Health management target generation method and system based on man-machine conversation

Abstract

The invention relates to the technical field of health management, and particularly discloses a human-machine dialogue-based health management target generation method and system, and the method comprises the steps: starting a health dialogue, and collecting initial information; performing deep semantic analysis on initial information input by a user by utilizing a natural language processing and sentiment analysis technology, and extracting key information; fusing the key information with historical behavior data from the wearable device and the health APP, and constructing a current comprehensive health state portrait of the user; generating a short-term target and a long-term target of the current situation based on a preset health management knowledge graph and a trained generative AI target recommendation model; displaying the generated short-term target and the long-term target of the current situation to the user in a natural language form, and waiting for the feedback of the user; the feedback of the user is recorded as context information; and according to a user execution target and a feedback result, recording a completion condition, inputting the completion condition as historical context information in a next round of dialogue, and optimizing a target generation strategy.

Description

Technical Field

[0001] This invention relates to the field of health management technology, specifically to a method and system for generating health management goals based on human-computer dialogue. Background Technology

[0002] In the current field of health management, traditional health management systems generally follow a rather rigid and passive operating model. They mainly rely on static health data collection and one-way goal generation mechanisms. Typically, when a user first uses the system, they are required to manually input structured data such as age, weight, and medical history. The system then combines this with a preset rule engine to generate health goals for fixed periods. Furthermore, goal adjustments are extremely passive, requiring the user to manually trigger a reassessment process before the system will modify the goals. This traditional system has many obvious shortcomings. With the rapid development of artificial intelligence technology, generative AI, edge computing, and multimodal interaction technologies have achieved groundbreaking applications. Generative AI can generate creative and targeted content based on massive amounts of data, edge computing enables rapid local data processing and reduces transmission latency, and multimodal interaction technologies can integrate multiple information input methods and provide a more natural and convenient interactive experience. However, existing health management technologies have not fully integrated these innovative achievements.

[0003] Therefore, by constructing a multimodal dialogue interaction engine and a generative AI target recommendation model, and combining it with an edge-cloud collaborative computing architecture, a leap from "passive response" to "proactive health intervention" has been achieved, significantly improving the real-time, personalization, feasibility, and privacy security of health management target generation, and providing a more efficient, accurate, and sustainable solution for the health management field. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for generating health management goals based on human-computer dialogue, and to solve the following technical problems: How can we leverage semantic understanding, knowledge reasoning, and generative AI technologies to construct dynamic health profiles of users and generate personalized, actionable, and sustainably optimized short-term and long-term health management goals based on these profiles?

[0005] The objective of this invention can be achieved through the following technical solutions: A method for generating health management goals based on human-computer dialogue, the method comprising: Step S1: Initiate a health dialogue and collect initial information; Step S2: Utilize natural language processing and sentiment analysis techniques to perform deep semantic analysis on the initial information input by the user and extract key information; Step S3: Integrate key information with historical behavioral data from wearable devices and health apps to construct a comprehensive profile of the user's current health status; Step S4: Based on the preset health management knowledge graph and the trained generative AI goal recommendation model, generate short-term and long-term goals for the current situation; Step S5: Display the generated short-term and long-term goals of the current situation to the user in natural language, and wait for the user's confirmation, rejection, or suggestions for modification; the user's feedback is recorded as contextual information; Step S6: Based on the user's execution of the goal and feedback results, record the completion status and use it as historical context information input in the next round of dialogue to optimize the goal generation strategy.

[0006] Preferably, steps S1-S3 utilize a multimodal dialogue interaction engine to achieve multi-channel information collection and preliminary analysis of user status, specifically including: The voice input module is used to recognize and input user voice in noisy environments and dialect scenarios using an end-to-end speech recognition model based on Transformer. The NLP module is used to achieve intent recognition and entity extraction based on pre-trained language models and medical knowledge graphs. The sentiment analysis module is used to determine the user's emotional state by analyzing speech tone features and a text sentiment dictionary. The physiological signal analysis module is used to access data from smart devices in real time and, combined with threshold warning and pattern recognition algorithms, determine the user's physiological state.

