Medical follow-up visit management method and system based on large model cue word engineering

Abstract

The invention provides a medical follow-up visit management method and system based on large model cue word engineering, and relates to the technical field of medical management. Aiming at the problem of poorer pertinence of a follow-up visit strategy in the prior art, the method comprises the following steps: obtaining cue words in a cue word library according to a follow-up visit scene, and optimizing the cue words to obtain final cue words; collecting health state data and living habit data of the user; the health state data and the living habit data are input into a follow-up visit dialogue model, and the follow-up visit dialogue model generates a preset follow-up visit plan according to the input data; the follow-up conversation model adjusts a preset follow-up plan according to the final prompt word, and outputs a final follow-up plan; obtaining behavior data related to execution of the user on the final follow-up visit plan; and according to the health state data, the living habit data and the behavior data, updating the final cue word in real time. The follow-up visit strategy generated by the medical follow-up visit management method provided by the invention has relatively high pertinence.

Description

Technical Field

[0001] This invention relates to the field of medical management technology, and in particular to a medical follow-up management method and system based on large model prompt word engineering. Background Technology

[0002] Medical follow-up refers to the medical management practice of medical institutions continuously tracking the health status of patients after discharge or completion of a stage of treatment through regular inquiries, data collection, and health guidance. By dynamically monitoring patients' postoperative recovery, chronic disease control effectiveness, and treatment adherence, potential health risks can be identified in a timely manner, reducing readmission rates and the incidence of complications.

[0003] Patent CN20220933A discloses an intelligent follow-up method based on a large model, which includes acquiring patient follow-up data, generating patient follow-up questions based on preset follow-up questions corresponding to the patient's disease, and using instructional prompts to guide the large model to generate patient follow-up questions specific to the patient's disease. However, this patent only collects patient health status data and does not consider lifestyle data or patient behavioral feedback data regarding the follow-up process. Furthermore, the prompts cannot be updated in real time based on patient health status data, lifestyle data, and behavioral data.

[0004] Existing technologies do not take into account lifestyle data and patient feedback data on the follow-up process. Furthermore, the prompts cannot be updated in real time based on patient health status data, lifestyle data, and behavioral data. These issues make it difficult to adjust follow-up strategies in a targeted manner.

[0005] Therefore, developing a medical follow-up management method and system based on large model prompt word engineering is of great significance for improving the targeting of follow-up strategies. Summary of the Invention

[0006] To address the problem of poor targeting of follow-up strategies in existing technologies, this invention proposes a medical follow-up management method based on large-scale model prompt word engineering, which specifically includes the following steps: S1. Obtain prompt words from the prompt word library according to the follow-up scenario, and optimize the prompt words through an open source model. Specifically, use a small parameter open source NLP model to diagnose problems using the prompt words. The small parameter open source NLP model is obtained by fine-tuning and training based on a medical domain labeled dataset. Based on the problem diagnosis results, the content of the prompt words is automatically adjusted using the small-parameter open-source NLP model, including: testing the prompt words before and after adjustment respectively, and determining the final prompt words through a sampling algorithm; S2. Collect users' health status data and lifestyle data; S3. Input the health status data and the lifestyle data into the follow-up dialogue model, and the follow-up dialogue model generates a preset follow-up plan based on the input data; Specifically, the follow-up dialogue model embeds a dynamic decision-making algorithm, which integrates input data to predict the user's follow-up resource needs, including follow-up frequency needs, follow-up method needs, and follow-up duration needs. The follow-up dialogue model, combined with the follow-up resource requirements, generates a preset follow-up plan that includes a resource allocation scheme. S4. The follow-up dialogue model adjusts the preset follow-up plan according to the final prompt word and outputs the final follow-up plan; S5. Obtain user behavioral data related to the execution of the final follow-up plan; S6. Update the final prompt word in real time based on the health status data, the lifestyle data, and the behavior data.

[0007] Furthermore, the dynamic decision-making algorithm integrates input data to predict users' follow-up resource needs, including: standardizing the input data; assigning corresponding weights to the input data based on the degree of influence of the input data on the follow-up needs; mining the correlation features between the input data through a machine learning model to establish a prediction model of multi-dimensional features and follow-up resource needs; and outputting the follow-up resource needs through the prediction model.

[0008] Furthermore, after obtaining prompt words from the prompt word library based on the follow-up scenario, the process also includes optimizing the prompt words, specifically: diagnosing problems with the prompt words; adjusting the content of the prompt words based on the results of the problem diagnosis; testing the prompt words before and after adjustment respectively, and using the Thompson sampling algorithm to determine the final prompt words.

[0009] Furthermore, the follow-up scenario includes disease type, and obtaining prompt words from the prompt word library according to the follow-up scenario includes: obtaining prompt words associated with disease type; in S4, the follow-up dialogue model adjusts the preset follow-up plan according to the final prompt words, including: supplementing the preset follow-up plan with follow-up content for the disease type.

[0010] Furthermore, the follow-up scenario also includes patient type. Obtaining prompt words from the prompt word library according to the follow-up scenario includes: obtaining prompt words associated with patient type, wherein the prompt words include speech type; in S4, the follow-up dialogue model adjusts the preset follow-up plan according to the prompt words, including: adjusting the speech type in the preset follow-up plan.

