Patient follow-up visit management method and system

Abstract

The invention relates to a patient follow-up visit management method and system, and solves the problems that generation of follow-up visit content by means of a fixed template is lack of personalized consideration, and actual situations such as special requirements, living changes and current disease treatment stages of patients cannot be fully combined. The method comprises the following steps: performing feature coding on a first draft of follow-up content and personalized adjustment information, identifying semantic association features of sentence components through a semantic role labeling technology, calculating statement similarity of to-be-fused content by applying a semantic matching technology based on a cosine similarity algorithm and a word vector model, and optimizing a content logic structure so as to eliminate redundant information; and finally, the follow-up visit content of this time is formed. And based on a preset follow-up plan, the follow-up content of this time is combined with the features extracted from the health data of the patient, and a voice notification strategy is generated by adopting a reinforcement learning algorithm. The follow-up visit management method has the following effect that the accuracy and efficiency of follow-up visit management of the patient are improved.

Description

Technical Field

[0001] This invention relates to the field of medical information management, and in particular to a method and system for patient follow-up management. Background Technology

[0002] With the continuous improvement of medical standards and the increasing emphasis people place on health management, patient follow-up management plays an increasingly important role in the entire medical service process. It aims to provide patients with more targeted and consistent medical services by continuously tracking their health status, treatment progress, and other aspects, thereby helping to improve treatment outcomes, quality of life, and satisfaction with medical services.

[0003] Currently, some patient follow-up management methods rely primarily on medical staff manually recording patient information and conducting follow-up visits periodically via telephone or outpatient visits. Some systems also collect basic physiological data from patients and generate follow-up content according to fixed templates for notification. For example, some hospitals use patient treatment data recorded in electronic medical records systems to set uniform follow-up cycles and content frameworks, conducting routine health inquiries and reminders to patients.

[0004] Manual record-keeping and follow-up are inefficient and easily affected by factors such as the workload of medical staff and the accuracy of their memory, leading to problems such as untimely follow-ups and missing information. Relying on fixed templates to generate follow-up content lacks personalized consideration and cannot fully take into account the patient's specific needs, life changes, and current stage of disease treatment, making it difficult to truly meet the diverse health management needs of patients and affecting the improvement of patients' medical experience and treatment outcomes. Summary of the Invention

[0005] To improve the accuracy and efficiency of patient follow-up management, this application provides a patient follow-up management method and system.

[0006] Firstly, this application provides a patient follow-up management method, which adopts the following technical solution: A patient follow-up management method includes: Regularly acquire patients' health-related data and perform data preprocessing; Using a pre-defined feature extraction method, physiological indicators, lifestyle features, and treatment-related features are extracted from pre-processed health-related data. The patient's physiological characteristics, lifestyle characteristics, and treatment-related characteristics are input into the trained follow-up type classification model, and the follow-up type is output. Obtain the patient's disease type and current treatment stage, and use the follow-up type, patient's disease type and current treatment stage as input objects to input into the follow-up content draft generation model, and output the follow-up content draft; The system acquires patients' personalized information and inputs it into a personalized adjustment model based on causal inference, outputting personalized adjustment information for follow-up content. A bidirectional semantic coding model based on the Transformer architecture is used to encode the features of the initial draft of the follow-up content and personalized adjustment information. Semantic role labeling technology is used to identify the semantic association features of sentence components. Semantic matching technology based on cosine similarity algorithm and word vector model is used to calculate the sentence similarity of the content to be fused. Based on this, the logical structure of the content is optimized to eliminate redundant information, and finally the follow-up content is formed. Based on the pre-set follow-up plan, the content of this follow-up visit, combined with the features extracted from the patient's health data, is used to generate a voice notification strategy using a reinforcement learning algorithm. The system automatically matches the patient's communication method according to the voice notification strategy, triggers the voice notification at a specified time, dynamically adjusts the broadcast parameters, and collects the patient's feedback information through voice recognition technology after the notification ends.

[0007] Secondly, this application provides a patient follow-up management system, which adopts the following technical solution: A patient follow-up management system includes a memory, a processor, and a program stored in the memory and executable on the processor, the program being loaded and executed by the processor to implement the patient follow-up management method as described in the first aspect. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating a patient follow-up management method according to an embodiment of this application.

[0009] Figure 2 This is a flowchart illustrating the process by which the follow-up content is ultimately formed according to another embodiment of this application. Detailed Implementation

[0010] The present application will be further described in detail below with reference to the accompanying drawings.

[0011] Reference Figure 1 This application discloses a patient follow-up management method, comprising: Step S100: Periodically acquire the patient's health-related data and perform data preprocessing.

[0012] The health-related data includes physiological indicators, lifestyle and environmental data, and treatment-related data. Data preprocessing includes noise reduction, format standardization, and missing data imputation to improve data quality for subsequent analysis.

[0013] Specifically, physiological data includes: blood pressure (obtained via a blood pressure monitor), heart rate (commonly measured using an electrocardiogram (ECG) or wearable health monitoring devices), blood glucose levels (detected using a blood glucose meter), and various blood lipid indicators (usually obtained through analysis of blood samples using specialized biochemical testing equipment). Lifestyle and environmental data include: lifestyle data such as the patient's daily diet (collected by the patient themselves recording daily food types and intake, or using a diet tracking app); exercise habits (such as weekly exercise frequency, duration, and type); and environmental data including air quality in the patient's living environment (data obtained through professional air quality monitoring, such as PM2.5 concentration and harmful gas levels), and noise levels (measured using a noise meter). Treatment-related data includes the patient's current treatment methods (such as medication, physical therapy, or surgery, recorded by the doctor during the consultation), and medication details (specific drug names, dosages, and frequency of use).

[0014] Step S200: Using a preset feature extraction method, physiological indicator features, lifestyle features, and treatment-related features are extracted from the preprocessed health-related data.

[0015] In terms of physiological indicators, key elements that reflect the characteristics of the patient's physiological state are extracted from data such as blood pressure, heart rate, and blood glucose, such as blood pressure fluctuation characteristics and heart rate variability characteristics. These are extracted using methods such as wavelet transform and principal component analysis, and the data is processed with appropriate parameters to obtain the data.

[0016] In terms of lifestyle characteristics, we analyze lifestyle and environmental data such as diet, exercise, smoking and drinking to extract elements that reflect lifestyle habits, such as food diversity and exercise regularity. We use text mining technology combined with decision tree algorithms to extract these elements, and perform text preprocessing before constructing classification rules to determine the corresponding features.

[0017] In the treatment-related features section, key content reflecting the characteristics of the treatment process, such as medication adherence and the efficacy of different treatment methods, is summarized based on treatment-related data. Methods such as association rule mining and time series analysis are used to extract relevant features, such as analyzing drug combination and the trend of past treatment effects.

[0018] Step S300: Input the patient's physiological indicators, lifestyle characteristics, and treatment-related characteristics into the trained follow-up type classification model, and output the follow-up type for this time.

[0019] Follow-up types include routine follow-up, emergency follow-up, and specialist referral follow-up. Routine follow-up is suitable for patients with stable conditions, in the recovery period, or in the maintenance treatment stage of chronic diseases, and is conducted at predetermined time intervals and with routine examinations. Emergency follow-up is for patients with sudden changes in their condition, at risk of acute attacks, or experiencing serious adverse reactions, such as worsening symptoms in heart disease patients or severe bone marrow suppression in chemotherapy patients. In these cases, immediate follow-up is necessary for a detailed examination at the hospital, and the doctor will quickly adjust the treatment plan and take remedial measures. Specialist referral follow-up occurs when a patient's condition exceeds the scope of their current department or when the current treatment is ineffective. For example, a diabetic patient with retinopathy is referred to ophthalmology. The ophthalmologist follows up according to specialist guidelines, communicating and collaborating with the original department's doctors to ensure the patient's health.

[0020] Follow-up type classification models are often built based on machine learning or deep learning algorithms. For example, support vector machines learn to find the optimal classification boundary between features and follow-up types by learning from a large number of labeled training samples; neural networks use neurons to simulate the workings of the human brain and automatically learn complex nonlinear relationships to determine the follow-up type. During model training, parameters (such as the kernel function parameters of SVM, the weights and biases of the neural network, etc.) are continuously adjusted to improve accuracy and generalization ability.

[0021] Step S400: Obtain the patient's disease type and current treatment stage, and input the follow-up type, the patient's disease type, and the current treatment stage into the follow-up content draft generation model, and output the follow-up content draft.

[0022] Disease type refers to the specific classification name of the patient's disease, serving as a clear diagnostic marker for the patient's condition. It covers various clinical diseases, such as coronary heart disease and hypertension in the cardiovascular system, and diabetes in the endocrine system. Current treatment stage represents the patient's current treatment phase or status for their disease, reflecting the progress and focus of treatment, such as the recovery process after surgery, the stage of chemotherapy, and adjustments to drug dosage. Both the patient's disease type and current treatment stage can be obtained by querying the system or database for patient disease type and treatment stage information.

[0023] First, collect information on the aforementioned disease types and current treatment stages, and combine it with the follow-up type obtained in step S300. Organize this data according to the format required by the initial draft follow-up content generation model, such as encoding disease types and standardizing the data format for treatment stages. Then, input the organized data into a model built using machine learning (e.g., decision trees) or deep learning (e.g., recurrent neural networks) techniques. Based on pre-learned correlations between numerous different situations and follow-up content, the model generates an initial draft of follow-up content, including inquiries about the patient's condition, examination recommendations, and health guidance.

[0024] Step S500: Obtain the patient's personalized information and input the patient's personalized information into the personalized adjustment model based on causal inference, and output personalized adjustment information of follow-up content.

[0025] Personalized information includes patients' specific needs (such as in-home care for post-operative recovery), personal preferences (dietary restrictions, exercise preferences), allergy history (drug / food allergies), and life changes (moving, changing jobs). The methods for obtaining this information are as follows: 1. Electronic Medical Record System: Extracting structured data such as patients' allergy history and special medical needs recorded in disease diagnoses. 2. Doctor-Patient Communication Records: Reviewing text or voice records of past consultations and follow-ups to understand patients' specific needs and life changes. 3. Health Monitoring App: Analyzing dietary and exercise data actively entered by patients or automatically uploaded by devices to obtain their lifestyle preference information. 4. Follow-up Logs: Reviewing past follow-up records to understand recent changes in patients' life status and newly raised personalized needs.

[0026] The general process and model algorithm are disclosed as follows: Personalized information obtained from multiple sources is cleaned and formatted, then input into a personalized adjustment model based on causal inference. This model is built upon a structural causal model (SCM), defining causal relationships between variables through a directed acyclic graph (e.g., "moving → change in medical distance → adjustment of follow-up frequency"). Intervention analysis and counterfactual reasoning algorithms are used to simulate the impact of different personalized factors on follow-up content (e.g., removing recommendations containing allergens), ultimately outputting a targeted follow-up content adjustment plan.

