Intelligent follow-up method and system for chronic atrophic gastritis based on multi-source electronic medical records

Abstract

The present invention provides a method and system for intelligent follow-up of chronic atrophic gastritis based on multi-source electronic medical records, which relates to the technical field of intelligent disease follow-up, including preprocessing and feature extraction of structured, semi-structured and unstructured electronic medical record data, mining the inter-layer feature association relationship by a meta-path reasoning module, and generating comprehensive graph features of patients; identifying the causal relationship between features by a Bayesian causal network module, and building a disease risk prediction model in combination with a temporal causal graph module; and optimizing the follow-up plan by a Nash equilibrium algorithm based on risk prediction results and medical resource constraints; the present invention can effectively predict the disease risk of chronic atrophic gastritis, and optimize the follow-up plan in combination with medical resources, thereby improving the follow-up efficiency and facilitating the early diagnosis and treatment of chronic atrophic gastritis.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent disease follow-up, and in particular to an intelligent follow-up method and system for chronic atrophic gastritis based on multi-source electronic medical records. Background Art

[0002] Chronic atrophic gastritis (CAG) is a common digestive system disease with a prolonged course, poor prognosis, and may even develop into gastric cancer. Therefore, long-term and standardized follow-up of CAG patients is crucial to detect changes in the condition in a timely manner, take intervention measures, and reduce the risk of disease progression.

[0003] Traditional CAG follow-up mainly relies on the doctor's experience and limited clinical indicators, and has some defects and shortcomings: traditional follow-up plans are usually based on pre-set time intervals, such as an endoscopy every six months or one year. This method lacks personalization and cannot adjust the follow-up frequency according to the patient's specific condition and risk level, resulting in some patients being over-followed up, while other high-risk patients may miss the best time for intervention. Traditional follow-up focuses on a few clinical indicators, such as gastroscopy results and serological indicators, while ignoring other important information about the patient, such as medical history, symptoms, medication, etc. This makes the assessment of the patient's condition not comprehensive enough and it is difficult to accurately predict the risk of disease progression. With the popularization of electronic medical records, medical institutions have accumulated a large amount of patient data, but most of this data is stored in an unstructured or semi-structured form and is difficult to use directly for clinical decision-making. Traditional follow-up methods cannot effectively utilize these valuable data resources, limiting the improvement of follow-up efficiency and accuracy. The present invention can solve the problems in the prior art. Summary of the invention

[0004] The embodiments of the present invention provide a method and system for intelligent follow-up of chronic atrophic gastritis based on multi-source electronic medical records, which can solve the problems in the prior art.

[0005] According to a first aspect of the embodiments of the present invention,

[0006] Provide an intelligent follow-up method for chronic atrophic gastritis based on multi-source electronic medical records, including:

[0007] Divide the multi-source electronic medical record data into a structured data layer, a semi-structured data layer and an unstructured data layer according to the data structure, pre-process the clinical monitoring data in the structured data layer using an outlier detection algorithm and a time series filling algorithm to obtain structured features, use a medical word embedding model to perform standardized mapping on the diagnosis and treatment instruction data in the semi-structured data layer to obtain semi-structured features, use a medical entity recognition model to extract key information from the medical process record data in the unstructured data layer to obtain unstructured features, input the structured features, the semi-structured features and the unstructured features into a meta-path reasoning module, mine the inter-layer feature association relationship, and generate a comprehensive patient graph feature;

[0008] Input the comprehensive graph features of the patient into the Bayesian causal network module, identify the causal relationship between the features through the structural learning algorithm, set the causal path weight, use the expectation maximization algorithm to solve the latent variable distribution, input the causal relationship and the latent variable distribution into the temporal causal graph module, divide the progression of chronic atrophic gastritis into an early stage, a middle stage and a late stage, construct a temporal causal link, and calculate the causal effect strength, construct a disease risk prediction model according to the causal effect strength, and generate a disease risk prediction result of chronic atrophic gastritis in patients;

[0009] Based on the patient's disease risk prediction results, a risk level probability distribution matrix was constructed. According to the risk level probability distribution matrix, the follow-up priority at different time points was calculated. The time series planning model was used to optimize the follow-up priority to obtain the initial follow-up plan. The medical resource constraint analysis was performed on the initial follow-up plan and the resource competition cost was calculated. The Nash equilibrium algorithm was used to perform multi-objective optimization of the follow-up priority and resource competition cost to generate an optimized follow-up plan.

[0010] In an optional embodiment,

[0011] The multi-source electronic medical record data is divided into a structured data layer, a semi-structured data layer and an unstructured data layer according to the data structure. The clinical monitoring data in the structured data layer is preprocessed by using an outlier detection algorithm and a time series filling algorithm to obtain structured features. The diagnosis and treatment instruction data in the semi-structured data layer is standardized and mapped by using a medical word embedding model to obtain semi-structured features. The medical process record data in the unstructured data layer is extracted by using a medical entity recognition model to obtain unstructured features. The structured features, the semi-structured features and the unstructured features are input into a meta-path reasoning module to mine the inter-layer feature association relationship to generate a comprehensive patient graph feature, including:

[0012] For the clinical monitoring data at each time point in the structured data layer, the time series correlation weight between the clinical monitoring data at adjacent time points is calculated, the time series correlation weight is integrated into the local reachable density calculation to obtain a local outlier factor, and outliers are detected based on the local outlier factor; the outliers are filled using an adaptive Gaussian process regression algorithm with dynamically selected kernel function parameters to obtain structured features;

[0013] For the diagnosis and treatment instruction data in the semi-structured data layer, a domain knowledge tree including a hierarchical relationship of medical concepts is constructed, the hierarchical relationship of medical concepts is encoded into a word vector, and the semantic similarity and semantic normativity of the word vector are simultaneously optimized through a multi-task learning framework, and the weight of the word vector is dynamically adjusted according to the context through an attention mechanism to obtain a semi-structured feature;

[0014] For the medical process record data in the unstructured data layer, a bidirectional encoder is constructed to learn entity representation from the text dimension and the knowledge graph dimension respectively, and the medical ontology knowledge is used as the entity type constraint to guide the entity boundary recognition, extract the basic information in the medical process record data, and obtain the unstructured features;

[0015] A heterogeneous feature graph is constructed according to the semantic association between the structured features, the semi-structured features and the unstructured features, a meta-path is adaptively generated based on the heterogeneous feature graph, and a multi-level attention aggregation network is used to perform feature fusion to obtain comprehensive graph features of the patient.

[0016] In an optional embodiment,

[0017] A heterogeneous feature graph is constructed according to the semantic association between the structured feature, the semi-structured feature and the unstructured feature. A meta-path is adaptively generated based on the heterogeneous feature graph, and a multi-level attention aggregation network is used to perform feature fusion to obtain the patient comprehensive graph features including:

[0018] Receive structured features, semi-structured features and unstructured features as input features, wherein the input features form feature pairs in pairs, calculate the cosine similarity of the feature pairs, and obtain the semantic relevance of the feature pairs; based on a preset semantic relevance threshold, screen the semantic relevance, extract feature pairs greater than the semantic relevance threshold, establish connection relationships, and construct a heterogeneous feature graph based on the connection relationships;

[0019] In the heterogeneous feature graph, a graph node corresponding to the structured feature is selected as a starting node, and the heterogeneous feature graph is traversed through a depth-first search to obtain an access sequence of graph nodes; a pattern is extracted from the access sequence, and access sequences with the same node type transfer pattern are extracted as candidate meta-paths; the cumulative value of the semantic association between adjacent graph nodes on each candidate meta-path is calculated, and the candidate meta-path with the largest cumulative value is selected as the target meta-path;

[0020] Construct a multi-level attention aggregation network to perform aggregation operations and obtain initial fusion features;

[0021] The initial fusion features are divided into anchor features and contrast features, and data enhancement processing is performed on the anchor features and contrast features respectively to obtain enhanced anchor features and enhanced contrast features; the mutual information value between the enhanced anchor features and the enhanced contrast features is calculated, and based on the maximization criterion of the mutual information value, the feature representation parameters are updated to obtain the final patient comprehensive atlas features.

[0022] In an optional embodiment,

[0023] Construct a multi-level attention aggregation network for aggregation operation, and obtain the initial fusion features including:

[0024] The multi-level attention aggregation network includes a node-level attention aggregator, a path-level attention aggregator, and a semantic-level attention aggregator, which perform aggregation operations in sequence, including:

[0025] Calculating node attention scores between adjacent graph nodes on the target meta-path through the node-level attention aggregator, taking the node attention scores as the first weight, performing weighted aggregation on features of adjacent graph nodes, and obtaining node-level fusion features;

[0026] Calculating the path importance scores of different target meta-paths through the path-level attention aggregator, taking the path importance scores as the second weight, performing weighted aggregation on the node-level fusion features on each target meta-path to obtain the path-level fusion features;

[0027] The semantic importance scores of different types of features are calculated through the semantic-level attention aggregator, and the semantic importance scores are used as the third weight to perform weighted aggregation on the path-level fusion features of each type to obtain the initial fusion features.

