Weight management program adjustment method and system based on multi-party feedback

Abstract

The present application relates to the technical field of AI health management, in particular to a weight management scheme adjustment method and system based on multi-party feedback, which constructs a user health state vector by using multi-source health data, takes the health state vector as the environment state, takes the user diet and exercise scheme as the action, corrects the candidate action output by the current state of the strategy network by the interaction of the intelligent agent and the environment and by using expert knowledge and safety constraints to obtain the best strategy, wherein the reward function is composed of weight trend reward, execution rate reward, expert feedback reward and risk penalty, and the safety constraint is a pre-set weight management safety constraint boundary set. The present application converts medical record style free text feedback into structured reward signals and strategy correction amounts, and generates an interpretable, reviewable and sustainable updated personalized weight loss scheme by combining real-time physiological feedback, behavior execution data and metabolic safety constraints of weight loss users, so as to realize scientific and healthy management of the weight of the user.

Description

Technical Field

[0001] This invention relates to the field of AI health management technology, and in particular to a method and system for adjusting weight management programs based on multi-party feedback. Background Technology

[0002] Obesity and overweight management typically requires long-term dietary control, exercise intervention, sleep management, and behavioral support. Traditional weight loss programs often rely on fixed templates or infrequent manual follow-ups, making it difficult to respond promptly to changes in a user's weight, exercise load, sleep fluctuations, dietary deviations, and psychological state. While existing health management apps can record weight, steps, and diet, most only provide statistical displays or rule-based reminders, failing to continuously translate doctors' experience into a calculable, traceable, and iterative dynamic strategy update mechanism.

[0003] In recent years, wearable devices, large language models, and reinforcement learning technologies have been used in health management scenarios. However, simple model recommendations have two problems: First, the model output lacks supervision from medical experts and is prone to generating excessive calorie deficits, excessive exercise intensity, or plans that are not suitable for special populations. Second, doctors' opinions are often in the form of natural language. If they are only saved as medical records or notes, they cannot be directly entered into the model update process, and the expert experience is difficult to be continuously absorbed by the system.

[0004] Reinforcement learning is suitable for dynamically adjusting weight loss programs because weight loss management is not a one-time classification or static recommendation problem, but a continuous decision-making problem. The system observes the user's status daily or in each intervention cycle, outputting dietary and exercise recommendations. It then rewards or penalizes users based on weight trends, adherence rates, sleep patterns, heart rate changes, and doctor feedback, updating subsequent strategies accordingly. Compared to single-batch supervised learning predictions, reinforcement learning can express "the impact of current adjustments on future multi-day adherence and health outcomes." However, directly applying ordinary reinforcement learning to health management carries exploratory risks. Therefore, there is an urgent need for a technology that can dynamically adjust weight loss programs, translate doctors' natural language opinions into machine-executable updates, and output the final plan within metabolic safety boundaries to achieve scientific health management for weight loss users. Summary of the Invention

[0005] To address the problems of existing weight management programs being static, lacking clear boundaries, insufficient personalization, and having limited room for exploration, this invention provides a method and system for adjusting weight management programs based on multi-party feedback. Without altering the natural writing habits of doctors and other experts, it transforms free-form text feedback in the style of medical records into structured reward signals and strategy correction quantities. Combined with real-time physiological feedback, behavioral execution data, and metabolic safety constraints from weight loss users, it generates interpretable, verifiable, and continuously updatable personalized weight loss programs to achieve scientific and healthy weight management for users.

[0006] According to the design scheme provided by the present invention, on the one hand, a method for adjusting a weight management plan based on multi-party feedback is provided, comprising:

[0007] The user's health data is acquired from multiple sources based on user terminals and wearable devices, and the data is preprocessed to obtain the user's health status vector.

[0008] The user's health status vector is used as the environment state, and the user's diet and exercise plan is used as the action. The user's weight management plan is dynamically adjusted and modeled as a Markov decision process by combining the reward function. Through the interaction between the agent and the environment in reinforcement learning, and by using expert knowledge and safety constraints, the candidate actions of the policy network and the current state output are modified to obtain the best weight management strategy for the user. The reward function consists of weight trend reward, execution rate reward, expert feedback reward and risk penalty. The safety constraints are a pre-set set of weight management safety constraint boundaries.

