Recommendation method, training method, electronic device and medium for medical care and rehabilitation service

Abstract

The application provides a medical care service recommendation method, a training method, an electronic device and a medium. Medical care services and at least two user labels are used as training data to train an intelligent agent. The intelligent agent can complete the construction of a user demand portrait, the calculation of a user label correlation degree, and medical care service prediction based on the calculation. The training of the intelligent agent can make the medical care service more accurate, efficient and intelligent.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to a medical, nursing and health service recommendation method, a training method, an electronic device and a medium. BACKGROUND

[0002] The needs of the elderly population for health promotion are increasingly diversified, and they not only need medical care but also need health and pension, and more diversified integrated services combining medical care, health and pension. However, the existing intelligent medical, nursing and health services face the problem of effective and efficient matching of supply and demand, which leads to insufficient aging, poor experience, lack of follow-up services, and difficulty in landing the "last mile", and affects the quality and efficiency of services. SUMMARY

[0003] The present application provides a medical, nursing and health service recommendation method, a training method, an electronic device and a medium, which can solve the problem of accurate, efficient and intelligent medical, nursing and health services.

[0004] To achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0005] In a first aspect, a training method of a medical, nursing and health service recommendation intelligent agent is provided, and the training method comprises:

[0006] obtaining training data, the training data comprising: medical, nursing and health services, and at least two user labels related to the medical, nursing and health services; and

[0007] training an intelligent agent using the training data, the intelligent agent being specifically used for: constructing a user demand portrait using the at least two user labels; calculating the correlation degree between the user labels; and predicting medical, nursing and health services based on the user demand portrait and the correlation degree.

[0008] Based on the above technical solutions, the medical, nursing and health services and the at least two user labels are used as training data to train the intelligent agent, and the intelligent agent can complete the construction of the user demand portrait, the calculation of the user label correlation degree, and the prediction of the medical, nursing and health services. The training of such an intelligent agent can make the medical, nursing and health services more accurate, efficient and intelligent.

[0009] In a possible design manner of the first aspect, the at least two user labels are used to construct a user demand portrait, specifically comprising:

[0010] performing feature fusion on the at least two user labels to obtain a high-dimensional feature vector;

[0011] obtaining at least two demand predictions based on the user labels; and

[0012] The user demand portrait is obtained by using a reinforcement learning method taking the high-dimensional feature vector and the demand prediction as inputs.

[0013] In a possible design of the first aspect, the reinforcement learning method comprises an attention mechanism and a task adaptive reward mechanism.

[0014] In a possible design of the first aspect, the correlation degree between the user labels is calculated, and specifically, the calculation comprises:

[0015] structured data in the user labels is extracted; and

[0016] The structured data is processed by using a graph neural network to obtain the correlation degree.

[0017] In a possible design of the first aspect, an attention mechanism is added in the graph neural network processing, and a long short-term memory network is combined, where the long short-term memory network is used to complete dynamic expression of the correlation degree according to unstructured data and time-series data in the user labels.

[0018] In a possible design of the first aspect, medical care and health services are predicted based on the user demand portrait and the correlation degree, and specifically, a reinforcement learning method comprising an attention mechanism and a task adaptive reward mechanism and combined with an A* algorithm is used to obtain the medical care and health prediction services from the user demand portrait and the correlation degree.

[0019] In a possible design of the first aspect, the medical care and health services are constrained by QoS, the user labels are health labels, behavior labels, environment labels, or service preference labels, and the user labels are obtained by collecting user data, data cleaning, noise reduction, normalization, and analysis.

[0020] In a second aspect, a medical care and health service recommendation method is provided, and the recommendation method comprises:

[0021] current user labels are obtained; and

[0022] The current medical care and health prediction services are obtained by using the intelligent agent trained as described above.

[0023] In a third aspect, an electronic device is provided, which comprises a processor and a memory coupled with the processor, where the memory is configured to store a computer program, and the processor is configured to execute the computer program stored in the memory, so that the electronic device performs the training method in any possible implementation manner of the first aspect or performs the recommendation method in the second aspect.

[0024] In a fourth aspect, a computer readable storage medium is provided, including computer programs or instructions, which, when executed on a computer, cause the computer to perform the training method according to any possible implementation of the first aspect, or perform the recommendation method according to the second aspect. BRIEF DESCRIPTION OF DRAWINGS

[0025] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments or related description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0026] Figure 1 is a whole flowchart provided by the embodiments of the present application;

[0027] Figure 2 is a medical data preprocessing flowchart provided by the embodiments of the present application;

[0028] Figure 3 is a rehabilitation data preprocessing flowchart provided by the embodiments of the present application;

[0029] Figure 4 is a behavior data preprocessing flowchart provided by the embodiments of the present application;

[0030] Figure 5 is an environmental data preprocessing flowchart provided by the embodiments of the present application;

[0031] Figure 6 is a decision-level fusion flowchart provided by the embodiments of the present application;

[0032] Figure 7 is a structured data processing flowchart provided by the embodiments of the present application;

[0033] Figure 8 is an unstructured data processing flowchart provided by the embodiments of the present application;

[0034] Figure 9 is a time series data processing flowchart provided by the embodiments of the present application;

[0035] Figure 10 is a service recommendation flowchart provided by the embodiments of the present application;

[0036] Figure 11 is a feedback optimization mechanism flowchart provided by the embodiments of the present application;

[0037] Figure 12 is a service chain management and execution flowchart provided by the embodiments of the present application. DETAILED DESCRIPTION

[0038] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not intended to limit the present application.

