A method, apparatus, equipment, and medium for dynamic optimization of health plans based on a large model.

Abstract

This application relates to a method, apparatus, device, and medium for dynamic optimization of health plans based on a large-scale model. The method includes: acquiring user-set health goals, user historical behavior data, and self-psychological assessment scores; the user historical behavior data includes historical health plan executions, historical plan execution status, and user interaction logs; constructing a current user personality vector based on the user's historical behavior data and self-psychological assessment score; constructing a set of failed personality vectors based on historical health plan executions and historical plan execution status; based on the health goals, conducting an intervention path game driven by personality adversarial interaction based on the current user personality vector and the set of failed personality vectors to obtain a health plan intervention path; the health plan intervention path includes behavioral suggestions, corresponding time period constraints, and behavioral contextual considerations. This method integrates personality modeling and adversarial game theory to optimize the personalization and resilience of health plan paths.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and in particular relates to a method, apparatus, device and medium for dynamic optimization of health plans based on a large model. Background Technology

[0002] With the development of health management technology, health plan development technology has emerged. Traditional technologies employ rule-based template-based health plan recommendation methods, which match corresponding health intervention paths based on the basic information provided by the user by pre-setting several health goal templates. Alternatively, intelligent development methods such as machine learning models and knowledge graphs can be used to generate personalized health suggestions and optimize health plans by analyzing user behavioral data.

[0003] The aforementioned methods lack modeling of user behavior failure mechanisms, attributing unmet user goals to insufficient execution or environmental interference, and lacking mechanisms for extracting, identifying, and counteracting failure modes. They also lack the ability to simulate future user behavior deviations; traditional technologies primarily rely on historical data regression to predict user states, lacking reasoning mechanisms for the evolution of user plan trajectories, making it difficult to capture potential deviation trends during plan execution in a timely manner. Furthermore, the rigid structure of health goals lacks variability and adaptability. In real-world scenarios, user goals may require dynamic adjustments due to execution pressure, changes in health status, or psychological factors. Existing technologies lack support for automatic generation mechanisms for goal evolution or alternative goals, leading to the easy failure or interruption of health plans. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, device, and medium for dynamic optimization of health plans based on a large model, which can perform in-depth modeling of user behavior and adversarial reasoning, to address the above-mentioned technical problems.

[0005] Firstly, this application provides a dynamic optimization method for health plans based on a large model, including:

[0006] Acquire user-defined health goals, user historical behavior data, and self-psychological assessment scores; user historical behavior data includes historical execution of health plans, historical plan execution status, and user interaction logs.

[0007] Construct a current user's personality vector based on the user's historical behavior data and self-psychological assessment score;

[0008] Construct a set of failed personality vectors based on historical health plan implementation and historical plan execution status;

[0009] Based on health goals, an intervention path game driven by personality conflict is conducted according to the current user's personality vector and the set of failed personality vectors to obtain the health plan intervention path; the health plan intervention path includes behavioral suggestions as well as corresponding time period constraints and behavioral context precautions.

[0010] In one embodiment, a set of failed personality vectors is constructed based on historical health plan execution and historical plan execution status, including:

[0011] Extract historical execution health plans and reasons for failures from the historical plan execution data;

[0012] Clustering modeling was performed based on multiple historical execution failures of health plans and the reasons for failure to obtain multiple potential failure personality models;

[0013] By conducting multi-round dialogue simulations on multiple potential failure personality models to create personality profiles, and extracting the failure latent variable representations for each personality profile, a set of failure personality vectors is obtained.

[0014] In one embodiment, based on health goals, an intervention path game driven by personality adversarial dynamics is performed using the current user's personality vector and the set of failed personality vectors to obtain a health plan intervention path, including:

[0015] Based on users' daily temporal behavior templates and external environmental constraints, health goals are semantically vectorized to obtain a set of target subtasks and constraint dimensions; the constraints include external environmental constraints.

[0016] Based on the current user personality vector and the set of failed personality vectors, the user id, the failed personality, and the health intervention coach are used as roles to engage in adversarial game, and multiple initial health plans are generated according to the target sub-task set and constraint dimensions.

[0017] Based on the evaluation function, the stress resistance score of each behavioral suggestion in each initial health plan is calculated for the set of failed personality vectors, resulting in a set of stress resistance scores; the set of stress resistance scores includes each initial health plan and its corresponding score sequence;

[0018] The initial health plan corresponding to the highest comprehensive score in the stress resilience score set is identified as the health plan intervention path.

[0019] In one embodiment, based on an evaluation function, a resilience score is calculated for each behavioral suggestion in each initial health plan for the set of failed personality vectors, resulting in a set of resilience scores, including:

[0020] The compressive strength score is obtained using the following formula:

[0021] ;

[0022] in, For the first in the initial health plan Recommendations for specific behaviors; For the first A failed personality vector; The behavioral semantic encoding vector for behavioral suggestions; Cosine similarity of vectors; External environmental constraints; The degree of contextual fit between the behavioral recommendations and the current external environmental constraints; This is the fitness weighting coefficient.

