Meal nutrition scientific evaluation and improvement method and system based on AI large model

Abstract

The invention discloses a meal nutrition scientific evaluation and improvement method and system based on an AI large model, relates to the technical field of scientific evaluation, and is used for solving the problem that the existing method generally adopts a fixed threshold value to judge whether food group intake reaches the standard or not and cannot dynamically adjust a judgment standard according to data integrity under the condition that records are incomplete, so that the efficiency is high. And misjudgment is easily caused. A linkage closed-loop food intake record optimization method is constructed by introducing an integrity credibility scoring mechanism and taking the integrity credibility scoring mechanism as core input. According to the method, data integrity evaluation, dynamic threshold gating, adaptive sampling intensity scheduling, uncertainty gating compensation, double-factor suggestion intensity mapping and parameter adaptive updating form a closely connected necessary chain, and system-level optimization is achieved.

Description

Technical Field

[0001] This invention relates to the field of scientific evaluation technology, and more specifically, to a method and system for scientific evaluation and improvement of meal nutrition based on AI large-scale models. Background Technology

[0002] With the rapid development of mobile internet and health management technologies, the accurate recording and analysis of food intake data has become a crucial requirement in the health management field. Existing mobile health applications typically record dietary data through manual user input or image recognition technology, and then use this data to conduct nutritional analysis and provide health recommendations. However, these technologies still face many challenges in data processing and algorithm optimization.

[0003] The existing technology has the following shortcomings: First, for cases with incomplete records, existing methods typically use fixed thresholds to determine whether food group intake meets standards, failing to dynamically adjust the criteria based on data completeness, which easily leads to misjudgments. Second, when assessing weekly food diversity, existing technologies struggle to effectively handle missing data, resulting in insufficient accuracy. Furthermore, existing bootstrap sampling intensities usually use fixed values, unable to adaptively adjust based on the degree of data loss, potentially wasting computational resources and affecting the reliability of results. More importantly, existing technologies lack effective closed-loop optimization mechanisms; the compensation strategy lacks linkage with uncertainty assessment, easily leading to overcompensation. Simultaneously, the lack of dynamic adjustment mechanisms for recommendation strength and prior parameter updates makes it difficult to achieve personalized and accurate recommendations.

[0004] To address the above problems, this invention proposes a solution. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method and system for scientific assessment and improvement of meal nutrition based on AI large models, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The method for scientific assessment and improvement of meal nutrition based on AI big data models includes the following steps: Step 1: Obtain and standardize dietary records, determine the number of items to be recorded each day and the credibility of each item based on the user's history, calculate the daily completeness-credibility score, and archive the data to the shared intermediate in the evaluation link after structuring the data; Step 2: Establish basic intake thresholds for each food group, and generate daily effective thresholds according to the segmented double-slope rule of the integrity-reliability score. The rule uses a larger adjustment coefficient in the low integrity range and a smaller adjustment coefficient in the high integrity range, and the effective thresholds do not increase monotonically with the score. Step 3: Estimate food diversity on a weekly scale. For each food group on missing days, combine the observational indications and prior occurrence probabilities and smooth them to obtain the weekly coverage probability. Combine the daily record completeness with the coverage probability and truncate the upper limit of each group to summarize the expected diversity score. Step 4: Determine the missing severity using the complement of the weekly mean of completeness, and set the bootstrap resampling intensity accordingly. Adaptively select the bootstrap interval type to obtain the intake confidence interval and its relative width. The selection depends on the number of observation days and the missing severity. Step 5: Perform dual-gated compensation and parameter update: When the interval width does not exceed the threshold, the recommended intake is higher than the actual intake, and the number of observation days reaches the preset lower limit, the compensation amount is determined according to the rule that it is proportional to the gap, decreases with uncertainty, and is constrained by the global upper limit; the suggestion strength is generated based on the mapping between the lower bound of the interval and the missing severity, and the prior update rate is adjusted according to the compliance rate and the interval convergence to update the prior occurrence probability parameter.

[0007] In a preferred embodiment, the integrity confidence score is determined by the number of food items to be recorded and the confidence level of the recorded items; the number of food items to be recorded is calculated based on the user's average daily record volume over a preset time window in the past; the confidence level is given by the recording method and can be revised with user confirmation, and the scoring weight is set in a configurable range and bound to the record source.

[0008] In a preferred embodiment, the segmented dual-slope rule divides the scoring interval into two segments, low integrity and high integrity, using a preset breakpoint. The low integrity segment uses a first slope coefficient, and the high integrity segment uses a second slope coefficient that is smaller than the first slope coefficient. The effective intake threshold remains monotonically non-increasing relative to the score, and the basic threshold and breakpoint parameters are recorded for each food group.

[0009] In a preferred embodiment, the prior occurrence probability is initialized hierarchically according to population statistics, including age, gender, region, and dietary preference dimensions; the smoothing coefficient is selected within a preset range and used to limit the magnitude of prior variation; the truncation process limits the upper limit of a single food group in the week and stores it in a unified structured format for subsequent updates.

[0010] In a preferred embodiment, the missing severity is calculated based on the mean of the daily integrity confidence scores over a week; the self-administered resampling intensity is determined by a power function mapping between the minimum and maximum number of resampling attempts, and rounded up to the nearest integer; the relevant mapping coefficients and value boundaries are registered in the metadata for auditing and recalculation.

[0011] In a preferred embodiment, the confidence interval calculation method is switched according to rules: when the number of observation days is less than the minimum observation threshold and the missing severity is not lower than the preset threshold, the BCa method is used; otherwise, the percentile method is used. The minimum observation threshold and the missing threshold are configurable, and the switching record is archived as an auxiliary parameter for interval calculation.

