Gynecological disease diagnosis and treatment method and system fused with electronic medical record

Abstract

The invention provides a gynecological disease diagnosis and treatment method and system fused with an electronic medical record, and relates to the technical field of medical informationization, and the method comprises the steps: collecting the multi-modal diagnosis and treatment data of a patient, and constructing a personalized medical database through a self-encoding network; deploying a plurality of agents based on a dynamic causal network, and generating an initial treatment scheme in combination with deep reinforcement learning; through real-time data acquisition and a multi-objective optimization algorithm, the scheme is dynamically optimized by fusing expert decision features. According to the invention, personalized precise treatment of gynecological disease diagnosis and treatment can be realized, the treatment effect is improved, and the complication risk is reduced.

Description

Technical Field

[0001] The present invention relates to medical informatization technology, and particularly to a diagnosis and treatment method and system for gynecological diseases integrating electronic medical records. Background Art

[0002] Traditional diagnosis and treatment methods for gynecological diseases mainly rely on doctors' clinical experience and professional knowledge to formulate treatment plans by analyzing patients' symptoms, signs, and examination results. The existing diagnosis and treatment technologies for gynecological diseases have the following deficiencies: First, the existing technologies lack the ability to deeply integrate and analyze multi-modal medical data, making it difficult to fully explore the correlation features between different types of data, resulting in an incomplete overall understanding of the patient's condition. Second, traditional diagnosis and treatment methods lack a systematic analysis of the mutual influence between various intervention measures during the treatment process, and cannot accurately evaluate the combined effects of different treatment means, affecting the formulation of treatment plans. Third, the existing technologies have limitations in the dynamic adjustment of treatment plans, making it difficult to quickly optimize the drug administration strategy based on the patient's real-time feedback, and also lacking the ability to accurately predict and evaluate the treatment trajectory. Summary of the Invention

[0003] Embodiments of the present invention provide a diagnosis and treatment method and system for gynecological diseases integrating electronic medical records, which can solve the problems in the existing technologies.

[0004] In the first aspect of the embodiments of the present invention,

[0005] A diagnosis and treatment method for gynecological diseases integrating electronic medical records is provided, including:

[0006] Collect multi-modal diagnosis and treatment data of patients, including time-series physiological index sequences, clinical manifestation feature sets, medical image sets, and test report sets, perform dimensionality reduction and denoising processing through an autoencoder network, construct an inter-modal feature mapping relationship, extract common and individual features of the data, and establish a personalized medical database for the patient;

[0007] Based on the personalized medical database, construct a dynamic causal network including disease development, treatment intervention, and complication prevention and control. The dynamic causal network analyzes the interaction effects between different treatment means through a time-series causal reasoning algorithm; based on the dynamic causal network, deploy a drug treatment agent, a surgical plan agent, and a nursing plan agent, form sub-plans through a multi-level reasoning mechanism by combining professional domain rules and historical case features using a deep reinforcement learning method, perform feasibility analysis on the sub-plans based on confidence, and perform conflict resolution to generate an initial treatment plan;

[0008] Collect the time-series data of the patient's medication dosage, metabolic indicators, clinical symptoms, and adverse reactions in real time. Combine the attention calculation method to identify treatment features, and use bidirectional recursive calculation to predict the treatment trajectory. Based on the patient's tolerance indicators, organ function indicators, and treatment effect indicators, use a multi-objective optimization algorithm to calculate the optimal dosing plan. Analyze the expert decision-making features in historical successful cases, establish a rule base for treatment timing, medication strategies, and complication prevention and control, and form decision criteria. Integrate the treatment trajectory prediction results, the optimal dosing plan, and the decision criteria to dynamically optimize the initial treatment plan and obtain the final treatment plan.

[0009] In an alternative implementation,

[0010] Based on the personalized medical database, constructing a dynamic causal network including disease development, treatment intervention, and complication prevention and control. The steps of analyzing the interactive effects between different treatment means through a time-series causal reasoning algorithm include:

[0011] Based on the personalized medical database, construct a patient state sequence including immediate features, cumulative features, and historical features. Extract the dynamic features of the patient state sequence through a recursive neural encoder, and use a time attention mechanism to fuse multi-scale time-series information to obtain a dynamic feature vector representing the disease development law.

[0012] Based on the dynamic feature vector, construct a time-series causal graph of treatment intervention and disease response. Through residual function parameterized state transition calculation, introduce latent variables to describe uncertainty, and establish a dynamic causal network including feature nodes, action nodes, and latent variable nodes.

[0013] Based on the dynamic causal network, calculate the complication risk prediction value, integrate dynamic features, latent variables, and treatment actions for risk assessment, and use the graph structure to calculate the correlation between complication types to obtain a treatment strategy considering risk constraints.

[0014] Conduct a two-layer counterfactual analysis on the treatment strategy. The first layer calculates the state difference of different treatment combinations to obtain the initial interaction influence value, and the second layer introduces an adaptive weight matrix based on the complication risk to dynamically weight the initial interaction influence value. The adaptive weight matrix is updated in real time with the complication risk prediction value.

[0015] Jointly constrain the error between the dynamic feature vector and its reconstruction result, the consistency of the state distribution of the latent variable, the hierarchical constraints between the initial interaction influence value and the weighted interaction influence value, and the deviation between the complication risk prediction value and the actual risk. Through gradient iteration optimization, obtain a unified description of disease development, treatment intervention, and complication prevention and control.

[0016] In an alternative implementation,

[0017] Based on the dynamic causal network, the steps of deploying drug treatment agents, surgical plan agents, and nursing plan agents, forming sub - plans through a multi - level reasoning mechanism by combining the deep reinforcement learning method with professional domain rules and historical case features, performing feasibility analysis on the sub - plans based on confidence, and resolving conflicts to generate an initial treatment plan include:

[0018] Deploy drug treatment agents, surgical plan agents, and nursing plan agents based on the dynamic causal network. Each agent has a dual - network architecture consisting of a policy network and a value network; the policy network processes patient state information through a multi - head attention mechanism, calculates the query matrix, key matrix, and value matrix of the patient state vector, and generates an action policy distribution by fusing professional domain rules through residual connections; the value network adopts a multi - layer perceptron structure, takes the concatenation of the patient state vector and historical case features as input, and obtains a policy evaluation value;

[0019] Build a hierarchical reasoning framework including an immediate response layer, an intermediate effect layer, and a long - term impact layer based on the action policy distribution and the policy evaluation value. Extract environmental features at each layer. The environmental features include vital sign indicators, organ function status, and prognosis evaluation information; use a long short - term memory network to process and model the inter - layer dependencies of the environmental features, take the policy evaluation value as the guiding information of the long short - term memory network, generate a sub - plan probability distribution, and adjust the sub - plan probability based on the action policy distribution to obtain the sub - plan decisions of each agent;

[0020] Calculate the three - dimensional confidence composed of policy credibility, rule compliance, and experience similarity for the sub - plan decisions, where the policy credibility is calculated based on the action policy distribution; sort the sub - plans according to the weighted scores of the three - dimensional confidence, identify the temporal conflicts, resource conflicts, and effect conflicts between the sub - plans that exceed the preset confidence threshold, and form a conflict - free initial treatment plan through iterative adjustment and replacement.

[0021] In an alternative embodiment,

[0022] The steps of calculating the three - dimensional confidence composed of policy credibility, rule compliance, and experience similarity for the sub - plan decisions, where the policy credibility is calculated based on the action policy distribution; sorting the sub - plans according to the weighted scores of the three - dimensional confidence, identifying the temporal conflicts, resource conflicts, and effect conflicts between the sub - plans that exceed the preset confidence threshold, and forming a conflict - free initial treatment plan through iterative adjustment and replacement include:

[0023] Calculate the three-dimensional confidence for the sub-schemes. The three-dimensional confidence includes policy credibility, rule compliance, and experience similarity. Among them, the policy credibility is calculated based on the action policy distribution and an adaptive temperature parameter is introduced for distribution smoothing. The rule compliance is calculated by hierarchically evaluating the rules and performing a weighted accumulation of the degree to which the sub-schemes meet the rules and the rule weights. The experience similarity is obtained by calculating the weighted score of the similarity between the sub-schemes and historical cases and the treatment effects. Construct a factor graph containing confidence nodes and dependency edges, define a potential function representing the strength of the relationship between nodes based on the factor graph, use the variational inference method to optimize the dimensional weight coefficients of the three-dimensional confidence to obtain the optimized weights for each dimension of confidence, and calculate the weighted score of the three-dimensional confidence according to the optimized weights. Sort the sub-schemes according to the weighted scores.

[0024] For sub-schemes that exceed the preset confidence threshold, construct a spatio-temporal dependency graph containing treatment activity nodes and spatio-temporal constraint edges. Calculate the spatio-temporal set intersection and union ratio between sub-schemes based on the spatio-temporal dependency graph to obtain temporal conflicts. Calculate the ratio of the resource usage of sub-schemes at each time point to the upper limit of the resource capacity to obtain resource conflicts. Construct a treatment effect interference matrix representing the degree of mutual influence between sub-schemes to calculate effect conflicts. Identify the temporal conflicts, resource conflicts, and effect conflicts between sub-schemes.