[0023] In one embodiment, the method further includes:

[0024] In response to the obtained actual user execution data, the actual user execution data is compared with the health plan intervention path to obtain the deviation sequence;

[0025] Based on a preset long time window, the deviation sequence is extracted to obtain continuous deviation values;

[0026] If the continuous deviation value exceeds the preset deviation threshold, the health plan intervention path will be reconstructed by replacing the target group.

[0027] In one embodiment, the deviation is compared between the user's actual execution data and the health plan intervention path to obtain a deviation sequence, including:

[0028] Generate an ideal digital avatar for the user based on the health plan intervention path;

[0029] By using the user's ideal digital avatar to simulate the user's daily behavior path based on the user's schedule data, social events, and daily workload, the ideal behavior trajectory can be obtained;

[0030] Transform users' actual execution data into structured actual behavior trajectories;

[0031] The deviation sequence is obtained by calculating the distance between the behavior codes of the actual behavior trajectory and the ideal behavior trajectory at each time period.

[0032] In one embodiment, if the continuous deviation value exceeds a preset deviation threshold, the health plan intervention path is reconstructed by replacing the target group, including:

[0033] Based on the challenge value, willingness-to-benefit ratio, and semantic consistency of the initial health goal, a set of alternative goals is generated by multi-dimensional alternative goals according to the goal sub-task set and constraint dimensions.

[0034] Determine the reasons for the failure of health plan intervention pathways based on continuous deviation values;

[0035] Determine the optimal alternative target from the set of alternative targets based on the reasons for failure;

[0036] Health plan intervention pathways are generated based on the optimal alternative goals as new health goals.

[0037] Secondly, this application also provides a dynamic optimization device for health plans based on a large model, comprising:

[0038] The data acquisition module is used to acquire user-set health goals, user historical behavior data, and self-psychological assessment scores; user historical behavior data includes historical health plan execution, historical plan execution status, and user interaction logs.

[0039] The large-scale user personality construction module is used to construct the current user's personality vector based on the user's historical behavioral data and self-psychological assessment scores.

[0040] The large-scale model failure personality simulation module is used to construct a set of failure personality vectors based on historical health plan execution and historical plan execution performance.

[0041] The health plan planning module is used to obtain the health plan intervention path by engaging in a personality-driven intervention path game based on health goals, current user personality vectors, and the set of failed personality vectors.

[0042] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above-described dynamic optimization methods for health plans based on large models.

[0043] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above-described dynamic optimization methods for health plans based on large models.

[0044] The aforementioned method, apparatus, equipment, and media for dynamic optimization of health plans based on a large-scale model, through personalized personality vectors constructed from users' historical behavior and psychological self-assessment, can comprehensively characterize users' behavioral preferences, motivational intensity, and emotional susceptibility, achieving precise adaptation of intervention strategies. The introduction of a failure personality adversarial mechanism optimizes path design from a failure prior perspective, effectively improving plan robustness and intervention success rate. Through a path game model constrained by health goals and the personality adversarial mechanism, the optimal intervention path among multiple paths is dynamically selected and optimized, exhibiting stronger adaptability and stability. The output includes complete path results encompassing behavioral suggestions, time periods, and contextual attention, making it highly operable, with low barriers to user understanding and execution, thus improving the actual conversion rate of health plans. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart illustrating the dynamic optimization method for health plans based on a large model according to the present invention.

[0047] Figure 2 This is a flowchart illustrating the steps of step S103.

[0048] Figure 3 This is a flowchart illustrating the steps of step S104.

[0049] Figure 4 This is a structural diagram of the health plan dynamic optimization device based on a large model according to the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0051] In one embodiment, such as Figure 1 As shown, a dynamic optimization method for health plans based on a large model is provided. This embodiment illustrates the method by applying it to a terminal. It is understood that this method can also be applied to a server, and to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0052] S101. Obtain the user's set health goals, user historical behavior data, and self-psychological assessment score; user historical behavior data includes historical execution of health plans, historical plan execution status, and user interaction logs.

[0053] As an example, users can set current health goals, either proactively or passively. Health goals refer to indicators of physical, behavioral, or psychological improvement that users hope to achieve within a specific time period. Examples include completing 30 minutes of moderate-intensity exercise daily for one month, reducing bedtime phone usage to less than 30 minutes for two weeks, or maintaining a continuous meditation record for seven days.

[0054] Furthermore, the system simultaneously acquires users' historical behavioral data, which can come from records of users' past health plan execution trajectories. This includes, but is not limited to, health plan items that users set and attempted to execute, specific execution records and completion status of each plan, as well as interaction logs formed during the interaction between users and the big model. In particular, it refers to the textual interactions between users and the health big model, such as multiple rounds of question-and-answer sessions, suggestions and feedback, and goal adjustments, which can reveal key personality clues such as users' acceptance of the plan, resistance psychology, cognitive biases, and motivations for goal changes.

[0055] Optionally, the self-psychological assessment scores filled in or participated in by users can be obtained simultaneously. These scores can be derived from standardized psychological measurement questionnaires or mental health indicators generated by large model inference, serving as auxiliary signals for personality state or psychological resilience modeling.

[0056] S102. Construct the current user's personality vector based on the user's historical behavior data and self-psychological assessment score.

