All-specialty collaborative diagnosis and treatment method and system based on medical intelligent agent middleware

Abstract

The application provides a full-specialty collaborative diagnosis and treatment method and system based on medical intelligent agent middleware, relates to the technical field of medical intelligent agents, and comprises the following steps: constructing a basic medical knowledge graph by using reinforcement learning, converting expert diagnosis and treatment experience into a diagnosis and treatment probability distribution by adopting a personalized treatment decision model, and forming an enhanced medical knowledge graph; calculating a matching probability of a patient symptom feature vector and a disease to generate a preliminary diagnosis decision, and selecting and adjusting a treatment scheme by reinforcement learning. The application can realize collaborative diagnosis and treatment of general practice and specialty, and improve the accuracy of medical decision and the efficiency of diagnosis and treatment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical agents, and particularly relates to a full-specialty collaborative diagnosis and treatment method and system based on a medical agent middleware. BACKGROUND

[0002] With the continuous development of medical technology, the scale and complexity of medical data are growing, and traditional medical diagnosis methods have been difficult to meet the needs of modern medical services. In recent years, artificial intelligence technology has been widely applied in the medical field, and medical intelligent diagnosis systems based on knowledge graphs, reinforcement learning and Monte Carlo tree search have gradually become a research hotspot. These technologies have brought revolutionary changes to the medical industry, and can assist doctors in making more accurate diagnoses and personalized treatment decisions. The full-specialty collaborative diagnosis and treatment mode is an important organizational form in medical practice, which integrates the knowledge and experience of doctors of different specialties to provide comprehensive and professional medical services for patients. However, existing medical intelligent systems still have some technical challenges in supporting full-specialty collaborative diagnosis and treatment.

[0003] Traditional medical intelligent systems lack effective medical text information extraction and knowledge integration mechanisms, and are difficult to accurately identify professional entities and complex relationships in medical texts, resulting in low-quality medical knowledge graphs that cannot fully reflect the diagnosis and treatment knowledge and clinical experience of medical experts.

[0004] Existing medical decision systems usually use static knowledge bases or fixed rules for diagnostic reasoning, lack personalized treatment decision models and dynamic learning capabilities, and cannot effectively simulate the diagnosis and treatment decision-making process of doctors, making it difficult to convert expert diagnosis and treatment experience into computable diagnosis and treatment probability distribution, thereby limiting the adaptability and accuracy of the system.

[0005] Existing medical intelligent systems lack effective technical support for full-specialty collaborative diagnosis and treatment, and cannot realize the automatic fusion and dynamic collaboration of different professional knowledge, resulting in diagnosis results that are often biased towards a single specialty field, making it difficult to provide patients with comprehensive and personalized diagnosis and treatment plans, especially for complex cases and multi-system disease handling capabilities. SUMMARY

[0006] The embodiments of the present application provide a full-specialty collaborative diagnosis and treatment method and system based on a medical agent middleware, which can solve the problems in the prior art.

[0007] In a first aspect, the embodiments of the present application provide a full-specialty collaborative diagnosis and treatment method based on a medical agent middleware, comprising:

[0008] The medical text is analyzed by reinforcement learning to extract entity information and relation information, the extracted medical text information is mapped to a unified semantic space, and the medical text information is analyzed based on a Monte Carlo tree search to construct a basic medical knowledge graph;

[0009] A personalized treatment decision model is used to simulate the doctor's diagnosis and treatment decision-making process in the basic medical knowledge graph, expert diagnosis and treatment experience is converted into a diagnosis and treatment probability distribution through reinforcement learning, and the diagnosis and treatment probability distribution is dynamically integrated into the basic medical knowledge graph using the Monte Carlo tree search to form an enhanced medical knowledge graph;

[0010] Based on the enhanced medical knowledge graph, the personalized treatment decision model is used to analyze the symptom association information in the enhanced medical knowledge graph to establish a symptom feature vector reflecting the characteristics of the patient's symptoms;

[0011] The symptom feature vector is input into the enhanced medical knowledge graph, and the matching probability of the patient's symptoms and diseases is calculated through the Monte Carlo tree search, and the current symptoms of the patient are diagnosed and analyzed according to the matching probability to generate a preliminary diagnosis decision;

[0012] According to the disease type in the preliminary diagnosis decision, reinforcement learning and the Monte Carlo tree search are used to select the corresponding treatment scheme from the enhanced medical knowledge graph, and the treatment scheme is continuously adjusted through reinforcement learning to improve the adaptability, and finally an optimal treatment scheme is generated.

[0013] The medical text is analyzed by reinforcement learning to extract entity information and relation information, the extracted medical text information is mapped to a unified semantic space, and the medical text information is analyzed based on a Monte Carlo tree search to construct a basic medical knowledge graph, comprising:

[0014] The medical text is analyzed by reinforcement learning to construct a state vector of medical entity information, a decision vector for entity extraction is determined based on the state vector, and a feature distribution is generated by dynamically adjusting the weights of the state vector and the decision vector;

[0015] The feature distribution is mapped to a unified semantic space to generate a neural feature vector, a rule feature vector is obtained by bidirectional feature transformation of the neural feature vector, and a fused feature vector is generated by weight optimization of the neural feature vector and the rule feature vector based on the dynamic weight adjustment mechanism;

[0016] The fused feature vector is analyzed by Monte Carlo tree search, the entity information in the fused feature vector is represented by constructing a search tree node, the node score is calculated based on the dynamic weight adjustment mechanism, the node score is used to guide the expansion direction of the search tree, and the association relationship between entities is determined according to the expansion result of the search tree.

[0017] convert the fusion feature vector into node information and edge information of a knowledge graph based on an analysis result of the Monte Carlo tree search, and construct a basic medical knowledge graph according to the node information and the edge information.

[0018] simulate a doctor's diagnosis and treatment decision-making process in the basic medical knowledge graph by using a personalized treatment decision-making model, convert expert diagnosis and treatment experience into a diagnosis and treatment probability distribution through reinforcement learning, and dynamically integrate the diagnosis and treatment probability distribution into the basic medical knowledge graph by using the Monte Carlo tree search to form an enhanced medical knowledge graph, including:

[0019] model the doctor's diagnosis and treatment decision-making process in the basic medical knowledge graph by using a personalized treatment decision-making model, construct a state vector represented by patient clinical features, and set an action vector of the doctor's diagnosis and treatment plan based on the state vector;

[0020] calculate a combined feature of the state vector and the action vector by using reinforcement learning, iteratively optimize the combined feature, and map the doctor's diagnosis and treatment experience into a diagnosis and treatment probability distribution based on the personalized treatment decision-making model;

[0021] input the diagnosis and treatment probability distribution into a Monte Carlo search tree structure to determine a node expansion direction, calculate a visit frequency and a cumulative value of an expanded tree node based on the node expansion direction, and select an optimal search path according to the visit frequency and the cumulative value;

[0022] extract a diagnosis and treatment path with credibility from the Monte Carlo search tree structure according to the optimal search path, dynamically write probability distribution information of the diagnosis and treatment path into the basic medical knowledge graph, and update edge weights of diagnosis and treatment relationships in the basic medical knowledge graph by using an adaptive weight method to form an enhanced medical knowledge graph.

[0023] input the diagnosis and treatment probability distribution into a Monte Carlo search tree structure to determine a node expansion direction, calculate a visit frequency and a cumulative value of an expanded tree node based on the node expansion direction, and select an optimal search path according to the visit frequency and the cumulative value, including:

[0024] input the diagnosis and treatment probability distribution into a Monte Carlo search tree structure, construct a Lorentz system including position coordinates, velocity changes, and acceleration components based on the diagnosis and treatment probability distribution, map the diagnosis and treatment probability distribution to a three-dimensional dynamic space based on the Lorentz system, obtain an evolution trajectory in the three-dimensional dynamic space through iterative calculation, and determine an initial node expansion direction according to the evolution trajectory;

[0025] Calculate a Lyapunov exponent sequence based on the initial node expansion direction, combine the Lyapunov exponent sequence and the diagnosis and treatment probability distribution by exponential weighting, construct an optimized node expansion probability, and expand the tree node based on the optimized node expansion probability as a selection basis;

[0026] Calculate the access frequency of the expanded tree node, construct a bifurcation metric function based on the node depth and the number of branches, and use the weighted product of the access frequency and the bifurcation metric function as the node weight to identify the decision node in the search path through the node weight;

[0027] Calculate the cumulative value of the tree node, extract the feature vector group corresponding to the evolution trajectory, use the Lyapunov exponent corresponding to the unstable mode of the feature vector group as a weight coefficient, and update the cumulative value by weighting, and select the optimal search path according to the access frequency and the cumulative value.

[0028] Calculate the matching probability of the patient's symptoms and diseases through the Monte Carlo tree search, and diagnose and analyze the patient's current symptoms according to the matching probability to generate a preliminary diagnosis decision, including:

[0029] Calculate the ratio of the node value parameter and the node access count parameter of the search node in the Monte Carlo search tree, determine the node expansion score in combination with the exploration coefficient and the logarithmic term of the parent node access count, and select the symptom combination path based on the node expansion score;

[0030] Probabilistic evaluation of the symptom combination path, calculation of the evaluation score of the symptom combination, update of the node value parameter using the evaluation score, and update of the node access count parameter according to the search depth; based on the updated node value parameter, calculate the matching probability of the symptoms and diseases, optimize the matching probability in combination with the symptom weight coefficient, and determine the preliminary diagnosis decision according to the optimized matching probability.

[0031] According to the disease type in the preliminary diagnosis decision, select the corresponding treatment scheme from the enhanced medical knowledge graph by reinforcement learning and the Monte Carlo tree search, continuously adjust the treatment scheme by reinforcement learning to improve adaptability, and finally generate an optimal treatment scheme, including:

[0032] According to the disease type, obtain a set of treatment schemes from the medical knowledge graph, calculate the similarity based on the feature vectors of the disease nodes in the medical knowledge graph and the disease type, and filter to generate a candidate treatment scheme according to the similarity;

[0033] construct a biological synapse dynamics model to analyze the candidate treatment scheme, map a treatment effect change trend based on the biological synapse dynamics model through an ion concentration gradient response function, and generate a synapse weight based on the ion concentration gradient response function and a time-varying attenuation coefficient;

[0034] search and optimize the candidate treatment scheme according to the adaptive strength of the synapse weight, determine an optimization direction based on an access count of a search node, and update a treatment parameter of the candidate treatment scheme by using the optimization direction and the adaptive strength of the synapse weight;

[0035] calculate a treatment effect evaluation value of the candidate treatment scheme based on the updated synapse weight, inherit the adaptive feature of the synapse weight according to the treatment effect evaluation value, and select an optimal treatment scheme from the candidate treatment scheme.

[0036] construct a biological synapse dynamics model to analyze the candidate treatment scheme, map a treatment effect change trend based on the biological synapse dynamics model through an ion concentration gradient response function, and generate a synapse weight based on the ion concentration gradient response function and a time-varying attenuation coefficient;

[0037] construct a biological synapse dynamics model, the biological synapse dynamics model receives a candidate treatment scheme, calculates an initial response feature based on a treatment parameter of the candidate treatment scheme, and the initial response feature reflects the action strength of a treatment intervention;

[0038] generate an ion concentration gradient response function, the ion concentration gradient response function calculates a change trend of an ion concentration inside and outside a cell based on the initial response feature, the change trend is adjusted by a time constant, and a treatment effect mapping of the candidate treatment scheme is established.