[0025] For the identified conflicting sub-schemes, construct the adjustment decisions of the sub-schemes as decision variables. With the goal of maximizing the weighted score of the sub-schemes, while satisfying the temporal constraint that the execution times of the sub-schemes do not overlap, the resource constraint that the total resource usage at each time point does not exceed the capacity limit, the conflict constraint that conflicting sub-schemes cannot be executed simultaneously, and the dependency relationship constraint between sub-schemes. Use the branch and bound method for calculation and solution. Detect conflicts through an iterative approach and add spatio-temporal overlap degree threshold constraints, resource competition degree upper limit constraints, and effect conflict degree limit constraints, and perform sub-scheme adjustment and replacement until a conflict-free set of schemes is obtained. Among them, the branch and bound method includes designing a pruning strategy based on heuristic rules to dynamically adjust the branching order. Generate an initial treatment plan according to the conflict-free set of schemes.

[0026] In an alternative implementation manner,

[0027] Collect the temporal data of the patient's drug dosage, metabolic indicators, clinical symptoms, and adverse reactions in real time. Combine the attention calculation method to identify treatment features and use bidirectional recursion to calculate the predicted treatment trajectory. Based on the patient's tolerance indicators, organ function indicators, and treatment effect indicators, use a multi-objective optimization algorithm to calculate the optimal dosing plan. Analyze the expert decision-making characteristics in historical successful cases and establish a rule base for treatment timing, medication strategies, and prevention and control of complications. The steps to form decision criteria include:

[0028] Collect the time-series data of the patient's medication dosage, metabolic indicators, clinical symptoms, and adverse reactions; process the time-series data using a multi-head attention mechanism, extract high-order feature representations by calculating the attention weight matrix between features and applying a graph attention network; construct a bidirectional recurrent neural network based on the high-order feature representations, design a state transition probability matrix by combining bidirectional feature representations and introducing a time attention mechanism to calculate the change trends of the medication dosage, metabolic indicators, clinical symptoms, and adverse reactions, and predict the state distribution at a preset time point to obtain the treatment trajectory;

[0029] Establish an evaluation system including patient tolerance indicators, organ function indicators, and treatment effect indicators, where the patient tolerance indicators include the incidence of adverse reactions and the degree of symptom burden, the organ function indicators include liver and kidney function indicators, cardiopulmonary function parameters, and hematopoietic function indicators, and the treatment effect indicators include disease control rate, symptom improvement degree, and survival benefit assessment; construct a multi-objective optimization function based on the evaluation system, map the patient tolerance indicators, organ function indicators, and treatment effect indicators to optimization objectives of maximizing treatment effect, minimizing adverse reactions, and maintaining stable organ function respectively, and use the drug dosage range, dosing interval, and safety indicators as constraint conditions, and solve using a non-dominated sorting genetic algorithm to obtain the optimal dosing plan;

[0030] Analyze historical successful cases, extract case feature vectors including patient baseline characteristics, treatment process characteristics, and outcome evaluation characteristics, and construct expert decision-making features based on decision time points, decision basis elements, and intervention measure selections; establish a rule base consisting of treatment timing rules, medication strategy rules, and complication prevention and control rules according to the expert decision-making features, where the treatment timing rules are based on disease progression status, adverse reaction degree, and organ function status, the medication strategy rules include dosage adjustment principles, combination medication strategies, and treatment course arrangement plans, and the complication prevention and control rules include risk warning indicators, preventive intervention measures, and emergency treatment plans; use an analytic hierarchy decision tree to integrate the rules in the rule base, set rule trigger conditions and rule priorities, and form decision-making criteria.

[0031] In an alternative implementation,

[0032] The steps of dynamically optimizing the initial treatment plan by integrating the treatment trajectory prediction results, the optimal dosing plan, and the decision-making criteria to obtain the final treatment plan include:

[0033] Construct a conditional variational flow field to map the treatment trajectory prediction results, the optimal dosing plan, and the decision-making criteria to a shared latent space, perform data alignment by establishing bidirectional optical flow constraints of the forward flow field and the reverse flow field, and optimize based on the variational lower bound of the reconstruction error and KL divergence to obtain the aligned data stream;

[0034] Construct a three - level cascaded memory structure based on the aligned data stream, establish an association with the initial treatment plan through the multi - head self - attention mechanism of the immediate memory layer, capture the execution status of the initial treatment plan using the time - aware attention mechanism of the short - term memory layer, store the historical feedback information of the initial treatment plan in the long - term memory layer, and perform adaptive weight fusion on the three - level memories through a meta - controller to obtain the dynamic optimization features of the initial treatment plan; construct a stochastic differential equation based on the dynamic optimization features, perform path sampling on the initial treatment plan in the feature space, calculate the path probability distribution in combination with the historical feedback information, and optimize the path probability distribution through a variational inference method. At the same time, dynamically adjust the optimization parameters based on the execution effect to obtain the optimized path of the initial treatment plan;

[0035] Construct a two - layer medical rule verification network. The two - layer medical rule verification network includes a rule inference layer and an execution optimization layer. Among them, the rule inference layer uses a conditional random field to model the correlation between medical rules and generates a rule verification sequence. The execution optimization layer dynamically executes the rule verification sequence based on a deep reinforcement learning framework, verifies and adjusts the optimized path, and obtains a set of optimized solutions that meet the medical rules; construct an optimal transport problem based on the Wasserstein distance on the Riemannian manifold, define a local metric through the geodesic distance, and perform projection optimization fusion on the set of optimized solutions and the initial treatment plan to generate a final treatment plan.

[0036] In an alternative implementation,

[0037] The steps of constructing a conditional variational flow field to map the treatment trajectory prediction result, the optimal drug administration plan, and the decision criterion to a shared latent space, aligning the data by establishing a bidirectional optical flow constraint between the forward flow field and the reverse flow field, and optimizing based on the variational lower bound of the reconstruction error and the KL divergence to obtain the aligned data stream include:

[0038] A mapping framework is constructed by using a conditional variational flow field. The prediction results of treatment trajectories, the optimal drug administration plans, and decision criteria are mapped to a shared latent space through an encoder network with a multi-layer perceptron structure. The encoder network generates mean and variance parameters through reparameterization. The conditional variational flow field introduces a conditional prior network to generate prior distribution parameters based on modality identifiers, and optimizes and learns the representation of the shared latent space by maximizing the variational lower bound. A bidirectional mapping network between modalities is constructed based on the latent representation of the shared latent space. A cyclic consistency constraint of the Wasserstein metric is introduced, and the Wasserstein distance is calculated by iteratively optimizing the transport matrix through the Sinkhorn algorithm. An energy function network including a feature encoding layer and multi-layer non-linear mappings is designed. The energy function network inputs feature pairs of different modalities, measures their dual consistency by calculating the energy scores of the feature pairs, and uses the energy function network as a discriminator to obtain the initial feature distribution through an adversarial training mechanism.

[0039] A forward flow field and a reverse flow field are constructed for the initial feature distribution. The forward flow field and the reverse flow field respectively include a mean component, a variance component, and a random perturbation term. A composite mapping relationship is established based on the forward flow field and the reverse flow field. The composite mapping relationship includes the reverse flow field mapping after the action of the forward flow field and the forward flow field mapping after the action of the reverse flow field. The composite mapping relationship is constructed as a bidirectional optical flow constraint, and a composite mapping loss is constructed based on the bidirectional optical flow constraint.

[0040] A spectral regularization mechanism based on the Lipschitz constraint is introduced. The spectral norm of the network parameters is calculated by the power iteration method, and the parameter scale is adjusted by an adaptive clipping strategy. The variational lower bound is constructed with the reconstruction error and the KL divergence and integrated with the spectral regularization term as the variational optimization objective to constrain the composite mapping loss. A cross-modal attention module is designed to calculate the correlation weights for the constrained feature distribution, and different modality features are fused based on the correlation weights. The fused feature distribution is integrated by the Bayesian deep ensemble method with uncertainty parameter weighting to obtain the aligned data stream.

[0041] In the second aspect of the embodiments of the present invention,

[0042] A gynecological disease diagnosis and treatment system integrating electronic medical records is provided, including:

[0043] A first unit for collecting multi-modal diagnosis and treatment data of patients, including a sequence of temporal physiological indicators, a set of clinical manifestation features, a set of medical images, and a set of test reports. The data is processed by an autoencoder network for dimensionality reduction and denoising, a feature mapping relationship between modalities is constructed, common and individual features of the data are extracted, and a personalized medical database of the patient is established.

[0044] A second unit is configured to construct a dynamic causal network including disease progression, treatment intervention, and prevention and control of complications based on the personalized medical database. The dynamic causal network analyzes the interaction effects between different treatment means through a time-series causal reasoning algorithm. Based on the dynamic causal network, a drug treatment agent, a surgical plan agent, and a nursing plan agent are deployed. Through a deep reinforcement learning method combined with professional domain rules and historical case features, a multi-level reasoning mechanism is adopted to form sub-schemes, the feasibility of the sub-schemes is analyzed based on confidence, and conflict resolution is performed to generate an initial treatment plan.