[0028] In an optional embodiment,

[0029] The patient's comprehensive graph features are input into the Bayesian causal network module, the causal relationship between the features is identified through the structural learning algorithm, and the causal path weight is set. The expectation maximization algorithm is used to solve the latent variable distribution, and the causal relationship and the latent variable distribution are input into the temporal causal graph module. The progression of chronic atrophic gastritis is divided into an early stage, a middle stage, and a late stage, a temporal causal link is constructed, and the causal effect strength is calculated. A disease risk prediction model is constructed according to the causal effect strength, and the disease risk prediction results of chronic atrophic gastritis in patients are generated, including:

[0030] Input the patient's comprehensive graph features into the Bayesian causal network, construct an optimization objective function of the structural learning algorithm, and the optimization objective function includes a structural fit scoring item and a structural complexity constraint item; based on the optimization objective function, the Bayesian causal network is iteratively searched to obtain an optimal network structure, and the weights of the causal paths in the optimal network structure are calculated; the expectation maximization algorithm is used to perform parameter learning on the optimal network structure, wherein the posterior probability distribution of the latent variables is calculated in each round of iteration, and the conditional probability parameters are updated until convergence, so as to obtain the causal relationship between the patient's comprehensive graph features and the latent variable distribution;

[0031] The causal relationship and the latent variable distribution are input into a temporal causal graph module, and the progression of chronic atrophic gastritis is divided into an early stage, a middle stage, and a late stage; the causal relationship is replicated within each stage to construct a local causal structure; the temporal correlation of nodes between adjacent stages is calculated, and a temporal causal link is established between nodes whose temporal correlation is greater than a preset correlation threshold; and the causal effect strength of the temporal causal link is obtained by causal intervention calculation;

[0032] A disease risk prediction model is constructed according to the causal effect strength, and the causal effect strength is used as a feature importance index to screen key features. A hierarchical predictor is constructed for the key features, and the hierarchical predictor includes an early to mid-term predictor and a mid-term to late-term predictor. Each hierarchical predictor integrates the local causal structure and temporal causal link information of the corresponding stage to generate a disease risk prediction result for patients with chronic atrophic gastritis.

[0033] In an optional embodiment,

[0034] Based on the patient's disease risk prediction results, a risk level probability distribution matrix is ​​constructed. According to the risk level probability distribution matrix, the follow-up priority at different time points is calculated. The time series planning model is used to optimize the follow-up priority to obtain the initial follow-up plan, which includes:

[0035] Receive the patient identification, risk level identification and disease risk prediction probability value in the patient's disease risk prediction results; receive the patient identification, follow-up time point and disease status information in the patient's historical follow-up data;

[0036] Using the patient identifier as a row identifier and the risk level identifier as a column identifier, constructing a risk level probability distribution matrix, filling the disease risk prediction probability value into the corresponding position in the risk level probability distribution matrix, and normalizing the disease risk prediction probability value of each row in the risk level probability distribution matrix;

[0037] Assigning an increasing risk weight coefficient to each of the risk level identifiers;

[0038] Using the disease status information as the row identifier and column identifier of the Markov state transfer matrix, calculating the state transfer probability value between two adjacent follow-up time points and filling them into the corresponding positions, and normalizing the state transfer probability value of each row in the Markov state transfer matrix;

[0039] According to a preset follow-up time interval, the Markov state transfer matrix is ​​subjected to a power operation to obtain a state transfer probability matrix; the transfer probability values ​​in the direction of disease aggravation are extracted from the state transfer probability matrix and summed to obtain a disease aggravation probability value;

[0040] Selecting a maximum disease worsening probability value from a preset time interval set, and dividing the disease worsening probability value by the maximum disease worsening probability value to obtain a dynamic time attenuation factor;

[0041] The disease risk prediction probability value corresponding to each patient identifier in the risk level probability distribution matrix is ​​multiplied by the corresponding risk weight coefficient and the sum is calculated to obtain the patient basic follow-up priority corresponding to each patient identifier;

[0042] Multiplying the basic follow-up priority of the patient by the dynamic time decay factor to obtain the comprehensive follow-up priority of the patient;

[0043] The inverse of the comprehensive follow-up priority of the patient is set as the cost coefficient, the maximum number of follow-up persons at each time point is set as the medical resource capacity constraint value, the minimum follow-up interval is set as the time interval constraint value, a follow-up schedule is generated for each patient identifier, and the follow-up schedule result is output;

[0044] Calculate the average waiting time in the follow-up schedule result, use the negative value corresponding to the average waiting time as a reward value, pass it into the reinforcement learning model, and update the parameter value in the Markov state transfer matrix;

[0045] The patient identifier in the follow-up schedule result is matched with the follow-up time point to generate an initial follow-up plan.

[0046] In an optional embodiment,

[0047] The initial follow-up plan was analyzed for medical resource constraints and the resource competition cost was calculated. The Nash equilibrium algorithm was used to perform multi-objective optimization on the follow-up priority and resource competition cost. The optimized follow-up plan was generated, including:

[0048] Receive the patient identification, follow-up time points, and comprehensive follow-up priority of the patient in the initial follow-up plan;

[0049] The number of follow-up patients at each follow-up time point is divided by the upper limit of the medical resource capacity to obtain the resource occupancy rate; the resource occupancy rate is subtracted from the preset resource occupancy threshold, and then squared and multiplied by the competition cost coefficient to obtain the resource competition cost at the time point; the resource competition costs at all the follow-up time points are accumulated to obtain the overall resource competition cost;

[0050] The patient identifier is set as a game participant, and the set of follow-up time points is set as a strategy space; the product of the comprehensive follow-up priority of the patient and the first weight coefficient is calculated to determine the first objective function; the product of the overall resource competition cost and the second weight coefficient is calculated to determine the second objective function; an initial follow-up time point is assigned to each of the patient identifiers; the patient identifier to be optimized is selected, and the follow-up time points of other patient identifiers are kept unchanged; the calculation result of the first objective function is subtracted from the calculation result of the second objective function to obtain the utility function value of the patient identifier to be optimized;

[0051] Based on the maximum utility function value, a corresponding follow-up time point is selected from the strategy space and updated to the optimal follow-up time point of the patient identification to be optimized;

[0052] Repeat the iteration for all patient identifiers. When the follow-up time points of all patient identifiers no longer change, the current follow-up time point is taken as the final follow-up time point.

[0053] Based on the patient identification and the final follow-up time point, the corresponding relationship is determined and the optimized follow-up plan is output.

[0054] According to a second aspect of the embodiments of the present invention,

[0055] Provided is an intelligent follow-up system for chronic atrophic gastritis based on multi-source electronic medical records, including:

[0056] The first unit is used to divide the multi-source electronic medical record data into a structured data layer, a semi-structured data layer and an unstructured data layer according to the data structure, pre-process the clinical monitoring data in the structured data layer by using an outlier detection algorithm and a time series filling algorithm to obtain structured features, standardize and map the diagnosis and treatment instruction data in the semi-structured data layer by using a medical word embedding model to obtain semi-structured features, extract key information from the medical process record data in the unstructured data layer by using a medical entity recognition model to obtain unstructured features, input the structured features, the semi-structured features and the unstructured features into a meta-path reasoning module, mine the inter-layer feature association relationship, and generate a comprehensive patient graph feature;

[0057] The second unit is used to input the comprehensive graph features of the patient into the Bayesian causal network module, identify the causal relationship between the features through the structural learning algorithm, set the causal path weight, use the expectation maximization algorithm to solve the latent variable distribution, input the causal relationship and the latent variable distribution into the temporal causal graph module, divide the progression of chronic atrophic gastritis into an early stage, a middle stage and a late stage, construct a temporal causal link, and calculate the causal effect strength, construct a disease risk prediction model according to the causal effect strength, and generate a disease risk prediction result of chronic atrophic gastritis in patients;

[0058] The third unit is used to construct a risk level probability distribution matrix based on the patient's disease risk prediction results, calculate the follow-up priority at different time points according to the risk level probability distribution matrix, and use the timing planning model to perform time series optimization on the follow-up priority to obtain the initial follow-up plan; perform medical resource constraint analysis on the initial follow-up plan and calculate the resource competition cost, use the Nash equilibrium algorithm to perform multi-objective optimization on the follow-up priority and resource competition cost, and generate an optimized follow-up plan.