[0009] The system translates the user's current weight management strategy into an executable language, enabling the user to execute the strategy with AI guidance and collect user health data based on actual execution feedback. This data is then used to model the next environmental state for reinforcement learning. Through continuous reinforcement learning interaction and action feedback, a personalized weight management action strategy dynamically adapted to the user is obtained.

[0010] As a weight management plan adjustment method based on multi-source feedback of the present invention, further, it acquires multi-source health data of the user based on the user terminal and wearable device, and preprocesses the multi-source health data, including:

[0011] The system acquires user data such as weight, body fat percentage, resting heart rate, heart rate variability, steps, sleep duration, sleep quality score, food intake, plan execution rate, subjective hunger, and fatigue within the current time period through user terminals and wearable devices.

[0012] The acquired data is time-aligned and outlier-handled to obtain a standardized, structured user health status vector.

[0013] As a weight management scheme adjustment method based on multi-party feedback in this invention, the user's health status vector is further represented as follows: ,in, The user's weight within the current time period t. To show the trend of weight change within a specified time window, This is the resting heart rate. For heart rate variability, For the number of steps, For sleep duration, For dietary intake, To achieve the plan execution rate, These are risk markers, set based on sleep quality, heart rate variability, plan execution rate, weight change trends, chronic disease labels, or special population labels.

[0014] As a weight management program adjustment method based on multi-party feedback according to the present invention, the reward function is further expressed as: ,in, These are the weighting coefficients. Reward users for their weight change trends. Reward users for their execution rate. Rewards for expert feedback As a risk penalty.

[0015] As a weight management scheme adjustment method based on multi-feedback of the present invention, further, by modifying the candidate actions output by the policy network according to the current state through the interaction between the agent and the environment in reinforcement learning and by utilizing expert knowledge and safety constraints, the method includes:

[0016] By using expert rules and historical health data to review samples for behavioral cloning, the policy network can pre-learn the action space decision boundary based on the cloned behavioral data.

[0017] The intelligent agent learns from environmental interactions with the user's individual health status and automatically generates candidate action plans for the current state.

[0018] Experts evaluate candidate action plans, and a pre-trained natural language processing model is used to quantify the expert evaluation results to obtain the expert feedback reward and the structured correction amount of the candidate action plan in the reward function.

[0019] The reward function is used to obtain real-time feedback on actions and update the policy network parameters. The structured correction amount is used to make corrections, and the corrected actions are projected onto the set of weight management safety constraint boundaries to generate the current best action policy when the safety constraint boundaries are met.

[0020] As part of the weight management plan adjustment method based on multi-party feedback of this invention, the method further quantifies the expert evaluation results using a pre-trained natural language processing model, including:

[0021] The expert evaluation results are preprocessed to obtain a text representation in a unified format. The preprocessing includes at least noise removal and text segmentation.

[0022] For text representation, a pre-trained natural language model is used for text classification to obtain the evaluation polarity representing the degree of expert approval of the candidate action plan, the role dimension representing the evaluation direction type of the candidate action plan, and the modification type and taboo entity type representing whether the evaluation direction of the candidate action plan needs to be adjusted. The modification objects corresponding to the modification types and the constraints corresponding to the taboo entity types are obtained by extracting entity objects from the expert evaluation results.

[0023] The reward vector and structured correction amount of candidate action schemes are constructed based on evaluation polarity, action dimension, modification type and modification object, and taboo entity type and constraints.

[0024] As a weight management scheme adjustment method based on multi-party feedback in this invention, the set of safety constraint boundaries includes, but is not limited to: energy deficit range within a period of time, upper limit of weight loss, upper limit of exercise intensity, continuous intensity exercise limit, exercise downgrade rule for sleep deprivation, composite rule for abnormal increase in resting heart rate, and specific user group identifiers.

[0025] Furthermore, this invention also provides a weight management program adjustment system based on multi-party feedback, comprising: a data acquisition module, a reinforcement learning module, and an AI coaching module, wherein...

[0026] The user's health data is acquired from multiple sources based on user terminals and wearable devices, and the data is preprocessed to obtain the user's health status vector.

