Multi-level attribution and recommendation method and system for medical management decision-making

Abstract

The present invention provides a multi-level attribution and recommendation method and system for medical management decisions, which relates to the technical field of medical management decision-making, including: through a multi-level medical data knowledge graph and a decision knowledge base, using a temporal graph convolutional network to extract the evolution characteristics of graph nodes, constructing a decision influence propagation graph through a cross-layer interactive attention network, and calculating the contribution of each decision node to the medical effect, generating attribution analysis results, determining a decision influence hierarchy system based on the attribution results, combining a multi-objective optimization network, a hierarchical reinforcement learning network and a Monte Carlo search tree network, generating and evaluating candidate decision paths, and combining a fuzzy comprehensive evaluation network to obtain the optimal decision recommendation plan. In summary, the present invention can realize accurate attribution analysis and multi-objective optimization of management decisions, effectively improve the efficiency of hospital resource allocation, reduce operating costs, provide scientific decision support for hospital management, and significantly enhance the hospital's refined management level and overall operational efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of medical management decision-making, and in particular to a multi-level attribution and recommendation method and system for medical management decision-making. Background Art

[0002] Medical management decisions have always been a core issue in the healthcare field. The goal is to effectively utilize medical resources and control operating costs while ensuring medical quality. Traditional medical management decisions rely mainly on experience and intuition, lacking data-driven and scientific decision-making support methods.

[0003] With the development of medical informatization, a large amount of clinical data, operational data, and resource allocation data has been accumulated, making data-driven medical management decisions possible. In recent years, the rapid development of artificial intelligence technology, especially machine learning and deep learning, has provided new tools and methods for medical management decision-making. However, existing medical management decision-making methods still have problems such as difficulty in conducting effective attribution analysis, inability to fully utilize data, and lack of support for multi-objective optimization.

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the Invention

[0005] The embodiments of the present invention provide a multi-level attribution and recommendation method and system for medical management decision-making, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a multi-level attribution and recommendation method for medical management decisions, comprising:

[0007] Acquire patient clinical data, medical institution operation data, and medical resource allocation data and combine them to obtain multidimensional medical data, extract temporal features and spatial features through the self-attention mechanism and generate a spatiotemporal representation vector, perform noise identification and anomaly detection on the spatiotemporal representation vector through a bidirectional gated recurrent unit, perform data repair through an autoencoder network based on the detection results, generate standardized data, and use a multi-head attention network to extract the correlation strength between different data elements to construct an initial relationship graph, add the initial relationship graph to the hierarchical graph structure generation network, construct a dynamic time sequence graph through the diagnosis and treatment process layer to record the diagnosis and treatment evolution trajectory and map it to the multidimensional tensor graph of the medical resource layer to establish resource allocation connections, add the resource allocation connections to the causal reasoning graph of the management decision layer to extract historical decision rules and corresponding constraints and effect evaluation indicators, generate a decision knowledge base, and establish inter-layer information channels based on the initial relationship graph to obtain a multi-level medical data knowledge graph;

[0008] The multi-level medical data knowledge graph and the decision knowledge base are added to the temporal graph convolutional network, the evolution characteristics of the graph nodes in different time windows are extracted through multi-scale convolution kernels, the historical decision rules are integrated to generate dynamic node representations, the dynamic node representations are added to the cross-layer interactive attention network, the direct influence intensity and indirect influence path between nodes are calculated in combination with the decision constraints, and a decision influence propagation graph is constructed, the decision influence propagation graph is added to the causal reasoning network, the contribution of each decision node in the decision influence propagation graph to the medical effect is calculated through back propagation based on the effect evaluation index, and the synergy effect and time lag effect between different nodes are calculated at the same time, the attribution analysis results are generated and added to the adaptive clustering network, similar influence patterns are determined and the decision nodes are grouped to obtain a decision influence hierarchy system;

[0009] The decision influence hierarchy system and the decision knowledge base are added to the multi-objective optimization network, and a decision optimization model is constructed based on the historical decision rules with medical quality, resource efficiency and operating costs as optimization objectives. The historical decision data and decision constraints are added to the hierarchical reinforcement learning network, and the long-term benefits of the decision are calculated through the distributed value evaluation network. The candidate decision actions are output in combination with the strategy generation network, and the decision results are predicted through the environment simulation network. A decision state transition model is constructed based on the predicted decision results and added to the Monte Carlo search tree network. Candidate decision paths are generated through iterative search, and the decision paths are subjected to effect prediction and constraint verification to obtain feasible solutions and add them to the fuzzy comprehensive network evaluation network. The optimal decision recommendation solution is obtained by combining the approximate ideal solution sorting method.

[0010] In an optional embodiment,

[0011] Acquire patient clinical data, medical institution operation data, and medical resource allocation data and combine them to obtain multidimensional medical data, extract temporal features and spatial features through the self-attention mechanism and generate a spatiotemporal representation vector, perform noise identification and anomaly detection on the spatiotemporal representation vector through a bidirectional gated recurrent unit, perform data repair through an autoencoder network based on the detection results, generate standardized data, and use a multi-head attention network to extract the correlation strength between different data elements to construct an initial relationship graph, add the initial relationship graph to the hierarchical graph structure generation network, construct a dynamic time sequence graph through the diagnosis and treatment process layer to record the diagnosis and treatment evolution trajectory and map it to the multidimensional tensor graph of the medical resource layer to establish resource allocation connections, add the resource allocation connections to the causal reasoning graph of the management decision layer to extract historical decision rules and corresponding constraints and effect evaluation indicators, generate a decision knowledge base, and establish inter-layer information channels based on the initial relationship graph to obtain a multi-level medical data knowledge graph including:

[0012] Acquire patient clinical data, medical institution operation data, and medical resource allocation data from electronic medical record systems, hospital management systems, and resource management systems, and combine the acquired data to generate multidimensional medical data;

[0013] The self-attention mechanism is used to extract the disease symptoms, test results, imaging features, and medication reaction time series features of the patient's clinical data; the personnel scheduling, equipment usage, and bed turnover space features of the medical institution's operational data; and the material inventory, equipment status, and expert resource distribution features of the medical resource allocation data. The extracted features are then concatenated and linearly transformed to generate a spatiotemporal representation vector.

[0014] Noise identification and anomaly detection are performed on the spatiotemporal representation vector using a forward processing unit and a backward processing unit of a bidirectional gated recurrent unit, data mutation points and data missing locations are marked using a forget gate, an input gate, and an output gate, the data mutation points and data missing locations are input into an autoencoder network for data repair, the input data is compressed into a low-dimensional feature space using a dimensionality reduction encoder to correct outliers, and the corrected features are restored to the original dimension using a dimensionality increase decoder to generate standardized data;

[0015] A multi-head attention network is used to extract the association strength between different data elements in the standardized data. The association strength of clinical data is extracted through the symptom diagnosis association head, the diagnosis and treatment association head, and the treatment effect association head. The association strength of operational data is extracted through the personnel and equipment association head and the equipment efficiency association head. The association strength of deployment data is extracted through the resource demand association head and the resource effect association head, and an initial relationship graph is constructed.

[0016] The initial relationship graph is added to the hierarchical graph structure generation network, a dynamic time sequence graph is constructed in the diagnosis and treatment process layer to record the diagnosis and treatment evolution trajectory, the diagnosis and treatment behavior is set as a node, the diagnosis and treatment time sequence relationship is set as an edge, the dynamic time sequence graph is mapped to the multidimensional tensor graph of the medical resource layer to establish a resource allocation connection, the medical resources are set as nodes, and the resource coordination relationship is set as an edge; the resource allocation connection is added to the causal reasoning graph of the management decision layer to extract historical decision rules and corresponding constraints and effect evaluation indicators, the decision behavior is set as a node, and the decision causal relationship is set as an edge to generate a decision knowledge base;

[0017] Based on the initial relationship graph, the graph attention network is used to calculate the correlation strength between nodes at different levels to establish an inter-layer information channel, and the diagnosis and treatment demand information is transferred from the diagnosis and treatment process layer to the medical resource layer to trigger resource allocation. The resource allocation results are transferred from the medical resource layer to the management decision layer for decision optimization, and a multi-level medical data knowledge graph is obtained.

[0018] In an optional embodiment,

[0019] The forward processing unit and the backward processing unit of the bidirectional gated recurrent unit are used to perform noise identification and anomaly detection on the spatiotemporal representation vector, and the forget gate, the input gate, and the output gate are used to mark the data mutation points and data missing locations.

[0020] The forget gate of the bidirectional gated recurrent unit is improved into a three-level discrimination structure. The first-level discrimination result is obtained by calculating the deviation value between the data point and the mean value in the window based on a fixed time window and comparing it with the initial deviation threshold. The second-level discrimination result is obtained by calculating the change rate based on adjacent data points and comparing it with the mutation threshold. The third-level discrimination result is obtained by calculating the continuity feature based on the data sequence and comparing it with the noise frequency threshold. When the first-level discrimination result, the second-level discrimination result and the third-level discrimination result all exceed the corresponding threshold, the forget gate outputs the noise point labeling information.

[0021] The input gate of the bidirectional gated recurrent unit is improved to a dynamic time window structure, the rate of change of the data sequence is obtained, the time window size is determined based on the rate of change, a larger time window is used for relatively stable data intervals, and a smaller time window is used for rapidly changing data intervals, the first-order difference value and the second-order difference value of the data within the time window are calculated, and the data mutation point is identified based on the first-order difference value and the second-order difference value, and the mutation point mark information is output;

[0022] The output gate of the bidirectional gated recurrent unit is improved into an integrity check structure, a data sampling interval standard is set, data points exceeding the data sampling interval standard are marked as potential missing positions, a segmented scan is performed on the data sequence to identify zero value positions and invalid value positions, and the zero value positions and the invalid value positions are marked as confirmed missing positions;

[0023] Scanning from the starting position of the data sequence by the forward processing unit, identifying noise points through the forget gate, identifying data mutation points through the input gate, and identifying data missing positions through the output gate in sequence, thereby obtaining a forward anomaly detection result;

[0024] The backward processing unit starts scanning backward from the end of the data sequence to verify the forward anomaly detection result, and supplements and identifies the missing position caused by data transmission delay to obtain the backward anomaly detection result. The forward anomaly detection result and the backward anomaly detection result are merged and processed. If there are multiple anomaly marks at the same position, the final anomaly type is determined based on the significance of the anomaly features, and a detection result including anomaly position information, anomaly type information and anomaly degree information is generated.

[0025] In an optional embodiment,

[0026] The multi-level medical data knowledge graph and the decision knowledge base are added to the temporal graph convolutional network, and the evolution characteristics of the graph nodes in different time windows are extracted through multi-scale convolution kernels. The historical decision rules are integrated to generate dynamic node representations, and the dynamic node representations are added to the cross-layer interactive attention network. The direct influence intensity and indirect influence path between nodes are calculated in combination with the decision constraints and a decision influence propagation graph is constructed. The decision influence propagation graph is added to the causal reasoning network, and the contribution of each decision node in the decision influence propagation graph to the medical effect is calculated through back propagation based on the effect evaluation index. At the same time, the synergy and time lag effects between different nodes are calculated, and the attribution analysis results are generated and added to the adaptive clustering network. Similar influence patterns are determined and the decision nodes are grouped. The decision influence hierarchy system obtained includes:

[0027] Importing a multi-level medical data knowledge graph and a decision knowledge base into a time-series graph convolutional network, setting short-term time windows, medium-term time windows, and long-term time windows in the time-series graph convolutional network, configuring an independent convolution kernel group for each type of time window, extracting immediate medical indicator features and drug response features through the convolution kernel group of the short-term time window, extracting treatment adjustment features and indicator change features through the convolution kernel group of the medium-term time window, and extracting disease outcome features and quality of life features through the convolution kernel group of the long-term time window;

[0028] The features extracted by the temporal graph convolutional network are enhanced using historical rules in the decision knowledge base, and the timeliness of the rules, the applicable conditions of the rules, and the constraints on the execution of the rules are mapped to the feature space to construct dynamic node representations of treatment plan information, administration method information, and combination drug information;

[0029] Input the dynamic node representation into the cross-layer interactive attention network, establish a medical rule constraint matrix, a timing constraint matrix and a resource constraint matrix, calculate the influence strength of the direct connection between nodes to obtain the direct influence feature, track the multi-hop path between nodes to obtain the indirect influence feature, construct a decision influence propagation graph based on the direct influence feature and the indirect influence feature, calculate the edge weight between nodes to represent the influence strength, determine the edge direction to represent the influence propagation direction, and measure the path length to represent the influence transmission complexity;

[0030] The decision-making influence propagation graph is input into a causal reasoning network. Starting from the effect evaluation index, the contribution of the decision node to the medical effect is calculated in reverse along the influence propagation path. The degree of effect enhancement or weakening when multiple decision nodes act together is analyzed to obtain the synergistic effect coefficient. The onset time, peak time, and duration from the execution of the decision to the manifestation of the effect are calculated to obtain the time lag effect characteristics.

[0031] The contribution degree, the synergy effect coefficient and the time lag effect characteristics are input into an adaptive clustering network, a dynamic clustering threshold is calculated according to the data distribution, the influence patterns of the decision nodes are clustered based on the dynamic clustering threshold, a decision influence hierarchy system is constructed according to the results of the influence pattern clustering, the decision nodes are divided into a primary intervention layer, a secondary intervention layer and an auxiliary intervention layer, and within each layer, the decision nodes are grouped according to the similarity of the influence patterns to obtain a hierarchical structure.