[0012] In a preferred embodiment, the dual-gating compensation includes a permission gate and a quota gate: the permission gate is determined based on the relative width, the relationship between the recommended intake and the actual intake, and the satisfaction of the number of observation days; the quota gate calculates the compensation amount according to the gap ratio and in combination with the uncertainty adjustment coefficient and the compensation upper limit when the permission gate passes, and performs version management of the input parameters and thresholds.

[0013] In a preferred embodiment, the recommendation strength is obtained by a two-factor weighted mapping, the two factors being the ratio of the lower bound of the intake range to the recommended intake and the severity of the missing information; the weight coefficients are configured between zero and one; the recommendation level is generated by mapping from a preset threshold table, and the threshold table and weights are stored independently to support separate maintenance and rollback.

[0014] In a preferred embodiment, the adaptive update of the prior parameters adopts a multiplicative learning rate scheme, and the learning rate is jointly adjusted by the user compliance rate and the convergence index of the intake interval; the learning rate is set with upper and lower boundaries and includes a linear combination coefficient, and the update rhythm is executed weekly; during the update, the old and new priors are smoothed with an exponential moving average and the timestamp is recorded.

[0015] In a preferred embodiment, a resource controller is configured to manage the computation budget and single-cycle cap; when the cumulative budget approaches a threshold or the self-service resampling intensity reaches its limit, a cap is applied to the resampling intensity, or the interval width threshold is temporarily increased, or the compensation calculation is postponed to the next cycle; the alternative strategy can be enabled independently and the decision basis is recorded.

[0016] In a preferred embodiment, a structured data output mode is established, in which all intermediate quantities and parameters include identifiers, units, time ranges, and source labels; when the suggested intensity mapping and the compensation calculation involve data from the same period, references are preferentially made using intervals and thresholds from the same batch, cross-batch mixing is prohibited, and batch relationships are stored in the form of foreign keys.

[0017] A system for scientific assessment and improvement of meal nutrition based on AI big data models, including: The data acquisition module is used to acquire dietary records formed by manual input and image recognition; The data preprocessing module, connected to the data acquisition module, is used for field validation, format standardization and alias mapping, and to generate record credibility. The integrity assessment module, connected to the data preprocessing module, is used to determine the recording benchmark based on user history and calculate the daily integrity credibility score by combining the credibility of each record. The dynamic threshold adjustment module, connected to the integrity assessment module, is used to read the basic intake threshold for each food group and adjust the effective threshold for the day according to the integrity confidence score based on the segmented double slope rule. The low integrity segment uses the first slope, the high integrity segment uses the second slope which is less than the first slope, and the effective threshold remains monotonically non-increasing relative to the score. The diversity assessment module, connected to the data preprocessing module and the dynamic threshold adjustment module, is used to smooth the comprehensive observation indicators and prior occurrence probabilities of food groups on missing days at a weekly granularity to obtain the weekly coverage probability, and to summarize the expected diversity score after truncating each food group in combination with the completeness of daily records. The self-help intensity adjustment module, connected to the diversity assessment module and the integrity assessment module, is used to calculate the missing severity based on the mean of the integrity confidence score within one week, and map the missing severity to the minimum and maximum number of resampling times to set the self-help resampling intensity, which is rounded to an integer; and selects the BCa interval method when the number of observation days is insufficient and the missing severity is not lower than the threshold, otherwise selects the percentile method. The uncertainty assessment and compensation module, connected to the bootstrap intensity adjustment module, is used to generate an intake confidence interval and calculate the relative width based on the resampling intensity; and to perform dual-gating compensation: the permission gate requires that the relative width is not greater than a threshold, the recommended intake is greater than the actual intake, and the number of observation days is not less than the lower limit; when the permission gate is satisfied, the quota gate determines the compensation amount proportionally to the gap, decreasing with uncertainty and constrained by the compensation upper limit. The suggestion generation module, connected to the uncertainty assessment and compensation module, is used to perform a weighted mapping of the ratio of the lower bound of the intake range to the recommended intake and the severity of the deficiency to obtain the suggestion strength. The weights are between zero and one, and the suggestion level is mapped by a threshold table. The parameter update module, connected to the diversity assessment module and the suggestion generation module, is used to update the prior parameters according to the weekly rhythm using a multiplicative learning rate. The learning rate is jointly adjusted by the user compliance rate and the convergence of the intake interval and is set with upper and lower boundaries. During the update, an exponential moving average smoothing is performed on the new and old priors. The resource controller, connected to the uncertainty assessment and compensation module and the self-help intensity adjustment module, is used to set the calculation budget and single-cycle upper limit, and when the upper limit is reached or close to the budget, to cap the resampling intensity, adjust the interval threshold, or delay compensation. The large model collaboration layer interacts with the data preprocessing module, the diversity assessment module, the uncertainty assessment and compensation module, the suggestion generation module, and the parameter update module in a candidate / draft manner. It is used to perform semantic standardization and knowledge retrieval alignment, generate structured candidates and explanation drafts, and does not directly modify the numerical parameters and thresholds confirmed by each algorithm module.

[0018] The technical effects and advantages of this invention, which is a method and system for scientific assessment and improvement of meal nutrition based on an AI large model, are as follows: This invention constructs a closed-loop optimization method for food intake records by introducing an integrity and reliability scoring mechanism as its core input. This method forms a tightly linked necessary chain encompassing data integrity assessment, dynamic threshold gating, adaptive sampling intensity scheduling, uncertainty gating compensation, two-factor suggestion intensity mapping, and adaptive parameter updates, achieving system-level optimization.

[0019] By employing a segmented / dual-slope threshold gating mechanism, different adjustment slopes are used in the low-quality recording range and the medium-to-high-quality recording range, effectively suppressing both the problems of "inflated standards" and "overly strict misjudgments," resulting in more precise and reasonable threshold adjustments. A resource manager is introduced, setting a weekly computation budget and a single-cycle upper limit, and providing alternative strategies when the upper limit is reached. This reflects a unique strategy of balancing system-level resources and accuracy, ensuring a balance between computational efficiency and user experience.