[0027] Step S600: A bidirectional semantic coding model based on the Transformer architecture is used to encode the features of the initial draft of the follow-up content and the personalized adjustment information. Semantic role labeling technology is used to identify the semantic association features of sentence components. Semantic matching technology based on cosine similarity algorithm and word vector model is used to calculate the sentence similarity of the content to be fused. Based on this, the logical structure of the content is optimized to eliminate redundant information, and finally the follow-up content is formed.

[0028] Among these, the following are key technologies: Bidirectional semantic encoding models based on the Transformer architecture: Deep learning models that encode text from both directions (front and back) to capture its semantic features. Semantic role labeling technology: Techniques used to identify the semantic relationships between sentence components and predicates, such as agent, patient, time, and location. Cosine similarity algorithm: A method to measure the semantic similarity of text by calculating the cosine of the angle between two vectors; the closer the value is to 1, the higher the similarity. Word vector models: Models that map words into vector form to represent word semantics and contextual relationships, such as Word2Vec and BERT.

[0029] For the specific process, please refer to steps S610 to S680, which will not be repeated here.

[0030] Step S700: Based on the preset follow-up plan, the follow-up content is combined with the features extracted from the patient's health data, and a reinforcement learning algorithm is used to generate a voice notification strategy.

[0031] Pre-defined follow-up plan: A set of pre-defined rules for patient follow-up time, frequency, and procedures, including basic settings such as regular follow-up cycles and emergency follow-up trigger conditions. Reinforcement learning algorithm: A machine learning algorithm that learns optimal strategies based on reward mechanisms through agent-environment interaction, continuously adjusting behavior to maximize long-term cumulative rewards. Voice notification strategy: A voice notification scheme generated based on patient follow-up needs, including parameter settings such as notification time, tone of voice, and content focus.

[0032] For details of the process, please refer to steps S710 to S760, which will not be elaborated here.

[0033] In step S800, the patient's communication method is automatically matched according to the voice notification strategy, a voice notification is triggered at a specified time, the broadcast parameters are dynamically adjusted, and the patient's feedback information is collected through voice recognition technology after the notification ends.

[0034] The communication methods include: pre-stored notification channels by the patient, such as SMS, telephone, email, and in-app messages. Dynamic adjustment of broadcast parameters involves real-time modification of parameters such as speech rate, tone, and volume based on patient feedback or real-time conditions. Speech recognition technology converts speech signals into text information to extract patient feedback after the voice notification.

[0035] The general process and example are as follows: Based on the voice notification strategy generated in step S700, the system automatically matches the corresponding communication method from the patient's file. For example, if the strategy is set to "emergency reminder" and the patient has provided a mobile phone number, a telephone voice notification is automatically triggered. During the voice notification playback, the system dynamically adjusts the playback parameters according to preset rules or real-time conditions. For example, it automatically increases the volume when loud ambient noise is detected and appropriately slows down the speech for elderly patients. After the notification ends, speech recognition technology is used to convert the patient's voice response into text.

[0036] Reference Figure 2 The final follow-up content includes: Step S610: Obtain data on the patient's home environment and perform data preprocessing.

[0037] Among them, home environment data acquisition: Patient home environment data is collected through various methods, covering physical environment (such as indoor temperature and humidity, noise level, which can be obtained with the help of temperature and humidity sensors and noise testers) and surrounding environment (such as the distribution of nearby hospitals and supermarkets, which can be determined through on-site inspections, online map queries, etc.) to comprehensively reflect the patient's living environment.

[0038] Data preprocessing includes data cleaning, format unification, data normalization or standardization, and data completion and imputation. Data completion and imputation: For missing critical information (such as information on the presence of hospitals near the home), data is supplemented through on-site visits and consultations with community personnel; non-critical estimated data (such as estimations based on house orientation when precise values ​​for lighting are unavailable) is appropriately filled in to ensure data completeness and provide a comprehensive and accurate data foundation for subsequent environmental feature extraction and risk assessment.

[0039] Step S620: Extract environmental features from the preprocessed household environmental data, including time series features, fluctuation features, and correlation features.

[0040] Feature extraction based on time series: For data that changes regularly over time, such as indoor temperature in a home, records values ​​at different time points to form a time series. Features such as daily average temperature, standard deviation of temperature change, time of highest temperature, and frequency of temperature periodic changes are extracted. Fluctuation feature extraction: This reflects the amplitude and trend of data fluctuations over a certain period, such as the fluctuation range and periodicity of indoor humidity during rainy and dry seasons, and the maximum fluctuation amplitude and frequent periods of household electricity consumption. Association feature extraction: Statistical methods such as correlation analysis are used to explore the correlations between different factors in the home environment, such as the positive correlation between indoor temperature and humidity, calculating their correlation coefficient; logical correlations between the completeness of home facilities and the health status of family members, and the convenience of transportation around the home and the frequency of patients seeking medical treatment.

[0041] Step S630: Input the extracted environmental features into the household environmental risk model and output the household environmental risk assessment level.

[0042] Home environment risk model: Constructed based on machine learning (such as decision trees, neural networks) or statistical methods. It involves collecting a large amount of sample data containing home environment characteristics and corresponding patient health status feedback, such as environmental characteristics data like temperature, humidity, and noise levels from numerous families, as well as records of health problems caused by environmental factors, such as respiratory disease recurrences and accidental injuries. After training with appropriate algorithms, such as decision trees to construct a tree-like decision structure, support vector machines to find the classification hyperplane, and neural networks to learn nonlinear relationships, it can accurately output risk assessment levels based on the input environmental characteristics.

[0043] The operation process is as follows: First, organize the family environment feature data extracted in step S620, unify the format (such as standardizing the number of decimal places in the numerical values), and normalize or standardize the data (such as using minimum-maximum normalization to map the feature values ​​to the 0-1 interval) so that feature data of different magnitudes and properties are in a fair comparison range when input into the model, which facilitates accurate calculation and judgment.

[0044] The processed data is then input into the pre-trained family environment risk model. The model calculates and judges based on the mapping relationship between internally learned features and risk levels. For example, in a neural network, feature data is processed by neurons in the hidden layer and then outputs the risk assessment level in the output layer (it may present the probability of each level as a probability and take the highest probability level; or it may directly output explicit labels such as "low risk", "medium risk", "high risk"). Finally, the family environment risk assessment level output by the model is obtained.

[0045] Step S640: Based on the preset mapping table of family environment risk assessment level and integration strategy, determine the follow-up content integration strategy that matches the current family environment risk assessment level.

[0046] Among them, the pre-set family environment risk assessment level and integration strategy mapping table: the pre-set table clarifies the follow-up content integration strategy corresponding to different family environment risk assessment levels. Follow-up content integration strategy: for different family environment risk assessment levels, it formulates methods and rules for effectively integrating the initial draft of follow-up content, personalized adjustment information and family environment risk assessment and other information.

[0047] The general process is described as follows: The system first obtains the current home environment risk assessment level from step S630, such as "high risk," "medium risk," or "low risk." Then, the system searches for a matching integration strategy in a preset home environment risk assessment level and integration strategy mapping table. For example, if the current home environment risk assessment level is "high risk," the corresponding integration strategy in the mapping table might be "highlighting personalized adjustment information related to the risk, strengthening monitoring suggestions for risk factors, and increasing the frequency and depth of follow-up visits."

[0048] Step S650: Using a multimodal fusion module based on a multi-head attention mechanism, and based on a determined fusion strategy, the initial draft of the follow-up content, personalized adjustment information and the family environment risk assessment level are feature-fused to enhance the semantic expression of content related to the risk level.

[0049] Multi-head attention mechanism: A mechanism used in deep learning to process sequential data. It allows the model to focus on different parts of the input sequence simultaneously, thereby better capturing information within the sequence. Attention weights are computed in parallel by multiple "heads," each focusing on a different aspect of the input, and these attention results are finally combined. Multimodal fusion module: A module used to fuse data from multiple different modalities (such as text, images, audio, etc.).

[0050] For details of the process, please refer to steps S651 to S658, which will not be elaborated here.

[0051] Step S660: Use the Transformer bidirectional encoder to perform semantic encoding on the fused content, and combine it with semantic role labeling technology to construct a semantic association graph.

[0052] Among them, the Transformer bidirectional encoder is an encoder based on the Transformer architecture that can encode text simultaneously from both forward and backward directions, thereby better capturing the contextual information of the text. Semantic encoding is the process of converting textual information into a semantic representation that computers can understand and process for subsequent analysis and operations. Semantic role labeling is a natural language processing technique used to identify the semantic roles of various components (such as subject, object, predicate, etc.) in a sentence, that is, their semantic functions and relationships within the sentence. Semantic association graph is a graphical structure representing the semantic relationships between words in text, where nodes represent words and edges represent the strength of semantic associations between words.

[0053] The general process is described as follows: 1. Input the fused content: The fused content obtained in step S650 is input into the Transformer bidirectional encoder. This fused content includes the initial draft of the follow-up content, personalized adjustment information, and information related to the family environmental risk assessment level. 2. Bidirectional encoding: The Transformer bidirectional encoder encodes the input fused content in both forward and backward directions. During the encoding process, it considers the contextual information of the text and calculates the representation of each position through a self-attention mechanism, thereby better capturing the semantics of the text. 3. Semantic role labeling: Using semantic role labeling technology, the encoded content is analyzed to identify the semantic roles of each component in the sentence. 4. Constructing a semantic association graph: Based on the results of semantic role labeling, a semantic association graph is constructed. In the graph, each word is treated as a node, and the semantic relationships between them are treated as edges. 5. Output the result: Finally, a result containing the semantic encoding of the fused content and the semantic association graph is obtained.

[0054] Step S670: Calculate sentence similarity based on cosine similarity algorithm and word vector model, and adjust the calculation weight according to the family environment risk assessment level.

[0055] Among them, word vector models are models that map words to vector spaces, so that words with similar meanings are close in distance in the vector space. Commonly used models include Word2Vec and GloVe. Sentence similarity is an indicator that measures the degree of semantic similarity between two sentences.

[0056] The general process is described as follows: Assume that the follow-up content after fusion in step S650 contains two sentences: Sentence A, "Measure blood pressure at 8 am every day"; and Sentence B, "It is recommended to measure blood pressure at a fixed time in the morning." First, the Word2Vec word vector model is used to convert the words in the two sentences into numerical vectors. For example, words such as "every day," "morning," "measure," and "blood pressure" will be converted into specific vectors, and then the vector representations of sentences A and B are obtained through averaging or other methods. Next, the cosine similarity algorithm is used to calculate the similarity between the two sentences. If the calculation result is close to 1, it means that the two sentences are semantically similar. For example, here the similarity might be calculated to be 0.9, indicating that they both emphasize the matter of measuring blood pressure in the morning.