[0039] It should be noted that although the functional modules are divided in the device schematic diagram, and the logical sequence is shown in the flowchart, in some cases, the steps shown or described can be performed in a manner different from the module division in the device or the sequence in the flowchart. The terms "first", "second", etc. in the specification and claims and the above-described drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence.

[0040] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0041] Before introducing the embodiments of the present application, a brief description of the present stage of technical research of the present application is given:

[0042] Demand profile: a multi-dimensional attribute description constructed for users through multi-source data analysis and fusion, including health status, behavior preference and environmental characteristics, etc.

[0043] Deep learning: a method of machine learning that uses multi-layer neural networks to learn high-order feature representations from data.

[0044] QoS (Quality of Service): a comprehensive index for measuring service response time, reliability, availability, etc. in service recommendation.

[0045] A* algorithm: a heuristic search algorithm used to find the shortest path from the starting point to the ending point in a weighted graph.

[0046] Reinforcement learning: a method of machine learning that learns behavior strategies that can achieve goals through interaction with the environment.

[0047] Multi-source heterogeneous data: data from different data sources, formats, structures and time distributions, such as medical records, behavior data and environmental data.

[0048] Below, the training method of the medical care and health service recommendation agent and the medical care and health service recommendation method of the embodiments of the present application are exemplarily described with reference to the accompanying drawings. Figures 1 to 12

[0049] ​The technical solution includes information collection, data preprocessing, portrait construction, service recommendation, feedback optimization and service chain management modules. Through the cooperation of these modules, a dynamic and accurate user demand portrait can be efficiently constructed, and personalized and real-time service recommendation can be realized through improved A* algorithm and QoS optimization. At the same time, the system also improves the recommendation accuracy through reinforcement learning and feedback optimization mechanism, and ensures the efficient execution of service chain and the reasonable scheduling of resources. The specific process is shown in Figure 1

[0050] S1: The data source is mainly based on the data collected by the medical care and rehabilitation service system, including medical data (such as electronic medical records, drug use records and examination reports), rehabilitation data (such as rehabilitation assessment records, treatment plans and training logs), behavior data (such as activity frequency, exercise data, especially data from wearable devices, and living habits), and environmental data (such as living location, air quality, noise level and community service distribution). Comprehensive analysis of these multi-dimensional data can comprehensively evaluate the health status of users and provide personalized health management suggestions.

[0051] S11: The collection of medical data UM is mainly obtained through the electronic health record system (EHR) of hospitals and clinics, including the electronic medical records, diagnosis records and treatment plans of users. In addition, the drug use records are extracted through the hospital information system (HIS), covering prescription information, drug use frequency and dosage, etc. The examination reports come from the hospital's inspection and detection system, including blood tests, imaging tests and other medical test results, to ensure comprehensive recording of user health information.

[0052] S12: The collection of rehabilitation data UR is mainly through the rehabilitation assessment records made by medical professionals. These assessments usually use standardized tools (such as Barthel Index, FIM scale, etc.) to assess functional recovery and track disease progression. Treatment plans are obtained through electronic health records or rehabilitation management systems, based on doctors' diagnoses and individual needs. In addition, training logs are recorded by patients themselves or rehabilitation center staff, including daily training items, time and progress, through paper logs, mobile applications or spreadsheets, etc.

[0053] S13: The collection of behavior data UB is carried out through various ways, including smart phones, wearable devices (such as smart bands, smart watches, etc.) and motion sensors, to monitor users' activity frequency, steps, exercise time and exercise intensity, etc. in real time. Exercise data is provided by wearable devices, covering heart rate, exercise type, gait and other physiological indicators. Living habit data is collected through questionnaires, user self-reports or smart home devices (such as smart refrigerators, smart lighting, etc.), recording users' sleep patterns, eating habits and social activities, etc. ​

[0054] S14: The collection of environmental data of the UE is obtained by GPS or the positioning function of the mobile device to obtain the user's residential location information. Air quality data is collected by air quality monitoring instruments to record the levels of pollutants such as PM2.5, PM10, and carbon dioxide concentration, ensuring comprehensive monitoring of environmental health indicators. Noise levels are monitored by noise sensors or smart devices (such as smart speakers, smartphones) to record noise data in the living environment. Community service distribution data is obtained through public service data platforms or government information systems to provide location information and service distances for hospitals, pharmacies, gyms, parks, and other facilities.

[0055] S2: Preprocessing of the data collected in S1 to improve the quality and consistency of the data. This process includes data cleaning, noise reduction, normalization, and feature extraction.

