A hypertension management system

Abstract

The present invention discloses a hypertension management system, belonging to the field of medical data processing technology, comprising: a data acquisition module for acquiring vital sign data and text data; a data preprocessing module for cleaning and processing the data acquired by the data acquisition module; a multimodal data melting module for performing time series feature extraction on vital sign data to obtain vital sign feature vectors; further for performing feature extraction on text data to obtain text feature vectors; further for integrating the vital sign feature vectors and text feature vectors into a joint vector; and a question-answer generation module for acquiring the joint vector through a large language model to understand the patient's questions, and then extracting corresponding information from a pre-constructed knowledge graph and generating answers that match the patient's questions. The hypertension management system solves the problem that existing intelligent follow-up systems have difficulty performing multimodal fusion analysis, resulting in low accuracy and adaptability of the system output results.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data processing, and in particular to a hypertension management system. Background Art

[0002] Hypertension, one of the most common chronic diseases worldwide, can lead to serious consequences such as cardiovascular and cerebrovascular disease and kidney damage if left uncontrolled for a long time, placing a significant burden on patients and society. Traditional hypertension follow-up management often relies on regular checkups and manual record-keeping by medical staff. This model is not only time-consuming and inefficient, but also makes it difficult to implement personalized health interventions. In chronic disease management, factors such as patient condition changes, the impact of lifestyle habits, and the efficacy of medications all require real-time tracking and analysis.

[0003] With the continuous advancement of intelligent technology, intelligent medical follow-up management systems based on big data and artificial intelligence are becoming a key technical means to improve the efficiency of hypertension management. Existing intelligent follow-up systems mostly focus on automated health data collection, data analysis, and report generation. However, they still struggle with multimodal fusion analysis of different types of data (including health signs, lifestyle, and medical history), resulting in low accuracy and adaptability of the system output. Summary of the Invention

[0004] In order to overcome the defects of the prior art, the present invention provides a hypertension management system to solve the above problems.

[0005] The technical solution adopted by the present invention to solve the technical problem is: a hypertension management system, comprising:

[0006] A data acquisition module is used to acquire vital sign data and text data;

[0007] The data preprocessing module is used to clean and process the data acquired by the data acquisition module;

[0008] The multimodal data melting module is used to extract time series features from vital sign data using Transformer-XL to obtain vital sign feature vectors; it is also used to extract features from text data using Dual-Stream CNN to obtain text feature vectors; and it is also used to integrate the vital sign feature vectors and text feature vectors into a joint vector;

[0009] The question-answer generation module is used to obtain the joint vector through a large language model to understand the patient's question, and then extract the corresponding information from the pre-built knowledge graph and generate an answer that meets the patient's question.

[0010] Optionally, the knowledge graph is obtained by executing a medical term recognition module and a knowledge graph construction module; wherein the medical term recognition module is used to recognize medical terms in text data and extract the association relationship between different medical terms; and the knowledge graph construction module is used to construct a knowledge graph based on the association relationship between different medical terms.

[0011] It is worth noting that the medical terminology recognition module uses the context-aware embedding generated by RoBERTa as the basis for node embedding of the graph neural network (GNN), combined with the structured message passing of the graph neural network (GNN) to extract the semantic relationships between different medical terms.

[0012] Preferably, in the RoBERTa, the input layer of RoBERTa captures the contextual semantic information of medical terms in their corresponding text data as a context-aware embedding representation of the medical terms;

[0013] In the graph neural network (GNN), each medical term is regarded as a node in the graph, and the node embedding is updated through the message passing mechanism of the graph neural network (GNN) itself, so that the representation of each node integrates its own contextual information and the semantic relationship of neighboring nodes to enhance the understanding of the association between nodes.

[0014] Specifically, in the RoBERTa, medical terms After processing by RoBERTa, medical terminology is generated Context-aware embedding representation of ;

[0015] In the graph neural network GNN, medical terminology The corresponding embedding representation of node v The iterative update formula at the kth layer is ;in, is the weight matrix of the kth layer, is the set of neighbor nodes of node v, is the activation function.

[0016] It is worth noting that in the multimodal data melting module, the text data is converted into a word vector sequence through the embedding layer of the Dual-Stream CNN, and the core content of the text data is extracted through the convolution layer of the semantic stream and the syntactic stream. The formula is: ; ;in represents the output of the semantic flow, represents the output of the syntactic stream, Represents the input text data, represents the weight matrix of the semantic stream convolution layer, represents the weight matrix of the convolutional layer of the syntactic stream, and Used to control the feature extraction method, represents the bias term of the semantic stream convolution layer, represents the bias term of the convolutional layer of the syntactic stream, and Used to adjust the output of convolution calculation;

[0017] The text feature vector ,in, Used to control the weight ratio of semantic features, Used to control the weight ratio of syntactic features.

[0018] Preferably, in the multimodal data melting module, the calculation formula of the joint vector H is: ;in Indicates that the vital signs data are processed by Transformer-XL Extracted physical sign feature vector Indicates that text data is processed by Dual-Stream CNN Extracted text feature vector.

