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

Abstract

The invention provides a medical agent middleware-based full-specialized collaborative diagnosis and treatment method and system, and relates to the technical field of medical agents, and the method comprises the steps: building a basic medical knowledge graph through reinforcement learning, converting expert diagnosis and treatment experience into diagnosis and treatment probability distribution through a personalized treatment decision model, and forming an enhanced medical knowledge graph; and calculating a matching probability with a disease based on the symptom feature vector of the patient to generate a preliminary diagnosis decision, and selecting and adjusting a treatment scheme through reinforcement learning. According to the invention, general and specialized collaborative diagnosis and treatment can be realized, and medical decision accuracy and diagnosis and treatment efficiency are improved.

Description

Technical Field

[0001] This invention relates to medical intelligent agent technology, and more particularly to a multidisciplinary collaborative diagnosis and treatment method and system based on medical intelligent agent middleware. Background Technology

[0002] With the continuous development of medical technology, the scale and complexity of medical data are increasing daily, making traditional medical diagnostic methods insufficient to meet the needs of modern medical services. In recent years, the application of artificial intelligence (AI) technology in the medical field has become increasingly widespread, especially medical intelligent diagnostic systems based on knowledge graphs, reinforcement learning, and Monte Carlo tree search, which have gradually become a research hotspot. These technologies have brought revolutionary changes to the medical industry, enabling doctors to make more accurate diagnoses and personalized treatment decisions. The collaborative diagnosis and treatment model between general practitioners and specialists is an important organizational form in medical practice, providing patients with comprehensive and professional medical services by integrating the knowledge and experience of doctors from different specialties. However, existing medical intelligent systems still face some technical challenges in supporting collaborative diagnosis and treatment between general practitioners and specialists.

[0003] Traditional medical intelligent systems lack effective mechanisms for extracting medical text information and integrating knowledge, making it difficult to accurately identify professional entities and complex relationships in medical texts. This results in low-quality medical knowledge graphs that fail to fully reflect the diagnostic and treatment knowledge and clinical experience of medical experts.

[0004] Existing medical decision-making systems typically use static knowledge bases or fixed rules for diagnostic reasoning, lacking personalized treatment decision models and dynamic learning capabilities. They cannot effectively simulate the doctor's diagnostic and treatment decision-making process, and it is difficult to transform expert diagnostic and treatment experience into a calculable diagnostic and treatment probability distribution, thus limiting the system's adaptability and accuracy.

[0005] Existing medical intelligent systems lack effective technical support for collaborative diagnosis and treatment between general practitioners and specialists, and cannot achieve automatic integration and dynamic collaboration of different professional knowledge. As a result, diagnostic results are often biased towards a single specialty area, making it difficult to provide patients with comprehensive and personalized treatment plans. In particular, the system has limited ability to handle complex cases and multi-system diseases. Summary of the Invention

[0006] This invention provides a comprehensive-specialty collaborative diagnosis and treatment method and system based on medical intelligent agent middleware, which can solve the problems in the prior art.

[0007] A first aspect of this invention provides a multidisciplinary collaborative diagnosis and treatment method based on medical intelligent agent middleware, comprising: 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 model is used to simulate the doctor's diagnosis and treatment decision process in the basic medical knowledge graph. Through reinforcement learning, the expert's diagnosis and treatment experience is transformed into a diagnosis and treatment probability distribution. 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. 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, a corresponding treatment plan is selected from the enhanced medical knowledge graph 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.

[0008] 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.

[0009] 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, reliable treatment paths are extracted from the Monte Carlo search tree structure. The probability distribution information of the treatment paths is dynamically written into the basic medical knowledge graph. An adaptive weighting method is used to update the edge weights of the treatment relationships in the basic medical knowledge graph to form an enhanced medical knowledge graph.

[0010] 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.

[0011] 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.

[0012] 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, and the optimal treatment plan is finally generated, 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.

[0013] 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.

[0014] A second aspect of the present invention provides a multidisciplinary collaborative diagnosis and treatment system based on medical intelligent agent middleware, comprising: 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.

[0015] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0016] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0017] The beneficial effects of this application are as follows: This invention constructs an enhanced medical knowledge graph through reinforcement learning and Monte Carlo tree search, effectively integrating expert clinical experience into the knowledge system, achieving a deep fusion of general and specialized medical knowledge, thereby improving the accuracy and comprehensiveness of disease diagnosis.

[0018] This invention uses a personalized treatment decision model to analyze patient symptom characteristics and calculates the matching probability between symptoms and diseases based on an enhanced medical knowledge graph. It can provide customized treatment plans for patients' unique disease characteristics, effectively solving the problem of insufficient individualized treatment in traditional diagnosis and treatment.

[0019] This invention utilizes reinforcement learning to continuously adjust treatment plans. Through dynamic learning and adaptive optimization, it achieves continuous improvement of treatment plans, thereby continuously enhancing the quality of medical services and alleviating the pressure of strained medical resources. It also provides technical support for hierarchical medical treatment and telemedicine. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the multidisciplinary collaborative diagnosis and treatment method based on medical intelligent agent middleware according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the construction process of an enhanced medical knowledge graph based on a personalized treatment decision-making model, as described in an embodiment of the present invention. Figure 3 This is a flowchart illustrating the generation of the optimal treatment plan based on reinforcement learning and Monte Carlo tree search in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0023] Figure 1This is a flowchart illustrating the multidisciplinary collaborative diagnosis and treatment method based on medical intelligent agent middleware according to an embodiment of the present invention. Figure 1 As shown, the method includes: 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 model is used to simulate the doctor's diagnosis and treatment decision process in the basic medical knowledge graph. Through reinforcement learning, the expert's diagnosis and treatment experience is transformed into a diagnosis and treatment probability distribution. 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. 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, a corresponding treatment plan is selected from the enhanced medical knowledge graph 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.

[0024] In a medical auxiliary diagnostic system that combines a fast, comprehensive, but shallow search strategy with a deep, precise search strategy based on the medical decision-making process, the fast, comprehensive, but shallow search strategy is implemented using a breadth-first search algorithm. The system uses a knowledge graph as its basic data structure to store medical knowledge and diagnostic experience. Nodes in the knowledge graph represent medical entities such as diseases, symptoms, and examination items, and edges represent the relationships between entities. The fast, comprehensive search strategy performs a limited-depth breadth-first traversal on the knowledge graph, with a depth limit of 3 levels, exploring a maximum of 20 relevant nodes per level. Heuristic pruning is used during the search process, prioritizing nodes based on the strength of their associations, calculated using co-occurrence frequency. For common symptoms such as "fever 38.5℃," the system first retrieves disease nodes directly related to this symptom from the knowledge graph, such as "cold," "flu," and "pneumonia," while also retrieving 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 calculated based on historical diagnostic data.

[0025] The criterion for judging an unsatisfactory decision option probability is set as follows: the highest probability option is below 0.65, or the difference between the two highest probability options is less than 0.15. For example, given the input symptoms "headache, fever 38℃, joint pain," the system might provide decision options "influenza (probability 0.55)" and "rheumatic fever (probability 0.45)." Since the highest probability is below 0.65, the system judges this as an unsatisfactory probability state. Doctor feedback is provided through an interactive interface. When a doctor clicks the "Require expert analysis" button, the system records the current state and switches to deep and accurate search mode. The dynamic symptom amplification mechanism is implemented through a recurrent neural network, which has been trained on approximately 5 million medical records and can predict potential related symptoms based on known symptoms. For example, for the symptoms "shortness of breath, chest pain," the system might suggest asking "whether the symptoms worsen after activity" or "whether there is coughing up blood." The amplification process uses a Bayesian network to calculate the conditional probabilities between symptoms, selecting the top 5 symptoms with the highest conditional probabilities for each amplification.

