Physical sign data monitoring method and system based on medical internet of things

Abstract

The invention provides a physical sign data monitoring method and system based on a medical internet of things, and the method comprises the steps: firstly converting sparsely collected first medical interaction scene data into a target medical service feature domain through employing a first graph generation model, effectively solving a data sparsity problem, and guaranteeing the consistency of data in different scenes. And then, based on the converted graph feature representation data, knowledge point prediction is carried out through a basic medical service knowledge point prediction network, network parameters are continuously optimized through an error feedback mechanism, and finally, a high-performance medical service knowledge point prediction network is generated. The medical service knowledge point prediction network not only can accurately predict the medical service knowledge points of the remote session trajectory data from the densely collected second medical interaction scene, but also can automatically generate a personalized physical sign data monitoring scheme based on the prediction result, thereby providing scientific and efficient decision support for doctors; and meanwhile, the precision and intelligence of health management of the patient are also promoted.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and more specifically, to a method and system for monitoring vital sign data based on the medical Internet of Things. Background Art

[0002] With the rapid development of medical technology and the widespread application of Internet of Things technology, the Internet of Medical Things (IoMT) has shown great potential in the fields of health monitoring, disease diagnosis and management. By integrating various sensors, wearable devices, telemedicine systems, etc., the Internet of Medical Things realizes the real-time collection, transmission and analysis of patients' vital signs data, providing strong support for personalized medicine and precise health management. However, in actual applications, vital sign data in different medical interaction scenarios have problems such as different collection density, diverse data formats, and lack of knowledge annotation, which seriously restricts the accuracy and efficiency of vital sign data monitoring.

[0003] Traditional vital sign data monitoring methods often rely on data from a single scenario and are unable to effectively address the challenges posed by cross-scenario data differences. Especially in remote areas or resource-constrained medical interaction scenarios, the collection density of vital sign data is low and the data quality is uneven, which greatly reduces the performance of knowledge discovery and prediction models based on these data. At the same time, vital sign data in different medical interaction scenarios often follow their own unique feature distribution and service model. Directly migrating models for cross-scenario predictions often has poor results and may even lead to misleading results.

[0004] In order to overcome the above problems, relevant technologies have begun to explore the use of advanced technologies such as graph neural networks (GNNs) to model and analyze medical IoT data. Graph neural networks represent the relationship between data by constructing graph structures, which can effectively capture feature information in complex networks and provide new ideas for cross-scenario data fusion and knowledge prediction. However, most existing technologies focus on data modeling in a single scenario and lack effective methods for multi-scenario data fusion and conversion. Especially in the face of medical interaction scenarios with significant differences in collection density, how to construct a unified target feature domain and achieve efficient knowledge prediction is still a technical problem that needs to be solved urgently. Summary of the invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present application, the embodiment of the present application provides a method for monitoring vital sign data based on the medical Internet of Things, the method comprising: Acquire first remote conversation trajectory training data originating from a first medical interaction scenario, wherein the first remote conversation trajectory training data carries first medical service knowledge point annotation data, and the first medical service knowledge point annotation data represents actual medical service knowledge points of the first remote conversation trajectory training data; Using a first graph generation model to perform graph feature representation on the first remote conversation trajectory training data to generate first graph feature representation data, the first graph generation model is used to convert the remote conversation trajectory data originating from the first medical interaction scenario into a target medical service feature domain, the target medical service feature domain is a medical service feature domain where the graph feature representation data corresponding to the remote conversation trajectory data originating from the second medical interaction scenario is located, the first medical interaction scenario and the second medical interaction scenario are different medical interaction scenarios, and the collection density of the remote conversation trajectory data originating from the first medical interaction scenario is less than the collection density of the remote conversation trajectory data originating from the second medical interaction scenario; Based on the first graph feature representation data, using the basic medical service knowledge point prediction network to predict knowledge points and generate predicted medical service knowledge points; Based on the error between the predicted medical service knowledge point and the first medical service knowledge point annotation data, the network parameter information of the basic medical service knowledge point prediction network is updated, and a medical service knowledge point prediction network is generated, so that the medical service knowledge point prediction network is used to predict the medical service knowledge points of the remote conversation trajectory data originating from the second medical interaction scenario, and the corresponding vital sign data monitoring plan is generated based on the medical service knowledge points of the remote conversation trajectory data originating from the second medical interaction scenario.

[0006] On the other hand, an embodiment of the present application also provides a medical Internet of Things service system, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0007] Based on the above aspects, the embodiment of the present application realizes data feature conversion and knowledge prediction across different medical interaction scenarios by combining a graph generation model with a deep learning network, significantly improving the accuracy and generalization ability of medical service knowledge point prediction. Specifically, the first graph generation model is first used to convert the sparsely collected first medical interaction scenario data into the target medical service feature domain, which effectively solves the data sparsity problem and ensures the consistency of data between different scenarios. Subsequently, based on the converted graph feature representation data, knowledge point prediction is performed through the basic medical service knowledge point prediction network, and the network parameters are continuously optimized through the error feedback mechanism, and finally a high-performance medical service knowledge point prediction network is generated. The medical service knowledge point prediction network can not only accurately predict the medical service knowledge points of the remote conversation trajectory data from the densely collected second medical interaction scenario, but also automatically generate personalized vital sign data monitoring solutions based on the prediction results, providing doctors with scientific and efficient decision support, and also promoting the precision and intelligence of patient health management. As a result, the efficiency and quality of medical services are greatly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 It is a schematic diagram of the execution flow of the vital sign data monitoring method based on the medical Internet of Things provided in an embodiment of the present application.

[0009] Figure 2 It is a schematic diagram of the hardware architecture of the medical Internet of Things service system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0010] The present application will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of a vital sign data monitoring method based on the medical Internet of Things provided by an embodiment of the present application. The vital sign data monitoring method based on the medical Internet of Things is introduced in detail below.

[0011] Step S110, obtaining first remote conversation trajectory training data originating from a first medical interaction scenario, wherein the first remote conversation trajectory training data carries first medical service knowledge point annotation data, and the first medical service knowledge point annotation data represents actual medical service knowledge points of the first remote conversation trajectory training data.

[0012] In this embodiment, in the medical field, different medical interaction scenarios will generate a large amount of remote conversation trajectory data, which contains rich medical service information and interaction details between patients and doctors. As the execution subject, the server first obtains the first remote conversation trajectory training data from the first medical interaction scenario. This process usually involves extracting data from a medical information system (HIS), an electronic medical record system (EMR) or a telemedicine platform.

[0013] For example, suppose that the first medical interaction scenario is an online medical consultation platform in a remote area. Due to geographical and resource limitations, the remote conversation data between users (patients) of the online medical consultation platform and doctors is relatively small and sparse. The server regularly pulls remote conversation records from the database of the online medical consultation platform. These remote conversation records include the patient's basic information, chief complaint, doctor's consultation process, diagnosis suggestions and follow-up treatment suggestions. Each remote conversation record is regarded as a first remote conversation trajectory training data, and each first remote conversation trajectory training data is accompanied by first medical service knowledge point annotation data, which are manually annotated by professional medical personnel according to the content of the conversation. For example, according to the content of the conversation, the key medical knowledge points involved are manually annotated, such as "respiratory infection management", "chronic disease follow-up guidance", etc. The server associates these first medical service knowledge point annotation data with the corresponding first remote conversation trajectory data to form a training data set.

[0014] Step S120, using a first graph generation model to perform graph feature representation on the first remote conversation trajectory training data to generate first graph feature representation data, wherein the first graph generation model is used to convert the remote conversation trajectory data originating from the first medical interaction scenario into a target medical service feature domain, wherein the target medical service feature domain is a medical service feature domain in which the graph feature representation data corresponding to the remote conversation trajectory data originating from the second medical interaction scenario is located, wherein the first medical interaction scenario and the second medical interaction scenario are different medical interaction scenarios, and the collection density of the remote conversation trajectory data originating from the first medical interaction scenario is less than the collection density of the remote conversation trajectory data originating from the second medical interaction scenario.

[0015] In this embodiment, since the data collection density of the first medical interaction scenario is relatively low, directly using these first remote conversation trajectory training data to train the knowledge point prediction model may not be effective. Therefore, the server uses the first graph generation model to convert these sparse remote conversation trajectory data into a more general target medical service feature domain, which is usually constructed based on the second medical interaction scenario data with higher collection density.

[0016] For example, the server first loads the pre-trained first graph generation model, which is based on the graph neural network (GNN) architecture and can capture the complex relational structure in the remote conversation trajectory. The input of the first graph generation model is the first remote conversation trajectory training data in the first medical interaction scenario. Each first remote conversation trajectory training data is represented as a graph structure, where nodes can be patients, doctors, medical events, etc., and edges represent the relationship between them (such as consultation, diagnosis, advice, etc.). The first graph generation model extracts high-level features in the graph structure through multi-layer graph convolution operations and generates corresponding first graph feature representation data. These first graph feature representation data are mapped to the target medical service feature domain, which is based on an online medical consultation platform (second medical interaction scenario) of a large urban hospital. The data collection density of this online medical consultation platform is much higher than that of platforms in remote areas.

[0017] Step S130, based on the first graph feature representation data, using the basic medical service knowledge point prediction network to perform knowledge point prediction and generate predicted medical service knowledge points.

[0018] In this embodiment, after obtaining the converted graph feature representation data, the server then uses the basic medical service knowledge point prediction network to predict knowledge points. The basic medical service knowledge point prediction network is usually a deep learning model, such as a combination of a convolutional neural network (CNN) or a recurrent neural network (RNN) and an attention mechanism, which can extract key information from the graph feature representation data and predict the corresponding medical service knowledge points.