[0057] This example illustrates how large language models with contextual understanding capabilities, such as the ChatGPT (Chat Generative Pre-trained Transformer) series or Claude, can be used. These models leverage their semantic understanding, contextual induction, and psychological tendency abstraction abilities to vectorize long-term user behavior data and interaction logs, thereby constructing personality vectors describing the user's current behavioral and psychological patterns. Specifically, by interpreting the context of user interaction logs, the large model identifies the user's typical communication style, motivational statements, emotional expressions, and behavioral intentions. Furthermore, by combining historical behavioral data, the model identifies behavioral patterns such as behavioral delays, execution interruptions, and frequent goal changes, extracting user behavioral tendency dimensions, including executive ability, self-efficacy, and procrastination preferences, as well as the user's emotion regulation abilities, such as anxiety sensitivity and frequency of mood swings.

[0058] Optionally, to quantify the user's personality state, a set of personality semantic vector bases is introduced. This is achieved through a high-dimensional psychobehavioral embedding space based on the Big-Five model. In this space, each personality factor is quantified using scores generated by the large model, forming the user's current personality vector. ,in The personality vector dimension, where vector values ​​reflect the intensity of personality traits on that dimension, categorizes personality states into extraversion, agreeableness, conscientiousness, emotional stability, and openness. For example, if a user's behavioral data shows multiple instances of interrupted execution but expressions of guilt, the larger model might determine that the user exhibits weak expression in the conscientiousness dimension but high emotional sensitivity.

[0059] S103. Construct a set of failed personality vectors based on the historical implementation of health plans and the historical implementation status of these plans.

[0060] Specifically, a series of task nodes marked as failed are selected from historical health plan execution records. Failure nodes refer to planned events where the user failed to consistently execute them or achieved very low goal attainment. For each failure node, the interaction logs and psychological assessment values ​​corresponding to the time of occurrence are extracted and input into a large model to retrospectively analyze the psychological state characteristics at that time, thereby obtaining a personality vector representation of the failed behavior. In order to construct a set of failed personality vectors ,in This indicates the number of past aggregation failure types. For example, if a user frequently fails in a diet control plan and exhibits avoidance strategies, self-abandonment, and adversarial language in their interaction logs, then this failure personality vector will highlight characteristics such as low self-discipline and high resistance.

[0061] S104. Based on health goals, an intervention path game driven by personality conflict is conducted according to the current user's personality vector and the set of failed personality vectors to obtain the health plan intervention path; the health plan intervention path includes behavioral suggestions as well as corresponding time period constraints and behavioral context precautions.

[0062] Indicatively, the personality adversarial game mechanism is used to simulate the potential adversarial and slip paths between the current user's personality vector and the set of failed personality vectors. Through game evolution, it automatically generates a health plan intervention path with the highest psychological acceptability and lowest failure risk. Specifically, the health goal is considered the objective function, the personality vector serves as the psychological weights for the execution strategy, and the intervention path is the strategy sequence. This is achieved by minimizing the following risk-efficacy loss function. ,in, As candidate intervention pathways, For the behavioral cost function under the health goal, This indicates the probability that the path will slide into a failed personality zone. Through virtual reasoning using a large model, the interaction between different paths and the user's psychological state is simulated, generating multiple intervention strategy branches, such as gradual incentive-based, constraint-driven, and emotion-guided approaches, and assessing their stability within the personality space. Ultimately, the path with the highest stability, lowest failure risk, and optimal psychological acceptability is selected as the intervention path for the current period's health plan. This path includes clear behavioral recommendations, time constraints for the recommended behaviors, and implementation prompts or contextual considerations for specific personality traits.

[0063] The aforementioned dynamic optimization method for health plans based on a large model enhances the accuracy of characterizing users' execution habits and psychological states by acquiring historical user behavior data and self-psychological assessment scores, providing a personalized input foundation for path planning. Constructing a current user personality vector enables quantitative modeling of user behavior patterns, motivational stability, and cognitive responses, facilitating the development of more behaviorally aligned intervention paths. Building a set of failure personality vectors, using historical failure patterns as a reference, identifies potential failure-prone paths and weak points, improving predictive and avoidance capabilities. A personality-driven intervention path game based on health goals generates more robust and resilient intervention paths, considering risky personalities, thus increasing the plan's success rate. The intervention path includes behavioral suggestions, time constraints, and contextual considerations, providing specific and actionable recommendations, clarifying time and environmental requirements, and enhancing user feasibility and sustainability.

[0064] In one embodiment, such as Figure 2 As shown, a set of failed personality vectors is constructed based on the historical implementation of health plans and the historical plan implementation status, including:

[0065] S201. Extract historical execution health plans and reasons for failure from the historical plan execution status.

[0066] Specifically, from the historical health plan database, samples of historically executed health plans marked as failed or partially failed are selected, and the corresponding failure reason information is extracted. These failure reasons include, but are not limited to, emotional factors mentioned in the user's subjective feedback text, objective indicators such as time deviation, low execution frequency, and task skipping recorded by the behavior monitoring module, and quantitative assessment items such as self-awareness reports or motivation scores collected by the questionnaire system. For each failed sample, structured data consisting of health plan content, execution trajectory, failure behavior characteristics, failure complaint text, and failure label is constructed to support subsequent cluster analysis and personality profile modeling.

[0067] S202. Cluster modeling is performed based on the historical execution health plans and reasons for failures to obtain multiple potential failure personality models.