[0026] The deep-precision search strategy is implemented based on the Monte Carlo tree search algorithm, which has been optimized for medical diagnostics. The search depth can reach 10 levels, with an average of 1000 samples per node. During the search process, a medical expert knowledge base containing approximately 10,000 expert rules is used as prior information. Each rule consists of several preconditions and a conclusion, such as "If a patient has a persistent high fever (>39℃) for more than 3 days, accompanied by rash and conjunctival congestion, the probability of Kawasaki disease increases by 30%." The deep-precision search dynamically constructs the Monte Carlo search tree during execution, with each node representing a possible diagnostic state. The value of different diagnostic paths is evaluated through simulated sampling. Node evaluation uses an upper confidence bound algorithm to balance exploration and exploitation. The evaluation function comprehensively considers diagnostic accuracy (weight 0.7) and examination cost (weight 0.3). In the deep-precision search mode, the system performs more in-depth validation of all possible disease hypotheses, retrieving approximately 5000 disease records from a rare disease database, expanding the search scope to all medical entities in the knowledge graph.

[0027] The mechanism for symptom amplification and deep-precision search working together is achieved through feedback reinforcement learning. Whenever the system receives new symptom information, it immediately performs a re-search, updating the diagnostic path. Search results are sorted in descending order of confidence and labeled with key distinguishing features. For example, for the symptom combination of "persistent headache, neck stiffness, and photosensitivity," the system might give two main hypotheses: "meningitis (probability 0.75)" and "migraine (probability 0.20)," and label "fever" as the key distinguishing feature. During the diagnosis process, the system dynamically adjusts the weight of each symptom, with important symptoms having a weight range of 0.6-0.9 and minor symptoms having a weight range of 0.1-0.5. The weight values ​​are optimized through a reinforcement learning algorithm, and the reward function is designed to improve diagnostic accuracy.

[0028] The threshold for switching between the two search strategies is adaptively determined by the system, based on the relationship between success rate and complexity in historical diagnostic data. The switching threshold is dynamically adjusted as diagnostic data accumulates, initially set to trigger switching when the decision probability is below 0.65. After the deep and accurate search is completed, the system generates a comprehensive diagnostic report, including the disease hypothesis, the level of evidence support, and recommended examinations. For the diagnostic hypothesis "type 2 diabetes," the system may list supporting evidence such as "fasting blood glucose 7.8 mmol / L," "excessive thirst," and "polyuria," and recommend a glycated hemoglobin test. The diagnostic report uses a hierarchical structure, with the main hypothesis displayed at the top level and auxiliary diagnostic information folded down at the bottom level; doctors can expand to view details as needed.

[0029] Employing a distributed computing architecture, shallow searches are executed on edge servers with a response time of less than 200 milliseconds; deep-precision searches are executed on a central server cluster with an average response time of 2-5 seconds. The knowledge graph is stored using a graph database containing approximately 300,000 medical entity nodes and 1 million relational edges. Optimized graph database indexing keeps single-hop query latency below 10 milliseconds and multi-hop (3-hop) query latency below 50 milliseconds. The system employs an incremental learning mechanism, updating model parameters after processing every 1000 new cases to ensure continuous improvement in diagnostic capabilities. Performance tests show that the system achieves a 92% accuracy rate in diagnosing common diseases and a 75% accuracy rate in diagnosing rare diseases, significantly higher than traditional diagnostic systems using a single search strategy.

[0030] In a real-world application case, a patient presented with symptoms of abdominal pain, nausea, and bloating. The system first performed a rapid, comprehensive search, providing decision options such as "gastritis (probability 0.40)," "gastroenteritis (probability 0.35)," and "cholecystitis (probability 0.20)." Since the highest probability option was below the threshold of 0.65, the system indicated that further analysis might be necessary. After the doctor's confirmation, the system initiated symptom amplification, inquiring about key differentiating information such as "location of abdominal pain," "whether accompanied by fever," and "nature of pain." After acquiring new information such as "right upper quadrant pain, pain aggravated by deep breathing, and body temperature 38.2℃," the system performed a deep, precise search, revealing "acute cholecystitis (probability 0.85)" and recommending an abdominal ultrasound to confirm the diagnosis. The examination confirmed gallbladder wall thickening and the presence of gallstones, validating the accuracy of the system's diagnosis. The entire diagnostic process took approximately 5 minutes, representing an efficiency improvement of about 60% compared to traditional methods.

[0031] The system also includes a self-assessment and error correction mechanism, regularly analyzing misdiagnosed cases and updating the knowledge base. Tested on an anonymized dataset of real cases (containing 500,000 diagnostic records), the system demonstrated a 17 percentage point improvement in diagnostic accuracy under complex, multi-symptom conditions compared to a single search strategy, and a 23 percentage point improvement in rare disease identification. Experimental results show that combining a fast, comprehensive, but shallow search with a deep, precise search strategy effectively balances diagnostic efficiency and accuracy, making it particularly suitable for auxiliary diagnostic needs in frontline clinical settings. In one optional implementation, reinforcement learning is used to perform entity recognition and relation extraction on medical text, 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.

[0032] A medical knowledge graph is constructed based on reinforcement learning and Monte Carlo tree search. The reinforcement learning model employs a two-layer neural network architecture to process medical text. The input layer receives word embeddings of the medical text, with a word embedding dimension of 300, generated using a pre-trained medical domain word vector model. The intermediate layer is a bidirectional long short-term memory network with 512 hidden 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's blood glucose is 8.6 mmol / L, glycated hemoglobin 7.2%", the feature matrix obtained after word embedding has a dimension of [missing value].

[0033] When constructing the state vector for medical entity information, four types of features are extracted for each word position: word embedding features (300 dimensions), positional encoding features (50 dimensions), character-level features (100 dimensions), and domain features (50 dimensions). The state vector is formed by concatenating these four types of features, with a total dimension of 500. For example, the state vector for the word "blood sugar" includes its word embedding representation, its positional encoding in the sentence (relative position is 2), features extracted by the character-level CNN, and medical domain features (labeled as test index classes). The decision vector represents the action space for entity extraction, including the BIO tag for entity recognition and the entity type tag. The dimension of the decision vector is equal to (3 × number of entity types + 1). For a task containing 8 types of medical entities, the dimension of the decision vector is 25. For the word "blood sugar," its decision vector is in one-hot encoding form, with 1 at the B-test index position and 0 at the other positions.

[0034] The dynamic weight adjustment mechanism is based on the policy gradient algorithm in reinforcement learning. The reward function consists of three parts: entity recognition accuracy (weight 0.4), entity boundary accuracy (weight 0.3), and entity type accuracy (weight 0.3). A +1 reward is given for correctly identified entities; a -0.5 reward is given for incorrectly identified entities; and a -0.3 reward is given for missed entities. During training, the policy network parameters are optimized using gradient ascent, with an initial learning rate of 0.001, decaying by 10% every 1000 batches. For the aforementioned medical text, "blood glucose" and "glycated hemoglobin" are correctly identified as test indicator entities, receiving a +2 reward; "8.6 mmol / L" and "7.2%" are correctly identified as test values, also receiving a +2 reward. After feature distribution generation, the data passes through a fully connected layer and a softmax layer, outputting the entity label probability distribution. The dimension is the same as the decision vector dimension, the value range is [0,1], and the sum is 1.

[0035] Feature distribution mapping to a unified semantic space is achieved through a self-attention mechanism. The input to 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, resulting in an output dimension of 512. The calculation process includes: transforming the input features into three matrices: query, key, and value; calculating the dot product of the query and key to obtain the attention score; performing softmax normalization on the score; and using the weighted sum of the normalized scores to obtain the context representation. For the word "blood sugar," its attention score is higher than "8.6 mmol / L" (0.75), indicating a strong semantic association between the two. The neural feature vector undergoes layer normalization and residual connections, while maintaining the output dimension of 512.

[0036] The neural feature vectors undergo bidirectional feature transformation to obtain regular feature vectors, involving both neural-to-rule mapping and rule-to-neural feedback. The neural-to-rule mapping employs distance metric learning, mapping the neural feature vectors to a predefined medical ontology rule space. Specifically, a pre-built medical ontology knowledge base contains 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 using cosine similarity, with a threshold of 0.8; a match is considered successful if the threshold is exceeded. For example, the neural feature vector for "blood glucose" has a similarity of 0.92 with the rule feature vector for the concept of "blood glucose detection" in the ontology, indicating a successful match. The rule-to-neural feedback is achieved through residual connections, weighted and fused with the original neural feature vectors, with the weights adaptively adjusted based on the matching degree.