[0019] For example, the server inputs the first graph feature representation data into the basic medical service knowledge point prediction network. The basic medical service knowledge point prediction network first further processes the graph feature representation data through several layers of convolutional layers or loop layers to extract features related to medical service knowledge points. Then, the attention mechanism is used to focus on the most relevant features to generate predicted medical service knowledge points.

[0020] Step S140, based on the error between the predicted medical service knowledge point and the first medical service knowledge point annotation data, update the network parameter information of the basic medical service knowledge point prediction network, generate a medical service knowledge point prediction network, so that the medical service knowledge point prediction network is used to predict the medical service knowledge points of the remote conversation trajectory data originating from the second medical interaction scenario, and generate a corresponding vital sign data monitoring plan based on the medical service knowledge points of the remote conversation trajectory data originating from the second medical interaction scenario.

[0021] In this embodiment, in order to improve the accuracy of the basic medical service knowledge point prediction network, the server will calculate the error between the predicted medical service knowledge point and the first medical service knowledge point annotation data, and update the network parameter information of the basic medical service knowledge point prediction network based on this error. This process is usually implemented through a back propagation algorithm, which aims to minimize the prediction error so that the model can predict medical service knowledge points more accurately.

[0022] For example, the server compares the predicted knowledge points with the annotated knowledge points and calculates the error between them (such as cross entropy loss). Then, using optimization algorithms such as gradient descent, the weights and biases of the prediction network are updated according to the error back propagation. After multiple iterations of training, the performance of the prediction network gradually improves and the error gradually decreases. In the end, the server obtains a trained medical service knowledge point prediction network that can not only process sparse data in the first medical interaction scenario, but also adapt to dense data in the second medical interaction scenario, providing strong support for the subsequent generation of vital sign data monitoring solutions.

[0023] Therefore, the first remote conversation trajectory training data from the sparse acquisition scenario is converted into a general graph feature representation, and this knowledge is used to train a high-performance medical service knowledge point prediction network, which can not only improve the efficiency and quality of remote medical consultation, but also provide a scientific basis for the subsequent generation of vital sign data monitoring plans, thereby promoting the rational allocation of medical resources and the effective management of patient health.

[0024] For example, in order to generate a vital sign data monitoring plan, the server needs to predefine a set of vital sign monitoring templates. These vital sign monitoring templates are designed according to different medical service knowledge points and contain vital sign monitoring indicators, monitoring frequency, monitoring methods and other information for different diseases or health problems. For example, for the knowledge point of "hypertension management", the vital sign monitoring template can include respiratory tract monitoring indicators, monitoring frequency (once in the morning and once in the evening), monitoring methods, etc.

[0025] The server matches the predicted medical service knowledge points with the vital sign monitoring templates and finds the vital sign monitoring templates corresponding to the knowledge points.

[0026] Although the vital sign monitoring template provides a basic vital sign monitoring framework, the specific circumstances of different patients may vary. Therefore, the server also needs to personalize the monitoring plan based on the patient's personal information (such as age, gender, medical history, etc.) and the specific information in the session trajectory (such as symptom description, doctor's advice, etc.). For example, for elderly patients or patients with more serious conditions, it may be necessary to increase the monitoring frequency or add additional monitoring indicators. After matching and personalization, the server finally generates a vital sign data monitoring plan for each remote session trajectory, which is sent to patients or medical staff in the form of electronic documents or application notifications to guide them to monitor and record vital sign data. At the same time, the server can also integrate a vital sign data collection system, allowing patients to upload monitoring data in real time through wearable devices, mobile applications, etc. The server analyzes and processes these data in real time, and immediately issues an alarm once an abnormal situation is found, so that medical staff can intervene and handle it in time.

[0027] Based on the above steps, the embodiment of the present application realizes data feature conversion and knowledge prediction across different medical interaction scenarios by combining the graph generation model with the deep learning network, significantly improving the accuracy and generalization ability of medical service knowledge point prediction. Specifically, the first graph generation model is first used to convert the sparsely collected first medical interaction scene data into the target medical service feature domain, which effectively solves the data sparsity problem and ensures the consistency of data between different scenarios. Subsequently, based on the converted graph feature representation data, the knowledge point prediction is performed through the basic medical service knowledge point prediction network, and the network parameters are continuously optimized through the error feedback mechanism, and finally a high-performance medical service knowledge point prediction network is generated. The medical service knowledge point prediction network can not only accurately predict the medical service knowledge points of the remote conversation trajectory data from the densely collected second medical interaction scene, but also automatically generate personalized vital sign data monitoring solutions based on the prediction results, providing doctors with scientific and efficient decision support, and also promoting the precision and intelligence of patient health management. As a result, the efficiency and quality of medical services are greatly improved.

[0028] In a possible implementation, the method further includes: Step A110, obtaining the remote conversation trajectory data to be predicted from the second medical interaction scenario.

[0029] In this embodiment, the server regularly pulls new remote conversation trajectory data to be predicted from the database of the second medical interaction scenario (the online medical consultation platform of a large urban hospital) as the prediction object. These remote conversation trajectory data to be predicted have not yet been labeled with knowledge points, but the format is similar to the training data, including the patient's basic information, chief complaint, doctor's consultation process, etc.

[0030] In detail, the server establishes a connection with the database of the second medical interaction scenario, and according to a preset time interval or data update flag, the server performs a data pulling operation, and the pulled data is temporarily stored in the local database of the server for subsequent processing.

[0031] Step A120, using a second graph generation model to perform graph feature representation on the remote conversation trajectory data to be predicted to generate second graph feature representation data, wherein the second graph generation model is used to convert the remote conversation trajectory data originating from the second medical interaction scenario into the target medical service feature domain.

[0032] Different from the first graph generation model in the first medical interaction scenario, the second graph generation model is specifically used to process the remote conversation trajectory data in the second medical interaction scenario. The second graph generation model is also based on the graph neural network (GNN) architecture, but has been optimized according to the data characteristics of the second medical interaction scenario.

[0033] In detail, the server loads the pre-trained second graph generation model and inputs the remote session trajectory data to be predicted into the second graph generation model. Each remote session trajectory data to be predicted is converted into a graph structure, and the definitions of nodes and edges are consistent with the training data. The second graph generation model extracts high-level features in the graph structure through multi-layer graph convolution operations and generates corresponding second graph feature representation data. These second graph feature representation data are also mapped to the target medical service feature domain to ensure consistency with the training data.

[0034] Step A130: Based on the second graph feature representation data, knowledge point prediction is performed using the medical service knowledge point prediction network to generate medical service knowledge points for the remote conversation trajectory data to be predicted.

[0035] The second graph feature representation data converted by the second graph generation model is now input into the trained medical service knowledge point prediction network, which can predict the corresponding medical service knowledge points based on the graph feature representation data.

[0036] In detail, the server inputs the second graph feature representation data into the medical service knowledge point prediction network. The medical service knowledge point prediction network further processes the graph feature representation data through multiple convolutional layers or loop layers to extract features related to medical service knowledge points. The attention mechanism is used to focus on the most relevant features and generate predicted medical service knowledge points. The prediction results are output in the form of labels or text, indicating the predicted medical service knowledge points, such as "hypertension management", "diabetes diet guidance", etc.

[0037] For example, suppose there is a new remote conversation trajectory data in the second medical interaction scenario, which is a regular follow-up record of a patient with hypertension. The server pulls the follow-up record of the patient from the database, including blood pressure measurement results, patient-reported symptoms, doctor's consultation process, etc. The second graph generation model converts the follow-up record into a graph structure, and extracts key features such as "persistently high blood pressure" and "self-reported dizziness" through graph convolution operations to generate the corresponding second graph feature representation data. The medical service knowledge point prediction network receives the second graph feature representation data, and after processing, predicts "hypertension management" as the main medical service knowledge point. The server outputs the prediction results in text form, suggesting that the doctor strengthen the patient's hypertension management, including adjusting the medication regimen, increasing the monitoring frequency, etc.

[0038] In another example, the server processes a comprehensive care record for a chronic disease patient (such as a diabetic patient). The processing flow is similar to the above example, but the predicted medical service knowledge points may be more complex and comprehensive, such as "comprehensive management of diabetes", involving multiple aspects such as diet guidance, exercise advice, and blood sugar monitoring.

[0039] In one possible implementation, the first graph generation model includes a first interaction scenario model and a second interaction scenario model, the first interaction scenario model being used to convert remote session trajectory data originating from the first medical interaction scenario into the target medical service feature domain, and the second interaction scenario model being used to convert remote session trajectory data originating from the second medical interaction scenario into the target medical service feature domain.

[0040] Step S120 includes: performing graph feature representation on the first remote conversation trajectory training data using the first interaction scenario model to generate first graph feature representation data.

[0041] The method further comprises: Step B110, obtaining the remote conversation trajectory data to be predicted from the second medical interaction scenario.

[0042] Step B120: Performing graph feature representation on the to-be-predicted remote conversation trajectory data using the second interaction scenario model to generate second graph feature representation data.

[0043] Step B130: Based on the second graph feature representation data, the medical service knowledge point prediction network is used to perform knowledge point prediction to generate the medical service knowledge points of the remote conversation trajectory data to be predicted.

[0044] In this embodiment, a first graph generation model is pre-defined and trained in the server. The first graph generation model includes two sub-models: a first interaction scenario model and a second interaction scenario model. The two models are respectively optimized for data characteristics of two different medical interaction scenarios.

[0045] In detail, during the training phase, the server uses a large amount of historical data from the first medical interaction scenario and the second medical interaction scenario to train the first interaction scenario model and the second interaction scenario model respectively. These models are based on the graph neural network (GNN) architecture and can capture the complex relationship structure in the remote conversation trajectory and convert it into a unified target medical service feature domain. After the training is completed, the server integrates these two sub-models into the first graph generation model so that they can be flexibly switched when processing data from different sources.