[0056] S21: Data preprocessing for the medical data UM collected in S1. Medical data typically includes electronic medical records, test reports, and medication usage records, which may contain duplicates, missing or invalid values. First, remove duplicate records and invalid information (such as expired medication information) through automated algorithms, and fill in missing values. Missing values will be filled with industry standard values. In terms of format standardization, different hospitals or systems may use different units or formats, so preprocessing is needed to ensure data consistency. The preprocessed data is UMs2 = {UMs, UMns}, where the subscript s is structured data, and ns is unstructured data. For example, different medical institutions use disease diagnosis codes (such as ICD codes), which need to be standardized to ensure uniform management and analysis of data. At the same time, the uniformity of drug names and the coordination of test result units also need to be standardized to ensure data quality. The specific process is shown in Figure 2 .

[0057] S22: Data preprocessing for the rehabilitation data UR collected in S1. Rehabilitation data includes rehabilitation assessment records, treatment plans, and training logs, which may contain missing values or measurement errors. In the data cleaning stage, first remove invalid records (such as incorrect assessment results or lack of treatment records). Missing data is filled in by the mean interpolation method. In the noise reduction stage, the motion data in the rehabilitation training log may be affected by device measurement errors, and the sliding average method is used to smooth the data and remove noise. The preprocessed data is URs2 = {URs, URns, URts}, where the subscript s is structured data, ns is unstructured data, and ts is time series data. As shown in Figure 3 .

[0058] S23: Data preprocessing for behavior data UB collected in S1. Behavior data, such as gait and movement frequency, are usually derived from wearable devices or sensors, which are prone to noise interference (e.g., device errors or external environmental factors). In the data cleaning process, missing data points or outliers (e.g., unreasonable step frequency or heart rate) are first removed. In the noise reduction process, Kalman filter is used to smooth the heart rate and motion trajectory data, reducing the impact of noise on the data. In the normalization process, due to the scale difference of behavior data (e.g., step count, heart rate, exercise intensity, etc.), it is necessary to unify these data to the same range. By normalizing the data to the 0-1 interval and ensuring that the data has zero mean and unit variance, the difference in different scales is avoided to affect the training effect of the subsequent machine learning model. The preprocessed data is UBs2 = {UBs, UBns, UBts}, where the subscript s is structured data, ns is unstructured data, and ts is time series data. As shown in Figure 4

[0059] S24: Data preprocessing for environmental data UE collected in S1. Environmental data such as air quality, noise level and community service distribution may contain missing values or abnormal fluctuations. In the data cleaning stage, invalid records are removed and missing values are filled by mean interpolation. For air quality and noise data, noise reduction processing is particularly important, and smoothing algorithms need to be used to remove short-term sudden fluctuations. In the normalization process, the dimension of environmental data (such as PM2.5 concentration, noise decibel, etc.) is unified to a standardized range to facilitate comprehensive analysis with other health data. The preprocessed data is UEs2 = {UEs, UEns}, where the subscript s is structured data and ns is unstructured data. As shown in Figure 5

[0060] S3: Constructing a multi-dimensional demand portrait label system for clean data processed in S2. By analyzing health, behavior, environment and service preference data, a multi-dimensional user portrait label system is constructed.

[0061] S31: Input health data UMs2, URs2 obtained from the preprocessing stage in S2, including user's disease type, treatment record and key health indicators (such as blood pressure, blood sugar, etc.). These data help to fully depict the user's health status and disease management needs, providing basic information for personalized recommendation. Through the analysis of health data, the user's health status label HL (such as hypertension, hyperglycemia, diabetes, healthy population, etc.), disease management needs (such as whether long-term monitoring is needed, whether regular treatment is needed, etc.), and health risk assessment (such as potential risk assessment based on health indicators) are output.

[0062] ​​S32: Input the behavior data UBs2 obtained from the S2 preprocessing stage, including the user's lifestyle habits, interests and hobbies, and daily activity patterns. By monitoring and analyzing the user's behavior patterns, output the user lifestyle label WL (such as healthy lifestyle, sports preference, social activity, etc.), behavior pattern recognition (such as high-risk behavior, low activity, etc.), personalized health recommendations (such as increasing exercise, adjusting diet, etc.).

[0063] S33: Input the environmental data UEs2 obtained from the S2 preprocessing stage, including the user's geographic location information, service accessibility and traffic conditions, etc. By analyzing these environmental data, output the user's environmental impact label EL (such as convenient transportation, rich service resources, etc.), surrounding service recommendation (such as nearby hospitals, rehabilitation centers, etc.), environmental optimization suggestions (such as recommending convenient transportation solutions, health facilities, etc.).

[0064] S34: (input is feedback data UF obtained in S7), including the user's historical service usage records and priority ranking of different service types. By analyzing the user's past service usage patterns and frequency, output the user's service preference label SL (such as preferring remote medical care, regular physical examination, etc.), personalized service recommendation (such as preferentially recommending the user's preferred service type), user demand change trend (predicting future demand based on historical records).

[0065] S4: Through data fusion technology, integrate the multi-dimensional demand portrait labels constructed in S3 stage, and finally form a comprehensive and accurate user demand portrait. This process includes feature-level fusion, decision-level fusion, weighting and label optimization, etc.