[0057] Finally, the similarity weight can be adjusted according to the family environment risk assessment level, referring to steps S671 to S675.

[0058] Step S680: Construct a content logic topology graph based on a graph neural network, using sentences as nodes and semantic relevance as edges. Optimize the logical structure through graph convolution algorithm, automatically delete redundant sentence branches with similarity exceeding the threshold, and generate final follow-up content that fits the family environment risk.

[0059] Among them, Graph Neural Networks (GNNs) are neural networks that run directly on graph-structured data. They update node features through information transmission between nodes and edges, and are used to mine relationships between data. Content Logic Topology Graphs are graphical structures that use sentences in the followed content as nodes and semantic relationships between sentences as edges, intuitively displaying the logical relationships within the content. Graph Convolution Algorithm is the core algorithm in graph neural networks, which optimizes the graph structure by aggregating information from neighboring nodes to update the features of the current node.

[0060] The general process is described as follows: Taking a specific scenario as an example, assume that the follow-up content processed in step S670 includes the following statements: A. "Measure blood pressure at a fixed time every day"; B. "It is recommended to monitor blood pressure at a fixed time every day"; C. "Keep quiet when measuring in a high-noise environment"; D. "Choose a quiet time for measurement". First, using statements A, B, C, and D as nodes, the statement similarity calculated in step S670 (e.g., A and B are similar to 0.9, C and D are similar to 0.8) is used as the edge weights to construct a content logical topology graph. At this point, A and B are closely connected due to their semantic similarity, and C and D are also associated. Next, a graph convolution algorithm is used to optimize the logical structure. The algorithm traverses each node and aggregates information from adjacent nodes. For example, for node A, it absorbs the feature information of adjacent node B and finds that A and B have highly repetitive semantics. Based on the set similarity threshold (e.g., 0.85), A and B are determined to be redundant content, and one of the statement branches is automatically deleted (e.g., B is deleted). Similarly, C and D are processed to retain key information and generate logically clear and repetitive final follow-up content: "Measure blood pressure at regular times every day, especially in noisy environments; choose quiet times for measurement." In this way, based on graph neural networks and graph convolution algorithms, the logical structure of the follow-up content is optimized and redundancy is eliminated.

[0061] The initial draft of the follow-up data, personalized adjustments, and the family environmental risk assessment level were integrated to include the following features: Step S651: Construct a risk-level adapted multimodal attention fusion framework. Preset high, medium and low family environment risk assessment levels to correspond to independent attention module groups. The high-risk module group includes a gated multi-head attention module with dynamic routing and is configured with risk-sensitive weight adjustment parameters and semantic processing logic. The medium and low-risk module groups include an adaptive weight balancing attention module and are configured with semantically balanced weight adjustment parameters and semantic processing logic.

[0062] The multimodal attention fusion framework integrates a fusion model that combines initial drafts of follow-up content (text modality), personalized adjustment information (text / numerical modality), and family environment risk assessment levels (numerical modality), dynamically allocating weights for different modalities through an attention mechanism. Risk grading adaptation: Independent attention module groups are pre-set for high, medium, and low levels of family environment risk, matching the semantic processing logic under different risk scenarios. Gated multi-head attention module with dynamic routing: A dedicated module for high-risk situations, it selects key semantic paths through a dynamic routing algorithm, filters non-core information using a gating mechanism, and strengthens the attention allocation for risk-related content. Adaptive weight balancing attention module: A dedicated module for medium and low-risk situations, it dynamically adjusts the weights of core and non-core content based on a balanced loss function, ensuring semantic coherence and information balance.

[0063] Weighting parameters: Obtained through training on historical risk-annotated data, such as using a follow-up content dataset containing different risk levels, with risk semantic density as the objective function to optimize the parameters. Semantic processing logic: A built-in rule base combined with a deep learning model, such as forcibly prioritizing the processing of core risk statements in high-risk scenarios, and restricting the excessive prominence of risk content in medium- and low-risk scenarios.

[0064] The general process is described below: Framework architecture design: A three-layer structure is established. The input layer receives the initial draft of the follow-up content, personalized adjustment information (such as the patient's allergy history and medication habits), and risk level (1-3). The middle layer calls the corresponding module group according to the risk level (high risk → dynamic routing gating module, medium / low risk → adaptive balancing module). The output layer generates the fused content feature vector.

[0065] Module Differentiation Configuration: High-Risk Module: Enables gated multi-head attention with dynamic routing. A risk-sensitive mask is generated through a gating mechanism (e.g., the Sigmoid function) to suppress non-core content (e.g., daily greetings) and strengthen the attention weight of core risk statements (e.g., "High temperature warning, reduce outdoor activities"). (Example weight allocation: 80% for core content, 20% for non-core). Medium-Low Risk Module: Employs adaptive weight balancing attention. The weight ratio of core to non-core content is dynamically adjusted through a balancing loss function (e.g., KL divergence) to ensure semantic balance between risk warnings and daily suggestions. Parameter Initialization: Presets the dynamic routing parameters for high-risk modules (e.g., routing decision threshold 0.7), the suppression threshold for gating modules (non-core content attention weight ≤ 0.3), and the balancing loss weight coefficient for medium- and low-risk modules (core content weight fluctuation range ±10%).

[0066] Step S652: Based on the preset risk keyword library and content classification model, the follow-up content is divided into core risk response content and non-core content.

[0067] The system comprises several components: a risk keyword database (a pre-compiled set of terms related to home environmental risks, such as "gas leak," "heatstroke," and "dampness and mold"), used to identify risk-related content; and a content classification model (a machine learning-based model that automatically categorizes follow-up content into core risk response content and non-core content based on the correlation between text content and the risk keyword database). Core risk response content provides key information on prevention and response measures directly addressing home environmental risks, such as "open windows for ventilation immediately" and "suspend the use of gas appliances." Non-core content contains information with lower relevance to risks, such as daily health reminders and medication advice.

[0068] The process is briefly described as follows: Keyword Matching: The follow-up content is compared with a risk keyword database, and the frequency and location of risk keywords in each piece of content are statistically analyzed. For example, if a text contains keywords such as "gas leak" or "close valve," it is marked as potentially risk-related. Model Classification: The text processed by keyword matching is input into a trained content classification model, which outputs classification results (core risk response content or non-core content). For example, the TextCNN model extracts text features through convolutional layers and outputs classification probabilities through fully connected layers. If the probability of the core risk category exceeds 0.7, it is determined to be core risk response content. Content Segmentation: Based on the model output, the follow-up content is divided into two categories: core risk response content and non-core content, for use in subsequent fusion steps.

[0069] Step S653: When the risk assessment level of the home environment is high risk, activate the gated multi-head attention module with dynamic routing, generate a risk-sensitive attention mask, apply a preset suppression threshold to non-core content, adjust the order of statements according to the preset risk occurrence probability priority queue for core risk response content, force the core prompt statements with high semantic intensity to be placed first, assign a preset gain threshold to words in the risk keyword library, strengthen the semantic association of the context through multi-head attention calculation, and output a content sequence that highlights the core risk information.

[0070] Among them, the risk-sensitive attention mask is a binary matrix generated based on high-risk scenarios, used to suppress the attention weight of non-core content and highlight core risk information. The preset suppression threshold is a pre-set value used to limit the weight of non-core content in the attention calculation, such as limiting the attention weight of non-core content to below 0.2. The preset risk occurrence probability priority queue is a queue formed by sorting the occurrence probabilities of various risks based on historical data and expert experience, used to adjust the sentence order of core risk response content. The preset gain threshold is a pre-set value used to enhance the weight of risk keywords in the attention calculation, such as increasing the attention weight of risk keywords by 1.5 times.

[0071] The process is briefly described as follows: 1. Module Activation: When the home environment risk assessment level is determined to be high risk, the gated multi-head attention module with dynamic routing is activated. 2. Mask Generation and Suppression: The module generates a risk-sensitive attention mask and applies a preset suppression threshold to the non-core content, reducing its weight in the attention calculation. For example, the attention weight of non-core content is reduced from 0.5 to 0.1. 3. Content Sorting: For core risk response content, the order of statements is adjusted according to a preset risk occurrence probability priority queue. For example, if the probability of fire risk is higher than that of slipping, the fire response statement is placed before the slipping response statement. 4. Keyword Enhancement: Preset gain thresholds are assigned to words in the risk keyword library. Through multi-head attention calculation, the semantic relevance of these keywords to the context is enhanced. For example, the attention weight of the keyword "gas leak" is increased from 0.3 to 0.45. 5. Sequence Output: After the above processing, a content sequence highlighting core risk information is output, ensuring that high-risk key information is presented first.

[0072] Step S654: When the risk assessment level of the family environment is medium or low risk, the adaptive weight balance attention module is activated. Based on the preset balance loss function, the weight ratio of core and non-core content is dynamically adjusted, and the connection between non-core content and core prompts is optimized through the semantic smoothing mechanism to output a semantically coherent sequence that balances core and non-core content.

[0073] Predefined Balanced Loss Function: A predefined loss function used to measure the balance between core and non-core content in semantic expression, guiding weight adjustments. Semantic Smoothing Mechanism: A mechanism that optimizes the semantic transition between non-core content and core prompts, ensuring a natural and smooth connection between the two.

[0074] The process is briefly as follows: 1. Module Activation: When medium or low risk levels are detected, the adaptive weighted attention module is activated. 2. Dynamic Weight Adjustment: A weighted cross-entropy loss function is used. ,in The labels are real (1 indicates key information, 0 indicates non-key information). To predict probabilities, The balancing coefficient is set at 0.6 for medium risk and 0.4 for low risk. The loss function is minimized using the backpropagation algorithm, dynamically adjusting the weights of core and non-core content in attention calculation to achieve a more balanced semantic expression. 2. Semantic smoothing: If core and non-core content are detected to be adjacent, a transitional phrase template is selected from the rule base, such as "In addition, daily attention should also be paid to...", and inserted between the two types of content; or the tone intensity of the core prompt is adjusted to ensure a natural transition with the non-core content. 3. Coherent sequence output: The adjusted content is integrated to output a balanced and semantically coherent follow-up content sequence.

[0075] Step S655: Using a context-aware sliding window mechanism, when a preset number of consecutive non-core contents are detected in the output content sequence of all risk levels, risk warning phrases are generated according to the risk level and the insertion probability is adapted to form fused content.