[0032] In an optional embodiment,

[0033] The decision-making influence propagation graph is input into the causal reasoning network. Starting from the effect evaluation index, the contribution of the decision node to the medical effect is calculated backward along the influence propagation path. The degree of effect enhancement or weakening when multiple decision nodes act together is analyzed to obtain the synergistic effect coefficient, including:

[0034] Convert the decision-making influence propagation graph into an adjacency matrix and construct a node feature matrix. The node feature matrix integrates node attribute information, temporal features, and topological features. Sparse processing is performed on the adjacency matrix to remove connections below the influence threshold. The node feature matrix is normalized to eliminate dimensional differences in feature dimensions.

[0035] Inputting the adjacency matrix and the node feature matrix into a multi-layer causal feature extraction network, extracting local causal features through a first layer sliding window convolution operation, extracting mid-range causal features through a second layer stacked convolution layer with expanded receptive field range, and extracting global causal features through a third layer global pooling operation;

[0036] The treatment effect index, safety index, and economic index are constructed as an objective function, the gradient value of the objective function with respect to the local causal feature, the mid-range causal feature, and the global causal feature is calculated, and an attention mechanism is introduced to weight the gradient information of different paths;

[0037] Normalize the gradient value passed to the decision node to obtain the initial contribution. According to the interval from the decision-making time to the effect generation time, introduce a time series attenuation factor to perform time series weighting on the initial contribution. Use sliding average to build a contribution smoothing mechanism.

[0038] constructing a decision combination feature vector from feature information of multiple decision nodes, inputting the decision combination feature vector into a synergistic effect predictor to obtain an expected synergistic effect coefficient, and updating the synergistic effect predictor based on a deviation between the expected synergistic effect coefficient and an actual effect;

[0039] The synergistic effect coefficients are stored in a synergistic effect knowledge base according to the effect intensity, and a decision type index and an applicable condition index are established. The synergistic effect knowledge base is dynamically updated according to the reliability and representativeness of the newly added data.

[0040] In an optional embodiment,

[0041] The decision influence hierarchy system and the decision knowledge base are added to the multi-objective optimization network. Based on the historical decision rules, a decision optimization model is constructed with medical quality, resource efficiency and operating cost as optimization objectives. The historical decision data and decision constraints are added to the hierarchical reinforcement learning network. The long-term benefits of the decision are calculated through the distributed value evaluation network. The candidate decision actions are output in combination with the strategy generation network. The decision results are predicted through the environment simulation network. A decision state transition model is constructed based on the predicted decision results and added to the Monte Carlo search tree network. Candidate decision paths are generated through iterative search. The decision paths are subjected to effect prediction and constraint verification to obtain feasible solutions and add them to the fuzzy comprehensive network evaluation network. The optimal decision recommendation solution is obtained by combining the approximate ideal solution sorting method, including:

[0042] Assigning the highest weight value to the node information of the primary intervention layer in the decision-making influence hierarchy system, assigning a medium weight value to the node information of the secondary intervention layer, and assigning a basic weight value to the node information of the auxiliary intervention layer. The weighted decision-making influence hierarchy system is introduced into the multi-objective optimization network. Clinical pathway specification information, expert experience rule information, and treatment plan template information are extracted from the decision-making knowledge base to construct historical decision rules. The historical decision rules are used to establish a decision optimization model with medical quality, resource efficiency, and operating cost as optimization goals.

[0043] Performing temporal alignment on historical decision data to extract decision state features, decision action features, and decision result features, and importing the decision state features, decision action features, and decision result features along with medical specification constraint information, resource capacity constraint information, and temporal logic constraint information into a hierarchical reinforcement learning network;

[0044] Construct a dual-channel value assessment architecture, using the immediate value assessment channel to calculate the immediate benefits of treatment effect information, resource consumption information, and medical risk information, and using the long-term value assessment channel to predict the cumulative benefits of treatment cycle information, recovery process information, and prognosis information. The immediate benefits and cumulative benefits are input into the distributed value assessment network to calculate the long-term benefits of decision-making;

[0045] A multi-module strategy generator is used to generate a set of treatment plan adjustment actions, an examination arrangement action set, and a follow-up plan action set. An adaptive sampling mechanism is used to dynamically adjust the exploration range parameters. A hierarchical action filter is used to sequentially perform medical safety checks, resource feasibility checks, and timing rationality checks, and output candidate decision actions.

[0046] A state transition predictor is constructed using a deep learning framework to predict the next state feature distribution. A medical environment stochastic simulator is introduced to simulate individual patient differences, treatment response fluctuations, and complication risk information. A multi-scale prediction engine is used to predict short-term state change information, mid-term complication risk information, and long-term prognosis trend information.

[0047] A hierarchical search tree structure is constructed to store decision state information, action information, and evaluation value information. A decision state transition model is established using a value-prioritized node expansion strategy and an adaptive search depth control mechanism. The decision state transition model is imported into a Monte Carlo search tree network. Candidate decision paths are generated through iterative search in the Monte Carlo search tree network. The decision path execution process is simulated using a multi-dimensional simulation verification platform. A hierarchical constraint verification mechanism is used to verify medical standard compliance information, resource scheduling feasibility information, and execution timing rationality information, thereby screening out feasible solutions.

[0048] Establish a hierarchical evaluation index system for medical effect information, resource efficiency information, and economic benefit information, and use a dynamic weight adjustment mechanism to conduct a fuzzy comprehensive evaluation of the feasible plans;

[0049] A positive and negative ideal solution generator is constructed based on pre-acquired historical optimal solutions and expert knowledge. A multi-objective distance calculation module is used to calculate the medical effect distance, resource efficiency distance and economic cost distance. The candidate solutions are ranked according to the medical effect distance, resource efficiency distance and economic cost distance, and the optimal decision recommendation solution is output.

[0050] In an optional embodiment,

[0051] Constructing a hierarchical search tree structure to store decision state information, action information, and evaluation value information, establishing a decision state transition model using a value-priority-based node expansion strategy and an adaptive search depth control mechanism, importing the decision state transition model into a Monte Carlo search tree network, and generating candidate decision paths through iterative search in the Monte Carlo search tree network includes:

[0052] Constructing a hierarchical search tree structure, collecting patient medical examination data, symptom data, medication data, and complication data to construct decision state information, constructing a multidimensional state matrix based on the physiological indicator dimension, symptom and sign dimension, medication status dimension, and complication dimension of the decision state information, standardizing and converting the data in the multidimensional state matrix, and organizing the standardized decision state information into a feature vector;

[0053] Collecting medical intervention plan data to construct action information, dividing the action information into medication adjustment actions, examination arrangement actions, and rehabilitation training actions, using an action coding mechanism to input medication number information, dosage change information, administration route information, and administration frequency information for the medication adjustment actions, input examination item number information, examination time information, and examination frequency information for the examination arrangement actions, and input training item number information, training intensity information, and training cycle information for the rehabilitation training actions;

[0054] Collecting medical effect data, resource consumption data, and economic cost data to construct evaluation value information, extracting symptom control information, complication occurrence information, and quality of life information to construct the medical effect data, counting hospitalization days information, examination frequency information, and drug usage information to construct the resource consumption data, calculating medical expense information, equipment depreciation information, and labor cost information to construct the economic cost data, and performing standardization on the evaluation value information;

[0055] Using the decision state information to construct a search tree root node, using the action information to construct a search tree intermediate node, and using the evaluation value information to construct a search tree leaf node;

[0056] Calculate the access frequency score, historical revenue score, and exploration potential score based on the value-first node expansion strategy, and construct a node score value by weighted summing the access frequency score, historical revenue score, and exploration potential score;

[0057] Adopting an adaptive search depth control mechanism, the system calculates the disease complexity index by counting information on the number of complications and the severity of symptoms, calculates the treatment difficulty index by analyzing information on previous treatment effects and drug responsiveness, and calculates the environmental uncertainty index by evaluating information on compliance levels and the degree of perfection of support systems. The system dynamically adjusts the search depth based on the disease complexity index, the treatment difficulty index, and the environmental uncertainty index.

[0058] Establish a decision-making state transition model, extract features from historical decision data to obtain state transition features, analyze the state transition features to obtain state evolution rules, import individual patient characteristics for personalized adjustment, and output the target state expected value, state distribution variance, and transition probability confidence interval;

[0059] Importing the decision state transition model into a Monte Carlo search tree network, selecting expansion nodes based on the node score values, calling the decision state transition model to generate child nodes, evaluating the long-term benefits of the decision path, updating node statistics, and simultaneously exploring different decision branches through multi-threaded parallel computing;

[0060] Candidate decision paths are generated by iteratively searching in the Monte Carlo search tree network, a directed decision graph is constructed to store the decision state information, the action information and the evaluation value information, the decision benefit value of each path in the directed decision graph and the similarity value between any two paths in the directed decision graph are calculated, and the directed decision graph is optimized according to a preset benefit threshold and similarity threshold to obtain the candidate decision path.

[0061] A second aspect of an embodiment of the present invention provides a multi-level attribution and recommendation system for medical management decisions, including:

[0062] The first unit is used to obtain patient clinical data, medical institution operation data and medical resource allocation data and combine them to obtain multi-dimensional medical data, extract temporal features and spatial features through the self-attention mechanism and generate a spatiotemporal representation vector, perform noise identification and anomaly detection on the spatiotemporal representation vector through a bidirectional gated recurrent unit, perform data repair through an autoencoder network according to the detection results, generate standardized data, and use a multi-head attention network to extract the correlation strength between different data elements to construct an initial relationship graph, add the initial relationship graph to the hierarchical graph structure generation network, construct a dynamic time sequence graph through the diagnosis and treatment process layer to record the diagnosis and treatment evolution trajectory and map it to the multi-dimensional tensor graph of the medical resource layer to establish resource allocation connections, add the resource allocation connections to the causal reasoning graph of the management decision layer to extract historical decision rules and corresponding constraints and effect evaluation indicators, generate a decision knowledge base, and establish inter-layer information channels based on the initial relationship graph to obtain a multi-level medical data knowledge graph;

[0063] The second unit is used to add the multi-level medical data knowledge graph and the decision knowledge base to the temporal graph convolution network, extract the evolution characteristics of the graph nodes in different time windows through multi-scale convolution kernels, fuse historical decision rules to generate dynamic node representations, add the dynamic node representations to the cross-layer interactive attention network, calculate the direct influence intensity and indirect influence path between nodes in combination with decision constraints and construct a decision influence propagation graph, add the decision influence propagation graph to the causal reasoning network, calculate the contribution of each decision node in the decision influence propagation graph to the medical effect based on the effect evaluation index through back propagation, and simultaneously calculate the synergy effect and time lag effect between different nodes, generate attribution analysis results and add them to the adaptive clustering network, determine similar influence patterns and group the decision nodes to obtain a decision influence hierarchy system;

[0064] The third unit is used to add the decision influence hierarchy system and the decision knowledge base to the multi-objective optimization network, construct a decision optimization model based on the historical decision rules with medical quality, resource efficiency and operating cost as optimization objectives, add historical decision data and decision constraints to the hierarchical reinforcement learning network, calculate the long-term benefits of the decision through the distributed value evaluation network, output candidate decision actions in combination with the strategy generation network, predict the decision results through the environment simulation network, construct a decision state transition model based on the predicted decision results and add it to the Monte Carlo search tree network, generate candidate decision paths through iterative search, perform effect prediction and constraint verification on the decision path to obtain a feasible solution and add it to the fuzzy comprehensive network evaluation network, and combine the approximate ideal solution sorting method to solve and obtain the optimal decision recommendation solution.

[0065] According to a third aspect of the embodiments of the present invention,

[0066] An electronic device is provided, comprising:

[0067] processor;

[0068] a memory for storing processor-executable instructions;

[0069] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0070] According to a fourth aspect of the embodiments of the present invention,

[0071] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0072] In the present invention, through multi-level attribution analysis, the contribution, synergy and lag effects of different decision nodes to medical effects are revealed, avoiding the limitations of traditional methods that only focus on single factors or linear relationships. It can evaluate the impact of decisions more comprehensively and accurately, thereby improving the scientificity and effectiveness of decisions. By constructing a decision influence propagation diagram and a multi-objective optimization model, key decision nodes and resource bottlenecks can be effectively identified, and resource allocation and operation processes can be optimized in a targeted manner, thereby improving the efficiency of medical resource utilization and reducing operating costs. Based on historical data and knowledge base, combined with reinforcement learning and fuzzy comprehensive evaluation methods, the optimal decision recommendation plan can be generated, and personalized adjustments can be made according to the specific situation of the patient, realizing intelligent medical decision support, and ultimately improving medical quality and patient satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 A flowchart of a multi-level attribution and recommendation method for medical management decision-making according to an embodiment of the present invention;

[0074] Figure 2 Schematic diagram of the structure of the multi-level attribution and recommendation system for medical management decision-making according to an embodiment of the present invention. DETAILED DESCRIPTION

[0075] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0076] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0077] Figure 1 FIG. 1 is a flow chart of a multi-level attribution and recommendation method for medical management decision-making according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0078] S1. Acquire patient clinical data, medical institution operation data, and medical resource allocation data and combine them to obtain multidimensional medical data, extract temporal features and spatial features through the self-attention mechanism and generate a spatiotemporal representation vector, perform noise identification and anomaly detection on the spatiotemporal representation vector through a bidirectional gated recurrent unit, perform data repair through an autoencoder network based on the detection results, generate standardized data, and use a multi-head attention network to extract the correlation strength between different data elements to construct an initial relationship graph, add the initial relationship graph to the hierarchical graph structure generation network, construct a dynamic time sequence graph through the diagnosis and treatment process layer to record the diagnosis and treatment evolution trajectory and map it to the multidimensional tensor graph of the medical resource layer to establish resource allocation connections, add the resource allocation connections to the causal reasoning graph of the management decision layer to extract historical decision rules and corresponding constraints and effect evaluation indicators, generate a decision knowledge base, and establish inter-layer information channels based on the initial relationship graph to obtain a multi-level medical data knowledge graph;

[0079] The multi-head attention network is a neural network structure that uses multiple groups of attention mechanisms to process different information channels in parallel, so as to more efficiently capture the global dependencies in sequence or graph data. The association strength is an importance indicator that measures the correlation between nodes or features, and is used to quantify the degree of mutual influence between different elements. The dynamic timing graph is a graph structure that combines time changes and is used to model dynamic relational data with time dependence. The diagnosis and treatment evolution trajectory is an evolutionary pattern in the medical field that records the patient's diagnosis and treatment process over time, and is used to analyze changes in the disease and treatment effects. The multidimensional tensor graph is a form that represents multidimensional data as a graph structure, and models and analyzes complex relationships through tensor operations.

