training.

[0041] Step S103: Unsupervised learning of the management data matrix is ​​performed using a restricted Boltzmann machine to output the probability distribution model of the indicators and the connection weight matrix between the indicators.

[0042] In this embodiment, a generative stochastic neural network model called a Restricted Boltzmann Machine (RBM) is used to automatically discover the underlying structure and patterns of data from a regular managed data matrix. The RBM learns the joint probability distribution of the input data in an unsupervised manner (i.e., without requiring manual labeling) through the probabilistic interactions between neurons in its visible and hidden layers.

[0043] The output of the Restricted Boltzmann Machine (RBM) includes two key products: an indicator probability distribution model, which reveals the potential statistical dependencies between various management indicators (such as medical revenue and costs), i.e., how a change in one indicator affects the probability of other indicators. This provides data distribution constraints for subsequent generative adversarial networks to simulate scenarios; and an inter-indicator connection weight matrix, which quantifies the correlation strength between any two management indicators. For example, it may learn that there is a strong positive correlation between "nurse ratio" and "patient satisfaction" and assign a high weight to this relationship.

[0044] Step S104: Construct a hospital management graph structure based on the inter-indicator connection weight matrix, where nodes represent management entities, edge weights are assigned by the inter-indicator connection weight matrix, and the management importance score of each node is calculated using the improved PageRank algorithm.

[0045] In this embodiment of the application, a hospital management graph can be constructed based on the connection weight matrix obtained in the previous step. Nodes represent specific management entities (such as clinical departments, administrative departments, key equipment, or even key personnel), while edges represent the management, business, or resource flow relationships between entities. The weight of the edges is directly assigned by the connection weight matrix, reflecting the strength of the relationship.

[0046] However, association strength alone is insufficient to assess the importance of an individual node. Therefore, this embodiment introduces an improved PageRank algorithm. The classic PageRank algorithm measures webpage importance, but it assumes that the network structure satisfies equal voting and random jumps, which is inconsistent with the hierarchical and fixed processes of hospital management graphs. The improvement in this embodiment lies in adjusting the convergence factor according to the characteristics of hospital management (such as hierarchical depth and fixed processes), so that influence decays reasonably according to the depth of organizational hierarchy during transmission, thereby calculating a more realistic management importance score. This score identifies which departments or units are hubs in hospital operations (e.g., the emergency department or operating room may receive a high score due to their extensive connectivity), providing a crucial basis for optimal resource allocation.

[0047] Step S105: Based on the preset management target configuration parameters and management importance scores, generate node embedding vectors through graph embedding algorithm, and use adaptive clustering analysis to classify and label the nodes, outputting a set of node classification labels;

[0048] The purpose of graph embedding algorithms is to map the high-dimensional and complex connections of each node (management entity) in the graph into a low-dimensional, continuous vector space, generating a node embedding vector. This vector is like the node's DNA, condensing the node's own attributes, its structural position in the network, and its interaction patterns with other nodes.

[0049] Then, based on the preset management objectives (such as identifying abnormal departments or grouping by function), adaptive clustering analysis is employed. For example, if the objective is anomaly detection, a density-based clustering algorithm (such as DBSCAN) is used, which can effectively identify outliers that deviate from the majority of node groups (such as departments with extremely low or high operational efficiency); if the objective is hierarchical division, a hierarchical clustering algorithm is used to discover natural organizational clusters within the hospital (such as clustering all surgical departments into one category). The final output set of node classification labels provides managers with a new data-driven classification perspective that transcends traditional administrative divisions.

[0050] Step S106: Combine the inter-indicator connection weight matrix, management importance score and node classification label set to calculate the comprehensive management evaluation score by category group;

[0051] Among them, a multi-dimensional, weighted aggregation evaluation system is constructed. This evaluation system does not simply average the scores of all nodes, but follows the principle of "classification first, evaluation second, and synthesis third".

[0052] In this embodiment, firstly, based on the node classification label set obtained in the previous step, the management entities are grouped (e.g., into clinical frontline departments, medical technology support departments, and administrative and logistical departments). Then, within each group, a weighted average score is calculated, using the management importance score of each node as a weight, combined with the influence reflected by its connection weights with other nodes. Finally, considering that different groups are not isolated (e.g., the efficiency of administrative departments can affect clinical departments), the scores of all groups are globally summed using the correlation strength between groups as an adjustment coefficient to obtain the final comprehensive management evaluation score. This score is a unified and quantifiable metric that can objectively and comprehensively reflect the overall or specific operational management level of the hospital.

[0053] Step S107: Input the indicator probability distribution model, comprehensive management evaluation score, node classification label set and external intervention strategy parameters into the pre-configured generative adversarial network, and output a dynamic management scenario prediction report and optimization strategy scheme after adversarial training.

[0054] In this embodiment of the application, the Generative Adversarial Network (GAN) consists of two neural networks: a generator and a discriminator.

[0055] The generator receives external intervention strategy parameters (such as "increase pediatric investment by 10%" or "reduce administrative expenses by 5%) as input, along with existing indicator probability distribution models and management importance scores, to simulate and generate a series of possible management decision sequences and their resulting changes in management status. The discriminator, based on patterns learned from historical data and comprehensive management evaluation scores, judges whether the simulated scenarios generated by the generator are realistic and effective. Through repeated adversarial training, the generator continuously learns how to generate more reasonable and optimized strategies, while the discriminator continuously improves its discrimination capabilities.

[0056] Ultimately, the system can output dynamic management scenario prediction reports, intuitively showing the future results that different decisions may bring (such as the risk of financial improvement but quality decline), and provide optimized strategy solutions that have been verified through adversarial training, thereby moving from static and retrospective evaluation to dynamic and forward-looking decision support.

[0057] In the above implementation, scattered and static management data is gradually transformed into a probabilistic model of deeply correlated indicators, node scores reflecting structural importance, and a set of labels reflecting functional clustering. Finally, through a generative adversarial network, a digital twin system capable of predicting and optimizing complex management interventions is constructed. This technical solution not only provides a multi-dimensional and quantifiable comprehensive assessment score of the hospital's historical and current status, but also serves as a powerful decision support tool, simulating the potential chain reactions and comprehensive effects of various management strategies in the future. This enhances the predictability, scientific rigor, and precision of hospital management, effectively mitigates decision-making risks, and guides resources to achieve optimal allocation.

[0058] In practical applications, a typical example of this application is its application in the scenario of "refined cost control and resource allocation optimization" in large general hospitals. Hospital management often faces the challenges of continuously rising drug and consumable costs and uneven resource utilization efficiency across departments. Therefore, they hope to find a systematic cost optimization path and dynamic resource allocation solution without affecting medical quality and patient safety.

[0059] First, the system collects operational indicators (such as daily drug costs, consumable ratio, and equipment uptime) from all departments across the hospital, staffing (medical and nursing hours and shift schedules), financial data (departmental cost structure and revenue details), and quality data (complication rate and patient prognosis scores). After cleaning, alignment, and standardization, this data forms a management data matrix covering all departments and monthly cycles. Subsequently, through unsupervised learning of this matrix using a restricted Boltzmann machine, the system automatically discovers deep-seated statistical patterns. For example, it outputs a distribution model showing a probabilistic correlation between "high-value consumable usage rate" and "risk of specific surgical complications," and quantifies the connection weight between "nurse nighttime staffing ratio" and "patient fall adverse events," thereby revealing the implicit interrelationships between cost, efficiency, and quality.

[0060] Based on the learned inter-indicator connection weight matrix, the system constructs a hospital management graph, where nodes represent clinical departments, pharmacies, consumables warehouses, and support departments. Using an improved PageRank algorithm (which adjusts the influence decay factor based on hospital administrative hierarchy and business reporting relationships), nodes with high resource consumption and strong business interdependencies, such as the "central operating room," "intensive care unit (ICU)," and "cardiovascular interventional catheterization lab," are identified as having the highest management importance scores, thus recognizing them as key hubs for cost control and resource allocation.

[0061] Next, based on the management goal of identifying resource usage patterns, the complex relationships (such as resource sharing and patient transfer) of each node in the high-dimensional graph are condensed into low-dimensional vectors using graph embedding algorithms. Then, cluster analysis is used to automatically divide all departments in the hospital into groups with similar resource behavior characteristics, such as "high-cost-high-efficiency surgical group", "high-cost-risk-sensitive internal medicine group", "platform-supported medical technology group" and "cost-controllable administrative group", forming a set of node classification labels.

[0062] Based on the aforementioned correlation weights, node importance scores, and classification labels, the system calculates a "comprehensive cost-benefit assessment score" for each department by group. For example, for the "high-cost-high-efficiency surgical group," the score not only reflects its own resource consumption but also incorporates its resource intensity in important hub departments such as the surgical platform and anesthesia recovery room, as well as its medical quality output, thus obtaining a quantitative evaluation that balances cost, efficiency, and contribution.

[0063] Finally, the administrator inputs external intervention strategy parameters (such as "reducing the hospital's average drug ratio by 3 percentage points" and "optimizing the transfer standards for each department while ensuring that the ICU bed occupancy rate is not less than 85%), along with the generated indicator probability distribution model, comprehensive evaluation scores, and classification labels, into the generative adversarial network for adversarial training. The generator simulates and proposes a series of specific and differentiated resource allocation adjustment schemes (such as setting tiered consumable control targets for departments in group A and optimizing the scheduling model for group B), while the discriminator judges whether these simulated schemes will cause unacceptable quality risks or system disorder based on historical patterns and comprehensive evaluation standards.

