Customer health data intelligent management system and method

Abstract

The invention discloses a customer health data intelligent management system and method, and belongs to the technical field of health data management. Basic data and health data are collected; the client inputs a daily menu and calculates nutrients; establishing a daily nutrient matrix based on nutrients, and identifying a diet mode; defining a taste space, and extracting taste features based on the seasoning use condition of the menu every day; integrating to form an input vector; establishing a personalized nutrition demand prediction model, and training to obtain nutrition constraints; retrieving the candidate recipes based on nutrition constraints, and performing taste matching to obtain recommended recipes; based on the menu selection records of the similar clients, the recommended menu is optimized; dynamically adjusting the recommendation strategy according to the score of the customer for the recommended menu to obtain an optimal menu; when the customer eats the optimal menu, the change of health data before and after diet intervention is calculated, and the intervention effect is quantified; nutrients needed by the client are adjusted based on the intervention effect, and the optimal menu is correspondingly adjusted.

Description

Technical Field

[0001] The present invention relates to the technical field of health data management, and in particular to an intelligent management system and method for customer health data. Background Art

[0002] As people's health awareness continues to rise, health management has gradually become an important part of life. The popularity of wearable devices, home health monitoring instruments, and the widespread use of various diet recording apps have led to an explosive growth in personal health and diet data. At the same time, the rapid development of artificial intelligence technology has provided technical support for the intelligent management of health data.

[0003] In existing technologies, users' health data are scattered across different devices and platforms, and the formats of various data sources are not unified, which makes data integration difficult and makes it difficult to form a complete user health portrait; most nutrition recommendation systems are based on general nutrition standards or simple user basic information, and do not fully consider the individual health status, eating patterns and taste preferences of users, and the recommendation results lack specificity; traditional health intervention effect evaluation mostly uses subjective questionnaires or regular physical examinations, lacks quantitative, data-based evaluation methods, and cannot accurately measure the relationship between dietary intervention and changes in health indicators. It is difficult to judge the effectiveness of intervention measures. Even if the intervention effect is found to be unsatisfactory, it is impossible to quickly and automatically adjust the nutrition recommendation strategy. Summary of the Invention

[0004] The purpose of the present invention is to provide a customer health data intelligent management system and method to solve the problems raised in the prior art.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] In a first aspect, the present application provides a method for intelligently managing customer health data, comprising the following steps:

[0007] Basic data is collected through registration forms, health data management cycles are set, and health monitoring equipment is used to monitor customers and obtain health data; customers enter their daily menus and nutrients are calculated;

[0008] A daily nutrient matrix is ​​established based on nutrients, and dietary patterns are identified using the K-means clustering algorithm. A taste space is defined, and taste characteristics are extracted based on the seasoning usage of daily recipes. Basic data, health data, and dietary patterns are integrated to form an input vector.

[0009] All numerical features are Z-score standardized to eliminate the impact of dimension. Relative value features such as the nutrient ratio of the dietary pattern are Min-Max normalized to scale the value range to the [0,1] interval, and basic data, health data, and dietary patterns are integrated. The integrated input vector is in the form of a multidimensional feature vector with the following structure: InputVector = [basic data features; health data features; dietary pattern features; position encoding]; the position encoding uses the sine-cosine position encoding of the Transformer architecture to capture the temporal characteristics of the data.

[0010] A personalized nutritional needs prediction model was built using the Transformer architecture. The dataset was divided for training, and the daily nutrient intake range was predicted based on the input vector, serving as a nutritional constraint. Based on the nutritional constraint, candidate recipes were retrieved, and the cosine similarity between the candidate recipes' flavor profiles and the customer's was calculated to perform taste matching and generate recommended recipes.

[0011] Based on the recipe selection records of similar customers, recommended recipes are optimized; using the Q-Learning algorithm, the recommendation strategy is dynamically adjusted according to the customer's rating of the recommended recipe to obtain the optimal recipe; when the customer consumes the optimal recipe, the Pearson correlation coefficient is used to calculate the changes in health data before and after the dietary intervention to quantify the intervention effect; based on the intervention effect, the nutrients required by the customer are adjusted and the optimal recipe is adjusted accordingly.