[0045] A third unit is configured to collect time-series data of a patient's drug dosage, metabolic indicators, clinical symptoms, and adverse reactions in real time, identify treatment features by combining an attention calculation method, and predict a treatment trajectory by using bidirectional recursive calculation. Based on the patient's tolerance index, organ function index, and treatment effect index, a multi-objective optimization algorithm is used to calculate an optimal drug administration plan. The expert decision-making features in historical successful cases are analyzed, a rule base for treatment timing, drug use strategies, and prevention and control of complications is established, and decision-making criteria are formed. By integrating the treatment trajectory prediction results, the optimal drug administration plan, and the decision-making criteria, the initial treatment plan is dynamically optimized to obtain a final treatment plan.

[0046] In a third aspect of the embodiments of the present invention,

[0047] an electronic device is provided, including:

[0048] a processor;

[0049] a memory for storing instructions executable by the processor;

[0050] wherein the processor is configured to call the instructions stored in the memory to execute the method described above.

[0051] In a fourth aspect of the embodiments of the present invention,

[0052] a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0053] By collecting multi-modal medical data, performing dimensionality reduction and denoising processing, establishing an inter-modal feature mapping relationship, extracting common and individual features of the data, and forming a personalized medical database, the present invention effectively integrates various clinical data of patients and provides a reliable data basis for subsequent precise diagnosis and treatment.

[0054] Based on a dynamic causal network, the present invention constructs multiple intelligent agent systems, adopts deep reinforcement learning combined with professional rules and historical case features, and forms a treatment plan through multi-level reasoning, realizing a systematic analysis of disease progression, treatment intervention, and prevention and control of complications, and improving the scientificity and feasibility of the treatment plan.

[0055] By collecting various indicators during the patient's treatment process in real time, combining attention calculation to predict the treatment trajectory, using a multi-objective optimization algorithm to calculate the optimal drug administration plan, and integrating expert decision-making features to form decision-making criteria, the present invention realizes the dynamic optimization and personalized adjustment of the treatment plan, significantly improving the treatment effect and the prognosis of the patient. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a schematic flowchart of a gynecological disease diagnosis and treatment method integrating electronic medical records according to an embodiment of the present invention;

[0057] Figure 2 It is a schematic structural diagram of a gynecological disease diagnosis and treatment system integrating electronic medical records according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0059] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0060] Figure 1 It is a schematic flowchart of a gynecological disease diagnosis and treatment method integrating electronic medical records according to an embodiment of the present invention, as Figure 1 shown, the method includes:

[0061] S1. Collect multi-modal diagnosis and treatment data of the patient, including a time-series physiological index sequence, a clinical manifestation feature set, a medical image set, and a test report set, perform dimensionality reduction and denoising processing through an autoencoder network, construct an inter-modal feature mapping relationship, extract common and individual features of the data, and establish a personalized medical database for the patient;

[0062] S2. Based on the personalized medical database, construct a dynamic causal network including disease development, treatment intervention, and complication prevention and control. The dynamic causal network analyzes the interactive effects between different treatment methods through a temporal causal reasoning algorithm; Based on the dynamic causal network, deploy drug treatment agents, surgical plan agents, and nursing plan agents. Through the deep reinforcement learning method, combined with professional domain rules and historical case features, adopt a multi-level reasoning mechanism to form sub-schemes, conduct feasibility analysis on the sub-schemes based on confidence, and perform conflict resolution to generate an initial treatment plan;

[0063] S3. Real-time collect the temporal data of the patient's drug dosage, metabolic indicators, clinical symptoms, and adverse reactions, identify treatment features by combining the attention calculation method, and use bidirectional recursive calculation to predict the treatment trajectory; Based on the patient's tolerance indicators, organ function indicators, and treatment effect indicators, use a multi-objective optimization algorithm to calculate the optimal drug administration plan; Analyze the expert decision-making characteristics in historical successful cases, establish a rule base for treatment timing, drug use strategies, and complication prevention and control, and form decision-making criteria; Integrate the treatment trajectory prediction results, the optimal drug administration plan, and the decision-making criteria to dynamically optimize the initial treatment plan to obtain the final treatment plan.

[0064] In an alternative implementation,

[0065] The steps of constructing a dynamic causal network including disease development, treatment intervention, and complication prevention and control based on the personalized medical database, and analyzing the interactive effects between different treatment methods through a temporal causal reasoning algorithm include:

[0066] Construct a patient state sequence including instant features, cumulative features, and historical features based on the personalized medical database, extract the dynamic features of the patient state sequence through a recursive neural encoder, and use the time attention mechanism to fuse multi-scale temporal information to obtain a dynamic feature vector representing the disease development law;

[0067] Construct a temporal causal graph of treatment intervention and disease response based on the dynamic feature vector. Through the parameterization of the residual function for state transition calculation, introduce latent variables to describe uncertainty, and establish a dynamic causal network including feature nodes, action nodes, and latent variable nodes;

[0068] Calculate the complication risk prediction value based on the dynamic causal network, integrate dynamic features, latent variables, and treatment actions for risk assessment, and use the graph structure to calculate the correlation between complication types to obtain a treatment strategy considering risk constraints;

[0069] Perform a two - layer counterfactual analysis on the treatment strategy. In the first layer, calculate the state differences of different treatment combinations to obtain the initial interaction effect values. In the second layer, introduce an adaptive weight matrix based on the complication risk to dynamically weight the initial interaction effect values, and the adaptive weight matrix is updated in real - time according to the complication risk prediction value;

[0070] Jointly constrain the error between the dynamic feature vector and its reconstruction result, the consistency of the state distribution of the latent variable, the hierarchical constraints between the initial interaction effect value and the weighted interaction effect value, and the deviation between the complication risk prediction value and the actual risk. Through gradient iteration optimization, a unified description of disease development, treatment intervention, and complication prevention and control is achieved.

[0071] Exemplarily, first construct a patient state sequence. Extract the patient's immediate features (such as current vital signs, test indicators, etc.), cumulative features (such as cumulative drug dosage, treatment duration, etc.), and historical features (such as past medical history, family history, etc.) from the personalized medical database. Use a long - short - term memory network as a recursive neural encoder to encode the patient state sequence. By setting multiple time windows, extract feature information at different time scales respectively, and then use the attention mechanism to weight - fuse the features at different time scales to obtain a dynamic feature vector that can represent the law of disease development.

[0072] Then construct a temporal causal graph. Based on the above - mentioned dynamic feature vector, establish a causal relationship graph between treatment intervention and disease response. Calculate state transitions through a residual network, introduce latent variables described by probability distributions such as Gaussian distribution to characterize uncertainty, and finally form a dynamic causal network containing feature nodes (such as various test indicators), action nodes (such as medication regimens), and latent variable nodes. For example, for hypertensive patients, a causal relationship can be established between nodes such as blood pressure indicators, medication regimens, and complication risks.

[0073] Then perform complication risk prediction. Take the dynamic feature vector, latent variable state, and treatment plan as inputs, and calculate the risk of complication occurrence through a graph neural network. Use the graph structure to analyze the correlation between different complications, such as the degree of association between diabetic nephropathy and retinopathy. Based on the predicted risk value, constrain and adjust the treatment plan.

[0074] After that, conduct a two - layer counterfactual analysis. In the first layer, calculate the initial interaction between treatment plans by comparing the differences in patient states under different treatment combinations. In the second layer, introduce an adaptive weight matrix based on the complication risk to dynamically weight the initial interaction. The weight matrix is updated in real - time according to the complication risk prediction value, and the higher the risk, the greater the weight. For example, when the predicted risk of cardiovascular complications increases, increase the weight of relevant treatment plans.

[0075] Finally, joint optimization is carried out. Unified constraints are imposed on the reconstruction error of the dynamic feature vector, the consistency of the latent variable distribution, the hierarchical constraint of the interaction effect, and the accuracy of risk prediction. Through the iterative optimization algorithm, the overall modeling of disease development, treatment intervention, and complication prevention and control is realized.

[0076] By constructing a patient state sequence containing multi-dimensional features and integrating multi-scale time-series information, the present invention can comprehensively describe the law of disease development and improve the prediction accuracy of the changes in the condition; adopting a two-layer counterfactual analysis framework and combining an adaptive weight mechanism, it can accurately evaluate the interaction effects of different treatment plans and provide a reliable basis for the optimization of clinical treatment plans; by establishing a dynamic causal network and conducting joint optimization, the unified modeling of disease development, treatment intervention, and complication prevention and control is realized, and the scientificity and effectiveness of personalized treatment plans are improved.

[0077] In an alternative embodiment,

[0078] Based on the dynamic causal network, the steps of deploying drug treatment agents, surgical plan agents, and nursing plan agents, forming sub-plans through a multi-level reasoning mechanism by combining the deep reinforcement learning method with professional field rules and historical case features, conducting a feasibility analysis of the sub-plans based on the confidence level, and resolving conflicts to generate an initial treatment plan include:

[0079] Based on the dynamic causal network, drug treatment agents, surgical plan agents, and nursing plan agents are deployed. Each agent includes a dual-network architecture of a policy network and a value network; the policy network processes the patient state information through a multi-head attention mechanism, calculates the query matrix, key matrix, and value matrix of the patient state vector, and generates an action policy distribution by fusing professional field rules through a residual connection; the value network adopts a multi-layer perceptron structure, and inputs the concatenation of the patient state vector and historical case features to obtain a policy evaluation value;

[0080] Based on the action policy distribution and the policy evaluation value, a hierarchical reasoning framework including an immediate response layer, an intermediate effect layer, and a long-term impact layer is constructed. Environmental features are extracted at each layer. The environmental features include vital sign indicators, organ function status, and prognosis evaluation information; the long short-term memory network is used to process and model the environmental features to establish inter-layer dependencies, and the policy evaluation value is used as the guiding information of the long short-term memory network to generate a sub-plan probability distribution, and the sub-plan probability is adjusted based on the action policy distribution to obtain the sub-plan decision of each agent;

[0081] For the three-dimensional confidence level composed of the credibility, rule compliance, and experience similarity of the sub-scheme decision-making calculation strategy, where the strategy credibility is calculated based on the action strategy distribution; sort the sub-schemes according to the weighted scores of the three-dimensional confidence level, and identify the temporal conflicts, resource conflicts, and effect conflicts between the sub-schemes that exceed the preset confidence threshold, and form an initial treatment plan without conflicts through iterative adjustment and replacement.