[0011] Furthermore, in step S5, acquiring user behavioral data related to the execution of the final follow-up plan includes: acquiring the user's feedback status on the follow-up plan; and acquiring the update status of the health status data.

[0012] Furthermore, in step S6, the final prompt word is updated in real time based on the health status data, including: when the health status data shows that the user's health indicators fluctuate beyond a preset range, adding a health risk warning-related phrase to the final prompt word; and synchronously saving the updated final prompt word to the prompt word library.

[0013] Furthermore, in step S6, the final prompt word is updated in real time based on the behavioral data, including: when the behavioral data shows that the user's feedback status on the follow-up plan does not meet the preset requirements, adding content related to points incentive guidance to the final prompt word; when the behavioral data shows that the user's feedback status on the follow-up plan meets the preset requirements, adding content related to affirming the effectiveness of health management to the final prompt word; and synchronously saving the updated final prompt word to the prompt word library.

[0014] This invention also provides a medical follow-up management system based on large model prompt word engineering, the system being used to execute the medical follow-up management method based on large model prompt word engineering described above, the system comprising: The prompt word acquisition module is used to acquire prompt words from the prompt word library according to the follow-up scenario, and optimize the prompt words to obtain the final prompt words; The data collection module is used to collect users' health status data and lifestyle data; A follow-up plan generation module, connected to the data collection module, is used to input the health status data and the lifestyle data into the follow-up dialogue model, and the follow-up dialogue model generates a preset follow-up plan based on the input data; The follow-up plan adjustment module is connected to the prompt word acquisition module and the follow-up plan generation module, and is used to control the follow-up dialogue model to adjust the preset follow-up plan according to the final prompt word and output the final follow-up plan. The behavioral data acquisition module is used to acquire user behavioral data related to the execution of the final follow-up plan; The real-time update module is connected to the behavior data acquisition module, the collection module, and the prompt word acquisition module, and is used to update the final prompt word in real time based on the health status data, the lifestyle data, and the behavior data.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: Firstly, this invention obtains prompt words from a prompt word library based on the follow-up scenario and optimizes them to obtain the final prompt words; it collects the user's health status data and lifestyle habit data; it inputs the health status data and lifestyle habit data into a follow-up dialogue model, which generates a preset follow-up plan based on the input data; specifically, the follow-up dialogue model embeds a dynamic decision-making algorithm, which integrates the input data to predict the user's follow-up resource needs, including follow-up frequency, follow-up method, and follow-up duration requirements; the follow-up dialogue model combines these resource needs to generate a preset follow-up plan containing a resource allocation scheme; the follow-up dialogue model adjusts the preset follow-up plan based on the final prompt words and outputs the final follow-up plan; it acquires user behavior data related to the execution of the final follow-up plan; and it updates the final prompt words in real time based on the health status data, lifestyle habit data, and behavior data. By accurately matching follow-up resource needs through the dynamic decision-making algorithm, the follow-up frequency, method, and duration are fully adapted to the user's individual capabilities and the scenario, avoiding both excessive follow-up that wastes medical resources and insufficient follow-up that misses health risks. In addition, by simultaneously acquiring health status data, lifestyle habit data, and behavioral data, and updating the final prompt words in real time based on these data, the system takes into account changes in the user's physiological state and the scenario, as well as execution feedback. This makes the follow-up strategy more compatible with the patient's willingness to cooperate, thus improving the effectiveness of the follow-up strategy.

[0016] Secondly, by diagnosing problems with the prompts, the content of the prompts is adjusted based on the diagnostic results. The prompts before and after adjustment are tested separately, and a sampling algorithm is used to determine the final prompts. This approach can accurately identify and correct issues such as missing fields, formatting errors, and misunderstandings in the prompts at the source. It also allows for targeted content optimization based on the follow-up scenario, ensuring that the prompts accurately convey personalized needs. Furthermore, data-driven algorithms dynamically select the optimal version, avoiding the limitations of static, fixed prompts. Standardized processes reduce manual intervention costs, improve optimization efficiency and reusability, ultimately providing high-quality, highly adaptable core guidance for the follow-up dialogue model. This ensures that the subsequently generated follow-up plans not only comply with clinical guidelines but also align with the user's individual health status, cognitive ability, and cooperation level, effectively supporting the overall solution's core objective of improving the targeting of follow-up strategies.

[0017] Third, the prompts are updated in real time based on the behavioral data, including: when the behavioral data shows that the user's feedback on the follow-up plan does not meet the preset requirements, adding incentive-based guidance content to the final prompt; when the behavioral data shows that the user's feedback on the follow-up plan meets the preset requirements, adding affirmation of health management effectiveness to the final prompt. Adding incentive-based guidance content to the final prompt when the feedback does not meet the preset requirements helps improve the patient's adherence. This personalized incentive design for patients with poor adherence can accurately match the cooperation willingness of such patients, further strengthening the targeting of the follow-up strategy. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart of a medical follow-up management method based on large model prompt word engineering provided by an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a medical follow-up management system based on large model prompt word engineering provided by an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0021] The specific embodiments of the present invention will be described below.