[0064] After repeated iterations, the system finally outputs a dynamic prediction report, showing the simulated trends of key indicators for each department over the next six months under different cost control strategies. It also recommends a validated and executable optimization strategy, such as "prioritizing the implementation of clinical pathway information management in the 'high-cost-risk-sensitive internal medicine group' to reduce variable costs, while intelligently scheduling equipment reservations for the 'platform-supported medical technology group' to improve throughput." This enables scientific decision-making for precise cost control and optimized resource allocation at the system level.

[0065] Reference Figure 2 As one implementation of step S103, the step of performing unsupervised learning on the management data matrix using a restricted Boltzmann machine to output the index probability distribution model and the inter-index connection weight matrix includes:

[0066] Step S201: Input the management data matrix into the visible layer of the restricted Boltzmann machine, and generate visible layer reconstruction data through bidirectional probability mapping between the visible layer and the hidden layer of the restricted Boltzmann machine;

[0067] In this embodiment, the Restricted Boltzmann Machine (RBM) consists of a visible layer and hidden layers. The number of neurons in the visible layer is consistent with the dimension of the input management data matrix, and it is used to receive specific values. The hidden layer is a collection of neurons, typically fewer in number, whose function is to learn and represent the internal features of the data. When the normalized management data matrix is ​​input into the visible layer, the model does not perform a deterministic transformation, but instead calculates the probability of each hidden layer neuron being activated based on the current inter-layer connection weights and bias parameters.

[0068] Specifically, given the visible layer state (input data), a sigmoid-type probability function is used to calculate the probability that each hidden unit will take the value of 1, and random sampling is performed based on this probability to obtain the activation state of the hidden layer. This process can be understood as extracting some abstract and comprehensive "management themes" or "operating models" (such as "high workload and low efficiency model" or "high efficiency and high quality model") from specific and observable management indicators (such as "high outpatient volume" or "cost exceeding the standard").

[0069] Subsequently, based on the sampled hidden layer states, the visible layer data is reconstructed in the same way, i.e., the probability of each visible unit's value is calculated. By repeatedly executing this process, the model learns the most representative high-order feature combinations in the original data distribution, making the data generated or reconstructed from the hidden layer states as close as possible to real hospital management data in terms of probability distribution. For example, through this mechanism, the model may learn the probability relationship between the two visible unit states, "high bed turnover rate" and "high drug ratio," occurring simultaneously under a specific hidden mode (which can be understood as an "operational pressure mode").

[0070] Step S202: Based on the contrast divergence between the reconstructed data and the management data matrix of the visible layer, iteratively optimize the inter-layer connection weight parameters of the restricted Boltzmann machine to generate a set of node feature vectors and an inter-layer connection weight matrix.

[0071] Specifically, this step uses the contrastive divergence algorithm. The core of this algorithm is to update the parameters by comparing the gradient of the energy function defined by the model in two states. These two states are: (1) Data distribution state: the real data matrix is ​​clamped in the visible layer and a forward propagation is performed to obtain the hidden layer sample. This state represents the data reality; (2) Model distribution state: the hidden layer sample obtained in the previous step is used as the starting point, a reverse reconstruction is performed to obtain the visible layer sample, and then a forward propagation is performed. This state represents the current imagination of the model.

[0072] In this embodiment, an approximate value for the gradient of the connection weight update can be obtained by calculating the difference between the expected values ​​of the products of the visible layer and hidden layer neuron states in these two states. Through iterative optimization, the model parameters are continuously adjusted to reduce the KL divergence between the data distribution and the model's own generated distribution. After training, the inter-layer connection weight matrix is ​​fixed. Meanwhile, for each sample in the input data matrix (such as data from a department in a certain month), its corresponding hidden layer activation probability vector constitutes a set of node feature vectors. Each vector is a deep feature encoding of a data sample, while the inter-layer connection weight matrix quantifies the contribution of each visible layer indicator (input feature) to each hidden layer feature (abstract concept).

[0073] Step S203: Based on the set of node feature vectors, fit the probability distribution model of each management indicator in the management data matrix through Gaussian mixture modeling to obtain the indicator probability distribution model;

[0074] The distribution of hospital management indicators (such as costs and satisfaction scores) is often not monomorphic but multimodal. For example, the average cost per visit in different specialties may form multiple cluster centers, which cannot be accurately described by a simple Gaussian distribution. Therefore, this application uses a Gaussian mixture model for fitting.

[0075] In this embodiment, the Gaussian mixture model assumes that the data is composed of multiple Gaussian distributed components mixed with certain weights, and each component can represent a potential state or category in hospital operations (such as "outpatient-dominated department" or "inpatient surgical department"). Using the expectation-maximization algorithm, the mean, covariance, and mixing weight of each Gaussian component can be iteratively estimated. For each management indicator, its marginal probability distribution function can be accurately characterized by a linear combination of these Gaussian components.

[0076] Ultimately, the obtained probability distribution model of the indicators not only describes the statistical characteristics of individual indicators (such as mean and variance), but more importantly, it characterizes the probability of different indicator values ​​and the complex dependencies between indicators (reflected by the covariance matrix of the mixed components). This provides the generator in the subsequent generative adversarial network with the necessary prior constraints on the data distribution, ensuring that the simulated management scenario is statistically realistic and credible.

[0077] Step S204: Extract cross-index correlation weights from the inter-layer connection weight matrix;

[0078] The inter-layer connection weight matrix obtained during training has rows corresponding to the management indicators of the original visible layer and columns corresponding to the learned hidden features.

[0079] In this embodiment, the original weight matrix reflects the association between each specific indicator and each abstract hidden feature. Therefore, in order to obtain a matrix that directly represents the strength of the potential association between indicators, information extraction is required. For example, calculating the correlation or covariance of indicators in the hidden feature space can be achieved through some operation on the weight matrix (such as calculating its Gram matrix or through the dot product of weight vectors).

[0080] Ultimately, the extracted cross-indicator correlation weights preliminarily quantify the non-linear correlation strength between any two management indicators mediated by implicit features. For example, it may reveal a negative correlation between "medical staff ratio" and "hospital infection rate," even though this relationship may be masked by other factors or be non-linear in the original data.

[0081] Step S205: Combining the pre-constructed medical management knowledge graph, perform redundant connection pruning and normalization on the cross-index association weights, and output the inter-index connection weight matrix.

[0082] Among them, the pre-constructed medical management knowledge graph contains deterministic or strongly correlated relationships from medical procedures, management theories, and clinical pathways (such as "surgical volume" is necessarily strongly correlated with "anesthesiology workload", and "outpatient volume" and "drug revenue" have expected management correlation).

[0083] In this embodiment, the cross-index association weights extracted in the previous step are compared with this knowledge graph to perform redundant connection pruning, that is, removing weak or anomalous connections that lack reasonable support in domain knowledge or clearly violate common sense. Subsequently, the retained association weights are normalized (e.g., using Softmax or row normalization) to transform them into a probabilistic form or relative strength matrix that is easy to interpret and use, ultimately outputting the inter-index connection weight matrix. This matrix not only encapsulates the statistical patterns mined from the data but also integrates the structured knowledge of domain experts, thus more robustly and reliably depicting the interaction network between various elements in the hospital management system, providing accurate edge weight data for the subsequent construction of a high-fidelity hospital management graph.

[0084] In the above implementation, the well-organized but superficial management data matrix is ​​transformed into a probability distribution model that deeply describes the dynamic statistical laws of each indicator, and a connection weight matrix that accurately quantifies the complex interdependencies between indicators. This technical solution can capture the nonlinear correlations that are prevalent in medical management data, ensuring that subsequent graph construction, importance assessment, and dynamic scenario simulation are based on objective laws learned from the data, rather than subjective empirical assumptions, thereby improving the scientific rigor and accuracy of the entire hospital management simulation and evaluation system.

[0085] Reference Figure 3 As one implementation of step S104, the steps of constructing a hospital management graph structure based on the inter-indicator connection weight matrix, where nodes represent management entities, edge weights are assigned by the inter-indicator connection weight matrix, and the management importance score of each node is calculated using an improved PageRank algorithm include:

[0086] Step S301: Based on the inter-indicator connection weight matrix, the hospital management entity is mapped to the graph node, and the inter-indicator connection weight is assigned to the edge between the corresponding node to generate the hospital management graph.

[0087] In this context, each element in the inter-indicator connection weight matrix quantifies the degree of mutual influence between any two management indicators.

[0088] In this embodiment, the construction process of the hospital management graph includes two key mappings: First, specific management entities (such as cardiology, radiology, pharmacy, nursing units, etc.) are defined as nodes in the graph, and these nodes are the subjects and objects of management activities. Second, the corresponding weight values ​​in the inter-indicator connection weight matrix are assigned to the edges connecting these nodes.

[0089] For example, if data analysis reveals a strong correlation between "outpatient visits" and "drug inventory turnover," and these two indicators belong to the "Outpatient Department" node and the "Pharmacy Department" node, respectively, then an edge will be established between these two nodes, with its weight representing the previously learned correlation strength. In this way, the entire hospital management system is abstracted into a complex, weighted network graph. Its structure directly reflects the dynamic interaction relationships between departments and elements based on actual operational data, providing an ideal model foundation for subsequent mathematical analysis.