[0082] Exemplarily, first deploy three core agents: a drug treatment agent, a surgical plan agent, and a nursing plan agent. Each agent adopts a dual-network architecture, including a policy network and a value network. In the policy network, an eight-head attention mechanism is used to process the patient's state information, and the various index data of the patient are converted into a query matrix, a key matrix, and a value matrix. Taking vital sign indicators such as blood pressure, heart rate, and blood oxygen as an example, after obtaining the weight distribution through attention calculation, it is fused with the medication guidelines, surgical indications, nursing specifications, etc. in the professional rule library to generate an action strategy distribution. The value network adopts a five-layer fully connected network, and the input layer includes the patient's state vector and historical case features. Taking a lung cancer patient as an example, the state information such as his age, course of disease, and stage is spliced and input together with the features of historical cured cases, and the strategy evaluation value is obtained through network calculation.

[0083] Then construct a three-layer inference framework. The immediate response layer focuses on real-time vital sign indicators such as blood pressure and heart rate; the medium-term effect layer focuses on organ state indicators such as liver and kidney functions; the long-term impact layer focuses on prognostic indicators such as five-year survival rate. Use a long short-term memory network to model the inter-layer dependencies, with 128 memory units and 3 hidden layers. Taking the above-mentioned lung cancer patient as an example, input the environmental features such as his vital signs, organ functions, and prognostic evaluations into the network, and at the same time use the strategy evaluation value as the guiding information to generate the sub-scheme probability distribution of each agent. The drug agent generates a targeted drug treatment plan, the surgical agent generates a minimally invasive surgical plan, and the nursing agent generates a perioperative nursing plan.

[0084] Finally, calculate the three-dimensional confidence level of the sub-scheme. The strategy credibility is calculated based on the action strategy distribution; the rule compliance is obtained by comparing with the clinical guidelines; the experience similarity is obtained by comparing with historical successful cases. For the treatment plan of the above-mentioned lung cancer patient, set the confidence threshold to 0.8, and screen out the sub-schemes with a confidence level higher than the threshold. Identify the temporal conflicts between the plans, such as the timing of medication before and after surgery; resource conflicts, such as the occupancy of the operating room; effect conflicts, such as drug interactions. Through iterative adjustment and plan replacement, finally form an initial treatment plan without conflicts.

[0085] Through the multi-agent collaborative decision-making and multi-level reasoning mechanism, the present invention improves the scientificity and reliability of the treatment plan, reduces the risk of medical errors, ensures the feasibility and safety of the treatment plan based on the confidence evaluation and conflict resolution mechanism, enhances the clinical implementation effect, combines professional rules and historical experience, realizes the intelligent generation of personalized treatment plans, and improves the medical decision-making efficiency.

[0086] In an alternative embodiment,

[0087] The steps of calculating the three-dimensional confidence of the sub-scheme decision, which consists of the strategy confidence, rule compliance, and experience similarity, where the strategy confidence is calculated based on the action strategy distribution; sorting the sub-schemes according to the weighted scores of the three-dimensional confidence, and for the sub-schemes exceeding the preset confidence threshold, identifying the temporal conflicts, resource conflicts, and effect conflicts between them and forming a conflict-free initial treatment plan through iterative adjustment and replacement include:

[0088] Calculating the three-dimensional confidence for the sub-scheme, where the three-dimensional confidence includes the strategy confidence, rule compliance, and experience similarity; the strategy confidence is calculated based on the action strategy distribution and an adaptive temperature parameter is introduced for distribution smoothing; the rule compliance adopts a hierarchical rule evaluation method and is calculated by the weighted accumulation of the degree to which the sub-scheme meets the rules and the rule weights; the experience similarity is obtained by calculating the weighted scores of the similarity between the sub-scheme and historical cases and the treatment effect; constructing a factor graph containing confidence nodes and dependency edges, defining a potential function representing the relationship strength between nodes based on the factor graph, using the variational inference method to optimize the dimension weight coefficients of the three-dimensional confidence to obtain the optimized weights of each dimension confidence, and calculating the weighted scores of the three-dimensional confidence according to the optimized weights, and sorting the sub-schemes according to the weighted scores;

[0089] For the sub-schemes exceeding the preset confidence threshold, constructing a spatio-temporal dependence graph containing treatment activity nodes and spatio-temporal constraint edges, calculating the spatio-temporal set intersection and union ratio between sub-schemes based on the spatio-temporal dependence graph to obtain temporal conflicts; calculating the ratio of the resource usage of the sub-scheme at each time point to the upper limit of the resource capacity to obtain resource conflicts; constructing a treatment effect interference matrix representing the degree of mutual influence between sub-schemes to calculate effect conflicts; identifying the temporal conflicts, resource conflicts, and effect conflicts between sub-schemes;

[0090] For the identified conflicting sub - solutions, construct the adjustment decision of the sub - solutions as decision variables. With the goal of maximizing the weighted score of the sub - solutions, while satisfying the timing constraint that the execution times of the sub - solutions do not overlap, the resource constraint that the total resource usage at each time point does not exceed the capacity limit, the conflict constraint that conflicting sub - solutions cannot be executed simultaneously, and the dependency relationship constraint between sub - solutions; use the branch - and - bound method for calculation and solution. Detect conflicts through an iterative approach and add the spatio - temporal overlap degree threshold constraint, the resource competition degree upper - limit constraint, and the effect conflict degree limit constraint, and perform sub - solution adjustment and replacement until a conflict - free solution set is obtained; where the branch - and - bound method includes designing a pruning strategy based on heuristic rules to dynamically adjust the branch order; generate an initial treatment plan according to the conflict - free solution set.

[0091] Exemplarily, first calculate three - dimensional confidence levels of credibility of the calculation strategy, compliance with rules, and similarity of experience for the sub - solutions. The calculation of the credibility of the strategy is performed by smoothing the action strategy distribution by introducing an adaptive temperature parameter. The initial value of the temperature parameter is set to 0.8 and is dynamically adjusted according to the entropy value of the strategy distribution. The larger the entropy value, the smaller the temperature parameter. Taking the treatment plan for endometriosis as an example, the distribution probabilities of different treatment strategies such as hormone therapy and surgical treatment are 0.4, 0.3, etc. After temperature smoothing, the corrected credibility of the strategy is obtained.

[0092] The compliance with rules adopts a hierarchical evaluation method. The clinical guideline rules are divided into multiple levels such as medication taboos, dosage ranges, and administration routes. The rule weights are determined by expert scoring. Taking the hormone therapy plan as an example, the weight of the medication taboo is 0.4, the weight of the dosage range is 0.3, and the weight of the administration route is 0.3. Evaluate the satisfaction degree of each rule for the sub - solution, with the value range of 0 - 1, and obtain the compliance with rules through weighted accumulation.

[0093] The similarity of experience is calculated based on historical cases. Select features such as age, pregnancy and childbirth history, and lesion site to calculate the similarity between the sub - solution and historical cases, and perform weighting in combination with the treatment effects of historical cases. Taking the treatment plan for endometriosis as an example, if the similarity between a certain sub - solution and a historical case is 0.8, and the treatment effect score of this historical case is 0.9, then the score of this dimension is 0.72.

[0094] Construct a factor graph to optimize the weights of the three - dimensional confidence levels. The nodes in the graph represent the confidence levels of each dimension, and the edges represent the relationships between dimensions. Through variational inference and iterative optimization, the weights of the credibility of the strategy, compliance with rules, and similarity of experience are 0.4, 0.35, and 0.25 respectively. Calculate and sort the weighted scores of the three - dimensional confidence levels, and set the confidence level threshold to 0.7.

[0095] Identify conflicts for sub - solutions exceeding the threshold. Construct a spatio - temporal dependency graph, where nodes represent treatment activities and edges represent spatio - temporal constraints. Taking the time arrangement of hormone therapy and surgical therapy as an example, calculate the activity time overlap to identify temporal conflicts. If the overlap exceeds 0.3, it is determined that there is a conflict. Statistically calculate the ratio of the usage amount to the capacity of resources such as the operating room and medical staff at each time point to identify resource conflicts. If the ratio exceeds 0.8, it is determined that there is a conflict. Construct an effect interference matrix to calculate the degree of treatment effect conflict. For example, the impact of hormone therapy on the surgical timing, if the interference degree exceeds 0.5, it is determined that there is a conflict.