[0039] In a second aspect of the embodiment of the present application, a full-specialty collaborative diagnosis and treatment system based on a medical intelligent agent middleware is provided, which comprises:

[0040] A first unit is configured to perform entity recognition and relationship extraction on medical texts by using reinforcement learning, map extracted medical text information to a unified semantic space, and analyze the medical text information to construct a basic medical knowledge graph based on Monte Carlo tree search;

[0041] A second unit is configured to simulate a doctor diagnosis and treatment decision-making process in the basic medical knowledge graph by using a personalized treatment decision-making model, convert expert diagnosis and treatment experience into a diagnosis and treatment probability distribution by reinforcement learning, and dynamically integrate the diagnosis and treatment probability distribution into the basic medical knowledge graph by using the Monte Carlo tree search to form an enhanced medical knowledge graph;

[0042] The third unit is configured to analyze the symptom correlation information in the enhanced medical knowledge graph based on the enhanced medical knowledge graph and the personalized treatment decision model, and establish a symptom feature vector reflecting the patient's disease characteristics.

[0043] The fourth unit is configured to input the symptom feature vector into the enhanced medical knowledge graph, calculate the matching probability of the patient's symptoms and diseases through the Monte Carlo tree search, perform diagnostic analysis on the patient's current symptoms according to the matching probability, and generate a preliminary diagnosis decision.

[0044] The fifth unit is configured to select a corresponding treatment scheme from the enhanced medical knowledge graph based on the disease type in the preliminary diagnosis decision and the reinforcement learning and Monte Carlo tree search, continuously adjust the treatment scheme through reinforcement learning to improve adaptability, and finally generate an optimal treatment scheme.

[0045] In a third aspect, the embodiment of the present application provides an electronic device, comprising:

[0046] a processor;

[0047] a memory for storing processor-executable instructions;

[0048] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0049] In a fourth aspect, the embodiment of the present application provides a computer-readable storage medium having computer program instructions stored thereon, wherein the computer program instructions are executed by a processor to implement the method described above.

[0050] The present application has the following beneficial effects:

[0051] The present application constructs an enhanced medical knowledge graph through reinforcement learning and Monte Carlo tree search, effectively integrates expert diagnosis and treatment experience into the knowledge system, realizes the deep integration of general and specialized medical knowledge, and improves the accuracy and comprehensiveness of disease diagnosis.

[0052] The present application analyzes the patient's symptom characteristics based on a personalized treatment decision model, calculates the matching probability of symptoms and diseases based on an enhanced medical knowledge graph, and can provide customized diagnosis and treatment schemes for the unique disease characteristics of patients, effectively solving the problem of insufficient individualized treatment in traditional diagnosis and treatment.

[0053] The present application continuously adjusts the treatment scheme through reinforcement learning, realizes the continuous improvement of diagnosis and treatment schemes through dynamic learning and adaptive optimization, continuously improves the quality of medical services, and at the same time relieves the pressure of medical resource shortage, provides technical support for hierarchical diagnosis and treatment and remote medical treatment. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 A flowchart of a full-specialty collaborative diagnosis and treatment method based on a medical intelligent agent middleware according to an embodiment of the present application is shown in FIG. 1.

[0055] Figure 2 A flowchart of constructing an enhanced medical knowledge graph based on a personalized treatment decision model according to an embodiment of the present application is shown in FIG. 2.

[0056] Figure 3 A flowchart of generating an optimal treatment plan based on reinforcement learning and Monte Carlo tree search according to an embodiment of the present application is shown in FIG. 3. DETAILED DESCRIPTION

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in connection with the drawings of the embodiments of the present application. Obviously, the described embodiments are only a part but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0058] The technical solutions of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes can not be described in some embodiments.

[0059] Figure 1 A flowchart of a full-specialty collaborative diagnosis and treatment method based on a medical intelligent agent middleware according to an embodiment of the present application is shown in FIG. 1. Figure 1 The method includes:

[0060] Performing entity recognition and relationship extraction on medical text using reinforcement learning, mapping the extracted medical text information to a unified semantic space, and analyzing the medical text information based on Monte Carlo tree search to construct a basic medical knowledge graph;

[0061] Simulating the doctor diagnosis and treatment decision process in the basic medical knowledge graph using a personalized treatment decision model, converting expert diagnosis and treatment experience into a diagnosis and treatment probability distribution through reinforcement learning, and dynamically integrating the diagnosis and treatment probability distribution into the basic medical knowledge graph using the Monte Carlo tree search to form an enhanced medical knowledge graph;

[0062] Based on the enhanced medical knowledge graph, analyzing the symptom association information in the enhanced medical knowledge graph using the personalized treatment decision model to establish a symptom feature vector reflecting the patient's disease characteristics;

[0063] Input the symptom feature vector into the enhanced medical knowledge graph, calculate the matching probability of the patient's symptoms and diseases through the Monte Carlo tree search, perform diagnostic analysis on the patient's current symptoms according to the matching probability, and generate a preliminary diagnosis decision;

[0064] According to the disease type in the preliminary diagnosis decision, a corresponding treatment scheme is selected from the enhanced medical knowledge graph by reinforcement learning and the Monte Carlo tree search, the treatment scheme is continuously adjusted by reinforcement learning to improve adaptability, and finally an optimal treatment scheme is generated.

[0065] In a medical auxiliary diagnosis system that combines a fast and comprehensive but shallow search strategy with a deep and accurate search strategy in a medical decision-making process, the fast and comprehensive but shallow search strategy is implemented through a breadth-first search algorithm. The system uses a knowledge graph as a basic data structure to store medical knowledge and diagnostic experience. The nodes in the knowledge graph represent diseases, symptoms, examination items, and other medical entities, and the edges represent the association between entities. The fast and comprehensive search strategy performs a breadth-first traversal with a limited depth on the knowledge graph, with a depth limit of 3 layers and a maximum of 20 related nodes explored per layer. During the search process, a heuristic pruning method is used to prioritize nodes based on their association strength, which is calculated based on co-occurrence frequency. For common symptoms such as "fever 38.5℃", the system first retrieves disease nodes directly related to the symptom from the knowledge graph, such as "cold", "flu", and "pneumonia", and also retrieves other symptom nodes related to these diseases, such as "cough" and "sore throat". The search results are organized in the form of a decision tree, with each decision option having a probability distribution, and the probability values are calculated based on historical diagnostic data.

[0066] The decision option probability is not ideal, and the judgment standard is set as the highest probability option being lower than 0.65 or the difference between the top two high probability options being less than 0.15. For example, for the input symptoms "headache, fever 38℃, joint pain", the system may give the decision options "flu (probability 0.55)" and "rheumatic fever (probability 0.45)", as the highest probability is lower than 0.65, the system determines that the probability is not ideal. Doctor feedback requirements are achieved through an interactive interface, and when the doctor clicks the "need expert-level analysis" button, the system records the current state and switches to the deep and accurate search mode. The symptom dynamic amplification mechanism is implemented through a recurrent neural network, which is trained on approximately 5 million medical records and can predict potential related symptoms based on known symptoms. For example, for the symptoms "rapid breathing, chest pain", the system may suggest asking "does the symptom worsen after activity" or "is there any coughing up blood". The amplification process uses a Bayesian network to calculate the conditional probability between symptoms, and selects the top 5 symptoms with the highest conditional probability for each amplification.

[0067] The deep search strategy is implemented based on the Monte Carlo tree search algorithm, which is optimized in the medical diagnosis field, with a search depth of up to 10 layers and an average sampling frequency of 1000 times per node. During the search process, a medical expert knowledge base is used as prior information, which contains about 10,000 expert rules. Each rule consists of several premise conditions and a conclusion, such as "If a patient has persistent high fever (> 39°C) for more than 3 days, accompanied by rash and conjunctival congestion, consider the possibility of Kawasaki disease increased by 30%". The deep search dynamically builds a Monte Carlo search tree during execution, with each node representing a possible diagnosis state. The value of different diagnosis paths is evaluated through simulated sampling. The node evaluation uses the upper confidence bound algorithm to balance exploration and utilization. The evaluation function considers both diagnostic accuracy (weight 0.7) and examination cost (weight 0.3). In the deep search mode, the system conducts more in-depth verification of all possible disease hypotheses, retrieves about 5000 disease records in the rare disease database, and expands the search range to all medical entities in the knowledge graph.

[0068] The mechanism of symptom amplification working with deep search is realized through feedback reinforcement learning. Whenever the system receives new symptom information, it immediately performs a re-search and updates the diagnosis path. The search results are ranked in descending order of confidence and labeled with key distinguishing features. For example, for the symptom combination "persistent headache, neck stiffness, light sensitivity", the system may give two main hypotheses "meningitis (probability 0.75)" and "migraine (probability 0.20)", and label "fever" as a key distinguishing feature. During diagnosis, the system dynamically adjusts the weight of each symptom, with important symptoms ranging from 0.6 to 0.9 and secondary symptoms ranging from 0.1 to 0.5. The weight value is optimized through a reinforcement learning algorithm, with the reward function designed to improve the accuracy of diagnosis.

[0069] The threshold for switching between the two search strategies is determined adaptively by the system based on the relationship between success rate and complexity in historical diagnosis data. The switching threshold is dynamically adjusted as the diagnosis data accumulates, with an initial setting of triggering the switch when the decision probability is below 0.65. After the deep search is completed, the system generates a comprehensive diagnosis report, including disease hypotheses, evidence support, and recommended examination items. For the diagnosis hypothesis "type 2 diabetes", the system may list supporting evidence such as "fasting blood glucose 7.8 mmol / L", "thirst and polydipsia", "polyuria", etc., and recommend a glycosylated hemoglobin test. The diagnosis report adopts a hierarchical structure, with main hypotheses displayed on the top layer and auxiliary diagnosis information folded in the lower layer. Doctors can expand and view details as needed.

[0070] With a distributed computing architecture, the shallow search is executed on the edge server with a response time below 200 milliseconds, and the deep search is executed on the central server cluster with an average response time of 2-5 seconds. The knowledge graph storage uses a graph database, containing about 300,000 medical entity nodes and 1 million relationship edges. Index optimization of the graph database controls the single-hop query delay within 10 milliseconds and the multi-hop query (3 hops) delay within 50 milliseconds. The system uses an incremental learning mechanism, updating model parameters after processing 1000 new cases to ensure continuous improvement of diagnostic ability. Performance tests show that the system has a diagnosis accuracy rate of 92% for common diseases and 75% for rare diseases, significantly higher than that of traditional single search strategy diagnosis systems.

[0071] In an actual application case, a patient presents symptoms of "abdominal pain, nausea, abdominal distension", the system first performs a fast and comprehensive search and gives decision options of "gastritis (probability 0.40)", "gastroenteritis (probability 0.35)", and "cholecystitis (probability 0.20)". Since the highest probability option is below the threshold of 0.65, the system prompts that in-depth analysis may be needed. After the doctor confirms, the system starts symptom amplification and asks for key distinguishing information such as "abdominal pain location", "whether accompanied by fever", "pain nature", etc. After obtaining the new information "right upper abdominal pain, pain aggravated during deep breathing, body temperature 38.2°C", the system performs a deep search and the analysis result shows "acute cholecystitis (probability 0.85)", while recommending an "abdominal ultrasound" examination to confirm the diagnosis. The examination result confirms the existence of gallbladder wall thickening and gallstones, verifying the accuracy of the system's diagnosis. The entire diagnosis process takes about 5 minutes, about 60% more efficient than traditional methods.