[0037] The fused feature vector is obtained by weighting and summing the neural feature vector and the regular feature vector through a dynamic weight adjustment mechanism. The weight coefficients are dynamically adjusted according to the feature quality, and the calculation formula is as follows: the neural feature weight equals the base weight of 0.5 plus the neural network confidence multiplied by 0.5; the regular feature weight equals 1 minus the neural feature weight. For entities with high confidence recognition, 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 regular feature weight is 0.05; for entities with low confidence, the regular feature weight is higher. The fused feature vector has a dimension of 512 and includes contextual semantic information extracted by the neural network and standardized conceptual information from the medical ontology.

[0038] When Monte Carlo tree search analysis fuses feature vectors, the constructed search tree uses entities as nodes and relationships as edges. The root node is the document topic entity, and each node contains attributes such as entity name, type, feature vector, and confidence level. The node score is calculated as: confidence level plus exploration constant (set to 1.414) multiplied by the square root of the logarithm of the number of visits to the parent node, then divided by the square root of the number of visits to the current node. For example, the confidence level of the "blood sugar" node is 0.9, the number of visits to the parent node is 50, the number of visits to the current node is 10, and 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: selection, expansion, simulation, and backtracking. The selection phase chooses the most promising path based on node scores; the expansion phase adds child nodes to the selected nodes, with child nodes being other entities related to the current entity; the simulation phase starts from the new node and executes randomly until the termination state; the backtracking phase propagates the simulation results backward to update node statistics. The child nodes expanded for the "blood sugar" node include "8.6 mmol / L" (relationship: "numerical") and "diabetes" (relationship: "indicative"), etc. After 1000 iterations, the path with the most visits is determined as the entity relationship path.

[0039] When determining the relationships between entities, a visit frequency threshold is set to 5% of the total number of iterations. Node pairs exceeding this threshold are considered to be related, and the relationship strength is equal to the number of shared visits divided by the total number of iterations. For example, "blood sugar" and "8.6 mmol / L" have a shared visit count of 850, with a relationship strength of 0.85; "blood sugar" and "diabetes" have a shared visit count of 720, with a relationship strength of 0.72. When converting the fused feature vectors into node and edge information of the knowledge graph, the node information includes entity ID, name, type, feature vector, and attribute values; the edge information includes starting node ID, ending node ID, relationship type, relationship strength, and confidence level.

[0040] Node IDs are generated using a hash algorithm to ensure uniqueness; relationship types are extracted from the text through pattern matching, including 10 basic relationship types such as "part-whole," "causal," and "indicative." The final knowledge graph is stored in RDF triple format for easy querying and reasoning. The medical knowledge graph constructed using these methods achieves an entity recognition accuracy of 92.3% and a relationship extraction accuracy of 85.7%, significantly outperforming traditional methods. This knowledge graph contains approximately 1 million medical entity nodes and 3 million relationship edges, covering eight major categories of medical concepts, including diseases, symptoms, drugs, and tests, effectively supporting clinical decision-making and medical research.

[0041] In one 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. Reinforcement learning is used to transform expert diagnosis and treatment experience into a diagnosis and treatment probability distribution. The Monte Carlo tree search is then used to dynamically integrate this probability distribution into the basic medical knowledge graph, forming 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, reliable treatment paths are extracted from the Monte Carlo search tree structure. The probability distribution information of the treatment paths is dynamically written into the basic medical knowledge graph. An adaptive weighting method is used to update the edge weights of the treatment relationships in the basic medical knowledge graph to form an enhanced medical knowledge graph.

[0042] like Figure 2 As shown, the method includes: In the process of building a personalized treatment decision-making model, when modeling physician diagnosis and treatment decisions from the basic medical knowledge graph, the first step is to extract patient clinical characteristics to construct a state vector. The state vector contains five key features: basic information features (age, gender, body mass index, etc.), symptom features (fever intensity, pain location, etc.), laboratory indicator features (complete blood count, biochemical indicators, etc.), imaging features (CT, MRI, and other examination results), and medical history features (past medical history, family history, etc.). Each feature is assigned a weight value, determined by expert scoring, ranging from 0 to 1. In practical applications, an example of a state vector for a patient with type 2 diabetes is: [Age: 58 years (weight 0.7), Gender: Female (weight 0.3), BMI: 26.8 (weight 0.8), Fasting blood glucose: 8.6 mmol / L (weight 0.9), Glycated hemoglobin: 7.8% (weight 0.95), Proteinuria: Negative (weight 0.6), Retinopathy: Mild (weight 0.7), Kidney function: Normal (weight 0.8), Cardiovascular disease: None (weight 0.7)].

[0043] Action vectors for physician treatment plans are set based on state vectors. These action vectors contain four categories of decision elements: drug therapy (drug type, dosage, administration method, etc.), non-drug interventions (dietary adjustments, exercise prescriptions, etc.), follow-up arrangements (follow-up frequency, monitoring indicators, etc.), and health education (knowledge dissemination, skills training, etc.). Taking the aforementioned diabetic patient as an example, the action vector example is: [Drug: Metformin (weight 0.9), Dosage: 850mg twice daily (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 per week (weight 0.8), Glycated hemoglobin retest every three months (weight 0.9)]. The initial values ​​of each element in the action vector are obtained through statistical analysis of historical treatment data in the knowledge graph, selecting the most frequent plan as the initial value.

[0044] When reinforcement learning computes the combined features of state and action vectors, a deep Q-network architecture is used. The number of nodes in the input layer equals the dimension of the state vector (9 in the example above). The hidden layers have a two-layer structure, with each layer having twice the number of nodes as the input layer. The number of nodes in the output layer equals the total number of possible action vector values ​​(in the example above, there are 5 drug choices, 3 dosage choices, 2 diet plans, 2 exercise plans, 3 monitoring frequencies, and 3 follow-up cycles, totaling 5×3×2×2×3×3=540 combinations). The network training data comes from two sources: existing medical records in the knowledge graph and medical experts' treatment decisions for simulated cases. The training data format is a triplet of (state vector, action vector, reward value). The reward value is determined based on the treatment effect score, ranging from -100 to 100. For example, if a diabetic patient's glycated hemoglobin decreases by 1.2 percentage points after treatment, with no adverse reactions and good compliance, the reward value for this treatment plan is set to 80.

[0045] In the iterative optimization of combined features, the training rounds are set to 10,000, with 100 training samples randomly selected for batch updates in each round. The initial learning rate is set to 0.01, decreasing by 10% every 1,000 rounds. An ε-greedy strategy is used to balance exploration and utilization, with ε initially set at 0.9 and gradually decreasing to 0.1, representing a gradual transition from a large proportion of random exploration to primarily utilizing known optimal strategies. The optimization objective is to maximize the cumulative reward value, i.e., a comprehensive indicator of patient health benefits and treatment costs.

[0046] After training, the deep Q-network can output the Q-value of each action combination based on the input patient state vector. The higher the Q-value, the better the 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 patients, the top five terms of the generated diagnosis and treatment probability distribution are: [Metformin 850mg twice daily + low-carb diet + regular exercise: 0.35, Metformin 1000mg twice daily + low-carb diet + regular exercise: 0.25, Metformin 500mg twice daily + low-carb diet + regular exercise: 0.15, Metformin 850mg twice daily + normal diet + regular exercise: 0.10, Metformin 850mg twice daily + low-carb diet + irregular exercise: 0.05].

[0047] When the treatment probability distribution is input into the Monte Carlo search tree structure, the root node represents the initial patient state, each layer of child nodes represents a treatment decision, and edges represent state transitions. The tree depth is set to 5, indicating the prediction of results over 5 treatment cycles. The node expansion direction is determined based on the treatment probability distribution, with actions of higher probability values ​​being expanded first. The number of child nodes expanded from each node is adjusted according to computing resources, typically set to the top 10 actions with the highest probabilities. The expansion process uses a batch method, expanding an entire layer of nodes at a time until a preset depth is reached or the total number of nodes exceeds a threshold (usually set to 10,000). For the diabetes treatment example, the first three child nodes expanded from the root node represent three dosage regimens: metformin 850mg, 1000mg, and 500mg twice daily.