[0019] Optionally, for the vital sign data, the data preprocessing module is further used to analyze outliers in the vital sign data by Z-score and remove outliers;

[0020] The formula for Z-score analysis is: ,in is the vital sign data to be analyzed, is the mean of the data set where the physical sign data are located, is the standard deviation of the data set where the vital signs data are located, express and The distance between The physical sign data whose absolute value is greater than the preset analysis threshold is regarded as an abnormal value.

[0021] Specifically, the hypertension management system further includes an evaluation and optimization module, which includes an automatic optimization mechanism and an expert feedback mechanism;

[0022] The automatic optimization mechanism is used to use an evolutionary algorithm to simulate the treatment plan generated by the question-answer generation module that meets the answers to the patient's questions and optimize the treatment plan;

[0023] In the evolutionary algorithm, multiple candidate treatment plans are first obtained from the question-answer generation module. During each optimization of the evolutionary algorithm, the evolutionary algorithm gradually adjusts the variables in the candidate plans by simulating patient feedback, and selects the optimal treatment plan through multiple iterations. In the evolutionary algorithm, the population size represents the number of candidate treatment plans generated in each generation, the crossover rate represents the probability of controlling the exchange of information between two treatment plans, the mutation rate represents the probability of treatment plan mutation in each generation, and the fitness function represents the evaluation of the quality of the treatment plan.

[0024] The expert feedback mechanism is used to obtain expert feedback text content for the treatment plan generated in the question and answer generation module that meets the answers to the patient's questions, convert the expert feedback text content from natural language into machine language input through natural language processing, perform semantic analysis on the expert feedback text content converted into machine language input, extract medical guidance opinions, and then convert the medical guidance opinions into variable adjustment commands to optimize the treatment plan generated in the question and answer generation module that meets the answers to the patient's questions.

[0025] Preferably, the hypertension management system further comprises a complexity discrimination module, which is used to evaluate the complexity of the patient's condition through multi-dimensional weighted scoring. ; Complexity Score is the comprehensive complexity score, The first The weight of the factors, It is The value of a factor.

[0026] The beneficial effects of the present invention lie in: the hypertension management system utilizes a large language model for intelligent question-answering, generating precise health recommendations based on a patient's vital sign data, text data, and knowledge graph. This integration not only improves the efficiency of health management but also enables dynamic adjustments based on real-time patient changes, ensuring each patient receives the most appropriate treatment plan. The multimodal data fusion module effectively integrates multimodal information from diverse data sources, including vital sign data and text data. By employing a deep learning architecture combining Transformer-XL with a dual-stream convolutional neural network (Dual-Stream CNN), it simultaneously processes time-series vital sign data and natural language text data, providing patients with more comprehensive and accurate health assessments and management. This multimodal fusion approach empowers the hypertension management system with enhanced comprehensive capabilities in identifying health issues, enabling it to capture complex underlying connections between different data sources. The question-answer generation module, combining the large language model with the knowledge graph, generates highly customized health recommendations based on a patient's personalized health data. Compared to traditional health consultation systems, this intelligent question-answering system based on a large language model not only understands patient needs in real time but also deeply integrates personal health records and medical knowledge to provide practical and actionable recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a schematic diagram of the architecture of a hypertension management system in one embodiment of the present invention;

[0028] Figure 2 Schematic diagram of the architecture of a data acquisition module and a data preprocessing module in one embodiment of the present invention;

[0029] Figure 3 This is a schematic diagram of the architecture of a medical terminology recognition module and a knowledge graph construction module in one embodiment of the present invention;

[0030] Figure 4 Schematic diagram of the architecture of a multimodal data fusion module and a question-answer generation module in one embodiment of the present invention;

[0031] Figure 5 FIG. 1 is a schematic diagram of the architecture of a complexity determination module and an urgency determination module in one embodiment of the present invention. DETAILED DESCRIPTION

[0032] The following is a further description of specific embodiments of the present invention in conjunction with the accompanying drawings. It should be noted that the description of these embodiments is intended to facilitate understanding of the present invention and does not constitute a limitation of the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0033] like Figure 1-5 As shown, a hypertension management system includes: a data acquisition module, a data preprocessing module, a medical term recognition module, a knowledge graph construction module, a multimodal data melting module, and a question and answer generation module;

[0034] The data acquisition module is used to obtain vital sign data and text data; vital sign data include blood pressure, heart rate and body temperature, etc., and vital sign data are collected in real time from wearable devices (such as smart bracelets and blood pressure monitors); text data includes living habits, diet and medication use, and text data are obtained through questionnaires or patient self-reports; by obtaining vital sign data and text data, it is ensured that the data covers all aspects of health management; these vital sign data and text data as multimodal data provide comprehensive health data support for subsequent analysis, and the patient's health information is transmitted to the hypertension management system in real time to form a continuously updated database to ensure that the health status of each patient can be monitored comprehensively and in real time.

[0035] The data preprocessing module is used to clean and process the data obtained by the data acquisition module to improve data quality and consistency; for example, duplicate records are deleted through a deduplication algorithm to ensure data accuracy; missing data is supplemented through interpolation methods (such as linear interpolation or model-based interpolation); in this embodiment, vital sign data are ensured to be accurate and complete through operations such as deduplication, outlier detection, and missing value filling; at the same time, all vital sign data will be unified and normalized to ensure consistency of data from different sources during the analysis process; text data is optimized through preprocessing such as word segmentation, removal of stop words, and stemming to lay the foundation for subsequent semantic analysis.