[0090] Step S302: Calculate the hierarchical depth and topological clustering coefficient of each node in the hospital management graph;

[0091] Hierarchical depth refers to the shortest path number of steps required to reach a specific node from the root node of the graph (usually defined as the highest decision-making level in the hospital). This indicator quantifies the inherent hierarchical management structure of the hospital. For example, a general ward node has a greater hierarchical depth than its directly subordinate department nodes.

[0092] The topological clustering coefficient measures the tightness of connections between a node's neighboring nodes, thus characterizing the "small group" or "cluster" effect of the network. For example, nodes within a surgical system (such as the operating room, anesthesiology department, and surgical ward) may be very tightly connected, resulting in a high clustering coefficient.

[0093] By calculating these two parameters, the aim is to mathematically describe two core characteristics of the hospital management map: the hierarchical nature of its command chain transmission and the community-based nature of its business unit collaboration.

[0094] Step S303: Set the path decay weight according to the hierarchical depth, and dynamically configure the convergence factor parameters of the PageRank algorithm based on the topological clustering coefficient.

[0095] Specifically, an adaptive improvement to the classic PageRank algorithm is made to address the unique structure of hospital management graphs. Classic PageRank assumes a relatively equal and random network structure, using a fixed damping factor to simulate random jump behavior. However, hospital management systems have a strict hierarchical structure and specific collaboration patterns, and directly applying the classic algorithm can lead to distortion in importance scoring. Therefore, dynamically configuring the convergence factor parameter is necessary.

[0096] In this embodiment of the application, the specific formula for dynamically configuring the PageRank convergence factor α is as follows:

[0097] ;

[0098] Among them, C i For nodes The aggregation coefficient, , These represent the global mean and standard deviation.

[0099] Specifically, hierarchy depth is used to set path decay weights, reflecting the natural attenuation of influence or information when transmitted across different organizational levels in real-world management (for example, a hospital director's instruction has a much greater impact on a department head than on a regular nurse). Simultaneously, the topological clustering coefficient is used to adjust the probability distribution of random jumps, reflecting that random access or intervention is more likely to occur within closely related business clusters in reality (for example, a question about surgical quality is more likely to be discussed and resolved within high-cluster clusters such as the operating room or anesthesiology department, rather than randomly jumping to a support department). This dynamic configuration makes the random walk behavior of the PageRank algorithm more closely aligned with the actual operational logic of hospital management.

[0100] Step S304: Obtain the standardized scores of key indicators in the medical quality evaluation data as the inherent importance weight of the node;

[0101] The ultimate management importance of a node should not be solely determined by its connectivity; its inherent value or basic performance is equally crucial. This can be achieved by introducing standardized scores from key indicators in medical quality evaluation data. For example, core quality and efficiency indicators such as "success rate in rescuing critically ill patients," "in-hospital infection control rate," and "score for scientific research results transformation" for each department can be standardized (e.g., Min-Max normalization or Z-score standardization) to become a score between 0 and 1 or conforming to a specific distribution, serving as the inherent importance weight of each node. This weight represents the independent value of the management entity based on its final output after removing network effects.

[0102] Step S305: Using the improved PageRank algorithm, combining the convergence factor parameter and the inherent importance weight of the node, the management importance score of each node is iteratively calculated until convergence.

[0103] In this embodiment, after the improved PageRank algorithm starts running, it simulates a "virtual manager" performing a random walk in the graph. This walker moves along connecting edges, and its movement probability is influenced by the edge weights (i.e., the strength of the correlation between indicators). Simultaneously, its random jumping behavior is no longer unbiased but guided by the convergence factor parameters dynamically configured in the previous step, making it more inclined to move between nodes of similar hierarchy or with clustered business functions. During the iteration process, the importance score of a node depends on the importance of the other nodes connected to it, as well as the weights of these connecting edges. The iteration continues until the management importance scores of all nodes are less than a preset convergence threshold, indicating that the system has reached a stable state.

[0104] Ultimately, the output management importance score is a composite metric that integrates a node’s network centrality (determined by the original connection weights and dynamic jump rules), organizational structure position (adjusted by the hierarchy depth), local cooperation strength (adjusted by the clustering coefficient), and its own business quality (reflected by the inherent weights).

[0105] In the above implementation, the indicator association weights obtained from data mining are transformed into a concrete management graph. Two key topological features, hierarchical depth and topological clustering coefficient, are introduced to dynamically configure the core parameters of the PageRank algorithm. This allows the classic algorithm to accurately adapt to complex organizations like hospitals, which have distinct hierarchical structures and business clusters. The final calculated management importance score is no longer a simple count of connections, but a precise quantitative assessment that deeply integrates organizational hierarchy decay effects and internal collaboration patterns. This allows for a scientific and objective revelation of strategically valuable key nodes within the hospital management system, improving the accuracy and foresight of management decisions.

[0106] Reference Figure 4 As one implementation of step S105, the steps of generating node embedding vectors using a graph embedding algorithm based on preset management target configuration parameters and management importance scores, and classifying and labeling nodes using adaptive clustering analysis to output a set of node classification labels include:

[0107] Step S401: Receive the hospital management graph structure, preset management target configuration parameters, and management importance scores of nodes;

[0108] The hospital management graph structure (i.e., the set of nodes and the connections between weighted edges) is the ontology to be analyzed, encoding all interactions between entities. The preset management target configuration parameters are high-level instructions set by decision-makers, defining the ultimate purpose of this analysis. For example, is the goal to optimize the allocation efficiency of medical resources, or to identify departments with abnormal operating patterns? This parameter directly determines the optimization direction and specific implementation path of the subsequent algorithm model. The management importance score quantifies the structural influence and coreness of each management entity (node) within the overall hospital network.

[0109] Step S402: Establish a corresponding graph embedding optimization objective function based on the preset management target configuration parameters; wherein, if the preset management target configuration parameters include a medical resource optimization objective, construct a loss function that prioritizes edge weight retention; if the preset management target configuration parameters include an abnormal behavior detection objective, construct a loss function that prioritizes topology retention.

[0110] Specifically, a quantifiable criterion is constructed to guide the graph embedding algorithm in learning the node vector representation that best meets the expected purpose. The core task of graph embedding is to map high-dimensional, complex graph structures to a low-dimensional vector space, but this mapping can have multiple focuses. The preset management target configuration parameters directly determine the specific form of the optimization objective function.

[0111] In some embodiments, if the management objective is "medical resource optimization," the objective function is set to prioritize edge weight preservation. This means that the algorithm will strive to maintain the proportional relationship of connection weights between nodes in the original graph during dimensionality reduction. Because edge weights represent the intensity of resource flow or collaboration, preserving them helps to clearly identify which departments have the highest resource collaboration efficiency in the low-dimensional space. For example, the loss function based on edge weight reconstruction error will, during optimization, focus on forcing node pairs with large edge weights in the original graph (i.e., departments with close resource interaction, such as the operating room and anesthesiology department) to have closer vector distances in the low-dimensional embedding space. This ensures that the embedded vectors reflect the tightness of resource collaboration to the greatest extent, facilitating subsequent identification of resource communities or optimization of resource flow paths.

[0112] Conversely, if the management objective is "abnormal behavior detection," the objective function will be set to prioritize topology preservation. This means the algorithm will focus more on maintaining the local connectivity (neighbor structure) between nodes. Abnormal nodes often exhibit connection patterns drastically different from most nodes, and this setting amplifies the anomalousness of their topology, facilitating subsequent detection. An abnormal node (such as a department whose operational data differs significantly from other departments) often has a unique connection pattern. By prioritizing topology preservation, the vector position of such nodes in the embedding space will significantly deviate from the dense region composed of most normal nodes, thus creating favorable separability conditions for subsequent anomaly detection algorithms.

[0113] Step S403: Using the graph embedding optimization objective function as a constraint, the hospital management graph structure is input into the graph embedding algorithm, and the node weights are weighted by combining the management importance score to generate a low-dimensional node embedding vector.

[0114] In this embodiment, the established graph embedding optimization objective function is used as a constraint, and a graph embedding algorithm (such as a graph convolutional network-based method) is used to perform nonlinear transformation and compression on the input hospital management graph structure. During the iterative optimization process, the algorithm not only considers the connection relationships (edges) between nodes, but also injects the management importance score as a node weight into the calculation.

[0115] Specifically, during message passing or gradient updates, nodes with higher importance scores have a stronger influence on their neighbors, or their own vector representations are more protected from information loss during optimization. This process is similar to a social network where people are located not only based on their social circles but also on their reputation; those with higher reputations receive greater weight and attention in determining their location coordinates. Through multiple nonlinear transformations, the algorithm can aggregate high-order neighbor information layer by layer, ultimately generating a low-dimensional node embedding vector for each node.

[0116] Understandably, this low-dimensional node embedding vector comprehensively encodes the node's structural role in the graph, connection strength pattern, and its own prior importance. Furthermore, its encoding method is strictly guided by the management objective of the first step input, giving the generated vector set a distinct goal-oriented characteristic and providing high-quality, targeted feature representations for the next classification task.

[0117] Step S404: Based on the clustering task type identifier in the preset management target configuration parameters, dynamically select the clustering analysis algorithm; wherein, if the clustering task type identifier is departmental collaboration relationship analysis, call the spectral clustering algorithm; if the clustering task type identifier is management mode anomaly detection, call the outlier sensitive clustering algorithm.

[0118] Among them, the clustering task type identifier is a concretization of the management target configuration parameters. It clearly indicates the intention of grouping. Different grouping intentions require different mathematical principles to be implemented. The system can respond to this identifier through a dynamic selection mechanism.