[0027] The user's health status vector is used as the environment state, and the user's diet and exercise plan is used as the action. The user's weight management plan is dynamically adjusted and modeled as a Markov decision process by combining the reward function. The interaction between the agent and the environment in reinforcement learning and the candidate actions output by the current state of the policy network are modified by using expert knowledge and safety constraints to obtain the best weight management strategy for the user. The reward function consists of weight trend reward, execution rate reward, expert feedback reward and risk penalty. The safety constraints are a pre-set set of weight management safety constraint boundaries.

[0028] The AI ​​coaching module is used to translate the user's current weight management strategy into executable language, enabling the user to execute the current weight management strategy under the AI ​​coaching. It also collects user health data based on the actual execution feedback. The health data based on the actual execution feedback is used to model the next round of environmental state for reinforcement learning. Through continuous reinforcement learning interaction and action feedback, it obtains personalized weight management action strategies that are dynamically adapted to the user.

[0029] The beneficial effects of this invention are:

[0030] This invention confines the reinforcement learning process within expert feedback and metabolic safety constraints, enabling the optimization of long-term weight management strategies using continuous execution results while avoiding the health risks associated with unconstrained exploration. Doctors and other experts can continue to use case-style-free text feedback, utilizing natural language processing models for automatic semantic parsing and quantification mapping to construct expert feedback rewards and structured correction quantities, reducing the workload for doctors. Expert feedback is no longer merely stored as a log but is transformed into reward vectors and policy correction quantities for collaborative model updates and policy generation. The metabolic safety constraint boundary set limits the output range of candidate action plans, reducing the risks of excessive calorie deficit, excessive exercise, and incompatibility with specific populations. AI-assisted training provides explanations, reminders, and psychological support within the final action strategy scope, separating interactive functions from high-risk decision boundaries, forming a closed loop from real-time status, expert feedback, safety projection to user execution feedback, which is beneficial for long-term personalized weight loss management. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the weight management plan adjustment method based on multi-party feedback in the embodiment;

[0032] Figure 2 This is a schematic diagram illustrating the closed-loop principle of the weight management scheme adjustment in the embodiment;

[0033] Figure 3 This is a schematic diagram of the reinforcement learning process in the embodiment;

[0034] Figure 4 This is a schematic diagram of the expert free text feedback quantification process in the embodiment;

[0035] Figure 5 This is a schematic diagram of the metabolic safety constraints and strategy projection process in the embodiment;

[0036] Figure 6 This is a schematic diagram of the closed-loop execution process of the AI ​​tutoring in the embodiment;

[0037] Figure 7 This is a schematic diagram of the dynamic update process of the weight reduction scheme in the embodiment. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer and more understandable, the invention will be further described in detail below with reference to the accompanying drawings and technical solutions.

[0039] To address the shortcomings of existing weight management programs, such as static approaches, unclear boundaries, lack of personalization, and limited room for exploration, this invention provides an embodiment, see [link to embodiment]. Figure 1 As shown, a method for adjusting a weight management plan based on multi-party feedback is provided, including:

[0040] S101. Obtain multi-source health data of users based on user terminals and wearable devices, and preprocess the multi-source health data to obtain the user's health status vector.

[0041] Specifically, user weight, body fat percentage, resting heart rate, heart rate variability, steps, sleep duration, sleep quality score, food intake, plan execution rate, subjective hunger and fatigue within the current time period can be obtained through user terminals and wearable devices; the obtained data is then time-aligned and outlier-processed to obtain a standardized structured user health status vector.

[0042] On day t, data generated by user terminals and wearable devices is received, and time alignment and outlier handling are performed on data from different sources to form a unified state vector. The user's health state vector can be represented as: ,in, The user's weight within the current time period t. To show the trend of weight change within a specified time window, This is the resting heart rate. For heart rate variability, For the number of steps, For sleep duration, For dietary intake, To achieve the plan execution rate, These are risk markers, set based on sleep quality, heart rate variability, plan execution rate, weight change trends, chronic disease labels, or special population labels.

[0043] S102. The user's health status vector is taken as the environment state, and the user's diet and exercise plan is taken as the action. The user's weight management plan is dynamically adjusted and modeled as a Markov decision process in combination with the reward function. The candidate action output by the current state of the policy network is corrected by the interaction between the agent and the environment in reinforcement learning and by using expert knowledge and safety constraints, so as to obtain the best weight management strategy for the user. The reward function consists of weight trend reward, execution rate reward, expert feedback reward and risk penalty. The safety constraints are a pre-set set of weight management safety constraint boundaries.