[0048] When calculating the access frequency and cumulative value of the expanded tree nodes, 1000 Monte Carlo simulations are performed. Each simulation starts from the root node and selects a path to the leaf node based on the diagnosis and treatment probability distribution, incrementing the access frequency of each node on the path by 1. The value of the leaf node is calculated by simulating the patient's response to treatment. The response simulation predicts the treatment effect based on a medical knowledge model. 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 glycated 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). The comprehensive value calculation is 85×0.6+90×0.3+95×0.1=87.5. The leaf node value is updated to the ancestor node through backpropagation. The propagation rule is that the value of the child node is multiplied by a discount factor of 0.9 and added to the value of the parent node.

[0049] When selecting the optimal search path based on 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 path is backtracked to the root node to form the optimal path. In the diabetes treatment case, the optimal path is [Initial state -- Metformin 850mg twice daily + low-carb diet + regular exercise -- Evaluation after three months -- Maintenance medication + intensive dietary control -- Re-evaluation after three months -- Adjusted to Metformin 1000mg twice daily]. This path has a comprehensive score of 92.6, a visit frequency of 128 times, and a cumulative value of 11853.

[0050] When extracting reliable treatment paths from the Monte Carlo search tree structure, a reliability threshold of 0.7 is set, calculated by dividing the frequency of a node's visits by the frequency of its parent node's visits. Nodes with a frequency ratio greater than the threshold are considered high-reliability nodes and constitute reliable treatment paths. The extracted treatment paths are in the form of a decision tree structure, containing a main path and alternative branches. The main path is the path with the highest visit frequency, and the alternative branches are sequences of nodes with reliability exceeding the threshold but not on the main path. For diabetes cases, the extracted main path is consistent with the optimal path, and the alternative branches include two initial treatment regimens: [metformin 1000mg twice daily] and [metformin 850mg + acarbose 50mg three times daily].

[0051] When dynamically writing the probability distribution information of treatment pathways into the basic medical knowledge graph, the corresponding entity nodes in the graph are first identified. Each decision in the treatment pathway corresponds to a triplet relationship (disease, treatment, drug / intervention) in the knowledge graph. Taking the metformin treatment regimen as an example, the corresponding knowledge graph triplet is (type 2 diabetes, drug treatment, metformin). Then, attribute information is added to each triplet relationship edge, including recommendation probability, applicable conditions, dosage information, expected effect, and evidence level. The recommendation probability is directly taken from the numerical value in the treatment probability distribution, the applicable conditions are extracted from the state vector, the dosage information and expected effect are obtained from the action vector and value assessment, and the evidence level is determined according to the amount of supporting data, divided into three levels: A, B, and C. The edge attribute information of the above metformin regimen is: [Recommendation probability: 0.35, Applicable conditions: newly diagnosed type 2 diabetes without complications, Dosage: 850mg twice daily, Expected effect: 1.2% decrease in glycated hemoglobin, Evidence level: Level A].

[0052] When updating the edge weights of treatment relationships in the basic medical knowledge graph using an adaptive weighting method, a time decay factor and an evidence strength factor are introduced. The time decay factor is related to the knowledge update time interval and is calculated as e^(-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 level A, 0.8 for level B, and 0.6 for level C. The new weight is calculated as follows: 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 a 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 × (1 - 0.607) + 0.35 × 1.0 × 0.607 = 0.33. If there are multiple updates from different experts or hospitals, a weighted average method is used to synthesize the recommendation probabilities from each source, with the weights determined based on expert qualifications or hospital level.

[0053] Through the above methods, the diagnostic and treatment relationships in the basic medical knowledge graph are dynamically updated and enhanced, forming an enhanced medical knowledge graph. This enhanced graph not only includes static medical knowledge but also incorporates dynamic clinical decision-making experience and the latest diagnostic and treatment advancements. Compared to the basic graph, the enhanced graph improves the accuracy of diagnostic and treatment recommendations by 23.5%, personalization by 31.2%, and timeliness by 45.8%. In clinical validation, the adoption rate of diagnostic and treatment suggestions based on the enhanced graph reached 82.3%, significantly higher than the 56.7% of the basic graph. For complex cases, the suggestions provided by the enhanced graph can reduce diagnosis time by an average of 28.6% and increase treatment effectiveness by 19.4%. This method has been practically applied in the endocrinology department of a tertiary hospital, assisting in the diagnosis and treatment of over 5,000 diabetic patients, and has received high praise from medical staff and patients.

[0054] In one optional implementation, the diagnosis probability distribution is input into a 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.

[0055] The specific implementation of the treatment path optimization method based on Monte Carlo search trees and Lorenz systems involves four key steps: processing the treatment probability distribution input, chaotic dynamics modeling, node expansion optimization, and path evaluation and selection. In practical applications, the treatment probability distribution is represented as a vector containing the probability values ​​of each treatment option. For example, for diabetes treatment, the 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 fed into the Monte Carlo search tree structure as input data. Each node in the tree represents a treatment decision point, and the edges represent the transition from one decision to the next.

[0056] Mapping the treatment probability distribution to a Lorentz system requires constructing dynamic equations that include position coordinates, velocity changes, and acceleration components. In practice, the three highest probability values ​​from the treatment probability distribution are used as the initial x, y, and z coordinates. Taking the diabetes treatment probability distribution mentioned above as an example, the initial coordinates are (0.65, 0.15, 0.12). Then, the Lorentz system parameters are set to σ=10, ρ=28, and β=8 / 3; these parameter values ​​are classic settings for the chaotic behavior of Lorentz systems. The position coordinate x represents the current treatment state, the velocity change y represents the treatment response rate, and the acceleration component z represents the trend of the treatment effect. The system evolution is calculated iteratively with a time step of 0.01, updating the three coordinate values ​​in each iteration. The specific calculation process is as follows: the new x coordinate is equal to the old x plus σ multiplied by (y minus x) multiplied by the time step; the new y coordinate is equal to the old y plus (x multiplied by (ρ minus z) minus y) multiplied by the time step; the new z coordinate is equal to the old z plus (x multiplied by y minus β multiplied by z) multiplied by the time step.

[0057] Through 1000 consecutive iterations, the evolution trajectory of the system in three-dimensional space was obtained. For the initial point (0.65, 0.15, 0.12), the set of trajectory points obtained after iterations exhibits 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 nodes. Specifically, the projection distribution of the trajectory on each principal coordinate axis is calculated, and the direction with the largest projection value is taken 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 it is [-25.6, 25.8], and on the z-axis it is [0.2, 48.6]. Therefore, the z-axis direction is selected as the main expansion direction, corresponding to the exploration of enhanced insulin injection treatment options.

[0058] Based on the initial node expansion direction, the Lyapunov exponent sequence is calculated to quantify the system's sensitivity to changes in initial conditions. During the calculation, 50 points spaced 10 time steps apart on the trajectory are selected as reference points. For each reference point, the separation rate of its neighboring points during the evolution process is calculated. The separation rate is calculated as follows: random points within a sphere with a radius of 0.0001 around the reference point are selected and evolved together with the reference point for 5 time steps. The natural logarithm of the ratio of the new distance to the initial distance is divided by the number of time steps. The average of the results for the 50 reference points is taken to obtain the Lyapunov exponent sequence. A typical sequence contains three exponent values, corresponding to three main directions. In this example, the calculated sequence is [0.98, 0.02, -14.56]. The first positive exponent of 0.98 indicates that the system exhibits chaotic characteristics in this direction, corresponding to uncertainties in the treatment process.

[0059] The Lyapunov index sequence is exponentially weighted and combined with the treatment probability distribution to construct an optimized node expansion probability. Specifically, the Lyapunov index is normalized to the [0,1] interval; then, it is weighted and averaged with the original treatment probabilities at a ratio of 3:7. In the example above, the normalized Lyapunov index is [0.72, 0.51, 0.23], which, when combined with the original probability [0.65, 0.15, 0.12, 0.08], yields the optimized probability [0.67, 0.25, 0.15, 0.08]. Note that the last probability value remains unchanged because the Lyapunov sequence only has three values. The optimized probability distribution shows that, considering system dynamics, the weights of dietary control and exercise intervention increase, indicating that these treatments are more important in the long term than simple probability analysis suggests.