[0066] S41: Fuse the dimensional labels (health HL, behavior WL, environment EL and service preference SL) generated in S3 stage at the feature level. By concatenating these labels into a high-dimensional feature vector, the user's various needs and health status can be fully expressed. The health label (such as high blood pressure, high blood sugar, etc.) will be combined with the behavior label (such as healthy lifestyle, sports preference) and the environment label (such as convenient transportation, rich service resources) to generate a comprehensive user portrait feature vector.

[0067] (S41 here does not need feature extraction, only needs to concatenate the user portrait labels of S3)

[0068] S42: On the basis of feature-level fusion, further integrate the results of multiple prediction models using decision-level fusion method. First, train multiple sub-models for each dimension (health, behavior, environment, service preference) to make dimension-level predictions.

[0069] For all sub-models, the method of logistic regression is used, and the input is the user portrait label of S3 (health HL, behavior WL, environment EL, and service preference SL), and the output includes health model prediction HP: predicting the health status of the user (such as diabetes, hypertension, etc.). Behavior model prediction BP: predicting the behavior mode of the user (such as whether to like sports, eating habits, etc.). Environment prediction model EP: predicting the environmental needs of the user according to the geographical information of the user (such as whether there is convenient medical service, transportation, etc.). Service preference prediction model SP: predicting the preferred medical service type of the user (such as telemedicine, physical examination, etc.).

[0070] Then, according to the prediction results of different models, a comprehensive decision is made through the improved reinforcement learning fusion method. The result of the comprehensive decision is FD.

[0071] After combining the adaptive decision strategy with the attention mechanism and adding it to the reinforcement learning, the improvement steps are as follows: training the agent under the meta-learning framework, while introducing the attention mechanism to weight the key features in each task stage. And dynamically adjusting the update strategy of the Q value function according to different tasks and environmental states, while using the attention mechanism to weight the contribution of different features. Design a reward function that contains attention weights and task adaptability, so that the agent not only focuses on the prediction error, but also rewards behaviors that can effectively identify important features. After combining, the Q value update process of the agent is represented as:

[0072]

[0073] where, is the current Q value (i.e. expected return) of performing action a in state s; R(s, a) is the reward function, representing the immediate reward obtained by the agent after performing action a in state s; θ is the parameter of the policy function in reinforcement learning, by optimizing these parameters, the agent can improve its decision-making ability. The reward function takes into account the importance of the current task and whether the key features are correctly identified; γ is the discount factor, representing the degree of discounting future rewards, taking values between [0, 1]. It measures the agent's emphasis on future rewards;

[0074]

[0075] is the maximum Q value of all possible actions a' in the next state s', representing the maximum expected return that the agent can obtain in the next state; α is the learning rate, controlling the step size of each update. It determines the agent's emphasis on the current experience.

[0076] Q value update after combining the attention mechanism:

[0077]

[0078] where Q final (s, a) is the final Q-value, obtained by weighting the Q-values of multiple sub-models; Q i (s, a) is the Q-value of the i-th sub-model, representing the expected return of performing action a in state s. Each sub-model corresponds to a different dimension prediction (health, behavior, environment, service preference); a is the attention weight, representing the contribution of each sub-model Q-value in the final decision-making. The attention mechanism dynamically adjusts these weights by learning the key features in different task stages; n: the number of sub-models, usually equal to the number of task dimensions.

[0079] The final decision strategy selects the optimal action by maximizing the weighted Q-value:

[0080]

[0081] where a* is the optimal action selected by the agent;

[0082]

[0083] represents the sum of the weighted Q-values, representing the expected return of all sub-model Q-values after weighting in state s.

[0084] By combining the adaptive decision strategy and the attention mechanism, the performance of the reinforcement learning model is significantly improved. The adaptive decision strategy enables the agent to flexibly respond to different task requirements through meta-learning, while the attention mechanism helps the agent focus on key input features, thereby avoiding information overload and redundancy.

[0085] S5: Input the label data (health HL, behavior WL, environment EL, and service preference SL) of the user demand image obtained in S3 stage, and use a deep learning model to support the construction and data processing of multi-dimensional demand portrait, to more efficiently mine the complex features and correlations in user data.

[0086] S51: input the label data obtained in S3 stage, including the structured data in health HL, behavior WL, environment EL and service preference SL. Through the improved graph neural network (GNN), the application models the complex non-Euclidean relationship between various features of the user, thereby improving the understanding and prediction ability of the user demand. Then the health, behavior, environment and other characteristics of the user are modeled as a graph structure, wherein each feature (such as disease type, health index, behavior habit, etc.) is a node in the graph, and the relationship (such as the association between disease and environmental factors, health data and behavior pattern) between them is an edge in the graph. Then the improved graph neural network (GNN) is used to learn the complex non-Euclidean relationship between the nodes, especially the association between the user health data and the environmental factors. Finally, the graph structure relationship UserGraph_Structure between the user health data and the behavior data based on the GNN is output, enhancing the accuracy and personalization of the demand portrait. The specific process is as shown in Figure 7 .