[0046] Therefore, the server obtains a batch of remote conversation trajectory training data from the first medical interaction scene, and prepares to use the first interaction scene model to represent its graph features.

[0047] In detail, the server first preprocesses the first remote conversation trajectory training data, including operations such as cleaning, formatting, and standardization to ensure data quality. Then, each preprocessed first remote conversation trajectory data is converted into a graph structure, in which nodes represent entities in the conversation (such as patients, doctors, medical events, etc.), and edges represent relationships between entities (such as consultation, diagnosis, advice, etc.). Using the first interaction scenario model, the server performs multi-layer graph convolution operations on the constructed graph structure to extract high-level features in the graph structure, which can characterize the key information in the remote conversation trajectory. After feature extraction, the server generates corresponding first graph feature representation data, which are mapped to the target medical service feature domain for subsequent knowledge point prediction tasks.

[0048] As the system runs, the server needs to obtain new remote conversation trajectory data from the second medical interaction scene in real time to predict knowledge points. Specifically, the server establishes a real-time synchronization mechanism with the database of the second medical interaction scene to ensure that the latest remote conversation trajectory data can be obtained in a timely manner. From the synchronized data, the server filters out the remote conversation trajectory data to be predicted that needs to be predicted for knowledge points. These remote conversation trajectory data to be predicted may be filtered based on specific conditions or rules, such as only selecting conversation trajectories containing specific disease or symptom records.

[0049] For the filtered remote conversation trajectory data to be predicted, the server uses the second interaction scenario model to represent its graph features. Similarly, the server first converts the remote conversation trajectory data to be predicted into a graph structure. But this time, the second interaction scenario model is used, which is optimized for the data characteristics of the second medical interaction scenario. Through multi-layer graph convolution operations, the server extracts high-level features in the graph structure. Similarly, these features are mapped to the target medical service feature domain to generate the second graph feature representation data.

[0050] After obtaining the second graph feature representation data, the server uses the trained medical service knowledge point prediction network to predict the knowledge points. For example, the second graph feature representation data can be input into the medical service knowledge point prediction network, and the input data can be further processed through multiple convolutional layers or loop layers, and the attention mechanism can be used to focus on the most relevant features. Finally, the network outputs the predicted medical service knowledge points, such as "respiratory infection management", "chronic disease follow-up guidance", etc. Thus, the prediction results are output in text or other forms for reference by doctors or patients.

[0051] In one possible implementation, the method includes: Step C110, obtaining multiple remote conversation trajectory training data combinations, wherein the remote conversation trajectory training data combinations include first member remote conversation trajectory training data originating from the first medical interaction scenario and second member remote conversation trajectory training data originating from the second medical interaction scenario, and the first member remote conversation trajectory training data and the second member remote conversation trajectory training data originating from the same remote conversation trajectory training data combination are the same medical service data.

[0052] In this embodiment, in this scenario, the server is responsible for processing and optimizing the remote conversation trajectory data of two different medical interaction scenarios (the first medical interaction scenario and the second medical interaction scenario). The server trains and optimizes the two interaction scenario models (the first interaction scenario model and the second interaction scenario model) to ensure that these models can accurately convert the remote conversation trajectory data in their respective scenarios into a unified target medical service feature domain.

[0053] For example, remote conversation trajectory data can be collected from two different medical interaction scenarios and combined. Each combination contains two members: a first member remote conversation trajectory training data from the first medical interaction scenario, and a second member remote conversation trajectory training data from the second medical interaction scenario. The two members share the same medical service data, that is, they involve the same medical event or patient case.

[0054] In detail, the server establishes connections with the databases of the first medical interaction scenario and the second medical interaction scenario respectively, and pulls the remote conversation trajectory data therefrom. The data from different scenarios are matched according to the medical service data (such as patient ID, medical event type, etc.) to form multiple remote conversation trajectory training data combinations. The matched data combinations are stored in the local database of the server for subsequent processing.

[0055] Step C120: for a target remote conversation trajectory training data combination among the multiple remote conversation trajectory training data combinations, a basic first interaction scenario model is used to perform graph feature representation on the first member remote conversation trajectory training data in the target remote conversation trajectory training data combination to generate first member graph feature representation data.

[0056] For each remote conversation trajectory training data combination, the server first regards it as a target combination, and uses the basic first interaction scene model and the basic second interaction scene model to perform graph feature representation on the first member and the second member in the combination, respectively.

[0057] For example, the server converts the remote conversation trajectory training data of the first member and the remote conversation trajectory training data of the second member in the first data combination into graph structures respectively. The basic first interaction scene model is used to perform multi-layer graph convolution operations on the graph structure of the first member to extract the first member graph feature representation data. The basic second interaction scene model is used to perform multi-layer graph convolution operations on the graph structure of the second member to extract the second member graph feature representation data.

[0058] Step C130 , using the basic second interaction scenario model to perform graph feature representation on the second member remote conversation trajectory training data in the target remote conversation trajectory training data combination, to generate second member graph feature representation data.

[0059] Step C140 : Using the plurality of remote conversation trajectory training data combinations as the target remote conversation trajectory training data combinations, respectively, to generate a plurality of first member graph feature representation data and a plurality of second member graph feature representation data.

[0060] The server repeatedly applies the above process to all remote conversation trajectory training data combinations, and generates corresponding first member graph feature representation data and second member graph feature representation data for each combination.

[0061] In detail, the server traverses all remote session trajectory training data combinations and repeatedly performs the above graph structure construction and feature extraction operations until all combinations are processed.

[0062] Step C150, based on the network optimization goal of minimizing the feature distance of the target member graph feature representation data combination and maximizing the feature distance of other member graph feature representation data combinations, update the network parameter information of the basic first interaction scenario model and the network parameter information of the basic second interaction scenario model, generate the first interaction scenario model and the second interaction scenario model, the first member graph feature representation data and the second member graph feature representation data covered by the target member graph feature representation data combination are generated based on the same remote conversation trajectory training data combination, and the first member graph feature representation data and the second member graph feature representation data covered by the other member graph feature representation data combinations are not generated based on the same remote conversation trajectory training data combination.

[0063] In this embodiment, the server uses the generated multiple groups of first member graph feature representation data and second member graph feature representation data to optimize the network parameters of the basic first interaction scenario model and the basic second interaction scenario model. The optimization goal is to minimize the distance between the first member graph feature representation data and the second member graph feature representation data from the same remote session trajectory training data combination in the feature space, and maximize the distance between the graph feature representation data from different combinations.

[0064] In detail, the server defines a loss function that simultaneously considers minimizing the feature distance of the target member graph feature representation data combination and maximizing the feature distance of the other member graph feature representation data combination. Then, a gradient descent or other optimization algorithm is used to update the network parameters of the basic first interaction scene model and the basic second interaction scene model to minimize the loss function. The above process is repeated multiple times until the loss function converges to a stable value, indicating that the model has been fully trained.

[0065] After multiple iterations of training, the server obtains the optimized first interaction scenario model and the second interaction scenario model. The first interaction scenario model and the second interaction scenario model are now able to better convert the remote conversation trajectory data in their respective scenarios into a unified target medical service feature domain. The trained models are saved to a local or remote repository for use in subsequent knowledge point prediction tasks. At the same time, the server can also further evaluate and test the first interaction scenario model and the second interaction scenario model to ensure that their performance and accuracy meet the expected requirements.

[0066] In a possible implementation, if the remote conversation trajectory data originating from the first medical interaction scenario is sparse feature data, and the remote conversation trajectory data originating from the second medical interaction scenario is dense feature data, the method further includes: Step D110, obtain x second medical service knowledge point annotation data and a first guiding data sample including a knowledge point placeholder, wherein the second medical service knowledge point annotation data is used to describe the medical service knowledge points of the remote conversation trajectory data originating from the second medical interaction scenario, and the first guiding data sample represents a deep learning network to generate multiple dense feature semantic data, and the multiple medical image feature fragments described by the multiple dense feature semantic data constitute the target dense feature data related to the medical service knowledge point annotation data corresponding to the knowledge point placeholder, and x is a positive integer.

[0067] Step D120: Load the x second medical service knowledge point annotation data into the knowledge point placeholder in the first guide data sample to generate first guide data.

[0068] Step D130: Based on the first guide data, the first remote conversation trajectory training data is generated by using the deep learning network, where the first remote conversation trajectory training data is dense feature semantic data related to the x second medical service knowledge point annotation data.

[0069] Step D140: output the x second medical service knowledge point annotation data as first medical service knowledge point annotation data of the first remote conversation trajectory training data.

[0070] In this scenario, the server is faced with two medical interaction scenarios with significant differences in data sparsity and density. In order to use the rich dense feature data in the second medical interaction scenario to enhance the training effect of the sparse feature data in the first medical interaction scenario, the server adopts a data enhancement and conversion method.

[0071] For example, the server obtains x second medical service knowledge point annotation data from the database of the second medical interaction scenario, which describes in detail the medical service knowledge points in the remote conversation trajectory, such as "diabetes management", "hypertension control", etc. At the same time, the server also has a predefined first guide data sample, which is a template for guiding the deep learning network on how to generate dense feature semantic data related to a specific knowledge point.

[0072] In detail, the server extracts x second medical service knowledge point annotation data from the database of the second medical interaction scenario through SQL query or other data retrieval methods. Then, the first guide data sample is loaded from the local storage or remote warehouse, and the first guide data sample includes multiple placeholders, wherein the knowledge point placeholder is used for subsequent insertion of specific medical service knowledge point annotation data.