[0068] This study illustrates how to semantically encode the content of a health plan, extracting task types such as exercise, diet, and sleep patterns, as well as structural features like task intensity and target time windows. It then uses multimodal feature fusion encoding to address failure reasons, including embedding the complaint text into a large model, behavioral frequency vectors, and failure type encoding. Furthermore, a clustering algorithm is used to semantically cluster the multimodal failure samples, resulting in multiple cluster centers. Each cluster center represents a potential failure personality model, characterizing typical failure paths under specific behavioral motivations, such as high planning enthusiasm but low execution stability, social distraction, and short-term high-pressure abandonment. Each personality model exhibits strong semantic consistency and structural similarity in failure behaviors.

[0069] S203. By conducting multi-round dialogue simulations on multiple potential failure personality models, and extracting the failure latent variable representations for each type of personality model, a set of failure personality vectors is obtained.

[0070] Furthermore, a large-scale model is invoked to simulate typical personality profiles, generating their behavioral and feedback semantics in various health scenarios through multi-turn dialogues to enhance the completeness of their semantic expression. Specifically, personality labels, typical failure scenarios, and failure behavior trajectories are input into the large-scale model. A simulated dialogue generator is then activated to generate dialogue records between the user and a virtual health assistant or system guide. The simulated dialogues showcase the behavioral feedback, procrastination speech, and emotional expression of this personality type in different health tasks. The dialogue text is vectorized and combined with behavioral trajectory encoding and failure mode labels to construct a multimodal failure latent variable representation for each personality type, forming a high-dimensional vector representation. The generated set of failure personality vectors includes multiple vectors representing typical failure behavior patterns. Each vector possesses stable semantic cohesion and individual behavior inference capabilities, which can serve as a reference item for adversarial roles in the subsequent intervention path generation.

[0071] In one embodiment, such as Figure 3 As shown, based on health goals, and using the current user's personality vector and the set of failed personality vectors, an intervention path game driven by personality adversarial dynamics is conducted to obtain the health plan intervention path, including:

[0072] S301. Based on the user's daily time-series behavior template and external environmental constraints, semantic vectorize the health goals to obtain a set of target sub-tasks and constraint dimensions; the constraints include external environmental constraints.

[0073] In a schematic manner, health goals are used as input semantic content. By calling a large language model, the goal structure is decomposed and vectorized to extract key sub-tasks and construct their logical order and dependencies. At the same time, combined with the user's daily temporal behavior templates, such as commuting time, family routine, and device availability, as well as external environmental constraints, such as climate, device limitations, social event arrangements, and medical advice, the goal tasks are constrained and mapped to obtain a constraint dimension matrix.

[0074] S302. Based on the current user personality vector and the set of failed personality vectors, an adversarial game is played with the user id, the failed personality, and the health intervention coach as roles, and multiple initial health plans are generated according to the target sub-task set and constraint dimensions.

[0075] Specifically, the personality adversarial game-generating framework employs three virtual agent roles: the Ego Agent, which takes the current user's personality vector as input and reflects their true behavioral preferences and motivational responses; the Failure Agent, which extracts multiple representative vectors from the set of failure personality vectors to simulate potential obstacles and derailment paths; and the Coach Agent, which simulates the ideal path and guidance logic based on a health knowledge base and intervention strategy model. In the adversarial process, each agent engages in path-playing games in the vector semantic space based on target sub-tasks and constraint dimensions. In each iteration, the failure personality attempts to disrupt the planned path, the user's ego attempts to correct or compromise, and the coach guides the goal to be maintained. Several feasible health path plans are generated through reinforcement learning-based dialogue generation and a rule-based game tree, forming multiple candidate initial health plans. Each initial health plan includes structured fields such as sub-task execution time, frequency, form, and dynamic adaptation strategies.

[0076] S303. Based on the evaluation function, calculate the stress resistance score of each behavioral suggestion in each initial health plan for the set of failed personality vectors, and obtain the stress resistance score set; the stress resistance score set includes each initial health plan and its corresponding score sequence.

[0077] Furthermore, a multi-dimensional evaluation function is designed to score the resilience performance between the initial health plan and the failed personality vector. Specifically, the failed personality vector is input into an adversarial simulator and subjected to multiple rounds of interference tests with the health plan path to observe the plan's resilience; the breakdown points under critical values ​​of parameters such as task time, task frequency, and psychological load are identified; and the performance is combined with an execution acceptability function. Path maintainability function Behavioral incentive resistance coefficient The score sequence for each initial health plan is calculated comprehensively. .

[0078] S304. Determine the initial health plan corresponding to the highest comprehensive score in the stress resilience score set as the health plan intervention path.

[0079] Optionally, a comprehensive analysis of the stress resilience score set can be performed. Methods such as weighted averaging, min-maximum balancing, or fuzzy integrals can be used to aggregate the score sequences of each health plan, obtaining a comprehensive stress resilience score for that plan. The health plan with the highest score is selected as the final health plan intervention path generated in the current round. This path not only meets the health goal requirements but also maximizes adaptability and intervention flexibility to failed behaviors, ensuring a high success rate of plan implementation in real-world user environments.