[0020] A "dual-gating" compensation strategy is adopted, breaking down compensation into a "permission gate" and a "quota gate," which are linked to the uncertainty, magnitude of the gap, and risk level of food intake. This is not a simple threshold switch, but a joint governance of statistical security and user risk, significantly reducing the risk of overcompensation. The "two-factor mapping" of recommendation strength ensures that the recommendation strength is more consistent with the mechanism of user behavior and health risk, making the recommendations more targeted and effective.

[0021] The adaptive parameter update, triggered by both "compliance changes" and "interval convergence changes," achieves a joint optimization of long-term stability and individualized learning speed, enabling the system to continuously learn and adapt to user behavior. This allows the invention to provide more accurate assessments, more reasonable compensation, and more personalized suggestions even with incomplete or missing data, significantly improving the effectiveness of health management. Attached Figure Description

[0022] Figure 1 This is a schematic diagram illustrating the closed-loop linkage between meal nutrition assessment and improvement based on a large AI model. Figure 2 A schematic diagram of the dynamic threshold segmented double slope adjustment mechanism; Figure 3 This is a schematic diagram of the resource controller's workflow; Figure 4 This is a schematic diagram of a dual-gating compensation strategy; Figure 5 This is a schematic diagram of the proposed intensity two-factor mapping; Figure 6 A schematic diagram of the parameter adaptive update process; Figure 7 This is a schematic diagram of the overall process of the method; Figure 8 This is a diagram illustrating the system architecture modules and their collaborative relationships. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] The following is a corresponding symbol table and related terminology from an embodiment of the present invention: Symbol / Unit Table: Daily integrity reliability score, unitless, value range [0.0, 1.0].

[0025] : Dynamically adjusted effective intake threshold for food groups, in grams (g) or kcal (kcal).

[0026] s: Missing severity, unitless, range [0.0, 1.0].

[0027] B: Strength of self-help method, unitless, positive integer.

[0028] : Relative width of uncertainty in food intake, unitless, range [0.0, 1.0].

[0029] δ: Uncertainty threshold, unitless, range [0.0, 1.0].

[0030] Compensation amount, in grams (g) or kilocalories (kcal).

[0031] Maximum compensation amount, in grams (g) or kilocalories (kcal).

[0032] L: Lower bound of the confidence interval, in grams (g) or kilocalories (kcal).

[0033] Recommended intake, in grams (g) or kcal (kcal).

[0034] Actual intake, in grams (g) or kcal (kcal).

[0035] Intensity: Suggested intensity, unitless, range [0.0, 1.0].

[0036] ρ: User compliance rate, unitless, range [0.0, 1.0].

[0037] UL: Confidence Interval Width, in grams (g) or kilocalories (kcal).

[0038] U+L: The sum of the upper and lower bounds of the confidence interval, in grams (g) or kilocalories (kcal).

[0039] Terms and definitions: AI large models: refer to pre-trained general semantic models (language or multimodal) used for semantic standardization, knowledge retrieval alignment, suggestion and explanation text generation, and do not directly determine the numerical value.

[0040] This embodiment provides a method for scientific assessment and improvement of food nutrition based on an AI large model. This method solves the problems of inaccurate assessment when food intake records are incomplete or missing, unreasonable allocation of computing resources, and lack of effective closed-loop optimization mechanism in the prior art by constructing a tightly coupled linkage closed-loop system.

[0041] Figure 1 This diagram illustrates the closed-loop linkage of the AI-based large-scale model-based scientific assessment and improvement method for food nutrition provided in this embodiment of the invention. The core of this method lies in organically linking each processing step to form an "input-processing-output-feedback" closed-loop system. This ensures that the diagnostic results of the previous step directly influence the judgment criteria, sampling intensity, compensation strategy, and recommendation strength of the subsequent step. Simultaneously, the uncertainty assessment of the subsequent step inversely constrains the compensation magnitude and prior parameters of the previous step.

[0042] AI Large Model Collaboration Layer: This invention preferably introduces an AI large-scale model collaboration layer as a horizontal enhancement module, without altering the decision-making power of the main process. This layer may include a semantic normalization submodule, a knowledge retrieval and alignment submodule (RAG), a suggested text generation submodule, and an interpretability generation submodule. Its input may include structured elements of the original food text / image, ,s, , , , And user profiles. Its output may include normalized naming and synonym mapping, and The system initializes / revises candidate data, provides draft suggestion text and structured intent slots, and draft explanation signals. The AI ​​large model collaboration layer does not directly rewrite numerical results and thresholds; all values ​​are still calculated or confirmed by each module. The large model only provides candidate values / text and ontology alignment. The AI ​​large model collaboration layer can be deployed locally or in the cloud, requires no large-scale retraining, and can be fine-tuned. Its RAG data source can be public nutrition guidelines and ontology libraries. When the large model cannot provide effective mappings or suggestions, the system will fall back to the preset rule base or prompt for manual confirmation.

[0043] Figure 1 and Figure 8 In the process, the AI ​​large model collaboration layer interacts with the data preprocessing module, diversity assessment module, uncertainty assessment module, suggestion generation module, and parameter update module in a candidate / draft manner, and the final value is confirmed by the modules.

[0044] The necessary links and input-output relationships of this closed loop are as follows: 1. Data Integrity Assessment (Integrity Assessment Module): Input: Raw diet record data.

[0045] Output: Completeness and reliability score of daily food intake records .

[0046] Impact on downstream industries: It is directly used as input to the dynamic threshold adjustment module to determine the adjustment range of the threshold; at the same time, The weekly average value is used to calculate the missing severity s, which in turn affects the self-help strength adjustment module.

[0047] 2. Threshold gating (dynamic threshold adjustment module): Input: Integrity credibility score Basic intake threshold for food groups .

[0048] Output: Dynamically adjusted effective intake thresholds for each food group .