[0073] Next, the server loads the acquired x second medical service knowledge point annotation data into the knowledge point placeholder position in the first guide data sample, thereby generating the first guide data containing specific knowledge point information.

[0074] For example, the server traverses x second medical service knowledge point annotation data and inserts them one by one into the knowledge point placeholders in the first guide data sample. After the insertion is completed, the server verifies the format of the first guide data to ensure that all placeholders are correctly filled and the data format meets the requirements of subsequent processing.

[0075] Then, the server uses a pre-trained deep learning network (such as a generative adversarial network GAN, a variational autoencoder VAE, etc.) to process the first guided data to generate dense feature semantic data related to the x second medical service knowledge point annotation data. These data are richer and denser in features, similar to the remote conversation trajectory data in the second medical interaction scenario.

[0076] For example, the server loads a pre-trained deep learning network model, passes the first guide data as input to the deep learning network, and generates dense feature semantic data based on the input first guide data. The data is similar in form to the remote conversation trajectory training data, but richer and denser in features. Thus, the generated dense feature semantic data is organized into the same format and structure as the first remote conversation trajectory training data.

[0077] Finally, the server directly outputs the x second medical service knowledge point annotation data originally used to generate the first guidance data as the first medical service knowledge point annotation data of the first remote conversation trajectory training data, and these annotation data are used in subsequent training and evaluation processes.

[0078] For example, the server associates x second medical service knowledge point annotation data with the generated dense feature semantic data (i.e., the first remote session trajectory training data), and then outputs the associated data in a standard format to the server's local storage or remote warehouse for subsequent use.

[0079] Through the above steps, the rich and dense feature data knowledge in the second medical interaction scenario is introduced into the sparse remote conversation trajectory data training in the first medical interaction scenario, thereby enhancing the diversity and richness of the training data, which is expected to improve the accuracy and generalization ability of the medical service knowledge point prediction model.

[0080] In a possible implementation, if the first guidance data sample also represents that the deep learning network generates other medical service knowledge point annotation data related to the target dense feature data, then step D140 includes: outputting the x second medical service knowledge point annotation data and the other medical service knowledge point annotation data as the first medical service knowledge point annotation data of the first remote session trajectory training data.

[0081] In this scenario, the server not only uses the dense feature data in the second medical interaction scenario to generate the first remote conversation trajectory training data related to specific knowledge points, but the first guide data sample also instructs the deep learning network to generate additional medical service knowledge point annotation data associated with the target dense feature data, which means that in addition to the original x second medical service knowledge point annotation data, the deep learning network can also automatically discover and annotate other potential related knowledge points.

[0082] For example, the server first confirms the design of the first guiding data sample, which not only includes a placeholder for inserting specific knowledge point annotation data, but also has instructions or structures for guiding the deep learning network to generate additional relevant knowledge point annotation data.

[0083] In detail, the server reviews the document or metadata of the first guide data sample to confirm that it can support the function of generating additional medical service knowledge point annotation data. If necessary, the server may fine-tune or configure the first guide data sample to ensure that it meets this requirement.

[0084] The server loads the x second medical service knowledge point annotation data into the first guide data sample and triggers the deep learning network to start processing. During the processing, the deep learning network not only generates target dense feature data based on the given knowledge points, but also automatically explores and generates other medical service knowledge point annotation data associated with these data.

[0085] In detail, the server fills the x second medical service knowledge point annotation data into the corresponding placeholders of the first guide data sample. The deep learning network is started, and the filled first guide data is passed to the network as input. When processing the input data, the deep learning network not only generates target dense feature data related to the given knowledge point (i.e., the first remote session trajectory training data), but also automatically discovers and annotates other potential related medical service knowledge points based on the patterns and features in the data.

[0086] After the deep learning network is processed, the server collects all the generated medical service knowledge point annotation data, including the original x second medical service knowledge point annotation data and the additional knowledge point annotation data automatically generated by the network. Then, the server outputs these data as the first medical service knowledge point annotation data of the first remote conversation trajectory training data.

[0087] For example, the server extracts all generated medical service knowledge point annotation data from the output of the deep learning network, which may exist in the form of a list, array, or other data structure. The server integrates the original x second medical service knowledge point annotation data with the additional knowledge point annotation data automatically generated by the network to form a complete first medical service knowledge point annotation data set. Finally, the server outputs the integrated first medical service knowledge point annotation data set in an appropriate format to local storage or a remote warehouse for use in subsequent training and evaluation processes.

[0088] Through this process, not only the dense feature data in the second medical interaction scenario is utilized to enhance the richness of the first remote conversation trajectory training data, but also additional potential knowledge points are discovered through the automatic exploration capability of the deep learning network, further improving the diversity and accuracy of the training data.

[0089] In a possible implementation, the first guidance data sample further characterizes that when the x second medical service knowledge point annotation data are contradictory, the deep learning network intercepts and generates first remote conversation trajectory training data related to the target dense feature data, and the method further includes: extracting the x second medical service knowledge point annotation data from y second medical service knowledge point annotation data, where y is an integer greater than x.

[0090] In this embodiment, in this scenario, the data processed by the server is more complex because there may be contradictions or inconsistencies between the second medical service knowledge point annotation data. In order to deal with this situation, the first guide data sample is designed to guide the deep learning network to suspend the generation of the first remote session trajectory training data when a data contradiction is detected, and the server needs to carefully select x non-contradictory annotation data from a larger data set (y annotation data) for training.

[0091] The server has a data set containing y annotated data of the second medical service knowledge points, where y is an integer greater than x. The data set is much larger than the x annotated data originally used to generate the first guide data, providing more selection space to avoid data contradictions. For example, the server retrieves the complete y annotated data sets from the database of the second medical interaction scenario and stores them in a local or remote warehouse for subsequent processing.

[0092] Before extracting x out of y labeled data, the server needs to clean and detect contradictions in the entire data set. This is to ensure that the selected x labeled data are logically consistent and do not contradict each other. For example, the server performs a series of data cleaning operations, such as removing duplicates, correcting incorrect annotations, filling missing values, etc., to improve data quality. Then, specialized algorithms or logic are used to detect possible contradictions in the data set, which may involve comparing information between different annotations to check whether they conflict or are inconsistent with each other.

[0093] After confirming that there are no contradictions or all contradictions have been resolved in the y labeled data sets, the server starts to extract x labeled data from them to generate the first remote session trajectory training data.

[0094] For example, the server may use random sampling or sampling based on specific strategies (such as the importance of knowledge points, data diversity, etc.) to extract x from y labeled data. After the extraction is completed, the server may verify the selected x labeled data again to ensure that there are no contradictions between them and that they meet the training requirements.

[0095] Using the extracted x non-contradictory annotated data, the server generates the first guide data and triggers the deep learning network to process. During this process, the first guide data sample monitors the operation of the deep learning network, and once a contradiction in the annotated data is detected (although it should not occur in theory in this scenario, as a preventive measure), the process of generating the first remote session trajectory training data is intercepted.

[0096] For example, the server generates the first guide data according to the previous steps and loads x non-contradictory annotated data into it. The deep learning network is started and the first guide data is passed in for processing. In the process of generating the target dense feature data, the deep learning network monitors the consistency of the annotated data according to the instructions of the first guide data sample.

[0097] Although in this specific scenario, the server has taken measures to ensure that there are no contradictions between the selected x labeled data, as a robust design, the first guide data sample still contains the logic of contradiction interception. If (theoretically unlikely) the deep learning network detects a contradiction during processing, it will suspend the generation of the first remote session trajectory training data according to the instructions of the first guide data sample.

[0098] If the deep learning network does detect a contradiction during the generation process (although this is unlikely to happen in the current scenario), it will suspend operation according to the instructions of the first guide data sample and send an error report to the server. After receiving the error report, the server will conduct further investigation and processing, which may include rechecking the consistency of the labeled data, adjusting the parameters or logic of the deep learning network, etc.

[0099] However, in most cases, since the server has cleaned and detected contradictions in the y labeled data in advance and selected x non-contradictory labeled data to generate the first guide data, the deep learning network is unlikely to encounter contradictions in the labeled data during the processing process. In this way, the server can smoothly use the deep learning network to generate high-quality first remote conversation trajectory training data for subsequent medical service knowledge point prediction tasks.

[0100] In a possible implementation, step S130 includes: Step S131, based on the first graph feature representation data, using an adaptation network to perform transformation to generate first adapted graph feature representation data, wherein the adaptation network is used to adapt the graph feature representation data from the target medical service feature domain to the network feature domain, and the network feature domain is the medical service feature domain adapted by the medical service knowledge point prediction network.

[0101] Step S132: Based on the first adaptation graph feature representation data, the basic medical service knowledge point prediction network is used to perform knowledge point prediction to generate predicted medical service knowledge points.

[0102] The method further comprises: Step E110, obtaining the remote conversation trajectory data to be predicted from the second medical interaction scenario.

[0103] Step E120 , using the first graph generation model to perform graph feature representation on the remote conversation trajectory data to be predicted, to generate second graph feature representation data.

[0104] Step E130: Based on the second graph feature representation data, the adaptation network is used to perform transformation to generate second adapted graph feature representation data.

[0105] Step E140: Based on the second adaptation graph feature representation data, the medical service knowledge point prediction network is used to perform knowledge point prediction to generate the medical service knowledge points of the remote conversation trajectory data to be predicted.

[0106] In this scenario, the server has trained and optimized multiple models through the previous steps, including the first graph generation model (including the first interaction scenario model and the second interaction scenario model), the adaptation network, and the basic medical service knowledge point prediction network. Now, the server needs to use these models to process the new remote session trajectory data to generate predicted medical service knowledge points.