[0036] For text data, the data preprocessing module is used to remove noise information through text preprocessing technology, and the text preprocessing technology includes removing stop words, word segmentation and stemming;

[0037] For the vital sign data, the data preprocessing module is used to unify the units of the vital sign data (such as the blood pressure unit is unified into mmHg) and perform normalization processing on the vital sign data.

[0038] For vital sign data, the data preprocessing module is further used to analyze and remove outliers in the vital sign data using Z-score analysis; Z-score analysis (standard score) is a standardized indicator used to measure the degree of deviation of a data point from the mean of a data set;

[0039] The formula for Z-score analysis is: ,in is the vital sign data to be analyzed, is the mean of the data set where the physical sign data are located, is the standard deviation of the data set where the vital signs data are located, express and If the distance between If it is positive, it means that the physical sign data is higher than the mean. Negative means the physical sign data is lower than the mean; filter out The physical sign data whose absolute value is greater than the preset analysis threshold is regarded as an abnormal value.

[0040] The multimodal data melting module is used to extract time series features from vital sign data using Transformer-XL to obtain vital sign feature vectors. Since indicators such as blood pressure and heart rate in hypertension follow-up data are time-dependent, and the follow-up period may last for weeks or even months, Transformer-XL's extended memory mechanism can capture long-term changes in health status. Specifically, Transformer-XL uses the self-attention mechanism to identify short-term fluctuations and long-term trends in vital sign data. The self-attention formula is: , , is the key, is the value matrix, is the key dimension, and T is the transpose of K. This ensures that the vital sign data at each moment can be related to the vital sign data before the current moment. In addition, the extended memory mechanism of Transformer-XL retains the historical state during long-term follow-up, enabling the multimodal data melting module to better understand the changes in patient vital signs over time. The calculation formula for extended memory is: ;in Indicates the The extended memory of the layer is used to process and memorize the long-term dependencies from the vital sign data by extracting temporal features from the vital sign data. It contains historical information from the previous layer. Indicates stopping gradient propagation (i.e. stopping updating historical information); It is the output of the previous layer and provides the context of the time series data. Therefore, this extended memory mechanism can effectively capture long-term dependencies in the analysis of health data over a long period of time. For example, monitoring the changing trends of physiological parameters such as blood pressure or heart rate can help the system more accurately understand the dynamic changes of the patient's health status over time.

[0041] The multimodal data fusion module is designed to extract features from text data using a dual-stream CNN to generate text feature vectors. Follow-up text data typically contains descriptive information, such as the patient's lifestyle and diet. To deconstruct and capture the key word meanings and contextual information in the text data, the dual-stream CNN processes semantic and syntactic information separately. The embedding layer of the dual-stream CNN converts the text data into a sequence of word vectors. This is then passed through the convolutional layers of the semantic and syntactic streams to extract the core content of the text data. The formula is: in represents the output of the semantic flow, represents the output of the syntactic stream, Represents the input text data, represents the weight matrix of the semantic stream convolution layer, represents the weight matrix of the convolutional layer of the syntactic stream, and Used to control the feature extraction method, represents the bias term of the semantic stream convolution layer, represents the bias term of the convolutional layer of the syntactic stream, and Used to adjust the output of convolution calculation. The semantic stream can effectively represent the semantic features of the text by capturing the contextual relevance of the text content and the deep semantic connection between words; while the syntactic stream focuses on extracting the grammatical structure information of the text, such as the dependency relationship of sentences or the modification relationship between words. Ultimately, the output features of these two streams are and Will be combined in the fusion layer to form the final text feature vector , specifically: ,in, and are the learnable parameters of the fusion layer, Used to control the weight ratio of semantic features, This is used to control the weighting of syntactic features. Through this fusion mechanism, Dual-Stream CNN not only provides a deep understanding of semantic information but also captures the syntactic regularities of text. During hypertension follow-up, patients may describe changes in their symptoms or multi-dimensional lifestyle information, such as "I've been feeling very stressed lately, and my blood pressure feels a bit high." This statement combines semantic and syntactic features. The semantic stream of the Dual-Stream CNN extracts the impact of diet and mood on health, while the syntactic stream analyzes the causal relationships within the sentence. This allows for more precise management of hypertension by combining semantic and syntactic features (e.g., blood pressure and mood).

[0042] The multimodal data melting module is further used to integrate the physical sign feature vector and the text feature vector into a joint vector; the calculation formula of the joint vector H is: ;in represents the sign feature vector, Used to input vital signs data Convert to embedded representation , and then processed by the multi-layer self-attention mechanism of Transformer-XL to capture the long-range dependencies and short-term fluctuations in the vital sign data, and finally generated the final vital sign feature vector through the fully connected layer and nonlinear activation function. ; represents the text feature vector, Used to input text data Convert to word vector sequence , then processed by the semantic stream and syntactic stream of Dual-Stream CNN to extract the semantic features and syntactic features in the text respectively, and finally the output of the two streams are combined to generate the final text feature vector through the fully connected layer and nonlinear activation function ; In order to cope with the high-dimensional characteristics of the follow-up data, the system will perform the joint vector Normalize the rows and use principal component analysis (PCA) to reduce the dimension when necessary. This step can reduce noise and redundant features while keeping the information intact, so as to enhance the input quality of the subsequent large language model.