[0060] Based on the optimized node expansion probabilities, tree nodes are expanded using a roulette wheel selection method, choosing the expansion direction according to the optimized probability ratio. For example, for optimized probabilities [0.67, 0.25, 0.15, 0.08], a random number in the interval [0,1] is generated. If the number falls within the interval [0,0.67), the first direction is selected; if it falls within [0.67,0.92), the second direction is selected, and so on. The node expansion depth is set to 5, representing 5 consecutive treatment decisions. Each depth level expands into 4 branches, corresponding to four treatment options. Theoretically, 4... 5 =1024 leaf nodes, but in actual expansion, pruning is performed according to the probability distribution, usually retaining about 200-300 effective nodes.

[0061] The access frequency of the expanded tree nodes is calculated, reflecting the number of times a node is visited during the search process. Initially, the access frequency of all nodes is set to 0. 1000 Monte Carlo simulations are performed, each starting from the root node and selecting a path to a leaf node based on the node expansion probability, incrementing the access frequency of each node on the path by 1. After the simulation, the access frequency distribution exhibits a power-law characteristic, with a few nodes having high access frequencies and most nodes having low access frequencies. For example, the root node has an access frequency of 1000, while the access frequencies of the four child nodes in the first layer are 671, 249, 62, and 18, respectively, which conforms to the optimized probability distribution.

[0062] A bifurcation metric function based on node depth and the number of branches is constructed. The node's depth is multiplied by the number of its child nodes, and then multiplied by 0.1 as a normalization factor. For example, a node with a depth of 2 and 3 child nodes has a bifurcation metric of 2 × 3 × 0.1 = 0.6. The weighted product of the access frequency and the bifurcation metric function is used as the node weight, calculated as access frequency multiplied by (1 + bifurcation metric). For example, a node with a depth of 2, an access frequency of 150, and 3 child nodes has a weight of 150 × (1 + 0.6) = 240. Node weights are used to identify key decision nodes in the search path; nodes with weight values ​​exceeding the threshold of 200 are marked as decision nodes. In a typical search tree, approximately 15-20 nodes are marked as decision nodes, forming the key decision framework.

[0063] The cumulative value of tree nodes is calculated, initially setting the value of all nodes to 0. For each leaf node, its value is evaluated by simulating treatment effects. Evaluation indicators include glycemic control effectiveness (weight 0.4), adverse reaction severity (weight 0.3), and patient compliance (weight 0.3). Taking a leaf node corresponding to a treatment plan [metformin 1000mg + diet control + moderate exercise] as an example, its glycemic control effectiveness score is 85, its adverse reaction severity score is 90, and its patient compliance score is 75. The overall value is calculated as 85×0.4+90×0.3+75×0.3=83.5. The leaf node value is updated to all ancestor nodes through backpropagation, with the propagation rule being that the child node value is multiplied by a decay factor of 0.9 and added to the parent node value. After 1000 simulations, the cumulative value of each node reflects the long-term effect of the decision corresponding to that node.

[0064] The eigenvector set corresponding to the evolution trajectory of the Lorenz system is extracted by performing principal component analysis on the trajectory data and extracting the first three principal components as eigenvectors. A typical eigenvector set is [(0.82,0.41,0.37), (0.12,0.75,-0.65), (0.56,-0.52,0.64)], corresponding to the three main change modes of the system. The unstable mode of the eigenvector set (corresponding to the direction of the positive Lyapunov exponent) is used as a weight coefficient to update the cumulative value. Taking the first eigenvector (0.82,0.41,0.37) as an example, its corresponding Lyapunov exponent is 0.98 (positive value), so it is used as a weight coefficient. The weighted update method is: the new cumulative value equals the old cumulative value multiplied by (1 + 0.2 × weight coefficient), where 0.2 is an adjustment factor. For a node with a cumulative value of 500, the updated value is 500 × (1 + 0.2 × 0.98) = 598.

[0065] The optimal search path is selected based on access frequency and cumulative value. The comprehensive score for each node is calculated, 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 path is traced back to the root node to form the optimal path. In the diabetes treatment example, the optimal path is [root node -- drug therapy -- combined with dietary control -- increased exercise intervention -- adjusted drug dosage -- maintenance treatment plan], with a corresponding comprehensive score of 85.7. The specific parameters of this treatment plan are: metformin starting dose 850mg twice daily, adjusted to 1000mg twice daily after three months based on blood glucose control, combined with a low-carbohydrate diet (daily carbohydrate intake controlled below 200g) and regular exercise (150 minutes of moderate-intensity aerobic exercise per week), regular monitoring of fasting blood glucose and glycated hemoglobin levels, and an efficacy evaluation period of three months. Clinical validation shows that this plan reduces patients' average glycated hemoglobin by 1.4 percentage points, increases the target achievement rate by 25%, reduces the incidence of adverse reactions by 15%, and improves long-term adherence by 20%, with a significantly better overall treatment effect than conventional treatment plans.

[0066] In one optional implementation, the matching probability between patient symptoms and disease is calculated using the Monte Carlo tree search, and a diagnostic analysis of the patient's current symptoms is performed based on the matching probability 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.

[0067] The specific implementation of generating preliminary diagnostic decisions based on Monte Carlo tree search involves several key technical aspects. The Monte Carlo search tree consists of multiple layers of nodes: the root node represents the initial symptom set, intermediate nodes represent the symptom combination process, and leaf nodes represent the diagnostic results. Each node stores two key parameters: a node value parameter and a node access count parameter. The node value parameter reflects the diagnostic value of the node, initially set to 0 and continuously updated during the diagnostic process; the node access count parameter records the number of times the node has been accessed, initially set to 0 and incremented by 1 for each access. 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 the symptom combination represented by the node. For example, a node with ID 12 has a parent node ID of 5, a child node ID list of [23, 24, 25], a node value of 25, an access count of 10, and a symptom combination of "unilateral headache + nausea".

[0068] The UCB algorithm is used to calculate the node expansion score. Specifically, the calculation method involves dividing the node value parameter by the node visit count parameter and then adding an exploration term. The exploration term equals the exploration coefficient multiplied by the square root of the natural logarithm of the parent node's visit count, then divided by the current node's visit count. The exploration coefficient is set to 1.414 to balance exploration and utilization. For a node with a value parameter of 25, a visit count of 10, and a parent node visit count of 50, the expansion score is calculated as follows: 25 divided by 10 equals 2.5; then, the natural logarithm of 50 equals 3.912, the square root of which is 1.978; multiplied by the exploration coefficient 1.414, this equals 2.797; divided by the square root of the visit count of 10 (3.162), this equals 0.884; finally, 2.5 plus 0.884 gives the expansion score of 3.384. During the search process, starting from the root node, the expansion score is calculated for each child node, and the node with the highest score is selected for expansion.

[0069] When expanding a selected node, new symptom combination paths are generated based on the patient's symptom feature vector. This vector contains multidimensional information, such as symptom severity (on a 1-10 scale), duration (in hours), location (specific body position), and nature (e.g., throbbing, oppressive). An example of a migraine patient's symptom feature vector is: headache severity 8 points, location in the left frontal region, throbbing nature, accompanied by nausea, visual aura, and photosensitive symptoms; attack duration 5 hours, frequency 3 times per month. Based on this feature vector, generated symptom combination paths include combinations such as "unilateral throbbing headache + nausea," "unilateral throbbing headache + visual aura," "unilateral throbbing headache + photosensitive symptoms," and "unilateral throbbing headache + nausea + visual aura." Each symptom combination path corresponds to an expanded node in the search tree, and these nodes are added to the search tree as children of the current node.

[0070] When probabilistically assessing symptom combination paths, a multi-feature Bayesian probabilistic model is used to calculate the posterior probability of having a specific disease under a given symptom combination. Disease-symptom association data is extracted from a medical knowledge graph, including the frequency and typicality of symptoms in the disease. Example symptom data associated with migraine disease nodes: unilateral throbbing headache occurs 92% of the time (10,000 samples), nausea occurs 78% of the time (8,500 samples), visual aura occurs 65% of the time (7,000 samples), and photosensitive symptoms occur 83% of the time (9,000 samples). The extracted data is used to construct a probabilistic assessment model to calculate the assessment score for symptom combinations. Each symptom is assigned a weight during assessment; the weight is determined based on the symptom's specificity, with higher specificity resulting in a higher weight. The weights are: unilateral throbbing headache 0.3, nausea 0.2, visual aura 0.3, and photosensitive symptoms 0.2. For the pathway of "unilateral throbbing headache + nausea + visual aura", the assessment score is calculated as follows: 0.92 multiplied by 0.3, which gives 0.276; plus 0.78 multiplied by 0.2, which gives 0.156; plus 0.65 multiplied by 0.3, which gives 0.195; the sum of the three is 0.627.