[0059] According to a third aspect of the embodiments of the present invention,

[0060] An electronic device is provided, comprising:

[0061] processor;

[0062] a memory for storing processor-executable instructions;

[0063] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0064] A fourth aspect of the embodiments of the present invention is:

[0065] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0066] In an embodiment of the present invention, the accuracy of follow-up of chronic atrophic gastritis is improved: by integrating multi-source electronic medical record data (structured, semi-structured and unstructured data), and using meta-path reasoning, Bayesian causal networks and temporal causal graphs and other technologies, the patient's condition can be more comprehensively understood, the causal relationship between key features can be identified, and the disease risk can be more accurately predicted, thereby achieving personalized and accurate follow-up; based on the disease risk prediction results and the risk level probability distribution matrix, combined with medical resource constraints and Nash equilibrium algorithms, the follow-up plan can be optimized, medical resources can be reasonably allocated, the follow-up efficiency can be improved, and resource waste can be avoided; the intelligent follow-up plan can assist doctors in making decisions, timely identify high-risk patients, and conduct early intervention, thereby improving the management level of chronic atrophic gastritis and delaying disease progression. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a flow chart of an intelligent follow-up method for chronic atrophic gastritis based on multi-source electronic medical records according to an embodiment of the present invention;

[0068] Figure 2 It is a structural schematic diagram of an intelligent follow-up system for chronic atrophic gastritis based on multi-source electronic medical records according to an embodiment of the present invention. DETAILED DESCRIPTION

[0069] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0070] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0071] Figure 1 FIG. 4 is a flow chart of an intelligent follow-up method for chronic atrophic gastritis based on multi-source electronic medical records according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0072] S101. Divide the multi-source electronic medical record data into a structured data layer, a semi-structured data layer and an unstructured data layer according to the data structure, pre-process the clinical monitoring data in the structured data layer using an outlier detection algorithm and a time series filling algorithm to obtain structured features, perform standardized mapping on the diagnosis and treatment instruction data in the semi-structured data layer using a medical word embedding model to obtain semi-structured features, extract key information from the medical process record data in the unstructured data layer using a medical entity recognition model to obtain unstructured features, input the structured features, the semi-structured features and the unstructured features into a meta-path reasoning module, mine the inter-layer feature association relationship, and generate a comprehensive patient graph feature;

[0073] In this embodiment, the electronic medical record data is divided into three levels: structured, semi-structured, and unstructured, which helps to select appropriate processing methods for different data types, thereby improving data processing efficiency and accuracy; structured data (such as clinical monitoring data) are processed through outlier detection and time series filling algorithms, which can clean data and fill missing information, thereby improving the reliability of structured features; a medical word embedding model is used to standardize and map the diagnosis and treatment instruction data, so that semi-structured data can be converted into standardized features for further analysis, thereby improving its effectiveness in the model; key information is extracted from unstructured data (such as medical process records) through a medical entity recognition model, and text data is effectively converted into structured features, thereby providing useful semantic information for subsequent analysis; the meta-path reasoning module integrates the features of different data layers and mines the correlation between layers, which can comprehensively construct the comprehensive map features of patients and improve the depth and accuracy of overall data analysis.

[0074] S102. Input the comprehensive graph features of the patient into the Bayesian causal network module, identify the causal relationship between the features through the structural learning algorithm, set the causal path weight, use the expectation maximization algorithm to solve the latent variable distribution, input the causal relationship and the latent variable distribution into the temporal causal graph module, divide the progression of chronic atrophic gastritis into an early stage, a middle stage and a late stage, construct a temporal causal link, and calculate the causal effect strength, construct a disease risk prediction model according to the causal effect strength, and generate a disease risk prediction result of chronic atrophic gastritis in the patient;

[0075] In this embodiment, through the Bayesian causal network module and the structural learning algorithm, the causal relationship between the comprehensive graph features of the patient can be effectively identified, the potential influence path between different variables can be revealed, and a clear causal structure can be provided for subsequent disease prediction; the expectation maximization algorithm is used to solve the latent variable distribution, which can handle potential factors in the data that are not easily directly observed, thereby enhancing the predictive ability of the causal network and the robustness of the model; the progression of chronic atrophic gastritis is modeled in stages through the time series causal graph module, capturing the causal effects of the disease over time, and calculating the causal effect strength at different stages, thereby improving the model's adaptability to time series data; a disease risk prediction model is constructed based on the causal effect strength, which can quantify the risk of chronic atrophic gastritis and provide accurate prediction results based on patient characteristics and disease stages, which is helpful for early intervention and personalized treatment.

[0076] S103. Based on the patient's disease risk prediction results, a risk level probability distribution matrix is ​​constructed. According to the risk level probability distribution matrix, the follow-up priority at different time points is calculated. The time series planning model is used to optimize the follow-up priority to obtain the initial follow-up plan. The initial follow-up plan is subjected to medical resource constraint analysis and the resource competition cost is calculated. The Nash equilibrium algorithm is used to perform multi-objective optimization on the follow-up priority and the resource competition cost to generate an optimized follow-up plan.

[0077] In this embodiment, based on the patient's disease risk prediction results, a risk level probability distribution matrix is ​​constructed, which can accurately assess the patient's risk level, thereby providing a scientific basis for the subsequent follow-up priority setting; the time series optimization of the follow-up priority through the timing planning model can ensure that the follow-up arrangement is more reasonable at different time points, give priority to high-risk patients, and improve the accuracy and effectiveness of the follow-up strategy; after generating the initial follow-up plan, a medical resource constraint analysis is performed and the resource competition cost is calculated to ensure that the follow-up plan can be reasonably allocated and arranged under limited resources to avoid resource waste; the Nash equilibrium algorithm is used to perform multi-objective optimization of the follow-up priority and the resource competition cost, which can balance the patient's risk and the use of medical resources, generate the optimal follow-up plan, and achieve efficient use of medical resources; the optimized follow-up plan generated by comprehensive risk prediction, time series optimization and resource allocation can improve the efficiency of medical services and the accuracy of patient management, and improve the overall effect of chronic disease management.

[0078] In an optional implementation, the multi-source electronic medical record data is divided into a structured data layer, a semi-structured data layer and an unstructured data layer according to the data structure, the clinical monitoring data in the structured data layer is preprocessed using an outlier detection algorithm and a time series filling algorithm to obtain structured features, the diagnosis and treatment instruction data in the semi-structured data layer is standardized and mapped using a medical word embedding model to obtain semi-structured features, the medical process record data in the unstructured data layer is extracted using a medical entity recognition model to obtain unstructured features, the structured features, the semi-structured features and the unstructured features are input into a meta-path reasoning module, the inter-layer feature association relationship is mined, and the patient comprehensive graph features are generated, including:

[0079] For the clinical monitoring data at each time point in the structured data layer, the time series correlation weight between the clinical monitoring data at adjacent time points is calculated, the time series correlation weight is integrated into the local reachable density calculation to obtain a local outlier factor, and outliers are detected based on the local outlier factor; the outliers are filled using an adaptive Gaussian process regression algorithm with dynamically selected kernel function parameters to obtain structured features;

[0080] For the diagnosis and treatment instruction data in the semi-structured data layer, a domain knowledge tree including a hierarchical relationship of medical concepts is constructed, the hierarchical relationship of medical concepts is encoded into a word vector, and the semantic similarity and semantic normativity of the word vector are simultaneously optimized through a multi-task learning framework, and the weight of the word vector is dynamically adjusted according to the context through an attention mechanism to obtain a semi-structured feature;

[0081] For the medical process record data in the unstructured data layer, a bidirectional encoder is constructed to learn entity representation from the text dimension and the knowledge graph dimension respectively, and the medical ontology knowledge is used as the entity type constraint to guide the entity boundary recognition, extract the basic information in the medical process record data, and obtain the unstructured features;

[0082] A heterogeneous feature graph is constructed according to the semantic association between the structured features, the semi-structured features and the unstructured features, a meta-path is adaptively generated based on the heterogeneous feature graph, and a multi-level attention aggregation network is used to perform feature fusion to obtain comprehensive graph features of the patient.

[0083] In a specific implementation, first, the clinical monitoring data in the structured data layer is preprocessed. For example, the patient's body temperature, blood pressure, heart rate and other physiological index data. Taking body temperature data as an example, assume that a patient's five-day body temperature records are 36.5°C, 36.8°C, 39.5°C, 36.6°C and 36.7°C respectively. The temporal correlation weight is calculated by comparing the difference between the body temperatures of two adjacent days. For example, the body temperature difference between the first and second days is 0.3°C, and the weight can be set to 0.9; the body temperature difference between the second and third days is 2.7°C, and the weight can be set to 0.1, and so on. Then, according to the body temperature value at each time point and its temporal correlation weight with the adjacent time points, the local outlier factor of the time point is calculated. For example, the body temperature value of 39.5°C on the third day is significantly higher than the body temperature values ​​of other days, and the temporal correlation weights with the body temperatures of the adjacent two days are low, so its local outlier factor will be high and it will be identified as an outlier. For the detected outliers, an adaptive Gaussian process regression algorithm with dynamically selected kernel function parameters is used for filling. For example, if the body temperature of 39.5℃ on the third day is identified as an outlier, the algorithm can be used to predict the body temperature value of the third day based on the body temperature data of the previous two days. For example, the predicted value is 37.1℃, and the outlier value 39.5℃ is replaced by 37.1℃. After outlier detection and filling, structured features are obtained.

[0084] Secondly, the diagnosis and treatment instruction data in the semi-structured data layer is standardized and mapped. For example, data such as medical orders and examination reports. Taking the medical order of "500ml of intravenous saline drip" as an example, first construct a domain knowledge tree containing the hierarchical relationship of medical concepts. For example, "normal saline" is a subclass of "drugs" and "intravenous drip" is a subclass of "route of administration". Then, the hierarchical relationship of these medical concepts is encoded into word vectors. For example, the word vector of "normal saline" contains the information of its parent class "drugs". The semantic similarity and semantic normativeness of word vectors are optimized simultaneously through the multi-task learning framework. For example, the word vectors of "infusion" and "intravenous drip" are semantically similar, and the word vectors of "sodium chloride injection" and "normal saline" are semantically normative. Finally, through the attention mechanism, the weight of the word vector is dynamically adjusted according to the context. For example, in the medical order of "500ml of intravenous saline drip", the word vector weights of "normal saline" and "500ml" are higher, while the word vector weight of "intravenous drip" is relatively low. Finally, semi-structured features are obtained.