[0080] In one embodiment, based on an evaluation function, a resilience score is calculated for each behavioral suggestion in each initial health plan for the set of failed personality vectors, resulting in a set of resilience scores, including:

[0081] The compressive strength score is obtained using the following formula:

[0082] ;

[0083] in, For the first in the initial health plan Recommendations for specific behaviors; For the first A failed personality vector; The behavioral semantic encoding vector for behavioral suggestions; Cosine similarity of vectors; External environmental constraints; The degree of contextual fit between the behavioral recommendations and the current external environmental constraints; This is the fitness weighting coefficient.

[0084] In one embodiment, the method further includes:

[0085] S41. In response to the obtained actual user execution data, compare the actual user execution data with the health plan intervention path to obtain the deviation sequence.

[0086] As an example, during the health plan execution phase, continuous data feedback is received from the user's terminal or wearable device, including behavior completion time, action frequency, physiological status indicators, interaction response logs, etc., forming a real execution data set. This dataset is compared with the target task parameters in the previously generated health program intervention pathway. Conduct a one-to-one comparison: For each behavioral suggestion, calculate its plan execution deviation, including time difference, frequency difference, intensity difference, etc.

[0087] Constructing the deviation function This yields a sequence of deviation vectors in sequence form. This sequence is used to describe the degree of gradual deviation between the user's execution process and the original path.

[0088] S42. Based on a preset long time window, the deviation sequence is extracted to obtain continuous deviation values.

[0089] To avoid misleading the public about the stability of the health path due to short-term fluctuations, the system is set with a fixed-time window. The bias sequence is locally aggregated using a sliding method; specifically, continuous bias values ​​are extracted within each window. Based on the set statistical rules such as mean, maximum value, and trend slope, the continuous deviation index value within the window is calculated. Continuous Deviation Index Time Series It is used to assess the deviation trend of user behavior. This mechanism can effectively detect the gradual deviation pattern, that is, the situation where users are close to the execution plan in the early stage but gradually give up or adjust the execution pace.

[0090] S43. If the continuous deviation value exceeds the preset deviation threshold, the health plan intervention path will be reconstructed by replacing the target group.

[0091] When the continuous deviation value within any window If the deviation exceeds the system's preset tolerance threshold, the current health plan is deemed ineffective or has lost its behavioral stickiness, triggering the plan reconstruction logic. Specifically, the system calls upon alternative target groups in the health goal map that are similar to the original goal's functional path but whose execution difficulty or incentive mechanism is closer to the user's current behavioral state. The user's current execution preferences are modeled against the alternative target groups to match a target path with higher behavioral acceptability. While maintaining the consistency of the original health goal framework, the health plan intervention path generation process is restarted based on the alternative goal tasks. This involves re-decomposing the goal semantics, engaging in personality adversarial games, assessing stress tolerance, and selecting the path, ultimately forming a new health plan intervention path.

[0092] In one embodiment, the deviation is compared between the user's actual execution data and the health plan intervention path to obtain a deviation sequence, including:

[0093] S51. Generate an ideal digital clone for the user based on the health plan intervention path.

[0094] Optionally, based on the current health program intervention pathway Based on the target behavior set, timing suggestions, and priority relationships in the data, and combined with the user's basic profile information, an ideal execution behavior agent is constructed, which is the user's ideal digital avatar. This agent can understand the semantic structure of the health plan goals and their sub-tasks, execute the user's possible optimal response behavior, and follow the established constraints.

[0095] S52. Using the user's ideal digital clone, simulate the user's daily behavior path based on the user's schedule data, social events, and daily workload to obtain the ideal behavior trajectory.

[0096] Ideal digital avatars access users' dynamic schedule data streams, including work schedules, social activities, and family responsibilities. By combining these with a rule model of health plan task requirements and behavioral responses, they generate a structured ideal behavioral trajectory. Its form is Each time period Corresponding to an ideal behavior code This indicates the optimal health choice that a user should make in a specific context.

[0097] S53. Transform the user's actual execution data into a structured actual behavior trajectory.

[0098] The raw execution data obtained from the user terminal device is subjected to multimodal behavior recognition and structured coding to generate the user's actual behavior trajectory. in Indicates the time period Actual behavior coding versus ideal behavior coding It maintains dimensional consistency and has semantic alignment capabilities.

[0099] S54. Calculate the distance between the behavior codes of the actual behavior trajectory and the ideal behavior trajectory at each time period to obtain the deviation sequence.

[0100] For each time period Using behavioral coding distance function The semantic deviation between actual behavior and ideal behavior is calculated. The behavior encoding can be represented by multi-dimensional feature embedding, and the distance function can include Euclidean distance, cosine similarity inversion, etc. The deviation values ​​of all time periods are aggregated to obtain a complete deviation sequence. By constructing an ideal digital clone of the user and comparing the encoded behavior trajectory, the semantic and structural level of the user's multi-dimensional behavior deviation can be accurately detected.

[0101] In one embodiment, if the continuous deviation value exceeds a preset deviation threshold, the health plan intervention path is reconstructed by replacing the target group, including:

[0102] S61. Based on the challenge value, willingness integration degree and semantic consistency of the initial health goal, generate multiple alternative goals according to the goal sub-task set and constraint dimension to obtain the alternative goal set.