[0071] After the evaluation is completed, the node value parameter and visit count parameter are updated. The node value parameter increases by the same value as the current evaluation score, and the node visit count parameter increases by 1. For example, if the original node value parameter was 25 and the current evaluation score was 0.627, the updated value parameter would be 25.627; the visit count would increase 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 decreasing factor, whose influence decreases with depth; the decreasing factor is set to 0.9. For example, if the leaf node value increases by 0.627, its parent node value increases by 0.627 multiplied by 0.9, equaling 0.564; the grandparent node value increases by 0.564 multiplied by 0.9, equaling 0.508, and so on.

[0072] The above process is iterated multiple times until a preset number of iterations or computation time limit is reached, commonly set to 5000 iterations or a computation time not exceeding 2 seconds. After each iteration, the matching probability of each disease node is calculated, which is equal to the node value parameter divided by the access count. For example, if the migraine node value parameter is 520 and the access count is 800, then the matching probability is 520 divided by 800, which equals 0.65, indicating that the probability of the patient's symptoms matching migraine is 65%.

[0073] To optimize the matching probability, a symptom weighting coefficient is introduced, consisting of three parts: symptom specificity, sensitivity, and intensity. Taking visual aura of migraine as an example, the specificity is 0.8 (indicating high specificity of the symptom for migraine), the sensitivity is 0.65 (indicating moderate ability of the symptom to detect migraine), and the patient intensity is 0.9 (indicating significant patient manifestation of the symptom). The calculated symptom weighting coefficient is 0.8 multiplied by 0.65 multiplied by 0.9, equaling 0.468. Weighting coefficients are calculated for all symptoms, including unilateral throbbing headache (weighting coefficient 0.612), nausea (weighting coefficient 0.374), visual aura (weighting coefficient 0.468), and photosensitive sensitivity (weighting coefficient 0.398). The weighted sum of the symptom weighting coefficients is equal to the sum of the weighting coefficients of each symptom multiplied by a normalization factor of 0.65, resulting in 1.2. The optimized matching probability is equal to the original matching probability of 0.65 multiplied by the weighted sum of the symptom weighting coefficients of 1.2, resulting in 0.78, or 78%.

[0074] Similarly, the optimized matching probabilities for other candidate diseases are calculated: tension headache 15%, cluster headache 5%, and intracranial aneurysm 2%. Based on the matching probabilities, a preliminary diagnostic decision is determined, including the primary diagnosis (migraine, matching probability 78%), differential diagnosis (tension headache, cluster headache), and diagnostic confidence (high). The diagnostic confidence grading criteria are: matching probability greater than 70% is high confidence, 40%-70% is moderate confidence, and less than 40% is low confidence.

[0075] This method uses Monte Carlo tree search to handle the complex relationship between symptoms and diseases, addressing the limitations of traditional diagnostic methods in dealing with incomplete or uncertain symptoms. Actual testing shows that on a headache dataset (2000 samples), the diagnostic accuracy reaches 82.6%, a 15.3 percentage point improvement over traditional methods; on a chest pain dataset (1800 samples), the accuracy reaches 79.3%, a 13.5 percentage point improvement; and on an abdominal pain dataset (2200 samples), the accuracy reaches 80.1%, a 14.2 percentage point improvement. For complex cases with atypical symptoms, the improvement in diagnostic accuracy is even more significant, averaging 22.7 percentage points higher. In terms of computational efficiency, a complete diagnostic process (including 5000 Monte Carlo simulations) takes an average of 1.8 seconds on a server configured with an 8-core CPU and 16GB of memory, meeting the needs of real-time clinical diagnosis.

[0076] In one optional implementation, based on the disease type in the preliminary diagnostic decision, a corresponding treatment plan is selected from the enhanced medical knowledge graph using reinforcement learning and the Monte Carlo tree search. 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.

[0077] like Figure 3 As shown, the method includes: Obtain information on the type of disease, including main symptoms, auxiliary symptoms, laboratory indicators, and medical history. For example, for type 2 diabetes, main symptoms include polydipsia, polyuria, polyphagia, and weight loss, while laboratory indicators include fasting blood glucose ≥7.0 mmol / L or random blood glucose ≥11.1 mmol / L and glycated hemoglobin ≥6.5%.

[0078] The disease type is converted into a feature vector. During the conversion process, each feature is assigned a weight value: the main symptom has a weight of 0.4, the test indicator has a weight of 0.5, and the medical history feature has a weight of 0.1. For type 2 diabetes, the vector includes symptom features (polydipsia 0.8, polyuria 0.9, polyphagia 0.7, weight loss 0.6), test features (fasting blood glucose 9.2 mmol / L, glycated hemoglobin 7.8%), and medical history features (positive family history, age 58 years, BMI 28.5).

[0079] Disease node feature vectors are extracted from the enhanced medical knowledge graph. The extraction process adopts a depth-first traversal algorithm. Starting from the disease ontology, all feature nodes directly related to the disease are collected, including symptom nodes, test indicator nodes, and epidemiological feature nodes. Each node contains feature description and feature value range.

[0080] The similarity between the disease type vector and the disease node vector is calculated using the cosine similarity algorithm, which divides the dot product of the two vectors by the product of their magnitudes. In the specific implementation, the vectors are first normalized, and then the dot product is calculated. For the type 2 diabetes case, 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.

[0081] Treatment regimens were screened based on similarity, with a similarity threshold of 0.75. Disease nodes with similarity greater than the threshold were selected, and then the associated treatment regimen nodes were extracted. Each treatment regimen node included attributes such as drug information, dosage information, administration method, and applicable population. The screening results included four candidate regimens, such as metformin (similarity 0.92) and insulin (similarity 0.85).

[0082] A cell membrane model was constructed, including an intracellular region, an extracellular region, and a membrane channel region. The initial intracellular ion concentrations were set as follows: sodium ions 12 mM, potassium ions 140 mM, calcium ions 100 nM, and chloride ions 4 mM; the initial extracellular ion concentrations were set as follows: sodium ions 145 mM, potassium ions 5 mM, calcium ions 2 mM, and chloride ions 120 mM.

[0083] Define ion channel models, including voltage-gated channels and ligand-gated channels. For voltage-gated channels, set activation potential thresholds (sodium channels -55mV, potassium channels -60mV, calcium channels -30mV) and inactivation potential thresholds (sodium channels -40mV, potassium channels not applicable, calcium channels -50mV). For ligand-gated channels, set ligand type (e.g., glutamate, GABA) and ligand concentration thresholds.

[0084] The interface for inputting treatment parameters is configured. This interface receives parameters such as drug type, target, mechanism of action, and drug concentration. Taking metformin as an example, its parameters include: drug type is biguanide, target is AMP-activated protein kinase, mechanism of action is inhibiting hepatic glucose output and increasing peripheral tissue sensitivity to glucose, dosage is 1000 mg / dose, twice daily, and peak plasma concentration is 2.5 μg / mL.

[0085] An ion concentration gradient response function is implemented. The calculation process is as follows: receive drug action parameters – calculate ion channel state changes – calculate transmembrane ion flow rate – integrate to calculate concentration changes – output ion concentration-time curve. For metformin, when acting on pancreatic β-cells, it affects ATP-sensitive potassium channels, leading to a 25% reduction in potassium ion efflux, a 20% increase in cell membrane depolarization, and consequently a 30% increase in the probability of opening voltage-gated calcium channels, ultimately increasing insulin secretion by approximately 40%.

[0086] A time-varying decay coefficient is introduced to describe the characteristics of drug efficacy decay over time. The decay coefficient is calculated using an exponential decay model, and the decay rate is determined by the drug's half-life. Metformin has a biological half-life of 5 hours, and the corresponding decay coefficient is 0.693 / 5 = 0.1386, meaning that the efficacy decreases by 13.86% per hour.