[0022] To address the issue of poor targeting in existing follow-up strategies, this invention retrieves prompt words from a prompt word library based on the follow-up scenario; collects user health status data and lifestyle habit data; inputs this data into a follow-up dialogue model, which generates a preset follow-up plan based on the input data; adjusts the preset follow-up plan according to the prompt words, and outputs a final follow-up plan; acquires user behavior data related to the execution of the final follow-up plan; and updates the prompt words in real time based on the health status data, lifestyle habit data, and behavioral data. The medical follow-up management method provided by this invention generates a highly targeted follow-up strategy.

[0023] Example 1 This invention provides a medical follow-up management method based on large model prompt word engineering. Figure 1 This is a flowchart of a medical follow-up management method based on large model prompt word engineering provided by an embodiment of the present invention, such as... Figure 1 As shown, the specific steps include the following: S1. Based on the follow-up scenario, obtain prompt words from the prompt word library and optimize the prompt words using an open-source model. Specifically, use a small-parameter open-source NLP model to diagnose problems using the prompt words. The small-parameter open-source NLP model is fine-tuned and trained based on a labeled dataset in the medical field. Based on the results of the problem diagnosis, automatically adjust the content of the prompt words using the small-parameter open-source NLP model, including testing the prompt words before and after adjustment respectively, and determining the final prompt words through a sampling algorithm.

[0024] Follow-up scenarios refer to the specific contextual characteristics involved in the medical follow-up process, including patient disease, follow-up stage, individual patient characteristics, and core health needs. These are key criteria for defining follow-up goals and content. Patient diseases include hypertension and diabetes, while individual patient characteristics include age, education level, and adherence. A cue word library is a pre-built database storing a large number of instructional texts for different follow-up scenarios. Cue words guide the large model to generate instructional texts that match the expected content, matching the follow-up scenario and directly driving the generation or adjustment of the follow-up plan.

[0025] For example, retrieving prompts from the prompt word library based on the disease type in the follow-up scenario includes retrieving prompts associated with the disease type. Retrieving prompts from the prompt word library based on the patient type (i.e., individual patient characteristics) in the follow-up scenario includes retrieving prompts associated with the patient type. By matching prompts with the scenario, it ensures that the extracted prompts closely address the core needs of the current scenario, reducing invalid or off-topic follow-up content. Moreover, in multi-scenario, high-volume follow-up tasks, it eliminates the need for medical staff to manually write or adjust prompts, significantly improving efficiency.

[0026] This embodiment uses a small-parameter open-source NLP model to optimize the prompt words. A small-parameter open-source NLP model refers to an open-source natural language processing model with a parameter count in the range of 7B, including but not limited to lightweight versions of open-source large language models such as the Qwen series (Alibaba) and the ERNIE series (Baidu Wenxin).

[0027] The process of building a small-parameter open-source NLP model includes: (1) Establishment of medical data sets This study collected and organized sample data of prompt words in the field of medical follow-up to construct a dedicated dataset for prompt word optimization. The dataset includes prompt word problem-annotated data, prompt word optimization pairing data, and medical imaging report prompt word data. Specifically, the prompt word problem-annotated data refers to a training set for problem diagnosis formed by manually annotating prompt words for issues such as missing fields, formatting errors, semantic deviations, and range errors; the prompt word optimization pairing data refers to a training set for content adjustment formed by collecting prompt word pairing samples before and after optimization, annotating optimization strategies and effects; and the medical imaging report prompt word data refers to a dedicated sample library of extracted prompt words for various medical imaging data, including blood routine test reports, biochemical test reports, electrocardiogram reports, X-ray diagnostic reports, CT reports, and MRI reports.

[0028] (2) Model fine-tuning training Based on the aforementioned medical dataset, a small-parameter open-source NLP model was trained using supervised fine-tuning (SFT). This fine-tuning enabled the model to acquire the following capabilities: Prompt word problem diagnosis capability: Identify problems in prompt words such as missing fields, non-standard format, inaccurate semantic expression, and unreasonable numerical range constraints; Prompt word content adjustment capability: Automatically generate optimization suggestions based on diagnostic results, including supplementing missing fields, correcting format constraints, optimizing semantic descriptions, and adjusting numerical ranges; Medical image data extraction prompt word optimization capability: For various medical image data extraction scenarios, the prompt words are optimized to more accurately guide the large model to identify and extract key medical data.

[0029] (3) Collaborative working mechanism between open source models and large models The small-parameter open-source NLP model and the follow-up dialogue model form a collaborative relationship with each other: The small-parameter open-source NLP model is responsible for the problem diagnosis and content optimization of prompt words, and has the advantages of fast response speed, low deployment cost and local operation. The follow-up dialogue model is responsible for performing core tasks such as extracting medical image data and generating follow-up plans based on optimized prompts, leveraging its powerful understanding and generation capabilities.

[0030] Through the aforementioned collaborative mechanism, both the real-time and cost-effectiveness of prompt word optimization are ensured, as well as the execution quality of core follow-up tasks.