[0103] This approach illustratively identifies the set of subtasks involved in the current health plan and their key constraint dimensions. Based on these constraints, a set of candidate alternative goals is constructed. The challenge value measures the cognitive, volitional, or time resource input required for the user to complete the goal, and can be assessed using standardized scales. The willingness-to-goal ratio indicates the degree of consistency between the goal and the user's current behavioral preferences and values, and can be modeled based on user profiles and historical task preferences. Semantic consistency refers to the semantic similarity between the original health goal and candidate goals. Through multi-goal ranking and fusion scoring across these three constraint dimensions, combined with task logic constraint filtering, a set of alternative goals that are substitutable and reasonably coupled with the tasks can be obtained.

[0104] S62. Determine the reasons for the failure of the health plan intervention path based on the continuous deviation value.

[0105] Based on pattern recognition and location distribution features of consecutive deviation segments in behavioral deviation sequences, combined with prior failure personality models and target task dependency graphs, the core reasons for failure in the original health plan intervention path are extracted. These include conflicts between the triggering period of a certain behavior and high-pressure work, frequent failures due to the need for cooperation from others in a certain sub-task, and mismatch between the current user state and the pre-set psychological state of the plan. Through attribution mapping and failure path annotation experience, root causes are classified into cognitive, resource-related, emotional, and environmental factors.

[0106] S63. Determine the optimal alternative target from the set of alternative targets based on the reasons for failure.

[0107] Based on the reasons for failure, the optimal alternative target is selected from the set of alternative targets that has the strongest resistance or higher compliance. Specifically, the optimal alternative target is selected from the set of alternative targets that meets the following conditions: it has weak or anti-coupling with the root cause of failure, has a higher probability of success under the failure personality model, and can be naturally embedded within the original path time-resource framework. Optionally, a weighted target scoring function can be used to optimize the ranking by comprehensively considering indicators such as accessibility, compliance, and health return rate.

[0108] S64. Generate a health plan intervention path based on the optimal alternative goal as the new health goal.

[0109] Specifically, with the optimal alternative goal as the core objective, the system calls upon the planning module of a large model or expert system, and combines the current user status, external context, and plan execution history to generate a new health plan intervention path. This path includes the execution strategy and task decomposition of the new goal, a time window that matches the user's original behavioral rhythm, and early warning buffer mechanisms and emotional support measures set for failure trigger points.

[0110] For example, suppose User A is a 30-year-old office worker who has set health goals to improve sleep quality and increase exercise habits. The Big Language Model obtained relevant health data from User A through questionnaires, behavior logs, and historical data collection. This included their historical health plan of going to bed before 11:00 PM every day and engaging in aerobic exercise three times a week. However, the historical plan execution showed that the user only completed the planned exercise tasks three times in the past month, and went to bed after 11:30 PM on several occasions. Furthermore, the interaction logs of User A during their use of the Big Language Model were also collected. In their dialogue with the health assistant within the Big Language Model, User A frequently expressed feelings of high work pressure and difficulty sticking to their plans. At the same time, their self-psychological assessment questionnaire showed a moderate level of behavioral motivation and a high degree of procrastination tendency. Furthermore, the GPT-4-based, health behavior-tuned version of the large language model inputs user A's historical behavioral data, self-assessment score, and interaction logs into the personality modeling submodule. It extracts features such as stable behavioral patterns, emotional response preferences, and execution will stability, and aligns these features with multiple typical user personality vectors summarized in the training set, thereby generating a current personality vector representing user A's individual behavioral tendencies. Simultaneously, the large model utilizes a large number of failed health plan samples from the training library and their semantic context of execution failure to construct a set of failed personality vectors. In this process, the health plan texts, failure feedback, and behavioral decision paths of multiple failed users are clustered, and various failed personality profiles are generated through multi-turn dialogues. For personality patterns where the execution time of a plan is repeatedly postponed, the failed personality vector will reflect a high procrastination factor and low sensitivity to immediate feedback. A personality adversarial mechanism is activated, with the large model engaging in an interactive game between the user's id, the failed personality, and the health coach. Internally, the model constructs multiple sets of possible health intervention paths through a class-based reinforcement learning path generation method. To ensure that the generated intervention path is both feasible and resilient to personality, each path generated by the game is matched and evaluated with a failing personality type. The path is stress-tested using a behavioral perturbation simulation method to obtain a stress resistance score sequence for the path under the influence of personality types such as distractibility and social procrastination. Finally, the path with the highest comprehensive score is selected as the health plan intervention path.

[0111] The large-scale model employed is a pre-trained language model system, such as a general language understanding and generation model based on the Transformer architecture, which has been fine-tuned and enhanced with multi-task training specifically for health behavior modeling tasks. The user data processing stage embedding involves structuring user historical behavior data, self-psychological assessment scores, and user interaction logs using a defined data format template, then feeding them into the large-scale model in a multi-turn dialogue format to stimulate its implicit abstraction ability of user behavior patterns and output quantitative personality characteristics. The failure personality modeling stage embedding involves inputting user text from an existing historical failure sample library into the large-scale model for emotion and behavioral intent extraction, generating personality vectors through vector clustering and mapping behavioral motivation labels, and establishing a failure personality semantic space. The intervention path game generation stage embedding involves the large-scale model initiating a game process involving a "self-other-coach" triangular role after the target task is input. It explores multiple path solutions through a policy generation network and outputs semantic descriptions of path generation and behavioral sequence suggestions. By utilizing the large-scale model's simulation reasoning capabilities, an interaction script between the failure personality vector and the intervention path is constructed under typical execution scenarios, and a stress resistance score distribution is generated for subsequent path optimization.