[0107] The server first processes the first remote conversation trajectory training data originating from the first medical interaction scenario, and generates predicted medical service knowledge points through a series of conversion and prediction steps.

[0108] In detail, the first interactive scenario model can be used to perform graph feature representation on the first remote conversation trajectory training data to generate first graph feature representation data, which are in the target medical service feature domain. Then, the first graph feature representation data is transformed using an adaptation network to adapt it from the target medical service feature domain to the network feature domain. This step is to ensure that the data format matches the requirements of the basic medical service knowledge point prediction network. The adapted data is called the first adapted graph feature representation data. Finally, the first adapted graph feature representation data is input into the basic medical service knowledge point prediction network, and the basic medical service knowledge point prediction network outputs the predicted medical service knowledge points.

[0109] Next, the server obtains the remote conversation trajectory data to be predicted from the second medical interaction scenario, and repeats similar processing steps to generate predicted medical service knowledge points for the data.

[0110] For example, the second interaction scenario model can be used to perform graph feature representation on the remote conversation trajectory data to be predicted, and generate second graph feature representation data. Similar to the first remote conversation trajectory training data, these data are also in the target medical service feature domain. Similarly, the server uses the adaptation network to transform the second graph feature representation data, adapting it from the target medical service feature domain to the network feature domain, and generating second adapted graph feature representation data. Finally, the second adapted graph feature representation data is input into the basic medical service knowledge point prediction network, and the basic medical service knowledge point prediction network outputs the predicted medical service knowledge points for the remote conversation trajectory data to be predicted.

[0111] Assume that the server obtains a first remote conversation trajectory training data about "respiratory tract infection management". The server first uses the first interaction scenario model to convert it into graph feature representation data, and then uses the adaptation network to adapt this data from the target medical service feature domain to the network feature domain. Finally, the basic medical service knowledge point prediction network receives the adapted data and outputs the predicted medical service knowledge point "respiratory tract infection management".

[0112] Subsequently, the server obtains a new remote conversation trajectory data from the second medical interaction scenario, which involves the regular follow-up record of a hypertensive patient. The server uses the second interaction scenario model to represent its graph features and adapts the data to the network feature domain through the adaptation network. Finally, the basic medical service knowledge point prediction network outputs the predicted medical service knowledge point "hypertension management", which provides valuable reference information for doctors.

[0113] As a result, data from different medical interaction scenarios can be flexibly processed, and the trained model can be used to generate accurate predictive medical service knowledge points, thereby improving the efficiency and quality of telemedicine services.

[0114] In a possible implementation, the method further includes: Step F110 , obtaining second remote conversation trajectory training data originating from the first medical interaction scenario, wherein the second remote conversation trajectory training data carries third medical service knowledge point annotation data.

[0115] Step F120 : Based on the second remote conversation trajectory training data, the first graph generation model is used to perform graph feature representation on the second remote conversation trajectory training data to generate third graph feature representation data.

[0116] Step F130: Based on the third graph feature representation data, a basic adaptation network is used to perform transformation to generate adapted graph feature representation training data.

[0117] Step F140 , based on the adaptation graph feature representation training data, using the basic medical service knowledge point prediction network to perform knowledge point prediction and generate knowledge point prediction data.

[0118] Step F150, based on the error between the knowledge point prediction data and the third medical service knowledge point annotation data, locking the network parameter information of the basic medical service knowledge point prediction network, updating the network parameter information of the basic adaptation network, and generating the adaptation network.

[0119] In this embodiment, the server retrieves new remote conversation trajectory data from the database of the first medical interaction scenario. These remote conversation trajectory data are called second remote conversation trajectory training data, and each piece of data carries the third medical service knowledge point annotation data.

[0120] For example, the server extracts the second remote conversation trajectory training data and its corresponding third medical service knowledge point annotation data from the database through a preset data retrieval interface or SQL query. The extracted data is temporarily stored in the local storage or memory of the server for subsequent processing.

[0121] The server uses the trained first graph generation model (especially the first interaction scenario model therein, because the data originates from the first medical interaction scenario) to perform graph feature representation on the second remote conversation trajectory training data to generate third graph feature representation data.

[0122] For example, the second remote conversation trajectory training data can be input into the first graph generation model. The first graph generation model extracts key features from the second remote conversation trajectory training data through multi-layer graph convolution and other operations, and generates third graph feature representation data. These second remote conversation trajectory training data are now in the target medical service feature domain, but have not yet been adapted to the network feature domain required by the basic medical service knowledge point prediction network.

[0123] Next, the basic adaptation network is used to transform the third graph feature representation data, adapting it from the target medical service feature domain to the network feature domain to generate adapted graph feature representation training data.

[0124] For example, the third graph feature representation data can be input into the basic adaptation network. The adaptation network transforms the data from the target feature domain to the network feature domain through a series of transformation operations (such as linear transformation, nonlinear activation, etc.). The transformed data (adapted graph feature representation training data) now matches the input requirements of the basic medical service knowledge point prediction network.

[0125] The server uses the basic medical service knowledge point prediction network to predict the knowledge points of the adaptation graph feature representation training data and generate knowledge point prediction data. For example, the server inputs the adaptation graph feature representation training data into the basic medical service knowledge point prediction network, extracts the features related to the medical service knowledge points through multi-layer convolution, loop or attention mechanism, and generates prediction results. The prediction results are recorded as knowledge point prediction data, that is, predicted medical service knowledge points.

[0126] The server evaluates the accuracy of the prediction by comparing the error between the knowledge point prediction data and the third-party medical service knowledge point annotation data, and optimizes the network parameter information of the basic adaptation network accordingly to generate the final adaptation network.

[0127] For example, the server calculates the error (such as cross entropy loss) between the knowledge point prediction data and the third medical service knowledge point annotation data. Using the back propagation algorithm, the server backpropagates the error to the basic adaptation network and updates its network parameter information (such as weights and biases). This process may require multiple iterations until the error converges to an acceptable range. Finally, the server generates an optimized adaptation network that can more accurately adapt the graph feature representation data from the target medical service feature domain to the network feature domain.

[0128] Therefore, not only the new training data is used to enhance the generalization ability of the model, but also the accuracy of knowledge point prediction is improved by optimizing the adaptation network, which is of great significance for improving the intelligence level of telemedicine services and patient experience.

[0129] In a possible implementation, the method further includes: Step G110, obtaining a second guide data sample and an adaptation request, wherein the second guide data sample includes a graph feature representation data placeholder and an adaptation request placeholder, and the adaptation request represents that the basic medical service knowledge point prediction network generates the knowledge point prediction data.

[0130] Step G120, loading the adaptation graph feature representation training data into the graph feature representation data placeholder in the second guide data sample, loading the adaptation request into the adaptation request placeholder in the second guide data sample, and generating second guide data.

[0131] Step F140 includes: based on the second guidance data, using the basic medical service knowledge point prediction network to perform knowledge point prediction to generate the knowledge point prediction data.

[0132] In this embodiment, the server obtains the second guidance data sample from the configuration file, database or remote service, and receives an adaptation request, which clearly indicates that the server should use the basic medical service knowledge point prediction network to generate knowledge point prediction data.

[0133] For example, the server retrieves a second guide data sample from local storage, database or remote service through a predefined interface or method call. The sample is a template for guiding the subsequent data loading and prediction process. At the same time, the server receives an adaptation request, which may be transmitted through an API, message queue, command line parameter or other communication mechanism. The adaptation request contains the instructions and parameters required to perform the prediction task.

[0134] The server loads the adapted graph feature representation training data and the adaptation request into the corresponding placeholders in the second guide data sample, respectively, to generate a complete second guide data. For example, the server first parses the second guide data sample and identifies the graph feature representation data placeholder and the adaptation request placeholder therein. Next, the server loads the adapted graph feature representation training data generated previously through a series of processing steps into the graph feature representation data placeholder. These data are obtained after the remote session trajectory data is transformed by the graph generation model and the adaptation network, and they are in a feature domain suitable for knowledge point prediction. At the same time, the server loads the received adaptation request into the adaptation request placeholder. The adaptation request contains specific instructions and parameters required to perform the prediction task, such as the predicted target knowledge point type, the predicted confidence threshold, etc. After the above operations, the server generates the second guide data containing the adapted graph feature representation training data and the adaptation request, which organizes all the information required for the prediction task in a structured manner.

[0135] Next, the server uses the basic medical service knowledge point prediction network to process the adaptation graph feature representation training data in the second guidance data to generate knowledge point prediction data.

[0136] For example, the server passes the second guidance data as input to the basic medical service knowledge point prediction network. After receiving the input, the basic medical service knowledge point network first parses the adaptation graph feature representation training data and the adaptation request in the second guidance data. Then, the basic medical service knowledge point network processes the adaptation graph feature representation training data according to the instructions and parameters in the adaptation request. The processing process may include multiple machine learning tasks such as feature extraction, classification, and regression.

[0137] Finally, the basic medical service knowledge point network outputs prediction results, namely, knowledge point prediction data, which may exist in the form of text, labels, probability distribution or other forms, depending on the requirements of the prediction task and the output design of the network.

[0138] For example, suppose the server is processing a remote conversation trajectory data about "follow-up of heart disease patients". After graph feature representation and adaptation network transformation, the server obtains the adapted graph feature representation training data. At this time, the server receives an adaptation request, requiring the use of a specific basic medical service knowledge point prediction network to predict the medical service knowledge points related to the conversation trajectory.