[0096] Use the branch - and - bound method to iteratively adjust the conflicting sub - solutions. Design pruning rules based on the surgical time window and medical resource utilization rate to dynamically adjust the branching order. Adjust the solution by adding constraints such as the spatio - temporal overlap degree less than 0.3, the resource competition degree less than 0.8, and the effect conflict degree less than 0.5. Replace the solutions that cannot be adjusted, select the alternative solutions with sub - optimal confidence, and finally form an initial conflict - free treatment plan.

[0097] The present invention effectively smooths the policy distribution, improves the stability and reliability of decision - making, and avoids the problem of overly extreme policy selection by introducing a method for calculating the credibility of strategies with an adaptive temperature parameter. By using a hierarchical rule evaluation and an experience similarity calculation method based on historical cases, it comprehensively considers clinical specification requirements and practical experience, improves the scientificity and feasibility of the treatment plan, reduces the treatment risk. The conflict identification method based on the spatio - temporal dependency graph and the effect interference matrix, combined with the iterative optimization strategy of the branch - and - bound algorithm, realizes the automatic coordination and optimization of the treatment plan, and improves the efficiency and quality of plan generation.

[0098] In an alternative embodiment,

[0099] Real - time collect the time - series data of the patient's drug dosage, metabolic indicators, clinical symptoms, and adverse reactions, identify treatment characteristics by combining an attention calculation method, and use bidirectional recursion to calculate the predicted treatment trajectory; based on the patient's tolerance index, organ function index, and treatment effect index, use a multi - objective optimization algorithm to calculate the optimal drug administration plan; analyze the expert decision - making characteristics in historical successful cases, establish a rule base for treatment timing, drug - using strategies, and complication prevention and control. The steps of forming decision - making criteria include:

[0100] Collect the time-series data of the patient's drug dosage, metabolic indicators, clinical symptoms, and adverse reactions; process the time-series data using a multi-head attention mechanism, extract high-order feature representations by calculating the attention weight matrix between features and applying a graph attention network; construct a bidirectional recurrent neural network based on the high-order feature representations, design a state transition probability matrix by combining bidirectional feature representations and introducing a time attention mechanism to calculate the change trends of the drug dosage, metabolic indicators, clinical symptoms, and adverse reactions, and predict the state distribution at a preset time point to obtain the treatment trajectory;

[0101] Establish an evaluation system including patient tolerance indicators, organ function indicators, and treatment effect indicators, where the patient tolerance indicators include the incidence of adverse reactions and the degree of symptom burden, the organ function indicators include liver and kidney function indicators, cardiopulmonary function parameters, and hematopoietic function indicators, and the treatment effect indicators include the disease control rate, the degree of symptom improvement, and the evaluation of survival benefit; construct a multi-objective optimization function based on the evaluation system, map the patient tolerance indicators, organ function indicators, and treatment effect indicators to the optimization objectives of maximizing the treatment effect, minimizing adverse reactions, and maintaining stable organ function respectively, use the drug dosage range, dosing interval, and safety indicators as constraint conditions, and solve using a non-dominated sorting genetic algorithm to obtain the optimal dosing plan;

[0102] Analyze historical successful cases, extract case feature vectors including patient baseline characteristics, treatment process characteristics, and outcome evaluation characteristics, and construct expert decision-making features based on the decision time point, decision basis elements, and intervention measure selection; establish a rule base consisting of treatment timing rules, medication strategy rules, and complication prevention and control rules according to the expert decision-making features, where the treatment timing rules are based on the disease progression state, the degree of adverse reactions, and the organ function state, the medication strategy rules include dosage adjustment principles, combination medication strategies, and treatment course arrangement plans, and the complication prevention and control rules include risk warning indicators, preventive intervention measures, and emergency treatment plans; use an analytic hierarchy decision tree to integrate the rules in the rule base, set rule trigger conditions and rule priorities, and form decision-making criteria.

[0103] Exemplarily, first perform time-series data collection and feature extraction. Record the patient's medication information in real time through the hospital information system, including the specific drug name, dosage (such as morphine 10mg / time), dosing frequency (such as once every 4 hours), etc. At the same time, collect various metabolic indicators of the patient, such as vital sign data such as blood glucose, blood pressure, and heart rate, clinical symptom scores (such as pain scores from 0 to 10), and adverse reaction records (such as nausea, drowsiness degree, etc.). Preprocess the collected time-series data, including operations such as outlier processing and missing value filling.

[0104] The multi-head attention mechanism is adopted to process time-series data. Different types of feature data are mapped to the same feature space to calculate the correlation weights between features. For example, analyze the degree of association between drug dosage and adverse reactions, or the correlation between symptom improvement and changes in metabolic indicators. Higher-order features are further extracted through the graph attention network to capture the complex interaction relationships between features.

[0105] A bidirectional recurrent neural network is constructed to predict the treatment trajectory. Based on the extracted higher-order features, recurrent connections are established in both the forward and backward directions respectively to fuse the bidirectional feature representations. The time attention mechanism is introduced to assign different weights to the states at different time points. A state transition probability matrix is designed to calculate the change trends of various indicators. For example, predict the change curve of the pain score within the next 24 hours, or predict the risk of adverse reactions occurring within a week.

[0106] A multi-dimensional evaluation system is established. The patient tolerance evaluation includes the statistics of the incidence of adverse reactions (such as the incidence of nausea < 30%) and the symptom burden score. The organ function evaluation includes liver function indicators (transaminases, bilirubin, etc.), kidney function indicators (creatinine, urea nitrogen, etc.), cardiopulmonary function parameters (heart rate, blood pressure, respiratory rate, etc.) and blood routine indicators. The treatment effect evaluation includes the disease control rate, symptom improvement measurement and survival analysis.

[0107] The multi-objective optimization algorithm is adopted to calculate the dosing scheme. The evaluation indicators are transformed into optimization objectives, such as maximizing the disease control rate while minimizing the incidence of adverse reactions. Constraint conditions such as the drug dose range (such as morphine 10 - 30 mg / time) and the dosing interval (4 - 6 hours) are set. Through the non-dominated sorting genetic algorithm for iterative optimization, a set of Pareto optimal solutions are obtained as candidate dosing schemes.

[0108] Analyze historical successful cases to construct a decision rule base. Extract the patient baseline characteristics (age, weight, underlying diseases, etc.), treatment process characteristics (dosing scheme, adverse reactions, symptom changes, etc.) and outcome evaluation characteristics (treatment effect, survival period, etc.). Summarize the expert decision-making characteristics, including the selection of treatment timing, adjustment of dosing strategies and complication treatment plans. Establish a rule base covering treatment timing rules (such as adjusting the scheme when the disease progresses), medication strategy rules (such as the principle of reducing the dose when adverse reactions occur) and complication prevention and control rules (such as the indications for prophylactic medication).

[0109] By collecting multi-dimensional clinical data in real time and applying deep learning methods, the present invention can accurately predict the changing trends of various indicators during the treatment process, provide an objective basis for clinical decision-making, and improve the accuracy and individualization level of the medication plan; adopt a multi-objective optimization algorithm to balance the treatment effect and safety, comprehensively consider the patient's tolerance, organ function and treatment effect, and ensure that the dosing plan can not only achieve the expected treatment goal, but also control the adverse reactions within an acceptable range; the expert decision rule base constructed based on historical successful cases systematizes and standardizes the experience and knowledge of clinical experts, facilitating clinical promotion and application. At the same time, through continuous accumulation and optimization, the intelligent level of the system is improved, realizing the dynamic optimization and personalized adjustment of the treatment plan.

[0110] In an alternative embodiment,

[0111] The steps of dynamically optimizing the initial treatment plan by integrating the treatment trajectory prediction result, the optimal dosing plan and the decision criterion to obtain the final treatment plan include:

[0112] Construct a conditional variational flow field to map the treatment trajectory prediction result, the optimal dosing plan and the decision criterion to a shared latent space, perform data alignment by establishing a bidirectional optical flow constraint between the forward flow field and the reverse flow field, and optimize based on the variational lower bound of the reconstruction error and KL divergence to obtain the aligned data stream;

[0113] Based on the aligned data stream, construct a three-level cascaded memory structure, establish an association with the initial treatment plan through the multi-head self-attention mechanism of the immediate memory layer, capture the execution state of the initial treatment plan by using the time-aware attention mechanism of the short-term memory layer, store the historical feedback information of the initial treatment plan by using the long-term memory layer, and perform adaptive weight fusion on the three-level memories through a meta-controller to obtain the dynamic optimization features of the initial treatment plan; construct a stochastic differential equation according to the dynamic optimization features, perform path sampling on the initial treatment plan in the feature space, calculate the path probability distribution in combination with the historical feedback information, and optimize the path probability distribution through a variational inference method. At the same time, dynamically adjust the optimization parameters based on the execution effect to obtain the optimized path of the initial treatment plan;

[0114] Construct a two - layer medical rule verification network. The two - layer medical rule verification network includes a rule reasoning layer and an execution optimization layer. Among them, the rule reasoning layer uses a conditional random field to model the correlation between medical rules and generates a rule verification sequence. The execution optimization layer dynamically executes the rule verification sequence based on a deep reinforcement learning framework, verifies and adjusts the optimization path, and obtains a set of optimization solutions that meet the medical rules; construct an optimal transport problem based on the Wasserstein distance on the Riemannian manifold, define a local metric through the geodesic distance, and project - optimize and fuse the set of optimization solutions with the initial treatment plan to generate a final treatment plan.