[0080] In an optional embodiment,

[0081] Acquire patient clinical data, medical institution operation data, and medical resource allocation data and combine them to obtain multidimensional medical data, extract temporal features and spatial features through the self-attention mechanism and generate a spatiotemporal representation vector, perform noise identification and anomaly detection on the spatiotemporal representation vector through a bidirectional gated recurrent unit, perform data repair through an autoencoder network based on the detection results, generate standardized data, and use a multi-head attention network to extract the correlation strength between different data elements to construct an initial relationship graph, add the initial relationship graph to the hierarchical graph structure generation network, construct a dynamic time sequence graph through the diagnosis and treatment process layer to record the diagnosis and treatment evolution trajectory and map it to the multidimensional tensor graph of the medical resource layer to establish resource allocation connections, add the resource allocation connections to the causal reasoning graph of the management decision layer to extract historical decision rules and corresponding constraints and effect evaluation indicators, generate a decision knowledge base, and establish inter-layer information channels based on the initial relationship graph to obtain a multi-level medical data knowledge graph including:

[0082] Acquire patient clinical data, medical institution operation data, and medical resource allocation data from electronic medical record systems, hospital management systems, and resource management systems, and combine the acquired data to generate multidimensional medical data;

[0083] The self-attention mechanism is used to extract the disease symptoms, test results, imaging features, and medication reaction time series features of the patient's clinical data; the personnel scheduling, equipment usage, and bed turnover space features of the medical institution's operational data; and the material inventory, equipment status, and expert resource distribution features of the medical resource allocation data. The extracted features are then concatenated and linearly transformed to generate a spatiotemporal representation vector.

[0084] Noise identification and anomaly detection are performed on the spatiotemporal representation vector using a forward processing unit and a backward processing unit of a bidirectional gated recurrent unit, data mutation points and data missing locations are marked using a forget gate, an input gate, and an output gate, the data mutation points and data missing locations are input into an autoencoder network for data repair, the input data is compressed into a low-dimensional feature space using a dimensionality reduction encoder to correct outliers, and the corrected features are restored to the original dimension using a dimensionality increase decoder to generate standardized data;

[0085] A multi-head attention network is used to extract the association strength between different data elements in the standardized data. The association strength of clinical data is extracted through the symptom diagnosis association head, the diagnosis and treatment association head, and the treatment effect association head. The association strength of operational data is extracted through the personnel and equipment association head and the equipment efficiency association head. The association strength of deployment data is extracted through the resource demand association head and the resource effect association head, and an initial relationship graph is constructed.

[0086] The initial relationship graph is added to the hierarchical graph structure generation network, a dynamic time sequence graph is constructed in the diagnosis and treatment process layer to record the diagnosis and treatment evolution trajectory, the diagnosis and treatment behavior is set as a node, the diagnosis and treatment time sequence relationship is set as an edge, the dynamic time sequence graph is mapped to the multidimensional tensor graph of the medical resource layer to establish a resource allocation connection, the medical resources are set as nodes, and the resource coordination relationship is set as an edge; the resource allocation connection is added to the causal reasoning graph of the management decision layer to extract historical decision rules and corresponding constraints and effect evaluation indicators, the decision behavior is set as a node, and the decision causal relationship is set as an edge to generate a decision knowledge base;

[0087] Based on the initial relationship graph, the graph attention network is used to calculate the correlation strength between nodes at different levels to establish an inter-layer information channel, and the diagnosis and treatment demand information is transferred from the diagnosis and treatment process layer to the medical resource layer to trigger resource allocation. The resource allocation results are transferred from the medical resource layer to the management decision layer for decision optimization, and a multi-level medical data knowledge graph is obtained.

[0088] The dimension-raising decoder is a model module used to reconstruct original data from a low-dimensional space or generate high-dimensional feature representations. The causal reasoning graph is a graph structure built based on causal relationships and is used to analyze the causal relationship and reasoning process between variables.

[0089] Obtain patient clinical data from electronic medical record systems, including disease symptoms, test results, imaging features, and medication reactions. Obtain operational data about medical institutions from hospital management systems, such as personnel scheduling, equipment usage, and bed turnover. Obtain medical resource allocation data from resource management systems, such as material inventory, equipment status, and expert resource distribution. Combine these three types of data to form multidimensional medical data.

[0090] For example, clinical data for patient Zhang San is obtained from the electronic medical record system, including symptoms: cough, fever; test results: elevated white blood cell count; imaging features: lung shadows; and medication reaction: mild dizziness. Data is obtained from the hospital management system: Respiratory physician Li Si is on duty, CT equipment utilization is 80%, and there are 5 beds remaining. Data is obtained from the resource management system: antibiotics are in stock, ventilators are available, and respiratory specialist Wang Wu is on standby.

[0091] Spatiotemporal feature extraction and representation. A self-attention mechanism is used to extract temporal features of patient clinical data, such as the temporal trends of disease symptoms and the dynamic changes in test results. Spatial features are extracted from medical institution operational data, such as resource utilization by department and the spatial distribution of equipment. Distributional features of medical resource allocation data are extracted, such as the inventory locations of different types of supplies and the geographic locations of expert resources. The extracted temporal and spatial features are concatenated and linearly transformed to generate a spatiotemporal representation vector.

[0092] For example, the duration and severity of patient Zhang San's cough symptoms are extracted as temporal features, and the specific location of lung shadows is extracted as spatial features. Respiratory physician Li Si's department and working hours are extracted as spatial features, as are the usage time and location of CT equipment. The storage location of antibiotic inventory is extracted as a spatial feature, as is the department of respiratory specialist Wang Wu. These features are concatenated and converted into a fixed-length vector, the spatiotemporal representation vector.

[0093] Next, the data is cleaned and standardized. A bidirectional gated recurrent unit is used to identify noise and detect anomalies in the generated spatiotemporal representation vector. The forward and backward processing units process the time series data, respectively, to identify mutation points and missing locations in the data. For example, a sudden rise or fall in a patient's body temperature, or missing test data for a certain period of time. The identified abnormal data points are input into the autoencoder network for data repair. The autoencoder network compresses the input data into a low-dimensional feature space using a dimensionality reduction encoder, removing noise and outliers. The dimensionality increase decoder then restores the corrected features to their original dimensions, generating standardized data.

[0094] For example, if patient Zhang's temperature data shows an abnormal increase at a certain point in time, the bidirectional gated recurrent unit will identify this abnormal point. This abnormal point is input into the autoencoder network, and through compression and restoration operations, the abnormal temperature value is corrected to a value within a reasonable range.

[0095] To construct the initial relationship graph, a multi-head attention network is used to extract the strength of associations between different data elements in the standardized data. For example, the "symptom-diagnosis association head" extracts the strength of association between a patient's cough and a pneumonia diagnosis; the "diagnosis-treatment association head" extracts the strength of association between a pneumonia diagnosis and antibiotic treatment; and the "treatment-effect association head" extracts the strength of association between antibiotic treatment and the patient's temperature returning to normal. Similarly, the strength of associations between different elements in the operational and allocation data is extracted. These strengths of association are connected with edges, with different data elements serving as nodes, to construct the initial relationship graph.

[0096] For example, based on patient Zhang San's cough, fever, lung shadow imaging features, and elevated white blood cell count test results, the strength of the association between these clinical data and the pneumonia diagnosis is extracted. The strength of the association between the pneumonia diagnosis and antibiotic treatment and ventilator-assisted therapy is also extracted. The strength of the association between antibiotic treatment and the patient's temperature returning to normal and coughing being alleviated is also extracted. These data elements and the strength of the association are constructed into an initial relationship graph.

[0097] The constructed initial relationship graph is added to the hierarchical graph structure generation network. At the diagnosis and treatment process layer, a dynamic time-series graph is constructed to record the evolution of diagnosis and treatment. Treatment behaviors are set as nodes, such as "diagnosed with pneumonia" and "treated with antibiotics." Time-series relationships are set as edges, such as "diagnosed with pneumonia" occurs before "treated with antibiotics." The dynamic time-series graph at the diagnosis and treatment process layer is mapped to the multidimensional tensor graph at the medical resource layer to establish resource allocation relationships. Medical resources are set as nodes, such as "antibiotics," "ventilators," and "respiratory physicians." Resource coordination relationships are set as edges, such as "antibiotics" and "ventilators" are used together to treat pneumonia. Resource allocation relationships are added to the causal reasoning graph at the management decision layer to extract historical decision rules, corresponding constraints, and effectiveness evaluation indicators. Decision behaviors are set as nodes, such as "increase respiratory physician scheduling" and "increase antibiotic inventory." Decision causal relationships are set as edges, such as "increasing respiratory physician scheduling" can "shorten patient wait times." Finally, a decision knowledge base is generated, and the correlation strength between nodes at different levels is calculated based on the initial relationship graph and graph attention network, an inter-layer information channel is established, and the transmission of diagnosis and treatment demand information between different levels is realized, and finally a multi-level medical data knowledge graph is obtained.

[0098] For example, the diagnosis and treatment process of patient Zhang San is constructed into a dynamic time series graph, with nodes including "cough and fever symptoms", "blood test", "lung imaging", "diagnosis of pneumonia", "antibiotic treatment", "ventilator-assisted treatment", etc., and edges represent the order in which these diagnosis and treatment behaviors occur. These diagnosis and treatment behaviors are linked to resources such as "antibiotics", "ventilators", and "respiratory doctors" in the medical resource layer. For example, "antibiotic treatment" requires the consumption of "antibiotics" resources, and "ventilator-assisted treatment" requires the use of "ventilator" resources, and the diagnosis and treatment process requires the participation of "respiratory doctors". Resource allocation links are added to the management decision-making layer. For example, based on historical data, if the "number of pneumonia patients increases", it is necessary to "increase the scheduling of respiratory doctors" and "increase the inventory of antibiotics".

[0099] In this embodiment, the quality of medical data is improved through noise identification, anomaly detection and data repair, providing a reliable data foundation for subsequent analysis and decision-making. By constructing a multi-level knowledge graph, we can better understand the relationship between the diagnosis and treatment process, resource allocation and decision-making, thereby optimizing the diagnosis and treatment process and improving diagnosis and treatment efficiency. By extracting historical decision-making rules and effect evaluation indicators, we can assist management decisions and improve the scientific nature and effectiveness of decisions.

[0100] In an optional embodiment,

[0101] The forward processing unit and the backward processing unit of the bidirectional gated recurrent unit are used to perform noise identification and anomaly detection on the spatiotemporal representation vector, and the forget gate, the input gate, and the output gate are used to mark the data mutation points and data missing locations.

[0102] The forget gate of the bidirectional gated recurrent unit is improved into a three-level discrimination structure. The first-level discrimination result is obtained by calculating the deviation value between the data point and the mean value in the window based on a fixed time window and comparing it with the initial deviation threshold. The second-level discrimination result is obtained by calculating the change rate based on adjacent data points and comparing it with the mutation threshold. The third-level discrimination result is obtained by calculating the continuity feature based on the data sequence and comparing it with the noise frequency threshold. When the first-level discrimination result, the second-level discrimination result and the third-level discrimination result all exceed the corresponding threshold, the forget gate outputs the noise point labeling information.

[0103] The input gate of the bidirectional gated recurrent unit is improved to a dynamic time window structure, the rate of change of the data sequence is obtained, the time window size is determined based on the rate of change, a larger time window is used for relatively stable data intervals, and a smaller time window is used for rapidly changing data intervals, the first-order difference value and the second-order difference value of the data within the time window are calculated, and the data mutation point is identified based on the first-order difference value and the second-order difference value, and the mutation point mark information is output;

[0104] The output gate of the bidirectional gated recurrent unit is improved into an integrity check structure, a data sampling interval standard is set, data points exceeding the data sampling interval standard are marked as potential missing positions, a segmented scan is performed on the data sequence to identify zero value positions and invalid value positions, and the zero value positions and the invalid value positions are marked as confirmed missing positions;

[0105] Scanning from the starting position of the data sequence by the forward processing unit, identifying noise points through the forget gate, identifying data mutation points through the input gate, and identifying data missing positions through the output gate in sequence, thereby obtaining a forward anomaly detection result;

[0106] The backward processing unit starts scanning backward from the end of the data sequence to verify the forward anomaly detection result, and supplements and identifies the missing position caused by data transmission delay to obtain the backward anomaly detection result. The forward anomaly detection result and the backward anomaly detection result are merged and processed. If there are multiple anomaly marks at the same position, the final anomaly type is determined based on the significance of the anomaly features, and a detection result including anomaly position information, anomaly type information and anomaly degree information is generated.