[0044] Based on state vector Generate candidate actions Candidate actions must include at least the daily total calorie target (cal_t), protein percentage (pro_t), carbohydrate percentage (carb_t), fat percentage (fat_t), exercise type (type_t), exercise duration (dur_t), exercise intensity (int_t), and reminder rhythm (remind_t). A candidate action can be represented as:

[0045]

[0046] The weight loss intervention is implemented using a constrained doctor-in-the-loop Actor-Critic reinforcement learning method, modeling it as a Markov decision process with the user's multi-source health status as the state. The action plan is the diet and exercise plan for the next day. The reward is a combination of weight trend reward, execution rate reward, doctor feedback reward, and risk penalty. State transitions are formed by physiological and behavioral feedback after user execution. Rule-based models or supervised learning models can serve as cold-start or action pre-screening components, but the final candidate actions are generated through a reinforcement learning policy network and released after being quantified by doctor feedback and projected with safety constraints.

[0047] like Figure 2 As shown, the dynamic adjustment process of the weight loss plan is designed as a constrained physician-in-the-loop reinforcement learning process. Each decision cycle can be one day, or it can be set to half a day, three days, or one week depending on the follow-up frequency. The state is acquired at the beginning of the cycle. , by policy network Output candidate actions The candidate actions are jointly corrected by the doctor's feedback quantification module and the safety constraint update module, and the next state is generated after the user executes the action. .

[0048] The state space S may include weight trend, heart rate, heart rate variability, steps, sleep, eating bias, execution rate, subjective hunger, fatigue, and risk markers. The action space A may include calorie goals, macronutrient ratios, exercise type, exercise duration, exercise intensity, and reminder cadence. To reduce exploration risks, the action space is bounded by basic rules before generation and must be processed by a set of constraints after generation. projection.

[0049] The reward function can be expressed as: ,in, These are the weighting coefficients. Reward users for their weight change trends. Reward users for their execution rate. Rewards for expert feedback The penalty is a risk triggered by insufficient sleep, abnormal resting heart rate, excessive weight loss, or chronic disease labels.

[0050] Specifically, by leveraging the interaction between the agent and the environment in reinforcement learning and utilizing expert knowledge and safety constraints to modify the candidate actions output by the policy network based on the current state, it can be designed as follows:

[0051] By using expert rules and historical health data to review samples for behavioral cloning, the policy network can pre-learn the action space decision boundary based on the cloned behavioral data.

[0052] The intelligent agent learns from environmental interactions with the user's individual health status and automatically generates candidate action plans for the current state.

[0053] Experts evaluate candidate action plans, and a pre-trained natural language processing model is used to quantify the expert evaluation results to obtain the expert feedback reward and the structured correction amount of the candidate action plan in the reward function.

[0054] The reward function is used to obtain real-time feedback on actions and update the policy network parameters. The structured correction amount is used to make corrections, and the corrected actions are projected onto the set of weight management safety constraint boundaries to generate the current best action policy when the safety constraint boundaries are met.

[0055] like Figure 3 As shown, during training, the current state, candidate actions, doctor's original text, expert reward, structured correction amount, final solution, execution result and next state are stored together in the training sample cache. In the early stage, behavior cloning can be performed using expert rules and historical review samples, so that the policy network can learn the basic medical decision boundary first. During the runtime, the Actor and Critic parameters are updated based on the advantage function or time difference error.

[0056] The key to this reinforcement learning process is not limiting itself to a specific algorithm name, but rather forming a closed loop of "policy network candidate generation - doctor text feedback quantification - security constraint projection - execution result feedback - value assessment update". Even if PPO, A2C, DDPG or other policy optimization algorithms are used, as long as the above-mentioned constrained doctor-in-the-loop update process is retained, they are all equivalent implementations of this embodiment.

[0057] Experts present candidate solutions to experts. The system displays the user's recent status trends, performance deviations, and risk indicators identified by the system. Experts do not need to input structured scores; they can input free text according to their medical record or follow-up record habits, such as: "Weight trend is acceptable, but dinner contains too much staple food. Maintain aerobic exercise, but it is not recommended to increase the intensity."