[0076] The context-aware sliding window mechanism is a text processing method that scans the content sequence using a fixed-length sliding window and judges content features based on contextual information. Preset quantity: A pre-defined threshold for the number of consecutive non-core content items, such as 3 consecutive non-core items. Risk level-adapted insertion probability: The probability of inserting risk warning phrases based on different risk levels, such as 80% for high-risk scenarios, 50% for medium-risk, and 30% for low-risk. Risk warning phrase: A concise risk reminder statement, such as "Pay attention to regularly checking gas equipment." The process is briefly described as follows: 1. Window scanning: A fixed-length sliding window (e.g., 5 items) slides across the content sequence output by the module to extract the content within the current window. 2. Content judgment: The content within the window is analyzed to determine if there is a preset number (e.g., 3) of consecutive non-core content. 3. Probabilistic insertion: If consecutive non-core content is detected, a risk warning phrase is randomly inserted based on the insertion probability corresponding to the current risk level. For example, in a high-risk scenario, there is an 80% probability of randomly selecting a phrase from the risk warning phrase library and inserting it at the current position. 4. Content Integration Generation: After completing all insertion operations, a sequence of integrated content is generated, which includes naturally integrated risk warnings.

[0077] Step S656: The discriminator network performs risk semantic adaptability verification on the fused content, and determines whether it meets the standard based on a preset semantic feature matching threshold. If yes, proceed to step S657; if no, proceed to step S658.

[0078] The discriminator network is a deep learning-based binary classification model used to determine whether the fused content meets the semantic adaptation requirements of the corresponding risk level, outputting a "compliant" or "non-compliant" result. The discriminator network is trained using a labeled fused content dataset (labeled for compatibility with the corresponding risk level) based on a convolutional neural network (CNN) or Transformer architecture, for example, using the TextCNN model. Risk semantic adaptation refers to the completeness, accuracy, and degree of matching between the risk-related information in the fused content and the current risk assessment level of the family environment. The semantic feature matching threshold is a pre-set quantitative indicator used to measure the similarity between the semantic features of the fused content and the semantic features of the target risk level; for example, a threshold of 0.7 (out of 1) is used.

[0079] The process is briefly described as follows: 1. Feature Extraction: The fused content is input into the discriminator network, and semantic feature vectors are extracted through convolutional layers (such as the multi-layer convolutional kernels of TextCNN) or attention mechanisms (such as the multi-head attention of Transformer). Examples include extracting risk keyword density and sentiment tendency of sentences. 2. Similarity Calculation: The extracted semantic feature vectors are compared with a preset target risk level semantic feature template (constructed by expert-annotated semantic features of typical risk scenarios) using cosine similarity calculation. The formula is: ;in, To fuse content feature vectors, 3. Threshold judgment: Compare the calculated similarity score with the preset semantic feature matching threshold (e.g., 0.7): If the score is ≥ the threshold, the semantic adaptability of the fused content risk is deemed to meet the standard, and proceed to step S657 (content format standardization); if the score is < the threshold, it is deemed not to meet the standard, and proceed to step S658 (reverse optimization parameters).

[0080] Step S657: Standardize the format of the merged content and adapt it to different follow-up channels.

[0081] Format standardization: Convert the merged content into a unified and standardized text structure, including paragraph division, punctuation correction, and terminology standardization. Follow-up channels: Methods for pushing follow-up content to patients, such as SMS, app messages, voice calls, and emails. Content adaptation: Adjust the presentation of content according to the characteristics of different channels (such as character limits and interaction methods) to ensure effective information delivery. Channel adaptation is as follows: 1. SMS: Break down long texts, simplify content, and highlight core risks. 2. App: Add icons, bold key information, and display in segments with collapsible sections. 3. Voice calls: Convert to conversational language and mark pauses. 4. Emails: Use HTML formatting, add color blocks and jump links.

[0082] Step S658: Apply a preset gradient penalty threshold to the output results that do not meet the standard, and drive the attention module group, content classification model and discriminator network in the multimodal attention fusion framework to optimize the parameters until the risk-related semantic density reaches the preset improvement level.

[0083] Among them, the preset gradient penalty threshold is a pre-set gradient penalty coefficient used to adjust the step size and direction of model parameter updates, such as setting it to 0.5. Risk-related semantic density is the proportion of risk-related semantic information in the fused content, quantified by calculating indicators such as the frequency of risk keywords and the weight of key sentences. The preset improvement margin is the minimum percentage that the risk-related semantic density needs to be increased, such as 15%.

[0084] The process is briefly described as follows: 1. Penalty Application: For outputs that do not meet the criteria, a gradient penalty term is added to the loss functions of the discriminator network, content classification model, and attention module group during backpropagation. The formula is: ,in The original loss function, Set a preset gradient penalty threshold (e.g., 0.5). For the discriminator network output, 1. The square of the gradient norm. 2. Parameter optimization: Use an optimization algorithm (such as Adam) to minimize the total loss function and update the parameters of the attention module group, content classification model, and discriminator network to make the model pay more attention to risk-related content. 3. Iterative training: Repeat the above process until the risk-related semantic density increases by a preset amount (such as 15%) compared to before optimization. At this point, the model is considered to have converged, and the optimized fused content is output.

[0085] The weighting is adjusted based on the household environmental risk assessment level, including: Step S671: Construct a quantitative mapping table between the three levels of high, medium, and low risk in the home environment and preset weight adjustment parameters.

[0086] Among them, the preset weight adjustment parameters are pre-defined parameters used to adjust the calculation weights under different risk levels. The quantitative mapping table is a table that correlates the household environmental risk assessment level with the preset weight adjustment parameters, facilitating quick retrieval of the corresponding weight adjustment parameters based on the risk level.

[0087] Step S672: Based on the quantitative mapping table, when the risk level is high, increase the weight of risk keywords by the first preset ratio and decrease the weight of non-risk content by the second preset ratio; when the risk level is medium, fine-tune the weight of risk keywords by the third preset ratio and fine-tune the weight of non-risk content by the fourth preset ratio; when the risk level is low, decrease the weight of risk keywords by the fifth preset ratio and increase the weight of daily content by the sixth preset ratio.

[0088] Risk keywords: Words directly related to home environmental risks in the follow-up content, such as "high noise," "dampness," and "electrical leakage." Non-risk content: General information in the follow-up content that does not involve home environmental risks, such as daily health tips and lifestyle suggestions. Daily content: General health advice related to daily life. First to sixth preset ratios: Pre-set numerical ratios used to adjust the weight of risk keywords, non-risk content, and daily content under different risk levels.

[0089] Step S673: Using a multi-head attention mechanism, under high-risk levels, the proportion of the number of attention heads for risk keywords to the total number of attention heads is set to a preset value.

[0090] Among them, risk keywords are important terms related to home environmental risks, such as "high noise," "dampness," and "flammable materials," which play a key role in describing the risk status of the home environment. Number of attention heads: In a multi-head attention mechanism, the number of "heads" determines how many different angles the model can focus on the input information. Each "head" can learn different attention patterns. Preset value: A pre-set percentage value used to specify the proportion of risk keyword attention heads to the total number of attention heads at high risk levels.

[0091] The general process is as follows: When the home environment risk assessment level output in step S630 is high risk, step S673 is triggered. In a multi-head attention mechanism, a fixed total number of attention heads is usually set, for example, a total of 8 attention heads. The number of risk keyword attention heads is determined based on a preset value. Assuming the preset value is 0.75, then the number of risk keyword attention heads is 6, and the remaining 2 attention heads are used to focus on other non-risk related information.

[0092] Step S674: Calculate sentence similarity using a pre-defined bidirectional encoder representation and word vector fusion model. Based on the family environment risk assessment level, dynamically adjust the word vector weights according to pre-defined rules for similarity calculation. The bidirectional encoder representation and word vector fusion model is a model that combines the representation capabilities of a bidirectional encoder with word vector information, converting text into vector representations for subsequent similarity calculation. Sentence similarity: Measures the semantic similarity between two sentences. Word vector weight: The weight of each word vector in calculating sentence similarity, reflecting the importance of the word in the sentence. Pre-defined rules: Pre-set rules for adjusting word vector weights based on the family environment risk assessment level.

[0093] Step S675: Through a preset reinforcement learning module, the weight adjustment parameters are optimized using the Q-learning algorithm in reinforcement learning, with the preset weighted sum of risk warning clicks, content browsing time, and user satisfaction rating as the reward function.

[0094] Among them, the risk warning click count refers to the number of times users click on risk warning information in the follow-up content, reflecting the user's level of attention to the risk warning. Content browsing time refers to the time users spend browsing the follow-up content, reflecting their interest and engagement with the content. User satisfaction rating refers to the ratings given by users regarding the quality and usability of the follow-up content, directly reflecting their satisfaction with the content. The reward function is a function used in reinforcement learning to measure the quality of agent behavior; in this step, it is defined as a weighted sum of the risk warning click count, content browsing time, and user satisfaction rating. The Q-learning algorithm is a model-free reinforcement learning algorithm that finds the optimal policy by learning the Q-value (expected reward) of action-state pairs.

[0095] The general process is described as follows: Define the reward function: Let C be the number of risk warning clicks, T be the content browsing time, and S be the user satisfaction rating. The preset weighting coefficients are as follows: , , The reward function R is defined as: ,For example, If, in a certain follow-up, the number of risk warning clicks is C=5, the content browsing time is T=10 minutes, and the user satisfaction score is S=8, then the reward function value is R=9.9.

[0096] The optimization steps of the Q-learning algorithm are as follows: 1. The agent (reinforcement learning module) takes the current weight adjustment parameters as state s and the adjustment operation on the weight adjustment parameters as action a. 2. In each interaction cycle, the agent selects action a based on the current state s and updates the weight adjustment parameters. 3. The system adjusts parameters based on the new weights to generate follow-up content, collects user risk warning click counts, content browsing duration, and user satisfaction ratings, and calculates the reward function value R. The agent uses the Q-learning algorithm's update formula to update the Q-value: ;in, It's the learning rate. It is a discount factor. This is the new state after performing action 'a'. By continuously repeating the above process, the agent learns the optimal weight adjustment strategy by maximizing the long-term cumulative reward, that is, finding the weight adjustment parameters that maximize the reward function value.

[0097] Based on the pre-set follow-up plan, the follow-up content, combined with features extracted from patient health data, is used to generate voice notification strategies using reinforcement learning algorithms. These strategies include: Step S710: Obtain patient preference-related information.

[0098] The preference-related information includes language, timbre, and speech rate. Language: The language the patient prefers for the voice notification, such as Chinese, English, or French. Timbre: The patient's preferred voice characteristics, such as gentle, calm, or lively. Speech rate: The patient's preferred speaking speed, such as fast, medium, or slow.

[0099] Step S720: Input the patient's preference-related information into the preset customized voice notification model, and output a customized voice notification configuration including voice language type, timbre parameters, and speech rate parameters.

