A method and system for evaluating congestion based on multi-modal and reinforcement learning

Abstract

The application discloses a kind of based on multimodal and reinforcement learning's blocking dynamic evaluation method and system, it is related to auxiliary medical technical field.The present application constructs the global fusion feature that accurately reflects the real-time condition of anesthesia block by named entity extraction, vectorization operation and weighted fusion, while building the matching blocking special medical knowledge graph to sort out various diagnosis and treatment entity categories and internal correlation, to this as semantic constraint carries out reinforcement learning iterative training to the transformed Markov decision process Generation exclusive evaluation agent, the whole scheme can be throughout anesthesia block onset, peak maintenance to drug efficacy recession whole cycle carries out uninterrupted intelligent monitoring and evaluation.The present application improves the safety, stability and intelligent control level of anesthesia block diagnosis and treatment, can also accumulate standardization diagnosis and treatment data and perfect knowledge system, provides data support for subsequent individualized anesthesia block optimization.

Description

Technical Field

[0001] This invention relates to the field of assistive medical technology, and more specifically to a method and system for dynamic evaluation of impairment based on multimodal and reinforcement learning. Background Technology

[0002] Assessing anesthetic blockade is a core component of perioperative medicine and pain management, as its effectiveness directly impacts surgical safety, patient comfort, and postoperative recovery quality. Currently, clinical assessment of anesthetic blockade primarily relies on traditional methods such as the anesthesiologist's subjective judgment, vital sign monitoring data, and timed tactile / motor scoring. These methods are limited in scope and have low temporal resolution, making it difficult to provide a continuous and accurate dynamic quantitative description of the entire process from onset and peak effect to deterioration. Furthermore, the blockade effect is influenced by multiple factors, including individual patient differences, medication regimen, administration method, and intraoperative interventions, resulting in a dynamic evolution that exhibits significant nonlinearity and time-varying characteristics. Existing assessment methods struggle to capture subtle changes in the blockade status throughout the entire process, particularly lacking early warning capabilities for abnormal blockades.

[0003] In recent years, with the development of medical IoT and multimodal monitoring technologies, multi-source heterogeneous data such as EEG, ECG, EMG, pulse oximetry, video pupil response, and electronic medical record text can be obtained intraoperatively, providing a rich information foundation for block assessment. Some studies have attempted to use machine learning to fuse and analyze the above data, but they still mainly rely on static feature splicing and fail to fully utilize temporal evolution information; moreover, existing models are mostly general medical models, lacking the structured expression and reasoning ability for anesthesia block-specific knowledge, resulting in assessment results that lack interpretability and adaptability.

[0004] Therefore, there is an urgent need to construct a dynamic evaluation method for the entire process of obstruction that can deeply integrate multimodal heterogeneous data, possess dynamic modeling and adaptive learning capabilities for the entire obstruction process, and combine with specialized knowledge constraints for intelligent reasoning, in order to make up for the shortcomings of existing technologies in terms of data fusion depth, temporal dynamic modeling, and domain knowledge utilization. Summary of the Invention

[0005] In view of this, the present invention provides a method and system for dynamic evaluation of obstruction based on multimodal and reinforcement learning, in order to solve the problems existing in the background art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A dynamic evaluation method for impediment based on multimodal and reinforcement learning includes the following steps: Real-time acquisition of multimodal heterogeneous data of the entire process of anesthesia blockade from onset, peak maintenance to drug effect decline, and splitting the multimodal heterogeneous data into data segments of blocked patients and data segments of diagnosis and treatment; Named entity extraction is performed on the blocked patient data segment and the diagnosis and treatment data segment respectively to obtain the blocked patient-diagnosis and treatment dataset, and vectorization analysis is performed to obtain the blocked patient-diagnosis and treatment embedding vector; the multimodal heterogeneous data and the blocked patient-diagnosis and treatment embedding vector are weighted and fused to obtain the global fusion feature representing the real-time state of blockage. Based on the patient-treatment dataset and patient-treatment embedding vectors, a medical knowledge graph for patients with respiratory arrest is constructed. The graph is used to label the category to which each entity belongs and to store the associations and correlations between entities. The dynamic changes of anesthesia blockade are modeled as a Markov decision process, and the Markov decision process is iteratively trained using the blockade-specific medical knowledge graph as semantic constraints to obtain a dedicated intelligent agent for blockade dynamic evaluation. The real-time global fusion features are input into the trained agent for iterative reasoning. Simultaneously, historical blockage cases similar to the current patient entity are retrieved from the blockage-specific medical knowledge graph. The blockage quantification level and drug efficacy change trend are dynamically updated, and the results of normal blockage status determination or abnormal risk warning are output simultaneously.