[0049] Impact on downstream industries: Used to determine whether the daily food intake meets the standard, its adjustment directly affects the initial diagnosis of the user's eating behavior and avoids misjudgment due to incomplete data.

[0050] 3. Sampling intensity scheduling (bootstrap intensity adjustment module): Input: Missing severity s.

[0051] Output: Strength of self-help method: B.

[0052] Impact on downstream applications: B directly determines the number of times the bootstrap method is resampled in the uncertainty assessment module, thus affecting the calculation accuracy of the confidence interval and the consumption of computing resources.

[0053] 4. Uncertainty Gating Compensation (Uncertainty Assessment Module): Input: Self-help intensity B, assessment results (e.g., weekly food intake, weekly food diversity) Recommended intake Actual intake .

[0054] Output: Confidence interval for food intake Its relative width Food diversity confidence interval Its relative width Compensation amount .

[0055] Impact on downstream industries: and It directly affects the strength of the suggestions generated by the suggestion generation module; and (Interval convergence) has a reverse effect on the learning rate adjustment of the parameter update module. Used for structural recommendations and strategy explanations.

[0056] 5. Suggested intensity mapping (suggested generation module): Input: Lower bound of the confidence interval for food intake Severity of deficiency (s), recommended intake .

[0057] Output: Suggested Intensity.

[0058] Impact on downstream applications: Intensity determines the content and intensity of suggestions displayed to users, directly affecting user experience and compliance.

[0059] 6. Adaptive parameter update (parameter update module): Input: User compliance rate ρ, convergence of food intake intervals Current prior learning rate .

[0060] Output: Updated prior learning rate .

[0061] Impact on downstream industries: Used to update prior parameters within the system (such as...) This affects the probability inference of the diversity assessment module, thus forming a closed-loop feedback.

[0062] The quantifiable loss to the overall effect due to the absence of any step: The lack of dynamic threshold gating will cause the system to use a fixed threshold when the record is incomplete, resulting in a large number of "falsely high standards" or "overly strict misjudgments", and the evaluation accuracy will drop significantly.

[0063] Lack of sampling intensity scheduling: The sampling intensity of the bootstrap method is fixed. If a low intensity is still used when there is a lot of missing data, the uncertainty interval will be inaccurate, resulting in distorted compensation and suggestions. If a high intensity is still used when the data is complete, computing resources will be wasted, affecting the response speed of mobile devices.

[0064] Lack of uncertainty gating compensation: This will cause the system to continue to compensate when the uncertainty of the evaluation result is high, resulting in overcompensation, which may cause user resentment or health risks.

[0065] The lack of a two-factor mapping for suggestion strength means that suggestion strength depends on only a single dimension, which cannot take into account both the conservatism of the estimate and the quality of the data. This leads to a decrease in the relevance of the suggestions and may reduce user compliance.

[0066] Lack of adaptive parameter updates: The system cannot learn and adjust based on user behavior and data quality. After long-term operation, the model performance will gradually decline, and the evaluation accuracy may decrease.

[0067] II. Detailed explanation of each component / method step: The components mentioned above will be explained in detail below.

[0068] (I) Regarding the data acquisition module and the data preprocessing module: Figure 8 The system architecture of an embodiment of the present invention is illustrated. The data acquisition module may include a manual input unit and an image recognition unit for acquiring the user's dietary record data. The data preprocessing module receives the raw dietary record data and cleans, standardizes, and converts its format.

[0069] Data fields and interface descriptions (as an implementation example): Input data fields: may include (string), (string), (date), (string), (enumeration), quantity (floating-point number), unit (string), (enumerate), (floating-point number) (Boolean).

[0070] The value range is predefined food categories, such as {"grains", "vegetables", "fruits", "meats", "dairy products", "oils", "soybean products"}.

[0071] The unit of quantity is grams (g) or milliliters (ml), and the value range is positive floating-point numbers.

[0072] The value range of unit is {"g", "ml", "piece", "portion"}.

[0073] The value range is [0.0, 1.0].

[0074] Default logic: If is missing, try to match the preset food library through ; if is missing, try to estimate through and ; if is missing, the default value for manual input is 0.95, and the default value for image recognition is 0.7.

[0075] Optional fields: may include (string), (string), (boolean, marking whether participated by the large model), (string array, storing the guide fragment IDs retrieved by RAG).

[0076] Intermediate quantity fields: may include (floating-point number), (floating-point number), s (floating-point number), B (integer), , (floating-point number), (floating-point number), , (floating-point number), (floating-point number), (floating-point number), (floating-point number).

[0077] Output fields: may include (floating-point number), (floating-point number), (enumeration), explanation_signals (string array). <00​​​​Output pattern: The output of the large model follows a predefined schema (intent slot / evidence ID / confidence); if the schema is not met, a fallback is triggered.

[0080] Optional AI large model collaboration: The data preprocessing module can incorporate the semantic standardization submodule from AI large model collaboration layer 1 to achieve... The mapping of aliases or colloquial expressions to the nutritional ontology outputs standardized names and synonym mappings, which can be corrected based on user confirmation or correction. For continuous learning.

[0081] (II) Regarding the integrity assessment module: The integrity assessment module is electrically connected to the data preprocessing module and is used to calculate the integrity reliability score of daily food intake records. .

[0082] 1. Integrity and Credibility Score Calculation: The integrity and credibility score of daily records is calculated using the following formula: ; in, The actual number of food items recorded on that day. To predict the number of food items to be recorded on a given day based on user historical data. Let α be the credibility score for the i-th record, and α be the weighting parameter.

[0083] Parameter: α default value 0.7. Preferably, the calculation is based on the user's average daily record count over the past 30 days.

