A chronic disease medical and preventive integrated collaborative management and control system

Abstract

The present invention discloses a chronic disease medical and preventive integrated collaborative management and control system, which belongs to the technical field of health data processing. It solves the problem that when the existing system extracts the user's electronic medical record, it cannot reflect the patient's health status changes in real time, resulting in inaccurate extraction of key features. The system includes a user terminal, a fusion cloud platform, a medical and preventive judgment module, and a follow-up customization module. The fusion cloud platform is used to take a normalized data set associated with the user as input, execute a risk assessment model, and output a risk assessment result. The present invention constructs a risk assessment model based on deep learning combined with an attention mechanism through the fusion cloud platform, so that the risk assessment model can identify and analyze the normalized data set collected by the user terminal, and the user terminal can collect multi-source heterogeneous normalized data sets with good timeliness and real-time reflection of the patient's health status, thereby ensuring that the risk assessment model provides an accurate assessment path for medical prevention and control of chronic diseases.

Description

Technical Field

[0001] The present invention belongs to the technical field of health data processing, and specifically relates to a chronic disease medical and preventive integrated collaborative management and control system. Background Art

[0002] Chronic diseases are short for chronic non-communicable diseases, which are a general term for a class of diseases with long course, persistent symptoms and complex causes. Common chronic diseases include cardiovascular and cerebrovascular diseases, diabetes, and chronic respiratory diseases. Due to the particularity of chronic diseases, it is generally recognized that the most effective means of treating chronic diseases are high-risk population screening, lifestyle intervention, and regular follow-up and monitoring.

[0003] At present, grassroots medical and health institutions have long been facing pain points such as lack of high-quality resources, imperfect management system, low patient compliance, and limited technology. At the same time, the phenomenon of "three highs (morbidity, mortality, and disability rate) and three lows (awareness rate, control rate, and treatment rate)" among patients with two chronic diseases is becoming increasingly serious. With the development of information and communication technology, "Internet + Medical Health" has become an indispensable part of the contemporary health service system. The construction of an integrated hospital-community-family medical and prevention management and control platform can provide individualized continuing care services for patients with chronic diseases such as diabetes, hypertension, cardiovascular disease, and stroke.

[0004] Chinese patent CN115881285A discloses a three-level collaborative medical and preventive integrated data management system and management method for chronic disease management. The method steps are as follows: obtain user condition information, record the condition information on the patient data management page, and upload it to the user information database; extract key features of the user condition information, output recognition results, and generate user electronic medical records from the user condition information; classify the user electronic medical records according to the disease group type based on the obtained recognition results; establish a collaborative information database, and upload the information to the collaborative information database; however, when the key feature extraction module of the existing system extracts the user electronic medical record files, the user electronic medical record files have poor timeliness and cannot reflect the changes in the patient's health status in real time, and ignores the patient's personal living habits, environmental factors and other personalized information, resulting in inaccurate extraction of key features, which in turn affects the assessment and prediction accuracy of chronic diseases. To address the above problems, we propose a chronic disease medical and preventive integrated collaborative management and control system. Summary of the invention

[0005] The purpose of the present invention is to address the shortcomings of the prior art and provide a chronic disease medical and preventive integrated collaborative management and control system, which solves the problem that when the existing system extracts user electronic medical records, the user's electronic medical records have poor timeliness and cannot reflect the changes in the patient's health status in real time, and ignores the patient's personal living habits, environmental factors and other personalized information, resulting in inaccurate extraction of key features, which in turn affects the assessment and prediction accuracy of chronic diseases.

[0006] The present invention is implemented as follows: a chronic disease medical and preventive integrated collaborative management and control system, the chronic disease medical and preventive integrated collaborative management and control system comprising:

[0007] The user end is used to collect multi-source heterogeneous user-related information and upload the user-related information to the fusion cloud platform in real time. The user-related information includes user basic health information, health follow-up information, related family doctor information, chronic disease control information, user medical record information, and user vital sign monitoring data;

[0008] Fusion cloud platform, load user-related information, process user-related information based on edge computing, obtain normalized data set, import normalized data set into pre-built disease-specific database, pre-build risk assessment model based on deep learning combined with attention mechanism, use normalized data set associated with user as input, execute risk assessment model, and output risk assessment result;

[0009] The medical prevention judgment module obtains the risk assessment results and determines whether the risk assessment results exceed the preset risk threshold. If so, the risk reminder instruction is triggered and the risk assessment results are shared and collaboratively pushed to the user end. If not, the risk assessment results are pushed to the follow-up customization module.

[0010] The follow-up customization module responds to the risk assessment results, traverses the disease-specific database, indexes the follow-up survey information associated with the user, combines the risk assessment results and follow-up survey information to generate an intelligent follow-up plan for chronic diseases, and pushes the intelligent follow-up plan to the user end.

[0011] Preferably, the user terminal includes:

[0012] The health monitoring unit is used to collect the user's vital signs monitoring data in a distributed manner and upload the user's vital signs monitoring data to the data acquisition unit in real time;

[0013] A follow-up management unit, in response to the intelligent follow-up plan pushed by the follow-up customization module, identifies the intelligent follow-up plan, triggers personalized follow-up instructions based on the intelligent follow-up plan, obtains and parses health follow-up information, and pushes the health follow-up information to the data acquisition unit;

[0014] The data acquisition unit is connected to the health monitoring unit, the follow-up management unit, and the medical institution information system for communication, and is used to acquire multi-source heterogeneous user association information from the health monitoring unit, the follow-up management unit, and the medical institution information system.

[0015] Preferably, the method for acquiring and parsing health follow-up information comprises:

[0016] Obtain health follow-up information, retrieve the user's basic health information from the disease-specific database, verify the user's identity in the health follow-up information based on the user's basic health information, and determine whether the user's identity authentication is passed;

[0017] If the user identity verification is passed, the semantic understanding instruction is triggered, the health follow-up information is loaded, and a multi-channel band-stop filter is used to extract the non-voice signal in the health follow-up information;

[0018] Performing short-time Fourier transform on the non-speech signal to obtain a non-speech signal spectrum sequence, selecting the non-speech signal center frequency based on the non-speech signal spectrum sequence, and using a Fourier transform algorithm to calculate the interference amplitude energy and interference phase of the non-speech signal spectrum sequence;

[0019] The interference amplitude energy and interference phase of the non-speech signal spectrum sequence are calculated by the following formula:

[0020] (1)

[0021] (2)

[0022] (3)

[0023] in, Represents the interference amplitude energy of the non-speech signal spectrum sequence, is the interference phase of the non-speech signal spectrum sequence, is the number of discretized signal points of the non-speech signal spectrum sequence, is the input representation of the non-speech signal spectrum sequence, is the center frequency of the non-speech signal, represents the cutoff frequency of the band-stop filter, Indicates the transformation result of the non-speech signal spectrum sequence based on the Fourier transform algorithm;

[0024] Obtain interference amplitude energy and interference phase of the non-speech signal spectrum sequence, use a band-stop filter to perform L-order harmonic suppression and attenuation on the non-speech signal spectrum sequence, eliminate interference noise information in the health follow-up information, and output health follow-up information containing speech characteristic frequencies;

[0025] The interference noise information is expressed as:

[0026] (4)

[0027] (5)

[0028] in, Indicates interference noise information, is the stop-band attenuation of the band-stop filter, Indicates the number of harmonics, is the insertion loss of the band-stop filter, are the sampling angle differences of the band-stop filter for the positive and negative frequency components of the harmonics, is the speech characteristic frequency;

[0029] Acquire health follow-up information containing speech feature frequencies, understand the semantics of the health follow-up information based on a support vector machine combined with an attention mechanism, output the semantically understood health follow-up information, and automatically collect the health follow-up information into a data acquisition unit.