[0031] By collecting execution results and quality metrics of the prompt words, common problems such as missing fields, format errors, and misunderstandings are identified, and optimization suggestion reports are generated. The effectiveness of the prompt words is evaluated from multiple dimensions, including calculating accuracy metrics, analyzing common error patterns, identifying missing fields, and assessing data quality. The prompt word content is automatically adjusted based on the problem diagnosis results, enhancing the prompt description for missing fields, strengthening constraints for format errors, adding validation explanations for range errors, and adding positive case examples to improve accuracy.

[0032] Two sets of prompt words, before and after optimization of a small-parameter open-source NLP model, were tested in parallel in a real-world follow-up scenario. A sampling algorithm was used to dynamically allocate traffic, gradually increasing the probability of selecting high-performing prompt words and gradually eliminating low-performing ones, thus determining the final prompt words. The sampling algorithm did not allocate test resources fixedly, but dynamically adjusted the calling probability of the two sets of prompt words based on their real-time performance. Through continuous dynamic testing and probability adjustment, inefficient prompt word versions were ultimately eliminated, and the best-performing prompt words became the final prompt words, completing the automated selection process. The final selected prompt words accurately convey personalized needs, improving the adaptability of the prompt words and providing high-quality guidance for the follow-up dialogue model.

[0033] Based on the above embodiments, a systematic prompt word abstraction framework is established, comprising three layers: **Business Scenario Abstraction Layer:** This layer structurally defines medical follow-up business. Each business scenario includes attributes such as scenario identifier, scenario type, key points of focus, and data requirements. For example, the hypertension follow-up scenario focuses on blood pressure control, medication adherence, and the effectiveness of lifestyle interventions. **Data Extraction Abstraction Layer:** This layer clarifies the data fields and rules to be extracted for each scenario. Each data field definition includes complete information such as field name, data type, extraction rules, validation rules, normal range, and anomaly judgment criteria. This layer transforms medical business requirements into structured data definitions. **Prompt Word Generation Layer:** This layer automatically generates prompt words that meet the requirements of the AI ​​model based on the business scenario definition and data extraction rules. The prompt words include complete content such as task description, data extraction requirements, output format specifications, and quality control requirements. The system supports manual review and adjustment of the generated prompt words.

[0034] For example, based on the follow-up scenario, corresponding prompts are generated through a three-layer prompt word abstraction model. First, the business scenario abstraction layer identifies the current follow-up scenario type. Scenario types include disease type (e.g., hypertension, diabetes, coronary heart disease), follow-up stage (e.g., postoperative follow-up, chronic disease management), and patient characteristics (e.g., elderly patients, pregnant women). Second, the data extraction abstraction layer retrieves the corresponding data extraction rules from the configuration library based on the scenario type. The data extraction rules define which health status data needs to be collected, the format requirements for each data field, and the reasonable range of values. Finally, the prompt word generation layer automatically assembles and generates structured prompts based on the business scenario definition and data extraction rules. The prompts include the following parts: Task background description: describing the medical background and purpose of the current follow-up scenario; Data collection instructions: clarifying the data fields that the AI ​​model needs to extract and their rules; Follow-up plan generation instructions: explaining the strategy for generating a follow-up plan based on the data; Communication style requirements: defining the language style and expression for communicating with patients; Output format specifications: specifying the structured format of the AI ​​model's output results; Quality control requirements: setting standards for data accuracy and the rationality of the follow-up plan.

[0035] The generated prompts are saved to a prompt library, which supports manual review and adjustment. Administrators can view the generated prompt content through the prompt management module, optimize and adjust it according to actual business needs, and update the version number and save the modified prompts.

[0036] S2. Collect users' health status data and lifestyle data.

[0037] Health status data reflects information such as a user's physiological health status and disease progression, including but not limited to physiological indicators such as blood pressure and blood sugar, disease-related data such as symptom descriptions, disease stages, and complications, and treatment-related information such as medication records and examination reports.

[0038] Specifically, collecting users' health status data includes: obtaining the health status data by collecting data entered by users, data from smart terminals, and data from medical reports.

[0039] User-entered data refers to health-related information actively submitted by users. Smart terminal data refers to health data automatically monitored and output by smart devices, including blood pressure monitors, blood glucose meters, electrocardiogram monitors, and smart bracelets. Data in medical reports (i.e., third-party data) refers to medical document data containing user medical information, including hospital lab reports, follow-up medical records, and imaging examination reports.

[0040] For example, user-entered data and data from medical reports can be entered through WeChat mini-programs. Smart terminals are linked to WeChat mini-programs; for instance, the mini-program can directly use the phone's Bluetooth to establish a near-field connection with the smart terminal and periodically upload monitoring data from the smart terminal. Data is submitted through the WeChat mini-program's user entry point, smart terminal synchronization channel, and medical report upload channel, forming a complete user health status dataset. The collected health status data is stored in a health status database for subsequent retrieval and analysis.

[0041] Lifestyle data refers to subjective and controllable factors in a user's daily behavior that may affect their health, such as implicit data, sleep patterns, and exercise data.

[0042] For example, meteorological data is collected by using user-authorized location information, and real-time data and forecast data for the next 3-7 days are synchronized regularly. Air quality data is collected by synchronizing real-time air quality data of the user's area, with a focus on increasing the collection frequency for users with respiratory and cardiovascular diseases.