[0007] Optionally, a medical knowledge graph specifically for arterial blockade can be constructed. Specifically, this involves using patient entities and anesthesia treatment entities extracted from all historical arterial blockade cases as graph nodes, and the association between patient entities and corresponding treatment entities in the same case as graph edges. The correlation between entities is calculated based on the distance of their embedding vectors and stored as edge weights. The dimensionality of the embedding vectors corresponding to each entity is reduced to generate a visualized medical knowledge graph specifically for arterial blockade.

[0008] Optionally, a dynamic stage discrimination step may also be included: Based on the global fusion features, a pre-trained stage discrimination model is used to identify the dynamic sub-stage to which the current blocking process belongs. The dynamic sub-stage includes at least three categories: the onset period, the peak maintenance period, and the drug effect decline period. The stage discrimination model outputs the stagnation stage label and its confidence score corresponding to the current moment, and encodes the stage label into a structured vector form; The labels representing the stages of stagnation are used as prior knowledge and are injected into the entity relationship reasoning path of the dedicated medical knowledge graph for stagnation, as well as into the state space of the reinforcement learning policy network. The injection into the knowledge graph is used to constrain the temporal logic matching range during graph retrieval, and the injection into the policy network is used to limit the action search space under the corresponding stage, so as to achieve dynamic evaluation with stage-based anchoring.

[0009] Optionally, the dynamic change process of anesthesia blockade is modeled as a Markov decision process. Specifically, an enhanced state space is constructed based on the globally fused features and the embedded set of similar patient entities obtained by knowledge graph retrieval. An action space containing blockade state discrimination and adaptive adjustment of feature weights is constructed based on similar anesthesia diagnosis and treatment entities obtained by knowledge graph matching. A composite reward function is constructed by combining evaluation accuracy, early warning performance and graph semantic matching degree.

[0010] Optionally, the process of constructing the enhanced state space includes: calculating the distance between the current patient embedding vector and the embedding vectors of each historical patient entity in the dedicated medical knowledge graph for blockade, filtering similar patient entities whose distance is less than a set threshold, and splicing the historical blockade time sequence state and drug efficacy change curve corresponding to the similar patient entities with the current global fusion features to form the enhanced state space.

[0011] Optionally, the composite reward function also incorporates an intrinsic curiosity algorithm to calculate the forward prediction error between the current state and the predicted next state, and quantifies the error as an intrinsic exploration reward to encourage the agent to fully explore rare state transition paths in the process of stall decay in a sparse external reward environment.

[0012] Optionally, when multiple similar anesthesia diagnosis and treatment entities are retrieved, the matching probability of each entity is calculated separately, and the entities are sorted from high to low probability as supplementary evidence for block classification and risk warning.

[0013] A dynamic evaluation system for impediments based on multimodal and reinforcement learning, comprising: Multi-source heterogeneous data acquisition module: used to acquire multimodal heterogeneous data of the entire process of anesthesia blockade from onset, peak maintenance to drug effect decline in real time, and to split the multimodal heterogeneous data into data segments of blocked patients and data segments of diagnosis and treatment; Multi-source heterogeneous data processing module: used to extract named entities from the blocked patient data segment and the diagnosis and treatment data segment respectively to obtain blocked patient-diagnosis and treatment dataset, and to perform vectorization analysis to obtain blocked patient-diagnosis and treatment embedding vector; and to perform weighted fusion of multimodal heterogeneous data and blocked patient-diagnosis and treatment embedding vector to obtain global fusion features representing the real-time state of blockage. Knowledge graph construction module: used to construct a medical knowledge graph for arteriosclerosis based on the arteriosclerosis patient-treatment dataset and arteriosclerosis patient-treatment embedding vectors, and to label the category to which each entity belongs in the graph and store the association and correlation between entities; Dynamic evaluation dedicated agent training module: used to model the dynamic change process of anesthesia blockade as a Markov decision process, and iteratively train the Markov decision process with the blockade-specific medical knowledge graph as semantic constraints to obtain a blockade dynamic evaluation dedicated agent. Risk warning output module: It is used to input real-time global fusion features into the trained agent for iterative reasoning, simultaneously retrieve historical block cases similar to the current patient entity in the block-specific medical knowledge graph, dynamically update the block quantification level and drug efficacy change trend, and simultaneously output the block normal state judgment or abnormal risk warning result.