[0007] Preferably, the sentiment analysis module includes constructing a sentiment analysis model: An LSTM network is used to process speech temporal features, and TextCNN is used to extract text sentiment features. A multimodal fusion layer is then used to output sentiment labels. The formula for calculating the sentiment analysis model is: in, This represents the output emotion classification probability vector, whose dimension is equal to the preset number of emotion categories. It is used to represent the probability distribution of the user's current emotional state, such as "anxiety", "calm", or "fatigue". This is a normalized activation function used to convert the fused feature vector into probability values; This is the weight matrix of the fusion layer. This is a temporal feature vector extracted from the user's speech signal by the Long Short-Term Memory Network, reflecting the acoustic characteristics of speech such as fundamental frequency, energy, and speech rate; This is the sentiment feature vector after semantic encoding of user input text by a convolutional neural network; This means concatenating the speech feature vector with the text feature vector to form a joint representation; This is the bias vector for the fusion layer, used to adjust the output offset.

[0008] Preferably, the generative AI goal recommendation model achieves intelligent generation and optimization of health management goals, and the implementation methods include: Integrate structured and unstructured data output from the multimodal interaction engine, and extract key features through feature engineering; A Transformer-based autoregressive generative model is adopted, combined with reinforcement learning to optimize the target generation strategy; The Transformer-based autoregressive generative model takes the user's health status and historical goal completion records as input and outputs the goal sequence. .

[0009] Preferably, the generative AI target recommendation model is based on an autoregressive generative mechanism that uses an attention mechanism to capture the dependency relationship between health status and the target. The calculation formula is as follows: in, Indicates the health status of a given user Under the given conditions, generate the target sequence The joint probability; For the generated target sequence, by Composed of word elements or sub-targets, Indicates the first in the target sequence Each word element; ; Indicates in generation The previously generated lexical sequences are used to provide contextual dependencies; This represents a vector representing the user's current overall health status. This represents the trainable parameters of the generative model, including attention mechanism weights and feedforward network parameters; Preferably, the reinforcement learning optimization method for the generative AI target recommendation model is as follows: The PPO algorithm is used, with goal completion rate, user compliance, and improvement in health metrics as reward functions, to optimize the goal generation strategy. The optimization function for the goal generation strategy is as follows:

[0010] in, The objective function representing the policy gradient is used to update the parameters of the generated model. ; This indicates the user's health status under the current strategy. Select Action The probability of; This represents the probability of the old policy in the previous iteration; The dominant function is represented by ϵ, which is the clipping coefficient and typically ranges from 0 to 1. .

[0011] Preferably, the method further includes: Edge-cloud collaborative computing architecture, whereby the cloud aggregates edge node model parameters through weighted average: in, This represents the aggregated global generative AI model parameters; This indicates the total number of edge devices participating in this round of federated learning; For edge devices, ; Indicates the first Model parameters uploaded by an edge device after local training; Indicates the first The number of health data samples used by each device during local training; This represents the total number of samples from all participating devices, i.e. .

[0012] A health management goal generation system based on human-computer dialogue, utilizing a human-computer dialogue method for generating health management goals, the system includes: The data collection unit is used to initiate a health dialogue and collect initial information. The extraction unit is used to perform deep semantic analysis on the initial information input by the user using natural language processing and sentiment analysis technologies, and to extract key information. The fusion unit is used to integrate key information with historical behavioral data from wearable devices and health apps to build a comprehensive profile of the user's current health status. The goal generation unit is used to generate short-term and long-term goals for the current situation based on a preset health management knowledge graph and a trained generative AI goal recommendation model. The display feedback unit is used to show the short-term and long-term goals of the generated current situation to the user in natural language, and wait for the user to confirm, reject or suggest modifications; the user's feedback is recorded as contextual information; The goal optimization unit is used to record the completion status based on the user's execution of the goal and feedback results, and to use this as historical context information input in the next round of dialogue to optimize the goal generation strategy.

[0013] The beneficial effects of this invention are: (1) Significantly improved the personalization and context-adaptability of health management goals. Traditional systems rely on fixed templates or structured questionnaires to generate goals, making it difficult to understand users' real-life situations. This invention, through multi-turn natural language dialogue (step S1) combined with multimodal semantic parsing (step S2), can identify users' subjective feelings (e.g., "recently stressed"), life events (e.g., "constantly working overtime"), and potential needs (e.g., "wanting to improve sleep"), and integrates objective physiological data collected by wearable devices (step S3) to construct a dynamic health profile. For example, when the system identifies that a user has "high subjective fatigue score + low HRV + average sleep of less than 6 hours in the past three days," it prioritizes recommending "increase lunch break" rather than "increase exercise," making the goal more relevant to the user's actual situation and improving acceptance and willingness to implement.