[0112] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0113] Based on the same inventive concept, this application also provides a large-model-based dynamic optimization device for implementing the large-model-based dynamic optimization method for health plans described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the large-model-based dynamic optimization device provided below can be found in the limitations of the large-model-based dynamic optimization method for health plans described above, and will not be repeated here.

[0114] In one exemplary embodiment, such as Figure 4 As shown, a dynamic optimization device for health plans based on a large model is provided, comprising:

[0115] The data acquisition module 401 is used to acquire the user's set health goals, user historical behavior data, and self-psychological assessment scores; the user historical behavior data includes the history of health plan execution, the history of plan execution status, and user interaction logs.

[0116] The large-scale user personality construction module 402 is used to construct the current user's personality vector based on the user's historical behavioral data and self-psychological assessment scores.

[0117] The large-scale model failure personality simulation module 403 is used to construct a set of failure personality vectors based on the historical execution of health plans and the historical plan execution status.

[0118] The health plan planning module 404 is used to obtain the health plan intervention path by engaging in a personality-driven intervention path game based on health goals, the current user's personality vector, and the set of failed personality vectors.

[0119] In one embodiment, the large model failure personality simulation module 403 is also used for:

[0120] Extract historical execution health plans and reasons for failures from the historical plan execution data;

[0121] Clustering modeling was performed based on multiple historical execution failures of health plans and the reasons for failure to obtain multiple potential failure personality models;

[0122] By conducting multi-round dialogue simulations on multiple potential failure personality models to create personality profiles, and extracting the failure latent variable representations for each personality profile, a set of failure personality vectors is obtained.

[0123] In one embodiment, the health plan planning module 404 is further configured to:

[0124] Based on users' daily temporal behavior templates and external environmental constraints, health goals are semantically vectorized to obtain a set of target subtasks and constraint dimensions; the constraints include external environmental constraints.

[0125] Based on the current user personality vector and the set of failed personality vectors, the user id, the failed personality, and the health intervention coach are used as roles to engage in adversarial game, and multiple initial health plans are generated according to the target sub-task set and constraint dimensions.

[0126] Based on the evaluation function, the stress resistance score of each behavioral suggestion in each initial health plan is calculated for the set of failed personality vectors, resulting in a set of stress resistance scores; the set of stress resistance scores includes each initial health plan and its corresponding score sequence;

[0127] The initial health plan corresponding to the highest comprehensive score in the stress resilience score set is identified as the health plan intervention path.

[0128] In one embodiment, a dynamic correction module is also included, for:

[0129] In response to the obtained actual user execution data, the actual user execution data is compared with the health plan intervention path to obtain the deviation sequence;

[0130] Based on a preset long time window, the deviation sequence is extracted to obtain continuous deviation values;

[0131] If the continuous deviation value exceeds the preset deviation threshold, the health plan intervention path will be reconstructed by replacing the target group.

[0132] In one embodiment, a trajectory deviation module is also included, for:

[0133] Generate an ideal digital avatar for the user based on the health plan intervention path;

[0134] By using the user's ideal digital avatar to simulate the user's daily behavior path based on the user's schedule data, social events, and daily workload, the ideal behavior trajectory can be obtained;

[0135] Transform users' actual execution data into structured actual behavior trajectories;

[0136] The deviation sequence is obtained by calculating the distance between the behavior codes of the actual behavior trajectory and the ideal behavior trajectory at each time period.

[0137] In one embodiment, a reconfiguration module is also included, for:

[0138] Based on the challenge value, willingness-to-benefit ratio, and semantic consistency of the initial health goal, a set of alternative goals is generated by multi-dimensional alternative goals according to the goal sub-task set and constraint dimensions.

[0139] Determine the reasons for the failure of health plan intervention pathways based on continuous deviation values;

[0140] Determine the optimal alternative target from the set of alternative targets based on the reasons for failure;

[0141] Health plan intervention pathways are generated based on the optimal alternative goals as new health goals.

[0142] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the above method embodiments.