[0014] As can be seen from the above technical solution, compared with the prior art, the present invention provides a dynamic assessment method and system for anesthesia block based on multimodal and reinforcement learning. It constructs a global fusion feature that accurately reflects the real-time status of anesthesia block through named entity extraction, vectorization operation, and weighted fusion. Simultaneously, it builds a supporting medical knowledge graph specifically for block to organize various diagnostic and treatment entity categories and their inherent relationships. Using this as semantic constraints, it conducts reinforcement learning iterative training on the transformed Markov decision process to generate a dedicated assessment agent. The entire solution can conduct uninterrupted intelligent monitoring and assessment throughout the entire cycle of anesthesia block onset, peak maintenance, and drug effect decline. On the one hand, it fully integrates multi-dimensional information such as physiological signs and diagnostic and treatment operations to avoid the one-sidedness of single data monitoring. On the other hand, it uses the knowledge graph to accumulate clinical historical cases and professional medical logic, approximately... This approach focuses on reinforcement learning training to reduce meaningless model exploration and improve the medical rationality and reliability of the agent's decision-making and reasoning. On the other hand, it relies on real-time feature input to dynamically calculate the quantitative level of blockade and the evolution of drug efficacy, and simultaneously corrects the evaluation results by comparing with similar historical cases within the benchmark map. This allows for timely identification of abnormal blockade states and issuance of risk warnings, overcoming the limitations of traditional anesthetic blockade relying on intermittent manual observation and judgment by physicians, reducing evaluation bias caused by differences in human experience, and achieving refined, continuous, and quantitative control of blockade states. It can also identify safety hazards such as excessive blockade depth and abnormal blockade regression in advance, optimize the rhythm of anesthetic drug administration, and improve the safety, stability, and intelligent control level of anesthetic blockade diagnosis and treatment. Furthermore, it can accumulate standardized diagnostic and treatment data to improve the knowledge system and provide data support for subsequent optimization of individualized anesthetic blockade protocols. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the method flow provided by the present invention; Figure 2 This is a schematic diagram of the system structure provided by the present invention. Detailed Implementation

[0017] 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.

[0018] This invention discloses a method for evaluating the dynamic evaluation of impairment based on multimodal and reinforcement learning, such as... Figure 1 As shown, it includes the following steps: Step 1: Real-time acquisition of multimodal heterogeneous data of the entire process of anesthesia blockade from onset, peak maintenance to drug effect decline, and splitting the multimodal heterogeneous data into patient data segment and diagnosis and treatment data segment; wherein, the multimodal heterogeneous data includes: sensor time-series modal data of nerve block ultrasound images, continuous monitoring data of local blood flow, muscle strength and pain sensation, local anesthetic drug infusion data, and vital sign time-series data; text modal data of anesthesia records, operation descriptions and medical history records; and tagged modal data of blockade type, drug dosage and patient basic information.

[0019] Step 2: Named entity extraction is performed on the blocked patient data segment and the treatment data segment respectively to obtain the blocked patient-treatment dataset, and vectorization analysis is performed to obtain the blocked patient-treatment embedding vector; the multimodal heterogeneous data and the blocked patient-treatment embedding vector are weighted and fused to obtain the global fusion feature representing the real-time state of the block; the blocked patient entity set includes at least one of the following: block onset time, muscle strength grade, numbness range, pain score, vital sign fluctuations, underlying diseases, and previous block history; the anesthesia treatment entity set includes at least one of the following: block puncture site, local anesthetic drug parameters, block grading criteria, abnormal risk threshold, remedial treatment plan, and complication management methods.

[0020] Step 3: Construct a medical knowledge graph specifically for patients with respiratory arrest based on the patient-treatment dataset and patient-treatment embedding vectors. Mark the category of each entity in the graph and store the associations and correlations between entities. Step 4: Model the dynamic changes of anesthesia blockade as a Markov decision process, and iteratively train the Markov decision process using the blockade-specific medical knowledge graph as semantic constraints to obtain a dedicated intelligent agent for blockade dynamic evaluation. Step 5: Input the real-time global fusion features into the trained agent for iterative reasoning, and simultaneously search for historical blockage cases similar to the current patient entity in the blockage-specific medical knowledge graph. Dynamically update the blockage quantification level and drug efficacy change trend, and simultaneously output the blockage normal state judgment or abnormal risk warning result.