[0072] The system also contains a self-evaluation and error correction mechanism, which regularly analyzes misdiagnosis cases and updates the knowledge base. Through a real case data set (containing 500,000 diagnosis records) after anonymization processing, the system's diagnosis accuracy rate in complex multi-symptom cases is improved by 17 percentage points compared to single search strategy, and its identification ability for rare diseases is improved by 23 percentage points. Experimental results show that the combination of fast and comprehensive but shallow search and deep search can effectively balance the efficiency and accuracy of diagnosis, and is particularly suitable for the auxiliary diagnosis needs in clinical first-line medical scenarios.

[0073] In an alternative embodiment, reinforcement learning is used for entity recognition and relation extraction of medical text, and the extracted medical text information is mapped to a unified semantic space, and the medical text information is analyzed based on Monte Carlo tree search to construct a basic medical knowledge graph, including:

[0074] The medical text is analyzed by using reinforcement learning, a state vector of medical entity information is constructed, a decision vector of entity extraction is determined based on the state vector, and a feature distribution is generated by dynamically adjusting the state vector and the decision vector through a weight adjustment mechanism;

[0075] The neural feature vector is generated by mapping the feature distribution to a unified semantic space, the rule feature vector is obtained by bidirectional feature transformation of the neural feature vector, and the fusion feature vector is generated by weight optimization of the neural feature vector and the rule feature vector based on the dynamic weight adjustment mechanism;

[0076] The fusion feature vector is analyzed by using Monte Carlo tree search, the entity information in the fusion feature vector is represented by constructing a search tree node, the node score is calculated based on the dynamic weight adjustment mechanism, the node score is used to guide the expansion direction of the search tree, and the association relationship between entities is determined according to the expansion result of the search tree;

[0077] The fusion feature vector is converted into node information and edge information of a knowledge graph based on the analysis result of the Monte Carlo tree search, and a basic medical knowledge graph is constructed according to the node information and the edge information.

[0078] A medical knowledge graph is constructed based on reinforcement learning and Monte Carlo tree search. The reinforcement learning model uses a double-layer neural network architecture to process medical text. The input layer receives the word embedding representation of the medical text, with a word embedding dimension of 300, generated by a pre-trained medical domain word vector model. The middle layer is a bidirectional long short-term memory network, with 512 hidden layer neurons, used to capture text sequence features. The output layer is a fully connected layer, with an output dimension equal to the number of entity types plus one. For the medical text "patient blood glucose is 8.6 mmol / L, glycosylated hemoglobin 7.2%", the feature matrix after word embedding has a dimension of.

[0079] When constructing the state vector of medical entity information, four types of features are extracted for each word position: word embedding features (dimension 300), position encoding features (dimension 50), character-level features (dimension 100), and domain features (dimension 50). The state vector is formed by concatenating these four types of features, with a total dimension of 500. For example, the state vector of the word "blood glucose" contains its word embedding representation, position encoding in the sentence (relative position 2), character-level CNN extracted features, and medical domain features (labeled as test indicator class). The decision vector represents the action space of entity extraction, containing B-I-O labels and entity type labels. The decision vector has a dimension equal to (3×number of entity types+1), and for a task containing 8 medical entities, the decision vector has a dimension of 25. For the word "blood glucose", its decision vector is in one-hot encoding form, with 1 in the B-test indicator position and 0 in the remaining positions.

[0080] The dynamic weight adjustment mechanism is implemented based on the policy gradient algorithm in reinforcement learning. The reward function is composed of three parts: entity recognition accuracy (weight 0.4), entity boundary accuracy (weight 0.3), and entity type accuracy (weight 0.3). For correctly identified entities, a +1 reward is given; for incorrectly identified entities, a -0.5 reward is given; for missed entities, a -0.3 reward is given. During the training process, the policy network parameters are optimized by gradient ascent method, and the learning rate is initially set to 0.001, decaying by 10% every 1000 batches. For the above medical text, "blood sugar" and "glycated hemoglobin" are correctly identified as test indicator class entities, obtaining +2 rewards; "8.6 mmol / L" and "7.2%" are correctly identified as test values, obtaining +2 rewards. After feature distribution generation, it passes through a fully connected layer and a softmax layer to output entity label probability distribution, with a dimension equal to the decision vector dimension and a value range of [0, 1], and a sum of 1.

[0081] The feature distribution is mapped to the unified semantic space through the self-attention mechanism. The input of the self-attention layer is a sequence of state vectors, and the output is a context-sensitive neural feature vector. The number of attention heads is set to 8, each with a dimension of 64, and the output dimension is 512. The calculation process includes: transforming the input features into query, key, and value matrices; calculating the dot product of the query and the key to obtain the attention score; performing softmax normalization on the score; and weighting and summing the value matrix to obtain the context representation. For the "blood sugar" word, its attention score is higher (0.75) than "8.6 mmol / L", indicating that the semantic association between the two is strong. The neural feature vector is normalized by layer and connected by residual, and the output dimension remains unchanged at 512.

[0082] The neural feature vector is transformed into a rule feature vector in both directions, and the neural-to-rule mapping and rule-to-neural feedback are performed. The neural-to-rule mapping uses distance metric learning to map the neural feature vector to a predefined medical ontology rule space. In the specific implementation, a medical ontology knowledge base is pre-constructed, containing 5000 medical concepts and their relationships. Each concept is represented by a rule feature vector with a dimension of 200. The matching degree between the neural feature vector and the rule feature vector is calculated by cosine similarity, and the threshold is set to 0.8. If the threshold is exceeded, it is considered a successful match. For example, the neural feature vector of "blood sugar" has a similarity of 0.92 with the rule feature vector of the "blood sugar detection" concept in the ontology, and the match is successful. The rule-to-neural feedback is implemented through residual connection, which weights and fuses the rule feature vector with the original neural feature vector, and the weight is adjusted adaptively according to the matching degree.

[0083] The fusion feature vector is obtained by weighting and summing the neural feature vector and the rule feature vector through a dynamic weight adjustment mechanism. The weight coefficient is dynamically adjusted according to the feature quality, and the calculation formula is: the neural feature weight is equal to the basic weight 0.5 plus the neural network confidence multiplied by 0.5; the rule feature weight is equal to 1 minus the neural feature weight. For high-confidence recognized entities, such as "blood sugar" with a confidence of 0.9, the neural feature weight is 0.5+0.9×0.5=0.95, and the rule feature weight is 0.05; for low-confidence entities, the rule feature weight is higher. The dimension of the fused feature vector is 512, which contains the context semantic information extracted by the neural network and the standardized concept information of the medical ontology.

[0084] When the Monte Carlo tree search analyzes the fused feature vector, the search tree is constructed with entities as nodes and relationships as edges. The root node is the document topic entity, and each node contains entity name, type, feature vector, and confidence, etc. The node score calculation formula is: confidence plus exploration constant (set to 1.414) multiplied by the square root of the logarithm of the parent node visit times divided by the square root of the current node visit times. For example, the "blood sugar" node confidence is 0.9, the parent node visit times is 50, and the current node visit times is 10, the score is 0.9+1.414×(ln(50) / 10) 0.5 =0.9+1.414×0.402=1.47. The search tree expansion follows a four-step cycle of selection, expansion, simulation, and backtracking. The selection stage selects the most potential path based on the node score; the expansion stage adds child nodes to the selected node, which are other entities related to the current entity; the simulation stage randomly executes from the new node to the termination state; the backtracking stage updates the node statistical information by propagating the simulation results in reverse. The expanded child nodes of the "blood sugar" node include "8.6mmol / L" (relationship "value"), "diabetes" (relationship "indication"), etc. After 1000 iterations, the path with the most visits is determined as the entity relationship path.

[0085] When determining the association relationship between entities, the access frequency threshold is set to 5% of the total iteration times, and nodes with access frequency exceeding the threshold are considered to have an association, and the association strength is equal to the common access times divided by the total iteration times. For example, the "blood sugar" and "8.6mmol / L" have a common access times of 850, and the association strength is 0.85; the "blood sugar" and "diabetes" have a common access times of 720, and the association strength is 0.72. When converting the fused feature vector into node information and edge information of the knowledge graph, the node information includes entity ID, name, type, feature vector, and attribute value, etc.; the edge information includes starting node ID, ending node ID, relationship type, relationship strength, and confidence, etc.

[0086] The node ID is generated by a hash algorithm to ensure uniqueness; the relationship type is extracted from the text by pattern matching, including 10 basic relationships such as "part-whole", "cause and effect", "indication", etc. The final constructed knowledge graph is stored in the RDF triple format, which is convenient for query and reasoning. The medical knowledge graph constructed based on the above method has an entity recognition accuracy of 92.3% and a relationship extraction accuracy of 85.7%, which is significantly better than traditional methods. The knowledge graph contains about 1 million medical entity nodes and 3 million relationship edges, covering 8 major medical concepts such as diseases, symptoms, drugs, and tests, which can effectively support clinical decision-making and medical research.

[0087] In an optional implementation, a personalized treatment decision model is used to simulate the doctor's diagnosis and treatment decision-making process in the basic medical knowledge graph, expert diagnosis and treatment experience is converted into diagnosis and treatment probability distribution through reinforcement learning, and the diagnosis and treatment probability distribution is dynamically integrated into the basic medical knowledge graph using Monte Carlo tree search to form an enhanced medical knowledge graph, including:

[0088] A personalized treatment decision model is used to model the doctor's diagnosis and treatment decision-making process in the basic medical knowledge graph, a state vector represented by the patient's clinical characteristics is constructed, and an action vector of the doctor's diagnosis and treatment plan is set based on the state vector;

[0089] The combined features of the state vector and the action vector are calculated using reinforcement learning, the combined features are iteratively optimized, and the doctor's diagnosis and treatment experience is mapped to a diagnosis and treatment probability distribution based on the personalized treatment decision model;

[0090] The diagnosis and treatment probability distribution is input into a Monte Carlo search tree structure to determine the node expansion direction, the access frequency and cumulative value of the expanded tree nodes are calculated based on the node expansion direction, and the optimal search path is selected according to the access frequency and the cumulative value;

[0091] According to the optimal search path, a diagnosis and treatment path with credibility is extracted from the Monte Carlo search tree structure, the probability distribution information of the diagnosis and treatment path is dynamically written into the basic medical knowledge graph, and the edge weight of the diagnosis and treatment relationship in the basic medical knowledge graph is updated using an adaptive weight method to form an enhanced medical knowledge graph.

[0092] As shown in Figure 2 The method comprises:

[0093] In the process of constructing the personalized treatment decision model, the doctor's diagnosis and treatment decisions in the basic medical knowledge graph are modeled. First, the patient's clinical characteristics are extracted to construct a state vector. The state vector contains five key features: basic information features (age, gender, body mass index, etc.), symptom features (fever level, pain location, etc.), laboratory index features (blood routine, biochemical index, etc.), imaging features (CT, MRI, etc.), and medical history features (previous medical history, family history, etc.). Each feature is assigned a weight value, which is determined by expert scoring and ranges from 0 to 1. In practical applications, for a patient with type 2 diabetes, the state vector example is: [age: 58 years old (weight 0.7), gender: female (weight 0.3), BMI: 26.8 (weight 0.8), fasting blood glucose: 8.6 mmol / L (weight 0.9), glycosylated hemoglobin: 7.8% (weight 0.95), urinary protein: negative (weight 0.6), retinopathy: mild (weight 0.7), kidney function: normal (weight 0.8), cardiovascular disease: none (weight 0.7)].