[0043] In this embodiment, a dedicated prompt template for medical report data collection is set up for the data collection process, achieving zero-code configuration. It transforms the medical report data collection requirements into natural language descriptions, which are understood and executed by the AI ​​model without requiring any coding. The implementation process includes the following steps: Step 1: Image Type Recognition and Matching with Medical Report Data Collection Specific Prompt Templates. Pre-set specific prompt templates for common medical image types are used, including those for blood routine test reports, biochemical test reports, electrocardiogram reports, X-ray diagnostic reports, prescriptions, and outpatient medical records. When a user uploads a new medical report image, the system automatically recognizes the image type and matches the corresponding specific prompt template from the prompt word library, providing the basic instructions for accurately extracting health status data.

[0044] Step Two: A visual configuration interface for dedicated prompts in medical report data collection. By providing a web-based configuration interface, users, without any coding experience, can customize their health status data extraction needs from medical reports through simple form operations. This includes defining image types, configuring the health status fields to be extracted, and adding special instructions. These configuration operations facilitate accurate acquisition of health status data.

[0045] Step 3: Automatically generate customized prompts for medical report data collection. Based on the user's configuration on the web interface, the system calls the basic instruction framework from the prompt library to automatically generate targeted prompts for medical report data collection. These prompts include: task description, field extraction requirements, special instructions, output format requirements, and quality requirements.

[0046] Step 4: Instant Testing and Adjustment. After completing the configuration, you can upload sample images of similar medical reports for instant testing. Input the customized medical report data collection prompts into the AI ​​model, which will recognize the image and extract health status data. The interface displays a table comparing the AI-extracted health status data with the user's expected health status data. If discrepancies exist, the user can return to the web configuration interface to adjust the configuration of the medical report data collection prompts and retest until the extracted health status data meets the accuracy and completeness requirements.

[0047] Step 5: Configuration Saving and Reuse. Once the sample test is successful, users can save the currently configured medical report data collection prompts and association rules as a medical report health data collection template, and simultaneously store it in the initial prompt library. When encountering similar medical reports later, this template can be directly called without repeated configuration, achieving one-time configuration and multiple reuses, significantly improving the efficiency of collecting health status data for medical reports.

[0048] S3. Input the health status data and lifestyle data into the follow-up dialogue model, and the follow-up dialogue model generates a preset follow-up plan based on the status data.

[0049] The follow-up dialogue model is based on an AI-powered algorithm that pre-trains the logic linking health data with follow-up needs. For example, if blood glucose levels are >7.0 mmol / L, dietary inquiries should be intensified; if there is no feedback on wound condition 3 days post-surgery, proactive reminders should be given. The follow-up dialogue model can output follow-up plans that conform to clinical standards and user needs based on the input health data. The preset follow-up plan refers to the follow-up arrangements directly generated by the follow-up dialogue model based on the user's current health status, including follow-up time, content, and methods.

[0050] For example, the follow-up dialogue model embeds a dynamic decision-making algorithm. This algorithm integrates input data to predict the user's follow-up resource needs, including follow-up frequency requirements, follow-up method requirements, and follow-up duration requirements. The follow-up dialogue model then combines these resource needs to generate a preset follow-up plan that includes a resource allocation scheme.

[0051] The dynamic decision-making algorithm integrates input data to predict users' follow-up resource needs, including: standardizing the input data; assigning corresponding weights to the input data based on their impact on follow-up needs; mining the correlation features between the input data through machine learning models to establish a prediction model of multi-dimensional features and follow-up resource needs; and outputting the follow-up resource needs through the prediction model.

[0052] The input data contains various types and magnitudes, such as blood pressure values ​​of 120-180 mmHg and lifestyle habits of smoking 0-5 cigarettes per day. Direct use in calculations would lead to analytical biases due to differences in format and magnitude. Input data is normalized to ensure that data of different types and magnitudes are on the same dimension. Different data have significantly different impacts on follow-up resource requirements. For example, blood pressure fluctuation data is more indicative of follow-up frequency than lifestyle data for hypertensive patients. Based on medical guidelines and historical data statistics, hypertensive patients are assigned the following weights: 60% for health status data and 40% for lifestyle data. Key features are extracted from the standardized data, and cross-features are generated. Machine learning models such as random forests and gradient boosting trees are used to learn the patterns of feature combinations and resource requirements in historical follow-up data, forming a multi-dimensional feature-based predictive model of follow-up resource requirements. This predictive model maps the feature combinations corresponding to the input data to specific follow-up resource requirement parameters. By exploring the synergistic effects of multiple factors, resource requirement predictions are made more closely aligned with real-world scenarios.

[0053] The collected health status and lifestyle data are input into the follow-up dialogue model. This follow-up dialogue model is an AI service built on a large language model, capable of understanding medical data and generating follow-up suggestions.