[0014] (2) Significantly improved the feasibility of health goals and user compliance. Existing technologies often generate goals that are too general or unrealistic, leading users to abandon them. This invention introduces a generative AI model and reinforcement learning optimization mechanism in the goal generation stage (step S4). It automatically assesses goal difficulty by combining historical behavioral data and improves operability through goal decomposition and feasibility verification strategies. For example, for the goal of "1 hour of exercise daily," the system can automatically break it down into "20 minutes of walking in the morning + 40 minutes of brisk walking after dinner," and suggest execution times based on the user's schedule. Empirical data shows that this mechanism increases the average daily goal completion rate by more than 38%, significantly enhancing the sustainability of health management.

[0015] (3) It has achieved the ability to dynamically optimize and evolve health management goals in a closed loop over a long period of time. Traditional systems lack feedback and adjustment mechanisms, and goals remain fixed once set. This invention forms a complete closed loop of "generation—execution—feedback—optimization" through a user feedback confirmation mechanism (step S5) and a dynamic tracking and update mechanism (step S6). When a user reports that "the goal is too difficult," the system can automatically reduce the intensity (e.g., adjusting "walking 8000 steps" to "6000 steps"); as the user continuously completes the goal, the system gradually increases the challenge level. This mechanism, combined with federated learning technology in an edge-cloud collaborative architecture, achieves the unification of individual strategy optimization and group model iteration, enabling health management strategies to continuously evolve over time.

[0016] (4) It reduces the reliance on professional personnel and supports intelligent health management for large-scale populations. Current personalized goal setting relies on manual intervention by doctors or health managers, which is costly and inefficient. This invention automates the entire process from information collection and status assessment to goal recommendation through the collaborative work of a multimodal dialogue interaction engine and a generative AI goal model. The system can serve tens of thousands of users simultaneously, with a single goal generation response time of less than 1.5 seconds, significantly reducing service costs. This feature is particularly suitable for scenarios requiring high concurrency and low latency responses, such as enterprise EAPs and chronic disease management platforms.

[0017] (5) Enhanced user privacy protection and data security Unlike traditional AI systems that centrally upload raw health data, this invention employs an edge-cloud collaborative computing architecture and a federated learning mechanism. It performs preliminary processing and local modeling of sensitive data (such as voice and HRV) on edge devices, uploading only encrypted model parameters to the cloud for aggregation and updates. This design avoids centralized storage and transmission of users' raw physiological and conversational data, complies with privacy regulations such as GDPR and HIPAA, and enhances system security while ensuring a high level of intelligence.

[0018] Of course, any product implementing this invention does not necessarily need to achieve all the advantages described above at the same time. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the steps of a health management goal generation method based on human-computer dialogue according to the present invention. Figure 2 This is a technical flowchart of the information perception, intelligent decision-making, and system architecture aspects of the present invention; Figure 3 This is a unit diagram of a health management goal generation system based on human-computer dialogue according to the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Traditional systems have many obvious drawbacks: Service passivity: It heavily relies on users actively inputting data and lacks multimodal interaction capabilities. For example, when a user complains of "recent insomnia," the system cannot identify the underlying mental health management needs and struggles to provide targeted intervention suggestions.

[0023] Data utilization gap: Only single-dimensional data such as blood pressure and steps are processed, without taking into account emotional state or device data, resulting in a single dimension of personalized recommendations that cannot meet the diverse health management needs of users.

[0024] Delayed goal generation: Relying on periodic data reporting or simple machine learning models, the system cannot respond promptly to sudden health events. For example, if a user experiences a sudden abnormal heart rate, the system cannot immediately adjust exercise goals or push emergency intervention suggestions.

[0025] The interaction mode is one-way: it only supports the system passively responding to user queries and lacks the ability to actively guide users. For example, if the system does not recognize the user's fatigue state, it will not actively reduce the exercise target, which may cause the user to over-exercise.

[0026] Based on the above issues, with the rapid development of artificial intelligence technology, generative AI, edge computing, and multimodal interaction technologies have achieved breakthrough applications. Generative AI can generate creative and targeted content based on massive amounts of data, edge computing can achieve rapid local data processing and reduce transmission latency, and multimodal interaction technology can integrate multiple information input methods and provide a more natural and convenient interactive experience.