[0107] The three-level discriminant structure is a multi-level discriminant model that improves the accuracy of classification or detection by gradually refining the discriminant process. The zero-value position refers to the position where the value is zero in the data or feature matrix, which is usually used to represent sparsity or as a special constraint condition. The abnormal feature significance is an indicator for evaluating the importance of abnormal features in the overall data, and is used to highlight the contribution of abnormal features to model decision-making.

[0108] Preprocess the data series, such as removing duplicate values and processing missing values, to ensure the quality of the input data.

[0109] The forget gate of the bidirectional gated recurrent unit is improved into a three-level discriminant structure. The first-level discriminant is based on a fixed time window, for example, the window size is set to 10. The deviation of each data point from the mean within the window is calculated. An initial deviation threshold is set, for example, 0.5. If the deviation of a data point from the mean is greater than the threshold, the next level of discrimination is performed. Otherwise, the data point is considered normal. Example: Suppose the data in the time window is [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] and the current data point is 12. The mean is 5.5 and the deviation is 6.5, which are greater than the threshold of 0.5. The second-level discriminant calculates the rate of change based on adjacent data points. A mutation threshold is set, for example, 2. If the rate of change is greater than the threshold, the next level of discrimination is performed. Otherwise, the data point is considered normal. Example: Suppose the previous data point is 10 and the current data point is 12. The rate of change is 0.2, which is less than the threshold of 2. The third-level discriminant calculates continuity features based on the data sequence, such as variance. A noise frequency threshold is set, for example, 0.1. If the continuity feature is greater than the threshold, the data point is considered a noise point, and the forget gate outputs the noise point label information. Otherwise, the data point is considered normal. For example, assuming that the variance of the data sequence is 0.15, which is greater than the threshold of 0.1, the data point is considered a noise point.

[0110] The input gate of the bidirectional gated recurrent unit is improved to a dynamic time window structure to obtain the rate of change of the data sequence. The time window size is determined based on the rate of change. A larger time window is used for relatively stable data intervals, for example, the window size is set to 20. A smaller time window is used for rapidly changing data intervals, for example, the window size is set to 5, and the first-order difference and second-order difference values of the data within the time window are calculated. Data mutation points are identified based on the first-order difference and second-order difference values. For example, if both the first-order difference and the second-order difference exceed the preset threshold, the point is considered to be a mutation point, and the input gate outputs the mutation point label information. For example, assuming that the data in the time window is [1, 2, 4, 7, 11], the first-order difference value is [1, 2, 3, 4], and the second-order difference value is [1, 1, 1]. If the preset thresholds are 0.5 and 0.2 respectively, the point is considered to be a mutation point.

[0111] The output gate of a bidirectional gated recurrent unit is modified into an integrity check structure. A data sampling interval is set, such as 1 second. Data points that exceed the data sampling interval are marked as potential missing locations. The data sequence is segmented and scanned to identify zero-value locations and invalid value locations, such as null values or non-numeric values. Zero-value locations and invalid value locations are marked as confirmed missing locations, and the output gate outputs the missing location marking information. For example, if the data sequence is [1, 2, null, 4, 5], then null values are marked as confirmed missing locations.

[0112] The forward processing unit starts scanning from the starting position of the data sequence, identifies noise points through the forget gate, identifies data mutation points through the input gate, and identifies data missing positions through the output gate, and obtains the forward anomaly detection results.

[0113] The backward processing unit starts scanning backward from the end of the data sequence, verifies the forward anomaly detection result, and supplements the identification of missing positions caused by data transmission delays to obtain backward anomaly detection results. For example, if the backward scan finds that a data point is missing, but the forward scan does not detect it, it is considered that the missing is caused by data transmission delay.

[0114] The forward anomaly detection results are combined with the backward anomaly detection results. If multiple anomaly markers are present at the same location, the final anomaly type is determined based on the significance of the anomaly features. For example, if the same location is marked as both a noise point and a mutation point, the final anomaly type is determined based on features such as the size or duration of the anomaly value, generating a detection result that includes anomaly location information, anomaly type information, and anomaly severity information.

[0115] In this embodiment, through multi-level discrimination and bidirectional scanning, various types of anomalies can be identified more accurately, reducing false alarms and missed alarms. The dynamic time window and integrity check structure can improve the algorithm's adaptability to different data patterns and noise. The detection results contain information such as the location, type, and degree of the anomaly, which helps users better understand and handle abnormal situations.

[0116] S2. Add the multi-level medical data knowledge graph and the decision knowledge base to the temporal graph convolutional network, extract the evolution characteristics of the graph nodes in different time windows through multi-scale convolution kernels, fuse historical decision rules to generate dynamic node representations, add the dynamic node representations to the cross-layer interactive attention network, calculate the direct influence intensity and indirect influence path between nodes in combination with decision constraints and construct a decision influence propagation graph, add the decision influence propagation graph to the causal reasoning network, calculate the contribution of each decision node in the decision influence propagation graph to the medical effect based on the effect evaluation index through back propagation, and simultaneously calculate the synergy effect and time lag effect between different nodes, generate attribution analysis results and add them to the adaptive clustering network, determine similar influence patterns and group the decision nodes to obtain a decision influence hierarchy system;

[0117] The evolutionary characteristics are characteristic changes that occur over time or as the system state changes, and are used to analyze trends or patterns in dynamic systems. The historical decision rules are rules summarized based on past decision-making experiences, and are used to guide current or future decision-making processes. The decision constraints are restrictions that must be met in the decision-making process to ensure the rationality and feasibility of the decision. The indirect impact path refers to the impact link transmitted through intermediary variables or factors, and is used to analyze the indirect relationship between variables in complex systems. The decision impact propagation diagram is a graph structure that describes the propagation path of the impact of the decision on each part of the system, and is used to intuitively display the scope and intensity of the impact. The synergistic effect is the combined effect produced by the interaction of multiple factors or decisions, which is usually greater than the simple superposition of each single factor. The time lag effect refers to the phenomenon that the impact of decisions or events in the system takes a certain amount of time to appear, and is often used to analyze delayed behavior in dynamic systems. The similar impact pattern refers to an impact pattern with similar structural or behavioral characteristics, and is used to identify and classify regular relationships in the system.

[0118] In an optional embodiment,

[0119] The multi-level medical data knowledge graph and the decision knowledge base are added to the temporal graph convolutional network, and the evolution characteristics of the graph nodes in different time windows are extracted through multi-scale convolution kernels. The historical decision rules are integrated to generate dynamic node representations, and the dynamic node representations are added to the cross-layer interactive attention network. The direct influence intensity and indirect influence path between nodes are calculated in combination with the decision constraints and a decision influence propagation graph is constructed. The decision influence propagation graph is added to the causal reasoning network, and the contribution of each decision node in the decision influence propagation graph to the medical effect is calculated through back propagation based on the effect evaluation index. At the same time, the synergy and time lag effects between different nodes are calculated, and the attribution analysis results are generated and added to the adaptive clustering network. Similar influence patterns are determined and the decision nodes are grouped. The decision influence hierarchy system obtained includes:

[0120] Importing a multi-level medical data knowledge graph and a decision knowledge base into a time-series graph convolutional network, setting short-term time windows, medium-term time windows, and long-term time windows in the time-series graph convolutional network, configuring an independent convolution kernel group for each type of time window, extracting immediate medical indicator features and drug response features through the convolution kernel group of the short-term time window, extracting treatment adjustment features and indicator change features through the convolution kernel group of the medium-term time window, and extracting disease outcome features and quality of life features through the convolution kernel group of the long-term time window;

[0121] The features extracted by the temporal graph convolutional network are enhanced using historical rules in the decision knowledge base, and the timeliness of the rules, the applicable conditions of the rules, and the constraints on the execution of the rules are mapped to the feature space to construct dynamic node representations of treatment plan information, administration method information, and combination drug information;

[0122] Input the dynamic node representation into the cross-layer interactive attention network, establish a medical rule constraint matrix, a timing constraint matrix and a resource constraint matrix, calculate the influence strength of the direct connection between nodes to obtain the direct influence feature, track the multi-hop path between nodes to obtain the indirect influence feature, construct a decision influence propagation graph based on the direct influence feature and the indirect influence feature, calculate the edge weight between nodes to represent the influence strength, determine the edge direction to represent the influence propagation direction, and measure the path length to represent the influence transmission complexity;

[0123] The decision-making influence propagation graph is input into a causal reasoning network. Starting from the effect evaluation index, the contribution of the decision node to the medical effect is calculated in reverse along the influence propagation path. The degree of effect enhancement or weakening when multiple decision nodes act together is analyzed to obtain the synergistic effect coefficient. The onset time, peak time, and duration from the execution of the decision to the manifestation of the effect are calculated to obtain the time lag effect characteristics.

[0124] The contribution degree, the synergy effect coefficient and the time lag effect characteristics are input into an adaptive clustering network, a dynamic clustering threshold is calculated according to the data distribution, the influence patterns of the decision nodes are clustered based on the dynamic clustering threshold, a decision influence hierarchy system is constructed according to the results of the influence pattern clustering, the decision nodes are divided into a primary intervention layer, a secondary intervention layer and an auxiliary intervention layer, and within each layer, the decision nodes are grouped according to the similarity of the influence patterns to obtain a hierarchical structure.

[0125] The convolution kernel group is a group of convolution kernels used in a convolutional neural network to extract multi-scale or multi-directional feature information. The medical indicator feature is a specific indicator or feature data related to medical care, such as heart rate, blood pressure, etc., which is used for diagnosis and treatment analysis. The drug response feature is the feature or response data exhibited by the patient after taking the drug, which is used for personalized treatment optimization. The constraint matrix is a matrix form used to represent the constraint conditions in decision-making or optimization problems, which is used to simplify mathematical modeling. The dynamic clustering threshold is a clustering algorithm threshold that is dynamically adjusted as the data features or system status changes, which is used to improve classification accuracy.

[0126] Construct a multi-layered medical data knowledge graph and decision-making knowledge base. Multi-layered medical data includes electronic medical records, medical images, laboratory test results, and demographic information. Using technologies such as entity recognition and relationship extraction, the knowledge graph is constructed. Nodes represent medical entities (such as diseases, symptoms, and medications), and edges represent relationships between entities. The decision-making knowledge base includes historical diagnosis and treatment plans, medication rules, and treatment guidelines. Each rule includes rule conditions, execution actions, and effect evaluation.

[0127] For example, a patient's medical data includes a diagnosis of "type 2 diabetes," symptoms of "thirst" and "polyuria," medication "metformin," and a blood glucose level of "10 mmol / L." In the knowledge graph, "type 2 diabetes" and "thirst" are nodes, and "manifested as" is the edge connecting them. The decision knowledge base contains a rule: If the patient is diagnosed with "type 2 diabetes" and their blood glucose level is above "7 mmol / L," then treat them with metformin.

[0128] The knowledge graph and decision knowledge base are imported into the temporal graph convolutional network. Three time windows are set: short-term, medium-term, and long-term, corresponding to different convolution kernel groups. The short-term window focuses on immediate indicators and drug reactions, such as the patient's current blood sugar, blood pressure, heart rate, and the immediate reaction after taking medication. The medium-term window focuses on treatment adjustments and indicator changes, such as blood sugar trends and drug dosage adjustments within a week. The long-term window focuses on disease outcomes and quality of life, such as whether the disease is under control and the patient's quality of life score. The convolution operation extracts the evolutionary characteristics of each node in different time windows.

[0129] For example, a short-term convolution kernel extracts features for a patient's current blood glucose level of "10 mmol / L" and a blood glucose level of "8 mmol / L" one hour after taking medication. A medium-term convolution kernel extracts features for the changing trend of blood glucose levels over the past week. A long-term convolution kernel extracts features for the patient's blood glucose control and quality of life score three months later.

[0130] Leverage historical rules from the decision knowledge base to enhance node features. Map the timeliness, applicability, and execution constraints of rules into the feature space. For example, map the rule "If the patient is diagnosed with 'Type 2 Diabetes' and their blood sugar is above '7mmol / L', then treat with 'Metformin'" to the features of the "Type 2 Diabetes" and "Metformin" nodes. Simultaneously, consider constraints such as the rule's time of use, applicable population, and dosage limits to construct a dynamic node representation.

[0131] Dynamic node representations are fed into a cross-layer interactive attention network. This network considers medical rule constraints, timing constraints, and resource constraints to calculate direct and indirect influences between nodes. Direct influence refers to the strength of the influence of direct connections between nodes, such as the direct impact of metformin on blood sugar. Indirect influence refers to the influence between nodes through multi-hop paths, such as the indirect impact of metformin on blood sugar through its impact on insulin resistance. A decision influence propagation graph is constructed based on direct and indirect influences. Edge weights represent influence strength, edge directions represent influence propagation directions, and path lengths represent the complexity of influence transmission.

[0132] For example, the edge weight from "metformin" to "blood glucose" is large, indicating a strong direct influence. The path from "metformin" to "insulin resistance" and then to "blood glucose" indicates an indirect influence, with a path length of 2.

[0133] The decision-making impact propagation diagram is fed into a causal inference network. Starting from effect evaluation indicators (e.g., blood sugar control and complication rate), the contribution of each decision node to the medical outcome is calculated in reverse. Synergy and lag effects between nodes are also calculated. Synergy refers to the degree to which the effect is enhanced or weakened when multiple decision nodes work together. Lag effects refer to the time characteristics from decision execution to the manifestation of the effect, including onset time, peak time, and duration.

[0134] For example, the contribution of metformin to blood sugar control and the synergistic effect of metformin and lifestyle intervention are calculated. The onset time, peak time, and duration of metformin's blood sugar lowering effect are also calculated.