[0139] The server obtains the second guidance data sample and the adaptation request according to the above steps, and loads the adaptation graph feature representation training data and the adaptation request into the sample to generate the second guidance data. Then, the server inputs the second guidance data into the basic medical service knowledge point prediction network. The network processes the adaptation graph feature representation training data according to the instructions in the adaptation request, and finally outputs the predicted medical service knowledge point "Precautions for heart disease drug treatment" as the knowledge point prediction data. This result helps doctors or patients better understand the content of the conversation trajectory and take corresponding management measures.

[0140] In a possible implementation, step S140 includes: updating the network parameter information of the basic medical service knowledge point prediction network based on the error between the predicted medical service knowledge point and the first medical service knowledge point annotation data, generating a medical service knowledge point prediction network, and updating the network parameter information of the adaptation network based on the error between the predicted medical service knowledge point and the first medical service knowledge point annotation data, generating an updated adaptation network.

[0141] In this embodiment, the server first calculates the error between the predicted medical service knowledge point and the first medical service knowledge point annotation data, which is an important indicator to measure the difference between the predicted result and the actual annotation. For example, the server traverses all training samples, and for each sample, compares the predicted medical service knowledge point with the corresponding first medical service knowledge point annotation data. Use an appropriate error measurement method (such as cross entropy loss, mean square error, etc.) to calculate the difference between the two and obtain the prediction error of the sample. The prediction errors of all samples are summarized or averaged to obtain the overall training error.

[0142] Next, the calculated prediction error is used to update the network parameter information (such as weights and biases) of the basic medical service knowledge point prediction network to optimize the prediction performance of the network. For example, the back propagation algorithm is used to back propagate the prediction error from the output layer of the network to the input layer layer by layer. During the back propagation process, the gradient of each layer of network parameters (i.e., the derivative of the error with respect to the parameter) is calculated according to the chain rule. The value of each layer of network parameters is updated using gradient descent or other optimization algorithms to reduce the prediction error.

[0143] The above process is repeated multiple times (i.e., multiple training iterations are performed) until the prediction error converges to an acceptable range or reaches a preset training round.

[0144] In addition to updating the basic medical service knowledge point prediction network, the server also needs to update the network parameter information of the adaptation network based on the prediction error. This is because the adaptation network is responsible for adapting the graph feature representation data from the target medical service feature domain to the network feature domain, and its performance directly affects the subsequent knowledge point prediction results.

[0145] Similar to updating the basic medical service knowledge point prediction network, the server first calculates the contribution error of the adaptation network during the training process, which can be obtained by back-propagating the error of the basic prediction network to the adaptation network layer. Then, the network parameter information of the adaptation network is updated using the same back-propagation and optimization algorithm as that used to update the basic prediction network.

[0146] It should be noted that when updating the adaptation network, the server may need to consider the combined impact of the adaptation error and the basic prediction error to find the best parameter update strategy.

[0147] After multiple iterations of training, the server generated an optimized medical service knowledge point prediction network and an updated adaptation network, both of which now have better prediction performance and generalization capabilities.

[0148] For example, the server saves the trained medical service knowledge point prediction network and the network parameter information of the updated adaptation network, which can be used in subsequent prediction tasks to generate accurate medical service knowledge point prediction results. The server can also deploy the trained model to the actual production environment to provide real-time medical service knowledge point prediction services for doctors and patients.

[0149] For example, assume that the server is processing remote session trajectory data about "diabetes management". After a series of processing steps, the server generates a predicted medical service knowledge point "diabetes diet guidance" and compares it with the corresponding first medical service knowledge point annotation data. The calculated prediction error shows that there is a certain difference between the prediction result and the actual annotation.

[0150] In order to optimize the prediction performance, the server uses the back propagation algorithm and the gradient descent optimization algorithm to update the network parameter information of the basic medical service knowledge point prediction network and the adaptation network. After multiple iterations of training, the server generated an optimized medical service knowledge point prediction network and an updated adaptation network. These two networks can now more accurately predict medical service knowledge points related to diabetes management, providing more reliable reference information for doctors and patients.

[0151] In a possible implementation, the method further includes: Step H110, obtaining auxiliary medical diagnosis data of the first remote conversation trajectory training data.

[0152] Step H120, extracting features from the auxiliary medical diagnosis data to generate auxiliary medical diagnosis features.

[0153] Step S130 includes: based on the first graph feature representation data and the auxiliary medical diagnosis features, using the basic medical service knowledge point prediction network to perform knowledge point prediction to generate the predicted medical service knowledge point.

[0154] In this embodiment, in this scenario, the server not only uses the graph feature representation data of the first remote session trajectory training data to predict knowledge points, but also additionally obtains auxiliary medical diagnosis data associated with the graph feature representation data. These auxiliary medical diagnosis data provide additional contextual information, which helps to improve the accuracy and comprehensiveness of knowledge point prediction.

[0155] For example, auxiliary medical diagnosis data associated with the first remote session trajectory training data are retrieved from the database of the first medical interaction scenario. The auxiliary medical diagnosis data may include the patient's medical history, laboratory test results, imaging examination reports, etc.

[0156] Next, the server extracts features from the acquired auxiliary medical diagnosis data and converts it into a feature representation suitable for knowledge point prediction.

[0157] For example, for text-based auxiliary medical diagnosis data (such as medical history descriptions, diagnostic reports, etc.), the server may use natural language processing technology to perform operations such as word segmentation, stop word removal, and word embedding to convert the text into a numerical vector or feature matrix.

[0158] For image-based auxiliary medical diagnostic data (such as X-rays, CT scan images, etc.), the server may use image processing and computer vision technologies to extract features, such as edge detection, texture analysis, and convolutional neural network feature extraction.

[0159] For numerical auxiliary medical diagnosis data (such as laboratory test results, vital signs monitoring data, etc.), the server may directly use it as a feature, or perform preprocessing operations such as normalization and standardization.

[0160] After feature extraction, the server generates auxiliary medical diagnosis features, which will be used together with the first graph feature representation data for subsequent knowledge point prediction.

[0161] Finally, the first graph feature representation data and auxiliary medical diagnosis features are used as input, and the basic medical service knowledge point prediction network is used to predict knowledge points to generate predicted medical service knowledge points.

[0162] For example, the first graph feature representation data and auxiliary medical diagnosis features can be merged or concatenated to form a comprehensive feature representation containing rich contextual information, which is input into the basic medical service knowledge point prediction network. The network processes these features through multi-layer convolution, loop, attention and other mechanisms to extract key information related to medical service knowledge points. Finally, the network outputs the prediction result, that is, the predicted medical service knowledge point, which comprehensively considers the information of remote conversation trajectory data and auxiliary medical diagnosis data, and is therefore more accurate and comprehensive.

[0163] For example, assume that the server is processing a first remote conversation track training data about "hypertension management". In addition to the conversation track data itself, the server also retrieves auxiliary medical diagnosis data associated with the patient from the database, including the patient's medical history record, the most recent blood pressure monitoring results, and the electrocardiogram report.

[0164] The server first extracts features from the auxiliary medical diagnosis data, converts the text description into a numerical vector, converts the image report into a feature matrix, and normalizes the numerical results. Then, the server merges these auxiliary medical diagnosis features with the first image feature representation data to form a comprehensive feature representation.

[0165] Next, the server inputs this comprehensive feature representation into the basic medical service knowledge point prediction network. The network extracts key information through a multi-layer processing mechanism and outputs the prediction result "Adjustment of hypertension drug treatment plan", which comprehensively considers the information of conversation trajectory data and auxiliary medical diagnosis data, providing valuable reference opinions for doctors.

[0166] In a possible implementation, the method further includes: A third guide data sample and an extended training request are obtained, wherein the third guide data sample includes a graph feature representation data placeholder and an extended training request placeholder, and the extended training request represents that the basic medical service knowledge point prediction network generates the predicted medical service knowledge point.

[0167] The first graph feature representation data is loaded into the graph feature representation data placeholder in the third guide data sample, and the extended training request is loaded into the extended training request placeholder in the third guide data sample to generate third guide data.

[0168] Step S130 includes: based on the third guidance data, using the basic medical service knowledge point prediction network to perform knowledge point prediction and generate predicted medical service knowledge points.

[0169] In this embodiment, in this scenario, the server not only uses basic data and models to predict knowledge points, but also introduces a third guiding data sample to further guide and optimize the prediction process. The sample contains specific placeholders for receiving graph feature representation data and extended training requests, thereby ensuring that the prediction process can be expanded and optimized in the desired manner.

[0170] For example, the server obtains a third guide data sample from a configuration file, a database, or a remote service, and receives an extended training request, which indicates that the server should consider certain additional training conditions or constraints when performing knowledge point prediction.

[0171] Next, the first graph feature representation data and the extended training request are respectively loaded into corresponding placeholders in the third guide data sample to generate complete third guide data.

[0172] For example, first parse the third guide data sample and identify the graph feature representation data placeholder and the extended training request placeholder. Then, the server loads the first graph feature representation data obtained through a series of processing steps into the graph feature representation data placeholder. These data are obtained after the remote session trajectory data is converted by the graph generation model, and they contain rich session information. At the same time, the server loads the received extended training request into the extended training request placeholder. The request may specify certain specific prediction objectives, constraints or optimization strategies. After the above operations, the server generates the third guide data containing the first graph feature representation data and the extended training request, which structuredly organizes all the information and instructions required for the prediction task.

[0173] Finally, the basic medical service knowledge point prediction network is used to process the first graph feature representation data in the third guide data, and the instructions and conditions in the extended training request are considered to generate predicted medical service knowledge points.