[0085] Secondly, extract key information from the medical process record data in the unstructured data layer. For example, text data such as admission records, discharge summaries, and surgical records. Taking "The patient complained of coughing and sputum for three days" as an example, first construct a bidirectional encoder to learn entity representations from the text dimension and the knowledge graph dimension respectively. For example, the entity representations of "cough" and "sputum" contain their information in the medical knowledge graph, such as disease type, symptoms, etc. Then, use medical ontology knowledge as entity type constraints to guide entity boundary identification. For example, based on medical ontology knowledge, it can be identified that "cough" and "sputum" are symptom entities. Finally, key information such as "cough" and "sputum" are extracted to obtain unstructured features.

[0086] Finally, the structured features, semi-structured features, and unstructured features are input into the meta-path reasoning module. For example, a meta-path "patient-temperature-time-doctor's order-drug" can be defined to represent the relationship between the patient's temperature at a certain point in time and the drugs they receive. A heterogeneous feature graph is constructed based on the semantic association between the features, and a meta-path is adaptively generated based on the graph. For example, if there is a strong correlation between the patient's temperature and the antipyretic drugs they take, the corresponding meta-path will be generated. A multi-level attention aggregation network is used for feature fusion to obtain the patient's comprehensive graph features.

[0087] In this embodiment, multi-source heterogeneous electronic medical record data can be effectively integrated, and the correlation information between different data layers can be fully mined, thereby improving data utilization; a variety of advanced natural language processing and machine learning technologies can be used to effectively extract and fuse multi-level features, thereby improving feature expression capabilities; based on meta-path reasoning, the correlation relationship between different features can be clearly displayed, thereby enhancing the interpretability of the model.

[0088] In an optional embodiment, a heterogeneous feature graph is constructed according to the semantic association between the structured feature, the semi-structured feature and the unstructured feature, a meta-path is adaptively generated based on the heterogeneous feature graph, and a multi-level attention aggregation network is used to perform feature fusion, so that the patient comprehensive graph features include:

[0089] Receive structured features, semi-structured features and unstructured features as input features, wherein the input features form feature pairs in pairs, calculate the cosine similarity of the feature pairs, and obtain the semantic relevance of the feature pairs; based on a preset semantic relevance threshold, screen the semantic relevance, extract feature pairs greater than the semantic relevance threshold, establish connection relationships, and construct a heterogeneous feature graph based on the connection relationships;

[0090] In the heterogeneous feature graph, a graph node corresponding to the structured feature is selected as a starting node, and the heterogeneous feature graph is traversed through a depth-first search to obtain an access sequence of graph nodes; a pattern is extracted from the access sequence, and access sequences with the same node type transfer pattern are extracted as candidate meta-paths; the cumulative value of the semantic association between adjacent graph nodes on each candidate meta-path is calculated, and the candidate meta-path with the largest cumulative value is selected as the target meta-path;

[0091] Construct a multi-level attention aggregation network to perform aggregation operations and obtain initial fusion features;

[0092] The initial fusion features are divided into anchor features and contrast features, and data enhancement processing is performed on the anchor features and contrast features respectively to obtain enhanced anchor features and enhanced contrast features; the mutual information value between the enhanced anchor features and the enhanced contrast features is calculated, and based on the maximization criterion of the mutual information value, the feature representation parameters are updated to obtain the final patient comprehensive atlas features.

[0093] In a specific embodiment, structured, semi-structured and unstructured features of a patient are received as input. Structured features such as the patient's age, gender, diagnosis code, etc. can be represented by numerical values ​​or categories. Semi-structured features such as medical advice, examination reports, etc. can be represented by key-value pairs or tables. Unstructured features such as medical record text, medical images, etc. can be represented by text or image data. Taking a 60-year-old male patient with hypertension and diabetes as an example, the structured feature can be {age: 60, gender: male, diagnosis: [hypertension, diabetes]}, the semi-structured feature can be {blood pressure: 160 / 90mmHg, blood sugar: 10mmol / L}, and the unstructured feature can be a medical record text describing the patient's symptoms and treatment process.

[0094] Pair the input features in pairs and calculate the semantic association between each pair of features. The semantic association can be obtained by calculating the cosine similarity between feature vectors. For example, to calculate the semantic association between the diagnosis of "hypertension" and the measurement of "blood pressure", they can be represented as vectors respectively, and then the cosine similarity between the two vectors can be calculated. Assume that the semantic association between "hypertension" and "blood pressure" is calculated to be 0.8, the semantic association between "hypertension" and "blood sugar" is 0.2, and the semantic association between "age" and "blood pressure" is 0.1.

[0095] Set a semantic association threshold, such as 0.5. Filter out feature pairs with semantic association greater than the threshold, and establish connections between these feature pairs to construct a heterogeneous feature graph. In the above example, the semantic association between "hypertension" and "blood pressure" is greater than 0.5, so a connection is established between them. The semantic associations of other feature pairs are all less than 0.5, so no connection is established. The resulting heterogeneous feature graph contains three nodes: "hypertension", "blood pressure" and "age", and there is an edge between "hypertension" and "blood pressure".

[0096] Select the graph node corresponding to the structured feature as the starting node, for example, select the "age" node. Starting from the starting node, traverse the heterogeneous feature graph through the depth-first search algorithm and record the order of accessing the nodes. For example, starting from the "age" node, the possible access sequences are ["age"], because there are no other nodes connected to it. If we choose "hypertension" as the starting node, the possible access sequences are ["hypertension"], ["hypertension", "blood pressure"].

[0097] Perform pattern extraction on all access sequences, and extract sequences with the same node type transition pattern as candidate meta-paths. For example, [“hypertension”] and [“age”] both belong to single-node paths, and [“hypertension”, “blood pressure”] belong to the path from diagnosis to measurement. Therefore, [“hypertension”] and [“age”] can be used as one type of candidate meta-path, and [“hypertension”, “blood pressure”] can be used as another type of candidate meta-path.

[0098] For each candidate meta-path, calculate the cumulative value of the semantic association between adjacent nodes on the path. For example, for the path ["hypertension", "blood pressure"], its cumulative value is the semantic association between "hypertension" and "blood pressure", which is 0.8. Select the candidate meta-path with the largest cumulative value as the target meta-path. In this example, ["hypertension", "blood pressure"] has the largest cumulative value, so it is selected as the target meta-path.

[0099] Construct a multi-level attention aggregation network to aggregate the initial features along the target meta-path to obtain the initial fusion features. For example, along the path ["hypertension", "blood pressure"], the features of "hypertension" and "blood pressure" are aggregated.

[0100] The initial fusion features are divided into anchor features and contrast features. Data enhancement is performed on the anchor features and contrast features, such as adding random noise or performing random masking. The mutual information value between the enhanced anchor features and contrast features is calculated. Based on the maximization criterion of the mutual information value, the feature representation parameters are updated to obtain the final patient comprehensive atlas features.

[0101] In this embodiment, by fusing structured, semi-structured and unstructured data, the method can more comprehensively characterize the patient's health status and avoid information loss, thereby improving the accuracy and completeness of patient representation; the adaptive meta-path generation mechanism can dynamically adjust the feature fusion strategy according to different patient data, thereby improving the robustness and generalization ability of the model, enabling it to adapt to different data distributions and application scenarios; through a multi-level attention aggregation network and feature representation learning based on mutual information maximization, the method can effectively capture the complex relationship between patient characteristics, thereby improving the accuracy and reliability of disease prediction.

[0102] In an optional implementation, a multi-level attention aggregation network is constructed to perform aggregation operations, and the initial fusion features obtained include:

[0103] The multi-level attention aggregation network includes a node-level attention aggregator, a path-level attention aggregator, and a semantic-level attention aggregator, which perform aggregation operations in sequence, including:

[0104] Calculating node attention scores between adjacent graph nodes on the target meta-path through the node-level attention aggregator, taking the node attention scores as the first weight, performing weighted aggregation on features of adjacent graph nodes, and obtaining node-level fusion features;

[0105] Calculating the path importance scores of different target meta-paths through the path-level attention aggregator, taking the path importance scores as the second weight, performing weighted aggregation on the node-level fusion features on each target meta-path to obtain the path-level fusion features;

[0106] The semantic importance scores of different types of features are calculated through the semantic-level attention aggregator, and the semantic importance scores are used as the third weight to perform weighted aggregation on the path-level fusion features of each type to obtain the initial fusion features.

[0107] In a specific implementation, a node-level attention aggregator is first constructed. For pairs of adjacent graph nodes on the target meta-path, the feature vector of each graph node is extracted. For each pair of adjacent graph nodes, their feature vectors are respectively inner-producted with the shared attention vector to obtain a node attention score. The node attention score is converted into a node attention weight by normalization. Taking the smart medical scenario as an example, for the meta-path of "patient-disease-symptom", the attention weights between the patient node and the disease node, and between the disease node and the symptom node are calculated. The feature vectors of adjacent graph nodes are multiplied by the corresponding attention weights and summed to obtain node-level fusion features.

[0108] Then, a path-level attention aggregator is constructed. The node-level fusion features on all target meta-paths are collected and input into the path-level attention layer. The path-level attention layer contains learnable semantic vectors to capture the semantic importance of different meta-paths. The node-level fusion features of each meta-path are similarly calculated with the semantic vector to obtain the path importance score. The path importance score is converted into the path importance weight through normalization. Taking the smart medical scenario as an example, multiple meta-paths such as "patient-disease-symptoms" and "patient-examination-results" can be considered at the same time to calculate the importance weight of each path. The node-level fusion features of each meta-path are multiplied and summed with the corresponding path importance weight to obtain the path-level fusion features.