[0058] The expert evaluation results can be preprocessed to obtain a standardized text representation. This preprocessing includes at least noise removal and text segmentation, such as filtering emojis, meaningless typos, and stop words; standardizing professional terminology, for example, standardizing "dropping weight" as "weight loss" and "lifting weights" as "strength training"; and semantically dividing long texts into multiple words to avoid semantic interference. For the text representation, a pre-trained natural language model is used for text classification to obtain evaluation polarities such as positive / neutral / negative, representing the expert's level of approval of the candidate action plan. For example, "This plan is easy to stick to" will output a positive evaluation polarity, while "The plan is a bit difficult" will output a negative evaluation polarity. Text classification also yields dimensions representing the expert's evaluation direction of the candidate action plan, such as the impact of diet / exercise / sleep / health. Multiple dimensions can be represented simultaneously in the evaluation text; for example, "Exercise will cause insomnia at night" will output two impact dimensions: exercise and sleep. By performing text classification, we obtain the modification types and taboo entity types that indicate whether the expert's evaluation direction of the candidate action plan needs adjustment. We then extract the entity objects from the expert evaluation results to obtain the modification objects corresponding to the modification types and the constraints corresponding to the taboo entity types. For example, "reduce carbohydrates and replace exercise with walking" has modification types of reduction and replacement, and modification objects of "carbohydrates" and "exercise." The resulting reverse modification can be represented as "reduce carbohydrates, replace exercise type." Similarly, "old knee injury prevents running" has a taboo entity type of "exercise taboo," and the corresponding taboo constraint is "running." Based on evaluation polarity, effect dimension, modification type and modification object, and taboo entity type and constraints, we construct the reward vector and structured correction quantity for the candidate action plan.

[0059] like Figure 4 As shown, by using a natural language processing module to perform hierarchical parsing of doctors' text, at least the following four types of intermediate results are obtained: evaluation polarity, modification direction, effect dimension, and risk flag. Evaluation polarity is used to indicate the doctor's overall level of acceptance of the candidate treatment plan; modification direction is used to indicate addition, reduction, maintenance, prohibition, or review; effect dimension is used to indicate diet, exercise, sleep, lifestyle, or risk control; risk flag is used to indicate contraindications, upper limits, lower limits, or conditions that require manual review.

[0060] The parsing results are mapped to a reward vector. and structured correction amount ,in It can be represented as:

[0061]

[0062] Expert feedback rewards can be calculated in the following manner:

[0063]

[0064] in, This indicates the consistency between the candidate solution and historical doctor opinions and user feedback. This calculation does not replace expert judgment, but rather transforms expert language into an intermediate representation that can be used for policy updates, log tracing, and model training.

[0065] like Figure 5 As shown, metabolic safety constraints are implemented before candidate solutions are released. Weight change is predicted based on user basal metabolic rate, activity expenditure, and intake estimates, using the following basic formula:

[0066]

[0067]

[0068] in, Indicates energy intake, Indicates total energy consumption. This represents the basal metabolic rate. Indicates energy expenditure during exercise. This indicates energy expenditure not related to physical activity. This is the conversion factor between energy and weight changes.

[0069] Safety constraint set The constraints include, but are not limited to, the daily net energy deficit range, the weekly weight loss limit, the exercise intensity limit, the limit on continuous high-intensity exercise, the exercise downgrade rules when sleep is insufficient, the review rules when the resting heart rate rises abnormally, and special templates for people with chronic diseases or special populations.

[0070] The final solution can be generated as follows:

[0071]

[0072] in, This means projecting the expert-corrected actions onto a set of safe executables. When the correction results exceed safety boundaries, they should be pruned, rolled back, or trigger a doctor's review, instead of being released directly to the user.

[0073] S103. Transform the user's current weight management strategy into a user-executable language so that the user can execute the current weight management strategy under the guidance of AI coaching and collect user health data based on the actual execution feedback. Use the health data based on the actual execution feedback to model the environment state for the next round of reinforcement learning. Through continuous reinforcement learning interaction and action feedback, obtain personalized weight management action strategies that are dynamically adapted to the user.