[0119] In some embodiments, when categorized as "departmental collaboration analysis," the goal is to discover closely connected, naturally formed collaborative groups whose shape in the vector space may be arbitrary (non-spherical). In this case, calling a spectral clustering algorithm is an ideal choice because it excels at discovering distributions of arbitrary shapes in the embedding space and can effectively identify departmental clusters formed based on actual collaborative relationships.

[0120] When flagged as "Management Mode Anomaly Detection," the core objective is to first separate the very few outliers with abnormal behavior patterns from the mainstream group, and then group the normal nodes. Therefore, the system will call an outlier-sensitive clustering algorithm, which focuses first on the density distribution of nodes, rather than rushing to assign all nodes to a certain cluster.

[0121] Step S405: The selected clustering analysis algorithm is used to group and label the low-dimensional node embedding vectors, and output a set of node classification labels.

[0122] The selected clustering algorithm is used to divide the points (i.e., node embedding vectors) in the low-dimensional vector space and assign a classification label to each node.

[0123] Specifically, if spectral clustering is used, the algorithm calculates the similarity matrix of all embedded vectors, finds the optimal segmentation dimension of the data through feature decomposition, and finally divides the nodes into several clusters with high internal vector similarity and low inter-group similarity, assigning each node a departmental collaboration group label. If outlier-sensitive clustering algorithm is used, it first calculates the local density factor of each node's embedded vector (usually referring to the number of other nodes in a given radius neighborhood) and compares it with the global density threshold. Nodes with local densities far below the threshold are isolated in space and are marked as "abnormal nodes." The remaining nodes are considered normal patterns and can be grouped according to the distance of the vectors to the centroid using centroid-based clustering methods (such as K-means) to form "normal management pattern category" labels.

[0124] Ultimately, the output set of node classification labels is no longer a simple administrative number, but a dynamic classification result with profound management implications, derived through complex calculations based on network structure, node importance, and specific management objectives. For example, in an anomaly detection task, the output may contain one "abnormal" label and several "normal pattern" labels; in a collaboration relationship analysis task, the output may contain labels such as "surgical collaboration group," "internal medicine collaboration group," and "medical technology support group."

[0125] In the above implementation, abstract management objectives are transformed into specific mathematical optimization functions, and node importance scores are incorporated as constraints into the graph embedding learning process, resulting in node vector representations rich in goal-oriented semantic information. Furthermore, a task-identifier-driven intelligent clustering algorithm selection mechanism ensures a high degree of alignment between the group discovery process and the final analytical intent. The final output set of node classification labels represents a dynamic knowledge system that deeply aligns with management logic and clearly reveals internal hospital collaboration patterns, resource flows, or abnormal risks, providing a scientific and reliable grouping basis for subsequent precise comprehensive assessments and generative adversarial simulations.

[0126] Reference Figure 5 As one implementation of step S106, the step of calculating the comprehensive management evaluation score by category grouping, combining the inter-indicator connection weight matrix, management importance score, and node classification label set, includes:

[0127] Step S501: Receive the inter-indicator connection weight matrix, management importance score, and node classification label set;

[0128] The inter-indicator connection weight matrix quantifies the knowledge base of the inherent correlation strength among various management indicators. The management importance score quantifies the structural importance and influence of each node in the hospital management map. The node classification label set is a dynamic community based on data-driven division of management entities (such as departments) according to their actual operating mode, collaborative relationship, or functional characteristics, such as "surgical collaboration group" and "internal medicine collaboration group".

[0129] Step S502: Divide the managed entities into multiple categories based on the node classification label set;

[0130] Specifically, the division based on node classification label sets essentially groups management entities with similar characteristics or belonging to the same collaborative circle into the same evaluation unit. For example, all surgical departments, their associated anesthesiology departments, and operating rooms are grouped into the "perioperative management group," while outpatient departments, emergency departments, and radiology departments are grouped into the "emergency outpatient treatment group." This grouping method transcends the physical or administrative boundaries of the hospital, forming a natural aggregation based on actual business processes and data performance.

[0131] Step S503: Calculate the average management importance score of nodes within each category group;

[0132] In this embodiment of the application, the importance score of node management can be differentially weighted according to the medical risk level parameter, and the arithmetic mean of the weighted values ​​can be calculated as the group average.

[0133] For example, high-risk groups like the "Emergency and Critical Care Group," which includes ICU and emergency departments, are assigned a higher weighting factor when calculating the average score. This means that even if the importance scores of individual nodes within a high-risk group are not high, the group as a whole will receive due attention in the evaluation due to the high-risk nature of the medical tasks it undertakes. The average importance score of the group calculated in this way not only reflects the structural influence of the group but also incorporates the risk attributes of the medical services themselves, making it a more suitable initial benchmark for the characteristics of the healthcare industry.

[0134] Step S504: Extract the inter-group correlation weights from the inter-index connection weight matrix to generate the inter-group influence factor matrix;

[0135] Specifically, the steps for generating the inter-group influence factor matrix include: identifying cross-group association edges in the inter-indicator connection weight matrix, using the inverse ratio of the association edge weight value to the association path length as the benchmark for influence measurement; and constructing an N×N weight factor matrix between groups (N being the number of groups).

[0136] For example, the efficiency of the "outpatient and emergency treatment group" directly affects the bed turnover of the "inpatient ward management group," while the service capacity of the "medical technology support group" (such as laboratory and radiology) restricts the treatment speed of the former two. This step identifies the correlation edges connecting different groups from the global indicator inter-indicator connection weight matrix and extracts their weights. However, instead of directly using these weights, they are combined with the length of the correlation path to generate an inter-group influence factor. For example, a high-weight correlation with a short path (such as outpatient prescription and pharmacy dispensing) indicates a strong coupling relationship between the two groups, resulting in a large influence factor; while a low-weight correlation or a long path (such as administrative office and operating room) has a small influence factor.

[0137] Ultimately, the generated inter-group influence factor matrix is ​​an N×N matrix (N being the number of groups), which accurately depicts the strength of the network relationship of interdependence and mutual constraints among the various functional modules within the hospital.

[0138] Step S505: Perform cross-weighted calculation based on the inter-group influence factor matrix and the average management importance score, and output the evaluation benchmark score for each group;

[0139] In this embodiment, the specific steps of cross-weighted calculation include: normalizing the inter-group influence factor matrix into a transition probability distribution matrix; and performing Markov iterative correction on the average management importance score, as shown in the formula:

[0140] ;

[0141] in, Vk is the corrected average score of group k in the next iteration (time t+1), which is the final output group evaluation benchmark score; Vk is the average management importance score of group k; Pk kj The influence probability from group k to j is obtained by normalizing the inter-group influence factor matrix, which quantifies the transmission strength of the state change of group j to group k. α is the corrected average score of other groups j in the current iteration round (time t), and this value is dynamically updated during iteration; α is a key smoothing coefficient (also known as a damping factor), whose value is between 0 and 1, used to adjust the weight distribution of "self-performance" and "external influence" in the final score: when α approaches 1, the score mainly depends on the inherent level of the group itself, and the system tends to be statically isolated for evaluation; when α approaches 0, the score is dominated by the inter-group influence network, which better reflects the dynamic correlation of the system.

[0142] Understandably, each group's initial assessment baseline score is its average management importance score, but this score is isolated. The correction process in this embodiment takes into account the influence of other groups on the current group: a group's final assessment score depends partly on its own initial score and partly on the assessment scores of all other groups that can influence it, using an inter-group influence factor as the transition probability. Through iterative calculations with a damping factor, strongly correlated group scores will mutually reinforce each other, while weakly correlated groups remain relatively independent. This process dynamically reflects the cascading effect of hospital management; for example, an efficiency improvement in a medical technology department will, through this model, transfer its positive impact to the assessment scores of clinical departments that depend on it. The final output group assessment baseline score is a more realistic score adjusted for intra-system interactions.

[0143] Step S506: Generate a comprehensive management assessment score by combining the benchmark scores of all groups.

[0144] Specifically, based on the dimension weight vector in the preset medical management target configuration parameters, the benchmark scores of each group are globally linearly combined to generate a comprehensive management evaluation score.

[0145] Specifically, the adjusted benchmark scores of each group are combined into a single quantitative indicator that represents the overall management level of the hospital. This synthesis process is not a simple summation and averaging, but rather a global linear combination based on a pre-defined dimensional weight vector in the medical management goal configuration parameters. This weight vector reflects the strategic priorities of decision-makers within a certain period. For example, if the current focus is on medical quality and safety, then groups containing core quality indicators (such as the "Medical Quality Control Group") will be given higher weights; if the focus is on operational efficiency, then the "Process Optimization Group" will have a higher weight. Through this weighted synthesis, the final comprehensive management evaluation score not only objectively reflects the hospital's current overall operational status but also incorporates the strategic direction of management.

[0146] In the above implementation, management entities are dynamically grouped according to a data-driven model, and medical risk weighting and Markov-based iterative correction based on inter-group influence factors are introduced into the calculation. This ensures that the final comprehensive management evaluation score not only reflects the independent value of each component but also profoundly captures the dynamic game relationships of interrelationships and mutual influences among the functional modules within a complex hospital system. This comprehensive management evaluation score provides hospital managers with a more comprehensive, accurate, and strategically significant global effectiveness diagnostic report, providing an indispensable and scientifically reliable benchmark for subsequent management strategy simulation and optimization using generative adversarial networks.