[0174] For example, the third guide data can be passed as input to the basic medical service knowledge point prediction network. After receiving the input, the network first parses the first graph feature representation data and the extended training request in the third guide data. The network adjusts its prediction strategy according to the instructions and conditions in the extended training request, which may include changing the prediction target, applying specific optimization algorithms, adjusting model parameters, etc. Next, the network processes the first graph feature representation data, extracts features related to the medical service knowledge points, and generates prediction results based on these features and the conditions in the extended training request. Finally, the network outputs the predicted medical service knowledge point, which is not only based on the graph feature representation of the remote session trajectory data, but also fully considers the additional information and conditions in the extended training request.

[0175] For example, assume that the server is processing a first remote session trajectory data about "follow-up of heart disease patients". After the graph generation model conversion, the server obtains the first graph feature representation data. At this time, the server receives an extended training request, requiring special attention to the aspect of "drug treatment effect evaluation" when predicting knowledge points.

[0176] The server obtains the third guide data sample and the extended training request according to the above steps, and loads the first graph feature representation data and the extended training request into the sample to generate the third guide data. Then, the server inputs this third guide data into the basic medical service knowledge point prediction network. During the processing, the network pays special attention to the features related to the drug treatment effect, and adjusts the prediction strategy according to the conditions in the extended training request. In the end, the network outputs the predicted medical service knowledge point "Evaluation and Adjustment Recommendations of Drug Treatment Effects for Heart Disease", which is more in line with the specific requirements in the extended training request.

[0177] In a possible implementation, the third boot data sample further includes an auxiliary medical diagnosis data placeholder, and the method further includes: Acquire auxiliary medical diagnosis data of the first remote conversation trajectory training data.

[0178] The auxiliary medical diagnosis data is characterized to generate auxiliary medical diagnosis features.

[0179] The step of loading the first graph feature representation data into a graph feature representation data placeholder in the third guide data sample, and loading the extended training request into an extended training request placeholder in the third guide data sample, to generate third guide data, comprises: The first graph feature representation data is loaded into the graph feature representation data placeholder in the third guide data sample, the extended training request is loaded into the extended training request placeholder in the third guide data sample, and the auxiliary medical diagnosis feature is loaded into the auxiliary medical diagnosis data placeholder in the third guide data sample to generate third guide data.

[0180] In this embodiment, in this scenario, when processing the first remote session trajectory training data, the server not only considers its graph feature representation data, but also additionally obtains auxiliary medical diagnosis data associated with the session trajectory. In order to more effectively utilize this information, the server introduces a third guide data sample containing multiple placeholders, which are used to receive graph feature representation data, extended training requests, and auxiliary medical diagnosis features, respectively.

[0181] For example, the server retrieves auxiliary medical diagnosis data corresponding to the first remote session trajectory training data from the database of the first medical interaction scenario, and these data may include the patient's medical records, laboratory test results, imaging examination reports, etc.

[0182] Then, the server extracts and converts the acquired auxiliary medical diagnosis data to generate auxiliary medical diagnosis features suitable for knowledge point prediction.

[0183] For example, for text-based auxiliary medical diagnosis data (such as medical records, diagnostic reports, etc.), the server uses natural language processing technology to perform operations such as word segmentation, stop word removal, and word embedding to convert the text into a numerical vector or feature matrix.

[0184] For image-based auxiliary medical diagnostic data (such as X-rays, CT scan images, etc.), the server uses image processing and computer vision technology to extract features and generate image feature vectors.

[0185] For numerical auxiliary medical diagnosis data (such as laboratory test results, vital signs monitoring data, etc.), the server may directly use it as a feature, or perform preprocessing operations such as normalization and standardization.

[0186] After the above processing, the server generates auxiliary medical diagnosis features, which will be used in the subsequent knowledge point prediction process.

[0187] Finally, the server loads the first graph feature representation data, the extended training request, and the auxiliary medical diagnosis features into corresponding placeholders in the third guide data sample, respectively, to generate complete third guide data.

[0188] For example, the server first parses the third guide data sample and identifies the graph feature representation data placeholder, the extended training request placeholder, and the auxiliary medical diagnosis data placeholder. Then, the server loads the first graph feature representation data into the graph feature representation data placeholder. These data are obtained after the remote session trajectory data is converted by the graph generation model, and contain the graph structure information of the session. At the same time, the server loads the extended training request into the extended training request placeholder. The request may specify additional training conditions, optimization goals, or constraints to guide the knowledge point prediction process. Finally, the server loads the auxiliary medical diagnosis features into the auxiliary medical diagnosis data placeholder. These features provide additional information about the patient's health status, which helps to improve the accuracy and comprehensiveness of the knowledge point prediction. After the above operations, the server generates the third guide data containing all the necessary information, which organizes all the data and instructions required for the prediction task in a structured manner.

[0189] For example, assume that the server is processing a first remote conversation track training data about "diabetes management". In addition to the conversation track data itself, the server also retrieves auxiliary medical diagnosis data associated with the patient from the database, including the most recent blood glucose monitoring results, glycosylated hemoglobin level, and the doctor's diagnosis opinion.

[0190] The server first performs feature representation on the auxiliary medical diagnosis data, converts the text-type diagnostic opinions into numerical vectors, and normalizes the numerical blood glucose monitoring results and glycosylated hemoglobin levels. Then, the server loads these auxiliary medical diagnosis features together with the first graph feature representation data and the extended training request (for example, requiring special attention to the prediction of "blood glucose control strategy") into the third guide data sample to generate the third guide data.

[0191] Finally, the server uses the basic medical service knowledge point prediction network to process the third guidance data, comprehensively considers the graph structure information of the session trajectory, the specific requirements in the extended training request, and the patient health status information in the auxiliary medical diagnosis data, and generates the predicted medical service knowledge point "Personalized Diabetes Diet and Exercise Plan", providing valuable reference opinions for doctors and patients.

[0192] Figure 2 The hardware structure of the medical Internet of Things service system 100 provided in the embodiment of the present application for implementing the above-mentioned medical Internet of Things-based vital sign data monitoring method is shown, as shown in FIG. Figure 2 As shown, the medical Internet of Things service system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0193] In one possible design, the medical Internet of Things service system 100 can be a single server or a server group. The server group can be centralized or distributed (for example, the medical Internet of Things service system 100 can be a distributed system). In some embodiments, the medical Internet of Things service system 100 can be local or remote. For example, the medical Internet of Things service system 100 can access information and / or data stored in a machine-readable storage medium 120 via a network. For another example, the medical Internet of Things service system 100 can be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In some embodiments, the medical Internet of Things service system 100 can be implemented on a medical Internet of Things service system. By way of example only, the medical Internet of Things service system may include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, etc. or any aggregation thereof.

[0194] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data acquired from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions used by the medical Internet of Things service system 100 to execute or use to complete the exemplary methods described in this application.

[0195] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the vital sign data monitoring method based on the medical Internet of Things in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0196] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned medical Internet of Things service system 100. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0197] In addition, an embodiment of the present application also provides a readable storage medium, in which computer executable instructions are set. When the processor executes the computer executable instructions, the above-mentioned vital sign data monitoring method based on the medical Internet of Things is implemented.

[0198] It should be noted that in order to simplify the description disclosed in this application and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof. Similarly, it should be noted that in order to simplify the description disclosed in this application and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof.

Claims

1. A method for monitoring vital sign data based on medical Internet of Things, characterized in that: The method comprises: Acquire first remote conversation trajectory training data originating from a first medical interaction scenario, wherein the first remote conversation trajectory training data carries first medical service knowledge point annotation data, and the first medical service knowledge point annotation data represents actual medical service knowledge points of the first remote conversation trajectory training data; Using a first graph generation model to perform graph feature representation on the first remote conversation trajectory training data to generate first graph feature representation data, the first graph generation model is used to convert the remote conversation trajectory data originating from the first medical interaction scenario into a target medical service feature domain, the target medical service feature domain is a medical service feature domain where the graph feature representation data corresponding to the remote conversation trajectory data originating from the second medical interaction scenario is located, the first medical interaction scenario and the second medical interaction scenario are different medical interaction scenarios, and the collection density of the remote conversation trajectory data originating from the first medical interaction scenario is less than the collection density of the remote conversation trajectory data originating from the second medical interaction scenario; Based on the first graph feature representation data, using the basic medical service knowledge point prediction network to predict knowledge points and generate predicted medical service knowledge points; Based on the error between the predicted medical service knowledge point and the first medical service knowledge point annotation data, the network parameter information of the basic medical service knowledge point prediction network is updated to generate a medical service knowledge point prediction network, so that the medical service knowledge point prediction network is used to predict the medical service knowledge points of the remote conversation trajectory data originating from the second medical interaction scenario, and then generates a corresponding vital sign data monitoring plan based on the medical service knowledge points of the remote conversation trajectory data originating from the second medical interaction scenario.

2. The method for monitoring vital sign data based on the medical Internet of Things according to claim 1, characterized in that: The method further comprises: Acquire remote conversation trajectory data to be predicted from the second medical interaction scenario; Performing graph feature representation on the remote conversation trajectory data to be predicted by using a second graph generation model to generate second graph feature representation data, wherein the second graph generation model is used to convert the remote conversation trajectory data originating from the second medical interaction scenario into the target medical service feature domain; Based on the second graph feature representation data, the medical service knowledge point prediction network is used to perform knowledge point prediction to generate the medical service knowledge points of the remote conversation trajectory data to be predicted.