[0043] In this embodiment, the multimodal data fusion module is designed specifically to process the complex data during hypertension follow-up, providing comprehensive preparation for the data input of the subsequent large language model. The follow-up process usually involves two main data sources: vital sign data and text data. The deep learning architecture of Transformer-XL and Dual-Stream CNN can effectively fuse these data and solve practical problems encountered in data preprocessing. This multimodal fusion process not only improves the system's ability to process complex data during hypertension follow-up, but also provides accurate and rich basic data for subsequent health management and personalized recommendations, enabling the large language model to provide more targeted health guidance and solutions for hypertensive patients based on multimodal information.

[0044] The multimodal data fusion module needs to fuse the patient's vital sign data and text data. Considering the long-term dependency of vital sign data such as blood pressure and heart rate and text data such as medication usage during follow-up, Transformer-XL is used to process these temporal features. Transformer-XL performs weighted fusion of vital sign data through the self-attention mechanism. The calculation formula is as follows: , , ;in, Representing the query, and the self-attention formula The Q in is consistent; Represents the key, and the self-attention formula The K in is consistent; Value matrix, and self-attention formula The V in is consistent; is the time step The hidden state of , which contains the time series data of the patient's health history, is fused through the attention mechanism to extract long-term and short-term trend information; yes The weight matrix, yes The weight matrix, yes Through these calculations, long-term and short-term trends can be extracted from historical health data, providing effective support for subsequent generation of answers.

[0045] In this example, by integrating vital sign data and text data, a comprehensive assessment of a patient's health status can be made, and personalized health recommendations can be generated based on the fused joint vector. In this process, Transformer-XL is used to extract temporal features from the vital sign data, while a convolutional neural network (CNN) is used to process semantic information from the text data, ensuring the effective fusion of multi-source heterogeneous data.

[0046] The question-answer generation module uses a large language model to obtain the joint vector to understand the patient's question, then extracts the corresponding information from the pre-built knowledge graph and generates a response that matches the patient's question. When a patient asks a health-related question, a large language model (such as ChatGPT) can understand the specific content of the patient's question and extract relevant information from the knowledge graph to generate health recommendations and treatment plans that meet the patient's actual needs.

[0047] The large language model generates the final answer by fusing multimodal data (such as patient vital signs, text data, and information in the knowledge graph) with the question intent. The generation process is based on the autoregressive mechanism of the large language model, and the calculation formula is as follows: ;in, Indicates the generated words, is the joint vector of the input, including the user's text data, the patient's vital signs data (such as blood pressure, heart rate, etc.) and the information in the knowledge graph, In the context and joint vector Under the conditions, the This generation process ensures that the language model can effectively combine multimodal data and provide personalized and accurate health recommendations through the knowledge graph.

[0048] The Q&A module can understand user health questions from multiple dimensions, integrate data, and perform knowledge inference to generate personalized and accurate health recommendations. This effectively improves the effectiveness of hypertension follow-up and management, providing patients with more scientific and customized health management solutions.

[0049] It is worth noting that the knowledge graph is obtained by executing the medical term recognition module and the knowledge graph construction module, wherein the medical term recognition module is used to identify medical terms in text data and extract the association relationship between different medical terms;

[0050] When the medical term recognition module identifies medical terms in text data, it first utilizes a self-supervised pre-training model to ensure high-accuracy recognition and classification of medical terms. In particular, the medical term recognition module adopts a domain-specific model such as Med-BERT and trains it through self-supervised learning. It then fine-tunes a wide range of medical datasets in combination with the characteristics of the medical knowledge graph, enabling the medical term recognition module to effectively identify and classify medical terms such as "chest pain" and "dizziness." In addition, the medical term recognition module also combines word embedding technology (Word Embedding) and regularization enhancement methods, which enables the medical term recognition module to more accurately perceive and understand medical semantics. Word embedding converts medical terms into vector representations, thereby improving the processing accuracy of symptoms and their related semantics.

[0051] Patients can ask the system health-related questions. The question-answer generation module, based on a large language model, combines the patient's text data and medical knowledge graph to generate accurate, personalized health advice. For example, a user asks, "How can I control my blood pressure?" or "My blood pressure has fluctuated significantly recently. What should I do?" When generating answers, the question-answer generation module performs semantic understanding and intent extraction on the patient's natural language questions to obtain keywords. Using natural language processing techniques (such as word embedding), these keywords (such as "blood pressure control") are converted into feature vectors. This process uses pre-trained models (such as Word2Vec or BERT) for embedding conversion, calculated using the following formula: ,in, Represents words in text data The embedding vector of is the word in the text data The original representation of is the mapping function of the pre-trained model. In this way, keywords such as "blood pressure" and "control" in the text data are converted into high-dimensional vectors. Together with the embedded representation of the vital sign data, they form a joint vector for the multimodal data fusion module and are input into the answer generation module to capture its deep semantic information and support the generation of personalized health recommendations.