[0012] In conjunction with the first aspect, in a first implementation of the first aspect of the present application, collecting basic data through a registration form, setting a health data management cycle, and using a health monitoring device to monitor the customer to obtain health data include:

[0013] Wearable device data is collected in seconds and synchronized to the cloud every minute; for home device data, users are advised to measure at fixed times and automatically identify valid measurement periods; for physical examination reports and laboratory test data, the data validity period is automatically marked; health data is timestamp aligned and data normalized. For wearable device data, by identifying the valid measurement period, the data within the period is assigned a corresponding timestamp; for physical examination reports and laboratory test data, the data validity period is marked and the timestamp is determined according to the report generation time; health data is arranged in chronological order through a unified time base to achieve timestamp alignment; Z-score normalization is used to standardize health data; the wearable device data includes heart rate, blood oxygen saturation, blood pressure, body temperature, calorie consumption and sleep data.

[0014] In combination with the first aspect, in a second embodiment of the first aspect of the present application, the customer inputs a daily menu and calculates nutrients, including:

[0015] Customers input their daily menus in text format, and NLP technology is used to automatically identify ingredient names and weight descriptions, matching them with standard ingredient units in the database. For photos of recipes taken, the YOLO algorithm is used to identify the main ingredients, and OCR technology is used to parse the recipe information on tableware or packaging. The app supports importing historical recipe data from the diet record app, and achieves structured data migration through the API interface to obtain recipe information. Nutrients are calculated for each ingredient, complex dishes are broken down into main ingredients, side dishes, and seasonings, the nutritional contribution of each part is calculated separately, and the nutrient absorption rate is adjusted according to the cooking method.

[0016] In combination with the first aspect, in a third embodiment of the first aspect of the present application, establishing a daily nutrient matrix based on nutrients and identifying dietary patterns using a K-means clustering algorithm include:

[0017] The elbow rule is used to calculate the moment of inertia, and the K value at the inflection point of the curve is selected; K non-repeated single-day data points are randomly selected as the initial centroids, and the K-means algorithm is used to calculate the Euclidean distance between each single-day data point and the K centroids, and assign them to the closest cluster; the mean vector of all data points in each cluster is calculated as the new centroid; the mean nutrient value of each cluster is calculated, compared with the recommended value in the dietary guidelines, and significant difference indicators are extracted; the frequency of occurrence of each cluster label for each user is counted, and dietary patterns are identified for verification and optimization.

[0018] In combination with the first aspect, in a fourth implementation of the first aspect of the present application, defining a taste space and extracting taste characteristics based on the use of seasonings in daily recipes includes:

[0019] We defined basic taste dimensions, including saltiness, sweetness, sourness, spiciness, umami, and bitterness. We used a Likert scale to quantify each dimension, representing each recipe as a multidimensional taste vector. We constructed a 3D taste space, where each recipe corresponds to a point in the space. Different dietary patterns form clustered distributions in the taste space.

[0020] A database covering all condiments is established, and the taste contribution weights of the condiments are normalized to ensure that the sum of the weights of each dimension is 1. The usage frequency data is updated regularly, and the weights are dynamically adjusted according to newly entered recipes. For each recipe, the taste characteristics are calculated based on the condiments used and their amounts. A taste perception threshold is set, and the dimension value of the taste characteristics below the threshold is set to 0. For compound condiments, the recipe decomposition method is used to calculate the taste contribution of each ingredient separately, and then the total taste characteristics are obtained by superposition.

[0021] In combination with the first aspect, in a fifth embodiment of the first aspect of the present application, the personalized nutritional requirement prediction model is established using a Transformer architecture, the data set is divided for training, and the daily nutrient intake range is predicted based on the input vector as a nutritional constraint, including:

[0022] In the input and embedding layers, for input vectors, numerical features are directly input, categorical features are converted to continuous vectors using a trainable embedding matrix, all features are merged and positional encoding is added to form the final input vector. In the encoder layer, the number of heads and the dimensions of each head are set, the query, key, and value matrices are calculated, and the attention score is calculated. For the two fully connected layers, the activation function is GELU. Residual connections are added after each sublayer, and layer normalization is performed to stabilize training. In the output layer, the time dimension is compressed through global average pooling, and the fully connected layer regresses to predict the nutrient intake range.

[0023] The dataset was stratified by health goals and BMI levels to ensure consistent distribution across subsets. During training, Huber Loss was used to balance the influence of outliers, the AdamW optimizer was used, and a cosine annealing strategy was used for the learning rate. Dropout was added to the encoder layer to prevent overfitting, the weight matrix was initialized using Xavier, and layer normalization was added before the activation function. Model training and validation were performed to predict the daily nutrient intake range as a nutritional constraint.