[0021] Furthermore, a medical knowledge graph specifically for arterial block is constructed. Specifically, this involves using patient entities and anesthesia treatment entities extracted from all historical arterial block cases as graph nodes, and the association between patient entities and corresponding treatment entities in the same case as graph edges. The correlation between entities is calculated based on the distance of their embedding vectors and stored as edge weights. The dimensionality of the embedding vectors corresponding to each entity is reduced to generate a visualized medical knowledge graph specifically for arterial block.

[0022] This embodiment also includes a dynamic stage discrimination step: Based on the global fusion features, a pre-trained stage discrimination model is used to identify the dynamic sub-stage to which the current blocking process belongs. The dynamic sub-stage includes at least three categories: the onset period, the peak maintenance period, and the drug effect decline period. The stage discrimination model outputs the stagnation stage label and its confidence score corresponding to the current moment, and encodes the stage label into a structured vector form; The labels representing the stages of stagnation are used as prior knowledge and are injected into the entity relationship reasoning path of the dedicated medical knowledge graph for stagnation, as well as into the state space of the reinforcement learning policy network. The injection into the knowledge graph is used to constrain the temporal logic matching range during graph retrieval, and the injection into the policy network is used to limit the action search space under the corresponding stage, so as to achieve dynamic evaluation with stage-based anchoring.

[0023] Furthermore, the dynamic change process of anesthesia blockade is modeled as a Markov decision process. Specifically, an enhanced state space is constructed based on the global fusion features and the embedded set of similar patient entities obtained by knowledge graph retrieval. An action space containing blockade state discrimination and adaptive adjustment of feature weights is constructed based on similar anesthesia diagnosis and treatment entities obtained by knowledge graph matching. A composite reward function is constructed by combining evaluation accuracy, early warning performance and graph semantic matching degree.

[0024] Furthermore, the process of constructing the enhanced state space includes: calculating the distance between the current patient embedding vector and the embedding vectors of each historical patient entity in the dedicated medical knowledge graph for drug blockade; filtering out similar patient entities whose distance is less than a set threshold; and concatenating the historical blockade time sequence state and drug efficacy change curve corresponding to the similar patient entities with the current global fusion features to form the enhanced state space.

[0025] Furthermore, the composite reward function includes an evaluation accuracy reward, an early warning performance reward, and a graph semantic matching reward. The assessment accuracy reward is determined based on the error between the quantitative grading results of the blockade and the clinical gold standard. The early warning performance bonus is determined based on the number of missed or false alarms of abnormal blocking events; The graph semantic matching reward is determined based on the degree of matching between the agent's output and highly relevant diagnostic entities within the graph.

[0026] Furthermore, the composite reward function also incorporates an intrinsic curiosity algorithm to calculate the forward prediction error between the current state and the predicted next state, and quantifies the error as an intrinsic exploration reward to encourage the agent to fully explore rare state transition paths in the process of stall decay in a sparse external reward environment.

[0027] In another embodiment, modeling the dynamic changes in anesthesia blockade as a Markov decision process specifically includes: State space construction: The global fusion feature at time t is concatenated with the neighborhood subgraph embedding vector corresponding to the current patient entity in the blockade-specific medical knowledge graph to form a high-dimensional state vector; Action space definition: Define a discrete action set, which includes blockage depth level fine-tuning actions and abnormal warning triggering actions, wherein the blockage depth level fine-tuning actions include at least three categories: "level increase", "level decrease" and "level maintenance". Reward function setting: A delayed sparse reward mechanism is adopted. If the error between the predicted decay time of the blockage and the actual clinical record is less than the preset time threshold, or the recall rate of the agent for abnormal blockage events is higher than the preset performance threshold, a positive reward is given; if there is a missed report or a false report, a negative penalty is given. Training strategy: Iterative training is performed using the Actor-Critic reinforcement learning framework, where the Critic network is used to evaluate the value of the current state-action pair, and the Actor network is used to output the action probability distribution. Semantic constraint mechanism: During the action selection phase, the temporal constraint rules and entity associations in the blockade-specific medical knowledge graph are used as action masks to block actions that do not meet clinical logic, thus limiting the exploration space of the agent.

[0028] Furthermore, when multiple similar anesthesia diagnosis and treatment entities are retrieved, the matching probability of each entity is calculated and sorted from high to low probability as auxiliary evidence for block classification and risk warning.