[0135] Contribution, synergy, and time lag effects are input into an adaptive clustering network. The clustering threshold is dynamically adjusted based on data distribution, and decision nodes are clustered according to their influence patterns. Based on the clustering results, a decision influence hierarchy is constructed, dividing nodes into primary, secondary, and auxiliary intervention layers. Within each layer, nodes are grouped based on the similarity of their influence patterns.

[0136] For example, nodes that directly affect blood sugar, such as "metformin" and "insulin", are divided into the primary intervention layer, nodes that indirectly affect blood sugar, such as "lifestyle intervention" and "patient education", are divided into the secondary intervention layer, and nodes that assist decision-making, such as "genetic testing", are divided into the auxiliary intervention layer.

[0137] In this embodiment, by analyzing the influence levels of decision nodes, personalized treatment plans can be formulated for different patients to improve treatment effects. This can help doctors quickly understand the potential impacts of different treatment plans and assist doctors in making more scientific decisions. By analyzing the synergistic effects and time lag effects between nodes, new treatment strategies and drug combinations can be discovered, thereby promoting the advancement of medical knowledge.

[0138] In an optional embodiment,

[0139] The decision-making influence propagation graph is input into the causal reasoning network. Starting from the effect evaluation index, the contribution of the decision node to the medical effect is calculated backward along the influence propagation path. The degree of effect enhancement or weakening when multiple decision nodes act together is analyzed to obtain the synergistic effect coefficient, including:

[0140] Convert the decision-making influence propagation graph into an adjacency matrix and construct a node feature matrix. The node feature matrix integrates node attribute information, temporal features, and topological features. Sparse processing is performed on the adjacency matrix to remove connections below the influence threshold. The node feature matrix is normalized to eliminate dimensional differences in feature dimensions.

[0141] Inputting the adjacency matrix and the node feature matrix into a multi-layer causal feature extraction network, extracting local causal features through a first layer sliding window convolution operation, extracting mid-range causal features through a second layer stacked convolution layer with expanded receptive field range, and extracting global causal features through a third layer global pooling operation;

[0142] The treatment effect index, safety index, and economic index are constructed as an objective function, the gradient value of the objective function with respect to the local causal feature, the mid-range causal feature, and the global causal feature is calculated, and an attention mechanism is introduced to weight the gradient information of different paths;

[0143] Normalize the gradient value passed to the decision node to obtain the initial contribution. According to the interval from the decision-making time to the effect generation time, introduce a time series attenuation factor to perform time series weighting on the initial contribution. Use sliding average to build a contribution smoothing mechanism.

[0144] constructing a decision combination feature vector from feature information of multiple decision nodes, inputting the decision combination feature vector into a synergistic effect predictor to obtain an expected synergistic effect coefficient, and updating the synergistic effect predictor based on a deviation between the expected synergistic effect coefficient and an actual effect;

[0145] The synergistic effect coefficients are stored in a synergistic effect knowledge base according to the effect intensity, and a decision type index and an applicable condition index are established. The synergistic effect knowledge base is dynamically updated according to the reliability and representativeness of the newly added data.

[0146] The time series attenuation factor is a factor used to reduce the weight of the impact of past data and is often used to emphasize recent trends in time series data analysis. The effect intensity refers to the relative size or significance of the impact of a certain factor or decision on the system or target variable, and is used to evaluate the effectiveness and scope of influence of the results.

[0147] Construct a decision-influence propagation graph. This graph uses nodes to represent medical decisions or events, and edges to represent the causal relationships between decisions. For example, a node might be "use medication A," "undergo surgery B," or "develop symptom C," while an edge represents the impact of the previous node on the next. The edge weight indicates the strength of the impact. For example, for a patient with hypertension and diabetes, the doctor makes three decisions: D1 - start taking antihypertensive medication A, D2 - start taking antidiabetic medication B, and D3 - recommend lifestyle changes (increase exercise and eat a healthy diet). These decisions and the subsequent changes in health indicators constitute the decision-influence propagation graph.

[0148] The decision influence propagation graph is converted into an adjacency matrix and a node feature matrix. The adjacency matrix uses numerical values to represent the connection relationship between nodes. For example, if node i affects node j, the (i, j) position of the matrix is the weight of the edge, otherwise it is 0. The node feature matrix contains the attribute information, temporal features and topological features of each node. Attribute information can be the type and intensity of the decision; temporal features can be the time when the decision occurs; topological features can be the degree and centrality of the node in the graph. For example, if D1 has a slight impact on D2, the value of the corresponding position in the adjacency matrix can be set to 0.2. The node feature matrix records the specific information of each decision. For example, the attribute of D1 can be "antihypertensive drug A", the temporal feature is "January 1, 2024", and the topological feature is the number of connections in the graph.

[0149] Preprocess the adjacency matrix and node feature matrix. Sparsify the adjacency matrix by removing connections below a set threshold to reduce the impact of noise. Normalize the node feature matrix to eliminate differences in the scale of different feature dimensions. For example, scale all eigenvalues to between 0 and 1. For example, if the influence threshold is set to 0.1, connections with a value less than 0.1 will be removed. All node eigenvalues, such as time, are scaled to the range of 0 to 1.

[0150] The processed adjacency matrix and node feature matrix are input into a multi-layer causal feature extraction network. The network consists of multiple convolutional layers and pooling layers to extract causal features at different scales. The first layer of sliding window convolution operation extracts local causal features, such as the relationship between a decision and its directly adjacent decisions. The second layer of stacked convolution layers with expanded receptive field range extracts medium-range causal features, such as the relationship between a decision and its indirectly adjacent decisions. The third layer of global pooling operation extracts global causal features, such as the overall pattern of all decisions. The network extracts local, medium-range and global causal features respectively. For example, the local features of D1 and D2 can be the direct influence relationship between them, the medium-range features of D1, D2 and D3 can be the indirect influence relationship between them, and the global features can be the overall pattern of all decisions and the relationship between them.

[0151] The treatment effect index, safety index and economic index are constructed as the objective function. For example, the blood pressure reduction value, adverse reaction incidence rate and treatment cost can be used as components of the objective function respectively. The gradient value of the objective function for the extracted local, medium-range and global causal features is calculated to indicate the influence of each feature on the objective function. The attention mechanism is introduced to weight the gradient information of different paths, for example, giving higher weights to paths with greater influence. For example, if the blood pressure reduction value is the main goal, the network will calculate the influence of each causal feature on the blood pressure reduction value and give higher weights to paths with greater influence.

[0152] The gradient values passed to the decision node are normalized to obtain an initial contribution, representing each decision's contribution to the objective function. A time-series decay factor is introduced based on the interval between the decision's occurrence and the time its effect occurs, weighting the initial contribution to account for the impact of time. A sliding average is used to smooth the contribution and reduce fluctuations. For example, if D1 occurred a long time ago, its contribution will decay over time.

[0153] The feature information of multiple decision nodes is used to construct a decision combination feature vector. This vector is input into the synergy effect predictor to obtain the expected synergy effect coefficient, which represents the degree to which the effect of multiple decisions is enhanced or weakened when combined. The synergy effect predictor is updated based on the deviation between the expected synergy effect coefficient and the actual effect, improving prediction accuracy. For example, the combined feature vector of D1 and D2 can include their attribute, time, and topological characteristics. The predictor predicts the synergistic effect when these two decisions are used together.

[0154] Synergy effect coefficients are stored in a synergy effect knowledge base based on the intensity of the effect. A decision type index and an applicable condition index are established to facilitate query and use. The synergy effect knowledge base is dynamically updated based on the reliability and representativeness of the newly added data to maintain the timeliness of the knowledge base. For example, the knowledge base can record that the combined use of D1 and D2 has a "strong synergistic" effect, which is applicable to the case of "comorbid hypertension and diabetes."

[0155] In this embodiment, accurate quantification of the contribution of each medical decision to the treatment effect helps to identify key decisions, optimize treatment plans, reveal the synergistic effects of multiple medical decisions when they work together, avoid decision conflicts, and improve treatment effects. The constructed synergistic effect knowledge base can provide decision support for clinicians and assist doctors in formulating more effective personalized treatment plans.

[0156] S3. Add the decision influence hierarchy system and the decision knowledge base to the multi-objective optimization network, construct a decision optimization model based on the historical decision rules with medical quality, resource efficiency and operating cost as optimization targets, add historical decision data and decision constraints to the hierarchical reinforcement learning network, calculate the long-term benefits of the decision through the distributed value evaluation network, output candidate decision actions in combination with the strategy generation network, predict the decision results through the environment simulation network, construct a decision state transition model based on the predicted decision results and add it to the Monte Carlo search tree network, generate candidate decision paths through iterative search, perform effect prediction and constraint verification on the decision paths to obtain feasible solutions and add them to the fuzzy comprehensive network evaluation network, and solve the optimal decision recommendation solution in combination with the approximate ideal solution sorting method.

[0157] The hierarchical reinforcement learning network is a hierarchical reinforcement learning model that improves learning efficiency and flexibility of task execution by decomposing tasks into multiple subtasks and learning layer by layer. The distributed value evaluation network is a network architecture that evaluates value functions in a distributed environment. It is suitable for processing large-scale data or complex decision-making problems. The long-term benefit of the decision refers to the expected total benefit accumulated from the current state during the decision-making process, which is usually used to evaluate the pros and cons of the strategy. The Monte Carlo search tree network is a network model that combines Monte Carlo simulation and tree search methods, which is used to optimize decision paths and is commonly used in games and planning problems. The fuzzy comprehensive network evaluation network is an evaluation model based on fuzzy logic and multi-dimensional information fusion, which is used to comprehensively evaluate complex systems or multi-attribute problems. The approximate ideal solution ranking method is a multi-criteria decision-making method that ranks solutions by calculating the distance between the solution and the ideal solution to select the optimal solution.

[0158] In an optional embodiment,

[0159] The decision influence hierarchy system and the decision knowledge base are added to the multi-objective optimization network. Based on the historical decision rules, a decision optimization model is constructed with medical quality, resource efficiency and operating cost as optimization objectives. The historical decision data and decision constraints are added to the hierarchical reinforcement learning network. The long-term benefits of the decision are calculated through the distributed value evaluation network. The candidate decision actions are output in combination with the strategy generation network. The decision results are predicted through the environment simulation network. A decision state transition model is constructed based on the predicted decision results and added to the Monte Carlo search tree network. Candidate decision paths are generated through iterative search. The decision paths are subjected to effect prediction and constraint verification to obtain feasible solutions and add them to the fuzzy comprehensive network evaluation network. The optimal decision recommendation solution is obtained by combining the approximate ideal solution sorting method, including:

[0160] Assigning the highest weight value to the node information of the primary intervention layer in the decision-making influence hierarchy system, assigning a medium weight value to the node information of the secondary intervention layer, and assigning a basic weight value to the node information of the auxiliary intervention layer. The weighted decision-making influence hierarchy system is introduced into the multi-objective optimization network. Clinical pathway specification information, expert experience rule information, and treatment plan template information are extracted from the decision-making knowledge base to construct historical decision rules. The historical decision rules are used to establish a decision optimization model with medical quality, resource efficiency, and operating cost as optimization goals.

[0161] Performing temporal alignment on historical decision data to extract decision state features, decision action features, and decision result features, and importing the decision state features, decision action features, and decision result features along with medical specification constraint information, resource capacity constraint information, and temporal logic constraint information into a hierarchical reinforcement learning network;

[0162] Construct a dual-channel value assessment architecture, using the immediate value assessment channel to calculate the immediate benefits of treatment effect information, resource consumption information, and medical risk information, and using the long-term value assessment channel to predict the cumulative benefits of treatment cycle information, recovery process information, and prognosis information. The immediate benefits and cumulative benefits are input into the distributed value assessment network to calculate the long-term benefits of decision-making;

[0163] A multi-module strategy generator is used to generate a set of treatment plan adjustment actions, an examination arrangement action set, and a follow-up plan action set. An adaptive sampling mechanism is used to dynamically adjust the exploration range parameters. A hierarchical action filter is used to sequentially perform medical safety checks, resource feasibility checks, and timing rationality checks, and output candidate decision actions.

[0164] A state transition predictor is constructed using a deep learning framework to predict the next state feature distribution. A medical environment stochastic simulator is introduced to simulate individual patient differences, treatment response fluctuations, and complication risk information. A multi-scale prediction engine is used to predict short-term state change information, mid-term complication risk information, and long-term prognosis trend information.

[0165] A hierarchical search tree structure is constructed to store decision state information, action information, and evaluation value information. A decision state transition model is established using a value-prioritized node expansion strategy and an adaptive search depth control mechanism. The decision state transition model is imported into a Monte Carlo search tree network. Candidate decision paths are generated through iterative search in the Monte Carlo search tree network. The decision path execution process is simulated using a multi-dimensional simulation verification platform. A hierarchical constraint verification mechanism is used to verify medical standard compliance information, resource scheduling feasibility information, and execution timing rationality information, thereby screening out feasible solutions.

[0166] Establish a hierarchical evaluation index system for medical effect information, resource efficiency information, and economic benefit information, and use a dynamic weight adjustment mechanism to conduct a fuzzy comprehensive evaluation of the feasible plans;

[0167] A positive and negative ideal solution generator is constructed based on pre-acquired historical optimal solutions and expert knowledge. A multi-objective distance calculation module is used to calculate the medical effect distance, resource efficiency distance and economic cost distance. The candidate solutions are ranked according to the medical effect distance, resource efficiency distance and economic cost distance, and the optimal decision recommendation solution is output.

[0168] The dual-channel value evaluation architecture is a network structure consisting of two parallel channels. It evaluates the value of strategies from multiple perspectives to improve the accuracy and robustness of predictions. The value-priority-based node expansion strategy is a strategy that prioritizes the expansion of high-value nodes in the search algorithm to improve search efficiency and decision quality. The adaptive search depth control mechanism is an algorithm that dynamically adjusts the search depth to optimize the search process based on problem complexity or environmental changes.