[0115] Exemplarily, in the process of dynamic optimization of the treatment plan, it is first necessary to construct a conditional variational flow field to achieve data alignment. By designing a forward flow field and a reverse flow field, map the treatment trajectory prediction result, the optimal drug administration plan, and the decision criterion to a shared latent space. The forward flow field adopts a multi - layer perceptron network structure. The number of nodes in the input layer is the original feature dimension. The hidden layer uses the LeakyReLU activation function, and the dimension of the output layer is the same as the dimension of the latent space. The reverse flow field adopts a symmetric network structure. In the training stage, calculate the reconstruction error based on paired data samples, introduce the KL divergence as a regularization term, and optimize the network parameters through stochastic gradient descent to finally obtain an aligned data stream.

[0116] Then construct a three - layer cascaded memory structure to achieve dynamic feature extraction. The immediate memory layer uses a multi - head self - attention mechanism. The number of heads is set to 8, and the dimension of each attention head is 64. Map the initial treatment plan to the query and key - value spaces through a linear projection matrix. The short - term memory layer introduces time encoding, inputs the treatment plan execution status sequence into a bidirectional long - short - term memory network, and the dimension of the hidden layer is set to 256. The long - term memory layer adopts a key - value storage structure with a storage capacity of 1000, and obtains similar historical cases through a nearest - neighbor retrieval mechanism. The meta - controller calculates the fusion weights of the three - layer memories based on the current state and uses softmax normalization to ensure that the weights sum to 1.

[0117] Subsequently, perform path sampling and optimization in the feature space. Construct a stochastic differential equation, use Brownian motion as the noise term, set the step size to 0.01, and the number of sampled trajectories to 100. For each sampled path, calculate the path probability score based on historical feedback and perform variational inference using the importance sampling method. Dynamically adjust the temperature parameter according to the execution effect to control the balance between exploration and exploitation.

[0118] Finally, a two - layer medical rule verification network is constructed for scheme optimization. The rule inference layer encodes medical rules into feature vectors and uses conditional random fields to model the dependencies between rules. The execution optimization layer is based on the deep Q - network framework, where the state space is the feature representation of the current treatment plan, the action space is the optional adjustment operations, and the reward function comprehensively considers the rule compliance and treatment effect. An optimal transport problem is constructed based on the Wasserstein distance on the Riemannian manifold, with the geodesic distance used as the local metric, and the final treatment plan is obtained through iterative optimization of the set of optimized plans and the initial treatment plan.

[0119] The present invention realizes the alignment and fusion of heterogeneous data through the conditional variational flow field, improves the consistency of treatment trajectory prediction results, optimal drug administration plans, and decision - making criteria, and provides a reliable data basis for subsequent optimization; by combining a three - level cascade memory structure and stochastic differential equations, it can make full use of historical experience and maintain sufficient exploration, having strong adaptability while ensuring the stability of the plan; based on the two - layer medical rule verification network and the optimal transport theory on the Riemannian manifold, it ensures that the optimized plan meets the medical rule constraints and realizes performance improvement on the premise of maintaining similarity with the initial plan.

[0120] In an alternative embodiment,

[0121] The steps of constructing a conditional variational flow field to map the treatment trajectory prediction results, optimal drug administration plan, and decision - making criteria to a shared latent space, aligning the data by establishing a bidirectional optical flow constraint between the forward flow field and the reverse flow field, and optimizing based on the variational lower bound of the reconstruction error and KL divergence to obtain the aligned data stream include:

[0122] Construct a mapping framework using the conditional variational flow field, map the treatment trajectory prediction results, optimal drug administration plan, and decision - making criteria to the shared latent space through an encoder network with a multi - layer perceptron structure. The encoder network generates mean and variance parameters through re - parameterization. The conditional variational flow field introduces a conditional prior network to generate prior distribution parameters based on the modality identifier and optimizes the representation of the shared latent space by maximizing the variational lower bound; construct an inter - modality bidirectional mapping network based on the latent representation of the shared latent space, introduce the cyclic consistency constraint of the Wasserstein metric, iteratively optimize the transport matrix to calculate the Wasserstein distance through the Sinkhorn algorithm, design an energy function network including a feature encoding layer and multi - layer non - linear mapping. The energy function network inputs feature pairs of different modalities, measures their dual consistency by calculating the energy scores of the feature pairs, and uses the energy function network as a discriminator to obtain the initial feature distribution through the adversarial training mechanism;

[0123] Construct a forward flow field and a reverse flow field for the initial feature distribution. The forward flow field and the reverse flow field respectively include a mean component, a variance component, and a random perturbation term. Establish a composite mapping relationship based on the forward flow field and the reverse flow field. The composite mapping relationship includes the reverse flow field mapping after the action of the forward flow field and the forward flow field mapping after the action of the reverse flow field. Construct the composite mapping relationship into a bidirectional optical flow constraint, and construct a composite mapping loss based on the bidirectional optical flow constraint.

[0124] Introduce a spectral regularization mechanism based on Lipschitz constraint. Calculate the spectral norm of network parameters by the power iteration method and adopt an adaptive clipping strategy to adjust the parameter scale. Construct a variational lower bound with the reconstruction error and KL divergence and integrate it with the spectral regularization term into a variational optimization objective to constrain the composite mapping loss. Design a cross-modal attention module to calculate the correlation weights of the constrained feature distribution, and fuse different modal features based on the correlation weights. Integrate the fused feature distribution using the Bayesian deep ensemble method weighted by uncertainty parameters to obtain an aligned data stream.

[0125] Exemplarily, for the data alignment method based on conditional variational flow fields, first construct a mapping framework for latent space representation learning of multi-modal data. Design an encoder network through a multi-layer perceptron structure. This network includes four fully connected layers, with the number of neurons in each layer being 512, 256, 128, and 64 respectively, and the ReLU activation function is used. The encoder network inputs data such as treatment trajectory prediction results, optimal dosing regimens, and decision criteria, and outputs a 64-dimensional latent representation. During the encoding process, the mean and variance parameters are generated through the reparameterization trick, where the mean parameter is obtained through a linear mapping, and the variance parameter passes through the softplus function to ensure non-negativity. The conditional prior network generates prior distribution parameters based on the identity vectors of different modalities, and the identity vectors are in the one-hot encoding form.

[0126] After obtaining the shared latent space representation, construct an inter-modal bidirectional mapping network to achieve feature alignment. Introduce the cyclic consistency constraint of the Wasserstein metric, and iteratively calculate the transport matrix through the Sinkhorn algorithm, with the number of iterations set to 100. The energy function network includes a feature encoding layer and three non-linear mapping layers. The feature encoding layer reduces the input features to 128 dimensions, and the output dimensions of the non-linear mapping layers are 64, 32, and 1 in sequence. The energy function network inputs feature pairs of different modalities and calculates the energy scores of the feature pairs to evaluate their dual consistency. Optimize the energy function network through an adversarial training mechanism, with the number of training rounds being 200 rounds.

[0127] Construct forward and reverse flow fields for fine alignment based on the initial feature distribution. The flow field consists of a mean component, a variance component, and a random perturbation term, where the random perturbation term follows a standard normal distribution. Establish a composite mapping relationship based on the forward and reverse flow fields, including two mapping methods: forward-reverse and reverse-forward. Construct the composite mapping relationship as a bidirectional optical flow constraint, and set the constraint strength coefficient to 0.1.

[0128] Introduce a spectral regularization mechanism based on Lipschitz constraint to ensure mapping stability. Calculate the spectral norm of the network parameters by the power iteration method, and the number of iterations is 5 times. Adopt an adaptive clipping strategy to adjust the parameter scale, and set the clipping threshold to 1.0. Construct a variational lower bound with the reconstruction error and KL divergence, and integrate it with the spectral regularization term as a variational optimization objective, where the coefficient of the spectral regularization term is 0.01.

[0129] Design a cross-modal attention module to calculate the feature correlation weights. The attention module includes a query matrix, a key matrix, and a value matrix, all with a dimension of 64. Calculate the correlation weights through the dot-product attention mechanism, and perform weighted fusion on the features of different modalities after softmax normalization. Finally, adopt the Bayesian deep ensemble method to integrate the feature distribution, with the number of integrated models being 5, and each model being assigned an uncertainty parameter weight.

[0130] The present invention realizes the unified representation of the treatment trajectory prediction result, the optimal drug administration plan, and the decision-making criterion in the shared latent space through the conditional variational flow field, overcomes the challenges brought by the heterogeneity of multi-modal data, and improves the consistency and interpretability of feature representation; realizes stable and controllable feature alignment based on the bidirectional optical flow constraint and the spectral regularization mechanism, effectively reduces the distribution difference between modalities, and enhances the robustness and generalization ability of the mapping result; adopts the cross-modal attention mechanism and the Bayesian deep ensemble method to realize the adaptive fusion of features, fully utilizes the complementary information between different modalities, and improves the quality and reliability of the aligned data stream.