[0024] In combination with the first aspect, in a sixth implementation of the first aspect of the present application, the Q-Learning algorithm is used to dynamically adjust the recommendation strategy based on the customer's rating of the recommended recipe to obtain the optimal recipe, including:

[0025] Based on health data, we extract BMI, body fat percentage, blood pressure, and average sleep duration in the past week as health features, dietary patterns and taste characteristics as dietary features, and the time, scene, and season of the customer's meal as contextual features. We then perform bucketing on continuous features and one-hot encoding on categorical features to form discrete feature vectors.

[0026] Each candidate recipe is mapped to a unique action ID, and the customer's rating of the recommended recipe is mapped to a standardized reward ranging from -1 to 1. Delayed rewards are set based on the next morning's weight change, blood pressure fluctuations, and improvement in nutrient compliance. A total reward is set based on the standardized rewards and the delayed rewards. Failure to adopt a recommended recipe or triggering a dietary risk rule is included in the penalty mechanism.

[0027] Initialize the Q-table and use the ε-greedy algorithm to randomly select candidate recipes and choose the action with the largest Q-value. When the Q-values ​​are the same, select from high to low according to the following priorities: small historical score standard deviation, short cooking time, and low ingredient repetition rate. Update the Q-value. Record the status, action, user score, and next-day health data of each recommendation, filter out invalid records that have not been rated by the user and have no health changes, perform real-time updates and offline batch training. Select the action A* with the highest Q-value as the optimal recipe.

[0028] In combination with the first aspect, in a seventh embodiment of the first aspect of the present application, when the customer consumes the optimal recipe, the Pearson correlation coefficient is used to calculate the change in health data before and after the dietary intervention to quantify the intervention effect, including:

[0029] In the Pearson correlation coefficient, the independent variable X is set as the actual daily nutrient intake during the intervention period, the dependent variable Y is set as the change in health data before and after the intervention, the null hypothesis H0 is set as no significant linear correlation between X and Y, and the alternative hypothesis H1 is set as a significant linear correlation between X and Y. The correlation coefficient ρ is calculated using the following formula:

[0030]

[0031] Where n is the number of data points in the dataset, X i is the nutrient intake on day i, represents the mean of X, Y i is the health data on day i, is the mean of Y;

[0032] Perform a significance test and calculate the statistic t. The formula is:

[0033]

[0034] Look up the table to determine whether t exceeds the critical value and test significance;

[0035] According to the absolute value of the correlation coefficient ρ, the intervention effect is divided into different levels to clarify the degree of correlation between dietary intervention and health data.

[0036] In combination with the first aspect, in an eighth embodiment of the first aspect of the present application, adjusting the nutrients required by the client based on the intervention effect and adjusting the optimal recipe accordingly include:

[0037] Based on the intervention effect, an upward adjustment strategy is implemented for positively correlated indicators, and a downward adjustment strategy is implemented for negatively correlated indicators. The adjustment range is controlled, and user feedback of discomfort triggers a rollback operation; the adjusted nutrient range is used as a new constraint condition, the personalized nutritional demand prediction model is updated, and the nutritional constraints are regenerated; for recipes that meet the adjusted nutritional range, a reward value is given in the Q-Learning algorithm according to the correlation strength; the adjustment effect is verified to obtain the optimal adjusted recipe.

[0038] In a second aspect, the present application provides a customer health data intelligent management system, comprising:

[0039] Health data management and nutrient calculation module: includes: basic data collection unit, health data collection unit and nutrient calculation unit; the basic data collection unit collects basic data through the registration form, the health data collection unit sets the health data management cycle, uses health monitoring equipment to monitor customers and obtain health data; the nutrient calculation unit calculates nutrients based on the customer's daily menu input;

[0040] Input vector generation module: This module includes a dietary pattern recognition unit, a taste feature extraction unit, and an input vector integration unit. The dietary pattern recognition unit establishes a daily nutrient matrix based on nutrients and identifies dietary patterns using a K-means clustering algorithm. The taste feature extraction unit defines a taste space and extracts taste features based on the seasoning usage of daily recipes. The input vector integration unit integrates basic data, health data, and dietary patterns to form an input vector.

[0041] The recommended recipe generation module includes a nutritional constraint calculation unit and a recommended recipe generation unit. The nutritional constraint calculation unit uses the Transformer architecture to establish a personalized nutritional requirement prediction model, divides the data set for training, and predicts the daily nutrient intake range based on the input vector as a nutritional constraint. The recommended recipe generation unit retrieves candidate recipes based on the nutritional constraints, calculates the cosine similarity between the taste characteristics of the candidate recipes and the customer's taste characteristics, performs taste matching, and obtains recommended recipes.