[0094] Based on the state vector, the doctor's diagnosis and treatment plan is set up as an action vector, which contains four types of decision elements: drug treatment (drug type, dosage, administration method, etc.), non-drug intervention (diet adjustment, exercise prescription, etc.), follow-up arrangement (follow-up frequency, monitoring indicators, etc.), and health education (knowledge popularization, skill training, etc.). Taking the above-mentioned diabetic patient as an example, the action vector example is: [drug: metformin (weight 0.9), dosage: 850 mg twice a day (weight 0.8), diet: low-carb diet (weight 0.7), exercise: 150 minutes of moderate-intensity aerobic exercise per week (weight 0.6), blood glucose monitoring: 3 times a week (weight 0.8), review of glycosylated hemoglobin every three months (weight 0.9)]. The initial values of each element in the action vector are obtained by statistical analysis of historical diagnosis and treatment data in the knowledge graph, and the most frequently selected scheme is chosen as the initial value.

[0095] The deep Q-network architecture is used to reinforce learning to calculate the combined features of the state vector and the action vector. The number of nodes in the network input layer is equal to the dimension of the state vector (9 in the above example), the hidden layer uses a two-layer structure, the number of nodes in each layer is twice the number of input layer nodes, and the number of nodes in the output layer is equal to the total number of action vector values (in the above example, 5 kinds of drug selection, 3 kinds of dosage, 2 kinds of diet plan, 2 kinds of exercise plan, 3 kinds of monitoring frequency, and 3 kinds of review period, totaling 5×3×2×2×3×3=540 combinations). The network training data comes from two parts: one is the existing diagnosis and treatment records in the knowledge graph, and the other is the diagnosis and treatment decisions of medical experts on simulated cases. The training data format is a three-tuple of (state vector, action vector, reward value), and the reward value is determined according to the treatment effect score, ranging from -100 to 100. For example, a diabetic patient's glycosylated hemoglobin decreased by 1.2 percentage points after treatment, with no adverse reactions and good compliance, and the reward value of this plan is set to 80.

[0096] During the iterative optimization of the combined features, the training rounds are set to 10,000, and 100 training samples are randomly selected for batch update each round. The initial learning rate is set to 0.01, and it is reduced by 10% every 1,000 rounds. The ε-greedy strategy is used for exploration and utilization balance, with an initial ε value of 0.9 gradually decreasing to 0.1, indicating a gradual transition from a high proportion of random exploration to mainly using known optimal strategies. The optimization goal is to maximize the cumulative reward value, which is a comprehensive indicator of patient health benefits and treatment costs.

[0097] After training, the deep Q-network can output the Q-value of each action combination according to the input patient state vector, and a higher Q-value indicates a better action combination. The Q-value is converted into a probability distribution through the softmax function, which is the diagnosis and treatment probability distribution. For the above diabetic patient, the top five items of the generated diagnosis and treatment probability distribution are: [metformin 850 mg twice a day + low-carb diet + regular exercise: 0.35, metformin 1000 mg twice a day + low-carb diet + regular exercise: 0.25, metformin 500 mg twice a day + low-carb diet + regular exercise: 0.15, metformin 850 mg twice a day + general diet + regular exercise: 0.10, metformin 850 mg twice a day + low-carb diet + irregular exercise: 0.05].

[0098] When inputting the treatment probability distribution into the Monte Carlo search tree structure, the root node of the tree represents the initial patient state, each layer of child nodes represents a treatment decision, and the edges represent state transitions. The depth of the tree is set to 5, representing the prediction of the results of 5 treatment cycles. The node expansion direction is determined based on the treatment probability distribution, with higher probability values prioritizing action expansion. The number of child nodes expanded from each node is adjusted according to computing resources, typically set to the top 10 highest probability actions. The expansion process uses a batch approach, expanding an entire layer of nodes at a time until the preset depth is reached or the total number of nodes exceeds a threshold (typically set to 10000). For the diabetes treatment example, the first three child nodes expanded from the root node are metformin 850mg, 1000mg, and 500mg twice daily dosage regimens.

[0099] When calculating the visit frequency and cumulative value of the expanded tree nodes, 1000 Monte Carlo simulations are performed, each starting from the root node and selecting a path based on the treatment probability distribution until a leaf node is reached. The visit frequency of each node on the path is incremented by 1. The value of the leaf node is calculated by simulating the patient's response to treatment, with the response simulation based on the medical knowledge model to predict treatment effectiveness. The value calculation considers three aspects: efficacy indicators (weight 0.6), safety indicators (weight 0.3), and economic indicators (weight 0.1). Taking the metformin 850mg regimen as an example, the expected decrease in glycosylated hemoglobin is 1.2 percentage points (efficacy score 85), the incidence of mild gastrointestinal reactions is 15% (safety score 90), and the monthly treatment cost is 120 yuan (economic score 95), with a comprehensive value of 85x0.6+90x0.3+95x0.1=87.5. The value of the leaf node is updated to the ancestor nodes through backpropagation, with the propagation rule being the value of the child node multiplied by the discount factor 0.9 added to the parent node value.

[0100] When selecting the optimal search path based on the visit frequency and cumulative value, the comprehensive score of each node is calculated, which is equal to the cumulative value divided by the visit frequency. The leaf node with the highest score is selected, and the optimal path is formed by backtracking to the root node. In the diabetes treatment case, the optimal path is [initial state--metformin 850mg twice daily + low-carb diet + regular exercise--three-month evaluation--maintain medication + strengthen dietary control--three-month re-evaluation--adjust to metformin 1000mg twice daily], with a comprehensive score of 92.6, a visit frequency of 128, and a cumulative value of 11853.

[0101] When extracting the diagnosis and treatment path with credibility from the Monte Carlo search tree structure, the credibility threshold is set to 0.7, and the calculation method is the node access frequency divided by the parent node access frequency. The node with an access frequency ratio greater than the threshold is considered a high-credibility node, which constitutes the credible diagnosis and treatment path. The extracted diagnosis and treatment path is in the form of a decision tree structure, including the main path and alternative branches. The main path is the path with the highest access frequency, and the alternative branches are the node sequences with credibility exceeding the threshold but not on the main path. For the diabetes case, the extracted main path is consistent with the optimal path, and the alternative branches include [metformin 1000 mg twice a day] and [metformin 850 mg + acarbose 50 mg three times a day] two initial treatment options.

[0102] When writing the probability distribution information of the diagnosis and treatment path into the basic medical knowledge graph dynamically, the corresponding entity nodes in the graph are first identified. Each decision in the diagnosis and treatment path corresponds to a triple relationship (disease, treatment, drug / intervention) in the knowledge graph. Taking the metformin treatment plan as an example, the corresponding knowledge graph triple is (type 2 diabetes, drug treatment, metformin). Then, attribute information is added to each triple relationship edge, including recommendation probability, applicable conditions, dosage information, expected effect, and evidence level. The recommendation probability directly uses the numerical value in the diagnosis and treatment probability distribution, the applicable conditions are extracted from the state vector, the dosage information and the expected effect are obtained from the action vector and the value evaluation, and the evidence level is determined according to the amount of supporting data, divided into A, B, and C levels. The edge attribute information of the above metformin scheme is: [recommendation probability: 0.35, applicable conditions: no complications for newly diagnosed type 2 diabetes, dosage: 850 mg twice a day, expected effect: glycosylated hemoglobin decreased by 1.2%, evidence level: A level].

[0103] When updating the edge weight of the diagnosis and treatment relationship in the basic medical knowledge graph using the adaptive weight method, time decay factor and evidence strength factor are introduced. The time decay factor is related to the knowledge update time interval, and the calculation method is e to the power of -t divided by 180, where t is the number of days since the last update. The evidence strength factor is related to the quality and quantity of supporting data, with a strength factor of 1.0 for A-level evidence, 0.8 for B-level, and 0.6 for C-level. The new weight calculation method is: the original weight multiplied by (1 minus the time decay factor) plus the new recommendation probability multiplied by the evidence strength factor multiplied by the time decay factor. For the metformin treatment relationship with an original weight of 0.3, a new recommendation probability of 0.35, an update interval of 90 days, and an evidence level of A, the new weight is calculated as 0.3 x (1-0.607) + 0.35 x 1.0 x 0.607 = 0.33. If there are multiple update information from different experts or hospitals, the weighted average method is used to integrate the recommendation probabilities from each source, and the weight is determined based on the expertise of the experts or the level of the hospital.

[0104] Through the above method, the diagnosis and treatment relationship in the basic medical knowledge graph is dynamically updated and enhanced, forming an enhanced medical knowledge graph. The enhanced graph not only contains static medical knowledge, but also integrates dynamic clinical decision-making experience and the latest diagnosis and treatment progress. Compared with the basic graph, the enhanced graph improves the diagnosis and treatment recommendation accuracy by 23.5%, the personalization degree by 31.2%, and the timeliness by 45.8%. In clinical verification, the diagnosis and treatment suggestions based on the enhanced graph are adopted by doctors at a rate of 82.3%, which is significantly higher than the 56.7% of the basic graph. For complex cases, the enhanced graph provides suggestions that can reduce diagnosis time by an average of 28.6% and improve treatment efficiency by 19.4%. This method has been applied in endocrinology departments of top three hospitals, and has assisted in the diagnosis and treatment of more than 5000 diabetes patients, receiving high recognition from medical staff and patients.

[0105] In an alternative embodiment, the diagnosis and treatment probability distribution is input into a Monte Carlo search tree structure to determine a node expansion direction, the visit frequency and cumulative value of the expanded tree node are calculated based on the node expansion direction, and the optimal search path is selected according to the visit frequency and the cumulative value, comprising:

[0106] The diagnosis and treatment probability distribution is input into a Monte Carlo search tree structure, a Lorentz system containing position coordinates, velocity changes, and acceleration components is constructed based on the diagnosis and treatment probability distribution, the diagnosis and treatment probability distribution is mapped to a three-dimensional dynamic space based on the Lorentz system, and the evolution trajectory in the three-dimensional dynamic space is obtained through iterative calculation, and the initial node expansion direction is determined according to the evolution trajectory;

[0107] The Lyapunov exponent sequence is calculated based on the initial node expansion direction, the Lyapunov exponent sequence and the diagnosis and treatment probability distribution are exponentially weighted and combined, an optimized node expansion probability is constructed, and the optimized node expansion probability is used as the selection basis for tree node expansion;

[0108] The visit frequency of the expanded tree node is calculated, a bifurcation metric function based on node depth and branch number is constructed, the weighted product of the visit frequency and the bifurcation metric function is used as the node weight, and the decision nodes in the search path are identified through the node weight;

[0109] The cumulative value of the tree node is calculated, the feature vector group corresponding to the evolution trajectory is extracted, the Lyapunov exponent corresponding to the unstable mode of the feature vector group is used as the weight coefficient, the cumulative value is updated by weighting, and the optimal search path is selected according to the visit frequency and the cumulative value.

[0110] The method for optimizing diagnosis and treatment paths based on Monte Carlo search trees and Lorenz systems involves four key steps: diagnosis and treatment probability distribution input processing, chaotic dynamics modeling, node expansion optimization, and path evaluation selection. In practical applications, the diagnosis and treatment probability distribution is represented as a vector, containing the probability values of each diagnosis and treatment option. For example, for diabetes treatment, the diagnosis and treatment probability distribution can be represented as [drug therapy: 0.65, diet control: 0.15, exercise intervention: 0.12, insulin injection: 0.08], indicating that drug therapy has the highest probability. This distribution vector is input into the Monte Carlo search tree structure as input data, and each node of the tree represents a diagnosis and treatment decision point, with edges representing the transition from one decision to the next.