[0030] Preferably, the method for semantic understanding of health follow-up information based on a support vector machine combined with an attention mechanism comprises:

[0031] Loading health follow-up information, dividing the health follow-up information into characters, words, and phrases, and identifying the speech feature frequencies corresponding to the characters, words, and phrases;

[0032] Based on the polynomial of the support vector machine, the speech feature frequency features corresponding to the characters, words and phrases are fused to obtain the fused feature vector;

[0033] Using the string kernel function of the support vector machine to index at least one set of library semantic vectors in the disease-specific database that have a high correlation with the fused feature vector;

[0034] The correlation between the fusion feature vector and the library semantic vector is calculated by the following formula:

[0035] (6)

[0036] The correlation between the fused feature vector and the library semantic vector is corrected by correcting the cosine similarity to obtain the similarity between the fused feature vector and the library semantic vector;

[0037] (7)

[0038] in, Indicates the similarity between the fusion feature vector and the library semantic vector, Represents the fused feature vector Semantic vector with library Relevance, Represent the fusion feature vectors Collection, library semantic vector A collection of is the string kernel function output representation of the fused feature vector, is the weight of the library semantic vector;

[0039] Load the similarity between the fused feature vector and the library semantic vector, and determine whether the similarity between the fused feature vector and the library semantic vector meets the preset similarity threshold. If the similarity threshold is met, select the library semantic vector with the highest similarity to replace the fused feature vector;

[0040] Load at least one set of library semantic vectors, perform weighted combination analysis on the library semantic vectors based on the weights of the library semantic vectors to obtain a follow-up score matrix, use a spectral clustering algorithm to classify the follow-up score matrix, integrate at least one set of library semantic vectors, obtain a library semantic vector set and a follow-up score classification associated with the library semantic vector set, and complete the semantic understanding of health follow-up information.

[0041] Preferably, the fusion cloud platform includes:

[0042] A fusion processing unit, which is used to load user association information, fuse and process the user association information based on edge computing to obtain a normalized data set, and import the normalized data set into a pre-built disease-specific database. The fusion processing unit includes a K-means clustering block, an NLP processing block, an AutoML learning block, a knowledge graph, and an Upgrade block.

[0043] The disease-specific database is used to load the normalized data set, associate the data in the normalized data set with the user based on the master data management (MDM) technology, generate a unique index code for the user, store and visualize the normalized data set;

[0044] The risk management and control unit pre-builds a risk assessment model based on deep learning combined with the attention mechanism, takes the normalized data set associated with the user as input, executes the risk assessment model, and outputs the risk assessment results.

[0045] Preferably, the method for processing user association information based on edge computing fusion includes:

[0046] Load user association information, process abnormal values ​​and missing values ​​of user association information, and obtain a noise information elimination set;

[0047] The NLP processing block is used to convert the data format, and the interference information elimination set is converted into a vector using one-hot encoding to obtain the associated information vector set;

[0048] Load the associated information vector set, and the AutoML learning block performs dimensionality reduction processing on the associated information vector set based on edge computing to obtain an information dimensionality reduction set;

[0049] Identify the one-hot encoding associated with the information reduction set, map the information reduction set to the Euclidean space through the one-hot encoding, integrate at least one set of Euclidean space mapping results to obtain a normalized data set, and expand the values ​​of the discrete features of the normalized data set to the knowledge graph.

[0050] Preferably, the method for pre-building a risk assessment model based on deep learning combined with an attention mechanism comprises:

[0051] The convolutional neural network is used as the initial model of the risk assessment model, and the initial model includes an input layer, a convolution module, a hidden layer, and an output layer;

[0052] The normalized data set in the disease-specific database is indexed based on the unique index code, and the normalized data set is divided into a training set and a test set;

[0053] Initialize the hyperparameters and loss function of the initial model, and set the initial model iterative training rounds and optimizer;

[0054] Obtain a training set and use it to iteratively train the initial model. During training, use a sensitivity analysis algorithm to evaluate the influencing features that are positively correlated with the risk assessment results, prune and compress the weights of influencing features with low sensitivity, adjust the hyperparameters of the initial model based on the sensitivity coefficients of the influencing features, and iterate the training until the initial model converges;

[0055] Load the test set, use the test set as input, execute the initial model, output the test results, and judge whether the test results are qualified based on the preset accuracy threshold. If the test results are qualified, output the converged risk assessment model.

[0056] Preferably, in the initial model, the input layer is connected to the convolution module, the convolution module is connected to the hidden layer, the hidden layer is connected to the output layer, the hidden layer is provided with three layers, the hidden layer nodes are 32, 64 and 128 respectively, and the hidden layer activation function is the Act_Logistic activation function;

[0057] The competitive adaptive weighted sampling method and channel attention mechanism module are introduced into the input layer to improve the input layer. The generative adversarial network is introduced into the convolution module. The convolution module consists of the first convolution layer, the second convolution layer, and the third convolution layer. The third convolution layer is frozen, and a multi-layer perceptron with feature parameter sharing is introduced between the second and third convolution layers. The global average pooling layer is used to replace the third convolution layer.

[0058] Preferably, the method for executing the risk assessment model and outputting the risk assessment results specifically includes:

[0059] The normalized data set associated with the user in the index disease database is used, and the input layer determines the chronic disease type of the user associated with the normalized data set based on the competitive adaptive weighted sampling method;

[0060] Based on the channel attention mechanism module, the normalized data set features are extracted through the maximum pooling operation combined with the average pooling operation to obtain the feature factor set;

[0061] Load the characteristic factor set and chronic disease types, set the weights of the influencing features associated with the chronic disease types based on the Delphi method, extract the influencing features associated with the chronic disease types, and calculate the sensitivity coefficient using the sensitivity analysis algorithm;

[0062] Determine whether the sensitivity coefficient of the influencing feature exceeds a preset sensitivity threshold. If it exceeds the preset sensitivity threshold, retain the current influencing feature, integrate at least one group of influencing features, and obtain a sensitive feature set;

[0063] Load the sensitive feature set. The convolution module performs convolution dimensionality reduction operations on the influencing features in the sensitive feature set through the first 7×7 convolution layer and the second convolution layer. The multi-layer perceptron performs element-by-element weighted fusion on the influencing features and outputs the feature fusion set.

[0064] The hidden layer uses the follow-up score classification as prior knowledge, combines the genetic algorithm to globally search the feature fusion set, and calculates the risk assessment result through the Act_Logistic activation function.