[0109] Finally, a semantic-level attention aggregator is constructed. Different types of path-level fusion features are aggregated, including patient features, disease features, symptom features, etc. A semantic-level attention network is designed, which contains multiple attention heads. Each attention head is responsible for learning the importance of features from a specific semantic perspective. Each type of feature is input into the semantic-level attention network, and the semantic importance scores under different attention heads are calculated. The semantic importance scores are converted into semantic importance weights through normalization. Taking the smart medical scenario as an example, attention heads of multiple semantic perspectives such as disease severity, symptom relevance, and examination indicators can be set. Each type of path-level fusion feature is multiplied by the corresponding semantic importance weight and summed to obtain the final initial fusion feature.

[0110] In the specific implementation, take the data of diabetes patients in a certain hospital as an example. For a newly admitted diabetic patient, the node-level attention aggregator is first used to analyze that the patient's correlation with the "type 2 diabetes" disease node is 0.8, and the correlation with the "hypertension" disease node is 0.3. Then, through the path-level attention aggregator analysis, it is found that the importance weight of the "patient-disease-symptom" path is 0.6, and the importance weight of the "patient-examination-result" path is 0.4. Finally, through the semantic-level attention aggregator analysis, it is found that the semantic weight from the perspective of disease severity is 0.5, the semantic weight from the perspective of complication risk is 0.3, and the semantic weight from the perspective of treatment effect is 0.2. Through multi-level attention aggregation, the comprehensive feature representation of the patient is finally obtained for subsequent personalized diagnosis and treatment plan recommendations.

[0111] In this embodiment, the node-level attention aggregator can accurately capture the correlation between different nodes (such as patients and diseases, diseases and symptoms, etc.) by calculating the attention weights between adjacent nodes, thereby better characterizing the characteristics of individual nodes; the path-level attention aggregator can dynamically adjust the contribution of different paths (such as "patient-disease-symptoms", "patient-examination-results") to the final feature representation by calculating the importance weights of meta-paths, which helps to highlight the role of key paths according to actual conditions; the semantic-level attention aggregator captures the importance of features from different semantic perspectives (such as disease severity, complication risk, treatment effect, etc.) through multiple attention heads, thereby achieving multi-dimensional feature fusion, making the final feature representation more comprehensive and accurate; through a multi-level attention mechanism, a personalized comprehensive feature representation can be generated for each patient, thereby improving the accuracy and effectiveness of subsequent personalized diagnosis and treatment plan recommendations.

[0112] In an optional embodiment, the patient's comprehensive graph features are input into a Bayesian causal network module, the causal relationship between the features is identified through a structural learning algorithm, and the causal path weight is set, the expectation maximization algorithm is used to solve the latent variable distribution, the causal relationship and the latent variable distribution are input into a temporal causal graph module, the progression of chronic atrophic gastritis is divided into an early stage, a middle stage, and a late stage, a temporal causal link is constructed, and the causal effect strength is calculated, and a disease risk prediction model is constructed according to the causal effect strength, and the disease risk prediction results of chronic atrophic gastritis in patients are generated, including:

[0113] Input the patient's comprehensive graph features into the Bayesian causal network, construct an optimization objective function of the structural learning algorithm, and the optimization objective function includes a structural fit scoring item and a structural complexity constraint item; based on the optimization objective function, the Bayesian causal network is iteratively searched to obtain an optimal network structure, and the weights of the causal paths in the optimal network structure are calculated; the expectation maximization algorithm is used to perform parameter learning on the optimal network structure, wherein the posterior probability distribution of the latent variables is calculated in each round of iteration, and the conditional probability parameters are updated until convergence, so as to obtain the causal relationship between the patient's comprehensive graph features and the latent variable distribution;

[0114] The causal relationship and the latent variable distribution are input into a temporal causal graph module, and the progression of chronic atrophic gastritis is divided into an early stage, a middle stage, and a late stage; the causal relationship is replicated within each stage to construct a local causal structure; the temporal correlation of nodes between adjacent stages is calculated, and a temporal causal link is established between nodes whose temporal correlation is greater than a preset correlation threshold; and the causal effect strength of the temporal causal link is obtained by causal intervention calculation;

[0115] A disease risk prediction model is constructed according to the causal effect strength, and the causal effect strength is used as a feature importance index to screen key features. A hierarchical predictor is constructed for the key features, and the hierarchical predictor includes an early to mid-term predictor and a mid-term to late-term predictor. Each hierarchical predictor integrates the local causal structure and temporal causal link information of the corresponding stage to generate a disease risk prediction result for patients with chronic atrophic gastritis.

[0116] In a specific embodiment, first, the comprehensive profile features of the patient are collected. These features may include endoscopic examination results, pathological histological features, blood indicators, gene expression data, and lifestyle information. For example, a 50-year-old male patient may have pale and thin gastric mucosa as shown by endoscopic examination, glandular atrophy and intestinal metaplasia as shown by pathological histological examination, decreased serum pepsinogen I / II ratio as shown by blood indicators, and a long-term smoking history.

[0117] Next, the collected comprehensive map features of the patients are input into the Bayesian causal network module. The optimization goal of the structural learning algorithm is constructed, which consists of two parts: the structural fit score item and the structural complexity constraint item. The structural fit score item is used to evaluate the degree of match between the network structure and the data, and the structural complexity constraint item is used to avoid the network structure being too complex. Through iterative search algorithms, such as greedy search or simulated annealing algorithms, the network structure is continuously adjusted to find the network structure that optimizes the optimization goal. In this example, the algorithm may find that there is a causal relationship between smoking and intestinal metaplasia, and there is also a causal relationship between a decrease in the pepsinogen I / II ratio and glandular atrophy.

[0118] After obtaining the optimal network structure, the weight of each causal path in the network is calculated. For example, the weight of smoking leading to intestinal metaplasia is 0.8, and the weight of reduced pepsinogen I / II ratio leading to glandular atrophy is 0.9. Then, the expectation maximization algorithm is used to learn the network parameters. The algorithm iteratively calculates the posterior probability distribution of latent variables (such as the degree of inflammation) and updates the conditional probability parameters until the algorithm converges. For example, the algorithm may infer that the patient has a high degree of chronic inflammation. Finally, the causal relationship and latent variable distribution between the comprehensive map features of the patient are obtained.

[0119] Then, the learned causal relationships and latent variable distributions are input into the temporal causal graph module. The progression of chronic atrophic gastritis is divided into three stages: early, middle, and late. The previously learned causal relationships are replicated within each stage to construct a local causal structure. For example, in the early stage, the causal relationship between smoking and intestinal metaplasia still exists. Next, the temporal correlation of nodes between adjacent stages is calculated. For example, the temporal correlation between intestinal metaplasia in the early stage and glandular atrophy in the middle stage is calculated. If the temporal correlation is greater than a preset threshold (e.g., 0.7), a temporal causal link is established between the two nodes. Then, the causal effect strength of each temporal causal link is calculated by the causal intervention method. For example, the causal effect strength of intestinal metaplasia in the early stage on glandular atrophy in the middle stage is 0.6.

[0120] Finally, a disease risk prediction model was constructed based on the calculated causal effect strength. The causal effect strength was used as a feature importance indicator to screen key features. For example, intestinal metaplasia and glandular atrophy were identified as key features. Then, a hierarchical predictor was constructed for these key features, which included an early to mid-term predictor and a mid-term to late predictor. Each predictor integrates the local causal structure and temporal causal link information of the corresponding stage. For example, the early to mid-term predictor integrates the causal relationship between smoking and intestinal metaplasia in the early stage, as well as the causal effect strength of intestinal metaplasia on glandular atrophy in the mid-term stage. Finally, the constructed prediction model was used to generate the disease risk prediction results of chronic atrophic gastritis in patients. For example, the patient's risk of developing moderate atrophic gastritis in the future was 80%, and the risk of developing severe atrophic gastritis was 30%.

[0121] In this embodiment, by constructing a time-series causal graph, the dynamic process of disease progression and the causal relationship between features are taken into account, so that the disease risk of chronic atrophic gastritis in patients can be predicted more accurately; risk factors that play a key role in disease progression can be identified, providing guidance for personalized prevention and treatment; based on causal inference, the reasons behind the predicted results can be explained, thereby enhancing the credibility and practicality of the model.