[0169] Construct a decision-making influence hierarchy system and a decision-making knowledge base. The decision-making influence hierarchy system is divided into the primary intervention layer, the secondary intervention layer and the auxiliary intervention layer, which correspond to different decision weights. For example, for the treatment plan decision of a certain disease, the selection of the main treatment drug belongs to the primary intervention layer, the selection of the auxiliary treatment drug belongs to the secondary intervention layer, and the arrangement of the rehabilitation plan belongs to the auxiliary intervention layer. The nodes of the primary intervention layer are assigned the highest weight value, such as 0.8; the nodes of the secondary intervention layer are assigned a medium weight value, such as 0.5; and the nodes of the auxiliary intervention layer are assigned a basic weight value, such as 0.2. The decision knowledge base contains clinical pathway specification information, expert experience rule information and treatment plan template information. For example, for the treatment of pneumonia, the knowledge base contains the clinical pathway of pneumonia, the antibiotic use rules recommended by experts, and common treatment plan templates.

[0170] Construct a multi-objective optimization model and a hierarchical reinforcement learning network. The multi-objective optimization model optimizes medical quality, resource efficiency, and operating costs. For example, medical quality can be measured by indicators such as cure rate and complication rate; resource efficiency can be measured by indicators such as average length of stay and bed turnover rate; and operating costs can be measured by indicators such as treatment costs and drug consumption. Historical decision rules guide model construction. For example, historical data analysis can reveal that certain drug combinations can effectively improve cure rates and reduce complication rates. These rules can serve as prior knowledge for the model. The inputs to the hierarchical reinforcement learning network include historical decision data (e.g., patient medical history, treatment plans, treatment outcomes), and decision constraints (e.g., medical standards and resource capacity constraints). Historical decision data must be time-series aligned, for example, by aligning the treatment processes of different patients in chronological order, so that the model can learn from the time series information. Decision state features can include the patient's condition and physiological indicators; decision action features can include the medications used and dosages; and decision outcome features can include treatment efficacy and adverse reactions.

[0171] A value assessment network and a policy generation network are constructed. The value assessment network adopts a dual-channel architecture. The immediate value assessment channel calculates short-term benefits, such as the immediate effect, resource consumption, and risks of a treatment plan; the long-term value assessment channel predicts long-term benefits, such as the patient's long-term recovery and prognosis. For example, for a pneumonia patient, antibiotics can quickly control the infection and bring immediate benefits; however, in the long term, the side effects of antibiotics and the patient's recovery need to be considered. The distributed value assessment network integrates immediate and long-term benefits to calculate the long-term value of the decision. The policy generation network generates candidate decision actions, such as adjusting treatment plans, scheduling examinations, and establishing follow-up plans. The policy generation network adopts a multi-module structure, with each module responsible for generating different types of actions. For example, the treatment adjustment module can adjust drug dosage based on the patient's condition; the examination scheduling module can schedule appropriate examination items based on the patient's condition. An adaptive sampling mechanism and hierarchical action filtering are used to improve search efficiency. For example, the exploration scope is dynamically adjusted based on the value of the current state, prioritizing more promising actions. Unsafe, infeasible, or unreasonable actions are filtered out through medical safety checks, resource feasibility checks, and temporal rationality checks.

[0172] An environmental simulation network and a Monte Carlo search tree network are constructed. The environmental simulation network predicts decision outcomes, such as a patient's response to treatment plans and the probability of complications. The environmental simulation network includes a state transition predictor and a medical environment stochasticity simulator. The state transition predictor predicts the characteristic distribution of the next state. For example, based on the current treatment plan and the patient's condition, it predicts the patient's next stage of condition change. The medical environment stochasticity simulator simulates factors such as individual patient differences, fluctuations in treatment response, and the risk of complications. For example, it simulates differences in sensitivity to the same drug among different patients and the potential complications that may arise during treatment. The Monte Carlo search tree network is used to generate candidate decision paths. The hierarchical search tree structure stores decision states, actions, and evaluation value information. For example, the root node of the tree represents the initial state, and each child node represents a possible action. Each node stores the corresponding state information, action information, and evaluation value information. A value-based node expansion strategy and an adaptive search depth control mechanism are used to improve search efficiency. For example, high-value nodes are expanded first, and the search depth is dynamically adjusted based on search progress. The multi-dimensional simulation verification platform simulates the execution process of the decision path and performs constraint verification, for example, verifying whether the decision path complies with medical standards, resource limitations and time constraints.

[0173] A fuzzy comprehensive evaluation network and an approximate ideal solution ranking method are constructed. The fuzzy comprehensive evaluation network evaluates feasible solutions, for example, by assessing their medical effectiveness, resource efficiency, and economic benefits. A hierarchical evaluation index system is used to comprehensively assess all aspects of the solution. For example, medical effectiveness can consider indicators such as cure rate and complication rate; resource efficiency can consider indicators such as average length of stay and bed turnover rate; and economic benefits can consider indicators such as treatment costs and drug consumption. A dynamic weight adjustment mechanism adjusts the weights of indicators based on different situations. For example, medical effectiveness is given a higher weight for critically ill patients. The approximate ideal solution ranking method is used to select the optimal solution. The positive and negative ideal solution generator generates positive and negative ideal solutions based on historical optimal solutions and expert knowledge. For example, based on historical data and expert experience, the optimal treatment effectiveness, resource efficiency, and economic benefit indicator values are determined as the positive ideal solution, and the worst indicator values are determined as the negative ideal solution. The multi-objective distance calculation module calculates the distance between the candidate solution and the positive and negative ideal solutions, for example, calculating the difference in medical effectiveness, resource efficiency, and economic benefits between the candidate solution and the positive and negative ideal solutions. The candidate solutions are sorted according to the distance, and the solution closest to the positive ideal solution and farthest from the negative ideal solution is selected as the optimal solution.

[0174] In this embodiment, by combining multi-objective optimization, reinforcement learning, and fuzzy comprehensive evaluation methods, a better treatment plan can be generated, the patient's cure rate can be improved, the complication rate can be reduced, and the quality of medical care can be improved. By considering resource constraints, the allocation of medical resources can be optimized, resource utilization efficiency can be improved, and medical costs can be reduced. By simulating individual differences among patients and fluctuations in treatment responses, a more personalized treatment plan can be generated to improve treatment effects.

[0175] In an optional embodiment,

[0176] Constructing a hierarchical search tree structure to store decision state information, action information, and evaluation value information, establishing a decision state transition model using a value-priority-based node expansion strategy and an adaptive search depth control mechanism, importing the decision state transition model into a Monte Carlo search tree network, and generating candidate decision paths through iterative search in the Monte Carlo search tree network includes:

[0177] Constructing a hierarchical search tree structure, collecting patient medical examination data, symptom data, medication data, and complication data to construct decision state information, constructing a multidimensional state matrix based on the physiological indicator dimension, symptom and sign dimension, medication status dimension, and complication dimension of the decision state information, standardizing and converting the data in the multidimensional state matrix, and organizing the standardized decision state information into a feature vector;

[0178] Collecting medical intervention plan data to construct action information, dividing the action information into medication adjustment actions, examination arrangement actions, and rehabilitation training actions, using an action coding mechanism to input medication number information, dosage change information, administration route information, and administration frequency information for the medication adjustment actions, input examination item number information, examination time information, and examination frequency information for the examination arrangement actions, and input training item number information, training intensity information, and training cycle information for the rehabilitation training actions;

[0179] Collecting medical effect data, resource consumption data, and economic cost data to construct evaluation value information, extracting symptom control information, complication occurrence information, and quality of life information to construct the medical effect data, counting hospitalization days information, examination frequency information, and drug usage information to construct the resource consumption data, calculating medical expense information, equipment depreciation information, and labor cost information to construct the economic cost data, and performing standardization on the evaluation value information;

[0180] Using the decision state information to construct a search tree root node, using the action information to construct a search tree intermediate node, and using the evaluation value information to construct a search tree leaf node;

[0181] Calculate the access frequency score, historical revenue score, and exploration potential score based on the value-first node expansion strategy, and construct a node score value by weighted summing the access frequency score, historical revenue score, and exploration potential score;

[0182] Adopting an adaptive search depth control mechanism, the system calculates the disease complexity index by counting information on the number of complications and the severity of symptoms, calculates the treatment difficulty index by analyzing information on previous treatment effects and drug responsiveness, and calculates the environmental uncertainty index by evaluating information on compliance levels and the degree of perfection of support systems. The system dynamically adjusts the search depth based on the disease complexity index, the treatment difficulty index, and the environmental uncertainty index.

[0183] Establish a decision state transition model, extract features from historical decision data to obtain state transition features, analyze the state transition features to obtain state evolution rules, import individual patient characteristics for personalized adjustment, and output the target state expected value, state distribution variance, and transition probability confidence interval;

[0184] Importing the decision state transition model into a Monte Carlo search tree network, selecting expansion nodes based on the node score values, calling the decision state transition model to generate child nodes, evaluating the long-term benefits of the decision path, updating node statistics, and simultaneously exploring different decision branches through multi-threaded parallel computing;

[0185] Candidate decision paths are generated by iteratively searching in the Monte Carlo search tree network, a directed decision graph is constructed to store the decision state information, the action information and the evaluation value information, the decision benefit value of each path in the directed decision graph and the similarity value between any two paths in the directed decision graph are calculated, and the directed decision graph is optimized according to a preset benefit threshold and similarity threshold to obtain the candidate decision path.

[0186] The compliance level is a measure of the degree to which an individual or system follows a certain rule, strategy or instruction, and is used to evaluate the execution effect. The directed decision graph is a structure that stores and represents decision paths or rules in the form of a directed graph, and is used to describe and optimize complex decision-making processes.

[0187] A hierarchical search tree structure is constructed. Medical examination data, such as blood pressure, oxygen saturation, and electrocardiogram results, are collected from hospital databases and electronic medical records. Symptom data, such as headache, fever, and cough, are collected. Medication data, such as drug name, dosage, and route of administration, are collected. Complication data, such as infection and bleeding, are collected. This data is then used to construct a multidimensional state matrix based on four dimensions: physiological indicators, symptoms and signs, medication status, and complications. For example, physiological indicators such as blood pressure and heart rate can be placed in the first dimension, symptoms and signs such as headache and fever can be placed in the second dimension, and so on. The data in the multidimensional state matrix is then normalized, for example, using a min-max normalization method to scale the data to between 0 and 1. Finally, the normalized decision state information is organized into a feature vector, which serves as the root node of the search tree. For example, a feature vector can be represented as [0.8, 0.5, 0.2, 0.1], representing the normalized values of blood pressure, heart rate, headache, and fever, respectively.

[0188] Collect medical intervention plan data to construct action information, and divide the action information into three categories: medication adjustment actions, inspection arrangement actions, and rehabilitation training actions. For medication adjustment actions, record information such as medication number, dosage change, route of administration, and frequency of administration. For example, "Drug A, increase the dose by 10 mg, orally, three times a day." For inspection arrangement actions, record information such as inspection item number, inspection time, and inspection frequency. For example, "blood routine examination, tomorrow morning at 9 am, once a week." For rehabilitation training actions, record information such as training item number, training intensity, and training cycle. For example, "walking training, moderate intensity, three times a week, 30 minutes each time." These action information will be used to construct the intermediate nodes of the search tree.

[0189] Medical effect data, resource consumption data and economic cost data are collected to construct evaluation value information. Medical effect data includes symptom control information, complication occurrence information and quality of life information. For example, symptom scores decrease, complications do not occur, and quality of life improves. Resource consumption data includes information such as the number of hospital days, number of examinations and drug usage. For example, the patient was hospitalized for 7 days, underwent 3 blood routine examinations, and used 5 drugs. Economic cost data includes information such as medical expenses, equipment depreciation and labor costs. For example, the total medical expenses are 10,000 yuan. These evaluation value information are standardized, for example, using the min-max normalization method. These evaluation value information will be used to construct the leaf nodes of the search tree.

[0190] A decision state transition model is established using a value-based node expansion strategy and an adaptive search depth control mechanism. The value-based node expansion strategy comprehensively considers the access frequency score, historical revenue score, and exploration potential score, and takes a weighted sum of these three factors to determine the node score. For example, the more times a node is visited, the higher its access frequency score; the higher the historical revenue of a node, the higher its historical revenue score; and the greater the exploration potential of a node, the higher its exploration potential score. The search depth control mechanism dynamically adjusts the search depth based on the disease complexity index, treatment difficulty index, and environmental uncertainty index. For example, the more complex the disease, the deeper the search depth; the greater the treatment difficulty, the deeper the search depth; and the greater the environmental uncertainty, the deeper the search depth. The decision state transition model analyzes the state transition characteristics of historical decision data and makes personalized adjustments based on individual patient characteristics. It outputs the target state expected value, the state distribution variance, and the confidence interval of the transition probability. For example, based on historical data, if a patient uses drug A, there is an 80% probability that their blood pressure will decrease and a 20% probability that their blood pressure will remain unchanged.

[0191] The decision state transition model is imported into a Monte Carlo search tree network. Extension nodes are selected based on node scores, the decision state transition model is invoked to generate child nodes, the long-term benefits of the decision path are evaluated, node statistics are updated, and different decision branches are explored simultaneously through multi-threaded parallel computing. An iterative search is performed within the Monte Carlo search tree network to generate candidate decision paths, construct a directed decision graph, and optimize the directed decision graph based on preset benefit and similarity thresholds to obtain the final candidate decision paths.