[0131] Figure 2 The structural schematic diagram of the gynecological disease diagnosis and treatment system integrating electronic medical records for the embodiment of the present invention is as Figure 2 shown, and the system includes:

[0132] The first unit is used to collect the multi-modal diagnosis and treatment data of patients, including the time-series physiological index sequence, the clinical manifestation feature set, the medical image set, and the test report set, perform dimensionality reduction and denoising processing through an autoencoder network, construct the feature mapping relationship between modalities, extract the common and individual features of the data, and establish a personalized medical database for patients;

[0133] A second unit is used to construct a dynamic causal network including disease development, treatment intervention, and prevention and control of complications based on the personalized medical database. The dynamic causal network analyzes the interaction effects between different treatment means through a time-series causal reasoning algorithm; based on the dynamic causal network, drug treatment agents, surgical plan agents, and nursing plan agents are deployed. Through a deep reinforcement learning method combined with professional field rules and historical case characteristics, a multi-level reasoning mechanism is adopted to form sub-schemes, the feasibility of the sub-schemes is analyzed based on confidence, and conflict resolution is performed to generate an initial treatment plan;

[0134] A third unit is used to collect real-time time-series data of patients' drug dosages, metabolic indicators, clinical symptoms, and adverse reactions, identify treatment characteristics by combining an attention calculation method, and predict treatment trajectories using bidirectional recursive calculations; based on patients' tolerance indicators, organ function indicators, and treatment effect indicators, a multi-objective optimization algorithm is used to calculate the optimal drug administration plan; analyze the expert decision-making characteristics in historical successful cases, establish a rule base for treatment timing, drug use strategies, and prevention and control of complications, and form decision-making criteria; integrate the treatment trajectory prediction results, the optimal drug administration plan, and the decision-making criteria to dynamically optimize the initial treatment plan to obtain a final treatment plan.

[0135] In the third aspect of the embodiments of the present invention,

[0136] Provided is an electronic device, including:

[0137] A processor;

[0138] A memory for storing instructions executable by the processor;

[0139] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0140] In the fourth aspect of the embodiments of the present invention,

[0141] Provided is a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0142] The present invention can be a method, a device, a system, and / or a computer program product. The computer program product can include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0143] 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 foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and 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 various embodiments of the present invention.

Claims

1. A gynecological disease diagnosis and treatment method integrating electronic medical records, characterized in that, Including: Collecting multi-modal diagnosis and treatment data of patients, including time-series physiological index sequences, clinical manifestation feature sets, medical image sets, and test report sets, performing dimensionality reduction and denoising processing through an autoencoder network, constructing inter-modal feature mapping relationships, extracting common and individual features of the data, and establishing a personalized medical database for patients; Based on the personalized medical database, constructing a dynamic causal network including disease development, treatment intervention, and complication prevention and control. The dynamic causal network analyzes the interactive effects between different treatment methods through a time-series causal reasoning algorithm. Based on the dynamic causal network, deploying drug treatment agents, surgical plan agents, and nursing plan agents, forming sub-plans through a multi-level reasoning mechanism by combining deep reinforcement learning methods with professional domain rules and historical case features, performing feasibility analysis on the sub-plans based on confidence, and resolving conflicts to generate an initial treatment plan; Real-time collecting time-series data of patients' drug dosages, metabolic indicators, clinical symptoms, and adverse reactions, identifying treatment features by combining an attention calculation method, and predicting treatment trajectories using bidirectional recursive calculations. Based on patients' tolerance indicators, organ function indicators, and treatment effect indicators, calculating the optimal drug administration plan using a multi-objective optimization algorithm. Analyzing the expert decision-making features in historical successful cases, establishing a rule base for treatment timing, drug use strategies, and complication prevention and control, and forming decision-making criteria. Integrating the treatment trajectory prediction results, the optimal drug administration plan, and the decision-making criteria to dynamically optimize the initial treatment plan to obtain the final treatment plan.

2. The method according to claim 1, wherein Based on the personalized medical database, the steps of constructing a dynamic causal network including disease development, treatment intervention, and complication prevention and control, where the dynamic causal network analyzes the interactive effects between different treatment methods through a time-series causal reasoning algorithm include: Constructing a patient state sequence including instant features, cumulative features, and historical features based on the personalized medical database, extracting the dynamic features of the patient state sequence through a recursive neural encoder, and using a time attention mechanism to fuse multi-scale time-series information to obtain a dynamic feature vector representing the disease development law; Constructing a time-series causal graph of treatment intervention and disease response based on the dynamic feature vector, parameterizing state transition calculations through a residual function, introducing latent variables to describe uncertainty, and establishing a dynamic causal network including feature nodes, action nodes, and latent variable nodes; Calculating the complication risk prediction value based on the dynamic causal network, integrating dynamic features, latent variables, and treatment actions for risk assessment, and calculating the correlation between complication types using a graph structure to obtain a treatment strategy considering risk constraints; Performing two-layer counterfactual analysis on the treatment strategy, where the first layer calculates the state differences of different treatment combinations to obtain the initial interactive effect value, and the second layer introduces an adaptive weight matrix based on the complication risk to dynamically weight the initial interactive effect value, and the adaptive weight matrix is updated in real time with the complication risk prediction value; Jointly constrain the error between the dynamic feature vector and its reconstruction result, the consistency of the state distribution of the latent variable, the hierarchical constraint between the initial interaction influence value and the weighted interaction influence value, and the deviation between the complication risk prediction value and the actual risk, and optimize the unified description of disease development, treatment intervention, and complication prevention and control through gradient iteration.

3. The method according to claim 1, wherein Based on the dynamic causal network, deploy drug treatment agents, surgical plan agents, and nursing plan agents. The steps of forming sub-plans through a multi-level reasoning mechanism by combining domain rules and historical case features using the deep reinforcement learning method, performing feasibility analysis on the sub-plans based on confidence, and resolving conflicts to generate an initial treatment plan include: Deploy drug treatment agents, surgical plan agents, and nursing plan agents based on the dynamic causal network. Each agent includes a dual network architecture of a policy network and a value network; the policy network processes patient state information through a multi-head attention mechanism, calculates the query matrix, key matrix, and value matrix of the patient state vector, and generates an action policy distribution by fusing domain rules through residual connections; the value network adopts a multi-layer perceptron structure, inputs the concatenation of the patient state vector and historical case features, and obtains a policy evaluation value. Construct a hierarchical reasoning framework including an immediate response layer, a medium-term effect layer, and a long-term impact layer based on the action policy distribution and the policy evaluation value. Extract environmental features at each layer. The environmental features include vital sign indicators, organ function status, and prognosis evaluation information; use a long short-term memory network to process and model the inter-layer dependencies of the environmental features, use the policy evaluation value as the guiding information of the long short-term memory network, generate a sub-plan probability distribution, and adjust the sub-plan probability based on the action policy distribution to obtain the sub-plan decisions of each agent. Calculate the three-dimensional confidence composed of policy credibility, rule compliance, and experience similarity for the sub-plan decisions, where the policy credibility is calculated based on the action policy distribution; sort the sub-plans according to the weighted scores of the three-dimensional confidence, identify the temporal conflicts, resource conflicts, and effect conflicts between the sub-plans that exceed the preset confidence threshold, and form a conflict-free initial treatment plan through iterative adjustment and replacement.

4. The method according to claim 3, wherein Calculate the three-dimensional confidence composed of policy credibility, rule compliance, and experience similarity for the sub-plan decisions, where the policy credibility is calculated based on the action policy distribution; sort the sub-plans according to the weighted scores of the three-dimensional confidence, identify the temporal conflicts, resource conflicts, and effect conflicts between the sub-plans that exceed the preset confidence threshold, and the steps of forming a conflict-free initial treatment plan through iterative adjustment and replacement include: Calculate the three-dimensional confidence for the sub-schemes, where the three-dimensional confidence includes policy credibility, rule compliance, and experience similarity; among which the policy credibility is calculated based on the action policy distribution and an adaptive temperature parameter is introduced for distribution smoothing; the rule compliance adopts a hierarchical rule evaluation method and is calculated by weighted accumulation of the degree to which the sub-scheme meets the rules and the rule weights; the experience similarity is obtained by calculating the weighted score of the similarity between the sub-scheme and historical cases and the treatment effect; construct a factor graph containing confidence nodes and dependency edges, define a potential function representing the strength of the relationship between nodes based on the factor graph, use the variational inference method to optimize the dimensional weight coefficients of the three-dimensional confidence to obtain the optimized weights of each dimension confidence, and calculate the weighted score of the three-dimensional confidence according to the optimized weights, and sort the sub-schemes according to the weighted scores; For sub-schemes exceeding the preset confidence threshold, construct a spatio-temporal dependency graph containing treatment activity nodes and spatio-temporal constraint edges, calculate the spatio-temporal set intersection and union ratio between sub-schemes based on the spatio-temporal dependency graph to obtain temporal conflicts; calculate the ratio of the resource usage of the sub-scheme at each time point to the upper limit of the resource capacity to obtain resource conflicts; construct a treatment effect interference matrix representing the degree of mutual influence between sub-schemes to calculate effect conflicts; identify temporal conflicts, resource conflicts, and effect conflicts between sub-schemes; For the identified conflicting sub-schemes, construct the adjustment decision of the sub-scheme as a decision variable, with the goal of maximizing the weighted score of the sub-scheme, while satisfying the temporal constraint that the execution times of the sub-schemes do not overlap, the resource constraint that the total resource usage at each time point does not exceed the capacity limit, the conflict constraint that the conflicting sub-schemes cannot be executed simultaneously, and the dependency relationship constraint between sub-schemes; use the branch and bound method for calculation and solution, detect conflicts through an iterative method and add spatio-temporal overlap degree threshold constraints, resource competition degree upper limit constraints, and effect conflict degree limit constraints, and perform sub-scheme adjustment and replacement until a conflict-free scheme set is obtained; among which the branch and bound method includes designing a pruning strategy based on heuristic rules and dynamically adjusting the branching order; generate an initial treatment plan according to the conflict-free scheme set.