[0065] Preferably, the sensitivity analysis algorithm is used to calculate the sensitivity coefficient using the following formula:

[0066] (8)

[0067] in, Indicates the sensitivity coefficient of the current influencing feature, represents the sensitivity constant of the sensitivity analysis algorithm, is the weight of the current influencing feature, Represents the average weight of the feature factor set that affects the feature;

[0068] The risk assessment result is calculated using the following formula:

[0069] (9)

[0070] in, represents the risk assessment level, Indicates the sensitivity coefficient of the current influencing feature, is the weight of the current influencing feature, and The classification level of follow-up score is represents the input representation of the feature fusion set, is the bias term of the Act_Logistic activation function, It is the number of iterative searches of the genetic algorithm.

[0071] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0072] In the embodiment of the present invention, a risk assessment model based on deep learning combined with an attention mechanism is constructed by integrating a cloud platform, so that the risk assessment model can identify and analyze the normalized data set collected by the user end, and the user end can collect multi-source heterogeneous normalized data sets with good timeliness and real-time reflection of the patient's health status, thereby ensuring that the risk assessment model provides an accurate assessment path for chronic disease medical prevention and control, and overcomes the problem that when the key feature extraction module of the existing system extracts the user's electronic medical case files, the user's electronic medical case files have poor timeliness and cannot reflect the changes in the patient's health status in real time, and ignores the patient's personal living habits, environmental factors and other personalized information, resulting in inaccurate extraction of key features, which in turn affects the assessment and prediction accuracy of chronic diseases.

[0073] In the embodiment of the present invention, a multi-channel band-stop filter is used to extract non-speech signals in the health follow-up information, and a short-time Fourier transform is performed on the non-speech signal, so that interference or noise within a specific frequency range can be effectively removed or suppressed, thereby overcoming the problem that most chronic disease patients are elderly people, there is a lot of communication noise during telephone follow-up, and the elderly speak non-standard Mandarin.

[0074] In the embodiment of the present invention, the semantic understanding of health follow-up information is achieved through the support vector machine combined with the attention mechanism, so that it has significant robustness to noise and outliers. The attention mechanism can further enhance this robustness by ignoring unimportant information, and the spectral clustering algorithm can be used to classify the follow-up score matrix, thereby improving the flexibility of follow-up information collection. The follow-up score classification can also provide prior knowledge for the risk assessment of the risk assessment model, further ensuring the assessment accuracy of the model, forming a full-scale integration of "data flow + business flow + resource flow", and solving a series of difficult problems such as the high proportion of chronic disease drugs in the elderly, low medical efficiency, and weak patient compliance.

[0075] In the embodiment of the present invention, unique-hot encoding is used to perform vector conversion on the interference information elimination set, and the dimension reduction processing of the associated information vector set is performed based on edge computing, thereby realizing the normalization processing of multi-source heterogeneous user associated information and avoiding the loss of feature information. It not only improves the efficiency and accuracy of data processing, but also enhances the consistency and comparability of data, providing a solid foundation for subsequent risk assessment model data analysis and machine learning tasks.

[0076] In an embodiment of the present invention, a risk assessment model is provided. The risk assessment model uses a convolutional neural network as the initial model, and introduces a competitive adaptive weighted sampling method, a channel attention mechanism, a generative adversarial network and a multi-layer perceptron. The channel attention mechanism is used to effectively extract the influencing features associated with different chronic diseases, so as to realize the evaluation and prediction of different chronic diseases, and overcome the limitation that the prior art can only predict and evaluate a single or two chronic diseases. At the same time, the risk assessment model is combined with a multi-source heterogeneous normalized data set to obtain highly timely and highly correlated features associated with chronic diseases from different dimensions and types of data, thereby reducing the loss of timeliness feature information, and realizing accurate assessment of the user's chronic disease risk while less increasing the complexity of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 It is a structural schematic diagram of the chronic disease medical and preventive integrated collaborative management and control system provided by the present invention.

[0078] Figure 2 A schematic diagram of the implementation process of the method for acquiring and parsing health follow-up information is shown.

[0079] Figure 3 The figure shows a schematic diagram of the implementation process of the method for semantic understanding of health follow-up information based on support vector machine combined with attention mechanism.

[0080] Figure 4 A schematic diagram of the implementation process of the method for processing user association information based on edge computing fusion is shown.

[0081] Figure 5 A schematic diagram of the implementation process of a method for pre-building a risk assessment model based on deep learning combined with an attention mechanism is shown.

[0082] Figure 6 A schematic diagram of the implementation process of executing a risk assessment model and outputting a risk assessment result method is shown.

[0083] In the figure: 100-user end, 110-health monitoring unit, 120-follow-up management unit, 130-data acquisition unit, 200-fusion cloud platform, 210-fusion processing unit, 220-special disease database, 230-risk management unit, 300-medical and prevention judgment module, 400-follow-up customization module, 410-follow-up information index unit, 420-plan customization unit. DETAILED DESCRIPTION

[0084] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of this application; the terms used in the specification of the application herein are only for the purpose of describing specific embodiments and are not intended to limit this application; the terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0085] When the key feature extraction module of the existing system extracts the user's electronic medical records, the user's electronic medical records have poor timeliness and cannot reflect the changes in the patient's health status in real time. It also ignores the patient's personal living habits, environmental factors and other personalized information, resulting in inaccurate extraction of key features, which in turn affects the assessment and prediction accuracy of chronic diseases. To address the above problems, we propose a chronic disease medical and preventive integrated collaborative management and control system. In short, the system consists of a user terminal 100, a fusion cloud platform 200, a medical and preventive judgment module 300, and a follow-up customization module 400. When the system is working, the user terminal 100 collects multi-source heterogeneous user-related information, and the fusion cloud platform 20 0 Based on edge computing fusion processing of user-related information, a normalized data set is obtained, the normalized data set is imported into the pre-built special disease database 220, and the normalized data set associated with the user is used as input to execute the risk assessment model and output the risk assessment result. Then the medical and preventive judgment module 300 determines whether the risk assessment result exceeds the preset risk threshold. If it does not exceed the preset risk threshold, the risk assessment result is pushed to the follow-up customization module 400. The follow-up customization module 400, in response to the risk assessment result, traverses the special disease database 220, indexes the follow-up survey information associated with the user, and generates a chronic disease intelligent follow-up plan in combination with the risk assessment result and the follow-up survey information. In the embodiment of the present invention, a risk assessment model based on deep learning combined with an attention mechanism is constructed by integrating the cloud platform 200, so that the risk assessment model can identify and analyze the normalized data set collected by the user terminal 100, and the user terminal 100 can collect multi-source heterogeneous normalized data sets with good timeliness and real-time reflection of the patient's health status, thereby ensuring that the risk assessment model provides an accurate assessment path for chronic disease medical prevention and control, and overcomes the problem that when the key feature extraction module of the existing system extracts the user's electronic medical case files, the user's electronic medical case files have poor timeliness and cannot reflect the patient's health status changes in real time, and ignores the patient's personal living habits, environmental factors and other personalized information, resulting in inaccurate extraction of key features, which in turn affects the assessment and prediction accuracy of chronic diseases.

[0086] The embodiment of the present invention provides a chronic disease medical and preventive integrated collaborative management and control system. Figure 1A schematic diagram of a chronic disease medical and preventive integrated collaborative management and control system is shown, and the chronic disease medical and preventive integrated collaborative management and control system specifically includes:

[0087] The user terminal 100 is used to collect multi-source heterogeneous user association information and upload the user association information to the fusion cloud platform 200 in real time.