[0052] Once a medical term is identified, the Medical Term Recognition Module enters the Medical Term Relationship Extraction phase, identifying the associations between different medical terms. For example, when a patient describes "recently increased salt intake, resulting in elevated blood pressure," the system can identify the causal relationship between salt intake and blood pressure. To accurately extract these relationships, the Medical Term Recognition Module uses a deep learning model based on RoBERTa and graph neural networks (GNNs) to extract the associations between different medical terms.

[0053] The medical term recognition module uses the context-aware embedding generated by RoBERTa as the basis for node embedding in the graph neural network (GNN). Combined with the structured message passing of the graph neural network (GNN), it extracts the semantic relationships between different medical terms, including causal relationships and correlation relationships.

[0054] In RoBERTa, its input layer captures the contextual semantic information of medical terms in their corresponding text data as context-aware embedding representations of the medical terms, providing initial features for the node representations of the graph neural network (GNN). In the GNN, each medical term is considered a node in the graph, and the node embeddings are updated through the GNN's message passing mechanism. This allows the representation of each node to incorporate its own contextual information and the semantic relationships of neighboring nodes, enhancing the understanding of associations between nodes. This process enables the medical term recognition module to efficiently and accurately identify complex relationships between terms.

[0055] In the RoBERTa, medical terms Processed by RoBERTa to generate medical terms Context-aware embedding representation of ;

[0056] In the graph neural network GNN, medical terms Corresponding nodes Embedded representation of In the The layer iterative update formula is: ;in, It is The weight matrix of the layer, For nodes The set of neighbor nodes of is the activation function.

[0057] During the query process of the medical knowledge graph, the system uses the graph neural network (GNN) to query relevant information in the graph based on the patient's health data and question intent. The GNN updates the status of new nodes through the message passing mechanism. The update formula is: ;in, Representation node In the Embedding after a round of message passing, is a node The neighbor set of and are the weight matrix and bias term respectively, is the activation function. Through this process, the system can accurately extract relevant treatment recommendations and plans from the medical knowledge graph.

[0058] The above two formulas correspond to different iterative update processes. Emphasize nodes The update depends not only on the neighbor nodes characteristics, including its own characteristics. It is another message passing mechanism that does not include the node's own features and only relies on the weighted sum of neighbor node features.

[0059] Finally, the medical terminology recognition module achieves accurate recognition of complex relationships between terms through the context-aware features of RoBERTa and the structured message passing of the graph neural network (GNN).

[0060] The knowledge graph construction module is used to construct a knowledge graph based on the associations between different medical terms; through the identified medical terms and their associations, the knowledge graph construction module constructs a comprehensive medical knowledge graph. This knowledge graph not only integrates basic information related to blood pressure management, such as treatment methods, medications, and symptoms, but also combines important information such as the patient's personal medical history and lifestyle habits. In this way, the hypertension management system can provide rich knowledge support for subsequent health management and personalized recommendations. With the help of this knowledge graph, the hypertension management system can perform reasoning and queries, and provide accurate health advice based on the patient's specific health status and symptoms. For example, when the hypertension management system analyzes the patient's blood pressure fluctuations, it can provide personalized treatment plans based on the knowledge graph, thereby realizing intelligent management of the patient's health.

[0061] In this example, Med-BERT, a natural language processing technology, is used to identify medical terms and accurately extract information related to hypertension, including symptoms, treatments, medications, and medical history. Graph Neural Network (GNN) technology is then used to extract the relationships between medical terms and construct a dynamic hypertension health knowledge graph. This knowledge graph not only encompasses known medical knowledge but also integrates individualized patient information to provide more precise health management recommendations.

[0062] The hypertension management system also includes an evaluation and optimization module, which is integrated into the question-answer generation module to further enhance the personalized health management capabilities of the hypertension management system. The evaluation and optimization module includes an automatic optimization mechanism and an expert feedback mechanism.

[0063] The automatic optimization mechanism is used to use an evolutionary algorithm to simulate the treatment plan generated by the question-answer generation module that meets the answers to the patient's questions and optimize the treatment plan; in this embodiment, the evolutionary algorithm is a genetic algorithm;

[0064] In the evolutionary algorithm, multiple candidate treatment plans are first obtained from the question-answer generation module. During each optimization of the evolutionary algorithm, the evolutionary algorithm gradually adjusts variables in the candidate plans, such as drug dosage and lifestyle habits, by simulating patient feedback. The optimal treatment plan is selected through multiple iterations to maximize the accuracy of patient health management.

[0065] In the evolutionary algorithm, the population size represents the number of candidate treatment plans generated in each generation, which is usually set to 100; the crossover rate represents the probability of controlling the exchange of information between two treatment plans, which is usually set to 0.7; the mutation rate represents the probability of treatment plan mutation in each generation, which is usually between 0.01 and 0.05; the fitness function represents the quality of the treatment plan, which can be evaluated based on factors such as blood pressure control effect or drug side effects.