[0042] Optimal recipe generation module: includes: recommended recipe optimization unit, optimal recipe generation unit, intervention effect quantification unit and optimal recipe adjustment unit; among them, the recommended recipe optimization unit optimizes the recommended recipes based on the recipe selection records of similar customers; the optimal recipe generation unit adopts the Q-Learning algorithm to dynamically adjust the recommendation strategy according to the customer's rating of the recommended recipe to obtain the optimal recipe; when the customer consumes the optimal recipe, the intervention effect quantification unit uses the Pearson correlation coefficient to calculate the changes in health data before and after the dietary intervention and quantify the intervention effect; the optimal recipe adjustment unit adjusts the nutrients required by the customer based on the intervention effect and adjusts the optimal recipe accordingly.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. This invention collects wearable device data in seconds and synchronizes it to the cloud every minute. It also supports automatic identification of effective measurement periods for home devices and automatic marking of validity periods for physical examination report data. In addition, it uses NLP technology, the YOLO algorithm, and OCR technology to achieve automatic recognition and extraction of multimodal recipe information, solving the problem of data fragmentation.

[0045] 2. This invention identifies user dietary patterns based on the K-means clustering algorithm, extracts taste features based on the defined multidimensional taste space, and uses the Transformer architecture to establish a personalized nutritional demand prediction model. It fully considers multi-dimensional information such as user basic data, health data, dietary patterns, and taste preferences to achieve accurate nutritional demand prediction and personalized recipe recommendations.

[0046] 3. The present invention uses the Pearson correlation coefficient to calculate the changes in health data before and after dietary intervention, and quantifies the association between dietary intervention and health indicators through statistical testing and correlation strength grading; based on the intervention effect, the recommended range of nutrients required by the customer is automatically adjusted, the personalized nutritional demand prediction model is updated, and the optimal recipe recommendation strategy is adjusted accordingly to achieve dynamic optimization of health management. Compared with existing technologies, the efficiency and effectiveness of health intervention are significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a schematic diagram of the steps of a customer health data intelligent management system and method of the present invention;

[0048] Figure 2 This is a system structure diagram of a customer health data intelligent management system and method of the present invention. DETAILED DESCRIPTION

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0050] Example: Figure 1-Figure 2 As shown, the present invention provides a technical solution.

[0051] like Figure 1 As shown in the schematic diagram of a customer health data intelligent management system and method, the present application provides a customer health data intelligent management method, including the following steps:

[0052] Step S100: Collect basic data through the registration form, set the health data management cycle, use health monitoring equipment to monitor the customer and obtain health data; the customer enters the daily menu and calculates nutrients;

[0053] Specifically, wearable device data is collected in seconds and synchronized to the cloud every minute; for home device data, users are advised to measure at fixed times and automatically identify valid measurement periods; for physical examination reports and laboratory test data, the data validity period is automatically marked; health data is timestamp aligned and data standardized. For wearable device data, by identifying the valid measurement period, the data within the period is assigned a corresponding timestamp; for physical examination reports and laboratory test data, the data validity period is marked, and the timestamp is determined according to the report generation time; through a unified time base, health data is arranged in chronological order to achieve timestamp alignment; Z-score standardization is used to standardize health data; the wearable device data includes heart rate, blood oxygen saturation, blood pressure, body temperature, calorie consumption and sleep data.

[0054] Furthermore, customers can input their daily menus in text, and the NLP technology will automatically identify the names and weight descriptions of the ingredients and match them with the standard ingredient units in the database. For the recipe photos taken, the YOLO algorithm will be used to identify the main ingredients, and the OCR technology will be combined to parse the recipe information on the tableware or packaging. It supports importing historical recipe data from the diet record APP, and realizes structured data migration through the API interface to obtain recipe information. Nutrients are calculated for each ingredient, and complex dishes are broken down into main ingredients, side dishes and seasonings. The nutritional contribution of each part is calculated separately, and the nutrient absorption rate is adjusted according to the cooking method.

[0055] Step S200: Establish a daily nutrient matrix based on nutrients and identify dietary patterns using a K-means clustering algorithm; define a taste space and extract taste features based on the seasoning usage of daily recipes; integrate basic data, health data, and dietary patterns to form an input vector;

[0056] Specifically, the elbow rule is used to calculate the moment of inertia, and the K value at the inflection point of the curve is selected; K non-repeated single-day data points are randomly selected as the initial centroids, and the K-means algorithm is used to calculate the Euclidean distance between each single-day data point and the K centroids, and assign them to the closest cluster; the mean vector of all data points in each cluster is calculated as the new centroid; the mean nutrient value of each cluster is calculated, compared with the recommended value in the dietary guidelines, and significant difference indicators are extracted; the frequency of occurrence of each cluster label for each user is counted, and the dietary pattern is identified for verification and optimization.