[0122] In an optional embodiment, based on the patient's disease risk prediction results, a risk level probability distribution matrix is ​​constructed, and the follow-up priority at different time points is calculated according to the risk level probability distribution matrix. The time series planning model is used to optimize the follow-up priority to obtain an initial follow-up plan, which includes:

[0123] Receive the patient identification, risk level identification and disease risk prediction probability value in the patient's disease risk prediction results; receive the patient identification, follow-up time point and disease status information in the patient's historical follow-up data;

[0124] Using the patient identifier as a row identifier and the risk level identifier as a column identifier, constructing a risk level probability distribution matrix, filling the disease risk prediction probability value into the corresponding position in the risk level probability distribution matrix, and normalizing the disease risk prediction probability value of each row in the risk level probability distribution matrix;

[0125] Assigning an increasing risk weight coefficient to each of the risk level identifiers;

[0126] Using the disease status information as the row identifier and column identifier of the Markov state transfer matrix, calculating the state transfer probability value between two adjacent follow-up time points and filling them into the corresponding positions, and normalizing the state transfer probability value of each row in the Markov state transfer matrix;

[0127] According to a preset follow-up time interval, the Markov state transfer matrix is ​​subjected to a power operation to obtain a state transfer probability matrix; the transfer probability values ​​in the direction of disease aggravation are extracted from the state transfer probability matrix and summed to obtain a disease aggravation probability value;

[0128] Selecting a maximum disease worsening probability value from a preset time interval set, and dividing the disease worsening probability value by the maximum disease worsening probability value to obtain a dynamic time attenuation factor;

[0129] The disease risk prediction probability value corresponding to each patient identifier in the risk level probability distribution matrix is ​​multiplied by the corresponding risk weight coefficient and the sum is calculated to obtain the patient basic follow-up priority corresponding to each patient identifier;

[0130] Multiplying the basic follow-up priority of the patient by the dynamic time decay factor to obtain the comprehensive follow-up priority of the patient;

[0131] The inverse of the comprehensive follow-up priority of the patient is set as the cost coefficient, the maximum number of follow-up persons at each time point is set as the medical resource capacity constraint value, the minimum follow-up interval is set as the time interval constraint value, a follow-up schedule is generated for each patient identifier, and the follow-up schedule result is output;

[0132] Calculate the average waiting time in the follow-up schedule result, use the negative value corresponding to the average waiting time as a reward value, pass it into the reinforcement learning model, and update the parameter value in the Markov state transfer matrix;

[0133] The patient identification in the follow-up schedule result is matched with the follow-up time point to generate an initial follow-up plan.

[0134] In a specific embodiment, the patient's disease risk prediction result and historical follow-up data are received. The disease risk prediction result includes a patient identifier, a risk level identifier, and a disease risk prediction probability value. The historical follow-up data includes a patient identifier, a follow-up time point, and a disease status information. For example, the risk prediction result of patient Zhang San is: the patient identifier is 001, the risk level identifier is high risk, and the disease risk prediction probability value is 0.8. The historical follow-up data of patient Li Si is: the patient identifier is 002, the follow-up time point is October 1, 2023, and the disease status is stable.

[0135] The risk level probability distribution matrix is ​​constructed with the patient identifier as the row identifier and the risk level identifier as the column identifier. The disease risk prediction probability value is filled into the corresponding position of the matrix. For example, if the risk level is divided into three levels: low risk, medium risk and high risk, then for patient Zhang San, the value corresponding to the high risk level is 0.8, and the values ​​corresponding to low risk and medium risk are initially 0. Next, the probability values ​​of each row in the matrix are normalized to ensure that the sum of the probability values ​​of each row is 1. Assuming that Zhang San's low-risk and medium-risk probability prediction values ​​are 0.1 and 0.1 respectively, after normalization, the probabilities of low risk, medium risk and high risk are 0.1, 0.1 and 0.8 respectively.

[0136] Assign an increasing risk weight factor to each risk level indicator. For example, the risk weight factors of low risk, medium risk and high risk are set to 1, 2 and 3 respectively.

[0137] The disease status information is used as the row and column identifiers of the Markov state transition matrix, and the state transition probability values ​​between two adjacent follow-up time points are calculated and filled into the corresponding positions. The probability values ​​of each row in the matrix are normalized. For example, assuming that the disease status is divided into three states: stable, worsening, and improved, and the disease status of patient Li Si on October 1, 2023 and November 1, 2023 is stable and worsening, respectively, then the probability of transitioning from a stable state to a worsening state is 1, and the probability of transitioning to other states is 0. After normalization, the probabilities of transitioning from a stable state to a stable, worsening, and improved state are 0, 1, and 0, respectively.

[0138] According to the preset follow-up time interval (e.g., one month, three months, six months), the Markov state transition matrix is ​​exponentially operated to obtain a new state transition probability matrix. The transition probability value in the direction of worsening of the disease is extracted from the new state transition probability matrix and summed to obtain the probability value of worsening of the disease. For example, assuming that the calculation is performed at an interval of one month, the probability of transitioning from a stable state to an worsening state is 0.2.

[0139] From the preset time interval set, select the maximum probability of disease progression. Divide the probability of disease progression calculated by the current time interval by the maximum probability of disease progression to obtain the dynamic time decay factor. For example, assuming the maximum probability of disease progression is 0.5, the current time interval is one month, and the probability of disease progression is 0.2, then the dynamic time decay factor is 0.2 / 0.5=0.4.

[0140] The disease risk prediction probability value corresponding to each patient identifier in the risk level probability distribution matrix is ​​multiplied by the corresponding risk weight coefficient and summed to obtain the patient basic follow-up priority corresponding to each patient identifier. For example, for patient Zhang San, his basic follow-up priority is 0.1×1+0.1×2+0.8×3=2.7.

[0141] Multiply the basic follow-up priority of the patient by the dynamic time decay factor to obtain the comprehensive follow-up priority of the patient. For example, the comprehensive follow-up priority of patient Zhang San is 2.7×0.4=1.08.

[0142] Set the inverse of the patient's comprehensive follow-up priority as the cost coefficient. Set the maximum number of follow-up visits at each time point as the medical resource capacity constraint value, and set the minimum follow-up interval as the time interval constraint value. Generate a follow-up schedule for each patient ID and output the follow-up schedule result. For example, the follow-up schedule for patient Zhang San is January 1, 2024.

[0143] Calculate the average waiting time in the follow-up schedule results. Use the negative value corresponding to the average waiting time as the reward value, pass it into the reinforcement learning model, and update the parameter value in the Markov state transfer matrix.

[0144] The patient ID in the follow-up schedule result is matched with the follow-up time point to generate an initial follow-up plan. For example, the initial follow-up plan for patient Zhang San is: patient ID 001, follow-up time January 1, 2024.

[0145] In this embodiment, by risk stratifying and prioritizing patients, limited medical resources can be allocated preferentially to high-risk patients, thereby improving the utilization efficiency of medical resources and avoiding waste of resources; through timely follow-up and intervention, the progression of the disease can be effectively controlled and the risk of disease worsening can be reduced, thereby improving patient prognosis and improving the patient's quality of life; through intelligent follow-up plan generation and optimization, manual intervention can be reduced, follow-up efficiency can be improved, follow-up costs can be reduced, and the overall follow-up process can be optimized.

[0146] In an optional implementation, the initial follow-up plan is subjected to a medical resource constraint analysis and the resource competition cost is calculated. The Nash equilibrium algorithm is used to perform multi-objective optimization on the follow-up priority and the resource competition cost. The generation of an optimized follow-up plan includes:

[0147] Receive the patient identification, follow-up time points, and comprehensive follow-up priority of the patient in the initial follow-up plan;

[0148] The number of follow-up patients at each follow-up time point is divided by the upper limit of the medical resource capacity to obtain the resource occupancy rate; the resource occupancy rate is subtracted from the preset resource occupancy threshold, and then squared and multiplied by the competition cost coefficient to obtain the resource competition cost at the time point; the resource competition costs at all the follow-up time points are accumulated to obtain the overall resource competition cost;

[0149] The patient identifier is set as a game participant, and the set of follow-up time points is set as a strategy space; the product of the comprehensive follow-up priority of the patient and the first weight coefficient is calculated to determine the first objective function; the product of the overall resource competition cost and the second weight coefficient is calculated to determine the second objective function; an initial follow-up time point is assigned to each of the patient identifiers; the patient identifier to be optimized is selected, and the follow-up time points of other patient identifiers are kept unchanged; the calculation result of the first objective function is subtracted from the calculation result of the second objective function to obtain the utility function value of the patient identifier to be optimized;

[0150] Based on the maximum utility function value, a corresponding follow-up time point is selected from the strategy space and updated to the optimal follow-up time point of the patient identification to be optimized;

[0151] Repeat the iteration for all patient identifiers. When the follow-up time points of all patient identifiers no longer change, the current follow-up time point is taken as the final follow-up time point.

[0152] Based on the patient identification and the final follow-up time point, the corresponding relationship is determined and the optimized follow-up plan is output.

[0153] In a specific embodiment, first, an initial follow-up plan is collected, which includes the identification of each patient, the planned follow-up time point, and the comprehensive follow-up priority evaluated based on factors such as condition and risk. For example, patient A, identified as 001, has a planned follow-up time point of the 3rd and 10th days, and a comprehensive follow-up priority rating of high, which is scored as 9 points; patient B, identified as 002, has a planned follow-up time point of the 5th and 10th days, and a comprehensive follow-up priority rating of medium, which is scored as 5 points; patient C, identified as 003, has a planned follow-up time point of the 3rd and 5th days, and a comprehensive follow-up priority rating of low, which is scored as 3 points. Assume that the upper limit of medical resource capacity is to receive 10 patients per day.

[0154] Next, calculate the resource utilization rate at each follow-up time point. For example, on the 3rd day, there are two patients A and C scheduled for follow-up, and the resource utilization rate is 2 / 10=0.2; on the 5th day, there are two patients B and C scheduled for follow-up, and the resource utilization rate is 2 / 10=0.2; on the 10th day, there are two patients A and B scheduled for follow-up, and the resource utilization rate is 2 / 10=0.2. The preset resource utilization threshold is 0.1, and the competitive cost coefficient is set to 10.