[0027] However, existing health management technologies have failed to fully integrate these innovative achievements. Therefore, this invention collects user health information through multi-turn natural language dialogue, combines it with objective physiological data from wearable devices and other sources, and utilizes semantic understanding, knowledge reasoning, and generative AI technologies to construct a dynamic health profile of the user. Based on this profile, personalized, actionable, and sustainably optimized short-term and long-term health management goals are generated. This invention is a method and system for generating health management goals based on human-computer dialogue. Example

[0028] Please see Figure 1 As shown, the method flow steps are as follows: Step S1: Initiate a health conversation and collect initial information The system proactively initiates health conversations via voice or text (e.g., "How are you feeling today?", "Have you been sleeping well lately?"), guiding users to express their current physical condition, emotions, and lifestyle habits. This step is driven by a multimodal dialogue interaction engine, supporting multiple input methods such as voice and text, improving the naturalness of interaction and the efficiency of information acquisition.

[0029] Step S2: Semantic parsing and key information extraction By leveraging Natural Language Processing (NLP) and sentiment analysis techniques to perform deep semantic analysis of user input, the following four key health elements were extracted: Subjective feelings (fatigue, anxiety, high energy, etc.); Life events (overtime work, travel, parties, etc.); Health needs (weight loss, sleep aid, relief of neck and shoulder pain, etc.); Obstacles (lack of time, easy to give up, etc.); This process is completed by the NLP and sentiment analysis modules in the multimodal dialogue interaction engine, ensuring that the system can not only "understand" the literal meaning, but also identify the user's potential emotions and real needs.

[0030] Step S3: Integrate multi-source data for state assessment The subjective information extracted in step S2 is integrated with historical behavioral data from wearable devices and health apps (such as average steps in the past 7 days, sleep duration, resting heart rate, HRV, etc.) to construct a comprehensive health profile of the user. This step relies on the physiological signal parsing module in the multimodal dialogue interaction engine and the data fusion layer in the generative AI target generation model to achieve a unified representation of subjective descriptions and objective indicators.

[0031] Step S4: Generation of short-term and long-term goals Based on a pre-set health management knowledge graph and a trained generative AI goal recommendation model, the system automatically generates short-term and long-term goals that match the user's current situation. This step is led by a generative AI goal generation model, which combines reinforcement learning strategies to optimize the feasibility of the goals and user compliance.

[0032] Step S5: Output the target and confirm user feedback The generated goals are presented to the user in natural language (e.g., "I suggest you take a 15-minute nap tomorrow, is that okay?"), and the user is asked to confirm, decline, or provide feedback. User feedback is recorded as contextual information for subsequent goal adjustments, forming a closed loop of human-machine collaborative decision-making.

[0033] Step S6: Dynamic Updates and Long-Term Tracking If the user completes the task and provides feedback (such as "completed" or "not completed"), the system will record the completion status and use it as historical context input in the next round of dialogue to optimize subsequent task generation strategies. This mechanism supports the continuous evolution and personalized refinement of tasks, improving long-term health management effectiveness.

[0034] The above Example 1 uses Ms. Zhang's sleep and stress management goal generation process as an example: User background information: Name: Ms. Zhang, Age: 32, Occupation: Product Manager at an Internet company, Devices used: Huawei Watch (collects HRV, sleep, and steps), iPhone 14 (runs system apps), Health concerns: Long hours of overtime work leading to decreased sleep quality, daytime fatigue, and anxiety; The specific implementation process is as follows: Step S1: Initiate a health conversation and collect initial information Triggering mechanism: The system automatically initiates a dialogue based on the user's historical behavior (usually starting a health dialogue around 21:00).

[0035] The system proactively pushed a message: "Ms. Zhang, how are you feeling today? Have you been sleeping well lately?" User voice input: "Sigh, I'm so tired lately. I only slept for a little over 5 hours last night, and my mind went completely blank during the meeting this morning." System response: The multimodal dialogue interaction engine immediately activates the voice input module to convert speech into text.