[0054] The follow-up dialogue model analyzes input data to identify the patient's health status characteristics, including: whether physiological indicators are within the normal range, whether there are abnormal fluctuation trends, medication adherence, and whether lifestyle habits are healthy. Based on the data analysis results, the follow-up dialogue model generates a preset follow-up plan. The preset follow-up plan includes: follow-up frequency: determining the follow-up interval based on health status, such as once a week or once a month; follow-up content: determining the questions to be asked, the data to be collected, and the health guidance to be provided; and precautions: highlighting key points and risk areas for medical staff. The preset follow-up plan is a preliminary plan generated based on general medical knowledge and the patient's current data.

[0055] S4. The follow-up dialogue model adjusts the preset follow-up plan based on the final prompt word and outputs the final follow-up plan.

[0056] The follow-up dialogue model analyzes the final prompts input to determine optimization directions, and then adjusts the preset follow-up plan item by item according to these directions. After the follow-up dialogue model completes the adjustments, the final follow-up plan is validated to ensure that every modification meets the requirements of the final prompts. The final follow-up plan is then output. This final follow-up plan is a directly executable solution formed by the follow-up dialogue model in conjunction with the adjusted prompts. It retains the core content of the preset follow-up plan that aligns with health data, while also adding personalized optimizations to suit different scenarios. The final follow-up plan serves as the ultimate action guide for subsequent actual follow-ups.

[0057] For example, the follow-up dialogue model adjusts the preset follow-up plan based on the final prompt word, including: supplementing the preset follow-up plan with follow-up content for the disease type. The final prompt word includes the disease type as type 2 diabetes. The follow-up dialogue model extracts type 2 diabetes from the prompt word and uses it as a keyword for the supplementary content. The follow-up dialogue model calls a built-in disease-specific content association library, using the extracted disease type as an index, to match the follow-up content that requires special attention for that disease from the association library. For example, for type 2 diabetes, it matches follow-up content such as whether pre-meal hypoglycemia has occurred recently, and daily carbohydrate intake. The follow-up dialogue model first analyzes the original structure of the preset follow-up plan, then embeds the matched disease type-specific content into the preset follow-up plan, thereby adjusting the preset follow-up plan and outputting the final follow-up plan.

[0058] Based on the above embodiments, the follow-up scenario includes patient types. According to the patient type in the follow-up scenario, prompt words of the corresponding dialogue type are obtained from the prompt word library. The follow-up dialogue model adjusts the preset follow-up plan based on the prompt words, including: adjusting the dialogue type within the preset follow-up plan.

[0059] For example, if the patient type is elderly patients with basic knowledge of hypertension, the search criteria are "elderly patients with basic knowledge of hypertension," and specific prompts are matched from the prompt word library. For instance, the prompts might include phrases like: using "high pressure" instead of "systolic pressure," "low pressure" instead of "diastolic pressure," and avoiding technical terms like "blood pressure fluctuation range" and "medication adherence." They might also break down multi-information questions into single-information short sentences, avoiding sentences containing two or more questions. The follow-up dialogue model then compares each sentence with the prompts, adjusting the pre-set follow-up plan's wording to be more colloquial and shorter. The final follow-up plan is output by the model after these adjustments.

[0060] S5. Obtain user behavioral data related to the execution of the final follow-up plan. Specifically, this includes: obtaining the user's feedback status on the follow-up plan; and obtaining the update status of the health status data.

[0061] Behavioral data reflects the status and results of users' actions in executing the final follow-up plan. This includes timeliness (whether the follow-up was completed within the planned timeframe), completeness (whether the questionnaire was fully completed and whether the user participated in the entire conversation), initiative (whether the user proactively updated their health data and completed the follow-up task ahead of schedule), and effectiveness (whether the data was submitted according to requirements and whether feedback was perfunctory). For example, user behavior can be automatically recorded through WeChat mini-programs, and the acquired behavioral data can be stored in a user behavior database for easy system access.

[0062] S6. Update the final prompt word in real time based on the health status data, lifestyle habit data, and behavioral data. A comprehensive judgment is made by combining health status data and behavioral data to ensure that the final prompt word update not only meets health needs but also adapts to the execution status.

[0063] The prompt word optimization method in the above embodiments continues to optimize the updated prompt words, establishes a prompt word version library, records the effect of each optimization, and supports parallel testing of multiple candidate versions to automatically converge to the optimal version.

[0064] For example, updating the final prompt word in real time based on the health status data includes: when the health status data shows that the user's health indicators fluctuate beyond a preset range, adding a health risk warning-related phrase to the prompt word; and synchronously saving the updated final prompt word to the prompt word library.

[0065] The system continuously collects and analyzes users' health status data, calculating the fluctuation range of health indicators within a preset period. For example, based on the health status data of a hypertensive patient, it calculates the fluctuation range of blood pressure over 24 hours, compares the calculated fluctuation range with a preset normal fluctuation range, and determines whether the fluctuation is abnormal. When the fluctuation range of a health indicator exceeds the preset range, a prompt word update process is triggered, generating health risk warning related statements such as: "Your systolic blood pressure has fluctuated beyond the normal range in the last 24 hours, which may cause dizziness and palpitations. Please pay attention to rest and medication." The system adds health risk warning related statements to the original final prompt word and synchronously saves the updated final prompt word to the prompt word library. By adding health risk warning related statements to the final prompt word and updating it in real time, the system ensures timely detection of health risks and avoids delays in risk assessment.