[0155] Then, the resource competition cost at each time point is calculated. Taking the third day as an example, the resource occupancy rate is 0.2 minus the threshold value 0.1, squared, and then multiplied by the competition cost coefficient 10, the resource competition cost at that time point is (0.2-0.1) 2 ×10=0.1. Using the same method, the resource competition cost on the 5th and 10th days is also 0.1. Adding up the resource competition cost at all time points, the total resource competition cost is 0.1+0.1+0.1=0.3.

[0156] Set the patient identifier as the game participant, and set all possible follow-up time points as the strategy space, for example, the strategy space is from day 1 to day 14. Set the first weight coefficient to 1 and the second weight coefficient to 0.5. The product of the patient's comprehensive follow-up priority and the first weight coefficient constitutes the first objective function, and the product of the overall resource competition cost and the second weight coefficient constitutes the second objective function.

[0157] Initially, the follow-up time points for patient A are day 3 and day 10. Now select patient A for optimization, and keep the follow-up time points for other patients unchanged. Assume that patient A chooses day 1 and day 7 for follow-up, recalculate the overall resource competition cost, and calculate the first objective function and the second objective function respectively. Subtract the result of the first objective function from the result of the second objective function to obtain the utility function value of patient A when choosing to follow up on day 1 and day 7.

[0158] Traverse all possible follow-up time point combinations in the strategy space and calculate the utility function value of patient A under each combination. Select the time point combination with the largest utility function value as the new follow-up time point for patient A. Repeat this optimization process for all patients until the follow-up time points of all patients no longer change, and finally determine the optimized follow-up time point for each patient.

[0159] Finally, based on the patient identifier and the final follow-up time points, the optimized follow-up plan is output. For example, the final follow-up time points for patient A are day 1 and day 7, the final follow-up time points for patient B are day 5 and day 12, and the final follow-up time points for patient C are day 3 and day 14.

[0160] In this embodiment, the individual needs of patients are taken into consideration, and follow-up of high-priority patients is given priority to meet the patients' expected follow-up time as much as possible; through the Nash equilibrium algorithm, the patient's needs and the limitations of medical resources are effectively balanced, excessive concentration or idleness of resources is avoided, and resource utilization is improved; by optimizing follow-up arrangements, resource competition and potential delays are reduced, thereby reducing overall medical costs and improving medical efficiency.

[0161] Figure 2 FIG. 1 is a schematic diagram of the structure of an intelligent follow-up system for chronic atrophic gastritis based on multi-source electronic medical records according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0162] The first unit is used to divide the multi-source electronic medical record data into a structured data layer, a semi-structured data layer and an unstructured data layer according to the data structure, pre-process the clinical monitoring data in the structured data layer by using an outlier detection algorithm and a time series filling algorithm to obtain structured features, standardize and map the diagnosis and treatment instruction data in the semi-structured data layer by using a medical word embedding model to obtain semi-structured features, extract key information from the medical process record data in the unstructured data layer by using a medical entity recognition model to obtain unstructured features, input the structured features, the semi-structured features and the unstructured features into a meta-path reasoning module, mine the inter-layer feature association relationship, and generate a comprehensive patient graph feature;

[0163] The second unit is used to input the comprehensive graph features of the patient into the Bayesian causal network module, identify the causal relationship between the features through the structural learning algorithm, set the causal path weight, use the expectation maximization algorithm to solve the latent variable distribution, input the causal relationship and the latent variable distribution into the temporal causal graph module, divide the progression of chronic atrophic gastritis into an early stage, a middle stage and a late stage, construct a temporal causal link, and calculate the causal effect strength, construct a disease risk prediction model according to the causal effect strength, and generate a disease risk prediction result of chronic atrophic gastritis in patients;

[0164] The third unit is used to construct a risk level probability distribution matrix based on the patient's disease risk prediction results, calculate the follow-up priority at different time points according to the risk level probability distribution matrix, and use the timing planning model to perform time series optimization on the follow-up priority to obtain the initial follow-up plan; perform medical resource constraint analysis on the initial follow-up plan and calculate the resource competition cost, use the Nash equilibrium algorithm to perform multi-objective optimization on the follow-up priority and resource competition cost, and generate an optimized follow-up plan.

[0165] According to a third aspect of the embodiments of the present invention,

[0166] An electronic device is provided, comprising:

[0167] processor;

[0168] a memory for storing processor-executable instructions;

[0169] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0170] A fourth aspect of the embodiments of the present invention is:

[0171] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0172] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0173] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent follow-up method for chronic atrophic gastritis based on multi-source electronic medical records, characterized in that: include: Divide the multi-source electronic medical record data into a structured data layer, a semi-structured data layer and an unstructured data layer according to the data structure, pre-process the clinical monitoring data in the structured data layer using an outlier detection algorithm and a time series filling algorithm to obtain structured features, use a medical word embedding model to perform standardized mapping on the diagnosis and treatment instruction data in the semi-structured data layer to obtain semi-structured features, use a medical entity recognition model to extract key information from the medical process record data in the unstructured data layer to obtain unstructured features, input the structured features, the semi-structured features and the unstructured features into a meta-path reasoning module, mine the inter-layer feature association relationship, and generate a comprehensive patient graph feature; Input the comprehensive graph features of the patient into the Bayesian causal network module, identify the causal relationship between the features through the structural learning algorithm, set the causal path weight, use the expectation maximization algorithm to solve the latent variable distribution, input the causal relationship and the latent variable distribution into the temporal causal graph module, divide the progression of chronic atrophic gastritis into an early stage, a middle stage and a late stage, construct a temporal causal link, and calculate the causal effect strength, construct a disease risk prediction model according to the causal effect strength, and generate a disease risk prediction result of chronic atrophic gastritis in patients; Based on the patient's disease risk prediction results, a risk level probability distribution matrix is ​​constructed. According to the risk level probability distribution matrix, the follow-up priority at different time points is calculated. The time series planning model is used to optimize the follow-up priority to obtain the initial follow-up plan, including: Receive the patient identification, risk level identification and disease risk prediction probability value in the patient's disease risk prediction results; receive the patient identification, follow-up time point and disease status information in the patient's historical follow-up data; Using the patient identifier as a row identifier and the risk level identifier as a column identifier, constructing a risk level probability distribution matrix, filling the disease risk prediction probability value into the corresponding position in the risk level probability distribution matrix, and normalizing the disease risk prediction probability value of each row in the risk level probability distribution matrix; Assigning an increasing risk weight coefficient to each of the risk level identifiers; Using the disease status information as the row identifier and column identifier of the Markov state transfer matrix, calculating the state transfer probability value between two adjacent follow-up time points and filling them into the corresponding positions, and normalizing the state transfer probability value of each row in the Markov state transfer matrix; According to a preset follow-up time interval, the Markov state transfer matrix is ​​subjected to a power operation to obtain a state transfer probability matrix; the transfer probability values ​​in the direction of disease aggravation are extracted from the state transfer probability matrix and summed to obtain a disease aggravation probability value; Selecting a maximum disease worsening probability value from a preset time interval set, and dividing the disease worsening probability value by the maximum disease worsening probability value to obtain a dynamic time attenuation factor; The disease risk prediction probability value corresponding to each patient identifier in the risk level probability distribution matrix is ​​multiplied by the corresponding risk weight coefficient and the sum is calculated to obtain the patient basic follow-up priority corresponding to each patient identifier; Multiplying the basic follow-up priority of the patient by the dynamic time decay factor to obtain the comprehensive follow-up priority of the patient; The inverse of the comprehensive follow-up priority of the patient is set as the cost coefficient, the maximum number of follow-up persons at each time point is set as the medical resource capacity constraint value, the minimum follow-up interval is set as the time interval constraint value, a follow-up schedule is generated for each patient identifier, and the follow-up schedule result is output; Calculate the average waiting time in the follow-up schedule result, use the negative value corresponding to the average waiting time as a reward value, pass it into the reinforcement learning model, and update the parameter value in the Markov state transfer matrix; Matching the patient identifier in the follow-up schedule result with the follow-up time point to generate an initial follow-up plan; The medical resource constraints of the initial follow-up plan were analyzed and the resource competition cost was calculated. The Nash equilibrium algorithm was used to perform multi-objective optimization on the follow-up priority and resource competition cost to generate an optimized follow-up plan.