[0084] (III) Regarding the dynamic threshold adjustment module: Segmented / Dual-slope mechanism: Figure 2 A segmented / dual-slope schematic diagram of the dynamic threshold adjustment module is shown. The dynamic threshold adjustment module is electrically connected to the integrity assessment module and is used to score integrity reliability. Dynamically adjust the effective intake thresholds for each food group .

[0085] In an optional implementation, the present invention employs a segmented / dual-slope mechanism to adjust the threshold. The aim is to: ensure record integrity. When the threshold is low, it should be increased at a faster rate to more rigorously determine whether the standard is met; when record integrity... When the threshold is high, it should be adjusted upwards at a slower rate to maintain the stability of the assessment.

[0086] 1. Dynamic Adjustment Calculation: The adjusted effective intake threshold is calculated using the following piecewise function: ; Among them, the function Defined as: ; Form constraints and invariants: This function satisfies That is, with the integrity score The increase in effective intake threshold It is monotonically non-increasing. At the same time, the double-slope interval satisfies β1>β2>0, ensuring a steeper adjustment in the low integrity interval, and the adjustment direction is always upward.

[0087] parameter: β1 is preferably 0.5. β2 is preferably 0.8. β3 is preferably 0.3.

[0088] (iv) Regarding the diversity assessment module: The diversity assessment module is electrically connected to the data preprocessing module and is used to assess weekly food diversity in the event of missing data.

[0089] 1. Weekly Coverage Probability Calculation: For food group g on missing days, the weekly coverage probability is calculated using the following formula. : ; Wherein, γ can preferably be 2.

[0090] Initialization and cold start: During the cold start phase or when a new user registers, It can be based on large-scale population dietary data statistics and perform population stratification initialization (such as by age, gender, region, and dietary preferences). During the cold start period, the system will use conservative priors and increase δ while decreasing Intensity.

[0091] Update rhythm and smoothing: The parameter update module performs adaptive adjustments, and the update rhythm is preferably weekly. To prevent short-term noise from entering, smoothing techniques such as exponential moving average (EMA) are used during updates, and the EMA smoothing coefficient is preferably [0.01, 0.1].

[0092] Optional collaboration with large AI models: During a cold start, initial candidate values ​​can be generated by the Knowledge Retrieval and Alignment (RAG) submodule in the AI ​​large model collaboration layer in conjunction with the nutrition knowledge base. During runtime, when updated by the parameter update module, the large model only provides literature or guideline evidence fragments and prior candidates, and the final acceptance is still confirmed by the algorithm-side EMA or constraint filtering.

[0093] (V) Regarding the self-help method intensity adjustment module: The self-help strength adjustment module is electrically connected to the diversity assessment module and is used to dynamically set the self-help strength B according to the missing severity s.

[0094] 1. Missing Severity Calculation: The missing severity s is calculated using the following formula: ; 2. Self-help method strength adjustment: The self-help method strength B is calculated using the following formula: ; parameter: The preferred value is 0. The preferred value is 1. k is preferably 1.5.

[0095] Bootstrap selection strategy is adaptive: when there are recorded days Less (e.g.) When the missing severity s is high (e.g., s>0.5), the system preferably uses the BCa method to calculate the confidence interval; otherwise, the percentile method is preferred.

[0096] Degenerate output strategy: When the number of observation days is too small (e.g., fewer than 10 days in a week are recorded). If the number of resampling attempts is insufficient (preferably 3 days) or the number of resampling attempts is insufficient, the system will output a message that "data is insufficient and the evaluation result is highly uncertain" and may set the confidence interval width to the maximum value to trigger a more conservative compensation and recommendation strategy.

[0097] (vi) Regarding the uncertainty assessment module: "Dual-gating" compensation strategy: Figure 4 A schematic diagram of the "dual-gating" compensation strategy is shown. The uncertainty assessment module is electrically connected to the bootstrap strength adjustment module to assess the uncertainty of the data and determine whether to trigger compensation.

[0098] In one embodiment, the present invention employs a "dual-gating" compensation strategy, splitting the compensation into a "permission gate" and a "quota gate," and relating it to the uncertainty of food intake (…). The magnitude of the gap is linked to the level of risk.

[0099] 1. Confidence Interval Calculation: Generate B samples using the bootstrap method, and calculate the confidence intervals for food intake (e.g., total energy, specific nutrient intake) for each sample. and its relative width and food diversity score ( ) confidence interval and its relative width .

[0100] Used for compensation and recommendation of strength.

[0101] Used for structural recommendations and strategy explanations.

[0102] 2. Permission gate (whether compensation is allowed): Safety invariant: If or or (Preferred) )but .

[0103] Parameter: δ is preferably 0.2 3. Quota threshold (maximum amount to be supplemented): Calculation formula: ; Safety invariants: ,right Monotonically non-increasing, opposite ( Monotonic and non-decreasing.

[0104] Parameter: η is preferably 0.3. Preferably, it is 5% of the total daily intake or 10 grams of a specific food group.

[0105] Optional output from the AI ​​large model collaboration: Uncertainty assessment module and Subsequently, the interpretability generation submodule in the AI ​​large model collaboration layer can translate the determined statistics into a readable draft of interpretive signals. Any compensation licenses or quotas are still calculated by the rules of the uncertainty assessment module.

[0106] (vii) Regarding the suggestion generation module: "Two-factor mapping": Figure 5 A schematic diagram of the "two-factor mapping" for recommendation strength is shown. The recommendation generation module is electrically connected to the uncertainty assessment module to generate personalized dietary recommendations.

[0107] In an alternative implementation, the present invention employs a "two-factor mapping" of the recommended strength, explicitly stating that the recommended strength is simultaneously subject to "estimated conservatism (using the lower bound of the confidence interval for food intake)". The impact of "data missing severity s" is reflected in the data.

[0108] 1. Suggested Intensity Calculation: The suggested intensity (Intensity) is calculated using the following formula: ; In the formula: λ is preferably 0.6.