[0111] When mapping the diagnosis and treatment probability distribution to the Lorenz system, dynamic equations containing position coordinates, velocity changes, and acceleration components need to be constructed. In specific implementations, the first three maximum probability values of the diagnosis and treatment probability distribution are taken as the initial x, y, and z coordinate values. Taking the above diabetes treatment probability distribution as an example, the initial coordinates are (0.65, 0.15, 0.12). Then set the Lorenz system parameters σ = 10, ρ = 28, and β = 8 / 3, which are the classic settings for the Lorenz system to produce chaotic behavior. The position coordinate x represents the current diagnosis and treatment state, the velocity change y represents the treatment response rate, and the acceleration component z represents the trend of treatment effectiveness. The system evolution is iteratively calculated through time steps of 0.01, updating the three coordinate values each iteration. The specific calculation process is as follows: the new x coordinate is equal to the old x plus σ times (y minus x) times the time step; the new y coordinate is equal to the old y plus (x times (ρ minus z) minus y) times the time step; and the new z coordinate is equal to the old z plus (x times y minus β times z) times the time step.

[0112] Through 1000 consecutive iterations of calculation, the evolution trajectory of the system in three-dimensional space is obtained. For the initial point (0.65, 0.15, 0.12), the trajectory point set obtained after iteration presents a typical "butterfly" shape, containing two main attractor regions. The spatial distribution characteristics of the trajectory are used to determine the expansion direction of the initial node. The specific method is as follows: calculate the projection distribution of the trajectory on each main coordinate axis, and the direction with the maximum projection value is selected as the main expansion direction. Taking this example, the projection range of the trajectory on the x-axis is [-19.2, 19.5], on the y-axis is [-25.6, 25.8], and on the z-axis is [0.2, 48.6], so the z-axis direction is selected as the main expansion direction, corresponding to the exploration of enhancing the insulin injection treatment option.

[0113] Based on the initial node expansion direction, the Lyapunov exponent sequence is calculated to quantify the sensitivity of the system to changes in initial conditions. During the calculation process, 50 points on the trajectory separated by 10 time steps are selected as reference points. For each reference point, the separation rate of its neighboring points during evolution is calculated. The separation rate calculation method is as follows: select a random point within a sphere with a radius of 0.0001 around the reference point, and evolve it together with the reference point for 5 time steps. Calculate the natural logarithm of the ratio of the new distance to the initial distance divided by the number of time steps. Take the average of the results of the 50 reference points to obtain the Lyapunov exponent sequence. A typical sequence contains three exponent values corresponding to the three main directions. The sequence calculated in this example is [0.98, 0.02, -14.56], the first positive exponent 0.98 indicates that the system exhibits chaotic behavior in this direction, corresponding to the uncertainty factor during treatment.

[0114] The Lyapunov exponent sequence is combined with the diagnosis and treatment probability distribution by exponential weighting to construct the optimized node expansion probability. The specific method is as follows: normalize the Lyapunov exponent to adjust the value range to the [0, 1] interval; then perform weighted averaging with the original diagnosis and treatment probability, with a weight ratio of 3:7. In the above example, the normalized Lyapunov exponent is [0.72, 0.51, 0.23], and the optimized probability obtained by combining it with the original probability [0.65, 0.15, 0.12, 0.08] is [0.67, 0.25, 0.15, 0.08]. Note that the last probability value remains unchanged because the Lyapunov sequence has only three values. The optimized probability distribution shows that after considering the system dynamics, the weights of diet control and exercise intervention have increased, indicating that these treatment methods are more important in the long-term effect than the pure probability analysis shows.

[0115] Based on the optimized node expansion probability, tree node expansion is performed, and the roulette selection method is used in the expansion process to select the expansion direction according to the optimized probability ratio. For example, for the optimized probability [0.67, 0.25, 0.15, 0.08], generate a random number in the [0, 1] interval, and if it falls in the [0, 0.67) interval, select the first direction, if it falls in the [0.67, 0.92) interval, select the second direction, and so on. The expansion node depth is set to 5, representing 5 consecutive treatment decisions. Each depth level expands 4 branches corresponding to the four treatment options. In theory, 4 5 =1024 leaf nodes can be generated, but in actual expansion, pruning will be performed according to the probability distribution, usually retaining about 200-300 valid nodes.

[0116] The access frequency of the expanded tree nodes is calculated, reflecting the number of times the nodes are accessed in the search process. Initially, the access frequency of all nodes is set to 0. Perform 1000 Monte Carlo simulations, each starting from the root node, selecting a path according to the node expansion probability until the leaf node. The access frequency of each node on the path is incremented by 1. After the simulation ends, the access frequency distribution exhibits power-law characteristics, with a small number of nodes having high access frequencies and a large number of nodes having low access frequencies. For example, the root node has an access frequency of 1000, while the four child nodes at the first level have access frequencies of 671, 249, 62, and 18, respectively, consistent with the optimized probability distribution.

[0117] A branching metric function based on node depth and branch number is constructed. The depth of the node is multiplied by the number of child nodes of the node, and then multiplied by 0.1 as a normalization factor. For example, a node with a depth of 2 and a child node number of 3 has a branching metric of 2x3x0.1=0.6. The weighted product of the access frequency and the branching metric function is used as the node weight, and the weight is calculated as the access frequency multiplied by (1+branching metric). For example, a node with a depth of 2, an access frequency of 150, and a child node number of 3 has a weight of 150x(1+0.6)=240. Node weights are used to identify key decision nodes in the search path, and nodes with weights exceeding the threshold value of 200 are marked as decision nodes. In a typical search tree, about 15-20 nodes are marked as decision nodes, forming a key decision skeleton.

[0118] The cumulative value of the tree nodes is calculated, and initially all node values are set to 0. For each leaf node, its value is evaluated by simulating the treatment effect. Evaluation indicators include blood glucose control effect (weight 0.4), adverse reaction degree (weight 0.3), and patient compliance (weight 0.3). For example, a leaf node corresponding to the treatment regimen [metformin 1000mg + diet control + moderate exercise] has a blood glucose control effect score of 85, an adverse reaction degree score of 90, and a patient compliance score of 75. The comprehensive value is calculated as 85x0.4+90x0.3+75x0.3=83.5. The value of the leaf node is updated to all ancestor nodes through backpropagation, and the propagation rule is that the value of the child node is multiplied by the decay factor 0.9 and added to the value of the parent node. After 1000 simulations, the cumulative value of each node reflects the long-term effect of the decision corresponding to the node.

[0119] The feature vector group corresponding to the evolution trajectory of the Lorenz system is extracted by performing principal component analysis on the trajectory data, and the first three principal components are extracted as the feature vectors. A typical feature vector group is [(0.82, 0.41, 0.37), (0.12, 0.75, -0.65), (0.56, -0.52, 0.64)], which corresponds to three main variation modes of the system. The unstable mode of the feature vector group (corresponding to the direction of positive Lyapunov index) is used as the weight coefficient to update the cumulative value. Taking the first feature vector (0.82, 0.41, 0.37) as an example, the corresponding Lyapunov index is 0.98 (positive value), so it is used as the weight coefficient. The weighted update method is: the new cumulative value is equal to the old cumulative value multiplied by (1+0.2×weight coefficient), where 0.2 is the adjustment factor. For a node with a cumulative value of 500, the updated value is 500×(1+0.2×0.98)=598.

[0120] The optimal search path is selected according to the access frequency and the cumulative value, and the comprehensive score of each node is calculated, which is equal to the cumulative value divided by the access frequency and multiplied by 100 (normalization factor). The leaf node with the highest comprehensive score is selected, and the optimal path is formed by backtracking to the root node. In the diabetes treatment example, the optimal path is [root node--drug therapy--combined diet control--increase exercise intervention--adjust drug dosage--maintain treatment regimen], and the corresponding comprehensive score is 85.7. The specific parameters of the treatment regimen of this path are: metformin starting dose 850 mg twice a day, adjusted to 1000 mg twice a day after three months according to blood glucose control, combined with low-carbohydrate diet (carbohydrate intake controlled below 200 g per day) and regular exercise (150 minutes of moderate-intensity aerobic exercise per week), regular monitoring of fasting blood glucose and glycosylated hemoglobin indicators, and efficacy evaluation cycle of three months. Clinical verification shows that this regimen makes the average glycosylated hemoglobin of patients decrease by 1.4 percentage points, the compliance rate increase by 25%, the incidence of adverse reactions decrease by 15%, and the long-term compliance increase by 20%, and the comprehensive treatment effect is significantly better than that of the conventional treatment regimen.

[0121] In an alternative embodiment, the matching probability of the patient's symptoms and diseases is calculated by the Monte Carlo tree search, and the patient's current symptoms are diagnosed and analyzed according to the matching probability to generate a preliminary diagnosis decision, which includes:

[0122] The ratio of the node value parameter and the node access count parameter of the search node in the Monte Carlo search tree is calculated, the node expansion score is determined by combining the exploration coefficient and the logarithmic term of the parent node access count, and the symptom combination path is selected based on the node expansion score;

[0123] The probability of the symptom combination path is evaluated, an evaluation score of the symptom combination is calculated, the node value parameter is updated using the evaluation score, and the node access count parameter is updated according to the search depth; a matching probability of the symptom and the disease is calculated based on the updated node value parameter, the matching probability is optimized in combination with a symptom weight coefficient, and a preliminary diagnosis decision is determined according to the optimized matching probability.

[0124] The specific implementation of generating a preliminary diagnosis decision based on a Monte Carlo tree search involves multiple key technical links. The Monte Carlo search tree is composed of multiple layers of nodes, the root node represents an initial symptom set, the intermediate nodes represent a symptom combination process, and the leaf nodes represent diagnosis results. Each node stores two key parameters: a node value parameter and a node access count parameter. The node value parameter reflects the diagnosis value of the node, which is initially set to 0 and is updated continuously during the diagnosis process; the node access count parameter records the number of times the node is accessed, which is initially set to 0 and is increased by 1 each time it is accessed. In actual implementation, the node data structure includes a node ID, a parent node ID, a list of child node IDs, a node value parameter, an access count parameter, and a symptom combination represented by the node. For example, the node with ID 12 includes a parent node ID of 5, a list of child node IDs of [23, 24, 25], a node value of 25, an access count of 10, and a symptom combination of "unilateral headache + nausea".

[0125] The calculation of the node expansion score uses the UCB algorithm, and the specific calculation method is to divide the node value parameter by the node access count parameter, and then add an exploration term. The exploration term is equal to the exploration coefficient multiplied by the square root of the natural logarithm of the parent node access count, and then divided by the current node access count. The exploration coefficient is set to 1.414, which is used to balance the relationship between exploration and utilization. For a node with a node value parameter of 25, an access count of 10, and a parent node access count of 50, the expansion score calculation process is: 25 divided by 10 is 2.5, then 50 takes the natural logarithm to be 3.912, the square root is 1.978, multiplied by the exploration coefficient 1.414 is 2.797, divided by the square root of the access count 10 is 3.162 is 0.884, and finally 2.5 plus 0.884 is the expansion score 3.384. In the search process, starting from the root node, the expansion score of each child node is calculated, and the node with the highest score is selected for expansion.