[0147] Reference Figure 6 As one implementation of step S107, the steps of inputting the indicator probability distribution model, comprehensive management evaluation score, node classification label set, and external intervention strategy parameters into a pre-configured generative adversarial network, and outputting a dynamic management scenario prediction report and optimized strategy scheme after adversarial training include:

[0148] Step S601: Receive the indicator probability distribution model, comprehensive management evaluation score, node classification label set, and external intervention strategy parameters;

[0149] The indicator probability distribution model characterizes the historical normal fluctuation range and statistical regularity of various hospital management indicators (such as bed occupancy rate and drug ratio), providing mathematical constraints on the distribution of real data for the subsequent generation process. The management importance score quantifies the core importance of each node in the hospital management map, serving as a key consideration for the discriminator when evaluating the generated results. The node classification label set defines data-driven management entity groupings (such as clinical department groups and medical technology support groups), providing a grouping label foundation for generating differentiated and customized strategies.

[0150] External intervention strategy parameters represent the manager's experience, assumptions, or strategic intentions (such as "focusing on improving emergency room efficiency"), and their introduction ensures that the algorithm's generated results do not completely deviate from the manager's decision-making framework. These four types of inputs make subsequent GAN training no longer simply data-driven, but a hybrid intelligent process that integrates historical patterns, structural features, business classifications, and expert priors.

[0151] Step S602: Construct the prior constraints of the generator network based on the index probability distribution model;

[0152] Specifically, the probability distribution model of the indicator is transformed into the probability distribution loss function of the hidden layer of the generator network, and the fluctuation range of the indicator of the generated decision sequence is constrained to be within the preset standard deviation range.

[0153] Specifically, this step involves setting boundaries and rules for the learning behavior of the generator network to ensure that the simulated decision sequences it generates are statistically reasonable and reliable. The goal of the generator network is to generate new, seemingly realistic management decision data; without constraints, it may generate absurd results that completely deviate from the actual operational patterns of the hospital. Therefore, it is necessary to transform the indicator probability distribution model into a probability distribution loss function for the generator's hidden layers.

[0154] Specifically, during the generator's training process, besides aiming to deceive the discriminator, the changes in indicators caused by the generated decision sequences must maintain a high likelihood with the probability distribution learned from historical data. For example, the generator can simulate the decision to "increase the number of surgical procedures by 10%," but the subsequent changes in indicators caused by this decision (such as increases in anesthesiology workload or drug consumption) must be constrained within the historical standard deviation range. This is equivalent to putting a bridle of historical data on the generator's imagination, preventing it from generating unrealistic scenarios, thus ensuring the seriousness and reference value of the simulation experiment.

[0155] Step S603: Configure the dynamic evaluation threshold parameters of the discriminator network according to the percentile distribution of the comprehensive management evaluation scores;

[0156] In this embodiment, a dynamic threshold is set based on the percentile distribution of the comprehensive management assessment scores:

[0157] ;

[0158] Where Qx is the importance score x quantile, and κ is the medical risk moderating factor.

[0159] Understandably, in a hospital management setting, decisions made at different stages have vastly different impacts. For example, a decision concerning a core, pivotal department (such as the operating room) is far more important than a decision concerning a general administrative department. Therefore, a dynamic evaluation threshold needs to be configured based on a comprehensive management assessment score.

[0160] In practice, the percentile of the comprehensive management assessment score (e.g., the 75th percentile Q75) is used as the benchmark, and a medical risk moderating factor κ is introduced for fine-tuning. This allows the discriminator to adopt stricter evaluation criteria (higher threshold τd) for the decision sequences generated by nodes with high importance scores (i.e., key management entities). This means that if the strategy proposed by the generator has a poor impact assessment on key departments, it will be difficult to pass the discriminator's review, thus guiding the generator to learn and formulate strategies that better ensure the stable operation of the hospital's core business.

[0161] Step S604: Extract the grouping identifier from the node classification label set and map it to the context feature layer of the generative adversarial network;

[0162] Different departments within a hospital have unique operational logic and needs; for example, optimization strategies suitable for outpatient departments may be completely unsuitable for inpatient departments. Therefore, it is necessary to encode the grouping identifiers (such as "outpatient group" and "surgical group") in the node classification label set into feature vectors and inject them into the context feature layer of the generative adversarial network (GAN). For the generator, this grouping identifier acts as a conditional signal, guiding it to generate targeted decisions for different groups. For instance, when the generator receives the identifier for the "surgical group," it is more inclined to generate strategies related to surgical scheduling and high-value consumable management; while when the identifier is for the "administrative group," it shifts to generating strategies related to process approval optimization. This makes the final generated strategy solutions no longer general but highly adapted to specific management objects, improving the feasibility and effectiveness of the strategies.

[0163] Step S605: Encode the external intervention strategy parameters into a policy vector and incorporate it into the input noise space of the generator network;

[0164] In this embodiment, the generator typically starts by generating data from a random noise vector. The innovation lies in encoding external intervention strategy parameters (i.e., the manager's subjective will or pilot strategy) into a strategy vector. This vector is then concatenated or fused with the original random noise and used as input to the generator. This guides the generator's exploration direction, making it less random and more aligned with the manager's interests. For example, if the manager inputs the strategy parameter "try to reduce the average length of hospital stay," the generator will, in its initial random exploration, be more inclined to generate decision sequences that might shorten the length of hospital stay for trial.

[0165] Step S606: Through the generator network, combined with prior constraints and context feature layers, a simulated management decision sequence is iteratively generated;

[0166] Specifically, the generator network (typically employing recurrent neural networks such as LSTM or its variants to model the temporal relevance of decisions) receives noise and contextual features that incorporate the policy intent, and gradually imagines a future simulated sequence of management decisions (e.g., "Month 1, increase the number of nurses in the cardiology department by 5%; Month 2, adjust the frequency of equipment inspections in the radiology department to once a week...").

[0167] Step S607: The validity of the simulated management decision sequence is verified by using a discriminator network with a dynamic evaluation threshold parameter, and the decision confidence score is output.

[0168] The validity verification includes mapping the decision sequence to the hospital management graph structure to calculate the topological consistency index and verifying the constraint boundary satisfaction of the medical quality evaluation data.

[0169] Specifically, the discriminator network (usually a classification or regression network) not only determines whether the simulated management decision sequence appears "realistic" (i.e., whether it conforms to historical data patterns), but more importantly, it uses dynamic evaluation threshold parameters to extrapolate and evaluate the potential consequences of executing the sequence. It executes the generated decision sequence in a virtual environment, predicts its impact on various management indicators and overall evaluation scores, and checks whether these impacts exceed thresholds, violate hard constraints on medical quality (such as the hospital infection rate not exceeding a certain value), and whether the changes in inter-node relationships caused by the decision are consistent with the topological consistency logic of the hospital management graph structure (e.g., whether a strategy of drastically reducing pharmacy inventory would lead to unreasonable and drastic changes in the connection weights with multiple clinical departments). After this series of complex validity verifications, the discriminator outputs a decision confidence score, which comprehensively reflects the feasibility, effectiveness, and risk level of the generated strategy.

[0170] Step S608: When the decision confidence score meets the preset verification threshold, the decision sequence is parsed to generate a dynamic management scenario prediction report and optimization strategy scheme.

[0171] Specifically, when the confidence score of one or more generated decision sequences meets a preset verification threshold (e.g., higher than 0.85), it is considered to have passed the adversarial training assessment. The system will analyze this high-confidence sequence: first, it will be transformed into a dynamic management scenario prediction report, which will visually display the predicted values ​​and fluctuation ranges of key management indicators (such as financial, efficiency, and quality) at different future points in time after executing the strategy sequence, in the form of charts, trend lines, etc., so that managers can foresee the potential chain reactions of the decision.

[0172] Secondly, the decision sequence is post-processed and optimized to generate optimized strategy solutions: high-frequency intervention operations are extracted from the decision sequence to form a basic strategy set. Based on the node classification label set, the basic strategy set is filtered for domain adaptability, eliminating those strategies that are obviously not applicable to a certain type of entity (for example, filtering out the strategy suggestion of "increasing scientific research funding" from the logistics department's plan). Necessary implementation details and risk warnings are added, and finally, a management optimization proposal that is both innovative and robust is formed.

[0173] In the above implementation, domain knowledge (such as indicator distribution, node importance, and business grouping) is deeply embedded into the training framework of the generative adversarial network as prior constraints and evaluation criteria. External intervention strategies are introduced to guide the exploration direction, ensuring that the generated management decision sequences conform to both historical objective laws and subjective management intentions, while also possessing business relevance and security. The final output of dynamic scenario prediction and optimization strategies provides hospital managers with a powerful tool to proactively evaluate decision effectiveness and scientifically formulate optimization paths, enhancing the foresight, scientific rigor, and risk resistance of management decisions, and realizing a shift from passive, reactive management to proactive, simulation-based management.

[0174] Reference Figure 7 As a further implementation of the hospital management scenario simulation and evaluation method, after the step of outputting the dynamic management scenario prediction report and optimization strategy in step S107, the method further includes:

[0175] Step S701: Access the hospital's real-time business monitoring data stream and extract the actual changes in management indicators within a preset period;

[0176] The real-time business monitoring data stream in hospitals is a complex dataset that continuously emerges from various parts of the hospital's information system, including but not limited to transaction logs from the Hospital Resource Planning (HRP) system, patient event records from the Clinical Information System (CIS), and physical resource status data transmitted from IoT devices. This data accurately reflects the actual changes in status resulting from the implementation of management strategies.