[0087] 1) Introduce a graph attention mechanism to enhance the model's attention to important features. First, calculate the attention coefficient between nodes, which is used to weight the aggregation of neighbor node feature information. Then calculate the attention weight e vu :

[0088] e vu =LeakyReLU(a T [Wh v ||Wh u ])

[0089] wherein hv and hu are the feature vectors of nodes v and u, W is a weight matrix, a is a weight vector for calculating attention, || represents a feature concatenation operation, T represents a transpose, and LeakyReLU represents a leaky rectified linear function.

[0090] The attention coefficient a vu is normalized by the following formula:

[0091]

[0092] wherein is the set of neighbor nodes of node v, and represents all nodes directly connected to node v; exp(e vu' ) applies an exponential function to amplify the attention coefficient, increasing the influence of the neighbor nodes with larger weights;

[0093]

[0094] represents the normalization of the attention coefficients of all neighbor nodes, ensuring that the sum is 1.

[0095] Finally, feature updates are performed:

[0096]

[0097] Among them, h' v N(v) represents the updated feature representation of node v; N(v) represents the set of neighboring nodes of node v; σ is the activation function, using ReLU or LeakyReLU; α vu The attention coefficients between node v and node u are used to weight and aggregate the features h of neighbor node u. u .

[0098] 2) Combined with LSTM, it facilitates the simultaneous handling of complex relationships in graph structures and dynamic changes in time series data. First, a Graph Attention Network (GAT) is used to extract features from the graph structure data, updating the representation of each node. Then, the node representations extracted by GAT are used as input to LSTM to capture dynamic changes in the time series data. Finally, the output of LSTM is used for user demand prediction.

[0099] The node representation hv′ obtained through GAT is used as the input sequence for LSTM:

[0100]

[0101] in, At time step t, the updated node representation of the graph neural network; {h'1,h'2,…,h' n} is the feature representation of all nodes 1, 2, ..., n at time step t.

[0102] LSTM dynamically models these feature vectors:

[0103]

[0104] in, It is the input representation of LSTM, which is usually obtained by further normalizing the node features updated by the graph neural network (GAT).

[0105] Use the output of the LSTM to predict user demand.

[0106]

[0107] in, This is the prediction result for time step t, W out and b out These are the output weights and biases used for prediction. This combination approach can handle the non-Euclidean structure of graph data while also capturing the temporal nature of the data, thereby improving the model's performance in tasks such as user demand prediction.

[0108] S52: Input the pre-processed unstructured data {UMns, URns, UBns, UEns} in S2 stage, including user-uploaded medical images (e.g. dermatological photos, CT scan images) or user's voice recordings (e.g. voice call records of health consultation). First is the encoder part, for image data, use convolutional neural network (CNN) to extract features of the image. CNN automatically learns and extracts discriminative local features (e.g. lesion area, texture, color, etc. of dermatological disease) from the image through multiple convolutional layers and pooling layers. For voice data, use recurrent neural network (RNN) to handle the time series information and extract time series features in the voice signal. Common variants of RNN such as LSTM (Long Short-Term Memory) can capture long-term dependencies in the voice (e.g. emotional fluctuations, health problem hint signals, etc.). Through the encoder part, the processed data (image or voice) is mapped to a low-dimensional latent space, generating a compact latent vector representing the core features of the input data. Then is the decoder part, for image data, use deconvolution layers to gradually restore the size of the image. Deconvolution layer performs upsampling operation to restore the low-dimensional feature vector to the original image size and try to preserve the structural information of the image. For voice data, use RNN (such as LSTM) to restore the latent vector to the time series structure of the voice signal, so as to reconstruct the original voice spectrum or waveform. The goal of the decoder is to reconstruct the output as close to the original input data as possible, i.e. to minimize the error between the input data and the reconstructed data. Autoencoder uses self-supervised learning method for training, i.e. in the absence of labeled data, the model parameters are optimized by minimizing the reconstruction error. In the training process, the system continuously adjusts the parameters of the encoder and decoder until the difference between the reconstructed data and the original input data is minimized, using mean square error (MSE) or other loss function for optimization. Output UserGraph_NStructure, the process is as shown in Figure 8 .

[0109] S53: Input the time series data {URts, UBts} preprocessed in S2 stage, including the user's health monitoring data (such as blood pressure, blood sugar, body temperature, etc.) and behavior data (such as exercise, sleep, etc.). First, use historical data to train the LSTM model, update the weights of the LSTM model by minimizing the loss function (such as mean square error). During the training process, LSTM will learn the relevance and long-term trends between health monitoring data and behavior data in the input sequence. After training, LSTM captures the long-term dependencies in these time series data, analyzes past health and behavior data, and predicts future health status and behavior patterns. For example, based on past blood sugar data, LSTM can predict the trend of future blood sugar levels to help identify potential health risks; at the same time, LSTM can also identify long-term trends in exercise and sleep data to provide personalized behavior recommendations. Through these predictions, the system can provide accurate health management recommendations for users and optimize personalized services to improve health management effectiveness. The process is shown in Figure 9

[0110] S6: Input the user demand portrait processed in S4 stage and the complex correlation obtained in S5, combine intelligent matching algorithm and real-time recommendation process to realize dynamic service recommendation and optimization.