[0109] 2. Suggestion Content Generation: Select the appropriate suggestion strength and content based on the calculated Intensity value. For example, Intensity < 0.1 is a mild reminder, 0.1 ≤ Intensity < 0.25 is a general suggestion, 0.25 ≤ Intensity < 0.4 is a strong suggestion, and Intensity ≥ 0.4 is a very strong suggestion.

[0110] Optional AI large model collaboration: After Intensity is given by the numerical link, it can be generated by the suggestion text generation submodule in AI large model collaboration layer 1 based on the Intensity, the missing severity s, and the... The template is filled with structured intent slots and the language is polished to generate a draft recommendation text. This draft text is then filtered by safety red lines and medical order priority rules to prevent unauthorized recommendations.

[0111] (viii) Regarding the parameter update module: Dual-trigger adaptive update: Figure 6 The diagram illustrates a dual-trigger design for adaptive parameter updates. The parameter update module is electrically connected to the uncertainty assessment module and is used to adjust system parameters based on user feedback and system performance to achieve closed-loop optimization.

[0112] In an optional implementation, the present invention employs a dual-trigger mechanism to adjust the prior learning rate. This means that it is simultaneously driven by "changes in compliance" and "changes in interval convergence", emphasizing the joint optimization of long-term stability and individualized learning speed.

[0113] 1. Learning rate adjustment: Calculate the prior learning rate using the following formula. : ; In the formula: 0.7 is an option. The value can be 0.1. The value can be set to 0.05.

[0114] Optional AI large model collaboration: When a nutrition guideline or ontology update is detected, the knowledge retrieval and alignment submodule in AI large model collaboration layer 1 can provide a change summary and a draft parameter impact report. and The numerical updates are still performed according to the established formula and threshold.

[0115] (ix) Regarding the resource controller: Figure 3 A schematic diagram of the resource controller's workflow is shown. The resource controller aims to optimize the allocation of computing resources.

[0116] 1. Calculation Budget Setting: Set a weekly calculation budget of 1010 (e.g., the maximum total CPU time) and a single-cycle budget of 1020 (e.g., the maximum self-help intensity B). ).

[0117] 2. Alternative strategy when hitting the upper limit 1030: When B≥ When this happens, the system will activate alternative strategy S, for example, forcing B to be set to... Alternatively, when the cumulative computational load approaches the budget, the uncertainty threshold δ may be temporarily increased, or compensation may be delayed.

[0118] (X) Abnormalities and Safety Red Lines: Global cap for high-risk groups: For special groups, the system will set a stricter global cap, such as a lower Intensity_max, a lower [missing information], etc. (Even 0), higher .

[0119] Safety invariant: Confidence level ≥ 95%, (e.g., 50), maximum delay compensation ≤ 7 days.

[0120] The "minimum action set" for the freeze policy: Once the freeze policy is triggered (e.g., user compliance rate is below 0.05 for N consecutive weeks), the freeze policy is triggered. ,or Higher than for N consecutive weeks If there are no valid records for N consecutive days, the system will enter a safe mode, allowing only: record prompts, data quality education, and medical / nutritionist advice; no further compensation calculations or automatic compensation actions will be performed, and no "strong" or "intense" level recommendations will be issued.

[0121] (xi) Explainability and traceability: The system outputs "decision explanation signals" to improve engineering controllability and compliance value. These explanation texts are preferably generated by the interpretability generation submodule in the AI ​​large model collaboration layer 1. For example: Reasons for raising the threshold: such as "due to low record integrity ( The threshold for food group Y has been raised by Z%.

[0122] Resource scheduling trigger reason: such as "Due to the shortage of computing resources, strategy A has been adopted in this assessment (e.g., reducing sampling intensity / delay compensation)".

[0123] Reasons for prohibition or limitation of compensation: such as "due to high uncertainty of assessment results ( This compensation is restricted / prohibited.

[0124] Furthermore, the specific workflow of this invention is as follows: Figure 7The overall system workflow diagram is shown. Below, using two complementary scenarios as examples, the overall workflow of the system provided in this embodiment will be explained.

[0125] Scenario 1: Normal recording, with partial missing data: The system starts and initializes, and the resource controller initializes the budget. Users record a week's worth of dietary data, with complete records for 5 days and missing records for 2 days. The data acquisition module acquires the data, and the data preprocessing module performs standardization processing.

[0126] In step S1, the integrity assessment module calculates the daily... Values, for example, are 0.85, 0.92, 0.78, 0.0, 0.88, 0.81, and 0.0.

[0127] In step S2, the dynamic threshold adjustment module adjusts according to... Value and segmented / double-slope mechanism adjust the effective intake threshold for each food group For example, for the vegetable group, its base threshold... For grams. In Date, adjusted effective intake threshold The weight was 319.8 grams, which triggered the slope of the medium-to-high integrity interval.

[0128] In step S3, the diversity assessment module calculates the desired food diversity score. It is 6.3.

[0129] In step S4, the bootstrap strength adjustment module first calculates the missing severity s as 0.3943, and then dynamically sets the bootstrap strength B to 322.3. At this point, B does not exceed... (For example).

[0130] In step S5, the uncertainty assessment module performs a Monte Carlo simulation using bootstrap strength B, and according to... Using the Percentile method, we calculate the confidence interval for food intake and its relative width. It is 0.25. Because... (0.25) is greater than the threshold δ (0.2), the permission gate fails, and the compensation amount is... The value is 0. The system outputs an explanation signal: "Due to high uncertainty in the evaluation results ( Since δ=0.25 > 0.2, this compensation is prohibited. In step S6, the suggestion generation module uses the lower bound of the confidence interval for food intake. With a weight of 450g and a deficiency severity (s) of 0.3943, the calculated intensity (Intensity) is 0.123. The system generates a "General Recommendation".

[0131] In step S7, the parameter update module adjusts the prior learning rate based on the user compliance rate ρ being 0.65 and the convergence of the food intake interval. Because the compliance rate was lower than the target, the learning rate remained unchanged.