[0066] For example, if a patient reports experiencing mild dizziness after taking Drug A, the hypertension management system leverages the automated optimization mechanism of the evaluation and optimization module to consider reducing the dosage of Drug A and increasing the frequency of Drug B during the next optimization run. This dynamic optimization allows the treatment plan to adapt to changes in the patient's health status. During the optimization process, a genetic algorithm is used as a global search strategy to explore a broad parameter space and identify potential optimization paths. Meanwhile, Bayesian optimization plays a key role in the local search. By further fine-tuning the parameter space explored by the genetic algorithm, a Gaussian process regression model is used to predict different parameter combinations and evaluate the performance of the hypertension management system. "Parameters" refer to variables in the patient's personalized treatment plan, such as drug dosage and frequency. In Bayesian optimization, the acquisition function combines the results of the Gaussian process predictions with uncertainty information to select the current optimal parameter settings, gradually approaching the final optimal solution.

[0067] The expert feedback mechanism is used to obtain expert feedback text content for the treatment plan generated in the question and answer generation module that meets the patient's question answer, convert the expert feedback text content from natural language into machine language input through natural language processing (NLP), perform semantic analysis on the expert feedback text content converted into machine language input, extract medical guidance opinions, and then convert the medical guidance opinions into variable adjustment commands to optimize the treatment plan generated in the question and answer generation module that meets the patient's question answer.

[0068] For example, in practice, expert feedback might include, for example, "Reduce the dose of drug A," which means adjusting the recommended dose parameter of drug A (e.g., 10mg) to a lower value (e.g., 5mg); or "Increase the frequency of drug B," which means increasing the recommended daily dose of drug B from 1 to 2. These adjustment parameters are converted into variable adjustment commands and input into the answer generation module. The answer generation module then recalculates and generates an optimized treatment plan based on the expert's advice. For example, it might modify the drug recommendation weights based on the feedback, reducing the weight of drug A and increasing the weight of drug B. Simultaneously, using a reinforcement learning method based on transfer learning, the expert's adjustments influence the training process of the answer generation module. Specifically, the hypertension management system first performs data preprocessing to extract feature variables from the patient's real-time health data (e.g., blood pressure, heart rate) and historical records. This data is then combined with patient feedback (e.g., adverse drug reactions) and expert advice text to convert this data into a standardized input vector. A policy network is then constructed based on a reinforcement learning algorithm (such as a deep Q-network (DQN)). Its inputs are the patient's feature vector and the current treatment status, and its output is the predicted reward value for different treatment adjustments. The effectiveness of each adjustment is evaluated using real-time and simulated patient feedback (e.g., predicted blood pressure fluctuations), and this feedback is used to update the weights of the policy network. During this process, the system periodically incorporates expert advice, using the recommended parameters (e.g., adjustments to medication dosage or frequency) as a correction term in the reinforcement learning model's reward function, thereby accelerating the policy optimization process. Ultimately, through continuous learning, treatment plans can be adjusted in real time based on the patient's individual health status. For example, if a patient's blood pressure drops too quickly during a treatment, the dose of antihypertensive medication can be reduced to avoid side effects. To ensure the scientific and applicable nature of the recommendations, simulation models (e.g., Gaussian process-based regression models) are used before generating treatment recommendations to predict the adjusted treatment effects and assess potential risks. The optimized plan, after verification with patient feedback, enters another iterative optimization process, enabling dynamic optimization of hypertension management and providing patients with personalized and scientific treatment plans.

[0069] To ensure that the generated recommendations have practical application value, the Evaluation and Optimization module regularly evaluates the effectiveness of treatment plans. The automatic optimization mechanism continuously adjusts the details of the treatment plan through evolutionary algorithms, optimizing drug dosage, frequency, and lifestyle interventions. This process is further enhanced by the expert feedback mechanism, which allows experts to provide feedback based on the patient's actual situation. The Evaluation and Optimization module then fine-tunes the recommended plan based on this feedback. Furthermore, the system has the ability to continuously learn. Through transfer learning and reinforcement learning, it can continuously adjust and optimize health management strategies based on the patient's long-term monitoring data and new medical research results.

[0070] The hypertension management system also includes a complexity discrimination module, which is used to evaluate the complexity of the patient's condition through multi-dimensional weighted scoring. ; Complexity Score is the comprehensive complexity score, The first The weight of each factor (such as hypertension and diabetes, etc.), It is The value of a factor (e.g., whether the patient has hypertension: 1 if yes, 0 if no);

[0071] The comprehensive complexity score (Complexity Score) assesses the hypertension management system's ability to understand and adapt to patients' health status, personalized needs, and data sources to influence the output of the hypertension management system.

[0072] In the multimodal data fusion module, Transformer-XL processes the temporal features of vital sign data (such as blood pressure fluctuations) and combines it with a dual-stream convolutional neural network (Dual-Stream CNN) to analyze patient text data (such as lifestyle habits and symptoms), thereby improving the accuracy of personalized recommendations. Based on the comprehensive complexity score, the hypertension management system dynamically adjusts the Transformer-XL weights and the attention mechanism of the Dual-Stream CNN to focus on more important information in the data fusion.