[0111] S61: When the user demand portrait is updated, the system automatically triggers the service recommendation process. If the user's health status changes (such as blood sugar rising, abnormal exercise, etc.), the system re-evaluates the user's demand based on the new health data and activates the recommendation process. For A* algorithm,

[0112] 1) Introduce Qos constraints, that is, in the original cost function, consider the quality of service (such as response time, reliability, bandwidth, etc.) to adjust the search strategy, so that the final path found is not only optimal in cost, but also meets the QoS constraint condition. Improvement steps: each node represents a service instance, and the state is represented by the QoS characteristics of the service (such as response time, cost, bandwidth, reliability, etc.). The edge represents the conversion or dependency relationship between services, and the cost is the switching cost or the relationship between service selection. The cost function is:

[0113]

[0114] where g(n) is the cost of the current path, representing the actual cost from the starting point to the current node; h(n) is the heuristic function, representing the expected cost (usually estimated distance or time) from the current node to the target node; QoS i (n) is the i-th QoS characteristic of node n (such as response time, bandwidth, availability, etc.), each QoS characteristic will affect the total cost; w i ​is the weight of each QoS characteristic, representing the contribution degree of the QoS characteristic to the cost function; m is the number of QoS characteristics considered.

[0115] 2) The heuristic function h(n) is static in the traditional A* algorithm, but in the improved version of the present application, h(n) should take into account the dynamic changes of service QoS characteristics, so the present application designs a dynamically adjusted heuristic function:

[0116]

[0117] wherein estimated.cost(n) is the expected cost from the current node to the target node estimated based on the heuristic method; and a and β are weights, controlling the relative importance of heuristic method estimation and QoS constraints.

[0118] 3) Reinforcement learning introduces an adaptive decision-making strategy. When QoS parameters change, reinforcement learning can dynamically adjust the service selection strategy, thereby achieving optimal service recommendation. State-action pairs and reward mechanisms are introduced, and the Q-learning algorithm is used to adjust the service selection strategy. At each time step t, the Q value is adjusted through environmental feedback (user feedback on services, QoS satisfaction):

[0119]

[0120] wherein St is the current state, At is the currently selected service, Rt is the immediate reward, and γ is the discount factor.

[0121] 4) Reinforcement learning is combined with A* algorithm, which can combine global optimization and local optimization. A* algorithm is used for global planning and searching for the optimal path, while reinforcement learning dynamically adjusts at the local level, enabling the system to make more intelligent and faster decisions when recommending services in real time.

[0122] The service recommendation process is shown in Figure 10 .

[0123] S62: In the service recommendation process, the system filters the service resources that meet the QoS constraints from the database. The filtering criteria include not only service quality (such as response time, success rate, treatment effect, etc.), but also resource limitations (such as service capacity, equipment availability). For example, if a user needs remote medical consultation, and the doctors of a certain platform have been fully booked, the system will dynamically adjust according to the available resources and recommend other platforms or doctors. In this way, users can obtain timely medical services and avoid delays due to resource limitations, ensuring the continuity and availability of services.

[0124] S63: When planning the service chain, the improved A* algorithm dynamically adjusts the recommended path according to the user's geographic location, service timeliness, and demand priority. This algorithm considers multiple factors such as distance, service availability, traffic conditions, etc., to optimize the order and process of services. For example, if the user lives in an area with poor transportation, the system will prioritize medical or rehabilitation centers that are closer and have higher timeliness, reducing the user's waiting time and travel burden. If the user's health condition requires emergency intervention, the system will ensure that the nearest available medical facility is reached in the shortest time, thereby improving service efficiency and user experience.

[0125] S64: After optimizing the service path, the system selects appropriate medical, rehabilitation, and elderly care services based on the user's profile and optimization results, and generates personalized service chains. For example, for users in the rehabilitation period, the system will recommend targeted rehabilitation training services; for users with chronic disease management needs, the system will recommend regular health monitoring, remote medical consultation, and other services. The service recommendation plan will generate a service chain based on the optimized path and output a personalized recommendation plan, covering detailed information for each service step, implementation time, priority order, and execution method. Through this process, users not only receive services that meet their health needs, but also enjoy personalized and precise service experiences in actual implementation.

[0126] S7: Feedback optimization mechanism, in this stage, the system continuously improves the accuracy and effectiveness of service recommendation through the collection of feedback data and model optimization, and dynamically updates the user demand profile label. The specific process is shown in Figure 11

[0127] S71: Collect feedback data, including objective data and subjective data. Objective data includes service completion, user health index changes, etc. These data can accurately reflect the effectiveness of service execution. For example, health monitoring data (such as blood sugar level, weight change, exercise data, etc.) can help evaluate the effectiveness of recommended services and determine whether the expected health improvement goal has been achieved. Subjective data includes user satisfaction scores, complaint information, etc. These data help understand the user's subjective feelings about the service and provide real-time feedback on service quality. For example, if the user is not satisfied with a medical service (such as remote consultation, health check), the system will collect the user's specific complaints and feedback to further analyze the pros and cons of the service. Through the collection of these two types of data, the system comprehensively understands the actual effectiveness of the service and provides basic data for subsequent model optimization.