[0177] In this embodiment, according to a preset time period (e.g., every 4 hours or every day), the system samples and calculates the specific changes in key management indicators (such as bed occupancy rate, average length of stay, and drug turnover rate) from these high-speed, heterogeneous data streams within that period. This process is not a simple data transfer, but involves cleaning, aligning, and aggregating multi-source data to ensure that the extracted actual change values ​​are a dimensionally unified quantitative vector that can be accurately compared later. For example, when a strategy aimed at optimizing outpatient processes is implemented, the system needs to extract the actual reduction in the "average waiting time from registration to consultation" indicator from multiple independent data sources such as the registration system, doctor workstations, and the laboratory.

[0178] Step S702: Perform deviation analysis between the actual changes in management indicators and the dynamic management scenario prediction report to generate a set of strategy execution deviation coefficients;

[0179] Among them, the dynamic management scenario prediction report output by the generative adversarial network (GAN) is a theoretical effect that the policy execution should achieve under ideal conditions, derived from historical data and algorithm models. This step is to test this theory in the real world.

[0180] Specifically, the core of deviation analysis lies in calculating the difference between the predicted value (V_pred) and the actual observed value (V_real) of each management indicator. However, simple absolute or relative differences may not accurately reflect the differences in importance between different indicators. Therefore, this embodiment introduces a more scientific standardization process and incorporates management importance scores as a weighting factor. That is, for nodes with high management importance scores (such as core medical units like ICUs and operating rooms), their indicator deviations will be given greater weight, thus occupying a more important position in the overall deviation assessment.

[0181] Ultimately, the generated "strategy execution deviation coefficient set" is a set of values ​​that quantifies the degree of deviation of the execution effect of each strategy in different management dimensions. It intuitively reveals which links did not achieve the expected results and the severity of the deviation, providing accurate location information for subsequent root cause analysis.

[0182] Step S703: Perform grouping and clustering analysis on the deviation coefficient set based on the node classification label set to identify high-deviation node groups;

[0183] A hospital is a complex system composed of numerous departments. While a deviation at a single node may be an isolated case, a collective deviation among a group of nodes with similar functions often indicates a systemic problem. This step utilizes a node classification label set (e.g., labeling all surgical wards as "surgical ward group" and all laboratory departments as "medical technology support group") to group nodes with the same or similar business attributes and their corresponding deviation coefficients together.

[0184] Building upon this foundation, clustering algorithms (such as the density-based DBSCAN algorithm) are employed for secondary analysis within each group. The aim is to identify outlier clusters with deviation coefficients significantly higher than the group's normal level. For example, within the "Internal Medicine Medication Management Group," several wards might exhibit abnormally high adverse drug reaction rates, forming a "high-deviation node group." This technical solution automatically identifies problems as not randomly distributed but concentrated in specific functional areas or workflows. This shifts the focus of analysis from "which department has a problem" to "which business model or management process has systemic risks," improving the efficiency and accuracy of problem localization.

[0185] Step S704: Based on the hospital management graph structure and the connection weight matrix between indicators, trace the key influence paths of high-bias node grouping;

[0186] The underlying logic of this step is to delve into the topology of the hospital management system and analyze the root cause transmission path. Essentially, it's a fault penetration analysis based on complex network theory. The previously constructed hospital management map vividly depicts the interactions between various management entities, while the connection weight matrix between indicators quantifies the strength of these relationships.

[0187] In this embodiment, the steps for tracing the key impact paths of high-bias node groups include: initiating a random walk algorithm within the high-bias node group; calculating the path influence score based on the inter-indicator connection weight matrix; and selecting the top K paths whose influence scores exceed a preset threshold. The formula for calculating the path influence score is as follows:

[0188] ;

[0189] In the above formula, w k Let d be the edge weight. k Where ρ is the path depth and ρ is the medical management attenuation factor;

[0190] Specifically, after identifying high-bias node groups, this step initiates a reverse tracing mechanism: starting from these high-bias nodes, it traces upstream nodes and paths that may influence them along the connecting edges in the management graph. In this process, the influence calculation of the path considers not only the weight of the connecting edges (relationship strength) but also introduces a medical attenuation factor to simulate the natural decay of influence during transmission (e.g., the influence weakens when crossing levels or passing through multiple departments). By calculating the combined influence of all possible backtracking paths and filtering out critical paths exceeding a specific threshold, the system can pinpoint the root cause or main influencing factor leading to high bias. For example, tracing back to find that the critical influencing path of inefficient operating rooms (high-bias node) may originate from the upstream node of "insufficient anesthesiologist scheduling" and propagate through the edge of "preoperative preparation process."

[0191] Step S705: Correct the external intervention strategy parameters based on the key impact paths and generate an optimized strategy enhancement package;

[0192] In some embodiments, the step of generating an optimization strategy enhancement package includes: mapping key impact paths to a constraint rule base of a medical management knowledge graph; injecting a medical quality control strategy template when the path node contains a clinical department; and injecting a resource scheduling strategy template when the path node contains an administrative department.

[0193] Specifically, based on the key impact paths identified through source tracing, the system can accurately understand which aspects of the existing strategy are deficient or uncovered. For example, if the key impact path indicates that the problem originates from inter-departmental collaboration processes, then the focus of correction will be on optimizing cross-departmental collaboration protocols. Modifying external intervention strategy parameters involves generating more targeted management instructions based on the identified root causes and transmission paths. This process is not blind experimentation but rather involves integrating prior knowledge from the field of healthcare management (such as clinical pathway guidelines and JCI standards for healthcare quality management) for verification and enrichment, ensuring that the generated "optimization strategy enhancement package" is not only highly targeted but also conforms to healthcare industry norms and best practices.

[0194] In addition, this optimization strategy enhancement package may include new resource allocation plans, process optimization suggestions, or risk management measures, with the aim of cutting off the transmission chain of problems at their source or strengthening weak links.

[0195] Step S706: Feed the optimized policy enhancement package back to the input layer of the generative adversarial network for iterative training.

[0196] Specifically, feeding the newly generated "optimization strategy enhancement package" back to the input layer of the GAN means that the problem discovered from the real world, the root cause analyzed, and the optimization scheme derived are used as new knowledge and constraints and re-injected into the model for training.

[0197] Subsequently, in the next adversarial training, the generator will attempt to generate policy sequences that incorporate this new knowledge, while the discriminator will adjust its evaluation criteria based on the latest feedback (including deviations in the actual effects of previous policies). Through this iterative training, the GAN model can continuously learn from actual performance, gradually correcting its internal simulated dynamic model of hospital management, making its predictions of future policy effects increasingly accurate, and the generated policy solutions increasingly aligned with the actual operating environment and constraints of the hospital.

[0198] In the above implementation, a real-time data access and deviation analysis channel is established. The forward-looking predictions of the generative adversarial network are compared with the complex operational realities. Utilizing prior knowledge such as node classification and management graphs, the system intelligently completes the attribution process from identifying deviation phenomena to tracing the systemic root causes. Furthermore, based on the accurate results of the attribution, targeted strategy enhancement packages are generated. Through a feedback loop-driven model, iterative optimization is performed, ultimately constructing an adaptive intelligent system capable of continuously learning from actual implementation results and constantly evolving its strategy generation capabilities. This improves the success rate of management strategy implementation, risk controllability, and long-term adaptability, laying a technical foundation for achieving refined and intelligent hospital operations.

[0199] Reference Figure 8 As a further implementation of the hospital management scenario simulation and evaluation method, after the step of outputting the dynamic management scenario prediction report and optimization strategy in step S107, the method further includes:

[0200] Step S801: Obtain the set of public health emergency cases from the pre-stored hospital emergency event history database;

[0201] Among them, the case study of public health emergencies is a high-quality data collection that has been rigorously screened and structured. Each case should contain a record of the complete life cycle of the event, from the initial trigger, development process, peak impact to subsequent recovery.

[0202] It should be noted that these cases need to have typical medical characteristics. For example, for infectious disease outbreaks, epidemiological correlation data should be included (such as the basic reproduction number R0 and cluster incidence characteristics); for large-scale trauma emergency events, the intensity of resource impact needs to be recorded (such as peak emergency room admissions and operating room overload); for systemic failures of critical equipment, the spatiotemporal propagation pattern of its impact and the resulting cascading failure paths of departments need to be clearly defined.

[0203] Understandably, the fundamental purpose of obtaining such a case set is to provide an empirical foundation for subsequent pattern recognition and model testing that is based on reality and covers a variety of extreme scenarios, ensuring that the risk patterns learned by the model are not imagined out of thin air, but rather highly representative samples of medical management vulnerabilities that have been tested in practice.

[0204] Step S802: Extract abnormal fluctuation pattern data of management indicators corresponding to each event in the case set of public health emergencies;

[0205] Among them, in-depth feature mining of historical emergency events transforms qualitative event descriptions into quantifiable and calculable abnormal fluctuation patterns, which essentially identifies the pathological characteristics of the hospital management system under extreme conditions.