[0029] In another embodiment, step two is as follows: Named entity extraction is performed on the blocked patient data segment and the anesthesiologist treatment data segment respectively to obtain the blocked patient entity set and the anesthesia treatment entity set; the corresponding entity sets are vectorized using the pre-trained patient vector transformation model and the doctor vector transformation model respectively to obtain the blocked patient embedding vector and the anesthesia treatment embedding vector; the multimodal original monitoring features and the two types of embedding vectors are weighted and fused to obtain the global fusion feature representing the real-time state of the block. Both the patient vector transformation model and the doctor vector transformation model are neural network models, which are pre-trained in the following manner: Multiple historical data related to doctor-patient blockade were collected as training samples. Named entity extraction was performed on each training sample to obtain patient entity samples and doctor entity samples. A patient vector transformation model is obtained by training patient entity samples, and a doctor vector transformation model is obtained by training doctor entity samples.

[0030] and Figure 1 Corresponding to the method shown, this invention also discloses a dynamic evaluation system for hindrance based on multimodal and reinforcement learning for evaluating... Figure 1 The implementation of the method, specifically its structure, is as follows: Figure 2 As shown, it includes: Multi-source heterogeneous data acquisition module: used to acquire multimodal heterogeneous data of the entire process of anesthesia blockade from onset, peak maintenance to drug effect decline in real time, and to split the multimodal heterogeneous data into data segments of blocked patients and data segments of diagnosis and treatment; Multi-source heterogeneous data processing module: used to extract named entities from the blocked patient data segment and the diagnosis and treatment data segment respectively to obtain blocked patient-diagnosis and treatment dataset, and to perform vectorization analysis to obtain blocked patient-diagnosis and treatment embedding vector; and to perform weighted fusion of multimodal heterogeneous data and blocked patient-diagnosis and treatment embedding vector to obtain global fusion features representing the real-time state of blockage. Knowledge graph construction module: used to construct a medical knowledge graph for arteriosclerosis based on the arteriosclerosis patient-treatment dataset and arteriosclerosis patient-treatment embedding vectors, and to label the category to which each entity belongs in the graph and store the association and correlation between entities; Dynamic evaluation dedicated agent training module: used to model the dynamic change process of anesthesia blockade as a Markov decision process, and iteratively train the Markov decision process with the blockade-specific medical knowledge graph as semantic constraints to obtain a blockade dynamic evaluation dedicated agent. Risk warning output module: It is used to input real-time global fusion features into the trained agent for iterative reasoning, simultaneously retrieve historical block cases similar to the current patient entity in the block-specific medical knowledge graph, dynamically update the block quantification level and drug efficacy change trend, and simultaneously output the block normal state judgment or abnormal risk warning result.

[0031] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0032] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A dynamic evaluation method for impediments based on multimodal and reinforcement learning, characterized in that, Includes the following steps: Real-time acquisition of multimodal heterogeneous data of the entire process of anesthesia blockade from onset, peak maintenance to drug effect decline, and splitting the multimodal heterogeneous data into data segments of blocked patients and data segments of diagnosis and treatment; Named entity extraction is performed on the blocked patient data segment and the diagnosis and treatment data segment respectively to obtain the blocked patient-diagnosis and treatment dataset, and vectorization analysis is performed to obtain the blocked patient-diagnosis and treatment embedding vector; the multimodal heterogeneous data and the blocked patient-diagnosis and treatment embedding vector are weighted and fused to obtain the global fusion feature representing the real-time state of blockage. Based on the patient-treatment dataset and patient-treatment embedding vectors, a medical knowledge graph for patients with respiratory arrest is constructed. The graph is used to label the category to which each entity belongs and to store the associations and correlations between entities. The dynamic changes of anesthesia blockade are modeled as a Markov decision process, and the Markov decision process is iteratively trained using the blockade-specific medical knowledge graph as semantic constraints to obtain a dedicated intelligent agent for blockade dynamic evaluation. The real-time global fusion features are input into the trained agent for iterative reasoning. Simultaneously, historical blockage cases similar to the current patient entity are retrieved from the blockage-specific medical knowledge graph. The blockage quantification level and drug efficacy change trend are dynamically updated, and the results of normal blockage status determination or abnormal risk warning are output simultaneously.

2. The method for dynamic evaluation of hindrance based on multimodal and reinforcement learning according to claim 1, characterized in that, Constructing a medical knowledge graph specifically for arterial blockade involves using patient entities and anesthesia treatment entities extracted from all historical arterial blockade cases as graph nodes, and the association between patient entities and corresponding treatment entities in the same case as graph edges. The correlation between entities is calculated based on the distance of their embedding vectors and stored as edge weights. The dimensionality of the embedding vectors corresponding to each entity is reduced to generate a visualized medical knowledge graph specifically for arterial blockade.