[0036] At the same time, the system synchronizes objective data from Huawei devices over the past 7 days: Average sleep duration: 5.4 hours (target: 7 hours) Mean HRV (heart rate variability): 48ms (normal range is 60-100ms, indicating high stress). Average daily steps: 6,200 steps (below the recommended 8,000 steps) Step S2: Semantic parsing and key information extraction NLP and sentiment analysis module running: Semantic understanding: Extracting the entity: "tired" → subjective feeling (fatigue). "Only slept for a little over 5 hours" → Health concern (sleep aid) "Blanking out during meetings" → Decreased cognitive function Sentiment analysis results: Emotional score: 78% probability of negative emotion. Emotional tags: Anxiety, Fatigue Output structured health events: { "subjective":{ "feelings":["fatigue","anxiety"], "events":["Worked overtime for 3 consecutive days"], "goals":["Improve sleep"], "barriers": ["High work stress", "Difficulty falling asleep"] }} Step S3: Integrate multi-source data for state assessment Data fusion layer builds a health profile: enter: Subjective data: fatigue, anxiety, sleep needs Objective data: HRV = 48ms (low), sleep = 5.4h (insufficient), steps = 6.2k (moderate) Feature engineering: Construct a composite characteristic: "Deteriorating sleep quality" = f(subjective insomnia + decreased HRV + reduced nighttime activity) Calculate the stress index: (High pressure) Output health state vector: S = [Fatigue = 0.85, Anxiety = 0.78, Sleep Deprivation = 0.92, Stress Index = 0.72, Lack of Exercise = 0.65] Step S4: Generation of short-term and long-term goals Generative AI goal generation model invocation: Input: Health status vector S + historical goal completion record (68% completion rate in the past 7 days) Combining knowledge graph rules: If HRV < 50 and sleep < 6 hours, then relaxation-related goals should be prioritized over exercise-related goals. Generate a goal: "It is recommended to go to bed 30 minutes earlier tonight and do 10 minutes of meditation" (short-term). The model uses autoregression to generate a natural language objective: "I suggest you go to bed 30 minutes earlier tonight and do 10 minutes of meditation, is that okay?" Step S5: Output the target and confirm user feedback The system output: "We have detected that your stress level is high and your HRV is low recently. We suggest you go to bed 30 minutes earlier tonight and do 10 minutes of meditation to relax and help you fall asleep. Do you think this suggestion is appropriate?" User feedback (voice): "Meditation is fine, but it's a bit difficult to do it 30 minutes in advance. Could you do it just 15 minutes in advance?" System response: The NLP module identifies "partial acceptance + modification requests". Trigger the target regeneration mechanism and input a new constraint: "Bedtime can only be 15 minutes earlier." The reinforcement learning optimization algorithm is used to generate a new objective: "It is recommended to go to bed 15 minutes earlier tonight and do 10 minutes of breathing exercises (shorter than meditation)." Step S6: Dynamic Updates and Long-Term Tracking Goal execution and feedback: When a user completes a new goal, the app records it as "completed". The following morning, the system synchronized the data: Sleep duration last night: 6.1 hours (+0.7h) Morning HRV: 54ms (+6ms) Target update strategy: In the next round of dialogue (Day 3), the system will retrieve the historical completion record: Target completion rate in the last 3 days: 83% → Increase the difficulty of the target The new long-term goal is: "To increase average sleep to 6.8 hours over the next four weeks, and to add 15 minutes to the time it takes to fall asleep each week." Evolution of swarm intelligence: The system fine-tunes the PPO strategy model locally and uploads the encrypted parameters to the cloud. The cloud then updates the global model using a federated learning aggregation formula, achieving a closed loop of "individual experience → group optimization." Table 1 below shows the implementation results and performance indicators:

[0037] ② Implementation of core technology modules To implement the above method and process, this invention has constructed three core technical modules, please refer to [link / reference]. Figure 2 As shown, these correspond to the information perception, intelligent decision-making, and system architecture levels, respectively: 1. Multimodal dialogue interaction engine Function: As the "perception layer" of the system, it is responsible for multi-channel information collection and preliminary analysis of user status, supporting the implementation of steps S1 to S3.

[0038] Technical Implementation: Voice input module: Employing an end-to-end speech recognition model based on Transformer, it can still accurately recognize user voice input even in noisy environments and dialect scenarios. For example, if a user describes "my knees hurt after exercising today," the system can recognize the keywords and trigger a joint health assessment process.

[0039] The NLP module, based on a pre-trained language model and combined with a medical knowledge graph, enables intent recognition and entity extraction. For example, if a user inputs "I've been having insomnia lately," the system can identify "insomnia" as a health problem entity and link it to a sleep management knowledge base.