[0088] It should be noted that the user-related information collected by the user terminal 100 includes but is not limited to the user's basic health information, health follow-up information, related family doctor information, chronic disease control information, user medical record information, and user vital sign monitoring data;

[0089] In this embodiment, the chronic diseases include but are not limited to hypertension, dyslipidemia, hyperglycemia, diabetes, coronary heart disease, atherosclerosis, chronic respiratory diseases, neurological diseases, mental and psychological diseases, etc. The basic health information of the user includes the user's name, age, gender, marital status, height, weight, blood type, education level, pre-retirement occupation, medical insurance type, chronic disease history, contact information, home address, and chronic disease control information includes but is not limited to medication compliance (regular, irregular), emotional management (good, poor), less oil, less salt, less sugar (yes, no), weight control, regular exercise (yes, no), knowledge awareness rate. The user's medical record information is obtained from the medical institution information system, and the medical institution information includes but is not limited to central hospitals, community grassroots hospitals, and medical research institutions.

[0090] In this embodiment, the user's vital sign monitoring data is collected by a vital sign collection device, which includes a sphygmomanometer, a thermometer, a wearable bracelet, a blood glucose tester, and a smart camera. The user's vital sign monitoring data includes but is not limited to blood pressure, body temperature, heart rate, blood oxygen saturation, blood glucose, urine volume, insulin, C-peptide level, expression video / image, and activity level.

[0091] The fusion cloud platform 200 loads the user-related information, processes the user-related information based on edge computing fusion, obtains a normalized data set, imports the normalized data set into the pre-built disease-specific database 220, pre-builds a risk assessment model based on deep learning combined with an attention mechanism, takes the normalized data set associated with the user as input, executes the risk assessment model, and outputs a risk assessment result;

[0092] The medical and preventive judgment module 300 obtains the risk assessment result, determines whether the risk assessment result exceeds the preset risk threshold, and if so, triggers the risk reminder instruction, and pushes the risk assessment result to the user terminal 100 for sharing and collaboration; if not, pushes the risk assessment result to the follow-up customization module 400; in this embodiment, the risk threshold can be set to 0.6-0.8.

[0093] The follow-up customization module 400 , in response to the risk assessment result, traverses the disease-specific database 220 , indexes the follow-up investigation information associated with the user, generates a smart follow-up plan for chronic diseases based on the risk assessment result and the follow-up investigation information, and pushes the smart follow-up plan to the user terminal 100 .

[0094] In this embodiment, the follow-up customization module 400 includes:

[0095] A follow-up information indexing unit 410, wherein the follow-up information indexing unit 410 traverses the disease-specific database 220 in response to the risk assessment result, and indexes the follow-up survey information associated with the user. Medical personnel or family doctors can set follow-up terms, follow-up time, and follow-up precautions for the user through the disease-specific database 220. The follow-up information indexing unit 410 imports the follow-up terms, follow-up time, and follow-up precautions into the follow-up form template to generate a personalized follow-up form;

[0096] The plan customization unit 420 is used to obtain a personalized follow-up form, and customize the follow-up period for different users in combination with the risk assessment value of the risk assessment result. In this embodiment, for users with a risk assessment value ≤ 0.1, the follow-up period can be 90 days, for users with a risk assessment value of 0.1 < ≤ 0.3, the follow-up period can be 60 days, for users with a risk assessment value of 0.3 < ≤ 0.5, the follow-up period can be 30 days, and for users with a risk assessment value of 0.5 < ≤ 0.6, the follow-up period can be 10 days.

[0097] It should be noted that, in this embodiment, a local area network or 5G communication connection is adopted between the user terminal 100, the integrated cloud platform 200, the medical and prevention judgment module 300, and the follow-up customization module 400 to realize the interactive transmission of data.

[0098] In the embodiment of the present invention, a risk assessment model based on deep learning combined with an attention mechanism is constructed by integrating the cloud platform 200, so that the risk assessment model can identify and analyze the normalized data set collected by the user terminal 100, and the user terminal 100 can collect multi-source heterogeneous normalized data sets with good timeliness and real-time reflection of the patient's health status, thereby ensuring that the risk assessment model provides an accurate assessment path for chronic disease medical prevention and control, and overcomes the problem that when the key feature extraction module of the existing system extracts the user's electronic medical case files, the user's electronic medical case files have poor timeliness and cannot reflect the patient's health status changes in real time, and ignores the patient's personal living habits, environmental factors and other personalized information, resulting in inaccurate extraction of key features, which in turn affects the assessment and prediction accuracy of chronic diseases.

[0099] In this embodiment, the user terminal 100 includes:

[0100] The health monitoring unit 110 is used for collecting the user's vital signs monitoring data in a distributed manner and uploading the user's vital signs monitoring data to the data acquisition unit 130 in real time;

[0101] The follow-up management unit 120, in response to the intelligent follow-up plan pushed by the follow-up customization module 400, identifies the intelligent follow-up plan, triggers personalized follow-up instructions based on the intelligent follow-up plan, obtains and analyzes health follow-up information, and pushes the health follow-up information to the data acquisition unit 130;

[0102] The data acquisition unit 130 is connected to the health monitoring unit 110 , the follow-up management unit 120 , and the medical institution information system for communication, and is used to acquire multi-source heterogeneous user association information from the health monitoring unit 110 , the follow-up management unit 120 , and the medical institution information system.

[0103] It should be noted that the data acquisition unit 130 can store data collected by the health monitoring unit 110 and the follow-up management unit 120, and the data acquisition unit 130 and the health monitoring unit 110 and the follow-up management unit 120 can use Bluetooth, WiFi, Zigbee, general packet radio service (GPRS) and other technical communications to achieve interactive data transmission.

[0104] In the embodiment of the present invention, the follow-up management unit 120 implements batch outbound calling, voice recognition, and follow-up scoring for users based on an intelligent outbound calling robot, thereby improving the timeliness of data and significantly reducing the workload of medical staff.

[0105] Most chronic disease patients are elderly people. During telephone follow-up, there are many communication noises and difficulties in the elderly's non-standard Mandarin. In order to solve the above problems, the embodiment of the present invention provides a method for obtaining and parsing health follow-up information. Figure 2 A schematic diagram of the implementation process of the method for obtaining and parsing health follow-up information is shown. The method for obtaining and parsing health follow-up information specifically includes:

[0106] Step S101, obtaining health follow-up information, retrieving basic health information of the user in the disease-specific database 220, and verifying the identity of the user in the health follow-up information based on the basic health information of the user;

[0107] It should be noted that when verifying the user's identity in the health follow-up information based on the user's basic health information, voice recognition, the last four digits of the ID number, facial recognition, and fixed password verification can be used.

[0108] Step S102, determining whether the user identity verification is passed;

[0109] Step S103, if the user identity verification is passed, a semantic understanding instruction is triggered, the health follow-up information is loaded, and a multi-channel band-stop filter is used to extract non-voice signals in the health follow-up information;

[0110] Step S104: If the user identity verification fails three times, the health follow-up process is terminated.