[0074] like Figure 6 As shown, the final solution for AI tutoring is received. The system integrates user status, historical execution records, and risk warnings, translating technical strategies into user-understandable natural language explanations. Interactive content includes plan explanations, dietary reminders, exercise reminders, psychological support, abnormal inquiry functions, and execution confirmation.

[0075] In one implementation, the AI ​​tutor only provides explanations and guidance, without directly bypassing expert feedback quantification and safety constraints to generate high-risk decisions. User feedback on the completion of the plan, subjective feedback, and any anomalies are fed back and used for behavioral evaluation, forming the state vector for the next round. .

[0076] Furthermore, based on the above method, this embodiment of the invention also provides a weight management program adjustment system based on multi-party feedback, comprising: a data acquisition module, a reinforcement learning module, and an AI coaching module, wherein,

[0077] The user's health data is acquired from multiple sources based on user terminals and wearable devices, and the data is preprocessed to obtain the user's health status vector.

[0078] The user's health status vector is used as the environment state, and the user's diet and exercise plan is used as the action. The user's weight management plan is dynamically adjusted and modeled as a Markov decision process by combining the reward function. The interaction between the agent and the environment in reinforcement learning and the candidate actions output by the current state of the policy network are modified by using expert knowledge and safety constraints to obtain the best weight management strategy for the user. The reward function consists of weight trend reward, execution rate reward, expert feedback reward and risk penalty. The safety constraints are a pre-set set of weight management safety constraint boundaries.

[0079] The AI ​​coaching module is used to translate the user's current weight management strategy into executable language, enabling the user to execute the current weight management strategy under the AI ​​coaching. It also collects user health data based on the actual execution feedback. The health data based on the actual execution feedback is used to model the next round of environmental state for reinforcement learning. Through continuous reinforcement learning interaction and action feedback, it obtains personalized weight management action strategies that are dynamically adapted to the user.

[0080] like Figure 7 As shown, multi-source state data of the user on day t is obtained, including weight, heart rate, steps, sleep, food intake, plan execution rate, and subjective feedback. The multi-source state data is cleaned, time-aligned, and outlier-marked to form a state vector. and risk indicators .

[0081] For example, if user A uploads their weight, steps, sleep, and diet records on day t, it can be calculated that: their weight has decreased by 0.45 kg in the past 7 days, they took 7600 steps the previous day, slept for 6.2 hours, their diet adherence rate was 78%, and their carbohydrate intake at dinner was higher than their target. A state vector is then generated. "High carbohydrate intake at dinner" and "slightly insufficient sleep" were marked as observation items.

[0082] according to Output the candidate actions for day t+1. This includes calorie goals, nutrient ratios, exercise type, exercise duration, exercise intensity, and alert pacing. The expert panel displays candidate exercises. It also includes corresponding user trends, execution deviations, and risk warnings, and provides expert input of medical record-style free text feedback.

[0083] For example, a candidate exercise plan for day t+1: total calories 1650 kcal, carbohydrates 42%, protein 28%, fat 30%, brisk walking for 40 minutes, moderate to low intensity, with a reminder after dinner. After the expert presents this candidate plan, the expert inputs: "The weight trend is acceptable, but dinner contains too much staple food. Maintain aerobic exercise, but do not increase the intensity."

[0084] By analyzing expert texts, we extract evaluation polarity, modification direction, impact dimensions, and taboos or upper limit constraints. The analysis results are then mapped to reward vectors. and structured correction amount It also preserves the correspondence between the doctor's original text, the analysis results, and the candidate solutions.

[0085] For example, the analysis reveals: the evaluation polarity is neutral to positive; the dietary dimension modification direction is to reduce carbohydrates at dinner; the exercise dimension is to maintain; and the risk dimension is to not increase intensity. Based on the analysis results, the following can be generated: = -0.4, =0, = -0.2, and form a correction amount. = [-120 kcal, 0, -8% carb, 0, 0, 0, 0].

[0086] Based on metabolic models and constraint rules, Project to safe set Generate the final solution The final solution is translated into a user-executable language, outputting plan explanations, behavioral reminders, and psychological support. User execution results and anomaly feedback are collected, the state vector is updated, and the process enters the next round of dynamic adjustment.