3. The method for dynamic evaluation of hindrance based on multimodal and reinforcement learning according to claim 1, characterized in that, It also includes a dynamic stage discrimination step: Based on the global fusion features, a pre-trained stage discrimination model is used to identify the dynamic sub-stage to which the current blocking process belongs. The dynamic sub-stage includes at least three categories: the onset period, the peak maintenance period, and the drug effect decline period. The stage discrimination model outputs the stagnation stage label and its confidence score corresponding to the current moment, and encodes the stage label into a structured vector form; The labels representing the stages of stagnation are used as prior knowledge and are injected into the entity relationship reasoning path of the dedicated medical knowledge graph for stagnation, as well as into the state space of the reinforcement learning policy network. The injection into the knowledge graph is used to constrain the temporal logic matching range during graph retrieval, and the injection into the policy network is used to limit the action search space under the corresponding stage, so as to achieve dynamic evaluation with stage-based anchoring.

4. The method for dynamic evaluation of hindrance based on multimodal and reinforcement learning according to claim 1, characterized in that, The dynamic process of anesthesia blockade is modeled as a Markov decision process. Specifically, an enhanced state space is constructed based on the global fusion features and the embedded set of similar patient entities obtained by knowledge graph retrieval. An action space containing blockade state discrimination and adaptive adjustment of feature weights is constructed based on similar anesthesia diagnosis and treatment entities obtained by knowledge graph matching. A composite reward function is constructed by combining evaluation accuracy, early warning performance and graph semantic matching degree.

5. The method for dynamic evaluation of hindrance based on multimodal and reinforcement learning according to claim 1, characterized in that, The process of constructing the enhanced state space includes: calculating the distance between the current patient embedding vector and the embedding vectors of each historical patient entity in the dedicated medical knowledge graph for blockade; filtering out similar patient entities whose distance is less than a set threshold; and concatenating the historical blockade time sequence state and drug efficacy change curve corresponding to the similar patient entities with the current global fusion features to form the enhanced state space.

6. The method for dynamic evaluation of hindrance based on multimodal and reinforcement learning according to claim 4, characterized in that, The composite reward function also incorporates an intrinsic curiosity algorithm to calculate the forward prediction error between the current state and the predicted next state, and quantifies the error as an intrinsic exploration reward to encourage the agent to fully explore rare state transition paths in the process of stall decay in a sparse external reward environment.

7. The method for dynamic evaluation of hindrance based on multimodal and reinforcement learning according to claim 1, characterized in that, When multiple similar anesthesia diagnosis and treatment entities are retrieved, the matching probability of each entity is calculated and sorted from high to low probability as auxiliary evidence for block classification and risk warning.

8. A dynamic evaluation system for impediments based on multimodal and reinforcement learning, characterized in that, include: Multi-source heterogeneous data acquisition module: used to acquire multimodal heterogeneous data of the entire process of anesthesia blockade from onset, peak maintenance to drug effect decline in real time, and to split the multimodal heterogeneous data into data segments of blocked patients and data segments of diagnosis and treatment; Multi-source heterogeneous data processing module: used to extract named entities from the blocked patient data segment and the diagnosis and treatment data segment respectively to obtain blocked patient-diagnosis and treatment dataset, and to perform vectorization analysis to obtain blocked patient-diagnosis and treatment embedding vector; and to perform weighted fusion of multimodal heterogeneous data and blocked patient-diagnosis and treatment embedding vector to obtain global fusion features representing the real-time state of blockage. Knowledge graph construction module: used to construct a medical knowledge graph for arteriosclerosis based on the arteriosclerosis patient-treatment dataset and arteriosclerosis patient-treatment embedding vectors, and to label the category to which each entity belongs in the graph and store the association and correlation between entities; Dynamic evaluation dedicated agent training module: used to model the dynamic change process of anesthesia blockade as a Markov decision process, and iteratively train the Markov decision process with the blockade-specific medical knowledge graph as semantic constraints to obtain a blockade dynamic evaluation dedicated agent. Risk warning output module: It is used to input real-time global fusion features into the trained agent for iterative reasoning, simultaneously retrieve historical block cases similar to the current patient entity in the block-specific medical knowledge graph, dynamically update the block quantification level and drug efficacy change trend, and simultaneously output the block normal state judgment or abnormal risk warning result.

Images