[0206] In this embodiment, the abnormal fluctuation pattern is a quantitative manifestation of the system's behavioral instability under pressure. The extraction process requires the use of medical statistics and time series data analysis methods to characterize it from multiple dimensions: In the time dimension, it is necessary to focus on the rate of change (first derivative) and acceleration (second derivative) of key management indicators (such as bed occupancy rate and emergency drug inventory) to capture the dynamic process of the system transitioning from normal to critical state; In the spatial dimension, it is necessary to calculate the covariance and correlation strength of indicators between different departments or functional units to reveal the transmission effect of shock within the system; In addition, it is also necessary to combine clinical medical knowledge to set critical safety thresholds for various indicators (such as the number of safe days for blood product inventory).

[0207] The innovation of this step lies in the detailed modeling of medical attenuation characteristics. For example, extreme value sequences of bed occupancy rates can be extracted from cases of nosocomial infection outbreaks; abrupt gradients in drug consumption rates can be extracted from large-scale emergency cases; and examination efficiency decay curves can be analyzed through medical equipment failure event logs, with attempts made to fit the key inflection points from normal to collapse states using complex mathematical models (such as fractional differential equations).

[0208] Step S803: Encode the abnormal fluctuation pattern data of management indicators into a spatiotemporal feature tensor;

[0209] The impact of hospital emergencies is multi-dimensional, involving temporal evolution, spatial propagation, and intertwined changes across multiple management dimensions. Spatiotemporal feature tensors can perfectly accommodate this complexity.

[0210] Specifically, a third-order or higher-order tensor can be constructed: one dimension represents the time series, sliced ​​at a fixed time granularity (e.g., 15 minutes), covering key time windows before and after the event; the second dimension represents spatial or topological coordinates, encoding various departments or management entities based on a previously constructed hospital management map to reflect the geographical or logical distribution of the impact; the third dimension represents different business indicator channels, such as medical quality, resource consumption, and financial indicators, which record the event's performance in different aspects in parallel. During the tensor construction process, evidence-based medicine rules can also be embedded. For example, events activated by the rule "excessive use of antimicrobial drugs" can be mapped to signal enhancement in a specific channel of the tensor. Through this encoding, a complex emergency event is transformed into a structured mathematical object containing rich spatiotemporal and semantic information, laying the foundation for accurate comparison with subsequent predictive models.

[0211] Step S804: Perform coupling analysis on the spatiotemporal feature tensor and the dynamic management scenario prediction report;

[0212] Specifically, a deep comparison is performed between the feature tensor (T_case) encoded from historical cases and the feature tensor (T_pred) corresponding to the prediction report generated by the Generative Adversarial Network (GAN) in normal mode. This comparison is typically achieved by calculating a similarity metric between tensors (such as distance based on the Frobenius inner product), resulting in a coupling coefficient C_s. This coefficient quantifies the degree to which the model's predictive behavior deviates from the historical crisis pattern.

[0213] Furthermore, this embodiment introduces a dynamic threshold mechanism called the medical risk tolerance coefficient α_s. This medical risk tolerance coefficient can be set differently according to the importance and sensitivity of different departments or risk types (for example, for key departments such as the ICU and emergency department, the tolerance coefficient α_s is set lower, requiring the model prediction to be highly consistent with historical crisis patterns). When the coupling coefficient C_s is greater than the tolerance coefficient α_s, it indicates that the model has significant deficiencies in simulating and responding to this type of historical crisis, which is equivalent to identifying an "immune deficiency blind spot" in the model's cognition, providing a clear target for subsequent targeted enhancement.

[0214] Step S805: Based on the coupling degree analysis results, identify the chain reaction risk paths not covered in the dynamic management scenario prediction report;

[0215] After identifying the cognitive blind spots of the model, the logical principle of this step is to deeply analyze the specific manifestations of these blind spots, that is, to find the key chain reaction risk paths that would occur in a real crisis but were not predicted by the model.

[0216] In this embodiment, based on the hospital management graph structure, starting from the highly coupled difference regions (such as the connecting edges of a certain department) identified in the previous step, and combining the relationship strength quantified by the connection weight matrix between indicators, and introducing attenuation factors that conform to medical laws (e.g., simulating the attenuation effect of pathogens spreading with increasing physical distance or process links), the most likely diffusion path of risk can be deduced. The ultimate goal is to mark those transmission chains that will eventually affect core medical resources or key medical processes. For example, a cascading failure path of "insufficient pressure in the central oxygen supply system -> postponement of elective surgery in the operating room -> increase in emergency room patients -> ICU bed overcrowding" may be identified, which was not fully simulated in the original dynamic management scenario prediction report.

[0217] Step S806: Integrate the chain reaction risk path into the external intervention strategy parameters to generate a risk mitigation strategy enhancement package;

[0218] This involves modifying external intervention strategy parameters based on traced chain reaction risk transmission paths to generate a risk mitigation strategy enhancement package. For example, for critical nodes on the path (such as resource bottlenecks), a "redundant resource buffer agreement" (such as reserving emergency drug reserves for the operating room) might be embedded; for vulnerable connections (such as inter-departmental collaboration processes), an "emergency collaborative workflow" or an "emergency response memorandum" signed with external suppliers might be inserted. The generation of all these strategies must adhere to healthcare industry standards and best practices (such as JCI standards and clinical pathway guidelines) to ensure their clinical feasibility and safety. The final enhancement package is a set of structured, executable instructions whose core purpose is to enhance the system's resilience in the face of specific historically similar crises, blocking or weakening known risk transmission chains.

[0219] In some embodiments, the specific steps for generating the risk mitigation strategy enhancement package include: injecting a redundant resource configuration strategy template into the clinical management node; inserting a supply chain resilience response protocol into the materials management node; and mapping the cascading reaction risk path as an additional evaluation dimension of the discriminator network.

[0220] Step S807: Perform incremental training by using the risk resistance strategy enhancement package as an auxiliary input constraint for the generative adversarial network.

[0221] In this regard, by integrating risk mitigation strategy enhancement packages as auxiliary input conditions or constraints into the incremental training process of generative adversarial networks (GANs), the wisdom of managers in dealing with historical crises and the weaknesses identified by the system are essentially transformed into guidance for the model training process.

[0222] Specifically, when learning to generate new policies, the generator network is required to optimize under these newly added defensive constraints, thus tending to produce policy solutions that can proactively avoid known risky paths. The discriminator network is also strengthened, adding dimensions to the evaluation of the policy solutions' resilience to risk.

[0223] Through this continuous, historical feedback-based incremental learning, GAN will gradually internalize response patterns to various typical emergency events. The management scenario prediction and optimization strategies it generates will naturally include avoidance mechanisms for similar historical risks, thereby enabling the entire hospital management simulation and evaluation system to more robustly cope with various potential future shocks.

[0224] In the above implementation, real-world public health emergencies are transformed into computable spatiotemporal feature tensors, and their coupling with the prediction reports of generative adversarial networks is analyzed in detail to accurately pinpoint blind spots and weaknesses in the model's risk perception. Furthermore, by tracing the risk transmission path, highly targeted risk mitigation strategy enhancement packages are generated, which drive incremental model training. Ultimately, this endows the system with the intrinsic ability to learn from historical crises and proactively avoid similar future risks. This technical solution upgrades hospital risk management from a passive, experience-driven traditional model to a data-driven, predictive, and adaptive intelligent model, enhancing the robustness and resilience of hospital management systems in the face of complex uncertainties.

[0225] This application also discloses a hospital management scenario simulation and evaluation system based on generative adversarial networks.

[0226] A hospital management scenario simulation and evaluation system based on generative adversarial networks, specifically including:

[0227] The multi-source data acquisition module is used to collect hospital management data from all dimensions, including departmental operation indicators, staffing parameters, financial revenue and expenditure data, and medical quality evaluation data.

[0228] The data processing module is used to perform spatiotemporal alignment and standardization on management data to generate a management data matrix;

[0229] The unsupervised learning module is used to perform unsupervised learning on the management data matrix through a restricted Boltzmann machine, and outputs the index probability distribution model and the connection weight matrix between indices.

[0230] The node importance quantification module is used to construct a hospital management graph structure based on the inter-indicator connection weight matrix, and calculate the management importance score of each node using an improved PageRank algorithm.

[0231] The dynamic clustering module is used to generate node embedding vectors through graph embedding algorithm based on preset management target configuration parameters and management importance scores, and to classify and label nodes using adaptive clustering analysis, outputting a set of node classification labels;

[0232] The comprehensive management assessment module is used to combine the weight matrix between indicators, management importance scores, and node classification label sets to calculate comprehensive management assessment scores by category.

[0233] The adversarial training output module is used to input the indicator probability distribution model, comprehensive management evaluation score, node classification label set and external intervention strategy parameters into a pre-configured generative adversarial network, and output a dynamic management scenario prediction report and optimization strategy scheme after adversarial training.

[0234] The hospital management scenario simulation and evaluation system based on generative adversarial networks according to the embodiments of this application can implement any of the above methods, and the specific working process of each module in the system can refer to the corresponding process in the above method embodiments.

[0235] In the several embodiments provided in this application, it should be understood that the provided methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for example, the division of a certain module is merely a logical functional division, and in actual implementation there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0236] This application also discloses a computer-readable storage medium.

[0237] A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above in any of the methods for simulating and evaluating hospital management scenarios based on generative adversarial networks.