[0111] Step S105, performing short-time Fourier transform on the non-speech signal to obtain a non-speech signal spectrum sequence, selecting a non-speech signal center frequency based on the non-speech signal spectrum sequence, and using a Fourier transform algorithm to calculate interference amplitude energy and interference phase of the non-speech signal spectrum sequence;

[0112] The interference amplitude energy and interference phase of the non-speech signal spectrum sequence are calculated by the following formula:

[0113] (1)

[0114] (2)

[0115] (3)

[0116] in, Represents the interference amplitude energy of the non-speech signal spectrum sequence, is the interference phase of the non-speech signal spectrum sequence, is the number of discretized signal points of the non-speech signal spectrum sequence, is the input representation of the non-speech signal spectrum sequence, is the center frequency of the non-speech signal, represents the cutoff frequency of the band-stop filter, Indicates the transformation result of the non-speech signal spectrum sequence based on the Fourier transform algorithm;

[0117] Step S106, obtaining interference amplitude energy and interference phase of the non-speech signal spectrum sequence, using a band-stop filter to perform L-order harmonic suppression and attenuation on the non-speech signal spectrum sequence, eliminating interference noise information in the health follow-up information, and outputting the health follow-up information containing the speech characteristic frequency;

[0118] The interference noise information is expressed as:

[0119] (4)

[0120] (5)

[0121] in, Indicates interference noise information, is the stop-band attenuation of the band-stop filter, Indicates the number of harmonics, is the insertion loss of the band-stop filter, are the sampling angle differences of the band-stop filter for the positive and negative frequency components of the harmonics, is the speech characteristic frequency, and L represents the number of times the band-stop filter filters the non-speech signal spectrum sequence. In this embodiment, the number may be 10-15 times;

[0122] It should be noted that the insertion loss of the band-stop filter is expressed in decibels (dB). The smaller the insertion loss, the better the performance of the filter, and the cut-off frequency is used to describe the frequency point at which the filter performance drops to a certain level. In this embodiment, the band-stop filter can be a Butterworth band-stop filter.

[0123] Step S107, obtaining health follow-up information containing speech feature frequencies, semantically understanding the health follow-up information based on a support vector machine combined with an attention mechanism, outputting the semantically understood health follow-up information, and automatically collecting the health follow-up information into the data acquisition unit 130.

[0124] In this embodiment, a band-stop filter is used to eliminate noise in the call speech. As an effective denoising tool, the band-stop filter plays an important role in the semantic recognition of voice calls. This application can significantly improve the quality of the voice signal and thus improve the accuracy of semantic recognition by reasonably designing and adjusting the band-stop filter.

[0125] In the embodiment of the present invention, a multi-channel band-stop filter is used to extract non-speech signals in the health follow-up information, and a short-time Fourier transform is performed on the non-speech signal, so that interference or noise within a specific frequency range can be effectively removed or suppressed, thereby overcoming the problem that most chronic disease patients are elderly people, there is a lot of communication noise during telephone follow-up, and the elderly speak non-standard Mandarin.

[0126] The embodiment of the present invention provides a method for semantic understanding of health follow-up information based on support vector machine combined with attention mechanism. Figure 3 The schematic diagram of the implementation process of the method for semantic understanding of health follow-up information based on support vector machine combined with attention mechanism is shown. The method for semantic understanding of health follow-up information based on support vector machine combined with attention mechanism specifically includes:

[0127] Step S1071, loading health follow-up information, dividing the health follow-up information into characters, words, and phrases, and identifying speech feature frequencies corresponding to the characters, words, and phrases;

[0128] Step S1072, based on the polynomial of the support vector machine, the speech feature frequency features corresponding to the characters, words, and phrases are checked and fused to obtain a fused feature vector;

[0129] Step S1073, using the string kernel function of the support vector machine to index at least one set of library semantic vectors in the disease-specific database 220 that have a high correlation with the fused feature vector;

[0130] The correlation between the fusion feature vector and the library semantic vector is calculated by the following formula:

[0131] (6)

[0132] The correlation between the fused feature vector and the library semantic vector is corrected by correcting the cosine similarity to obtain the similarity between the fused feature vector and the library semantic vector;

[0133] (7)

[0134] in, Indicates the similarity between the fusion feature vector and the library semantic vector, Represents the fused feature vector Semantic vector with library Relevance, Represent the fusion feature vectors Collection, library semantic vector A collection of is the string kernel function output representation of the fused feature vector, is the weight of the library semantic vector;

[0135] Step S1074, loading the similarity between the fused feature vector and the library semantic vector, and determining whether the similarity between the fused feature vector and the library semantic vector meets a preset similarity threshold;

[0136] Step S1075: If the preset similarity threshold is met, the library semantic vector with the highest similarity is selected to replace the fused feature vector;

[0137] In the embodiment of the present invention, the preset similarity threshold may be 0.95-1.

[0138] If it does not meet the preset similarity threshold, return to step S1072 to continue to perform speech feature frequency feature fusion corresponding to the word, phrase and phrase based on the polynomial verification of the support vector machine.

[0139] Step S1076, load at least one group of library semantic vectors, perform weighted combination analysis on the library semantic vectors based on the weights of the library semantic vectors to obtain a follow-up score matrix, use a spectral clustering algorithm to perform follow-up score classification on the follow-up score matrix, integrate at least one group of library semantic vectors, obtain a library semantic vector set and a follow-up score classification associated with the library semantic vector set, and complete the semantic understanding of health follow-up information.

[0140] In this embodiment, when the follow-up score matrix is ​​classified by the spectral clustering algorithm, the similarity between the data points can be calculated according to the data points in the follow-up score matrix. Commonly used similarity measurement methods include Gaussian kernel function, cosine similarity, etc. Then, a degree matrix based on the spectral clustering algorithm is constructed, the Laplace matrix is ​​calculated, the Laplace matrix is ​​decomposed, and then the matrix is ​​clustered using the K-means clustering algorithm, and finally a classification result with a follow-up score of 1-8 is obtained, where 1 is the lowest, indicating that the user's body is greatly affected by chronic diseases.

[0141] In the embodiment of the present invention, the semantic understanding of health follow-up information is achieved through the support vector machine combined with the attention mechanism, so that it has significant robustness to noise and outliers. The attention mechanism can further enhance this robustness by ignoring unimportant information, and the spectral clustering algorithm can be used to classify the follow-up score matrix, thereby improving the flexibility of follow-up information collection. The follow-up score classification can also provide prior knowledge for the risk assessment of the risk assessment model, further ensuring the assessment accuracy of the model, forming a full-scale integration of "data flow + business flow + resource flow", and solving a series of difficult problems such as the high proportion of chronic disease drugs in the elderly, low medical efficiency, and weak patient compliance.

[0142] The embodiment of the present invention provides a converged cloud platform 200. Figure 1 A schematic diagram of a converged cloud platform 200 is shown, wherein the converged cloud platform 200 specifically includes:

[0143] A fusion processing unit 210 is used to load user association information, fuse and process the user association information based on edge computing to obtain a normalized data set, and import the normalized data set into a pre-built disease-specific database 220. The fusion processing unit 210 includes a K-means clustering block, an NLP processing block, an AutoML learning block, a knowledge graph, and an Upgrade block;

[0144] It should be noted that in the fusion processing module, the K-means clustering block, NLP processing block, and AutoML learning block are connected, and the function of the NLP processing block is to perform word segmentation and feature extraction on multi-source heterogeneous data. The NLP processing block can be Hanlp, Stanford, ansj, and jieba word segmentation processing blocks, and the AutoML learning block adopts a fully connected model. The AutoML learning block consists of 300 nodes. The Upgrade block is responsible for managing the K-means clustering block, NLP processing block, AutoML learning block, knowledge graph updates, and version upgrades.