[0100] The customized voice notification model is a pre-trained model that outputs personalized voice notification configurations based on the patient's input, including information related to patient preferences. The customized voice notification configuration includes a set of settings for speech language type, timbre parameters, and speech rate parameters, used to generate voice notifications that match the patient's preferences. For example, if the patient's preference information shows a preference for Chinese language, a gentle timbre, and a medium speech rate, the model will process this input information.

[0101] The model likely employs deep learning algorithms, such as neural networks. During training, it learns the mapping between a large amount of patient preference information and corresponding appropriate voice notification configurations. When new patient preference information is received, it infers based on its learned knowledge. For language preference, the model directly identifies the speech language type as Chinese. For timbre preference, the model selects the parameter combination that best matches a gentle timbre from a pre-defined timbre parameter library. For speech rate preference, the model determines appropriate speech rate parameters based on the definition of medium speed, such as 150-180 words per second. Finally, the model outputs a customized voice notification configuration, for example, with the speech language type set to "Chinese," the timbre parameters set to a set of values ​​corresponding to a gentle timbre, and the speech rate parameter set to "160 words per second." This configuration will be used to generate voice notifications that meet the individualized needs of patients.

[0102] Step S730: Integrate patient health data features, the final follow-up content, and customized voice notification configuration to construct a multi-dimensional state vector.

[0103] Multidimensional state vector: A vector formed by integrating patient health data features, current follow-up content, and customized voice notification configuration, used in reinforcement learning algorithms to describe the current state.

[0104] First, the patient's health data features are quantified. For example, a patient's blood pressure of 120 / 80 mmHg and heart rate of 75 beats / minute can be converted into a numerical vector [120, 80, 75]. For the follow-up content, text vectorization methods such as bag-of-words model and word embedding can be used to convert it into a numerical vector. Assuming the follow-up content is "Remember to take your blood pressure medication on time," the processed vector is [0.1, 0.2, ..., 0.3]. The voice language type, timbre parameter, and speech rate parameter in the customized voice notification configuration are also converted into numerical forms. For example, a Chinese voice language type is encoded as 1, a gentle timbre parameter is 0.8, and a medium speech rate parameter is 160, converted to a standardized value of 0.6. Finally, these vectors are concatenated sequentially to form a multi-dimensional state vector. For example: [120, 80, 75, 0.1, 0.2, ..., 0.3, 1, 0.8, 0.6].

[0105] Step S740: Based on the constructed multi-dimensional state vector and the constraints of the actual application scenario, determine the set of executable actions, including the selection of discrete time points based on a preset follow-up time window, the selection of preset communication methods, the adjustment of broadcast parameters based on customized configuration, and the selection of content presentation format. The constraints of the actual application scenario include factors such as follow-up time window limitations, communication device compatibility, and patient preferences. The set of executable actions includes operation options in multiple dimensions, such as time selection, communication method, broadcast parameters, and content presentation format. A brief explanation of the process is as follows: Define options: Time: Discrete time points within the preset window (e.g., 9:00, 9:30). Communication: Available methods for patient registration (SMS, APP, telephone). Broadcast parameters: Speech rate, volume, etc. (e.g., medium speed, 70% volume). Content format: Presentation methods adapted to the channel, such as text, images, and voice. Filtering and combination: Based on the constraints (e.g., patients only receive notifications on weekdays), select feasible action combinations to form the final set of executable actions.

[0106] Step S750: Construct a comprehensive reward mechanism, and set quantitative rewards and penalties based on patient response feedback, preference matching degree, and the reasonableness of notification time.

[0107] The comprehensive reward mechanism is a set of rules used to measure the effectiveness of voice notification strategies. It sets quantified rewards and penalties based on multiple factors to guide the reinforcement learning algorithm to find the optimal strategy. Patient response feedback refers to the patient's specific behavioral reactions after receiving the voice notification, such as whether they reply promptly or act according to the notification content. Preference matching degree refers to the degree to which the various settings of the voice notification (such as language, tone, and speech rate) match the patient's preferences. Notification timing rationality refers to whether the timing of the voice notification aligns with the patient's daily routine and receiving habits. Quantified rewards and penalties convert factors such as patient response feedback, preference matching degree, and notification timing rationality into specific numerical rewards or penalties for strategy evaluation in the reinforcement learning algorithm.

[0108] Step S750 includes the following steps: Step S751: Construct a comprehensive reward mechanism, combining patient response feedback, preference matching degree, reasonable notification time and equipment energy consumption to set quantitative rewards and penalties.

[0109] Step S752: Using the NSGA-II algorithm, the quantized values ​​of the four objectives are used as input. Through selection, crossover, and mutation operations, the optimal solution set is found among multiple objectives. The NSGA-II algorithm is a non-dominated sorting genetic algorithm II, an algorithm that solves multi-objective optimization problems by simulating biological evolution, used to find the optimal set of strategies that balance multiple objectives. The general process is described as follows: 1. Population initialization: Randomly generate an initial population containing various voice notification strategies (such as different notification times, communication methods, and combinations of broadcast parameters), with each strategy corresponding to one individual. 2. Non-dominated sorting and crowding calculation: Non-dominated sorting is performed on the individuals in the population according to the four quantized objective values, dividing the individuals into different levels; simultaneously, the crowding degree of each individual in the population is calculated, reflecting the density of its surrounding individuals, used to preserve diversity during selection. 3. Selection operation: Based on the non-dominated level and crowding degree, a tournament selection method is used to randomly select several individuals from the population, choosing those with high levels and low crowding degrees to enter the next generation of the population. For example, each time, randomly select 5 individuals, choosing the one with the highest level (e.g., prioritizing first-level non-dominated individuals) and the lowest crowding. 4. Crossover and Mutation Operations: For the selected individuals, perform a crossover operation with a certain crossover probability (e.g., 0.8) to exchange some policy parameters and generate offspring individuals; perform a mutation operation with a lower mutation probability (e.g., 0.01) to randomly change some policy parameters, such as changing the notification time or adjusting the speech rate. 5. Iterative Optimization: Repeat the above non-dominated sorting, selection, crossover, and mutation operations to continuously update the population and gradually approach the optimal solution set until the stopping condition is met (e.g., reaching the preset number of iterations).

[0110] Step S753: The Transformer context encoder is used to encode information such as device status and patient schedules, dynamically adjusting the priority and method of voice notifications. The Transformer context encoder is a model component based on the Transformer architecture that encodes device status, patient schedules, and other information into computer-processable vectors, uncovering semantic relationships and potential patterns. Device status refers to the operational status of the device (such as a mobile phone or smartwatch) involved in the voice notification, including battery level, network connection, and silent mode. Patient schedule refers to the patient's daily activity arrangements, such as work, rest, and medical appointment time slots. The general process is described as follows: 1. Information Input and Encoding: Device status (e.g., "phone battery 15%", "network disconnected") and patient schedule (e.g., "9:00-12:00 meeting") are formatted into text and input into the Transformer context encoder. The encoder captures key features in the information through a multi-head attention mechanism, transforming them into feature vectors. For example, "phone battery 15%" is encoded as a vector containing a low battery warning. 2. Policy Adjustment Decision: A rule base is established to associate the encoded feature vectors with the notification policy. For example, when the device battery is below 20% and the patient is in a busy schedule, the notification priority is set to low, and a vibration plus a short text reminder is used; when the device is in normal condition and the patient is in an idle period, a detailed voice broadcast is used and the notification priority is increased. 3. Real-time adjustment: The system monitors device status and schedule changes in real time. Once the rule conditions are triggered, the system immediately adjusts the strategy of voice notifications that are being generated or are about to be sent. For example, if the patient adds a meeting to their schedule, the system automatically changes the originally planned voice notification to a vibration reminder.

[0111] Step S754: Based on the optimal notification strategy obtained by the NSGA-II algorithm and the context-aware scheduling results, combined with the quantitative scores of each objective, calculate the comprehensive reward to evaluate the merits of the voice notification strategy.

[0112] Optimal Notification Strategy: A voice notification scheme optimized by the NSGA-II algorithm that achieves a balance among multiple objectives, including patient response feedback, preference matching degree, notification timing rationality, and device energy consumption. Context-Aware Scheduling Result: The execution result of dynamically adjusting notification priority and method after encoding information such as device status and patient schedule by the Transformer context encoder. Comprehensive Reward: A numerical value obtained through quantitative calculation, used to measure the comprehensive performance of the voice notification strategy across multiple objective dimensions; the higher the value, the better the strategy. The general process is described as follows: 1. Data Integration: Collect the optimal notification strategy generated by the NSGA-II algorithm (e.g., notification time at 10:00 AM, using APP voice push, speech rate of 180 words / minute), context-aware scheduling results (e.g., volume reduction due to low device battery), and quantitative scores for each objective (e.g., patient response feedback 8 points, preference matching degree 7 points). 2. Weighted Calculation: Calculate and sum the scores of each objective according to preset weights. Assuming the weights for patient response, preference matching, notification timing rationality, and device energy consumption are 0.4, 0.3, 0.2, and 0.1 respectively, the comprehensive reward calculation formula is: Comprehensive Reward = 0.4 × Patient Response Score + 0.3 × Preference Matching Score + 0.2 × Notification Timing Rationality Score + 0.1 × Device Energy Consumption Score. Application: The calculated comprehensive reward value is used as the evaluation criterion; a higher value indicates that the voice notification strategy better meets multi-objective needs. For example, the calculated comprehensive reward is 8.2 points.

[0113] Step S760: A reinforcement learning algorithm is adopted, which is initialized with an ε-greedy policy with a preset ε value. The policy parameters are iteratively updated according to a preset learning rate and a preset discount factor, based on a multi-dimensional state vector and a comprehensive reward mechanism. In each iteration, the reinforcement learning algorithm tries to select different actions and adjusts the policy according to the reward feedback. When the rate of change of the update is less than a preset proportion for a preset number of consecutive rounds, it converges and finally generates the optimal voice notification policy that includes notification time, communication method and broadcast parameters.

[0114] Reinforcement learning algorithm: An algorithm in which an agent learns an optimal policy by interacting with the environment with the goal of maximizing cumulative reward. ε-Greedy policy: A policy that balances exploration (trying new actions) and exploitation (choosing the known best action). A preset ε value determines whether to randomly select an action (exploration) or choose the current best action (exploitation). Preset ε value: A pre-set probability value used in the ε-Greedy policy to control the ratio of exploration to exploitation. Preset learning rate: A parameter that controls the magnitude of policy parameter updates in the reinforcement learning algorithm, affecting the speed and stability of learning. Preset discount factor: Used to calculate the discounted value of future rewards, reflecting the importance attached to future rewards. Policy parameters: A set of parameters used in the reinforcement learning algorithm to describe the policy; the policy is optimized by continuously updating these parameters. Update change rate: The proportion of change in policy parameters between two consecutive updates, used to determine whether the algorithm has converged. Preset proportion: A pre-set threshold; when the update change rate for a preset number of consecutive rounds is less than this proportion, the algorithm is considered converged. Optimal voice notification strategy: After optimization by reinforcement learning algorithm, the voice notification strategy that maximizes the overall reward includes notification time, communication method, and broadcast parameters.