3. The method for monitoring vital sign data based on medical Internet of Things according to claim 1, characterized in that: The first graph generation model includes a first interaction scenario model and a second interaction scenario model, the first interaction scenario model is used to convert the remote conversation trajectory data originated from the first medical interaction scenario into the target medical service feature domain, and the second interaction scenario model is used to convert the remote conversation trajectory data originated from the second medical interaction scenario into the target medical service feature domain; The using the first graph generation model to perform graph feature representation on the first remote conversation trajectory training data to generate first graph feature representation data includes: Performing graph feature representation on the first remote conversation trajectory training data using the first interaction scenario model to generate first graph feature representation data; The method further comprises: Acquire remote conversation trajectory data to be predicted from the second medical interaction scenario; Using the second interaction scenario model to perform graph feature representation on the remote conversation trajectory data to be predicted, to generate second graph feature representation data; Based on the second graph feature representation data, using the medical service knowledge point prediction network to perform knowledge point prediction to generate the medical service knowledge points of the remote conversation trajectory data to be predicted; Wherein, the method comprises: Acquire multiple remote conversation trajectory training data combinations, wherein the remote conversation trajectory training data combination includes first member remote conversation trajectory training data originating from the first medical interaction scenario and second member remote conversation trajectory training data originating from the second medical interaction scenario, and the first member remote conversation trajectory training data and the second member remote conversation trajectory training data originating from the same remote conversation trajectory training data combination are the same medical service data; For a target remote conversation trajectory training data combination among the multiple remote conversation trajectory training data combinations, using a basic first interaction scenario model to perform graph feature representation on first member remote conversation trajectory training data in the target remote conversation trajectory training data combination to generate first member graph feature representation data; Using the basic second interaction scenario model, performing graph feature representation on the second member remote conversation trajectory training data in the target remote conversation trajectory training data combination to generate second member graph feature representation data; Using the plurality of remote conversation trajectory training data combinations as the target remote conversation trajectory training data combinations, respectively, to generate a plurality of first member graph feature representation data and a plurality of second member graph feature representation data; According to the network optimization goal of minimizing the feature distance of the target member graph feature representation data combination and maximizing the feature distance of other member graph feature representation data combinations, the network parameter information of the basic first interaction scenario model and the network parameter information of the basic second interaction scenario model are updated to generate the first interaction scenario model and the second interaction scenario model, the first member graph feature representation data and the second member graph feature representation data covered by the target member graph feature representation data combination are generated based on the same remote conversation trajectory training data combination, and the first member graph feature representation data and the second member graph feature representation data covered by the other member graph feature representation data combinations are not generated based on the same remote conversation trajectory training data combination.

4. The method for monitoring vital sign data based on medical Internet of Things according to claim 1, characterized in that: If the remote conversation trajectory data originating from the first medical interaction scenario is sparse feature data, and the remote conversation trajectory data originating from the second medical interaction scenario is dense feature data, the method further includes: Obtain x second medical service knowledge point annotation data and first guide data samples including knowledge point placeholders, wherein the second medical service knowledge point annotation data is used to describe the medical service knowledge points of the remote conversation trajectory data originating from the second medical interaction scenario, and the first guide data samples represent that a deep learning network generates a plurality of dense feature semantic data, and the plurality of medical image feature fragments described by the plurality of dense feature semantic data constitute target dense feature data related to the medical service knowledge point annotation data corresponding to the knowledge point placeholders, and x is a positive integer; Loading the x second medical service knowledge point annotation data into the knowledge point placeholder in the first guide data sample to generate first guide data; Based on the first guide data, using the deep learning network to generate the first remote conversation trajectory training data, where the first remote conversation trajectory training data is dense feature semantic data related to the x second medical service knowledge point annotation data; The x second medical service knowledge point annotation data are output as first medical service knowledge point annotation data of the first remote conversation trajectory training data.

5. The method for monitoring vital sign data based on medical Internet of Things according to claim 4, characterized in that: If the first guide data sample further represents that the deep learning network generates other medical service knowledge point annotation data related to the target dense feature data, then outputting the x second medical service knowledge point annotation data as first medical service knowledge point annotation data of the first remote conversation trajectory training data includes: Outputting the x second medical service knowledge point annotation data and the other medical service knowledge point annotation data as first medical service knowledge point annotation data of the first remote conversation trajectory training data; and The first guide data sample further represents that when the labeled data of the x second medical service knowledge points are inconsistent, the deep learning network intercepts and generates first remote conversation trajectory training data related to the target dense feature data, and the method further includes: The x second medical service knowledge point annotation data are extracted from y second medical service knowledge point annotation data, where y is an integer greater than x.

6. The method for monitoring vital sign data based on medical Internet of Things according to claim 1, characterized in that: The step of predicting knowledge points based on the first graph feature representation data and using a basic medical service knowledge point prediction network to generate predicted medical service knowledge points includes: Based on the first graph feature representation data, transforming using an adaptation network to generate first adapted graph feature representation data, wherein the adaptation network is used to adapt the graph feature representation data from the target medical service feature domain to a network feature domain, wherein the network feature domain is a medical service feature domain adapted by the medical service knowledge point prediction network; Based on the first adaptation graph feature representation data, using the basic medical service knowledge point prediction network to perform knowledge point prediction to generate predicted medical service knowledge points; The method further comprises: Acquire remote conversation trajectory data to be predicted from the second medical interaction scenario; Using the first graph generation model to perform graph feature representation on the remote conversation trajectory data to be predicted, to generate second graph feature representation data; Based on the second graph feature representation data, transform using the adaptation network to generate second adapted graph feature representation data; Based on the second adaptation graph feature representation data, the medical service knowledge point prediction network is used to perform knowledge point prediction to generate the medical service knowledge points of the remote conversation trajectory data to be predicted.

7. The method for monitoring vital sign data based on medical Internet of Things according to claim 6, characterized in that: The method further comprises: Acquire second remote conversation trajectory training data originating from the first medical interaction scenario, wherein the second remote conversation trajectory training data carries third medical service knowledge point annotation data; Based on the second remote conversation trajectory training data, using the first graph generation model to perform graph feature representation on the second remote conversation trajectory training data to generate third graph feature representation data; Based on the third graph feature representation data, transforming using a basic adaptation network to generate adapted graph feature representation training data; Based on the adaptation graph feature representation training data, using the basic medical service knowledge point prediction network to perform knowledge point prediction and generate knowledge point prediction data; Based on the error between the knowledge point prediction data and the third medical service knowledge point annotation data, locking the network parameter information of the basic medical service knowledge point prediction network, updating the network parameter information of the basic adaptation network, and generating the adaptation network; Wherein, the method further comprises: Acquire a second guide data sample and an adaptation request, wherein the second guide data sample includes a graph feature representation data placeholder and an adaptation request placeholder, and the adaptation request represents that the basic medical service knowledge point prediction network generates the knowledge point prediction data; Loading the adaptation graph feature representation training data into a graph feature representation data placeholder in the second guide data sample, and loading the adaptation request into an adaptation request placeholder in the second guide data sample, to generate second guide data; The method of representing the training data based on the adaptation graph features and using the basic medical service knowledge point prediction network to perform knowledge point prediction and generate knowledge point prediction data includes: Based on the second guidance data, using the basic medical service knowledge point prediction network to perform knowledge point prediction to generate the knowledge point prediction data; The updating of network parameter information of the basic medical service knowledge point prediction network based on the error between the predicted medical service knowledge point and the first medical service knowledge point annotation data to generate a medical service knowledge point prediction network includes: Based on the error between the predicted medical service knowledge point and the first medical service knowledge point annotation data, the network parameter information of the basic medical service knowledge point prediction network is updated to generate a medical service knowledge point prediction network; and based on the error between the predicted medical service knowledge point and the first medical service knowledge point annotation data, the network parameter information of the adaptation network is updated to generate an updated adaptation network.

8. The method for monitoring vital sign data based on medical Internet of Things according to claim 1, characterized in that: The method further comprises: Acquire auxiliary medical diagnosis data of the first remote conversation trajectory training data; Extracting features from the auxiliary medical diagnosis data to generate auxiliary medical diagnosis features; The step of predicting knowledge points based on the first graph feature representation data and using a basic medical service knowledge point prediction network to generate predicted medical service knowledge points includes: Based on the first graph feature representation data and the auxiliary medical diagnosis features, the basic medical service knowledge point prediction network is used to perform knowledge point prediction to generate the predicted medical service knowledge point.

9. The method for monitoring vital sign data based on medical Internet of Things according to claim 1, characterized in that: The method further comprises: Acquire a third guide data sample and an extended training request, wherein the third guide data sample includes a graph feature representation data placeholder and an extended training request placeholder, and the extended training request represents that the basic medical service knowledge point prediction network generates the predicted medical service knowledge point; Loading the first graph feature representation data into a graph feature representation data placeholder in the third guide data sample, and loading the extended training request into an extended training request placeholder in the third guide data sample, to generate third guide data; The step of predicting knowledge points based on the first graph feature representation data and using a basic medical service knowledge point prediction network to generate predicted medical service knowledge points includes: Based on the third guidance data, using the basic medical service knowledge point prediction network to perform knowledge point prediction and generate predicted medical service knowledge points; The third guide data sample further includes an auxiliary medical diagnosis data placeholder, and the method further includes: Acquire auxiliary medical diagnosis data of the first remote conversation trajectory training data; Performing feature representation on the auxiliary medical diagnosis data to generate auxiliary medical diagnosis features; The step of loading the first graph feature representation data into a graph feature representation data placeholder in the third guide data sample, and loading the extended training request into an extended training request placeholder in the third guide data sample, to generate third guide data, comprises: The first graph feature representation data is loaded into the graph feature representation data placeholder in the third guide data sample, the extended training request is loaded into the extended training request placeholder in the third guide data sample, and the auxiliary medical diagnosis feature is loaded into the auxiliary medical diagnosis data placeholder in the third guide data sample to generate third guide data.

10. A medical Internet of Things service system, characterized in that: The medical Internet of Things service system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the vital sign data monitoring method based on the medical Internet of Things as described in any one of claims 1 to 9.

Images