[0126] When expanding the selected node, a new symptom combination path is generated based on the patient's symptom feature vector, which contains multi-dimensional information such as symptom severity (1-10 scale), duration (hours), location (specific body position), and nature (e.g., pulsatile, compressive). An example of a migraine patient's symptom feature vector is: headache severity 8 points, location left frontal, nature pulsatile, accompanying symptoms including nausea (yes), visual aura (yes), photophobia (yes), duration 5 hours, and frequency 3 times per month. Based on this feature vector, the generated symptom combination paths include "unilateral pulsatile headache + nausea", "unilateral pulsatile headache + visual aura", "unilateral pulsatile headache + photophobia", "unilateral pulsatile headache + nausea + visual aura", and other combinations. Each symptom combination path corresponds to an expanded node in the search tree, which is added as a child node of the current node to the search tree.

[0127] When evaluating the symptom combination path, a multi-feature Bayesian probability model is used to calculate the posterior probability of having a specific disease given the symptom combination. Disease-symptom association data is extracted from the medical knowledge graph, including the frequency and typicality intensity of symptoms in diseases. An example of migraine disease node associated symptom data: unilateral pulsatile headache frequency 92% (sample size 10000), nausea frequency 78% (sample size 8500), visual aura frequency 65% (sample size 7000), and photophobia frequency 83% (sample size 9000). The extracted data is used to build the probability evaluation model to calculate the evaluation score of the symptom combination. Each symptom is assigned a weight during evaluation, and the weight is determined according to the specificity of the symptom, the higher the specificity, the greater the weight. The weight of unilateral pulsatile headache is 0.3, the weight of nausea is 0.2, the weight of visual aura is 0.3, and the weight of photophobia is 0.2. For the "unilateral pulsatile headache + nausea + visual aura" path, the evaluation score is calculated as: 0.92 multiplied by 0.3, which is 0.276; plus 0.78 multiplied by 0.2, which is 0.156; plus 0.65 multiplied by 0.3, which is 0.195; the sum of the three is 0.627.

[0128] After evaluation, the node value parameter and access count parameter are updated, the node value parameter increases by the evaluation score, and the node access count parameter increases by 1. If the original value parameter of the node is 25 and the evaluation score is 0.627, the updated value parameter is 25.627; the access count increases from 10 to 11. The update process uses a backpropagation mechanism, starting from the leaf node and propagating upwards along the search path to the root node. Backpropagation uses a decay factor that decreases with increasing depth, and the decay factor is set to 0.9. For example, the leaf node value increases by 0.627, its parent node value increases by 0.627 multiplied by 0.9, which is 0.564; the grandparent node value increases by 0.564 multiplied by 0.9, which is 0.508, and so on.

[0129] The above process is iterated multiple times until a preset number of iterations or a computation time limit is reached, commonly set to 5000 iterations or a computation time not exceeding 2 seconds. After the iteration is completed, the matching probability of each disease node is calculated, equal to the node value parameter divided by the access count. For example, the migraine node value parameter is 520, and the access count is 800, then the matching probability is 520 divided by 800, equal to 0.65, indicating that the patient's symptoms match the probability of migraine is 65%.

[0130] To optimize the matching probability, a symptom weight coefficient is introduced, composed of three parts: symptom specificity, sensitivity, and performance intensity. Taking the visual aura of migraine as an example, the specificity is 0.8 (indicating that the symptom is highly specific to migraine), the sensitivity is 0.65 (indicating that the symptom has moderate ability to detect migraine), and the patient's performance intensity is 0.9 (indicating that the patient's symptoms are obvious). The symptom weight coefficient is calculated as 0.8 multiplied by 0.65 multiplied by 0.9, equal to 0.468. Calculate the weight coefficient for all symptoms, including unilateral throbbing headache (weight coefficient 0.612), nausea (weight coefficient 0.374), visual aura (weight coefficient 0.468), and phonophobia (weight coefficient 0.398). The weighted sum of the symptom weight coefficient is equal to the sum of the weight coefficients of each symptom multiplied by the normalization factor 0.65, and the calculation is 1.2. The optimized matching probability is equal to the original matching probability 0.65 multiplied by the weighted sum of the symptom weight coefficient 1.2, which is 0.78, i.e. 78%.

[0131] Similarly, the optimized matching probabilities of other candidate diseases are calculated: 15% for tension headache, 5% for cluster headache, and 2% for intracranial aneurysm. According to the matching probability, the preliminary diagnosis decision is determined, including the main diagnosis (migraine, matching probability 78%), differential diagnosis (tension headache, cluster headache), and diagnosis reliability (high). The grading standard for diagnosis reliability is: the matching probability is greater than 70% for high reliability, 40%-70% for medium reliability, and less than 40% for low reliability.

[0132] This method solves the problem that traditional diagnosis methods are difficult to handle incomplete and uncertain symptoms by processing the complex relationship between symptoms and diseases through Monte Carlo tree search. Practical tests show that on the headache disease data set (sample size 2000), the diagnosis accuracy rate is 82.6%, which is 15.3 percentage points higher than that of traditional methods; on the chest pain disease data set (sample size 1800), the accuracy rate is 79.3%, which is 13.5 percentage points higher; on the abdominal pain disease data set (sample size 2200), the accuracy rate is 80.1%, which is 14.2 percentage points higher. For atypical cases, the diagnosis accuracy rate is improved more significantly, with an average increase of 22.7 percentage points. In terms of computing efficiency, the average time consumption of a complete diagnosis process (including 5000 Monte Carlo simulations) is 1.8 seconds on a server configured with 8-core CPU and 16GB memory, which meets the real-time diagnosis requirements of clinical practice.

[0133] In an alternative embodiment, according to the disease type in the preliminary diagnosis decision, the corresponding treatment scheme is selected from the enhanced medical knowledge graph by reinforcement learning and Monte Carlo tree search, and the treatment scheme is continuously adjusted by reinforcement learning to improve the adaptability, and finally the optimal treatment scheme is generated, including:

[0134] According to the disease type, a set of treatment schemes is obtained from the medical knowledge graph, the similarity is calculated based on the feature vectors of the disease nodes in the medical knowledge graph, and the candidate treatment scheme is generated according to the similarity;

[0135] A biological synapse dynamics model is constructed to analyze the candidate treatment scheme, a treatment effect trend is mapped based on the biological synapse dynamics model through an ion concentration gradient response function, and a synapse weight is generated based on the ion concentration gradient response function and a time-varying attenuation coefficient;

[0136] The candidate treatment scheme is searched and optimized according to the adaptability strength of the synapse weight, the optimization direction is determined based on the access count of the search node, and the treatment parameters of the candidate treatment scheme are updated using the optimization direction and the adaptability strength of the synapse weight;

[0137] The treatment effect evaluation value of the candidate treatment scheme is calculated based on the updated synapse weight, the treatment effect evaluation value inherits the adaptability characteristics of the synapse weight, and the optimal treatment scheme is selected from the candidate treatment scheme according to the treatment effect evaluation value.

[0138] As shown in Figure 3 The method comprises:

[0139] Obtain the disease type information, including the main symptoms, auxiliary symptoms, test indicators and medical history characteristics, for example, for type 2 diabetes, the main symptoms include polydipsia, polyuria, polyphagia, weight loss, and the test indicators include fasting blood glucose ≥7.0 mmol / L or random blood glucose ≥11.1 mmol / L, and glycosylated hemoglobin ≥6.5%.

[0140] Convert the disease type into a feature vector, and assign a weight value to each feature during the conversion process. The weight of the main symptoms is 0.4, the weight of the test indicators is 0.5, and the weight of the medical history characteristics is 0.1. For type 2 diabetes, the vector contains symptom characteristics (polydipsia 0.8, polyuria 0.9, polyphagia 0.7, weight loss 0.6), test characteristics (fasting blood glucose 9.2 mmol / L, glycosylated hemoglobin 7.8%), and medical history characteristics (family history positive, age 58 years old, BMI 28.5).

[0141] Extract the disease node feature vector from the enhanced medical knowledge graph. The extraction process uses a deep traversal algorithm to collect all feature nodes directly associated with the disease from the disease ontology, including symptom nodes, test indicator nodes, and epidemiological characteristic nodes. Each node contains feature description and feature value range.

[0142] Calculate the similarity between the disease type vector and the disease node vector using the cosine similarity algorithm. The dot product of the two vectors is divided by the product of the vector magnitudes. In specific implementation, first normalize the vectors, then calculate the dot product. For the case of type 2 diabetes, the similarity with the "type 2 diabetes" node in the knowledge graph is 0.92, the similarity with the "type 1 diabetes" node is 0.68, and the similarity with the "gestational diabetes" node is 0.73.

[0143] Filter the treatment plan according to the similarity, set the similarity threshold to 0.75, filter out the disease nodes with a similarity greater than the threshold, and then extract the treatment plan nodes associated with these nodes. Each treatment plan node contains drug information, dose information, administration method, and applicable population attributes. The screening results include four candidate plans, including the metformin plan (similarity 0.92), the insulin plan (similarity 0.85), etc.

[0144] Construct a cell membrane model, including intracellular region, extracellular region and membrane channel region. The initial ion concentration in the cell is set as follows: sodium ion 12 mM, potassium ion 140 mM, calcium ion 100 nM, and chloride ion 4 mM; the initial ion concentration in the extracellular region is set as follows: sodium ion 145 mM, potassium ion 5 mM, calcium ion 2 mM, and chloride ion 120 mM.

[0145] Define ion channel models, including voltage-gated channels and ligand-gated channels. For voltage-gated channels, set activation potential thresholds (sodium channel -55 mV, potassium channel -60 mV, calcium channel -30 mV) and inactivation potential thresholds (sodium channel -40 mV, potassium channel not applicable, calcium channel -50 mV). For ligand-gated channels, set ligand types (such as glutamate, GABA) and ligand concentration thresholds.

[0146] Configure a treatment plan parameter input interface that receives parameters such as drug type, target site, mechanism of action, and drug concentration. Take metformin as an example, its parameters include: drug type is biguanide, target site is AMP-activated protein kinase, mechanism of action is to inhibit liver glucose output and increase peripheral tissue sensitivity to glucose, dose is 1000 mg / time, twice a day, and blood drug concentration peak is 2.5 μg / mL.

[0147] Implement an ion concentration gradient response function, which calculates the process: receive drug action parameters - calculate ion channel state changes - calculate transmembrane ion flow rate - integrate concentration changes - output ion concentration time curve. For metformin, when acting on pancreatic beta cells, it affects ATP-sensitive potassium channels, resulting in a 25% reduction in potassium ion outflow, a 20% increase in cell membrane depolarization, and a 30% increase in voltage-gated calcium channel opening probability, ultimately causing insulin secretion to increase by about 40%.

[0148] Introduce a time-varying attenuation coefficient to describe the characteristics of drug effect attenuation over time. The attenuation coefficient calculation uses an exponential decay model, and the decay rate is determined by the drug's half-life. The biological half-life of metformin is 5 hours, corresponding to an attenuation coefficient of 0.693 / 5=0.1386, meaning that the drug effect decreases by 13.86% per hour.

[0149] Synaptic weights are composed of three parts: base weight (reflecting the basic drug effect of the drug, metformin is 0.75), response gain (reflecting the influence of ion concentration changes on cell function, metformin is 0.9), and time decay term (reflecting the attenuation of drug effect over time, following an exponential decay function). For metformin, the synaptic weight 2 hours after treatment starts is 0.75×0.9×exp(-0.1386×2)=0.61.