[0040] Sentiment analysis module: By analyzing voice tone characteristics (fundamental frequency, energy) and a text sentiment dictionary, the system determines the user's emotional state. For example, if a user's voice trembles or words such as "anxiety" or "frustration" appear in the text, the system can identify potential psychological stress and trigger a mental health intervention process.

[0041] Physiological signal analysis module: This module receives real-time data from smart devices (such as heart rate variability (HRV) and electrical skin response (EDA), and combines this data with threshold warning and pattern recognition algorithms to determine the user's physiological state. For example, if HRV remains below the threshold, the system can identify that the user is under stress and adjust their exercise goals accordingly.

[0042] Core Algorithm: Sentiment analysis model: An LSTM network is used to process temporal features of the speech, combined with TextCNN to extract sentiment features from the text, and a multimodal fusion layer is used to output sentiment labels. The formula is: in: This represents the output emotion classification probability vector, whose dimension is equal to the preset number of emotion categories. It is used to represent the probability distribution of the user's current emotional state, such as "anxiety", "calm", or "fatigue". This is a normalized activation function used to convert the fused feature vector into probability values; The weight matrix of the fusion layer has its dimension determined by the concatenation dimension of the speech feature vector and the text feature vector, and is a trainable parameter. This is a temporal feature vector extracted from the user's speech signal by the Long Short-Term Memory (LSTM) network, reflecting the acoustic characteristics of the speech such as fundamental frequency, energy, and speech rate; This is the sentiment feature vector generated by the convolutional neural network (TextCNN) after semantically encoding the user input text; This means concatenating the speech feature vector with the text feature vector to form a joint representation; This is the bias vector for the fusion layer, used to adjust the output offset.

[0043] The formula for calculating the probability vector of emotion classification is used to achieve the fusion processing of bimodal emotional information of speech and text, thereby improving the system's accuracy in recognizing users' subjective psychological states. For example, when a user verbally expresses "I'm fine" but their voice features show trembling, the system can still identify potential anxiety, thus triggering a mental health intervention mechanism.

[0044] 2. Generative AI Target Generation Model Function: As the "decision layer" of the system, it is responsible for the intelligent generation and optimization of health management goals, supporting the core logic of step S4.

[0045] Technical Implementation: Data Fusion Layer: This layer integrates structured data (such as blood pressure and step count) and unstructured data (such as voice text and sentiment tags) output by the multimodal interaction engine, and extracts key features through feature engineering. For example, it merges "insomnia," "decreased HRV," and "reduced nighttime step count" into a single feature called "decreased sleep quality." Generative model architecture: A Transformer-based autoregressive generative model is employed, combined with reinforcement learning (PPO) to optimize the target generation strategy. The model input consists of the user's health status (SS) and historical target completion records; the output is a sequence of targets. ; Target generation logic: Short-term goals: Focus on specific, actionable actions, such as "walk 8,000 steps today" or "meditate for 15 minutes before bed"; Long-term goals: Focus on improving health indicators, such as "reducing BMI to below 24 within 3 months" and "improving sleep efficiency to 85% within 6 months".

[0046] Core Algorithm: Goal generation model: Based on an autoregressive generation mechanism, it captures the dependency between health status and goals through an attention mechanism. The calculation formula is as follows: in: This represents the joint probability of generating the target sequence T given the user's health state S. The generated target sequence consists of several word elements or sub-targets, such as "It is recommended to go to bed 15 minutes earlier tonight"; Indicates the first in the target sequence Each word element; ; Indicates in generation The previously generated lexical sequences are used to provide contextual dependencies; This represents a vector indicating the user's current overall health status; it is a fusion of subjective descriptions (such as fatigue and anxiety) and objective physiological data (such as HRV and sleep duration). This represents the trainable parameters of the generative model, including attention mechanism weights and feedforward network parameters. This formula is used for natural language generation to achieve personalized health management goals. Through an autoregressive mechanism, the model can generate grammatically correct, semantically reasonable, and context-appropriate suggestion statements, significantly improving the naturalness of human-computer interaction and user compliance.