[0132] Finally, the system saves the analysis results, updates the cumulative computation of the resource controller, and awaits the data for the next cycle.

[0133] Scenario 2: Shared family dining plates, low-confidence image, limited resources: The system starts and initializes; the resource manager initializes the budget. Assume the current device is a low-end mobile device. The system is configured with a weekly CPU budget of 20 seconds. Users record their weekly dietary data, mostly through image recognition. Due to shared family plates, the confidence scores for image recognition are generally low, with three days of records missing. In this scenario, the system utilizes an AI large-scale model collaboration layer for semantic standardization and knowledge alignment.

[0134] In step S1, the integrity assessment module calculates the daily... Values ​​such as 0.45, 0.52, 0.38, 0.0, 0.48, 0.41, 0.0 generally indicate low r_d values.

[0135] In step S2, the dynamic threshold adjustment module adjusts according to... Value and segmented / double-slope mechanism adjust the effective intake threshold for each food group For example, for the grain group, the base threshold It weighs 250 grams. Date, adjusted effective intake threshold The value was 360 grams, which triggered the low integrity interval slope. The system output an explanation signal: "Due to low record integrity ( The target threshold for the grain group has been raised by 44%. In step S3, the diversity assessment module calculates the desired food diversity score. It is 5.8.

[0136] In step S4, the bootstrap strength adjustment module first calculates the missing severity s as 0.55, and then dynamically sets the bootstrap strength B to 467.2. At this point, B does not exceed... Meanwhile, the resource controller checks if the cumulative CPU time for the week is close to the weekly CPU budget. Assuming it's nearing the limit, the resource controller triggers an alternative strategy, and the system decides to temporarily increase the uncertainty threshold δ from 0.2 to 0.25. The system outputs an explanation signal: "Due to limited computing resources, the uncertainty threshold δ has been temporarily relaxed to 0.25 for this assessment." In step S5, the uncertainty assessment module performs a Monte Carlo simulation using bootstrap strength B (467.2). Because... (4 days) Less and s(0.55) higher, the system adaptively selects the BCa method to calculate the confidence interval and its relative width of food intake. It is 0.23. Because... If (0.23) is less than or equal to the adjusted threshold δ (0.25), the gate passes. Assume recommended intake. It is 0 kcal, and the actual intake is... Calculate the compensation amount for 1 kcal. It contains 12 calories.

[0137] In step S6, the suggestion generation module uses the lower bound of the confidence interval for food intake. With a value of 1 kcal and a missing severity (s) of 0.55, the calculated intensity (Intensity) is 0.4. The system generates a "Strong Recommendation".

[0138] In step S7, the parameter update module adjusts the prior learning rate based on the user compliance rate ρ being 0.5 and the convergence of the food intake interval. Due to low compliance, the learning rate remained unchanged.

[0139] Finally, the system saves the analysis results, updates the cumulative computation of the resource controller, and awaits the data for the next cycle.

[0140] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications 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.

[0141] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and inventive 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 implementation should not be considered beyond the scope of this application.

[0142] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0143] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0144] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for scientific assessment and improvement of meal nutrition based on AI large-scale models, characterized in that, Includes the following steps: Step 1: Obtain and standardize dietary records, determine the number of items to be recorded each day and the credibility of each item based on the user's history, calculate the daily completeness credibility score, and archive the data to the shared intermediate data of the evaluation link after data structuring; Step 2: Establish basic intake thresholds for each food group, and generate daily effective thresholds according to the segmented double-slope rule of the integrity credibility score. The rule uses a larger adjustment coefficient in the low integrity range and a smaller adjustment coefficient in the high integrity range, and the effective thresholds do not increase monotonically with the score. Step 3: Estimate food diversity on a weekly scale. For each food group on missing days, combine the observational indications and prior occurrence probabilities and smooth them to obtain the weekly coverage probability. Combine the daily record completeness with the coverage probability and truncate the upper limit of each group to summarize the expected diversity score. Step 4: Determine the missing severity using the complement of the weekly mean of completeness, and set the bootstrap resampling intensity accordingly. Adaptively select the bootstrap interval type to obtain the intake confidence interval and its relative width. The selection depends on the number of observation days and the missing severity. Step 5: Perform dual-gated compensation and parameter update: When the interval width does not exceed the threshold, the recommended intake is higher than the actual intake, and the number of observation days reaches the preset lower limit, the compensation amount is determined according to the rule that it is proportional to the gap, decreases with uncertainty, and is constrained by the global upper limit; the suggestion strength is generated based on the mapping between the lower bound of the interval and the missing severity, and the prior update rate is adjusted according to the compliance rate and the interval convergence to update the prior occurrence probability parameter.

2. The method for scientific assessment and improvement of meal nutrition based on AI large model according to claim 1, characterized in that: The integrity credibility score is determined by the number of food items to be recorded and the confidence level of the recorded items; the number of food items to be recorded is calculated based on the user's average daily record volume within a preset time window in the past; the confidence level is given by the recording method and can be revised by the user, and the scoring weight is set in a configurable range and bound to the record source.

3. The method for scientific assessment and improvement of meal nutrition based on AI large model according to claim 2, characterized in that: The segmented dual-slope rule divides the scoring interval into two segments, low integrity and high integrity, using a preset breakpoint. The low integrity segment uses a first slope coefficient, while the high integrity segment uses a second slope coefficient that is smaller than the first slope coefficient. The effective intake threshold remains monotonically non-increasing relative to the score, and the basic threshold and breakpoint parameters are recorded for each food group.

4. The method for scientific assessment and improvement of meal nutrition based on an AI large model according to any one of claims 1-3, Its characteristics are: The prior occurrence probability is initialized hierarchically according to population statistics, including age, gender, region and dietary preference dimensions; the smoothing coefficient is selected within a preset range and used to limit the magnitude of prior variation; the truncation process limits the upper limit of a single food group in the week and stores it in a unified structured format for subsequent updates.