[0128] ​S72: Through the collected feedback data, the system uses a reinforcement learning framework (DQN) to dynamically optimize the recommendation model. The update of the model follows a reward mechanism and a punishment mechanism. The reward mechanism is that when the user is satisfied with the recommended service or the service is completed well, the model will give a higher reward score to the recommendation strategy. This reward mechanism encourages the system to continuously recommend better service combinations. For example, if a specific service path (such as a remote medical service in a certain hospital) can improve the user's health indicators and receive high evaluation from the user, this service path will be considered as a priority selection item and will be preferentially recommended in the future. The punishment mechanism is that if the user's satisfaction is low or the service completion rate is low (for example, a health management service does not complete as planned, or the user is not satisfied with the response of remote medical treatment), the system will punish the corresponding service combination or path and reduce its priority in future recommendations. For example, if a service provider has slow service response and poor treatment effect, leading to user dissatisfaction, then this service provider will be reduced in priority, and the system will preferentially recommend other service providers with better performance. Through the reward and punishment mechanism, the system continuously learns from user feedback and gradually adjusts the recommendation strategy to improve service quality and user satisfaction.

[0129] S73: Dynamic updating of demand profile, by taking feedback data (including objective health data and subjective satisfaction data) as incremental input, the system will regularly update and optimize the user's demand profile label. This update can help the system better understand the user's health needs, service preferences and satisfaction with the service. For example, if a user feedbacks that the high blood sugar problem has been gradually controlled after using a health management service for a period of time, the system will adjust its health profile, update the disease management demand label, and accordingly recommend more efficient diabetes management services for it. With the continuous updating of user demand profile, the recommendation system can become more accurate and personalized, thereby improving the overall recommendation quality. Through dynamic updating of demand profile, the system can better respond to user health changes and feedback, and provide continuously optimized service recommendations.

[0130] S8: Service chain management and execution phase, to ensure standardized execution, efficient monitoring and cross-institutional resource coordination of service chain, to improve the accuracy and efficiency of service execution. As shown in Figure 12

[0131] ​S81: First, the system needs to establish a unified service process framework to ensure that each recommended service chain can be executed seamlessly. To this end, the system will define each service node in the service chain (such as medical services, rehabilitation services, and elderly care services, etc.), and clearly define the input and output requirements and time requirements of each node. For example, in the medical node, the input data may include health check data (such as blood sugar, blood pressure, etc.), and the output is a diagnosis and treatment plan and drug recommendations, and it is required to complete the preliminary diagnosis and treatment within a specified time; in the rehabilitation node, the input data is the user's health assessment result, and the output is a personalized rehabilitation training plan, which requires completing the training plan within a specified time. Through such a standardized process framework, it can ensure that the service execution of each node is carried out according to uniform specifications, avoiding service execution deviation.

[0132] S82: The system needs to monitor the execution of the service chain in real time, focusing on key indicators such as task completion time, user waiting time, and service resource utilization efficiency. Task completion time is used to ensure that each service node completes the task within the specified time, and user waiting time is used to monitor the time from the user's initiation of demand to the actual start of service, ensuring that the waiting time is within an acceptable range. At the same time, service resource utilization efficiency will evaluate the allocation and use of resources such as doctors, beds, and equipment, ensuring that resources are optimally configured. If an abnormal node execution is found during service execution (such as service timeout, task not completed, etc.), the system will start an emergency plan, automatically switch to backup service resources, adjust the service execution order, or recommend other available resources. For example, if a hospital cannot provide remote medical services due to equipment failure, the system can automatically switch to other hospitals to ensure the continuity and timeliness of services.

[0133] S83: In this stage, the system needs to build a service coordination platform based on micro-service architecture to support data sharing and service interaction between medical, rehabilitation, and elderly care institutions. The platform can ensure the flow of data between different institutions, avoiding information silos, thereby improving the efficiency of resource collaboration. By dynamically allocating resources, the system will flexibly adjust the resource allocation of each node according to the actual situation of service chain execution, ensuring that the service chain can be executed efficiently. For example, if a rehabilitation center is short of bed resources, the system will automatically transfer users to other available rehabilitation institutions nearby according to the user's health status and geographical location, thereby avoiding service delays due to insufficient resources and ensuring the efficiency and timeliness of services.

[0134] Other optional embodiments:

[0135] 1. In the process of demand profiling, LLMs model can be used to generate user profiles. By generating user profiles through LLMs, a more comprehensive and detailed understanding of users can be provided to the service recommendation system, not limited to basic health data and behavior habits, but also in-depth understanding of users' emotional background, life challenges and goals. This method enhances the personalization of the recommendation system, helps to improve the accuracy of service recommendation and user satisfaction, and promotes the continuous dynamic optimization of user demand.