[0073] In the medical terminology recognition module and knowledge graph construction module, pre-trained models (such as Med-BERT) and WordEmbedding technology are used to identify and construct semantic associations between terms. RoBERTa and graph neural networks (GNNs) are used to extract causal relationships between terms and construct the knowledge graph. When the overall complexity score (Complexity Score) exceeds a threshold, the hypertension management system conducts a deeper analysis of specific relationships in the knowledge graph to provide more targeted personalized recommendations. In the question-and-answer generation module, a large language model combines joint vectors and the knowledge graph to generate customized health management recommendations. When the overall complexity score (Complexity Score) exceeds a threshold, the hypertension management system generates more refined responses through the deep semantic understanding of the large language model and knowledge graph reasoning.

[0074] In the evaluation and optimization module, evolutionary algorithms and Bayesian optimization are used to dynamically adjust treatment plans to adapt to changing patient needs. Natural language processing (NLP) technology also transforms expert feedback into system optimization parameters. For plans with a Complexity Score above the threshold, the number of optimization iterations is increased and the acquisition function is adjusted to better suit the patient's health status. In terms of data preprocessing and standardization, data cleaning, missing value processing, and outlier detection are used to improve data quality. When the Complexity Score is above the threshold, the outlier detection algorithm is used more frequently during the preprocessing phase to ensure the accuracy of data analysis, thereby safeguarding the system's subsequent analysis results.

[0075] The hypertension management system further includes an emergency level determination module, which determines whether to issue an emergency alarm based on an emergency detection threshold: ; Wherein, Alert indicates whether to issue an emergency alert. When the value is 1, an emergency alert is issued, and when the value is 0, no emergency alert is issued. Blood pressure is the real-time blood pressure value, and threshold is the set emergency blood pressure threshold (for example, blood pressure greater than 180 / 120 mmHg is considered an emergency). The hypertension management system has a built-in emergency response mechanism. Once emergency symptoms related to hypertension (such as severe headache or chest pain) are detected, the patient and the medical team will be directly notified for timely intervention.

[0076] The hypertension management system uses a large language model for intelligent question-answering, generating precise health recommendations based on a patient's vital signs, text data, and knowledge graph. This combination not only improves the efficiency of health management but also enables dynamic adjustments based on real-time patient changes, ensuring each patient receives the most appropriate treatment plan.

[0077] The multimodal data fusion module effectively integrates multimodal information from different data sources, including vital sign data and text data. By employing a deep learning architecture combining Transformer-XL with a dual-stream convolutional neural network (Dual-StreamCNN), it can simultaneously process time-series vital sign data and natural language text data, providing patients with more comprehensive and accurate health assessments and management. This multimodal fusion approach empowers the hypertension management system with enhanced comprehensive capabilities in identifying health issues and capturing the complex underlying connections between different data sources.

[0078] By deeply integrating medical terminology recognition and relationship extraction, the Medical Term Recognition Module and the Knowledge Graph Construction Module utilize self-supervised learning models and graph neural network (GNN) technology to achieve efficient and accurate processing in medical term recognition and relationship extraction between terms. For example, when a patient describes their health condition, the Hypertension Management System can automatically identify and extract relevant medical terms and accurately identify the causal relationships between them, providing a solid foundation for subsequent health recommendations.

[0079] By incorporating evolutionary algorithms and expert feedback mechanisms, treatment plans can be optimized in real time based on patient feedback and health status. Specifically, automated optimization mechanisms dynamically adjust medication dosages and lifestyle recommendations to maximize treatment effectiveness, addressing the personalized needs of hypertensive patients. This dynamic optimization approach continuously improves the system's accuracy and adaptability, ensuring the ongoing effectiveness of treatment plans.

[0080] The Q&A generation module combines a large language model with a knowledge graph to generate highly customized health recommendations based on a patient's personalized health data. Compared to traditional health consultation systems, this intelligent Q&A based on a large language model not only understands patients' needs in real time but also deeply integrates personal health records and medical knowledge to provide patients with practical and actionable advice.

[0081] The hypertension management system comprehensively assesses a patient's health status, including blood pressure fluctuations, medication use, and lifestyle habits, and generates personalized health reports and treatment plans based on this information. Leveraging multimodal data fusion and deep learning technologies, the hypertension management system monitors changes in a patient's health in real time and adjusts treatment strategies accordingly, making health management more precise and personalized.

[0082] This solution combines a large-scale pre-trained language model, a medical knowledge graph, and multimodal data fusion technology to build a hypertension management system. The system efficiently integrates and analyzes a patient's vital signs, lifestyle, medical history, and other information to provide personalized, precise health management solutions, effectively supporting the intelligent management of chronic diseases. Through evolutionary algorithms and expert feedback mechanisms, the system continuously optimizes follow-up strategies, improves patient health management outcomes, reduces the burden on medical staff, and improves the efficiency, accuracy, and convenience of hypertension management, providing important technical support for innovation in chronic disease management.

[0083] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. It is apparent to those skilled in the art that various changes, modifications, substitutions, and variations to these embodiments may be made without departing from the principles and spirit of the present invention, and these changes and modifications still fall within the scope of protection of the present invention.