2. The method according to claim 1, characterized in that The multi-source electronic medical record data is divided into a structured data layer, a semi-structured data layer and an unstructured data layer according to the data structure. The clinical monitoring data in the structured data layer is preprocessed by using an outlier detection algorithm and a time series filling algorithm to obtain structured features. The diagnosis and treatment instruction data in the semi-structured data layer is standardized and mapped by using a medical word embedding model to obtain semi-structured features. The medical process record data in the unstructured data layer is extracted by using a medical entity recognition model to obtain unstructured features. The structured features, the semi-structured features and the unstructured features are input into a meta-path reasoning module to mine the inter-layer feature association relationship to generate a comprehensive patient graph feature, including: For the clinical monitoring data at each time point in the structured data layer, the time series correlation weight between the clinical monitoring data at adjacent time points is calculated, the time series correlation weight is integrated into the local reachable density calculation to obtain a local outlier factor, and outliers are detected based on the local outlier factor; the outliers are filled using an adaptive Gaussian process regression algorithm with dynamically selected kernel function parameters to obtain structured features; For the diagnosis and treatment instruction data in the semi-structured data layer, a domain knowledge tree including a hierarchical relationship of medical concepts is constructed, the hierarchical relationship of medical concepts is encoded into a word vector, and the semantic similarity and semantic normativity of the word vector are simultaneously optimized through a multi-task learning framework, and the weight of the word vector is dynamically adjusted according to the context through an attention mechanism to obtain a semi-structured feature; For the medical process record data in the unstructured data layer, a bidirectional encoder is constructed to learn entity representation from the text dimension and the knowledge graph dimension respectively, and the medical ontology knowledge is used as the entity type constraint to guide the entity boundary recognition, extract the basic information in the medical process record data, and obtain the unstructured features; A heterogeneous feature graph is constructed according to the semantic association between the structured features, the semi-structured features and the unstructured features, a meta-path is adaptively generated based on the heterogeneous feature graph, and a multi-level attention aggregation network is used to perform feature fusion to obtain comprehensive graph features of the patient.

3. The method according to claim 2, characterized in that A heterogeneous feature graph is constructed according to the semantic association between the structured feature, the semi-structured feature and the unstructured feature. A meta-path is adaptively generated based on the heterogeneous feature graph, and a multi-level attention aggregation network is used to perform feature fusion to obtain the patient comprehensive graph features including: Receive structured features, semi-structured features and unstructured features as input features, wherein the input features form feature pairs in pairs, calculate the cosine similarity of the feature pairs, and obtain the semantic relevance of the feature pairs; based on a preset semantic relevance threshold, screen the semantic relevance, extract feature pairs greater than the semantic relevance threshold, establish connection relationships, and construct a heterogeneous feature graph based on the connection relationships; In the heterogeneous feature graph, a graph node corresponding to the structured feature is selected as a starting node, and the heterogeneous feature graph is traversed through a depth-first search to obtain an access sequence of graph nodes; a pattern is extracted from the access sequence, and access sequences with the same node type transfer pattern are extracted as candidate meta-paths; the cumulative value of the semantic association between adjacent graph nodes on each candidate meta-path is calculated, and the candidate meta-path with the largest cumulative value is selected as the target meta-path; Construct a multi-level attention aggregation network to perform aggregation operations and obtain the initial fusion features; The initial fusion features are divided into anchor features and contrast features, and data enhancement processing is performed on the anchor features and contrast features respectively to obtain enhanced anchor features and enhanced contrast features; the mutual information value between the enhanced anchor features and the enhanced contrast features is calculated, and based on the maximization criterion of the mutual information value, the feature representation parameters are updated to obtain the final patient comprehensive atlas features.

4. The method according to claim 3, characterized in that Construct a multi-level attention aggregation network for aggregation operation, and obtain the initial fusion features including: The multi-level attention aggregation network includes a node-level attention aggregator, a path-level attention aggregator, and a semantic-level attention aggregator, which perform aggregation operations in sequence, including: Calculating node attention scores between adjacent graph nodes on the target meta-path through the node-level attention aggregator, taking the node attention scores as the first weight, performing weighted aggregation on features of adjacent graph nodes, and obtaining node-level fusion features; Calculating the path importance scores of different target meta-paths through the path-level attention aggregator, taking the path importance scores as the second weight, performing weighted aggregation on the node-level fusion features on each target meta-path to obtain the path-level fusion features; The semantic importance scores of different types of features are calculated through the semantic-level attention aggregator, and the semantic importance scores are used as the third weight to perform weighted aggregation on the path-level fusion features of each type to obtain the initial fusion features.

5. The method according to claim 1, characterized in that The patient's comprehensive graph features are input into the Bayesian causal network module, the causal relationship between the features is identified through the structural learning algorithm, and the causal path weight is set. The expectation maximization algorithm is used to solve the latent variable distribution, and the causal relationship and the latent variable distribution are input into the temporal causal graph module. The progression of chronic atrophic gastritis is divided into an early stage, a middle stage, and a late stage, a temporal causal link is constructed, and the causal effect strength is calculated. A disease risk prediction model is constructed according to the causal effect strength, and the disease risk prediction results of chronic atrophic gastritis in patients are generated, including: Input the patient's comprehensive graph features into the Bayesian causal network, construct an optimization objective function of the structural learning algorithm, and the optimization objective function includes a structural fit scoring item and a structural complexity constraint item; based on the optimization objective function, the Bayesian causal network is iteratively searched to obtain an optimal network structure, and the weights of the causal paths in the optimal network structure are calculated; the expectation maximization algorithm is used to perform parameter learning on the optimal network structure, wherein the posterior probability distribution of the latent variables is calculated in each round of iteration, and the conditional probability parameters are updated until convergence, so as to obtain the causal relationship between the patient's comprehensive graph features and the latent variable distribution; The causal relationship and the latent variable distribution are input into a temporal causal graph module, and the progression of chronic atrophic gastritis is divided into an early stage, a middle stage, and a late stage; the causal relationship is replicated within each stage to construct a local causal structure; the temporal correlation of nodes between adjacent stages is calculated, and a temporal causal link is established between nodes whose temporal correlation is greater than a preset correlation threshold; and the causal effect strength of the temporal causal link is obtained by causal intervention calculation; A disease risk prediction model is constructed according to the causal effect strength, and the causal effect strength is used as a feature importance index to screen key features. A hierarchical predictor is constructed for the key features, and the hierarchical predictor includes an early to mid-term predictor and a mid-term to late-term predictor. Each hierarchical predictor integrates the local causal structure and temporal causal link information of the corresponding stage to generate a disease risk prediction result for patients with chronic atrophic gastritis.

6. The method according to claim 1, characterized in that The initial follow-up plan was analyzed for medical resource constraints and the resource competition cost was calculated. The Nash equilibrium algorithm was used to perform multi-objective optimization on the follow-up priority and resource competition cost. The optimized follow-up plan was generated, including: Receive the patient identification, follow-up time points, and comprehensive follow-up priority of the patient in the initial follow-up plan; The number of follow-up patients at each follow-up time point is divided by the upper limit of the medical resource capacity to obtain the resource occupancy rate; the resource occupancy rate is subtracted from the preset resource occupancy threshold, and then squared and multiplied by the competition cost coefficient to obtain the resource competition cost at the time point; the resource competition costs at all the follow-up time points are accumulated to obtain the overall resource competition cost; The patient identifier is set as a game participant, and the set of follow-up time points is set as a strategy space; the product of the comprehensive follow-up priority of the patient and the first weight coefficient is calculated to determine the first objective function; the product of the overall resource competition cost and the second weight coefficient is calculated to determine the second objective function; an initial follow-up time point is assigned to each of the patient identifiers; the patient identifier to be optimized is selected, and the follow-up time points of other patient identifiers are kept unchanged; the calculation result of the first objective function is subtracted from the calculation result of the second objective function to obtain the utility function value of the patient identifier to be optimized; Based on the maximum utility function value, a corresponding follow-up time point is selected from the strategy space and updated to the optimal follow-up time point of the patient identification to be optimized; Repeat the iteration for all patient identifiers. When the follow-up time points of all patient identifiers no longer change, the current follow-up time point is taken as the final follow-up time point. Based on the patient identification and the final follow-up time point, the corresponding relationship is determined and the optimized follow-up plan is output.

7. An intelligent follow-up system for chronic atrophic gastritis based on multi-source electronic medical records, used to implement the method according to any one of claims 1 to 6, characterized in that: include: The first unit is used to divide the multi-source electronic medical record data into a structured data layer, a semi-structured data layer and an unstructured data layer according to the data structure, pre-process the clinical monitoring data in the structured data layer by using an outlier detection algorithm and a time series filling algorithm to obtain structured features, standardize and map the diagnosis and treatment instruction data in the semi-structured data layer by using a medical word embedding model to obtain semi-structured features, extract key information from the medical process record data in the unstructured data layer by using a medical entity recognition model to obtain unstructured features, input the structured features, the semi-structured features and the unstructured features into a meta-path reasoning module, mine the inter-layer feature association relationship, and generate a comprehensive patient graph feature; The second unit is used to input the comprehensive graph features of the patient into the Bayesian causal network module, identify the causal relationship between the features through the structural learning algorithm, set the causal path weight, use the expectation maximization algorithm to solve the latent variable distribution, input the causal relationship and the latent variable distribution into the temporal causal graph module, divide the progression of chronic atrophic gastritis into an early stage, a middle stage and a late stage, construct a temporal causal link, and calculate the causal effect strength, construct a disease risk prediction model according to the causal effect strength, and generate a disease risk prediction result of chronic atrophic gastritis in patients; The third unit is used to construct a risk level probability distribution matrix based on the patient's disease risk prediction results, calculate the follow-up priority at different time points according to the risk level probability distribution matrix, and use the time series planning model to optimize the follow-up priority to obtain the initial follow-up plan; The medical resource constraints of the initial follow-up plan were analyzed and the resource competition cost was calculated. The Nash equilibrium algorithm was used to perform multi-objective optimization on the follow-up priority and resource competition cost to generate an optimized follow-up plan.

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

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

Images