5. The method according to claim 1, wherein Real-time collect the temporal data of the patient's drug dosage, metabolic indicators, clinical symptoms, and adverse reactions, identify treatment features by combining the attention calculation method, and predict the treatment trajectory using bidirectional recursion; based on the patient's tolerance index, organ function index, and treatment effect index, use a multi-objective optimization algorithm to calculate the optimal dosing plan; Analyze the expert decision-making characteristics in historical successful cases, and the steps to establish a rule base for treatment timing, medication strategies, and complication prevention and control to form decision criteria include: Collect the time-series data of the patient's medication dosage, metabolic indicators, clinical symptoms, and adverse reactions; process the time-series data using a multi-head attention mechanism, extract high-order feature representations by calculating the attention weight matrix between features and applying a graph attention network; construct a bidirectional recurrent neural network based on the high-order feature representations, design a state transition probability matrix by combining bidirectional feature representations and introducing a time attention mechanism to calculate the change trends of the medication dosage, metabolic indicators, clinical symptoms, and adverse reactions, and predict the state distribution at a preset time point to obtain the treatment trajectory; Establish an evaluation system that includes patient tolerance indicators, organ function indicators, and treatment effect indicators. The patient tolerance indicators include the incidence of adverse reactions and the degree of symptom burden. The organ function indicators include liver and kidney function indicators, cardiopulmonary function parameters, and hematopoietic function indicators. The treatment effect indicators include the disease control rate, the degree of symptom improvement, and the evaluation of survival benefit; construct a multi-objective optimization function based on the evaluation system, map the patient tolerance indicators, organ function indicators, and treatment effect indicators to the optimization objectives of maximizing the treatment effect, minimizing adverse reactions, and maintaining stable organ function respectively, and use the non-dominated sorting genetic algorithm to solve for the optimal dosing plan with the drug dosage range, dosing interval, and safety indicators as the constraint conditions; Analyze historical successful cases, extract case feature vectors that include patient baseline characteristics, treatment process characteristics, and outcome evaluation characteristics, and construct expert decision-making features based on the decision time point, decision basis elements, and intervention measure selection; establish a rule base that consists of treatment timing rules, medication strategy rules, and complication prevention and control rules according to the expert decision-making features. The treatment timing rules are based on the disease progression state, the degree of adverse reactions, and the organ function state. The medication strategy rules include dosage adjustment principles, combination medication strategies, and treatment course arrangement plans. The complication prevention and control rules include risk warning indicators, preventive intervention measures, and emergency treatment plans; use an analytic hierarchy decision tree to integrate the rules in the rule base, set the rule trigger conditions and rule priorities, and form decision-making criteria.

6. The method according to claim 1, characterized in that The steps of dynamically optimizing the initial treatment plan by integrating the treatment trajectory prediction results, the optimal dosing plan, and the decision-making criteria to obtain the final treatment plan include: Construct a conditional variational flow field to map the treatment trajectory prediction results, the optimal dosing plan, and the decision-making criteria to a shared latent space, perform data alignment by establishing the bidirectional optical flow constraints of the forward flow field and the reverse flow field, and optimize based on the variational lower bound of the reconstruction error and the KL divergence to obtain the aligned data stream; Construct a three - level cascaded memory structure based on the aligned data stream, establish an association with the initial treatment plan through the multi - head self - attention mechanism of the immediate memory layer, capture the execution status of the initial treatment plan using the time - aware attention mechanism of the short - term memory layer, store the historical feedback information of the initial treatment plan in the long - term memory layer, and perform adaptive weight fusion on the three - level memories through a meta - controller to obtain the dynamic optimization features of the initial treatment plan; construct a stochastic differential equation based on the dynamic optimization features, perform path sampling on the initial treatment plan in the feature space, calculate the path probability distribution in combination with the historical feedback information, and optimize the path probability distribution through the variational inference method. At the same time, dynamically adjust the optimization parameters based on the execution effect to obtain the optimized path of the initial treatment plan; Construct a two - layer medical rule verification network. The two - layer medical rule verification network includes a rule inference layer and an execution optimization layer. Among them, the rule inference layer uses a conditional random field to model the correlation between medical rules and generates a rule verification sequence. The execution optimization layer dynamically executes the rule verification sequence based on a deep reinforcement learning framework, and verifies and adjusts the optimized path to obtain a set of optimized plans that meet medical rules; construct an optimal transport problem based on the Wasserstein distance on the Riemannian manifold, define a local metric through the geodesic distance, and perform projection optimization fusion on the set of optimized plans and the initial treatment plan to generate a final treatment plan.

7. The method according to claim 6, wherein Construct a conditional variational flow field to map the treatment trajectory prediction result, the optimal drug administration plan, and the decision criterion to a shared latent space. The steps to obtain the aligned data stream by establishing a two - way optical flow constraint between the forward flow field and the reverse flow field and optimizing based on the variational lower bound of the reconstruction error and the KL divergence include: Adopt a conditional variational flow field to construct a mapping framework. Map the treatment trajectory prediction result, the optimal drug administration plan, and the decision criterion to a shared latent space through an encoder network with a multi - layer perceptron structure. The encoder network generates mean and variance parameters through re - parameterization. The conditional variational flow field introduces a conditional prior network to generate prior distribution parameters based on the modality identifier, and optimizes the learning of the shared latent space representation by maximizing the variational lower bound; construct an inter - modal two - way mapping network based on the latent representation of the shared latent space, introduce the cyclic consistency constraint of the Wasserstein metric, calculate the Wasserstein distance by iteratively optimizing the transport matrix through the Sinkhorn algorithm, design an energy function network including a feature encoding layer and multiple non - linear mappings. The energy function network inputs feature pairs of different modalities, measures their dual consistency by calculating the energy scores of the feature pairs, and uses the energy function network as a discriminator to obtain the initial feature distribution through an adversarial training mechanism; Construct a forward flow field and a reverse flow field for the initial feature distribution. The forward flow field and the reverse flow field respectively include a mean component, a variance component, and a random perturbation term. Establish a composite mapping relationship based on the forward flow field and the reverse flow field. The composite mapping relationship includes the reverse flow field mapping after the action of the forward flow field and the forward flow field mapping after the action of the reverse flow field. Construct the composite mapping relationship as a bidirectional optical flow constraint, and construct a composite mapping loss based on the bidirectional optical flow constraint. Introduce a spectral regularization mechanism based on Lipschitz constraint. Calculate the spectral norm of network parameters by the power iteration method and adopt an adaptive clipping strategy to adjust the parameter scale. Construct a variational lower bound with the reconstruction error and KL divergence and integrate it with the spectral regularization term into a variational optimization objective to constrain the composite mapping loss. Design a cross-modal attention module to calculate the correlation weights for the constrained feature distribution, and fuse different modal features based on the correlation weights. Integrate the fused feature distribution using the Bayesian deep ensemble method with uncertainty parameter weighting to obtain the aligned data stream.

8. A gynecological disease diagnosis and treatment system integrating electronic medical records, for implementing the method described in any one of the foregoing claims 1-7, characterized in that, Comprising: The first unit is used to collect the multi-modal diagnosis and treatment data of the patient, including the time-series physiological index sequence, the clinical manifestation feature set, the medical image set, and the test report set. Perform dimensionality reduction and denoising processing through an autoencoder network, construct the feature mapping relationship between modalities, extract the common and individual features of the data, and establish a personalized medical database for the patient. The second unit is used to construct a dynamic causal network including disease development, treatment intervention, and complication prevention and control based on the personalized medical database. The dynamic causal network analyzes the interaction effects between different treatment methods through a time-series causal reasoning algorithm. Based on the dynamic causal network, deploy a drug treatment agent, a surgical plan agent, and a nursing plan agent. Use the deep reinforcement learning method combined with professional domain rules and historical case features, adopt a multi-level reasoning mechanism to form sub-schemes, perform feasibility analysis on the sub-schemes based on confidence, and perform conflict resolution to generate an initial treatment plan. The third unit is used to collect the time-series data of the patient's drug dosage, metabolic indicators, clinical symptoms, and adverse reactions in real time. Identify treatment features by combining the attention calculation method, and predict the treatment trajectory using bidirectional recursion. Calculate the optimal drug administration plan using a multi-objective optimization algorithm based on the patient's tolerance index, organ function index, and treatment effect index. Analyze the expert decision-making features in historical successful cases, establish a rule base for treatment timing, drug use strategy, and complication prevention and control, and form decision criteria. Integrate the treatment trajectory prediction results, the optimal drug administration plan, and the decision criteria to dynamically optimize the initial treatment plan to obtain the final treatment plan.

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

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

Images