[0192] For example, suppose a patient's initial blood pressure is 180 mmHg and their target blood pressure is 140 mmHg. A Monte Carlo search tree network search finds a candidate decision path: administering Drug A at a dose of 10 mg three times daily. After a period of treatment, the patient's blood pressure drops to 150 mmHg. According to the decision state transition model, continued use of Drug A is expected to further reduce the blood pressure to 140 mmHg.

[0193] In this embodiment, a personalized treatment plan is formulated based on the individual characteristics of the patient, which helps to improve the treatment effect, evaluate the resource consumption and economic costs of different decision paths, help doctors select the optimal treatment plan, help optimize the allocation of medical resources, and automatically generate candidate decision paths, which helps to reduce the workload of doctors and improve decision-making efficiency.

[0194] In an optional embodiment,

[0195] During the implementation of a management system at a tertiary hospital, financial data from 2022 to 2024 was collected. This included revenue data such as outpatient, inpatient, medical insurance, pharmaceutical, and medical equipment revenue, as well as expenditure data such as employee salaries, equipment depreciation, pharmaceutical procurement, and infrastructure maintenance. Operational data such as bed utilization by department, operating room usage, and physician outpatient visits were also collected, along with resource allocation data such as medical staff scheduling, equipment usage, and drug inventory. A self-attention mechanism was used to identify temporal characteristics of the raw data, such as seasonal fluctuations in revenue and expenditure and cyclical characteristics of human resource allocation.

[0196] The system analyzed the collaborative relationships between departments like internal medicine, surgery, operating rooms, and pharmacies, establishing rules for resource sharing and allocation across departments. Using a multi-head attention network, the system extracted the correlation strength between various management decisions and operational indicators. It found that personnel deployment decisions had an impact coefficient of 0.73 on labor costs, and that equipment procurement decisions had an improvement coefficient of 0.65 on diagnosis and treatment efficiency. The system also identified that budget adjustments had an average lag effect of two months on department performance.

[0197] During the decision-making optimization phase, the system built an optimization model with the goals of reducing operating costs, improving resource utilization, and optimizing revenue structure. Using a hierarchical reinforcement learning network, it was found that adjusting nurse scheduling could reduce labor costs by 12%, optimizing equipment utilization could increase equipment utilization by 15%, and improving performance appraisals could improve medical staff efficiency by 8%. Based on this, the system generated a comprehensive optimization plan encompassing staffing, equipment utilization, and budget allocation.

[0198] According to environmental simulation network predictions, the implementation of this optimization plan is expected to reduce the hospital's annual operating costs by approximately RMB 8.5 million, increase average resource utilization by 17%, and improve management efficiency by 23%.

[0199] Figure 2 FIG. 1 is a schematic diagram of the structure of a multi-level attribution and recommendation system for medical management decision making according to an embodiment of the present invention. Figure 2 As shown, the system includes:

[0200] The first unit is used to obtain patient clinical data, medical institution operation data and medical resource allocation data and combine them to obtain multi-dimensional medical data, extract temporal features and spatial features through the self-attention mechanism and generate a spatiotemporal representation vector, perform noise identification and anomaly detection on the spatiotemporal representation vector through a bidirectional gated recurrent unit, perform data repair through an autoencoder network according to the detection results, generate standardized data, and use a multi-head attention network to extract the correlation strength between different data elements to construct an initial relationship graph, add the initial relationship graph to the hierarchical graph structure generation network, construct a dynamic time sequence graph through the diagnosis and treatment process layer to record the diagnosis and treatment evolution trajectory and map it to the multi-dimensional tensor graph of the medical resource layer to establish resource allocation connections, add the resource allocation connections to the causal reasoning graph of the management decision layer to extract historical decision rules and corresponding constraints and effect evaluation indicators, generate a decision knowledge base, and establish inter-layer information channels based on the initial relationship graph to obtain a multi-level medical data knowledge graph;

[0201] The second unit is used to add the multi-level medical data knowledge graph and the decision knowledge base to the temporal graph convolution network, extract the evolution characteristics of the graph nodes in different time windows through multi-scale convolution kernels, fuse historical decision rules to generate dynamic node representations, add the dynamic node representations to the cross-layer interactive attention network, calculate the direct influence intensity and indirect influence path between nodes in combination with decision constraints and construct a decision influence propagation graph, add the decision influence propagation graph to the causal reasoning network, calculate the contribution of each decision node in the decision influence propagation graph to the medical effect based on the effect evaluation index through back propagation, and simultaneously calculate the synergy effect and time lag effect between different nodes, generate attribution analysis results and add them to the adaptive clustering network, determine similar influence patterns and group the decision nodes to obtain a decision influence hierarchy system;

[0202] The third unit is used to add the decision influence hierarchy system and the decision knowledge base to the multi-objective optimization network, construct a decision optimization model based on the historical decision rules with medical quality, resource efficiency and operating cost as optimization objectives, add historical decision data and decision constraints to the hierarchical reinforcement learning network, calculate the long-term benefits of the decision through the distributed value evaluation network, output candidate decision actions in combination with the strategy generation network, predict the decision results through the environment simulation network, construct a decision state transition model based on the predicted decision results and add it to the Monte Carlo search tree network, generate candidate decision paths through iterative search, perform effect prediction and constraint verification on the decision path to obtain a feasible solution and add it to the fuzzy comprehensive network evaluation network, and combine the approximate ideal solution sorting method to solve and obtain the optimal decision recommendation solution.

[0203] According to a third aspect of the embodiments of the present invention,

[0204] An electronic device is provided, comprising:

[0205] processor;

[0206] a memory for storing processor-executable instructions;

[0207] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0208] According to a fourth aspect of the embodiments of the present invention,

[0209] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0210] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0211] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-level attribution and recommendation method for medical management decision making, characterized by: include: Acquire patient clinical data, medical institution operation data, and medical resource allocation data and combine them to obtain multidimensional medical data, extract temporal features and spatial features through the self-attention mechanism and generate a spatiotemporal representation vector, perform noise identification and anomaly detection on the spatiotemporal representation vector through a bidirectional gated recurrent unit, perform data repair through an autoencoder network based on the detection results, generate standardized data, and use a multi-head attention network to extract the correlation strength between different data elements to construct an initial relationship graph, add the initial relationship graph to the hierarchical graph structure generation network, construct a dynamic time sequence graph through the diagnosis and treatment process layer to record the diagnosis and treatment evolution trajectory and map it to the multidimensional tensor graph of the medical resource layer to establish resource allocation connections, add the resource allocation connections to the causal reasoning graph of the management decision layer to extract historical decision rules and corresponding constraints and effect evaluation indicators, generate a decision knowledge base, and establish inter-layer information channels based on the initial relationship graph to obtain a multi-level medical data knowledge graph; The multi-level medical data knowledge graph and the decision knowledge base are added to the temporal graph convolutional network, the evolution characteristics of the graph nodes in different time windows are extracted through multi-scale convolution kernels, and the historical decision rules are integrated to generate dynamic node representations. The dynamic node representations are added to the cross-layer interactive attention network, and the direct influence intensity and indirect influence path between nodes are calculated in combination with the decision constraints to construct a decision influence propagation graph. The decision influence propagation graph is added to the causal reasoning network, and the contribution of each decision node in the decision influence propagation graph to the medical effect is calculated through back propagation based on the effect evaluation index. At the same time, the synergy effect and time lag effect between different nodes are calculated, and the attribution analysis results are generated and added to the adaptive clustering network. Similar influence patterns are determined and the decision nodes are grouped to obtain a decision influence hierarchy system, including: Importing a multi-level medical data knowledge graph and a decision knowledge base into a time-series graph convolutional network, setting short-term time windows, medium-term time windows, and long-term time windows in the time-series graph convolutional network, configuring an independent convolution kernel group for each type of time window, extracting immediate medical indicator features and drug response features through the convolution kernel group of the short-term time window, extracting treatment adjustment features and indicator change features through the convolution kernel group of the medium-term time window, and extracting disease outcome features and quality of life features through the convolution kernel group of the long-term time window; The features extracted by the temporal graph convolutional network are enhanced using historical rules in the decision knowledge base, and the timeliness of the rules, the applicable conditions of the rules, and the constraints on the execution of the rules are mapped to the feature space to construct dynamic node representations of treatment plan information, administration method information, and combination drug information; Input the dynamic node representation into the cross-layer interactive attention network, establish a medical rule constraint matrix, a timing constraint matrix and a resource constraint matrix, calculate the influence strength of the direct connection between nodes to obtain the direct influence feature, track the multi-hop path between nodes to obtain the indirect influence feature, construct a decision influence propagation graph based on the direct influence feature and the indirect influence feature, calculate the edge weight between nodes to represent the influence strength, determine the edge direction to represent the influence propagation direction, and measure the path length to represent the influence transmission complexity; The decision-making influence propagation graph is input into a causal reasoning network. Starting from the effect evaluation index, the contribution of the decision node to the medical effect is calculated in reverse along the influence propagation path. The degree of effect enhancement or weakening when multiple decision nodes act together is analyzed to obtain the synergistic effect coefficient. The onset time, peak time, and duration from the execution of the decision to the manifestation of the effect are calculated to obtain the time lag effect characteristics. Inputting the contribution degree, the synergistic effect coefficient, and the time lag effect feature into an adaptive clustering network, calculating a dynamic clustering threshold according to data distribution, clustering the influence patterns of decision nodes based on the dynamic clustering threshold, constructing a decision influence hierarchy system based on the results of the influence pattern clustering, dividing the decision nodes into a primary intervention layer, a secondary intervention layer, and an auxiliary intervention layer, and within each layer, grouping the decision nodes according to the similarity of the influence patterns to obtain a hierarchical structure; The decision influence hierarchy system and the decision knowledge base are added to the multi-objective optimization network, and a decision optimization model is constructed based on the historical decision rules with medical quality, resource efficiency and operating costs as optimization objectives. The historical decision data and decision constraints are added to the hierarchical reinforcement learning network, and the long-term benefits of the decision are calculated through the distributed value evaluation network. The candidate decision actions are output in combination with the strategy generation network, and the decision results are predicted through the environment simulation network. A decision state transition model is constructed based on the predicted decision results and added to the Monte Carlo search tree network. Candidate decision paths are generated through iterative search, and the decision paths are subjected to effect prediction and constraint verification to obtain feasible solutions and add them to the fuzzy comprehensive network evaluation network. The optimal decision recommendation solution is obtained by combining the approximate ideal solution sorting method.

2. The method according to claim 1, characterized in that Acquire patient clinical data, medical institution operation data, and medical resource allocation data and combine them to obtain multidimensional medical data, extract temporal features and spatial features through the self-attention mechanism and generate a spatiotemporal representation vector, perform noise identification and anomaly detection on the spatiotemporal representation vector through a bidirectional gated recurrent unit, perform data repair through an autoencoder network based on the detection results, generate standardized data, and use a multi-head attention network to extract the correlation strength between different data elements to construct an initial relationship graph, add the initial relationship graph to the hierarchical graph structure generation network, construct a dynamic time sequence graph through the diagnosis and treatment process layer to record the diagnosis and treatment evolution trajectory and map it to the multidimensional tensor graph of the medical resource layer to establish resource allocation connections, add the resource allocation connections to the causal reasoning graph of the management decision layer to extract historical decision rules and corresponding constraints and effect evaluation indicators, generate a decision knowledge base, and establish inter-layer information channels based on the initial relationship graph to obtain a multi-level medical data knowledge graph including: Acquire patient clinical data, medical institution operation data, and medical resource allocation data from electronic medical record systems, hospital management systems, and resource management systems, and combine the acquired data to generate multidimensional medical data; The self-attention mechanism is used to extract the disease symptoms, test results, imaging features, and medication reaction time series features of the patient's clinical data; the personnel scheduling, equipment usage, and bed turnover space features of the medical institution's operational data; and the material inventory, equipment status, and expert resource distribution features of the medical resource allocation data. The extracted features are then concatenated and linearly transformed to generate a spatiotemporal representation vector. Noise identification and anomaly detection are performed on the spatiotemporal representation vector using a forward processing unit and a backward processing unit of a bidirectional gated recurrent unit, data mutation points and data missing locations are marked using a forget gate, an input gate, and an output gate, the data mutation points and data missing locations are input into an autoencoder network for data repair, the input data is compressed into a low-dimensional feature space using a dimensionality reduction encoder to correct outliers, and the corrected features are restored to the original dimension using a dimensionality increase decoder to generate standardized data; A multi-head attention network is used to extract the association strength between different data elements in the standardized data. The association strength of clinical data is extracted through the symptom diagnosis association head, the diagnosis and treatment association head, and the treatment effect association head. The association strength of operational data is extracted through the personnel and equipment association head and the equipment efficiency association head. The association strength of deployment data is extracted through the resource demand association head and the resource effect association head, and an initial relationship graph is constructed. The initial relationship graph is added to the hierarchical graph structure generation network, a dynamic time sequence graph is constructed in the diagnosis and treatment process layer to record the diagnosis and treatment evolution trajectory, the diagnosis and treatment behavior is set as a node, the diagnosis and treatment time sequence relationship is set as an edge, the dynamic time sequence graph is mapped to the multidimensional tensor graph of the medical resource layer to establish a resource allocation connection, the medical resources are set as nodes, and the resource coordination relationship is set as an edge; the resource allocation connection is added to the causal reasoning graph of the management decision layer to extract historical decision rules and corresponding constraints and effect evaluation indicators, the decision behavior is set as a node, and the decision causal relationship is set as an edge to generate a decision knowledge base; Based on the initial relationship graph, the graph attention network is used to calculate the correlation strength between nodes at different levels to establish an inter-layer information channel, and the diagnosis and treatment demand information is transferred from the diagnosis and treatment process layer to the medical resource layer to trigger resource allocation. The resource allocation results are transferred from the medical resource layer to the management decision layer for decision optimization, and a multi-level medical data knowledge graph is obtained.