5. The method for scientific assessment and improvement of meal nutrition based on AI large model according to any one of claims 4, characterized in that: The severity of the missing data is calculated based on the mean of the daily integrity confidence scores over a week, and its complement is calculated accordingly. The intensity of the self-service resampling is determined by a power function mapping between the minimum and maximum number of resampling attempts, and rounded up to the nearest integer. The relevant mapping coefficients and value boundaries are registered in the metadata for auditing and recalculation, and the creator, creation time, and applicable version range are recorded on the parameter server.

6. The method for scientific assessment and improvement of meal nutrition based on an AI large model according to any one of claims 1, characterized in that, The confidence interval calculation method is switched according to rules: when the number of observation days is less than the minimum observation threshold and the missing severity is not lower than the preset threshold, the BCa method is used; otherwise, the percentile method is used. The minimum observation threshold and the missing threshold are configurable, and the switching record is archived as an auxiliary parameter for interval calculation and written into the interval calculation log to solidify the method, threshold and result summary.

7. The method for scientific assessment and improvement of meal nutrition based on an AI large model according to claim 5, characterized in that, The dual-gating compensation includes a permission gate and a quota gate: the permission gate is determined based on the relative width, the relationship between the recommended intake and the actual intake, and the satisfaction of the number of observation days, and verifies the consistency of the minimum number of observation days and the data batch; when the permission gate passes, the quota gate calculates the compensation amount according to the gap ratio and in combination with the uncertainty adjustment coefficient and the compensation upper limit, and the output value is rounded to the specified precision and includes the unit.

8. The method for scientific assessment and improvement of meal nutrition based on an AI large model according to claim 7, characterized in that, The recommended strength is obtained by a two-factor weighted mapping, the two factors being the ratio of the lower bound of the intake range to the recommended intake and the severity of the deficiency; Weight coefficients are configured between zero and one and stored in key-value format; the recommended level is generated by mapping from a preset threshold table, which contains the version and effective time, and the threshold table and weights are stored independently to support separate maintenance and rollback.

9. The method for scientific assessment and improvement of meal nutrition based on an AI large model according to claim 1, characterized in that, The adaptive update of prior parameters adopts a multiplicative learning rate scheme. The learning rate is jointly adjusted by the user compliance rate and the convergence index of the intake interval. The learning rate is set with upper and lower boundaries and includes a linear combination coefficient. The update rhythm is executed weekly. During the update, the new and old priors are smoothed with an exponential moving average and the timestamp is recorded. A clamping strategy is set for abnormal learning rate changes and the reason code is registered.

10. A system for scientific assessment and improvement of meal nutrition based on an AI large-scale model, used to implement the method described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire dietary records formed by manual input and image recognition; The data preprocessing module, connected to the data acquisition module, is used for field validation, format standardization and alias mapping, and to generate record credibility. The integrity assessment module, connected to the data preprocessing module, is used to determine the recording benchmark based on user history and calculate the daily integrity credibility score by combining the credibility of each record. The dynamic threshold adjustment module, connected to the integrity assessment module, is used to read the basic intake threshold for each food group and adjust the effective threshold for the day according to the integrity confidence score based on the segmented double slope rule. The low integrity segment uses the first slope, the high integrity segment uses the second slope which is less than the first slope, and the effective threshold remains monotonically non-increasing relative to the score. The diversity assessment module, connected to the data preprocessing module and the dynamic threshold adjustment module, is used to smooth the comprehensive observation indicators and prior occurrence probabilities of food groups on missing days at a weekly granularity to obtain the weekly coverage probability, and to summarize the expected diversity score after truncating each food group in combination with the completeness of daily records. The self-help intensity adjustment module, connected to the diversity assessment module and the integrity assessment module, is used to calculate the missing severity based on the mean of the integrity confidence score within one week, and map the missing severity to the minimum and maximum number of resampling times to set the self-help resampling intensity, which is rounded to an integer; and selects the BCa interval method when the number of observation days is insufficient and the missing severity is not lower than the threshold, otherwise selects the percentile method. The uncertainty assessment and compensation module, connected to the bootstrap intensity adjustment module, is used to generate an intake confidence interval and calculate the relative width based on the resampling intensity; and to perform dual-gating compensation: the permission gate requires that the relative width is not greater than a threshold, the recommended intake is greater than the actual intake, and the number of observation days is not less than the lower limit; when the permission gate is satisfied, the quota gate determines the compensation amount proportionally to the gap, decreasing with uncertainty and constrained by the compensation upper limit. The suggestion generation module, connected to the uncertainty assessment and compensation module, is used to perform a weighted mapping of the ratio of the lower bound of the intake range to the recommended intake and the severity of the deficiency to obtain the suggestion strength. The weights are between zero and one, and the suggestion level is mapped by a threshold table. The parameter update module, connected to the diversity assessment module and the suggestion generation module, is used to update the prior parameters according to the weekly rhythm using a multiplicative learning rate. The learning rate is jointly adjusted by the user compliance rate and the convergence of the intake interval and is set with upper and lower boundaries. During the update, an exponential moving average smoothing is performed on the new and old priors. The resource controller, connected to the uncertainty assessment and compensation module and the self-help intensity adjustment module, is used to set the calculation budget and single-cycle upper limit, and when the upper limit is reached or close to the budget, to cap the resampling intensity, adjust the interval threshold, or delay compensation. The large model collaboration layer interacts with the data preprocessing module, the diversity assessment module, the uncertainty assessment and compensation module, the suggestion generation module, and the parameter update module in a candidate / draft manner. It is used to perform semantic standardization and knowledge retrieval alignment, generate structured candidates and explanation drafts, and does not directly modify the numerical parameters and thresholds confirmed by each algorithm module.

Images