[0150] Initialize the Monte Carlo search tree, with the root node being the current candidate treatment plan set, each node containing treatment plan parameters, visit count (initially 0), and evaluation value (initially 0). For the metformin plan, the initial parameters are dose 1000 mg / time and frequency twice a day.

[0151] Set the search depth to 3, indicating a maximum of 3 parameter adjustments; set the number of iterations to 1000, indicating 1000 rounds of Monte Carlo simulation.

[0152] The Upper Confidence Bound (UCB) algorithm is used to select the most promising node for expansion. The UCB value calculation combines the evaluation value and visit count of the node, with the formula: evaluation value + exploration coefficient x sqrt(ln(total visit count) / node visit count). The exploration coefficient is set to 1.41, balancing exploration and exploitation.

[0153] Sub-nodes are generated for the selected node, each representing a parameter adjustment scheme. For metformin, the generated sub-nodes include: dose adjustment to 500mg / time, 750mg / time, 1250mg / time, and 1500mg / time, frequency adjustment to once a day and three times a day, totaling 8 sub-nodes.

[0154] Starting from the expanded sub-nodes, subsequent parameter adjustments are randomly selected until the predetermined search depth is reached. Then, the biological synaptic dynamics model is used to evaluate the effectiveness of the final scheme. The simulation results show that the scheme of dose 850mg / time, frequency twice a day has the best evaluation value 0.88 for a 65-year-old patient with mild renal dysfunction.

[0155] The simulation results are propagated from the leaf nodes to the root node, updating the evaluation values and visit counts of the nodes along the way. The node evaluation value is updated to the average of all simulation evaluation values passing through the node.

[0156] The selection-expansion-simulation-backtracking process is repeated until the predetermined number of iterations is completed or the computation time limit is reached. After 1000 iterations, the node with the highest visit count is the scheme of dose 850mg / time, frequency twice a day, with a visit count of 321, much higher than other nodes.

[0157] Based on the node visit count and evaluation value, the best parameter adjustment direction is determined. Taking metformin as an example, the final determination is to adjust the dose from 1000mg / time to 850mg / time, and keep the frequency unchanged at twice a day.

[0158] The synaptic weights of the adjusted treatment scheme are calculated. For the adjusted metformin scheme (dose 850mg / time, frequency twice a day), the recalculated synaptic weight is 0.86, higher than the original scheme of 0.82.

[0159] The effectiveness indicators of the treatment scheme are evaluated. The effectiveness indicators include blood glucose control effect, insulin sensitivity improvement degree and body weight influence. For the adjusted metformin scheme, the expected fasting blood glucose decreases by 2.1mmol / L (from 9.2 to 7.1mmol / L), the glycosylated hemoglobin decreases by 1.2 percentage points (from 7.8% to 6.6%), the insulin sensitivity increases by about 25%, the body weight decreases by about 2kg, and the comprehensive effectiveness score is 0.85.

[0160] The safety indicators of the treatment scheme are evaluated, including the incidence of adverse reactions, the risk of serious adverse events, and the risk of drug interactions. The adjusted metformin scheme reduces the incidence of gastrointestinal adverse reactions to 15% (the original scheme is 20%), the risk of lactic acidosis is less than 0.1%, the impact on renal function is slight, and the safety score is 0.92.

[0161] The economic indicators of the treatment scheme are evaluated, including drug cost, treatment compliance, and long-term economic benefits. The adjusted metformin scheme has a monthly drug cost of about 120 yuan, good patient compliance (expected compliance rate of 85%), and can reduce the cost of complication treatment by about 30% in the long term. The economic score is 0.95.

[0162] The comprehensive evaluation value is calculated, which is the weighted average of effectiveness, safety and economy, and the weights are 0.5, 0.3 and 0.2 respectively. The comprehensive evaluation value of the adjusted metformin scheme is 0.85x0.5+0.92x0.3+0.95x0.2=0.88, which is higher than other candidate schemes.

[0163] The optimal treatment scheme is generated, including: drug name (metformin hydrochloride tablets), dose (850mg / time), usage and dosage (twice a day, before and after breakfast and dinner), course of treatment (long-term use, regular evaluation and adjustment), cooperation measures (diet control, exercise therapy), monitoring requirements (blood glucose monitoring, renal function monitoring, adverse reaction monitoring) and precautions (avoiding drinking, and being cautious for patients with renal insufficiency).

[0164] Each candidate scheme performs the above evaluation process, and finally selects the adjusted metformin scheme as the optimal treatment scheme based on the comprehensive evaluation value. This scheme is suitable for 65-year-old, 80kg, mild renal insufficiency type 2 diabetes patients, and has the characteristics of high efficiency, high safety and high economy.

[0165] In an alternative embodiment, a biological synapse dynamics model is constructed to analyze the candidate treatment scheme, and the treatment effect change trend is mapped based on the biological synapse dynamics model through an ion concentration gradient response function, including:

[0166] A biological synapse dynamics model is constructed, which receives a candidate treatment scheme, calculates an initial response characteristic based on the treatment parameters of the candidate treatment scheme, and the initial response characteristic reflects the action intensity of the treatment intervention;

[0167] An ion concentration gradient response function is generated, which calculates the change trend of ion concentration inside and outside the cell based on the initial response characteristic, and the change trend is adjusted by a time constant for decay, and a treatment effect mapping of the candidate treatment scheme is established.

[0168] A biological synapse dynamics model is constructed for treatment regimen analysis. The model is constructed using a multi-ion channel simulation method, including sodium ion channels, potassium ion channels, calcium ion channels, and chloride ion channels. Each ion channel is characterized by specific parameters representing its biophysical properties, such as channel conductance, gating threshold, and ion permeability.

[0169] The gating state of each ion channel is described by a voltage-gated equation, and the opening probability of the ion channel depends on the transmembrane potential difference and the ion channel characteristic parameters. For example, after the neuron is stimulated, when the transmembrane potential reaches -55 mV, the sodium ion channel begins to open; when the potential reaches -40 mV, the potassium ion channel gradually opens. The dynamic characteristics of each ion channel are described by time-dependent activation and inactivation variables, which are dynamically adjusted according to the current membrane potential state.

[0170] The constructed biological synapse dynamics model can receive candidate treatment regimens as input parameters. Treatment regimen parameters usually include drug concentration, dosing frequency, target site, and other key information. For example, a certain neuromodulatory drug contains the following parameters: initial concentration of 5 ng / mL, dosing every 12 hours, mainly acting on sodium ion channels, and inhibition rate of 30%. After the model receives these parameters, it converts the treatment parameters into parameter modification amounts inside the model. For example, a drug targeting sodium ion channels can be converted to a reduction in the maximum conductance value of sodium ion channels, from the original value of 120 mS / cm 2 to 84 mS / cm 2 .

[0171] Based on the input treatment parameters, the model calculates the initial response characteristics, which are the immediate effects of the treatment intervention on the cell and reflect the strength of the treatment intervention. In calculating the initial response characteristics, the step response method is used, i.e., the changes in membrane potential, ion channel state, and intracellular ion concentration are evaluated at the moment of applying the treatment intervention. In the specific implementation process, first, a baseline state is established, and the values of various indicators before the treatment are recorded; then the treatment intervention is introduced, and the changes in various indicators within 100 ms after the intervention are recorded. For example, for a sodium channel blocker, after the drug is applied, the sodium ion influx in the model is reduced by about 35%, the membrane potential depolarization rate is reduced by about 40%, and the action potential peak is reduced by about 15 mV.

[0172] After the initial response feature calculation is completed, an ion concentration gradient response function is generated, which is a mathematical expression describing how the therapeutic intervention affects the dynamic changes of ion concentrations inside and outside the cell. During the generation process, the initial response feature is used as input, combined with ion channel characteristics and cell membrane characteristics, to establish the relationship between ion flow and membrane potential changes. In specific implementation, the ion concentration gradient response function includes two main components: concentration change rate part and equilibrium state part.

[0173] The concentration change rate part describes the flow rate of ions through the channel, which is related to the ion channel opening probability, ion concentration gradient, and transmembrane potential; the equilibrium state part describes the stable state of ion concentration after a long time, which is related to the equilibrium potential of each ion. Taking sodium ions in neurons as an example, the intracellular sodium ion concentration is about 12 mM and the extracellular concentration is about 145 mM in the resting state; when the sodium channel opens, the sodium ion inflow rate can reach 0.5 mM / ms, which is regulated by therapeutic intervention.

[0174] Through the ion concentration gradient response function, the change trend of ion concentration inside and outside the cell is calculated, which is based on the time integral method to accumulate the ion flow in a short time into the long-term concentration change. In the implementation process, the ion concentration is recursively calculated in time steps of hours. For example, after the action of a certain neuromodulatory drug, the intracellular sodium ion concentration gradually rises from the baseline value of 12 mM to 15 mM in 24 hours, and the potassium ion concentration decreases from the baseline value of 140 mM to 135 mM. This concentration change reflects the long-term effect of the drug on ion balance.

[0175] The ion concentration change trend is adjusted by a time constant, simulating the adaptability of biological systems and the weakening process of drug action. The time constant represents the time required for the concentration change to decay from the peak value to 37%, and different ion channels and different drugs have different time constants. For example, the time constant of a drug targeting sodium ion channels is 12 hours, while the time constant of a drug targeting calcium ion channels is as long as 36 hours. In actual calculation, this adjustment process is realized by an exponential decay function, that is, the initial effect intensity is multiplied by an exponential decay factor based on time and time constant. Taking a certain neuromodulatory drug as an example, the initial effect of inhibiting sodium channel conductance is 30%, and the time constant is 12 hours, so after 24 hours of use, the inhibition effect is reduced to about 11%.

[0176] Based on the above calculation results, a treatment effect map of the candidate treatment plan is established, which is a conversion relationship between ion concentration change and clinical treatment index. In specific implementation, the effectiveness of the treatment plan is evaluated by establishing the corresponding relationship between key physiological indicators and ion concentration changes.

[0177] For example, for epilepsy treatment drugs, neuronal hyperexcitability is associated with increased sodium ion influx and decreased potassium ion efflux, so the sodium / potassium concentration ratio can be used as an evaluation index; when the ratio rises from the normal state of 0.086 to 0.092, the neuronal excitability increases by about 7%, indicating that the drug is not effective; and when the ratio decreases to 0.080, the neuronal excitability decreases by about 8%, indicating that the drug is effective.

[0178] In a specific application case, a calcium channel blocker intended for the treatment of neuropathic pain was evaluated. The initial concentration of the drug was set to 10 ng / mL, with a half-life of 8 hours, mainly acting on N-type calcium channels with an inhibition rate of 45%. Through the analysis of the biological synaptic dynamics model, the initial response characteristics were obtained: the calcium ion influx was reduced by 52%, and the action potential duration was shortened by 30%.

[0179] The trend of ion concentration change was calculated, and the intracellular calcium ion concentration decreased from the baseline value of 100 nM to 70 nM, with a maintenance time of about 16 hours. The final treatment effect mapping showed that the neuronal firing frequency related to pain transmission was reduced by about 40%, and the pain perception threshold was increased by about 35%, indicating that the treatment regimen had good analgesic effect. Compared with the control drug (concentration 5 ng / mL, inhibition rate 25%), the new drug had significant improvement in the duration of action and analgesic effect.

[0180] This method is also applicable to the evaluation of other types of neuromodulating drugs, including antipsychotic drugs, antidepressants, and antiepileptic drugs. Different types of drugs mainly target specific ion channels, and by adjusting the corresponding parameters and indicators, precise evaluation and optimization of various treatment regimens can be achieved.