3. The method according to claim 2, characterized in that The forward processing unit and the backward processing unit of the bidirectional gated recurrent unit are used to perform noise identification and anomaly detection on the spatiotemporal representation vector, and the forget gate, the input gate, and the output gate are used to mark the data mutation points and data missing locations. The forget gate of the bidirectional gated recurrent unit is improved into a three-level discrimination structure. The first-level discrimination result is obtained by calculating the deviation value between the data point and the mean value in the window based on a fixed time window and comparing it with the initial deviation threshold. The second-level discrimination result is obtained by calculating the change rate based on adjacent data points and comparing it with the mutation threshold. The third-level discrimination result is obtained by calculating the continuity feature based on the data sequence and comparing it with the noise frequency threshold. When the first-level discrimination result, the second-level discrimination result and the third-level discrimination result all exceed the corresponding threshold, the forget gate outputs the noise point labeling information. The input gate of the bidirectional gated recurrent unit is improved to a dynamic time window structure, the rate of change of the data sequence is obtained, the time window size is determined based on the rate of change, a larger time window is used for relatively stable data intervals, and a smaller time window is used for rapidly changing data intervals, the first-order difference value and the second-order difference value of the data within the time window are calculated, and the data mutation point is identified based on the first-order difference value and the second-order difference value, and the mutation point mark information is output; The output gate of the bidirectional gated recurrent unit is improved into an integrity check structure, a data sampling interval standard is set, data points exceeding the data sampling interval standard are marked as potential missing positions, a segmented scan is performed on the data sequence to identify zero value positions and invalid value positions, and the zero value positions and the invalid value positions are marked as confirmed missing positions; Scanning from the starting position of the data sequence by the forward processing unit, identifying noise points through the forget gate, identifying data mutation points through the input gate, and identifying data missing positions through the output gate in sequence, thereby obtaining a forward anomaly detection result; The backward processing unit starts scanning backward from the end of the data sequence to verify the forward anomaly detection result, and supplements and identifies the missing position caused by data transmission delay to obtain the backward anomaly detection result. The forward anomaly detection result and the backward anomaly detection result are merged and processed. If there are multiple anomaly marks at the same position, the final anomaly type is determined based on the significance of the anomaly features, and a detection result including anomaly position information, anomaly type information and anomaly degree information is generated.

4. The method according to claim 1, wherein The decision-making influence propagation graph is input into the causal reasoning network. Starting from the effect evaluation index, the contribution of the decision node to the medical effect is calculated backward along the influence propagation path. The degree of effect enhancement or weakening when multiple decision nodes act together is analyzed to obtain the synergistic effect coefficient, including: Convert the decision-making influence propagation graph into an adjacency matrix and construct a node feature matrix. The node feature matrix integrates node attribute information, temporal features, and topological features. Sparse processing is performed on the adjacency matrix to remove connections below the influence threshold. The node feature matrix is normalized to eliminate dimensional differences in feature dimensions. Inputting the adjacency matrix and the node feature matrix into a multi-layer causal feature extraction network, extracting local causal features through a first layer sliding window convolution operation, extracting mid-range causal features through a second layer stacked convolution layer with expanded receptive field range, and extracting global causal features through a third layer global pooling operation; The treatment effect index, safety index, and economic index are constructed as an objective function, the gradient value of the objective function with respect to the local causal feature, the mid-range causal feature, and the global causal feature is calculated, and an attention mechanism is introduced to weight the gradient information of different paths; Normalize the gradient value passed to the decision node to obtain the initial contribution. According to the interval from the decision-making time to the effect generation time, introduce a time series attenuation factor to perform time series weighting on the initial contribution. Use sliding average to build a contribution smoothing mechanism. constructing a decision combination feature vector from feature information of multiple decision nodes, inputting the decision combination feature vector into a synergistic effect predictor to obtain an expected synergistic effect coefficient, and updating the synergistic effect predictor based on a deviation between the expected synergistic effect coefficient and an actual effect; The synergistic effect coefficients are stored in a synergistic effect knowledge base according to the effect intensity, and a decision type index and an applicable condition index are established. The synergistic effect knowledge base is dynamically updated according to the reliability and representativeness of the newly added data.

5. The method according to claim 1, characterized in that The decision influence hierarchy system and the decision knowledge base are added to the multi-objective optimization network. Based on the historical decision rules, a decision optimization model is constructed with medical quality, resource efficiency and operating cost as optimization objectives. The historical decision data and decision constraints are added to the hierarchical reinforcement learning network. The long-term benefits of the decision are calculated through the distributed value evaluation network. The candidate decision actions are output in combination with the strategy generation network. The decision results are predicted through the environment simulation network. A decision state transition model is constructed based on the predicted decision results and added to the Monte Carlo search tree network. Candidate decision paths are generated through iterative search. The decision paths are subjected to effect prediction and constraint verification to obtain feasible solutions and add them to the fuzzy comprehensive network evaluation network. The optimal decision recommendation solution is obtained by combining the approximate ideal solution sorting method, including: Assigning the highest weight value to the node information of the primary intervention layer in the decision-making influence hierarchy system, assigning a medium weight value to the node information of the secondary intervention layer, and assigning a basic weight value to the node information of the auxiliary intervention layer. The weighted decision-making influence hierarchy system is introduced into the multi-objective optimization network. Clinical pathway specification information, expert experience rule information, and treatment plan template information are extracted from the decision-making knowledge base to construct historical decision rules. The historical decision rules are used to establish a decision optimization model with medical quality, resource efficiency, and operating cost as optimization goals. Performing temporal alignment on historical decision data to extract decision state features, decision action features, and decision result features, and importing the decision state features, decision action features, and decision result features along with medical specification constraint information, resource capacity constraint information, and temporal logic constraint information into a hierarchical reinforcement learning network; Construct a dual-channel value assessment architecture, using the immediate value assessment channel to calculate the immediate benefits of treatment effect information, resource consumption information, and medical risk information, and using the long-term value assessment channel to predict the cumulative benefits of treatment cycle information, recovery process information, and prognosis information. The immediate benefits and cumulative benefits are input into the distributed value assessment network to calculate the long-term benefits of decision-making; A multi-module strategy generator is used to generate a set of treatment plan adjustment actions, an examination arrangement action set, and a follow-up plan action set. An adaptive sampling mechanism is used to dynamically adjust the exploration range parameters. A hierarchical action filter is used to sequentially perform medical safety checks, resource feasibility checks, and timing rationality checks, and output candidate decision actions. A state transition predictor is constructed using a deep learning framework to predict the next state feature distribution. A medical environment stochastic simulator is introduced to simulate individual patient differences, treatment response fluctuations, and complication risk information. A multi-scale prediction engine is used to predict short-term state change information, mid-term complication risk information, and long-term prognosis trend information. A hierarchical search tree structure is constructed to store decision state information, action information, and evaluation value information. A decision state transition model is established using a value-prioritized node expansion strategy and an adaptive search depth control mechanism. The decision state transition model is imported into a Monte Carlo search tree network. Candidate decision paths are generated through iterative search in the Monte Carlo search tree network. The decision path execution process is simulated using a multi-dimensional simulation verification platform. A hierarchical constraint verification mechanism is used to verify medical standard compliance information, resource scheduling feasibility information, and execution timing rationality information, thereby screening out feasible solutions. Establish a hierarchical evaluation index system for medical effect information, resource efficiency information, and economic benefit information, and use a dynamic weight adjustment mechanism to conduct a fuzzy comprehensive evaluation of the feasible plans; A positive and negative ideal solution generator is constructed based on pre-acquired historical optimal solutions and expert knowledge. A multi-objective distance calculation module is used to calculate the medical effect distance, resource efficiency distance and economic cost distance. The candidate solutions are ranked according to the medical effect distance, resource efficiency distance and economic cost distance, and the optimal decision recommendation solution is output.

6. The method according to claim 5, characterized in that Constructing a hierarchical search tree structure to store decision state information, action information, and evaluation value information, establishing a decision state transition model using a value-priority-based node expansion strategy and an adaptive search depth control mechanism, importing the decision state transition model into a Monte Carlo search tree network, and generating candidate decision paths through iterative search in the Monte Carlo search tree network includes: Constructing a hierarchical search tree structure, collecting patient medical examination data, symptom data, medication data, and complication data to construct decision state information, constructing a multidimensional state matrix based on the physiological indicator dimension, symptom and sign dimension, medication status dimension, and complication dimension of the decision state information, standardizing and converting the data in the multidimensional state matrix, and organizing the standardized decision state information into a feature vector; Collecting medical intervention plan data to construct action information, dividing the action information into medication adjustment actions, examination arrangement actions, and rehabilitation training actions, using an action coding mechanism to input medication number information, dosage change information, administration route information, and administration frequency information for the medication adjustment actions, input examination item number information, examination time information, and examination frequency information for the examination arrangement actions, and input training item number information, training intensity information, and training cycle information for the rehabilitation training actions; Collecting medical effect data, resource consumption data, and economic cost data to construct evaluation value information, extracting symptom control information, complication occurrence information, and quality of life information to construct the medical effect data, counting hospitalization days information, examination frequency information, and drug usage information to construct the resource consumption data, calculating medical expense information, equipment depreciation information, and labor cost information to construct the economic cost data, and performing standardization on the evaluation value information; Using the decision state information to construct a search tree root node, using the action information to construct a search tree intermediate node, and using the evaluation value information to construct a search tree leaf node; Calculate the access frequency score, historical revenue score, and exploration potential score based on the value-first node expansion strategy, and construct a node score value by weighted summing the access frequency score, historical revenue score, and exploration potential score; Adopting an adaptive search depth control mechanism, the system calculates the disease complexity index by counting information on the number of complications and the severity of symptoms, calculates the treatment difficulty index by analyzing information on previous treatment effects and drug responsiveness, and calculates the environmental uncertainty index by evaluating information on compliance levels and the degree of perfection of support systems. The system dynamically adjusts the search depth based on the disease complexity index, the treatment difficulty index, and the environmental uncertainty index. Establish a decision-making state transition model, extract features from historical decision data to obtain state transition features, analyze the state transition features to obtain state evolution rules, import individual patient characteristics for personalized adjustment, and output the target state expected value, state distribution variance, and transition probability confidence interval; Importing the decision state transition model into a Monte Carlo search tree network, selecting expansion nodes based on the node score values, calling the decision state transition model to generate child nodes, evaluating the long-term benefits of the decision path, updating node statistics, and simultaneously exploring different decision branches through multi-threaded parallel computing; Candidate decision paths are generated by iteratively searching in the Monte Carlo search tree network, a directed decision graph is constructed to store the decision state information, the action information and the evaluation value information, the decision benefit value of each path in the directed decision graph and the similarity value between any two paths in the directed decision graph are calculated, and the directed decision graph is optimized according to a preset benefit threshold and similarity threshold to obtain the candidate decision path.

7. A multi-level attribution and recommendation system for medical management decision making, for implementing the method according to any one of claims 1 to 6, characterized in that: include: The first unit is used to obtain patient clinical data, medical institution operation data and medical resource allocation data and combine them to obtain multi-dimensional medical data, extract temporal features and spatial features through the self-attention mechanism and generate a spatiotemporal representation vector, perform noise identification and anomaly detection on the spatiotemporal representation vector through a bidirectional gated recurrent unit, perform data repair through an autoencoder network according to the detection results, generate standardized data, and use a multi-head attention network to extract the correlation strength between different data elements to construct an initial relationship graph, add the initial relationship graph to the hierarchical graph structure generation network, construct a dynamic time sequence graph through the diagnosis and treatment process layer to record the diagnosis and treatment evolution trajectory and map it to the multi-dimensional tensor graph of the medical resource layer to establish resource allocation connections, add the resource allocation connections to the causal reasoning graph of the management decision layer to extract historical decision rules and corresponding constraints and effect evaluation indicators, generate a decision knowledge base, and establish inter-layer information channels based on the initial relationship graph to obtain a multi-level medical data knowledge graph; The second unit is used to add the multi-level medical data knowledge graph and the decision knowledge base to the temporal graph convolution network, extract the evolution characteristics of the graph nodes in different time windows through multi-scale convolution kernels, fuse historical decision rules to generate dynamic node representations, add the dynamic node representations to the cross-layer interactive attention network, calculate the direct influence intensity and indirect influence path between nodes in combination with decision constraints and construct a decision influence propagation graph, add the decision influence propagation graph to the causal reasoning network, calculate the contribution of each decision node in the decision influence propagation graph to the medical effect based on the effect evaluation index through back propagation, and simultaneously calculate the synergy effect and time lag effect between different nodes, generate attribution analysis results and add them to the adaptive clustering network, determine similar influence patterns and group the decision nodes to obtain a decision influence hierarchy system; The third unit is used to add the decision influence hierarchy system and the decision knowledge base to the multi-objective optimization network, construct a decision optimization model based on the historical decision rules with medical quality, resource efficiency and operating cost as optimization objectives, add historical decision data and decision constraints to the hierarchical reinforcement learning network, calculate the long-term benefits of the decision through the distributed value evaluation network, output candidate decision actions in combination with the strategy generation network, predict the decision results through the environment simulation network, construct a decision state transition model based on the predicted decision results and add it to the Monte Carlo search tree network, generate candidate decision paths through iterative search, perform effect prediction and constraint verification on the decision path to obtain a feasible solution and add it to the fuzzy comprehensive network evaluation network, and combine the approximate ideal solution sorting method to solve and obtain the optimal decision recommendation solution.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images