[0115] Step S760 uses a reinforcement learning algorithm to generate the optimal voice notification strategy. The specific process is as follows: 1. Initialization: Initialize using an ε-greedy policy with a preset ε value. For example, a preset ε value of 0.1 means that in each step, there is a 10% probability of randomly selecting an action for exploration and a 90% probability of selecting the currently considered optimal action. 2. Iterative Update: The agent selects an executable action using the ε-greedy policy based on the current multidimensional state vector. 3. After executing the action, a reward value is obtained according to the comprehensive reward mechanism. The policy parameters are updated using the update formula of a reinforcement learning algorithm (such as Q-learning) with a preset learning rate and a preset discount factor. For example, the update formula for Q-learning is... ,in It is the preset learning rate. It is a preset discount factor. Here, s is the reward value, s is the current state, and a is the action performed. This is the next state. 4. Convergence Judgment: After each update, calculate the rate of change of the strategy parameters. When the rate of change of the strategy parameters is less than a preset proportion (e.g., 0.01) for a consecutive preset number of rounds (e.g., 10 rounds), the algorithm is considered to have converged. 5. Strategy Generation: After the algorithm converges, the resulting strategy is the optimal voice notification strategy, which includes the notification time, communication method, and broadcast parameters. For example, the final strategy might be a notification time of 10:00 AM, a communication method of in-app voice push, and broadcast parameters of 70% volume and 160 words per second.

[0116] A patient follow-up management method further includes steps after acquiring patient feedback information through speech recognition technology, specifically as follows: Step SA00, using natural language processing technology, extract keywords and sentences related to health status, lifestyle, and environment from the patient feedback information.

[0117] Among them, natural language processing technology: a technology that uses a computer to process and analyze human language, including word segmentation, part-of-speech tagging, named entity recognition, etc. Keywords and sentences: words or phrases directly related to health status (such as "headache", "insomnia"), lifestyle (such as "exercise", "diet"), and environment (such as "humidity", "noise") in the feedback information.

[0118] The brief process is described as follows: 1. Text preprocessing: Clean the patient feedback information, including removing stop words (such as "de", "le"), special symbols, and unifying case. 2. Part-of-speech tagging and named entity recognition: Identify nouns and verbs in the text through a pre-trained medical NLP model (such as BioBERT), and label health-related entities (such as "hypertension", "diabetes"). 3. Keyword matching: Match the processed text with three preset keyword libraries for health, lifestyle, and environment. For example: Health keyword library: contains words such as "pain", "fatigue", "blood sugar", etc. Lifestyle keyword library: contains words such as "smoking", "exercise frequency", "sleep time", etc. Environment keyword library: contains words such as "temperature", "humidity", "ventilation", etc. 4. Sentence extraction: Extract complete sentences containing keywords, such as "The sleep quality has been very poor recently, and I can only sleep for 4 hours every night" (lifestyle category).

[0119] Step SB00, classify the extracted keywords and sentences, and the classification can be specifically divided into health status category, lifestyle category, and environment category.

[0120] Among them, classification: Map the extracted keywords and sentences to three preset categories of health status, lifestyle, and environment. Health status category: Information directly related to the patient's physical state, such as symptoms, diseases, and physiological indicators. Lifestyle category: Information related to the patient's daily habits, such as diet, exercise, and schedule. Environment category: Information related to the space where the patient lives or moves, such as temperature, humidity, and noise.

[0121] The process is briefly described as follows: 1. First, based on a keyword-category mapping rule base established by medical experts, keywords and phrases are quickly classified. For example, "cough" directly matches to the health status category, and "running every day" is categorized into the lifestyle category. 2. For content not covered by the rule base, a BERT or TextCNN multi-classification model trained on medical labeled data is used to predict and determine its category. If a keyword matches multiple categories simultaneously, the final classification is determined based on contextual semantic analysis or according to a weighted voting method set by experts.

[0122] Step SC00: For each category, the extracted keywords and phrases are matched and analyzed from the pre-established category-related device mapping database to obtain the mapped devices, which serve as the initial device set.

[0123] The category-related device mapping database is a pre-built database that has been validated by professional medical knowledge, health management experience, and other aspects. It stores the correspondence between different categories (i.e., health status categories, lifestyle categories, and environmental categories) and various devices that may help or are related to the patient's health.

[0124] The initial set of devices is a collection of devices obtained by matching and analyzing the mapping relationship between categories and associated devices in this step (step SC00) based on the different categories of content obtained after classifying and processing patient feedback information.

[0125] Step SD00: For the initial set of devices, a preset index quantification evaluation method is used to analyze and obtain the score assigned to each device under each index. A preset weight allocation method is used to calculate the comprehensive priority score, and all devices are sorted in descending order based on the calculated comprehensive priority score.

[0126] The pre-defined quantitative evaluation method is a set of pre-defined rules and calculation methods used to quantitatively measure the performance of equipment in different aspects. It typically considers multiple dimensions of indicators, such as the equipment's functional effectiveness (i.e., whether the equipment can accurately achieve its intended health monitoring and auxiliary improvement functions, such as the accuracy of blood pressure measurement by an electronic blood pressure monitor, quantified by values ​​such as the measurement error range), ease of use (e.g., the complexity of operating steps, quantified by the number of operations and the time required; the simpler the operation, the higher the score), reliability (the probability of equipment failure within a certain usage period, quantified by the proportion of failures to total usage), and cost-effectiveness (calculated using a specific cost-benefit analysis model, combining the equipment's price and the functional value it provides).

[0127] The preset weighting method is a pre-determined weighting approach based on the relative importance of each indicator in measuring the overall value and significance of the equipment. Different weight values ​​are assigned to different indicators based on a combination of factors, including medical expertise, health management experience, and the actual needs of patients.

[0128] The overall priority score is a numerical value calculated by comprehensively considering the quantitative scores of the device under various preset indicators and the preset weights of each indicator. It reflects the overall priority recommendation degree of the device. The higher the score, the more worthy the device is of priority recommendation to patients, considering factors such as meeting patients' health needs, ease of use, and cost-effectiveness.

[0129] The specific operation process is as follows: First, obtain the initial set of devices (derived from the operation result of step SC00), ensuring that the device information within the set is complete and accurate, including device name, model (if differentiated), basic function description, etc., to prepare for subsequent quantitative evaluation of indicators. Then, for each device in the initial set of devices, use the preset quantitative evaluation method to evaluate its score.

[0130] Taking a fitness tracker as an example, in terms of functional effectiveness: to evaluate the accuracy of its monitoring of exercise data (such as steps, distance, calories burned, etc.), the measurement results can be compared with those of professional sports testing equipment, and the error under a certain sample size can be statistically analyzed. Assuming that after testing, its step count monitoring error is within 5%, the fitness tracker is assigned a functional effectiveness index of 8 points according to the preset functional effectiveness quantification standard (such as 8 points for an error within 5%, 6 points for 5%-10%, etc.).

[0131] Regarding ease of use: Analyze its operation steps, such as whether it can be turned on with one click, switch function modes, etc., and the ease of pairing with mobile phones and other terminals. If the operation is simple, requiring only a few steps to complete basic settings and data synchronization, and in accordance with the quantitative rules for ease of use (e.g., 7 points for 5 or fewer operation steps, 5 points for 5-10 steps, etc.), its ease of use index can be assigned 7 points. Regarding reliability: Review the product quality report, user reviews, and other materials of the fitness tracker, and calculate its probability of failure within a certain period of use (e.g., one year). If the failure probability is low, meeting the score range corresponding to high reliability in the reliability quantitative standard (e.g., 8 points for failure probability below 1%, 6 points for 1%-3%, etc.), then its reliability index can be assigned 8 points. Regarding cost-effectiveness: Compare the price of similar fitness trackers on the market with the features of this product, and use the cost-effectiveness quantitative calculation model (e.g., comprehensively considering the number of functions, the level of advancement of functions, and the ratio of price, etc.) to obtain its cost-effectiveness quantitative score, assumed to be 6 points.

[0132] Following the same process, all other devices in the initial device set are quantitatively evaluated under various indicators such as functional effectiveness, ease of use, reliability, and cost-effectiveness, resulting in a specific score for each device under each indicator. Next, a preset weighting method is used to calculate the overall priority score. Assuming that for devices like fitness trackers, the preset weighting is 0.25 for functional effectiveness, 0.25 for ease of use, 0.25 for reliability, and 0.2 for cost-effectiveness. Combining the previously calculated scores for each indicator of the fitness tracker (functional effectiveness 8 points, ease of use 7 points, reliability 8 points, cost-effectiveness 6 points), the overall priority score is calculated as: 8 × 0.25 + 7 × 0.25 + 8 × 0.25 + 6 × 0.25 = 7.25 points. This calculation process is performed on each device in the initial device set to obtain its respective overall priority score. Finally, all devices are sorted in descending order based on their calculated overall priority scores.

[0133] One patient follow-up management method also includes a step following the ranking of all devices in descending order based on a calculated comprehensive priority score, as follows: Step SD01 employs a context-based Thompson sampling algorithm. This algorithm uses patient health data features, home environment risk assessment levels, and device usage history as contextual information to dynamically test different device combinations from the initial device recommendation list, probabilistically exploring potential high-quality solutions. The context-based Thompson sampling algorithm is an algorithm that considers contextual information in the multi-armed slot machine problem. By continuously trying different device combinations and adjusting the selection probability based on feedback results, it gradually finds the optimal device recommendation solution. Device usage history records the patient's past use of various devices, including usage frequency, usage time, and usage effects. Dynamic testing continuously adjusts the testing and selection of different device combinations in real time based on new feedback information to adapt to the patient's dynamic changing needs. Probabilistic approach: The algorithm determines which combination to test based on the expected reward probability of each device combination, rather than deterministically selecting a particular combination.

[0134] The general process is described as follows: 1. Initialization: Combining each device Initialize a Beta distribution , where x represents context information. and The initial values ​​are all set to 1, indicating that the initial uncertainty is the same for each device combination. 2. Sampling and Recommendation: For each context information x, from each device combination... A random number is sampled from the Beta distribution. Select the combination of devices with the largest sample value. 3. Feedback and Updates: Based on the patient's feedback on the recommended device combination. Based on feedback (acceptance or effectiveness), update the Beta distribution parameters for the device combination. If the feedback is positive, it is considered a success event. Add 1; if the feedback is poor or the recommendation is not accepted, it will be considered a failure event. Add 1.4. Iteration: Repeat the above sampling, recommendation, and update steps. Over time, the algorithm gradually favors selecting combinations of devices that have been proven to be better, thereby exploring potential high-quality solutions.