[0145] The disease-specific database 220 is used to load the normalized data set, associate the data in the normalized data set with the user based on the master data management (MDM) technology, generate a unique index code associated with the user, store and visualize the normalized data set;

[0146] In this embodiment, the disease-specific database 220 uses Excel 2019 software for data entry and SPSS 27.0 software for statistical analysis. The disease-specific database 220 supports retrieval tools with customized query conditions to perform full-text retrieval of all clinical information, and can also perform secondary retrieval on the retrieval results, export to Excel, and save and archive them in the scientific research project management system.

[0147] The risk management and control unit 230 pre-builds a risk assessment model based on deep learning combined with an attention mechanism, takes a normalized data set associated with a user as input, executes the risk assessment model, and outputs a risk assessment result.

[0148] The embodiment of the present invention provides a method for processing user association information based on edge computing fusion. Figure 4 The schematic diagram of the implementation process of the method for processing user association information based on edge computing fusion is shown. The method for processing user association information based on edge computing fusion specifically includes:

[0149] Step S201, loading user association information, processing abnormal values ​​and missing values ​​of the user association information, and obtaining an interference information elimination set;

[0150] In this embodiment, descriptive statistics (such as mean, standard deviation), box plots, scatter plots, etc. can be used to identify outliers in the data. For example, if the value of a certain attribute deviates significantly from other data points, or exceeds a preset reasonable range, it may be an outlier.

[0151] Step S202, using the NLP processing block to convert the data format, using one-hot encoding to perform vector conversion on the interference information elimination set, to obtain a correlation information vector set;

[0152] Step S203, loading the associated information vector set, and the AutoML learning block performs dimensionality reduction processing on the associated information vector set based on edge computing to obtain an information dimensionality reduction set;

[0153] In the embodiment of the present invention, when the AutoML learning block performs dimensionality reduction processing on the associated information vector set based on edge computing, the computing power of the edge computing device can be used to accelerate the processing. Edge computing devices (such as edge servers, smart gateways, etc.) have the advantages of being close to data sources, low latency, and strong real-time performance, and are very suitable for processing such computing-intensive tasks, thereby reducing the load on the system and modules.

[0154] Step S204, identify the one-hot encoding associated with the information dimensionality reduction set, map the information dimensionality reduction set to the Euclidean space through the one-hot encoding, integrate at least one set of Euclidean space mapping results to obtain a normalized data set, and expand the values ​​of the discrete features of the normalized data set to the knowledge graph.

[0155] In the embodiment of the present invention, unique-hot encoding is used to perform vector conversion on the interference information elimination set, and the dimension reduction processing of the associated information vector set is performed based on edge computing, thereby realizing the normalization processing of multi-source heterogeneous user associated information and avoiding the loss of feature information. It not only improves the efficiency and accuracy of data processing, but also enhances the consistency and comparability of data, providing a solid foundation for subsequent risk assessment model data analysis and machine learning tasks.

[0156] The embodiment of the present invention provides a method for pre-building a risk assessment model based on deep learning combined with an attention mechanism. Figure 5 A schematic diagram of the implementation process of a method for pre-building a risk assessment model based on deep learning combined with an attention mechanism is shown. The method for pre-building a risk assessment model based on deep learning combined with an attention mechanism specifically includes:

[0157] Step S301, using a convolutional neural network as an initial model of a risk assessment model, the initial model includes an input layer, a convolution module, a hidden layer and an output layer;

[0158] Step S302, indexing the normalized data set in the disease-specific database 220 based on the unique index code, and dividing the normalized data set into a training set and a test set;

[0159] Step S303, initializing the hyperparameters and loss function of the initial model, setting the initial model iterative training rounds and optimizer;

[0160] In this embodiment, the learning rate of the initial model is set to 0.001. The iteration rounds can be 500-540, the optimizer can be the Adam optimizer, the loss function can be the binary cross entropy loss function, the ratio of the training set to the test set can be 3:1, and the hyperparameters of the model are initialized using the He or Xavier initialization method.

[0161] Step S304, obtaining a training set, and using the training set to iteratively train the initial model. During the training, a sensitivity analysis algorithm is used to evaluate the influencing features that are positively correlated with the risk assessment results, and the weights of the influencing features with low sensitivity are pruned and compressed;

[0162] Step S305, adjusting the hyperparameters of the initial model based on the sensitivity coefficients affecting the features, and iteratively training until the initial model converges;

[0163] Step S306, loading the test set, taking the test set as input, executing the initial model, and outputting the test results;

[0164] Step S307, judging whether the test result is qualified based on a preset accuracy threshold, where the accuracy threshold is set to 0.98-0.985;

[0165] Step S308: If the test result is qualified, output a converged risk assessment model.

[0166] If the test result is unqualified, return to step S304 and continue to iteratively train the initial model.

[0167] In this embodiment, in the initial model, the input layer is connected to the convolution module, the convolution module is connected to the hidden layer, the hidden layer is connected to the output layer, the hidden layer is provided with three layers, the hidden layer nodes are 32, 64 and 128 respectively, and the hidden layer activation function is the Act_Logistic activation function;

[0168] The competitive adaptive weighted sampling method and channel attention mechanism module are introduced into the input layer to improve the input layer. The generative adversarial network is introduced into the convolution module. The convolution module consists of the first convolution layer, the second convolution layer, and the third convolution layer. The third convolution layer is frozen, and a multi-layer perceptron with feature parameter sharing is introduced between the second and third convolution layers. The global average pooling layer is used to replace the third convolution layer.

[0169] In an embodiment of the present invention, a risk assessment model is provided. The risk assessment model uses a convolutional neural network as the initial model, and introduces a competitive adaptive weighted sampling method, a channel attention mechanism, a generative adversarial network and a multi-layer perceptron. The channel attention mechanism is used to effectively extract the influencing features associated with different chronic diseases, so as to realize the evaluation and prediction of different chronic diseases, and overcome the limitation that the prior art can only predict and evaluate a single or two chronic diseases. At the same time, the risk assessment model is combined with a multi-source heterogeneous normalized data set to obtain highly timely and highly correlated features associated with chronic diseases from different dimensions and types of data, thereby reducing the loss of timeliness feature information, and realizing accurate assessment of the user's chronic disease risk while less increasing the complexity of the model.