[0238] The computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device; the program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0239] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A method for simulating and evaluating hospital management scenarios based on generative adversarial networks, characterized in that, The method includes: Collect comprehensive hospital management data, including departmental operation indicators, staffing parameters, financial revenue and expenditure data, and medical quality evaluation data; The management data is subjected to spatiotemporal alignment and standardization to generate a management data matrix; The management data matrix is ​​subjected to unsupervised learning using a restricted Boltzmann machine, and the output index probability distribution model and the inter-index connection weight matrix are generated. A hospital management graph structure is constructed based on the inter-indicator connection weight matrix, where nodes represent management entities, edge weights are assigned by the inter-indicator connection weight matrix, and the management importance score of each node is calculated using an improved PageRank algorithm. Based on the preset management target configuration parameters and the management importance score, a node embedding vector is generated by a graph embedding algorithm, and an adaptive clustering analysis is used to classify and label the nodes, outputting a set of node classification labels; Combining the inter-indicator connection weight matrix, management importance score, and node classification label set, the comprehensive management evaluation score is calculated by grouping by category; The probability distribution model of the indicators, the comprehensive management evaluation score, the set of node classification labels and the parameters of external intervention strategies are input into a pre-configured generative adversarial network. After adversarial training, a dynamic management scenario prediction report and an optimization strategy scheme are output.

2. The method for simulating and evaluating hospital management scenarios based on generative adversarial networks according to claim 1, characterized in that, The steps of performing unsupervised learning on the management data matrix using a restricted Boltzmann machine to output an indicator probability distribution model and an inter-indicator connection weight matrix include: The management data matrix is ​​input into the visible layer of the restricted Boltzmann machine, and the visible layer reconstruction data is generated through the bidirectional probability mapping between the visible layer and the hidden layer of the restricted Boltzmann machine. Based on the divergence between the reconstructed data of the visible layer and the management data matrix, the inter-layer connection weight parameters of the restricted Boltzmann machine are iteratively optimized to generate a set of node feature vectors and an inter-layer connection weight matrix. Based on the set of node feature vectors, the probability distribution model of each management indicator in the management data matrix is ​​fitted by Gaussian mixture modeling to obtain the indicator probability distribution model; Extract cross-index correlation weights from the inter-layer connection weight matrix; By combining a pre-constructed medical management knowledge graph, redundant connection pruning and normalization are performed on the cross-index association weights to output the inter-index connection weight matrix.

3. The method for simulating and evaluating hospital management scenarios based on generative adversarial networks according to claim 2, characterized in that, The steps for constructing a hospital management graph structure based on the inter-indicator connection weight matrix, where nodes represent management entities and edge weights are assigned by the inter-indicator connection weight matrix, and calculating the management importance score of each node using an improved PageRank algorithm, include: Based on the inter-indicator connection weight matrix, hospital management entities are mapped to graph nodes, and the inter-indicator connection weights are assigned to the edges between corresponding nodes to generate a hospital management graph. Calculate the hierarchical depth and topological clustering coefficient of each node in the hospital management graph; The path decay weight is set according to the level depth, and the convergence factor parameter of the PageRank algorithm is dynamically configured based on the topology clustering coefficient. Obtain standardized scores for key indicators in medical quality evaluation data, and use them as inherent importance weights for nodes; By using the improved PageRank algorithm, which combines the convergence factor parameter with the inherent importance weight of the node, the administrative importance score of each node is iteratively calculated until convergence.

4. The method for simulating and evaluating hospital management scenarios based on generative adversarial networks according to claim 3, characterized in that, The steps of generating node embedding vectors using a graph embedding algorithm based on preset management target configuration parameters and the management importance score, and then classifying and labeling nodes using adaptive clustering analysis to output a set of node classification labels include: Receive the hospital management map structure, preset management target configuration parameters, and the management importance score of each node; A corresponding graph embedding optimization objective function is established based on the preset management objective configuration parameters; wherein, if the preset management objective configuration parameters include a medical resource optimization objective, a loss function that prioritizes edge weight retention is constructed; if the preset management objective configuration parameters include an abnormal behavior detection objective, a loss function that prioritizes topology retention is constructed. Using the graph embedding optimization objective function as a constraint, the hospital management graph structure is input into the graph embedding algorithm, and the node weights are weighted by combining the management importance scores to generate a low-dimensional node embedding vector. Based on the clustering task type identifier in the preset management target configuration parameters, a clustering analysis algorithm is dynamically selected; wherein, if the clustering task type identifier is departmental collaboration relationship analysis, the spectral clustering algorithm is called; if the clustering task type identifier is management mode anomaly detection, the outlier sensitive clustering algorithm is called. The selected clustering analysis algorithm is used to group and label the low-dimensional node embedding vectors, and a set of node classification labels is output.

5. The method for simulating and evaluating hospital management scenarios based on generative adversarial networks according to claim 1, characterized in that, The steps for calculating the comprehensive management evaluation score by grouping by category, combining the inter-indicator connection weight matrix, management importance score, and node classification label set, include: Receive the inter-indicator connection weight matrix, management importance score, and node classification label set; The managed entities are divided into multiple categories based on the node classification label set; Calculate the average administrative importance score of nodes within each category group; Extract the inter-group correlation weights from the inter-indicator connection weight matrix to generate the inter-group influence factor matrix; The baseline score for each group is calculated by cross-weighting the inter-group impact factor matrix and the average management importance score. A comprehensive management assessment score is generated by combining the benchmark scores of all groups.

6. The method for simulating and evaluating hospital management scenarios based on generative adversarial networks according to any one of claims 1 to 5, characterized in that, The steps of inputting the indicator probability distribution model, comprehensive management evaluation score, node classification label set, and external intervention strategy parameters into a pre-configured generative adversarial network, and outputting a dynamic management scenario prediction report and optimized strategy scheme after adversarial training include: Receive the probability distribution model of the indicators, the comprehensive management evaluation score, the set of node classification labels, and the parameters of external intervention strategies; The prior constraints for constructing the generator network are based on the aforementioned index probability distribution model. Configure the dynamic evaluation threshold parameters of the discriminator network based on the percentile distribution of the comprehensive management evaluation scores; Extract the grouping identifiers from the node classification label set and map them to the context feature layer of the generative adversarial network; The external intervention policy parameters are encoded as policy vectors and incorporated into the input noise space of the generator network; The generator network, combined with the prior constraints and context feature layer, iteratively generates a simulated management decision sequence. The discriminator network uses the dynamic evaluation threshold parameter to verify the effectiveness of the simulated management decision sequence and outputs a decision confidence score. When the decision confidence score meets the preset verification threshold, the decision sequence is parsed to generate a dynamic management scenario prediction report and optimization strategy.

7. The method for simulating and evaluating hospital management scenarios based on generative adversarial networks according to claim 6, characterized in that, After the steps of outputting dynamic management scenario prediction reports and optimization strategy solutions, the following are also included: Access the hospital's real-time business monitoring data stream and extract the actual changes in management indicators within a preset period; The actual changes in the management indicators are compared with the dynamic management scenario prediction report to generate a set of strategy execution deviation coefficients. Based on the set of node classification labels, perform grouping and clustering analysis on the set of deviation coefficients to identify high-deviation node groups; Based on the hospital management graph structure and the connection weight matrix between indicators, the key influence paths of high-bias node grouping are traced. Based on the key impact paths, the external intervention strategy parameters are modified to generate an optimized strategy enhancement package; The optimized strategy enhancement package is fed back to the input layer of the generative adversarial network for iterative training.

8. The method for simulating and evaluating hospital management scenarios based on generative adversarial networks according to claim 6, characterized in that, After the steps of outputting dynamic management scenario prediction reports and optimization strategy solutions, the following are also included: Retrieve a set of public health emergency cases from a pre-stored hospital emergency event history database; Extract data on abnormal fluctuation patterns of management indicators for each event in the case set of public health emergencies; The abnormal fluctuation pattern data of the management indicators are encoded into a spatiotemporal feature tensor; The coupling degree between the spatiotemporal feature tensor and the dynamic management scenario prediction report is analyzed. Based on the coupling degree analysis results, chain reaction risk paths not covered in the dynamic management scenario prediction report are identified; The chain reaction risk path is integrated into the external intervention strategy parameters to generate a risk mitigation strategy enhancement package; The risk mitigation strategy enhancement package is used as an auxiliary input constraint for incremental training of the generative adversarial network.

9. A hospital management scenario simulation and evaluation system based on generative adversarial networks, characterized in that, The system includes: The multi-source data acquisition module is used to collect hospital management data from all dimensions, including departmental operation indicators, staffing parameters, financial revenue and expenditure data, and medical quality evaluation data. The data processing module is used to perform spatiotemporal alignment and standardization on the management data to generate a management data matrix; The unsupervised learning module is used to perform unsupervised learning on the management data matrix through a restricted Boltzmann machine, and outputs an index probability distribution model and an inter-index connection weight matrix. The node importance quantification module is used to construct a hospital management graph structure based on the connection weight matrix between the indicators, and to calculate the management importance score of each node using an improved PageRank algorithm. The dynamic clustering module is used to generate node embedding vectors through graph embedding algorithm based on preset management target configuration parameters and management importance scores, and to classify and label nodes using adaptive clustering analysis, outputting a set of node classification labels. The comprehensive management assessment module is used to combine the weight matrix between the indicators, the management importance score, and the set of node classification labels to calculate the comprehensive management assessment score by category. The adversarial training output module is used to input the index probability distribution model, comprehensive management evaluation score, node classification label set and external intervention strategy parameters into a pre-configured generative adversarial network, and output a dynamic management scenario prediction report and optimization strategy scheme after adversarial training.

10. A computer-readable storage medium, characterized in that: The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 8.

Images