[0047] Reinforcement learning optimization: The PPO algorithm is adopted, with the goal completion rate, user compliance, and improvement of health indicators as reward functions, to optimize the goal generation strategy. The optimization function of the goal generation strategy is as follows: in: The objective function representing the policy gradient is used to update the parameters of the generated model. ; This indicates the action to be selected in state s (i.e., user health status) under the current policy. (i.e., the probability of generating a certain type of health goal); This represents the probability of the old policy in the previous iteration; ϵ represents the advantage function, which measures the superiority of the current action relative to the average performance. Its value is calculated based on the goal completion rate, user feedback satisfaction, and the improvement in physiological indicators; ϵ is the clipping coefficient, which typically ranges from 0 to 1. This formula is used to limit the policy update step size and prevent instability during training. It enables continuous optimization of the health management goal generation strategy, allowing the system to automatically adjust recommendation preferences based on different users' response patterns, significantly improving the acceptability and long-term effectiveness of the goals.

[0048] 3. Edge-Cloud Collaborative Computing Architecture Function: As the "operational support layer" of the system, it achieves low-latency response and privacy protection, ensuring efficient and stable operation throughout the entire process.

[0049] Technical Implementation: Edge layer: Deployed on mobile phones, smartwatches, or local gateways, it is responsible for real-time data acquisition, preprocessing, and lightweight analysis. For example, edge nodes perform real-time threshold detection on HRV data, triggering alerts and uploading data to the cloud when anomalies are detected.

[0050] Cloud layer: Deploys generative AI models and big data analytics platforms, responsible for training complex models and analyzing overall health trends. For example, it aggregates multi-user data in the cloud to optimize the generalization ability of the target generative model.

[0051] Collaborative mechanism: Federated learning technology is adopted, where edge nodes train the model locally and upload parameter updates, and cloud aggregation updates realize model iteration, avoiding the transmission of raw data.

[0052] Core Algorithm: Federated learning aggregation: Edge node model parameters are aggregated in the cloud using a weighted average. in: This represents the aggregated global generative AI model parameters, used to update the local models of each edge device; This indicates the total number of edge devices participating in this round of federated learning; For edge devices, ; Indicates the first Model parameters uploaded by an edge device after local training; Indicates the first The number of health data samples used by each device during local training; This represents the total number of samples from all participating devices, i.e. Used to normalize aggregate weights.

[0053] This formula is used to achieve collaborative optimization of models while protecting privacy; it only uploads model parameters instead of raw data to avoid privacy leaks, while using group behavior data to improve the model's generalization ability. Example

[0054] Please see Figure 3 As shown, a health management goal generation system based on human-computer dialogue includes: The data collection unit is used to initiate a health dialogue and collect initial information. The extraction unit is used to perform deep semantic analysis on the initial information input by the user using natural language processing and sentiment analysis technologies, and to extract key information. The fusion unit is used to integrate key information with historical behavioral data from wearable devices and health apps to build a comprehensive profile of the user's current health status. The goal generation unit is used to generate short-term and long-term goals for the current situation based on a preset health management knowledge graph and a trained generative AI goal recommendation model. The display feedback unit is used to show the short-term and long-term goals of the generated current situation to the user in natural language, and wait for the user to confirm, reject or suggest modifications; the user's feedback is recorded as contextual information; The goal optimization unit is used to record the completion status based on the user's execution of the goal and feedback results, and to use this as historical context information input in the next round of dialogue to optimize the goal generation strategy.

[0055] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0056] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended documents. In some cases, the actions or steps described in this application may be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0057] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined in this application, they should all fall within the protection scope of the present invention.