[0170] The embodiment of the present invention provides a method for executing a risk assessment model and outputting a risk assessment result. Figure 6 The following is a schematic diagram of the implementation process of a method for executing a risk assessment model and outputting a risk assessment result. The method for executing a risk assessment model and outputting a risk assessment result specifically includes:

[0171] Step S401, indexing the normalized data set associated with the user in the disease-specific database 220, and the input layer determines the type of chronic disease associated with the user in the normalized data set based on the competitive adaptive weighted sampling method;

[0172] Step S402, extracting features of the normalized data set through a maximum pooling operation combined with an average pooling operation based on the channel attention mechanism module to obtain a feature factor set;

[0173] Step S403, loading the characteristic factor set and the chronic disease type, setting the weights of the influencing features associated with the chronic disease type based on the Delphi method, extracting the influencing features associated with the chronic disease type, and calculating the sensitivity coefficient using a sensitivity analysis algorithm;

[0174] It should be noted that when setting the weights of influencing features associated with chronic disease types in the Delphi method, a group (20 people) of experts with knowledge and experience in the field of chronic disease prevention and treatment can be selected. These experts should come from different backgrounds to ensure diversity and comprehensiveness. Then, according to the purpose and needs of the study, a questionnaire containing various potential influencing features is designed. These features may include lifestyle, genetic factors, environmental factors, socioeconomic status, etc., depending on the type of chronic disease being studied. The questionnaire is distributed to all selected experts, and their scores and opinions on the importance of each feature are collected. It is usually conducted anonymously to reduce bias and pressure. After collecting the results of the first round of surveys, the data are summarized and analyzed. Statistics such as the mean score, median, mode, etc. are calculated for each feature, and consensus and disagreement points are identified. The results of the first round are organized into a report, including the scores of each feature, a summary of expert opinions, and issues that need further discussion. This report is then distributed to all experts as a starting point for the second round of surveys. In the second round, experts can re-evaluate their scores based on new information and provide further opinions.

[0175] In the embodiment of the present invention, the influencing characteristics include but are not limited to the user's age, gender, smoking status, drinking status, physical activity, coronary heart disease, stroke, lung cancer, family history of hypertension, family history of diabetes, family history of dyslipidemia, white blood cell count, red blood cell count, total cholesterol, triglycerides, high-density lipoprotein, low-density lipoprotein, systolic blood pressure, diastolic blood pressure and other influencing characteristics.

[0176] Step S404, determining whether the sensitivity coefficient of the influencing feature exceeds a preset sensitivity threshold; if it exceeds the preset sensitivity threshold, retaining the current influencing feature, integrating at least one group of influencing features, and obtaining a sensitive feature set;

[0177] Step S405, loading the sensitive feature set, the convolution module performs a convolution dimension reduction operation on the influencing features in the sensitive feature set through the first 7×7 convolution layer and the second convolution layer, and the multi-layer perceptron performs element-by-element weighted fusion on the influencing features, and outputs a feature fusion set;

[0178] Step S406: The hidden layer uses the follow-up score classification as prior knowledge, combines the genetic algorithm to globally search the feature fusion set, and calculates the risk assessment result through the Act_Logistic activation function.

[0179] In this embodiment, the sensitivity analysis algorithm is used to calculate the sensitivity coefficient using the following formula:

[0180] (8)

[0181] in, Indicates the sensitivity coefficient of the current influencing feature, represents the sensitivity constant of the sensitivity analysis algorithm. In this embodiment, the sensitivity constant can be 2-8. is the weight of the current influencing feature. In this embodiment, the weight may be 0.1-0.85. Represents the average weight of the feature factor set that affects the feature;

[0182] The risk assessment result is calculated using the following formula:

[0183] (9)

[0184] in, represents the risk assessment level, Indicates the sensitivity coefficient of the current influencing feature, is the weight of the current influencing feature, and The classification level of follow-up score is represents the input representation of the feature fusion set, is the bias term of the Act_Logistic activation function, It is the number of iterative searches of the genetic algorithm, which can be 15-25 times.

[0185] In summary, the present invention provides a chronic disease medical and preventive integrated collaborative management and control system. In an embodiment of the present invention, a risk assessment model based on deep learning combined with an attention mechanism is constructed through the fusion cloud platform 200, so that the risk assessment model can identify and analyze the normalized data set collected by the user terminal 100, and the user terminal 100 can collect multi-source heterogeneous normalized data sets with good timeliness and real-time reflection of the patient's health status, thereby ensuring that the risk assessment model provides an accurate assessment path for chronic disease medical prevention and control, and overcomes the problem that when the key feature extraction module of the existing system extracts the user's electronic medical case files, the user's electronic medical case files have poor timeliness and cannot reflect the patient's health status changes in real time, and ignores the patient's personal living habits, environmental factors and other personalized information, resulting in inaccurate extraction of key features, which in turn affects the assessment and prediction accuracy of chronic diseases.

[0186] It should be noted that, for the above-mentioned embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should know that the present invention is not limited by the described order of actions, because according to the present invention, some steps may be performed in other orders or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.

[0187] In the several embodiments provided in the present application, it should be understood that the disclosed device can be implemented in other ways. For example, the device embodiments described above are only schematic, such as the division of the above-mentioned units, which is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be in the form of telecommunication or other forms.

[0188] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on these embodiments, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in this field can still combine, add, delete or make other adjustments to the features in the various embodiments of the present invention according to the circumstances without conflict, without making creative work, so as to obtain different other technical solutions that do not deviate from the concept of the present invention in essence, and these technical solutions also belong to the scope of protection of the present invention.