[0087] The above case illustrates that, through analysis and calculation, the revised plan remains within the safe range of daily net energy deficit and exercise intensity, and the final plan is output: total calories 1530 kcal, appropriately reducing the proportion of staple food at dinner, maintaining a brisk walk for 40 minutes, and no additional training. The AI ​​coach can explain to the user: "Tomorrow, reduce the staple food at dinner by about one-third, maintain a brisk walk for 40 minutes, and you are slightly sleep-deprived today, so no extra training is needed."

[0088] If user B's resting heart rate is higher than their baseline for three consecutive days and their sleep duration is less than 5 hours, even if the model generates candidate actions to increase exercise duration, the safety constraints will be updated according to... Triggering the exercise downgrade rule will revert high-intensity training to low-intensity walking or rest recommendations, requiring review by a physician or specialist. The physician or specialist can input "Insufficient recent recovery, suspend high-intensity training, prioritize sleep," which will be interpreted as prohibiting further increases in exercise intensity, increasing the priority of sleep, and triggering a risk review flag.

[0089] The training samples consist of each candidate solution, the original text from doctors and experts, the parsing results, the reward vector, the correction amount, the final solution, and the user's execution result. Initially, a full review by doctors and experts can be used to accumulate samples. Once the model is stable, low-risk solutions can be sampled for review, while high-risk solutions or those with abnormal indicators can be subject to mandatory review. Model updates can be performed periodically or after a preset sample size has been accumulated.

[0090] Without altering the core framework of this invention, candidate strategy generation can be implemented by a rule engine, a supervised learning model, a reinforcement learning model, or a combination of multiple models; the natural language processing module can be implemented by a classification model, a sequence labeling model, a large language model, or a combination thereof; metabolic safety constraints can be configured with different templates according to different population groups, chronic disease labels, or medical institution standards; and the AI ​​tutoring module can be deployed on cloud servers, mobile devices, or edge devices.

[0091] None of the above alternative methods affect the core mechanism of this invention, which transforms free text feedback from doctors and experts into structured rewards and corrections, and generates a final weight loss plan under metabolic safety constraints.

[0092] Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0093] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0094] The units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations are not considered to be beyond the scope of this invention.

[0095] Those skilled in the art will understand that all or part of the steps in the above methods can be implemented by a program instructing related hardware, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk. Optionally, all or part of the steps in the above embodiments can also be implemented using one or more integrated circuits. Accordingly, each module / unit in the above embodiments can be implemented in hardware or as a software functional module. This invention is not limited to any particular combination of hardware and software.

[0096] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for adjusting a weight management plan based on multi-party feedback, characterized in that, Include: The user's health data is acquired from multiple sources based on user terminals and wearable devices, and the data is preprocessed to obtain the user's health status vector. The user's health status vector is used as the environment state, and the user's diet and exercise plan is used as the action. The user's weight management plan is dynamically adjusted and modeled as a Markov decision process by combining the reward function. The interaction between the agent and the environment in reinforcement learning and the candidate actions output by the current state of the policy network are modified by using expert knowledge and safety constraints to obtain the best weight management strategy for the user. The reward function consists of weight trend reward, execution rate reward, expert feedback reward and risk penalty. The safety constraints are a pre-set set of weight management safety constraint boundaries. The system translates the user's current weight management strategy into an executable language, enabling the user to execute the strategy with AI guidance and collect user health data based on actual execution feedback. This data is then used to model the next environmental state for reinforcement learning. Through continuous reinforcement learning interaction and action feedback, a personalized weight management action strategy dynamically adapted to the user is obtained.

2. The method for adjusting a weight management plan based on multi-party feedback according to claim 1, characterized in that, This involves acquiring multi-source health data from users via user terminals and wearable devices, and preprocessing this data, including: The system acquires user data such as weight, body fat percentage, resting heart rate, heart rate variability, steps, sleep duration, sleep quality score, food intake, plan execution rate, subjective hunger, and fatigue within the current time period through user terminals and wearable devices. The acquired data is time-aligned and outlier-handled to obtain a standardized, structured user health status vector.