Claims

1. A health management goal generation method based on human-computer conversation, characterized by, The method comprises: S1, starting a health dialogue and collecting initial information; S2, using natural language processing and sentiment analysis technology to perform deep semantic analysis on the initial information input by the user, and extracting key information; S3, fusing the key information with historical behavior data from wearable devices and health APPs to construct a current comprehensive health status portrait of the user; S4, generating short-term and long-term goals for the current situation based on a pre-set health management knowledge graph and a trained generative AI target recommendation model; S5, displaying the generated short-term and long-term goals for the current situation to the user in natural language form, and waiting for the user to confirm, reject or propose modifications; the user's feedback is recorded as context information; Step S6, according to the user's execution of the goal and feedback, record the completion status, and input it as historical context information in the next round of dialogue to optimize the goal generation strategy. 2.The health management goal generation method based on human-computer dialogue according to claim 1, characterized in that, S1-S3 are realized by a multi-modal dialogue interaction engine to collect multi-channel information and preliminarily analyze the user's state, which specifically includes: A voice input module for recognizing and inputting user voice in noisy environments and dialect scenarios using an end-to-end speech recognition model based on Transformer; An NLP module for realizing intent recognition and entity extraction based on a pre-trained language model combined with a medical knowledge graph; An emotion analysis module for determining the user's emotional state through voice tone features and a text sentiment dictionary; A physiological signal analysis module for real-time access to smart device data, combining threshold warning and pattern recognition algorithms to determine the user's physiological state. 3.The health management goal generation method based on human-computer dialogue according to claim 2, characterized in that, The emotion analysis module includes building an emotion analysis model: Using LSTM network to process voice timing features, combining Text CNN to extract text sentiment features, and outputting emotion labels through a multi-modal fusion layer; The calculation formula of the emotion analysis model is: wherein, represents an output emotion classification probability vector, the dimension of which is equal to the number of preset emotion categories, used to represent the probability distribution of the user being in an "anxious", "calm", "tired" or other emotional state; is a normalization activation function, used to convert the fused feature vector into a probability value; is a weight matrix of the fusion layer, is a time sequence feature vector extracted by the long short-term memory network from the user's voice signal, reflecting the acoustic features of the voice such as fundamental frequency, energy, speech rate, etc. is a sentiment feature vector obtained by the convolutional neural network after semantic encoding of the user's input text; represents a concatenation operation of the voice feature vector and the text feature vector to form a joint representation; is a bias vector of the fusion layer, used to adjust the output offset. 4.The health management goal generation method based on human-computer dialogue according to claim 1, wherein, The generative AI target recommendation model realizes intelligent generation and optimization of health management goals, and the implementation includes: Integrating structured and unstructured data output by the multi-modal interaction engine to extract key features through feature engineering; Using a Transformer-based autoregressive generation model combined with reinforcement learning to optimize the goal generation strategy; The input of the Transformer-based autoregressive generation model is the user health status and historical target completion record, and the output is the target sequence .

5. The health management goal generation method based on human-computer dialogue according to claim 4, characterized in that, The generative AI target recommendation model is based on an autoregressive generation mechanism that captures the dependency between health status and goals through an attention mechanism, and the calculation formula is: wherein, denotes the joint probability of generating a target sequence given the user's health state ; is the generated target sequence, consisting of word pieces or sub-targets, denotes the word piece in the target sequence; ; denotes the sequence of word pieces that have been generated before generating , used to provide context dependency; denotes the user's current comprehensive health state vector; denotes the trainable parameters of the generative model, including attention mechanism weights, feed-forward network parameters.

6. The health management goal generation method based on human-computer dialogue according to claim 5, characterized in that, The reinforcement learning optimization method of the generative AI target recommendation model is: Using PPO algorithm to optimize the goal generation strategy with target completion rate, user adherence, and health index improvement as reward functions; the optimization function of the goal generation strategy is: where, represents the optimization objective function of the policy gradient for updating the generative model parameters ; represents the probability of selecting an action under the current policy for the user's health state ; represents the advantage function; ϵ is the clipping coefficient, usually in the range of .

7. The method of claim 1, wherein the health management goal is generated based on a dialogue between the user and the computer. The method further comprises: ​ Edge-cloud collaborative computing architecture, i.e. the cloud aggregates edge node model parameters through weighted average: wherein, denotes the global generative AI model parameters after aggregation; denotes the total number of edge devices participating in the current round of federated learning; is the index of the edge device, ; denotes the model parameters uploaded by the th edge device after local training; denotes the number of health data samples used by the th device during the local training process; denotes the total number of samples from all participating devices, i.e. .

8. A health management goal generation system based on human-machine conversation, characterized by, Using the health management goal generation method based on human-computer dialogue according to any one of claims 1-7, the system comprises: An acquisition unit for starting a health dialogue and collecting initial information; An extraction unit for using natural language processing and sentiment analysis technology to perform deep semantic analysis on the initial information input by the user, and extracting key information; A fusion unit is configured to fuse the key information with historical behavior data from the wearable device and the health APP to construct a comprehensive health status portrait of the user; A target generation unit is configured to generate short-term and long-term targets for the current situation based on a preset health management knowledge graph and a trained generative AI target recommendation model; A display feedback unit is configured to display the generated short-term and long-term targets for the current situation in a natural language form to the user and wait for the user to confirm, reject or provide modification suggestions; the feedback of the user is recorded as context information; A target optimization unit is configured to record the completion according to the user's execution of the target and feedback of the result, and input the historical context information in the next round of dialogue to optimize the target generation strategy.

Images