[0066] For example, updating the final prompt word in real time based on the behavioral data includes: when the behavioral data shows that the user's feedback status on the follow-up plan does not meet the preset requirements, adding content related to points incentive guidance to the final prompt word; when the behavioral data shows that the user's feedback status on the follow-up plan meets the preset requirements, adding content related to affirming the effectiveness of health management to the final prompt word; and synchronously saving the updated final prompt word to the prompt word library.

[0067] Continuously collect user behavior data related to the follow-up plan and compare it with preset requirements. For example, if the behavior data shows failure to upload blood glucose, a 50% questionnaire completion rate, or two missed visits in the past two weeks, it is considered as not meeting the preset requirements and is judged as low cooperation. If the behavior data shows timely uploading of blood glucose, a 100% questionnaire completion rate, and zero missed visits in the past two weeks, it is considered as fully meeting the preset requirements and is judged as high cooperation. Add corresponding guidance content to the existing final prompts. Synchronously save the updated final prompts to the prompt library. By adding a points incentive to guide the real-time updating of the final prompts, the enthusiasm of users with low cooperation can be improved.

[0068] In this embodiment, the follow-up scenario involves retrieving prompts from a prompt word library and optimizing them to obtain the final prompts; collecting the user's health status data and lifestyle habit data; inputting the health status data and lifestyle habit data into a follow-up dialogue model, which then generates a preset follow-up plan based on the input data; specifically, the follow-up dialogue model embeds a dynamic decision-making algorithm, which integrates the input data to predict the user's follow-up resource needs, including follow-up frequency, follow-up method, and follow-up duration requirements; the follow-up dialogue model, combined with these resource needs, generates a preset follow-up plan containing a resource allocation scheme; the follow-up dialogue model adjusts the preset follow-up plan based on the final prompts and outputs the final follow-up plan; acquiring user behavior data related to the execution of the final follow-up plan; and updating the final prompts in real time based on the health status data, lifestyle habit data, and behavior data. By accurately matching follow-up resource needs through the dynamic decision-making algorithm, the follow-up frequency, method, and duration are fully adapted to the user's individual capabilities and the scenario, avoiding both excessive follow-up that wastes medical resources and insufficient follow-up that misses health risks. In addition, by simultaneously acquiring health status data, lifestyle habit data, and behavioral data, and updating the final prompt words in real time based on these data, the system takes into account changes in the user's physiological state and the scenario, as well as execution feedback. This makes the follow-up strategy more compatible with the patient's willingness to cooperate, thus improving the effectiveness of the follow-up strategy.

[0069] Secondly, by diagnosing issues with the prompts, the content of the prompts is adjusted based on the diagnostic results. The prompts before and after adjustment are tested separately, and a sampling algorithm is used to determine the final prompts. This approach can accurately identify and correct issues such as missing fields, formatting errors, and misunderstandings in the prompts at the source. It also allows for targeted content optimization based on the follow-up scenario, ensuring that the prompts accurately convey personalized needs. Furthermore, data-driven algorithms dynamically select the optimal version, avoiding the limitations of static, fixed prompts. Standardized processes reduce manual intervention costs, improve optimization efficiency and reusability, ultimately providing high-quality, highly adaptable core guidance for the follow-up dialogue model. This ensures that the subsequently generated follow-up plans not only comply with clinical guidelines but also align with the user's individual health status, cognitive ability, and cooperation level, effectively supporting the overall solution's core objective of improving the targeting of follow-up strategies.

[0070] Finally, the prompts are updated in real time based on the behavioral data, including: when the behavioral data shows that the user's feedback on the follow-up plan does not meet the preset requirements, adding incentive-based guidance content to the final prompts; and when the behavioral data shows that the user's feedback on the follow-up plan meets the preset requirements, adding affirmation of health management effectiveness to the final prompts. Adding incentive-based guidance content to the final prompts when the feedback does not meet the preset requirements helps improve patient compliance. This personalized incentive design for patients with poor compliance can accurately match their willingness to cooperate, further strengthening the targeted nature of the follow-up strategy.

[0071] Example 2 This invention also provides a medical follow-up management system based on large model prompt word engineering. Figure 2 This is a schematic diagram of the structure of a medical follow-up management system based on large model prompt word engineering provided by an embodiment of the present invention, such as... Figure 2 As shown, the system includes: The prompt word acquisition module 110 is used to acquire prompt words from the prompt word library according to the follow-up scenario, and optimize the prompt words to obtain the final prompt words; The data collection module 120 is used to collect users' health status data and lifestyle data; The follow-up plan generation module 130 is connected to the data acquisition module 120 and is used to input the health status data and the lifestyle data into the follow-up dialogue model. The follow-up dialogue model generates a preset follow-up plan based on the input data. The follow-up plan adjustment module 140 is connected to the prompt word acquisition module 110 and the follow-up plan generation module 130, and is used to control the follow-up dialogue model to adjust the preset follow-up plan according to the final prompt word and output the final follow-up plan. The behavioral data acquisition module 150 is used to acquire user behavioral data related to the execution of the final follow-up plan. The real-time update module 160 is connected to the behavior data acquisition module 150, the collection module 120, and the prompt word acquisition module 110, and is used to update the final prompt word in real time based on health status data, lifestyle data, and the behavior data.