[0136] 2. In the process of cross-institutional resource collaboration and execution, a cloud service-based collaboration platform can be used to provide a unified data storage and service scheduling platform using cloud computing platforms such as AWS, Azure, etc. In cross-institutional resource collaboration, the cloud platform can provide a unified resource pool and scheduling mechanism to ensure that the data of each service node can be shared and the service execution can be quickly deployed and coordinated.

[0137] 3. In the process of service chain management and execution, a mature workflow management system such as Apache Airflow or Camunda can be used to manage and schedule each task node in the service chain. The workflow engine can assign tasks to each service node, monitor service status, and dynamically adjust according to execution conditions. In this way, the system can flexibly manage cross-service workflows and data transmission to ensure smooth completion of tasks.

[0138] The embodiments of the present application also provide an electronic device, including: a processor, and a memory coupled with the processor, the memory being configured to store a computer program; and the processor being configured to execute the computer program stored in the memory, so that the electronic device performs the method according to any one of the above embodiments.

[0139] The electronic device can be a desktop computer, a notebook computer, a palm computer, a cloud server, and other computing devices. The electronic device can include, but is not limited to, a processor and a memory.

[0140] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The processor is a control center of the electronic device, and connects various parts of the entire device through various interfaces and lines.

[0141] The memory can be used to store the computer program, and the processor realizes various functions of the electronic device by running or executing the computer program stored in the memory and calling data stored in the memory.

[0142] The memory can mainly include a program storage area and a data storage area. The program storage area can store an operating system, at least one application required by a function, etc. The data storage area can store data created according to the use of the mobile phone, etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state memory devices.

[0143] The embodiment of the present application further provides a storage medium, which is a computer readable storage medium, and the computer program is stored in the computer readable storage medium. When the computer program is executed by a processor, the steps of each method embodiment described above can be realized. The computer program includes computer program code, which can be in the form of source code, object code, executable files or some intermediate forms, etc. The computer readable medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal and software distribution medium, etc.

[0144] The above are preferred embodiments of the present application, it should be pointed out that, for those skilled in the technical field, without departing from the principles of the present application, can make several improvements and refinements, these improvements and refinements also considered the scope of protection of the present application.

Claims

1. A training method for an intelligent agent recommending medical, elderly care, and rehabilitation services, characterized in that, The training method includes: The training data includes: medical, elderly care, and rehabilitation services, and at least two user tags related to the medical, elderly care, and rehabilitation services; and Using the training data, the agent is trained, specifically to: construct a user demand profile using at least two types of user tags, whereby the user tags are health tags, behavior tags, environment tags, service preference tags, or satisfaction tags; calculate the correlation between the user tags; and, based on the user demand profile and the correlation, predict medical, elderly care, and rehabilitation services. Constructing a user demand profile using at least two user tags specifically includes: Feature fusion is performed on the at least two user tags to obtain a high-dimensional feature vector; Based on the user tags, at least two demand predictions are obtained; and The user demand profile is obtained by using a reinforcement learning method with the high-dimensional feature vector and the demand prediction as input. Calculating the correlation between the user tags specifically includes: Extract the structured data from the user tags; and The structured data is processed using a graph neural network to obtain the correlation degree. An attention mechanism is added to the graph neural network processing and combined with a long short-term memory network. The long short-term memory network is used to dynamically express the correlation degree based on the unstructured data and time-series data in the user tags. Based on the user demand profile and the correlation, medical, elderly care, and rehabilitation services are predicted. Specifically, a reinforcement learning method incorporating attention mechanisms, task-adaptive reward mechanisms, and the A* algorithm is used to obtain predicted medical, elderly care, and rehabilitation services from the user demand profile and the correlation. Specifically, the structured data is processed using a graph neural network to obtain the correlation degree, including: Model the complex non-Euclidean relationships between the features corresponding to the structured data; The model uses a graph structure, where each feature is a node in the graph, and the relationships between features are represented by edges. The complex non-Euclidean relations are learned using graph neural networks.

2. The training method as described in claim 1, characterized in that, The reinforcement learning method includes an attention mechanism and a task-adaptive reward mechanism.

3. The training method as described in claim 1, characterized in that, The medical, health, and rehabilitation services are subject to QoS constraints; the user tags are obtained through user data collection, data cleaning, noise reduction, normalization, and analysis.

4. A method for recommending medical, elderly care, and rehabilitation services, characterized in that, The recommendation method includes: Get the current user's tag; and Using the intelligent agent trained as described in any one of claims 1 to 3, the current recommended medical, health, and rehabilitation prediction service is obtained.

5. An electronic device, characterized in that, The electronic device includes: a processor, and a memory coupled to the processor. The memory is used to store computer programs; The processor is configured to execute the computer program stored in the memory, such that the electronic device performs the training method as described in any one of claims 1 to 3, or performs the recommended method as described in claim 4.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a computer program or instructions that, when executed on a computer, cause the computer to perform the training method as described in any one of claims 1 to 3, or to perform the recommended method as described in claim 4.

Images