[0143] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0144] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0145] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A large model-based health plan dynamic optimization method, characterized in that, The method comprises: obtaining a user-set health goal, user historical behavior data, and a self-psychological evaluation score; the user historical behavior data comprises historical health plan execution, historical plan execution, and user interaction logs; constructing a current user personality vector according to the user historical behavior data and the self-psychological evaluation score; extracting the historical health plans with execution failure and failure causes in the historical plan execution; performing clustering modeling on a plurality of the historical health plans with execution failure and the failure causes to obtain a plurality of potential failure personality models; extracting failure latent variable representations of each type of personality portrait through multi-round dialogue simulation personality portrait of a plurality of the potential failure personality models to obtain a failure personality vector set; performing semantic vectorization on the health goal based on user daily time-series behavior templates and external environment restrictions to obtain a target sub-task set and constraint dimensions; the constraint dimensions comprise external environment restriction conditions; performing an adversarial game with a user ego, a failure personality, and a health intervention coach based on the current user personality vector and the failure personality vector set, and generating a plurality of initial health plans according to the target sub-task set and the constraint dimensions; calculating a stress resistance score of each behavior suggestion in each of the initial health plans based on an evaluation function, and obtaining a stress resistance score set; the stress resistance score set comprises each of the initial health plans and a corresponding score sequence; determining the initial health plan corresponding to the highest comprehensive score in the score sequence of the stress resistance score set as the health plan intervention path; the health plan intervention path comprises behavior suggestions and corresponding time period constraints and behavior context considerations; generating a user ideal digital avatar according to the health plan intervention path in response to obtained user real execution data; obtaining an ideal behavior trajectory through user ideal digital avatar simulation according to user schedule data, social events, and daily behavior path simulation of daily load; converting the user real execution data into a structured actual behavior trajectory; calculating behavior encoding distances of each time period in the actual behavior trajectory and the ideal behavior trajectory to obtain a deviation sequence; extracting a continuous deviation value based on a pre-set long time window on the deviation sequence; if the continuous deviation value exceeds a pre-set deviation threshold, generating a multi-element replacement target based on the target sub-task set and the constraint dimensions according to a challenge value, a willingness combination degree, and an initial health goal semantic consistency to obtain a replacement target set; determining a failure cause of the health plan intervention path according to the continuous deviation value; determining an optimal replacement target in the replacement target set according to the failure cause; generating the health plan intervention path with the optimal replacement target as a new health goal.

2. The method of claim 1, wherein, The method of calculating a stress resistance score of each behavior suggestion in each of the initial health plans based on an evaluation function to obtain a stress resistance score set comprises: a stress resistance score is obtained through the following formula: ； wherein, is the first behavior recommendation in the initial health plan; is the first behavior recommendation in the initial health plan; is the first behavior recommendation in the initial health plan; is the first failure personality vector; is the behavior semantic encoding vector of the behavior recommendation; is the vector cosine similarity; is the external environment constraint condition; is the contextual adaptation degree of the behavior recommendation to the current external environment constraint condition; is the adaptation degree weight coefficient.

3. A large model-based health plan dynamic optimization apparatus, characterized by, The device comprises: The data acquisition module is configured to acquire a health target set by a user, historical behavior data of the user, and a self-psychological evaluation score; the historical behavior data of the user includes historical health plans, historical plan execution conditions, and a user interaction log; The large model user personality construction module is configured to construct a current user personality vector according to the historical behavior data of the user and the self-psychological evaluation score; The large model failure personality simulation module is configured to extract the historical health plans that fail to be executed and failure causes in the historical plan execution conditions; The large model failure personality simulation module is further configured to cluster and model the historical health plans that fail to be executed and the failure causes, to obtain a plurality of potential failure personality models; The large model failure personality simulation module is further configured to simulate personality portraits through multi-round dialogue simulation of the plurality of potential failure personality models, and extract failure latent variable representations of each type of personality portrait, to obtain a failure personality vector set; The health plan planning module is configured to perform semantic vectorization on the health target based on a user daily time sequence behavior template and external environment restrictions, to obtain a target sub-task set and a constraint dimension; the constraint dimension includes external environment restriction conditions; The health plan planning module is further configured to perform an adversarial game with a user ego, a failure personality, and a health intervention coach based on the current user personality vector and the failure personality vector set, and generate a plurality of initial health plans according to the target sub-task set and the constraint dimension; The health plan planning module is further configured to calculate a stress resistance score of each behavior suggestion in each of the initial health plans based on an evaluation function, to obtain a stress resistance score set; the stress resistance score set includes each of the initial health plans and a corresponding score sequence; The health plan planning module is further configured to determine that an initial health plan corresponding to a highest comprehensive score in the score sequence of the stress resistance score set is a health plan intervention path; the health plan intervention path includes behavior suggestions and corresponding time period constraints and behavior context considerations; The trajectory deviation module is configured to generate a user ideal digital avatar according to the health plan intervention path in response to acquired user real execution data; The trajectory deviation module is further configured to simulate an ideal behavior trajectory according to user schedule data, social events, and daily behavior paths of daily loads by using the user ideal digital avatar; The trajectory deviation module is further configured to convert the user real execution data into a structured actual behavior trajectory; The trajectory deviation module is further configured to calculate a behavior encoding distance of each time period in the actual behavior trajectory and the ideal behavior trajectory, to obtain a deviation sequence; The dynamic correction module is configured to extract deviations from the deviation sequence based on a preset long time window, to obtain continuous deviation values; The reconstruction module is configured to, if the continuous deviation values exceed a preset deviation threshold, perform multi-element replacement target generation according to the target sub-task set and the constraint dimension based on a challenge value, a willingness combination degree, and an initial health target semantic consistency, to obtain a replacement target set. The reconstruction module is further configured to determine a failure cause of the health plan intervention path according to the continuous deviation value; The reconstruction module is further configured to determine an optimal alternative target in the alternative target set according to the failure cause; The reconstruction module is further configured to generate the health plan intervention path according to the optimal alternative target as a new health target.

4. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor implements the method in any one of claims 1 and 2 when executing the computer program.

5. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by a processor, implements the method in any one of claims 1 and 2.

Images