3. The method for adjusting a weight management plan based on multi-party feedback according to claim 1 or 2, characterized in that, The user health status vector is represented as follows: ,in, The user's weight within the current time period t. To show the trend of weight change within a specified time window, This is the resting heart rate. For heart rate variability, For the number of steps, For sleep duration, For dietary intake, To achieve the plan execution rate, These are risk markers, set based on sleep quality, heart rate variability, plan execution rate, weight change trends, chronic disease labels, or special population labels.

4. The method for adjusting a weight management plan based on multi-party feedback according to claim 1, characterized in that, The reward function is expressed as follows: ,in, These are the weighting coefficients. Reward users for their weight change trends. Reward users for their execution rate. Rewards for expert feedback As a risk penalty.

5. The method for adjusting a weight management plan based on multi-party feedback according to claim 1 or 4, characterized in that, By leveraging the interaction between the agent and the environment in reinforcement learning and utilizing expert knowledge and safety constraints to refine the candidate actions output by the policy network based on the current state, the following methods are employed: By using expert rules and historical health data to review samples for behavioral cloning, the policy network can pre-learn the action space decision boundary based on the cloned behavioral data. The intelligent agent learns from environmental interactions with the user's individual health status and automatically generates candidate action plans for the current state. Experts evaluate candidate action plans, and a pre-trained natural language processing model is used to quantify the expert evaluation results to obtain the expert feedback reward and the structured correction amount of the candidate action plan in the reward function. The reward function is used to obtain real-time feedback on actions and update the policy network parameters. The structured correction amount is used to make corrections, and the corrected actions are projected onto the set of weight management safety constraint boundaries to generate the current best action policy when the safety constraint boundaries are met.

6. The method for adjusting a weight management plan based on multi-party feedback according to claim 5, characterized in that, The results of expert evaluations are quantified using a pre-trained natural language processing model, including: The expert evaluation results are preprocessed to obtain a text representation in a unified format. The preprocessing includes at least noise removal and text segmentation. For text representation, a pre-trained natural language model is used for text classification to obtain the evaluation polarity representing the degree of expert approval of the candidate action plan, the role dimension representing the evaluation direction type of the candidate action plan, and the modification type and taboo entity type representing whether the evaluation direction of the candidate action plan needs to be adjusted. The modification objects corresponding to the modification types and the constraints corresponding to the taboo entity types are obtained by extracting entity objects from the expert evaluation results. The reward vector and structured correction amount of candidate action schemes are constructed based on evaluation polarity, action dimension, modification type and modification object, and taboo entity type and constraints.

7. The method for adjusting a weight management plan based on multi-party feedback according to claim 5, characterized in that, The set of safety constraint boundaries includes, but is not limited to: energy deficit range within a period of time, upper limit of weight loss, upper limit of exercise intensity, continuous intensity exercise limit, exercise downgrade rule for sleep deprivation, composite rule for abnormal increase in resting heart rate, and specific user group identifiers.

8. A weight management program adjustment system based on multi-party feedback, characterized in that, It includes: a data acquisition module, a reinforcement learning module, and an AI tutoring module. The user's health data is acquired from multiple sources based on user terminals and wearable devices, and the data is preprocessed to obtain the user's health status vector. The user's health status vector is used as the environment state, and the user's diet and exercise plan is used as the action. The user's weight management plan is dynamically adjusted and modeled as a Markov decision process by combining the reward function. The interaction between the agent and the environment in reinforcement learning and the candidate actions output by the current state of the policy network are modified by using expert knowledge and safety constraints to obtain the best weight management strategy for the user. The reward function consists of weight trend reward, execution rate reward, expert feedback reward and risk penalty. The safety constraints are a pre-set set of weight management safety constraint boundaries. The AI ​​coaching module is used to translate the user's current weight management strategy into executable language, enabling the user to execute the current weight management strategy under the AI ​​coaching. It also collects user health data based on the actual execution feedback. The health data based on the actual execution feedback is used to model the next round of environmental state for reinforcement learning. Through continuous reinforcement learning interaction and action feedback, it obtains personalized weight management action strategies that are dynamically adapted to the user.

9. An electronic device, characterized in that, include: At least one processor, and a memory coupled to said at least one processor; The memory stores a computer program that can be executed by the at least one processor to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed, enables the implementation of the method as described in any one of claims 1 to 7.

Images