Claims

1. A chronic disease medical and preventive integrated collaborative management and control system, characterized in that: The chronic disease medical and preventive integrated collaborative management and control system includes: The user end is used to collect multi-source heterogeneous user-related information and upload the user-related information to the fusion cloud platform in real time. The user-related information includes user basic health information, health follow-up information, related family doctor information, chronic disease control information, user medical record information, and user vital sign monitoring data; Fusion cloud platform, load user-related information, process user-related information based on edge computing, obtain normalized data set, import normalized data set into pre-built disease-specific database, pre-build risk assessment model based on deep learning combined with attention mechanism, use normalized data set associated with user as input, execute risk assessment model, and output risk assessment result; The medical prevention judgment module obtains the risk assessment results and determines whether the risk assessment results exceed the preset risk threshold. If so, the risk reminder instruction is triggered and the risk assessment results are shared and collaboratively pushed to the user end. If not, the risk assessment results are pushed to the follow-up customization module. The follow-up customization module, in response to the risk assessment results, traverses the disease-specific database, indexes the follow-up survey information associated with the user, combines the risk assessment results and the follow-up survey information to generate an intelligent follow-up plan for chronic diseases, and pushes the intelligent follow-up plan to the user end; The method for obtaining and parsing health follow-up information includes: Obtain health follow-up information, retrieve the user's basic health information from the disease-specific database, verify the user's identity in the health follow-up information based on the user's basic health information, and determine whether the user's identity authentication is passed; If the user identity verification is passed, the semantic understanding instruction is triggered, the health follow-up information is loaded, and a multi-channel band-stop filter is used to extract the non-speech signal in the health follow-up information; Performing short-time Fourier transform on the non-speech signal to obtain a non-speech signal spectrum sequence, selecting the non-speech signal center frequency based on the non-speech signal spectrum sequence, and using a Fourier transform algorithm to calculate the interference amplitude energy and interference phase of the non-speech signal spectrum sequence; The interference amplitude energy and interference phase of the non-speech signal spectrum sequence are calculated by the following formula: （1） （2） （3） in, Represents the interference amplitude energy of the non-speech signal spectrum sequence, is the interference phase of the non-speech signal spectrum sequence, is the number of discretized signal points of the non-speech signal spectrum sequence, is the input representation of the non-speech signal spectrum sequence, is the center frequency of the non-speech signal, represents the cutoff frequency of the band-stop filter, Indicates the transformation result of the non-speech signal spectrum sequence based on the Fourier transform algorithm; Obtain interference amplitude energy and interference phase of the non-speech signal spectrum sequence, use a band-stop filter to perform L-order harmonic suppression and attenuation on the non-speech signal spectrum sequence, eliminate interference noise information in the health follow-up information, and output health follow-up information containing speech characteristic frequencies; The interference noise information is expressed as: （4） （5） in, Indicates interference noise information, is the stop-band attenuation of the band-stop filter, Indicates the number of harmonics, is the insertion loss of the band-stop filter, are the sampling angle differences of the band-stop filter for the positive and negative frequency components of the harmonics, is the speech characteristic frequency; Acquire health follow-up information containing speech feature frequencies, understand the semantics of the health follow-up information based on a support vector machine combined with an attention mechanism, output the semantically understood health follow-up information, and automatically collect the health follow-up information into a data acquisition unit; The method for semantic understanding of health follow-up information based on a support vector machine combined with an attention mechanism includes: Loading health follow-up information, dividing the health follow-up information into characters, words, and phrases, and identifying the speech feature frequencies corresponding to the characters, words, and phrases; Based on the polynomial of the support vector machine, the speech feature frequency features corresponding to the characters, words and phrases are fused to obtain the fused feature vector; Using the string kernel function of the support vector machine to index at least one set of library semantic vectors in the disease-specific database that have a high correlation with the fused feature vector; The correlation between the fusion feature vector and the library semantic vector is calculated by the following formula: （6） The correlation between the fused feature vector and the library semantic vector is corrected by correcting the cosine similarity to obtain the similarity between the fused feature vector and the library semantic vector; （7） in, Indicates the similarity between the fusion feature vector and the library semantic vector, Represents the fused feature vector Semantic vector with library Relevance, Represent the fusion feature vectors Collection, library semantic vector A collection of is the string kernel function output representation of the fused feature vector, is the weight of the library semantic vector; Load the similarity between the fused feature vector and the library semantic vector, and determine whether the similarity between the fused feature vector and the library semantic vector meets the preset similarity threshold. If the similarity threshold is met, select the library semantic vector with the highest similarity to replace the fused feature vector; Load at least one set of library semantic vectors, perform weighted combination analysis on the library semantic vectors based on their weights to obtain a follow-up score matrix, use a spectral clustering algorithm to classify the follow-up score matrix, integrate at least one set of library semantic vectors, obtain a library semantic vector set and a follow-up score classification associated with the library semantic vector set, and complete the semantic understanding of health follow-up information; The method of executing the risk assessment model and outputting the risk assessment results includes: The normalized data set associated with the user in the index disease database is used, and the input layer determines the chronic disease type of the user associated with the normalized data set based on the competitive adaptive weighted sampling method; Based on the channel attention mechanism module, the normalized data set features are extracted through the maximum pooling operation combined with the average pooling operation to obtain the feature factor set; Load the characteristic factor set and chronic disease types, set the weights of the influencing features associated with the chronic disease types based on the Delphi method, extract the influencing features associated with the chronic disease types, and calculate the sensitivity coefficient using the sensitivity analysis algorithm; Determine whether the sensitivity coefficient of the influencing feature exceeds a preset sensitivity threshold. If it exceeds the preset sensitivity threshold, retain the current influencing feature, integrate at least one group of influencing features, and obtain a sensitive feature set; Load the sensitive feature set. The convolution module performs convolution dimensionality reduction operations on the influencing features in the sensitive feature set through the first 7×7 convolution layer and the second convolution layer. The multi-layer perceptron performs element-by-element weighted fusion on the influencing features and outputs the feature fusion set. The hidden layer uses the follow-up score classification as prior knowledge, combines the genetic algorithm to globally search the feature fusion set, and calculates the risk assessment result through the Act_Logistic activation function.

2. The chronic disease medical and preventive integrated collaborative management and control system according to claim 1, characterized in that: The user terminal comprises: The health monitoring unit is used to collect the user's vital signs monitoring data in a distributed manner and upload the user's vital signs monitoring data to the data acquisition unit in real time; A follow-up management unit, in response to the intelligent follow-up plan pushed by the follow-up customization module, identifies the intelligent follow-up plan, triggers personalized follow-up instructions based on the intelligent follow-up plan, obtains and parses health follow-up information, and pushes the health follow-up information to the data acquisition unit; The data acquisition unit is connected to the health monitoring unit, the follow-up management unit, and the medical institution information system for communication, and is used to acquire multi-source heterogeneous user association information from the health monitoring unit, the follow-up management unit, and the medical institution information system.

3. The chronic disease medical and preventive integrated collaborative management and control system according to claim 2, characterized in that: The integrated cloud platform includes: A fusion processing unit, which is used to load user association information, fuse and process the user association information based on edge computing to obtain a normalized data set, and import the normalized data set into a pre-built disease-specific database. The fusion processing unit includes a K-means clustering block, an NLP processing block, an AutoML learning block, a knowledge graph, and an Upgrade block. The disease-specific database is used to load the normalized data set, associate the data in the normalized data set with the user based on the master data management (MDM) technology, generate a unique index code for the user, store and visualize the normalized data set; The risk management and control unit pre-builds a risk assessment model based on deep learning combined with the attention mechanism, takes the normalized data set associated with the user as input, executes the risk assessment model, and outputs the risk assessment results.

4. The chronic disease medical and preventive integrated collaborative management and control system according to claim 1, characterized in that: The method for pre-building a risk assessment model based on deep learning combined with an attention mechanism includes: The convolutional neural network is used as the initial model of the risk assessment model, and the initial model includes an input layer, a convolution module, a hidden layer, and an output layer; The normalized data set in the disease-specific database is indexed based on the unique index code, and the normalized data set is divided into a training set and a test set; Initialize the hyperparameters and loss function of the initial model, and set the initial model iterative training rounds and optimizer.

5. The chronic disease medical and preventive integrated collaborative management and control system according to claim 4, characterized in that: The method for pre-building a risk assessment model based on deep learning combined with an attention mechanism also includes: Obtain a training set and use it to iteratively train the initial model. During training, use a sensitivity analysis algorithm to evaluate the influencing features that are positively correlated with the risk assessment results, prune and compress the weights of influencing features with low sensitivity, adjust the hyperparameters of the initial model based on the sensitivity coefficients of the influencing features, and iterate the training until the initial model converges; Load the test set, use the test set as input, execute the initial model, output the test results, and judge whether the test results are qualified based on the preset accuracy threshold. If the test results are qualified, output the converged risk assessment model.

6. The chronic disease medical and preventive integrated collaborative management and control system according to claim 2, characterized in that: The sensitivity analysis algorithm is used to calculate the sensitivity coefficient using the following formula: （8） in, Indicates the sensitivity coefficient of the current influencing feature, represents the sensitivity constant of the sensitivity analysis algorithm, is the weight of the current influencing feature, Represents the average weight of the feature factor set that affects the feature; The risk assessment result is calculated using the following formula: （9） in, represents the risk assessment level, Indicates the sensitivity coefficient of the current influencing feature, is the weight of the current influencing feature, and The classification level of follow-up score is represents the input representation of the feature fusion set, is the bias term of the Act_Logistic activation function, It is the number of iterative searches of the genetic algorithm.

Images