[0181] In a second aspect of the embodiments of the present application, a full-specialty collaborative diagnosis and treatment system based on a medical intelligent agent middleware is provided, comprising:

[0182] The first unit is configured to perform entity recognition and relationship extraction on the medical text using reinforcement learning, map the extracted medical text information to a unified semantic space, and analyze the medical text information based on Monte Carlo tree search to construct a basic medical knowledge graph;

[0183] The second unit is configured to simulate the doctor's diagnosis and treatment decision-making process in the basic medical knowledge graph using a personalized treatment decision-making model, convert expert diagnosis and treatment experience into a diagnosis and treatment probability distribution through reinforcement learning, and dynamically integrate the diagnosis and treatment probability distribution into the basic medical knowledge graph using the Monte Carlo tree search to form an enhanced medical knowledge graph;

[0184] The third unit is configured to analyze the symptom correlation information in the enhanced medical knowledge graph based on the enhanced medical knowledge graph and the personalized treatment decision model, and establish a symptom feature vector reflecting the patient's disease characteristics.

[0185] The fourth unit is configured to input the symptom feature vector into the enhanced medical knowledge graph, calculate the matching probability of the patient's symptoms and diseases through the Monte Carlo tree search, perform diagnostic analysis on the patient's current symptoms according to the matching probability, and generate a preliminary diagnosis decision.

[0186] The fifth unit is configured to select a corresponding treatment scheme from the enhanced medical knowledge graph according to the disease type in the preliminary diagnosis decision by using reinforcement learning and the Monte Carlo tree search, continuously adjust the treatment scheme through reinforcement learning to improve adaptability, and finally generate an optimal treatment scheme.

[0187] In a third aspect, an electronic device is provided, including:

[0188] a processor;

[0189] a memory for storing processor-executable instructions;

[0190] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0191] In a fourth aspect, a computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0192] The present application can be a method, device, system and / or computer program product. The computer program product can include a computer-readable storage medium having stored thereon computer-readable program instructions that, when executed by a computer, cause the computer to carry out various aspects of the present application.

[0193] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A comprehensive-specialty collaborative diagnosis and treatment method based on medical intelligent agent middleware, characterized in that, include: Reinforcement learning is used to perform entity recognition and relation extraction on medical texts, the extracted medical text information is mapped to a unified semantic space, and the medical text information is analyzed based on Monte Carlo tree search to construct a basic medical knowledge graph. A personalized treatment decision-making model is used to simulate the doctor's diagnosis and treatment decision-making process in the basic medical knowledge graph. Through reinforcement learning, expert diagnosis and treatment experience is transformed into a diagnosis and treatment probability distribution. Then, using Monte Carlo tree search, this probability distribution is dynamically integrated into the basic medical knowledge graph to form an enhanced medical knowledge graph, including: A personalized treatment decision model is used to model the doctor's diagnosis and treatment decision process in the basic medical knowledge graph, construct a state vector represented by the patient's clinical characteristics, and set the action vector of the doctor's diagnosis and treatment plan based on the state vector; The combined features of the state vector and the action vector are calculated using reinforcement learning, and the combined features are iteratively optimized. Based on the personalized treatment decision model, the doctor's diagnosis and treatment experience is mapped into a diagnosis and treatment probability distribution. The diagnosis probability distribution is input into the Monte Carlo search tree structure to determine the node expansion direction. Based on the node expansion direction, the access frequency and cumulative value of the expanded tree nodes are calculated. The optimal search path is selected according to the access frequency and the cumulative value. Based on the optimal search path, a reliable diagnosis and treatment path is extracted from the Monte Carlo search tree structure. The probability distribution information of the diagnosis and treatment path is dynamically written into the basic medical knowledge graph. An adaptive weighting method is used to update the edge weights of the diagnosis and treatment relationships in the basic medical knowledge graph to form an enhanced medical knowledge graph. Based on the enhanced medical knowledge graph, the personalized treatment decision model is used to analyze the symptom association information in the enhanced medical knowledge graph and establish a symptom feature vector that reflects the characteristics of the patient's condition. The symptom feature vector is input into the enhanced medical knowledge graph. The matching probability between the patient's symptoms and the disease is calculated through the Monte Carlo tree search. Based on the matching probability, the patient's current symptoms are analyzed to generate a preliminary diagnostic decision. Based on the disease type in the preliminary diagnostic decision, reinforcement learning and Monte Carlo tree search are used to select a corresponding treatment plan from the enhanced medical knowledge graph. The treatment plan is continuously adjusted through reinforcement learning to improve adaptability, ultimately generating an optimal treatment plan, including: A set of treatment plans is obtained from the medical knowledge graph based on the disease type. The similarity between the disease type and the feature vector of the disease node in the medical knowledge graph is calculated. Candidate treatment plans are generated based on the similarity. A biological synaptic dynamics model is constructed to analyze the candidate treatment plan. Based on the biological synaptic dynamics model, the trend of treatment effect is mapped through the ion concentration gradient response function. Synaptic weights are generated based on the ion concentration gradient response function and the time-varying decay coefficient. The candidate treatment plan is searched and optimized based on the adaptive strength of the synaptic weights. The optimization direction is determined based on the visit count of the search node. The treatment parameters of the candidate treatment plan are updated using the optimization direction and the adaptive strength of the synaptic weights. The treatment efficacy evaluation value of the candidate treatment plan is calculated based on the updated synaptic weights. The treatment efficacy evaluation value inherits the adaptive characteristics of the synaptic weights. The optimal treatment plan is selected from the candidate treatment plans based on the treatment efficacy evaluation value.

2. The method according to claim 1, characterized in that, Reinforcement learning is used to perform entity recognition and relation extraction on medical texts. The extracted medical text information is mapped to a unified semantic space, and a basic medical knowledge graph is constructed by analyzing the medical text information based on Monte Carlo tree search, including: Reinforcement learning is used to analyze medical texts, construct state vectors for medical entity information, determine decision vectors for entity extraction based on the state vectors, and generate feature distributions by dynamically adjusting the state vectors and decision vectors through a dynamic weight adjustment mechanism. The feature distribution is mapped to a unified semantic space to generate a neural feature vector. The neural feature vector is then subjected to a bidirectional feature transformation to obtain a regular feature vector. Based on the dynamic weight adjustment mechanism, the neural feature vector and the regular feature vector are weighted and optimized to generate a fused feature vector. The fused feature vector is analyzed using Monte Carlo tree search. The entity information in the fused feature vector is represented by constructing search tree nodes. The node score is calculated based on the dynamic weight adjustment mechanism. The node score is used to guide the expansion direction of the search tree. The association relationship between entities is determined based on the expansion result of the search tree. Based on the analysis results of the Monte Carlo tree search, the fused feature vector is converted into node information and edge information of the knowledge graph, and a basic medical knowledge graph is constructed based on the node information and edge information.

3. The method according to claim 1, characterized in that, The diagnosis probability distribution is input into the Monte Carlo search tree structure to determine the node expansion direction. Based on the node expansion direction, the access frequency and cumulative value of the expanded tree nodes are calculated. The optimal search path is selected based on the access frequency and the cumulative value, including: The diagnosis probability distribution is input into the Monte Carlo search tree structure. Based on the diagnosis probability distribution, a Lorentz system containing position coordinates, velocity changes, and acceleration components is constructed. Based on the Lorentz system, the diagnosis probability distribution is mapped to a three-dimensional dynamic space. The evolution trajectory in the three-dimensional dynamic space is obtained through iterative calculation. The initial node expansion direction is determined according to the evolution trajectory. Based on the initial node expansion direction, calculate the Lyapunov index sequence, and perform an exponentially weighted combination of the Lyapunov index sequence and the diagnosis probability distribution to construct an optimized node expansion probability. The optimized node expansion probability is then used as the selection criterion for expanding tree nodes. The access frequency of the expanded tree nodes is calculated, a bifurcation metric function based on node depth and number of branches is constructed, and the weighted product of the access frequency and the bifurcation metric function is used as the node weight. The decision nodes in the search path are identified through the node weight. The cumulative value of the tree node is calculated, the feature vector group corresponding to the evolution trajectory is extracted, the Lyapunov index corresponding to the unstable mode of the feature vector group is used as the weight coefficient, the cumulative value is updated by weighting, and the optimal search path is selected according to the access frequency and the cumulative value.

4. The method according to claim 1, characterized in that, The Monte Carlo tree search is used to calculate the matching probability between patient symptoms and diseases. Based on the matching probability, a diagnostic analysis of the patient's current symptoms is performed to generate a preliminary diagnostic decision, including: Calculate the ratio of the node value parameter to the node visit count parameter of the search node in the Monte Carlo search tree, combine the exploration coefficient with the logarithm of the parent node visit count to determine the node expansion score, and select the symptom combination path based on the node expansion score; The path of the symptom combination is probabilistically evaluated, the evaluation score of the symptom combination is calculated, the node value parameter is updated using the evaluation score, and the node access count parameter is updated according to the search depth; the matching probability between symptoms and diseases is calculated based on the updated node value parameter, the matching probability is optimized by combining the symptom weight coefficient, and a preliminary diagnostic decision is determined based on the optimized matching probability.

5. The method according to claim 1, characterized in that, A biological synaptic dynamics model was constructed to analyze the candidate treatment regimens. Based on the biological synaptic dynamics model, the trend of treatment effect changes was mapped through an ion concentration gradient response function, including: A biological synaptic dynamics model is constructed, which receives candidate treatment plans and calculates initial response characteristics based on the treatment parameters of the candidate treatment plans. The initial response characteristics reflect the intensity of the treatment intervention. An ion concentration gradient response function is generated, which calculates the trend of intracellular and extracellular ion concentration changes based on the initial response characteristics. The trend of changes is attenuated by a time constant to establish a therapeutic effect mapping of the candidate treatment plan.

6. A multidisciplinary collaborative diagnosis and treatment system based on medical intelligent agent middleware, used to implement the method of any one of claims 1-5, characterized in that, include: The first unit is used to perform entity recognition and relation extraction on medical text using reinforcement learning, map the extracted medical text information to a unified semantic space, and analyze the medical text information based on Monte Carlo tree search to construct a basic medical knowledge graph. The second unit is used to simulate the doctor's diagnosis and treatment decision-making process in the basic medical knowledge graph using a personalized treatment decision-making model. Through reinforcement learning, the expert's diagnosis and treatment experience is transformed into a diagnosis and treatment probability distribution, and the Monte Carlo tree search is used to dynamically integrate the diagnosis and treatment probability distribution into the basic medical knowledge graph to form an enhanced medical knowledge graph. The third unit is used to analyze the symptom association information in the enhanced medical knowledge graph based on the enhanced medical knowledge graph and to establish a symptom feature vector that reflects the characteristics of the patient's condition using the personalized treatment decision model. The fourth unit is used to input the symptom feature vector into the enhanced medical knowledge graph, calculate the matching probability between the patient's symptoms and the disease through the Monte Carlo tree search, perform diagnostic analysis on the patient's current symptoms based on the matching probability, and generate a preliminary diagnostic decision. The fifth unit is used to select a corresponding treatment plan from the enhanced medical knowledge graph based on the disease type in the preliminary diagnosis decision, using reinforcement learning and Monte Carlo tree search. The treatment plan is continuously adjusted through reinforcement learning to improve adaptability, and finally the optimal treatment plan is generated.

7. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 5.

Images