[0072] The medical follow-up management system based on large model prompt word engineering provided in this embodiment executes the medical follow-up management method based on large model prompt word engineering described in any of the above embodiments, and has the beneficial effects of any of the above embodiments, which will not be repeated here.

[0073] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A medical follow-up management method based on large model prompt word engineering, characterized in that, include: S1. Obtain prompt words from the prompt word library according to the follow-up scenario, and optimize the prompt words through an open source model. Specifically, use a small parameter open source NLP model to diagnose problems using the prompt words. The small parameter open source NLP model is obtained by fine-tuning and training based on a medical domain labeled dataset. Based on the problem diagnosis results, the content of the prompt words is automatically adjusted using the small-parameter open-source NLP model, including: testing the prompt words before and after adjustment respectively, and determining the final prompt words through a sampling algorithm; S2. Collect users' health status data and lifestyle data; S3. Input the health status data and the lifestyle data into the follow-up dialogue model, and the follow-up dialogue model generates a preset follow-up plan based on the input data; Specifically, the follow-up dialogue model embeds a dynamic decision-making algorithm, which integrates input data to predict the user's follow-up resource needs, including follow-up frequency needs, follow-up method needs, and follow-up duration needs. The follow-up dialogue model, combined with the follow-up resource requirements, generates a preset follow-up plan that includes a resource allocation scheme. S4. The follow-up dialogue model adjusts the preset follow-up plan according to the final prompt word and outputs the final follow-up plan; S5. Obtain user behavioral data related to the execution of the final follow-up plan; S6. Update the final prompt word in real time based on the health status data, the lifestyle data, and the behavior data.

2. The medical follow-up management method based on large model prompt word engineering according to claim 1, characterized in that, The dynamic decision-making algorithm integrates input data to predict users' follow-up resource needs, including: Standardize the input data; The input data is assigned corresponding weights based on its impact on follow-up needs. By using machine learning models to uncover the correlation features between input data, a predictive model of multi-dimensional features and follow-up resource requirements is established. The predictive model outputs follow-up resource requirements.

3. The medical follow-up management method based on large model prompt word engineering according to claim 1, characterized in that, The follow-up scenario includes disease type. Obtaining prompt words from the prompt word library according to the follow-up scenario includes: obtaining prompt words associated with disease type; in S4, the follow-up dialogue model adjusts the preset follow-up plan according to the final prompt words, including: supplementing the preset follow-up plan with follow-up content for the disease type.

4. The medical follow-up management method based on large model prompt word engineering according to claim 3, characterized in that, The follow-up scenario also includes patient type. Obtaining prompt words from the prompt word library according to the follow-up scenario includes: obtaining prompt words associated with patient type, the prompt words including dialogue type; in S4, the follow-up dialogue model adjusts the preset follow-up plan according to the final prompt words, including: adjusting the dialogue type in the preset follow-up plan.

5. The medical follow-up management method based on large model prompt word engineering according to claim 1, characterized in that, In step S5, user behavioral data related to the execution of the final follow-up plan is acquired, including: Obtain user feedback on the follow-up plan; Obtain the updated status of the health status data.

6. The medical follow-up management method based on large model prompt word engineering according to claim 1, characterized in that, In step S6, the final prompt word is updated in real time based on the health status data, including: When the health status data shows that the user's health indicators fluctuate beyond the preset range, add a health risk warning message to the final prompt. The updated final prompt words will be saved synchronously to the prompt word library.

7. The medical follow-up management method based on large model prompt word engineering according to claim 5, characterized in that, In step S6, the final prompt word is updated in real time based on the behavioral data, including: When behavioral data shows that the user's feedback on the follow-up plan does not meet the preset requirements, add relevant content on incentive points to guide the user in the final prompt. When behavioral data shows that the user's feedback on the follow-up plan meets the preset requirements, add content related to the positive effect of health management to the final prompt. The updated final prompt words will be saved synchronously to the prompt word library.

8. A medical follow-up management system based on large-scale model prompt word engineering, characterized in that, The system is used to execute the medical follow-up management method based on large model prompt word engineering as described in any one of claims 1-7, the system comprising: The prompt word acquisition module is used to acquire prompt words from the prompt word library according to the follow-up scenario, and optimize the prompt words to obtain the final prompt words; The data collection module is used to collect users' health status data and lifestyle data; A follow-up plan generation module, connected to the data collection module, is used to input the health status data and the lifestyle data into the follow-up dialogue model, and the follow-up dialogue model generates a preset follow-up plan based on the input data; The follow-up plan adjustment module is connected to the prompt word acquisition module and the follow-up plan generation module, and is used to control the follow-up dialogue model to adjust the preset follow-up plan according to the final prompt word and output the final follow-up plan. The behavioral data acquisition module is used to acquire user behavioral data related to the execution of the final follow-up plan; The real-time update module is connected to the behavior data acquisition module, the collection module, and the prompt word acquisition module, and is used to update the final prompt word in real time based on the health status data, the lifestyle data, and the behavior data.

Images