[0135] Step SD02: Construct a digital twin model based on AnyLogic, inputting device functional parameters, patient health data characteristics, and home environment parameters, perform virtual simulation on the recommended device combination, and output the collaborative effectiveness score of the device combination.

[0136] The AnyLogic-based digital twin model is a digital mapping of a real-world device combination, patient health status, and home environment, built using AnyLogic software. It simulates their operational status and interactions under different conditions. Device functional parameters describe the various functional characteristics of the devices, such as the heart rate monitoring accuracy and measurement frequency of a smart bracelet, and the timed reminder accuracy of a smart pillbox. Collaborative effectiveness score is a quantitative evaluation score of the collaborative working effect of the device combination in the patient's home environment and health status, reflecting the degree of cooperation between devices and their comprehensive contribution to patient health management. The general process is as follows: 1. Model Construction: Using AnyLogic software, a digital twin model is constructed based on the working principles of the devices, the changing patterns of the patient's health status, and the interrelationships of home environmental factors. For example, simulating the data interaction and collaborative workflow between a smart pillbox and a health monitoring bracelet. 2. Data Input: The acquired device functional parameters, patient health data characteristics, and home environment parameters are input into the constructed digital twin model. For example, inputting the heart rate monitoring frequency of the bracelet, the patient's current heart rate data, and parameters such as indoor temperature and humidity. 3. Virtual Simulation: Different scenarios and conditions are set in the model to simulate the operation of the device combination over a period of time. For example, simulating how the device combination works collaboratively to monitor and manage a patient's health during different time periods throughout the day. 4. Collaborative Efficiency Score Output: Based on the model simulation results, the collaborative efficiency score of the device combination is calculated by evaluating multiple dimensions such as the accuracy of device data interaction, the effectiveness of patient health management, and energy consumption. For example, if the smart pillbox and wristband can accurately synchronize data, effectively remind patients to take medication on time and monitor their health, and have low energy consumption, then the collaborative efficiency score will be high.

[0137] Step SD03 incorporates the device combination recommendation effect and collaborative performance score into the comprehensive reward mechanism, and uses them as a new objective item to participate in the multi-objective optimization process of the NSGA-II algorithm.

[0138] Among them, the recommended equipment combination effect refers to the patient's acceptance of the recommended equipment combination and their feedback after use, such as whether the patient actually purchased and used the recommended equipment combination, and the improvement effect on their health status after use. The synergistic efficacy score is a quantitative score reflecting the synergistic working effect of the equipment combination, obtained by virtual simulation of the recommended equipment combination using a digital twin model based on AnyLogic.

[0139] The general process is described below: Expanded objectives: The effectiveness of equipment combination recommendations and collaborative efficiency scores will be added as evaluation objectives for the comprehensive reward mechanism.

[0140] Weighting: Assign weights to new objectives and update the overall reward calculation formula (e.g.) ).

[0141] Algorithm optimization: The updated reward mechanism is input into the NSGA-II algorithm, and the optimal policy solution set that takes into account multiple objectives is iteratively found through non-dominated sorting, selection, crossover and mutation operations.

[0142] Step SD04: Based on the optimization results of the NSGA-II algorithm, adjust the presentation of device recommendation content in the voice notification strategy, and generate the final device recommendation list to be sent to the patient's terminal.

[0143] Content Adjustment: Based on the analysis results, adjust the presentation of device recommendations in voice notifications. For example, if the optimization results show that a certain device combination has high collaborative efficiency, then the advantages of this combination will be emphasized in the voice notification; if patients have a high response rate to a specific type of device, then the recommendation duration for the corresponding device will be increased. 3. List Generation: Combine the adjusted device recommendation content with the comprehensive priority score to generate the final device recommendation list. 4. Terminal Transmission: Send the recommendation list to the patient's terminal via the communication method registered by the patient (such as mobile phone, smart terminal).

[0144] Step SE00: Organize the sorted devices into a recommendation list and send it to the patient's terminal.

[0145] One patient follow-up management method also includes a step parallel to matching and analyzing the mapped devices from a pre-established database of category-associated device mappings, as follows: Step Sa00 involves cleaning and standardizing the patient's feedback information, and then using an emotion analysis model to analyze the emotion category and emotion tendency score.

[0146] Cleaning: This is the initial processing of patient feedback information, aimed at removing irrelevant, erroneous, and improperly formatted information to make the data cleaner and more usable. Standardization: This is the process of standardizing the cleaned feedback information according to uniform formats and rules.

[0147] Sentiment analysis models are built using technologies such as machine learning and natural language processing. Their main function is to analyze text content (in this case, patient feedback) to determine the category and degree of sentiment it contains. Sentiment categories are typically divided into positive, negative, and neutral. For example, if a patient reports, "I recently adjusted my diet according to my doctor's advice and feel much better; I'm very happy," the model would classify it as positive. Conversely, a response like, "After taking this medication, I experienced severe side effects and felt very unwell," would be classified as negative. The sentiment tendency score further quantifies this degree of sentiment, usually taking a value within a specific range (e.g., 0-1, where 0 represents extremely negative, 1 represents extremely positive, and around 0.5 represents neutral). For instance, the positive sentiment example might have a sentiment tendency score close to 1, while the negative sentiment example would have a score closer to 0. This score provides a more nuanced representation of the patient's emotional state.

[0148] Step Sb00: Input the analyzed sentiment category and sentiment tendency score into the situation analysis-based device recommendation model, output the recommended devices, and add the recommended devices to the initial device set.

[0149] The situation-based device recommendation model is a specially constructed model designed to recommend suitable devices to patients based on the sentiment categories and sentiment tendency scores analyzed from patient feedback information, combined with other relevant factors (such as the patient's existing medical conditions, lifestyle characteristics, and other situational clues implied in the feedback information).

[0150] The specific operation process is as follows: First, obtain the sentiment category and sentiment tendency score output from step Sa00, ensuring the data is accurate. These two pieces of information are the key inputs for this step and will be extracted from the corresponding data storage location (generally in the temporary cache or data table corresponding to the system module performing sentiment analysis). A simple data format check is performed to ensure that it can be correctly recognized and processed by the situation-based device recommendation model. Then, the sentiment category and sentiment tendency score are used as the main inputs, along with other relevant situational information that can be mined from patient feedback, and are input into the situation-based device recommendation model. After a complex calculation and matching process, the model outputs a list of recommended devices.

[0151] Finally, add the recommended devices output by the model to the initial device set. The initial device set may already have some components through the previous steps. Incorporating the newly recommended devices here will expand the set, making it contain device options with more dimensions and more in line with the patient's current emotional state and overall situation, providing a richer data basis for subsequent operations such as quantitatively evaluating and sorting the device set, so as to ultimately compile the most suitable and targeted device recommendation list for the patient to assist the patient in better health management and life improvement.

[0152] Sort the sorted devices into a recommendation list and send it to the patient's terminal, including: Step SE10, for the sorted device recommendation list, input each pair of devices into the collaborative quantization index score evaluation model and output the collaborative quantization index score for each pair of devices.

[0153] Among them, the collaborative quantization index score evaluation model is constructed based on a large amount of device collaboration data and a professional technical index system. By collecting the performance data of different types of devices in actual collaboration scenarios, for example, in the medical device field, collecting the actual operation data of combinations such as various monitoring instruments and data processing systems, treatment devices and auxiliary devices, covering multiple dimensions such as data transmission rate, signal stability, device response time, and operation process coherence. Using statistical methods and machine learning algorithms to analyze and train these data, a quantitative relationship model between device performance indicators and collaboration effects is established. For example, regression analysis is used to determine the influence weights of different device parameters on collaboration efficiency, or neural network algorithms are used to mine complex non-linear relationships, enabling the model to accurately output reasonable collaborative quantization index scores according to the input device pair characteristics.

[0154] Model input factor consideration: When inputting each pair of devices into the model, the model will consider various factors. Taking smart home devices as an example, for the pair of devices of the smart lighting system and the smart curtain control system, the model will analyze their communication protocol compatibility to determine whether they can achieve smooth linkage control, such as simultaneously controlling the brightness, color of the lights, and the opening and closing of the curtains through a smart central control device; it will also consider the control delay of the devices, that is, the time interval from issuing an instruction to the actual response of the device. Excessive delay will seriously affect the user experience and device collaboration effect; in addition, factors such as the working frequency range and energy consumption mode of the devices are also within the evaluation scope because these factors may affect each other. For example, if the working frequencies of two devices are similar and the energy consumption is large, it may cause signal interference and unstable power supply, thereby affecting the quality of collaborative work.

[0155] Calculation Details: Internally, the model calculates various input factors for each pair of devices according to pre-defined algorithm rules. First, each factor is standardized to ensure it falls within the same dimension and numerical range, allowing for comparison and calculation on a unified scale. For example, data transmission rates are converted from different units (such as Mbps, KBps, etc.) to a single standard unit, and device response times are normalized to reflect response speed within a range of 0 to 1. Then, a weighted summation is performed based on the weights assigned to each factor in the model. These weights are determined through in-depth analysis of historical data and domain expertise, assigning higher weights to factors that significantly impact synergy. For instance, in an industrial automated production line, the degree of matching in processing precision and the synchronization of production cycles between devices might be given higher weights, as these factors directly relate to product quality and production efficiency. Through this weighted summation, a preliminary synergy quantification score is obtained.

[0156] Score Output and Significance: The final output score of the collaborative quantitative index is a quantitative value that reflects the collaborative work efficiency of the pair of devices. It is usually set within a specific range, such as between 0 and 1. The closer to 1, the better the collaborative effect between the devices, and the closer to 0, the worse the collaborative effect.

[0157] Step SE20: Using a pre-built functional complementarity scoring matrix, each pair of devices is scored and a functional complementarity score is obtained for each pair of devices.

[0158] The functional complementarity scoring matrix is ​​constructed by professionals based on their extensive industry experience and in-depth understanding of the functional characteristics of various devices. For devices in different fields, such as medical equipment, the functional complementarity between diagnostic and therapeutic equipment, and monitoring and rehabilitation equipment, is considered. In the smart home field, the functional synergy between lighting and temperature control equipment, and security and smart appliances, is analyzed. The rows and columns in the matrix represent different types of devices, and the score in each cell indicates the degree of functional complementarity between the corresponding device pairs. The score range can be set from 0 to 10, with higher scores indicating stronger functional complementarity. For example, in medical equipment, the functional complementarity score between cardiopulmonary resuscitation (CPR) equipment and electrocardiogram (ECG) monitoring equipment might be high because ECG monitoring equipment can monitor the patient's heart condition in real time, providing crucial data support and timing judgment for CPR operations. Therefore, their corresponding cell score in the matrix might be 8 points.