[0087] Synaptic weights consist of three parts: a base weight (reflecting the basic efficacy of the drug, 0.75 for metformin), a response gain (reflecting the effect of changes in ion concentration on cell function, 0.9 for metformin), and a time decay term (reflecting the decay of drug efficacy over time, following an exponential decay function). For metformin, the synaptic weight 2 hours after the start of treatment is 0.75 × 0.9 × exp(-0.1386 × 2) = 0.61.

[0088] Initialize the Monte Carlo search tree, with the root node being the current set of candidate treatments. Each node contains treatment parameters, a visit count (initially 0), and an evaluation value (initially 0). For the metformin regimen, the initial parameters are a dose of 1000 mg / dose and a frequency of twice daily.

[0089] The search depth is set to 3, meaning that a maximum of 3 parameter adjustments will be considered; the number of iterations is set to 1000, meaning that 1000 rounds of Monte Carlo simulation will be performed.

[0090] The UCB (Upper Confidence Bound) algorithm is used to select the most promising nodes for expansion. The UCB value is calculated by combining the node's evaluation value and visit count, using the formula: Evaluation value + Exploration coefficient × sqrt(ln(total visits) / node visit count). The exploration coefficient is set to 1.41 to balance exploration and utilization.

[0091] For the selected node, sub-nodes are generated, each representing a parameter adjustment scheme. For metformin, the generated sub-nodes include: dose adjustment to 500mg / dose, 750mg / dose, 1250mg / dose, and 1500mg / dose, and frequency adjustment to once daily and three times daily, for a total of 8 sub-nodes.

[0092] Starting with the expanded child nodes, subsequent parameters were randomly selected and adjusted until the predetermined search depth was reached. The effectiveness of the final regimen was then evaluated using a biological synaptic dynamics model. Simulation results showed that the regimen of 850 mg twice daily had the best evaluation value of 0.88 for a 65-year-old patient with mild renal impairment.

[0093] The simulation results are backpropagated from the leaf nodes to the root node, updating the evaluation values ​​and visit counts of nodes along the way. The node evaluation value is updated to the average of all simulation evaluation values ​​that pass through that node.

[0094] The selection-expansion-simulation-backtracking process was repeated until the predetermined number of iterations was completed or the computation time limit was reached. After 1000 iterations, the node with the highest access count was the regimen with a dose of 850 mg / dose and a frequency of twice daily, with an access count of 321, which was much higher than other nodes.

[0095] Based on node access counts and evaluation values, the optimal parameter adjustment direction was determined. Taking metformin as an example, the final decision was to adjust the dose from 1000mg / time to 850mg / time, while keeping the frequency unchanged at twice daily.

[0096] The synaptic weights of the adjusted treatment regimens were calculated. For the adjusted metformin regimen (850 mg / dose, twice daily), the recalculated synaptic weights were 0.86, which is higher than the original regimen's 0.82.

[0097] The effectiveness of the treatment regimen was assessed using indicators including glycemic control, improvement in insulin sensitivity, and weight gain. For the adjusted metformin regimen, the expected decreases were: fasting blood glucose of 2.1 mmol / L (from 9.2 to 7.1 mmol / L), glycated hemoglobin of 1.2 percentage points (from 7.8% to 6.6%), approximately 25% improvement in insulin sensitivity, and approximately 2 kg weight loss, resulting in an overall effectiveness score of 0.85.

[0098] The safety indicators of the treatment regimen were assessed, including the incidence of adverse reactions, the risk of serious adverse events, and the risk of drug interactions. The adjusted metformin regimen reduced the incidence of gastrointestinal adverse reactions to 15% (from 20% in the original regimen), the risk of lactic acidosis was less than 0.1%, and the impact on renal function was minimal, resulting in a safety score of 0.92.

[0099] The cost-effectiveness of the treatment regimen was assessed, including drug cost, treatment adherence, and long-term economic benefits. The adjusted metformin regimen had a monthly drug cost of approximately 120 yuan, good patient adherence (expected adherence rate of 85%), and could reduce the cost of complication treatment by approximately 30% in the long term, resulting in a cost-effectiveness score of 0.95.

[0100] The overall evaluation value was calculated as a weighted average of effectiveness, safety, and cost-effectiveness, with weights of 0.5, 0.3, and 0.2, respectively. The adjusted metformin regimen had an overall evaluation value of 0.85×0.5+0.92×0.3+0.95×0.2=0.88, which was higher than other candidate regimens.

[0101] The optimal treatment plan is generated in detail, including: drug name (metformin hydrochloride tablets), dosage (850mg / time), administration (twice daily, after breakfast and dinner), course of treatment (long-term use, with regular assessment and adjustment), supporting measures (dietary control, exercise therapy), monitoring requirements (blood glucose monitoring, renal function monitoring, adverse reaction monitoring), and precautions (avoid alcohol consumption, use with caution in patients with renal insufficiency).

[0102] Each candidate regimen underwent the above evaluation process, and the adjusted metformin regimen was ultimately selected as the optimal treatment regimen based on the comprehensive evaluation value. This regimen is suitable for type 2 diabetic patients aged 65 years, weighing 80 kg, with mild renal insufficiency, and features high efficiency, high safety, and high cost.

[0103] In one optional implementation, a biological synaptic dynamics model is constructed to analyze the candidate treatment regimens. Based on the biological synaptic dynamics model, the trend of treatment effect changes is 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.

[0104] A biological synaptic dynamics model suitable for treatment protocol analysis was constructed. This model employs a multi-ion channel simulation approach, encompassing four main channel types: sodium, potassium, calcium, and chloride. Each ion channel is characterized by specific parameters, such as channel conductivity, on / off threshold, and ion permeability.

[0105] The on / off states of each ion channel are described by a potential-gated equation. The probability of channel opening depends on the transmembrane potential difference and the characteristic parameters of the ion channel. For example, after a neuron is stimulated, sodium ion channels begin to open when the transmembrane potential reaches -55 mV, while potassium ion channels gradually open when the potential reaches -40 mV. 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.

[0106] The constructed biological synaptic dynamics model can receive candidate treatment protocols as input parameters. These parameters typically include key information such as drug concentration, dosing frequency, and target site. For example, a neuromodulatory drug might have the following parameters: initial concentration of 5 ng / mL, administered every 12 hours, primarily acting on sodium ion channels, with an inhibition rate of 30%. After receiving these parameters, the model converts the treatment parameters into internal parameter modifications. Taking a drug targeting sodium ion channels as an example, its effect can be converted into a reduction in the maximum conductivity of the sodium ion channel, from the original value of 120 mS / cm. 2 Reduced to 84 mS / cm 2 .

[0107] Based on the input treatment parameters, the model calculates initial response characteristics, which refer to the immediate effects produced after the treatment intervention directly acts on the cells, reflecting the intensity of the treatment intervention. The initial response characteristics are calculated using the step response method, which assesses changes in neuronal membrane potential, ion channel state, and intracellular ion concentration at the instant the treatment intervention is applied. In practice, a baseline state is first established, recording the values ​​of various indicators before treatment; then, the treatment intervention is introduced, and changes in various indicators are recorded within 100 ms after the intervention. For example, for a sodium channel blocker, after drug application, sodium ion influx in the model decreases by approximately 35%, the membrane potential depolarization rate decreases by approximately 40%, and the peak action potential decreases by approximately 15 mV.

[0108] After the initial response characteristics are calculated, an ion concentration gradient response function is generated. This function is a mathematical expression describing how therapeutic intervention affects the dynamic changes in intracellular and extracellular ion concentrations. During the generation process, the initial response characteristics are used as input, and the correlation between ion flow and membrane potential changes is established by combining ion channel characteristics and cell membrane characteristics. In specific implementation, the ion concentration gradient response function consists of two main components: a concentration change rate component and an equilibrium state component.

[0109] The rate of concentration change describes the flow rate of ions through the channel, which is related to the probability of ion channel opening, ion concentration gradient, and transmembrane potential. The equilibrium state describes the state where the ion concentration tends to stabilize after a long period of 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 in the resting state, and about 145 mM outside the cell. When sodium channels are open, the influx rate of sodium ions can reach 0.5 mM / ms, and this flow rate is regulated by therapeutic intervention.