Claims

1. A hypertension management system, characterized in that: include: A data acquisition module is used to acquire vital sign data and text data; The data preprocessing module is used to clean and process the data acquired by the data acquisition module; The multimodal data melting module is used to extract time series features from vital sign data using Transformer-XL to obtain vital sign feature vectors; it is also used to extract features from text data using Dual-Stream CNN to obtain text feature vectors; and it is also used to integrate the vital sign feature vectors and text feature vectors into a joint vector; In the multimodal data melting module, the text data is converted into a word vector sequence through the embedding layer of the Dual-Stream CNN, and the core content of the text data is extracted through the convolution layer of the semantic stream and the syntactic stream. The formula is: ; ;in represents the output of the semantic flow, represents the output of the syntactic stream, Represents the input text data, represents the weight matrix of the semantic stream convolution layer, represents the weight matrix of the convolutional layer of the syntactic stream, and Used to control the feature extraction method, represents the bias term of the semantic stream convolution layer, represents the bias term of the convolutional layer of the syntactic stream, and Used to adjust the output of convolution calculation; The text feature vector ,in, Used to control the weight ratio of semantic features, Used to control the weight ratio of syntactic features; The question-answer generation module is used to obtain the joint vector through a large language model to understand the patient's question, and then extract the corresponding information from the pre-built knowledge graph and generate an answer that meets the patient's question.

2. A hypertension management system according to claim 1, characterized in that: The knowledge graph is obtained by executing the medical term recognition module and the knowledge graph construction module; The medical term recognition module is used to identify medical terms in text data and extract the association relationships between different medical terms; the knowledge graph construction module is used to construct a knowledge graph based on the association relationships between different medical terms.

3. A hypertension management system according to claim 2, characterized in that: The medical term recognition module uses the context-aware embedding generated by RoBERTa as the basis for node embedding of the graph neural network (GNN), combined with the structured message passing of the graph neural network (GNN) to extract the semantic relationships between different medical terms.

4. A hypertension management system according to claim 3, characterized in that: In the RoBERTa, the input layer of RoBERTa captures the contextual semantic information of medical terms in their corresponding text data as a context-aware embedding representation of the medical terms; In the graph neural network (GNN), each medical term is regarded as a node in the graph, and the node embedding is updated through the message passing mechanism of the graph neural network (GNN) itself, so that the representation of each node integrates its own contextual information and the semantic relationship of neighboring nodes to enhance the understanding of the association between nodes.

5. A hypertension management system according to claim 4, characterized in that: In the RoBERTa, medical terminology After processing by RoBERTa, medical terminology is generated Context-aware embedding representation of ; In the graph neural network GNN, medical terminology The corresponding embedding representation of node v The iterative update formula at the kth layer is ;in, is the weight matrix of the kth layer, is the set of neighbor nodes of node v, is the activation function.

6. A hypertension management system according to claim 5, characterized in that: In the multimodal data melting module, the calculation formula of the joint vector H is: ;in Indicates that the vital signs data are processed by Transformer-XL The extracted physical sign feature vector, Indicates that text data is processed by Dual-Stream CNN The extracted text feature vector.

7. A hypertension management system according to claim 6, characterized in that: For the vital sign data, the data preprocessing module is further used to analyze the outliers in the vital sign data through Z-score and remove the outliers; The formula for Z-score analysis is: ,in is the vital sign data to be analyzed, is the mean of the data set where the physical sign data are located, is the standard deviation of the data set where the vital signs data are located, express and The distance between The physical sign data whose absolute value is greater than the preset analysis threshold is regarded as an abnormal value.

8. A hypertension management system according to claim 7, characterized in that: The hypertension management system further comprises an evaluation and optimization module, which comprises an automatic optimization mechanism and an expert feedback mechanism; The automatic optimization mechanism is used to use an evolutionary algorithm to simulate the treatment plan generated by the question-answer generation module that meets the answers to the patient's questions and optimize the treatment plan; In the evolutionary algorithm, multiple candidate treatment plans are first obtained from the question-answer generation module. During each optimization of the evolutionary algorithm, the evolutionary algorithm gradually adjusts the variables in the candidate plans by simulating patient feedback, and selects the optimal treatment plan through multiple iterations. In the evolutionary algorithm, the population size represents the number of candidate treatment plans generated in each generation, the crossover rate represents the probability of controlling the exchange of information between two treatment plans, the mutation rate represents the probability of treatment plan mutation in each generation, and the fitness function represents the evaluation of the quality of the treatment plan. The expert feedback mechanism is used to obtain expert feedback text content for the treatment plan generated in the question and answer generation module that meets the answers to the patient's questions, convert the expert feedback text content from natural language into machine language input through natural language processing, perform semantic analysis on the expert feedback text content converted into machine language input, extract medical guidance opinions, and then convert the medical guidance opinions into variable adjustment commands to optimize the treatment plan generated in the question and answer generation module that meets the answers to the patient's questions.

9. A hypertension management system according to claim 8, characterized in that: The hypertension management system also includes a complexity discrimination module, which is used to evaluate the complexity of the patient's condition through multi-dimensional weighted scoring. ; Complexity Score is the comprehensive complexity score, The first The weight of the factors, It is The value of a factor.

Images