[0159] Device pair rating operation: When rating each pair of devices in the sorted device recommendation list, the functional complementarity score of the pair of devices can be quickly obtained by finding the corresponding cell in the functional complementarity rating matrix.

[0160] Step SE30: Select device combinations with scores higher than the preset functional complementarity score threshold from the scoring matrix as recommended device combinations to be retained, and remove the remaining devices from the recommended list.

[0161] The preset functional complementarity score threshold is determined based on professional medical equipment knowledge, extensive clinical experience, and a precise understanding of patients' actual needs. For example, in the field of rehabilitation therapy, analysis and research of numerous real-world cases have shown that when the functional complementarity score of the equipment combination reaches 7 points (assuming a maximum score of 10 points) or higher, it can significantly improve patients' rehabilitation outcomes and treatment compliance.

[0162] Screening process: Each pair of devices is checked one by one from the constructed functional complementarity scoring matrix. If the functional complementarity score of a pair of devices is higher than the preset threshold, such as the combination of intelligent rehabilitation training equipment and rehabilitation effect evaluation software, which scores 8 points, higher than the set threshold of 7 points, then this pair of devices will be identified as having high functional complementarity value and will be retained as a recommended device combination. Conversely, if a pair of devices, such as ordinary massage equipment and simple health recording APP, has a functional complementarity score of only 4 points, which is lower than the threshold, it will be removed from the recommendation list.

[0163] Step SE40: Sort the retained recommended device combinations in descending order according to the collaborative quantitative index scores, and send them to the patient's terminal.

[0164] The recommended device combinations are sorted in descending order based on their collaborative quantitative index scores. The primary purpose is to provide patients with a clear and scientifically sound reference for device selection. The collaborative quantitative index score comprehensively reflects the devices' performance in actual collaborative work; a higher score indicates better data transmission efficiency, smoother functional collaboration, and overall operational stability between devices. For example, in a smart home health management scenario, a combination of a smart scale and a health and diet app, with a high collaborative quantitative index score, indicates that weight data can be quickly and accurately transmitted to the app, and that the app can accurately recommend dietary plans based on the weight data, better meeting the user's health management needs in actual use. After sorting, the sorted device combination list is pushed to the patient's terminal via a secure communication channel between the medical system and the patient's terminal (such as a smartphone app or smart wearable device).

[0165] Based on the same inventive concept, embodiments of the present invention provide a patient follow-up management system, including a memory and a processor, wherein the memory stores information that can run on the processor to implement, as described above. Figures 1 to 2 The procedure for any method.

[0166] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for patient follow-up management, characterized in that, include: Regularly acquire patients' health-related data and perform data preprocessing; Using a pre-defined feature extraction method, physiological indicators, lifestyle features, and treatment-related features are extracted from pre-processed health-related data. The patient's physiological characteristics, lifestyle characteristics, and treatment-related characteristics are input into the trained follow-up type classification model, and the follow-up type is output. Obtain the patient's disease type and current treatment stage, and use the follow-up type, patient's disease type and current treatment stage as input objects to input into the follow-up content draft generation model, and output the follow-up content draft; The system acquires patients' personalized information and inputs it into a personalized adjustment model based on causal inference, outputting personalized adjustment information for follow-up content. A bidirectional semantic coding model based on the Transformer architecture is used to encode the features of the initial draft of the follow-up content and personalized adjustment information. Semantic role labeling technology is used to identify the semantic association features of sentence components. Semantic matching technology based on cosine similarity algorithm and word vector model is used to calculate the sentence similarity of the content to be fused. Based on this, the logical structure of the content is optimized to eliminate redundant information, and finally the follow-up content is formed. Based on the pre-set follow-up plan, the content of this follow-up visit, combined with the features extracted from the patient's health data, is used to generate a voice notification strategy using a reinforcement learning algorithm. The system automatically matches the patient's communication method according to the voice notification strategy, triggers the voice notification at a specified time, dynamically adjusts the broadcast parameters, and collects the patient's feedback information through voice recognition technology after the notification ends.

2. The patient follow-up management method according to claim 1, characterized in that, The final follow-up content includes: Acquire data on the patient's home environment and perform data preprocessing; Environmental features are extracted from the preprocessed household environmental data, including time-series features, fluctuation features, and correlation features. The extracted environmental features are input into the household environmental risk model, and the household environmental risk assessment level is output. Based on the pre-set mapping table of family environment risk assessment levels and integration strategies, determine the follow-up content integration strategy that matches the current family environment risk assessment level; A multimodal fusion module based on a multi-head attention mechanism is adopted. Based on a defined fusion strategy, the initial draft of follow-up content, personalized adjustment information and family environment risk assessment level are fused to enhance the semantic expression of content related to risk level. The fused content is semantically encoded using a Transformer bidirectional encoder, and a semantic association graph is constructed by combining semantic role labeling technology. Sentence similarity is calculated based on cosine similarity algorithm and word vector model, and the calculation weights are adjusted according to the family environment risk assessment level; A content logic topology graph is constructed based on graph neural networks. With sentences as nodes and semantic relevance as edges, the logical structure is optimized through graph convolution algorithms. Redundant sentence branches with similarity exceeding the threshold are automatically deleted to generate final follow-up content that fits the family environment risk.

3. The patient follow-up management method according to claim 2, characterized in that, The initial draft of the follow-up data, personalized adjustments, and the family environmental risk assessment level were integrated to include the following features: A risk-level adaptive multimodal attention fusion framework is constructed. High, medium and low family environment risk assessment levels are preset to correspond to independent attention module groups. The high-risk module group includes a gated multi-head attention module with dynamic routing and is configured with risk-sensitive weight adjustment parameters and semantic processing logic. The medium and low-risk module groups include an adaptive weight balancing attention module and are configured with semantically balanced weight adjustment parameters and semantic processing logic. Based on a pre-set risk keyword library and content classification model, the follow-up content is divided into core risk response content and non-core content. When the home environment risk assessment level is high, the gated multi-head attention module with dynamic routing is activated to generate a risk-sensitive attention mask. A preset suppression threshold is applied to non-core content. For core risk response content, the order of statements is adjusted according to a preset risk occurrence probability priority queue. Core prompts with high semantic intensity are forced to the front. Preset gain thresholds are assigned to words in the risk keyword library. Multi-head attention calculation is used to strengthen the semantic connection of the context and output a content sequence that highlights the core risk information. When the home environment risk assessment level is medium or low, the adaptive weight balancing attention module is enabled. Based on a preset balance loss function, the weight ratio of core and non-core content is dynamically adjusted. A semantic smoothing mechanism is used to optimize the connection between non-core content and core prompts and output a semantically coherent sequence that balances core and non-core content. Using a context-aware sliding window mechanism, when a preset number of consecutive non-core contents are detected in the output content sequence of modules at all risk levels, risk warning phrases are generated according to the probability of insertion based on the risk level, thus forming integrated content; The discriminator network performs risk semantic adaptability verification on the fused content, and judges whether it meets the standard based on the preset semantic feature matching threshold. If so, the merged content will be formatted and adapted to different follow-up channels. If not, a preset gradient penalty threshold is applied to the output results that do not meet the standard, driving the attention module group, content classification model and discriminator network in the multimodal attention fusion framework to optimize parameters until the risk-related semantic density reaches the preset improvement level.

4. The patient follow-up management method according to claim 1, characterized in that, Based on the pre-set follow-up plan, the follow-up content, combined with features extracted from patient health data, is used to generate voice notification strategies using reinforcement learning algorithms. These strategies include: Obtain patient preference-related information, including language, tone of voice, and speech rate; Input the patient's preference-related information into the preset customized voice notification model, and output a customized voice notification configuration that includes voice language type, timbre parameters, and speech rate parameters; By integrating patient health data characteristics, the final follow-up content, and customized voice notification configurations, a multi-dimensional state vector is constructed. Based on the constructed multidimensional state vector and combined with the constraints of the actual application scenario, the set of executable actions is determined, including the selection of discrete time points based on the preset follow-up time window, the selection of preset communication methods, the adjustment of broadcast parameters based on customized configuration, and the selection of content presentation format. Establish a comprehensive reward mechanism that combines patient response feedback, preference matching degree, and the reasonableness of notification time to set quantitative rewards and penalties; The reinforcement learning algorithm is adopted and initialized with an ε-greedy policy with a preset ε value. The policy parameters are iteratively updated according to a preset learning rate and a preset discount factor, based on a multi-dimensional state vector and a comprehensive reward mechanism. In each iteration, the reinforcement learning algorithm tries to select different actions and adjusts the policy according to the reward feedback. When the rate of change of the update is less than a preset proportion for a preset number of consecutive rounds, it converges and finally generates the optimal voice notification policy that includes notification time, communication method and broadcast parameters.

5. A patient follow-up management method according to any one of claims 1 to 4, characterized in that, It also includes steps following the acquisition of patient feedback information through speech recognition technology, as follows: Natural language processing technology is used to extract keywords and phrases related to health status, lifestyle, and environment from patient feedback. The extracted keywords and statements are categorized into health status, lifestyle, and environment categories. For each category, the extracted keywords and phrases are matched and analyzed from a pre-established database of category-related device mappings to obtain the mapped devices, which serve as the initial device set. For the initial set of devices, a preset quantitative evaluation method is used to analyze and obtain the score assigned to each device under each indicator. A preset weight allocation method is used to calculate the comprehensive priority score. Based on the calculated comprehensive priority score, all devices are sorted in descending order. The sorted devices are compiled into a recommended list and sent to the patient's terminal.

6. A patient follow-up management method according to claim 5, characterized in that, It also includes a step that runs parallel to matching and analyzing the mapped devices from a pre-established database of category and associated device mappings, as follows: Patient feedback information is cleaned and standardized, and a sentiment analysis model is used to analyze sentiment categories and sentiment tendency scores. The analyzed sentiment categories and sentiment tendency scores are input into the situation-based device recommendation model, which outputs recommended devices and adds them to the initial device set.

7. A patient follow-up management method according to claim 6, characterized in that, The sorted devices are compiled into a recommended list and sent to the patient's terminal, including: For the sorted list of recommended devices, each pair of devices is input into the collaborative quantitative index score evaluation model, and the collaborative quantitative index score of each pair of devices is output. A pre-built functional complementarity scoring matrix is ​​used to score each pair of devices and obtain a functional complementarity score for each pair of devices; The device combinations with scores higher than the preset functional complementarity score threshold are selected from the scoring matrix and retained as recommended device combinations; the remaining devices are removed from the recommended list. The recommended device combinations will be sorted in descending order according to the collaborative quantitative index scores and sent to the patient's terminal.

8. A patient follow-up management system, characterized in that, It includes a memory, a processor, and a program stored in the memory and executable on the processor, which, when loaded and executed by the processor, implements a patient follow-up management method as described in any one of claims 1 to 7.

Images