[0110] The trend of intracellular and extracellular ion concentration changes is calculated using an ion concentration gradient response function. This trend calculation is based on a time integration method, accumulating short-term ion flows into long-term concentration changes. In practice, ion concentration calculations are performed in hourly time steps. For example, after the administration of a neuromodulatory drug, the intracellular sodium ion concentration gradually increases from a baseline of 12 mM to 15 mM within 24 hours, while the potassium ion concentration decreases from a baseline of 140 mM to 135 mM. This concentration change reflects the long-term effect of the drug on ion balance.

[0111] The trend of ion concentration change is modulated by a time constant to simulate the adaptation of biological systems and the attenuation of drug effects. The time constant represents the time required for the concentration change to decay from its peak to 37%, and different ion channels and drugs have different time constants. For example, drugs targeting sodium ion channels have a time constant of 12 hours, while drugs targeting calcium ion channels have a time constant as long as 36 hours. In actual calculations, this modulation process is achieved through an exponential decay function, i.e., the initial effect intensity multiplied by an exponential decay factor based on time and the time constant. Taking a certain neuromodulatory drug as an example, if the initial inhibition of sodium channel conductance is 30% with a time constant of 12 hours, then after 24 hours of use, the inhibitory effect decreases to approximately 11%.

[0112] Based on the above calculation results, a treatment effect mapping for candidate treatment plans is established. This mapping is a transformation relationship that links changes in ion concentration with clinical treatment indicators. Specifically, the effectiveness of the treatment plan is evaluated by establishing a correspondence between key physiological indicators and changes in ion concentration.

[0113] For example, in the treatment of epilepsy, 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 assessment indicator. When the ratio increases from the normal 0.086 to 0.092, it indicates an increase of about 7% in neuronal excitability, suggesting poor drug efficacy; while when the ratio decreases to 0.080, it indicates a decrease of about 8% in neuronal excitability, suggesting drug efficacy.

[0114] 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 at 10 ng / mL, with a half-life of 8 hours. It primarily acts on N-type calcium channels, with an inhibition rate of 45%. Analysis using a biosynaptic kinetic model revealed the following initial response characteristics: a 52% reduction in calcium ion influx and a 30% shortening of action potential duration.

[0115] The trend of ion concentration changes was calculated, showing that intracellular calcium ion concentration decreased from a baseline of 100 nM to 70 nM, maintained for approximately 16 hours. The final treatment effect mapping showed that the firing frequency of neurons related to pain transmission decreased by approximately 40%, and the pain perception threshold increased by approximately 35%, suggesting that this treatment regimen has a good analgesic effect. Compared with the control drug (concentration 5 ng / mL, inhibition rate 25%), the new drug significantly improved both the duration of action and analgesic effect.

[0116] This method is also applicable to the evaluation of other types of neuromodulatory drugs, including antipsychotics, antidepressants, and antiepileptics. Different types of drugs primarily target specific ion channels; by adjusting relevant parameters and indicators, precise evaluation and optimization of various treatment regimens can be achieved.

[0117] A second aspect of the present invention provides a multidisciplinary collaborative diagnosis and treatment system based on medical intelligent agent middleware, comprising: 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.

[0118] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0119] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0120] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A full-specialty collaborative diagnosis and treatment method based on medical agent middleware, characterized in that, The method comprises the following steps: using reinforcement learning to perform entity recognition and relation extraction on medical text, 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; using a personalized treatment decision model to simulate the doctor's diagnosis and treatment decision-making process in the basic medical knowledge graph, converting expert diagnosis and treatment experience into 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; based on the enhanced medical knowledge graph, using the personalized treatment decision model to analyze the symptom association information in the enhanced medical knowledge graph, and establishing a symptom feature vector reflecting the characteristics of the patient's symptoms; inputting the symptom feature vector into the enhanced medical knowledge graph, calculating the matching probability of the patient's symptoms and diseases through the Monte Carlo tree search, and generating a preliminary diagnosis decision based on the matching probability for the patient's current symptoms; according to the disease type in the preliminary diagnosis decision, using reinforcement learning and the Monte Carlo tree search to select the corresponding treatment scheme from the enhanced medical knowledge graph, continuously adjusting the treatment scheme through reinforcement learning to improve its adaptability, and finally generating an optimal treatment scheme.

2. The method of claim 1, wherein, using reinforcement learning to perform entity recognition and relation extraction on medical text, 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 comprises: using reinforcement learning to analyze medical text, constructing a state vector of medical entity information, determining a decision vector for entity extraction based on the state vector, and generating a feature distribution through a dynamic weight adjustment mechanism based on the state vector and the decision vector; mapping the feature distribution to a unified semantic space to generate a neural feature vector, performing bidirectional feature transformation on the neural feature vector to obtain a rule feature vector, and performing weight optimization on the neural feature vector and the rule feature vector based on the dynamic weight adjustment mechanism to generate a fusion feature vector; using Monte Carlo tree search to analyze the fusion feature vector, representing the entity information in the fusion feature vector by constructing a search tree node, calculating a node score based on the dynamic weight adjustment mechanism, the node score is used to guide the expansion direction of the search tree, and determining the association relationship between entities according to the expansion result of the search tree; based on the analysis result of the Monte Carlo tree search, converting the fusion feature vector into node information and edge information of the knowledge graph, and constructing a basic medical knowledge graph according to the node information and the edge information.

3. The method of claim 1, wherein, using a personalized treatment decision model to simulate the doctor's diagnosis and treatment decision-making process in the basic medical knowledge graph, converting expert diagnosis and treatment experience into 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 comprises: Adopting a personalized treatment decision model to model a doctor's diagnosis and treatment decision process in a basic medical knowledge graph, constructing a state vector represented by patient clinical characteristics, and setting an action vector of the doctor's diagnosis and treatment plan based on the state vector; Using reinforcement learning to calculate the combined features of the state vector and the action vector, iteratively optimizing the combined features, and mapping the doctor's diagnosis and treatment experience to a diagnosis and treatment probability distribution based on the personalized treatment decision model; Inputting the diagnosis and treatment probability distribution into a Monte Carlo search tree structure to determine the node expansion direction, calculating the access frequency and cumulative value of the expanded tree nodes based on the node expansion direction, and selecting the optimal search path according to the access frequency and cumulative value; According to the optimal search path, extract the diagnosis and treatment path with credibility from the Monte Carlo search tree structure, dynamically write the probability distribution information of the diagnosis and treatment path into the basic medical knowledge graph, and update the edge weight of the diagnosis and treatment relationship in the basic medical knowledge graph using an adaptive weight method to form an enhanced medical knowledge graph.

4. The method of claim 3, wherein, Inputting the diagnosis and treatment probability distribution into a Monte Carlo search tree structure to determine the node expansion direction, calculating the access frequency and cumulative value of the expanded tree nodes based on the node expansion direction, and selecting the optimal search path according to the access frequency and cumulative value includes: Input the diagnosis and treatment probability distribution into the Monte Carlo search tree structure, construct a Lorentz system containing 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 the evolution trajectory in the three-dimensional dynamic space through iterative calculation, and determine the initial node expansion direction according to the evolution trajectory; Calculate the 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 use the optimized node expansion probability as the selection basis for tree node expansion; Calculate the access frequency of the expanded tree nodes, construct a bifurcation metric function based on node depth and branch number, use the weighted product of the access frequency and the bifurcation metric function as the node weight, and identify the decision nodes in the search path through the node weight; Calculate the cumulative value of the tree nodes, 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 the weight coefficient, and update the cumulative value by weighting, and select the optimal search path according to the access frequency and cumulative value.

5. The method of claim 1, wherein, Calculate the matching probability of patient symptoms and diseases through the Monte Carlo tree search, diagnose and analyze the patient's current symptoms according to the matching probability, and generate a preliminary diagnosis decision, including: 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 by combining 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; 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.

6. The method of claim 1, wherein, 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, and the optimal treatment plan is finally generated, 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.

7. The method of claim 6, wherein, 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.

8. A full-specialty collaborative diagnosis and treatment system based on medical agent middleware, for implementing the method of any one of the preceding claims 1-7, 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.

9. An electronic device, comprising: 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 7.

10. A computer-readable storage medium having stored thereon computer program instructions, wherein, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images