[0057] Furthermore, we defined basic taste dimensions, including saltiness, sweetness, sourness, spiciness, umami, and bitterness. We used a Likert scale to quantify each dimension, representing each recipe as a multidimensional taste vector. A 3D taste space was constructed, with each recipe corresponding to a point in the space. Different dietary patterns formed clustered distributions in the taste space.

[0058] A database covering all condiments is established, and the taste contribution weights of the condiments are normalized to ensure that the sum of the weights of each dimension is 1. The usage frequency data is updated regularly, and the weights are dynamically adjusted according to newly entered recipes. For each recipe, the taste characteristics are calculated based on the condiments used and their amounts. A taste perception threshold is set, and the dimension value of the taste characteristics below the threshold is set to 0. For compound condiments, the recipe decomposition method is used to calculate the taste contribution of each ingredient separately, and then the total taste characteristics are obtained by superposition.

[0059] In a specific embodiment, customer C002 records nutrient intake data daily during a 30-day health data management cycle. Their dietary patterns are identified using the K-means algorithm: K values ​​ranging from 1 to 10 are tried, and the moment of inertia (the sum of the squared distances from points within the cluster to the centroid) of the clustering results under different K values ​​is calculated. For example, when K = 1, all data is grouped together, and the moment of inertia is 2856.3; when K = 2, the moment of inertia drops to 1234.5; and when K = 3, the moment of inertia is 654.2. A curve showing the moment of inertia changing with K value is plotted, and it is found that when K = 3, the curve has a clear inflection point, and the magnitude of the decrease in the moment of inertia decreases significantly. Therefore, K = 3 is determined to be the optimal number of clusters.

[0060] Randomly select three non-repeating single-day data points as the initial centroids. The nutrient data for day 5 (85g protein, 60g fat, 200g carbohydrates), day 12 (50g protein, 30g fat, 150g carbohydrates), and day 20 (120g protein, 80g fat, 250g carbohydrates) are selected as the initial centroids. Calculate the Euclidean distances between the remaining 27 days' daily nutrient data and the three centroids. For example, the Euclidean distances between the nutrient data for day 3 (70g protein, 50g fat, 180g carbohydrates) and the three centroids are:

[0061] The distance from centroid 1 (day 5) is 25.49, the distance from centroid 2 (day 12) is 36.06, and the distance from centroid 3 (day 20) is 91.65. Since the distance from centroid 1 is the smallest, the data on day 3 is assigned to the cluster where centroid 1 is located.

[0062] The mean vector of all data points within each cluster is calculated as the new centroid. For example, after the first round of assignment, the cluster containing centroid 1 contains 10 data points. Its mean is recalculated as (82g protein, 58g fat, 195g carbohydrates) and used as the new centroid 1. Repeat the data point assignment and centroid update steps. After five iterations, the centroid position stabilizes and the clustering results converge.

[0063] The mean nutrient values ​​for each cluster were calculated and compared with the recommended values ​​in the Dietary Guidelines. Cluster 1 had a mean nutrient value of (80g protein, 55g fat, 200g carbohydrates), Cluster 2 had a mean nutrient value of (55g protein, 35g fat, 160g carbohydrates), and Cluster 3 had a mean nutrient value of (110g protein, 75g fat, 240g carbohydrates). Comparison with the recommended values ​​(55-75g protein, 20-30g fat, 250-300g carbohydrates) revealed that: Cluster 1 had a high protein intake, while fat and carbohydrate intake were close to the recommended values; Cluster 2 had a low overall intake; and Cluster 3 had significantly high protein and fat intake, while low carbohydrate intake.

[0064] Based on these differences, we labeled the dietary patterns of the three clusters: Cluster 1: High-protein balanced; Cluster 2: Low-intake; Cluster 3: High-fat, high-protein, low-carb. Counting the frequency of each cluster label across the 30-day data revealed that "High-protein balanced" appeared 15 times (50%), "Low-intake" appeared 8 times (26.7%), and "High-fat, high-protein, low-carb" appeared 7 times (23.3%). Therefore, we determined that customer C002's primary dietary pattern was "High-protein balanced."

[0065] Six basic taste dimensions were defined: saltiness, sweetness, sourness, spiciness, umami, and bitterness, and quantified using a five-point Likert scale. Each recipe was mapped to a point in the three-dimensional space, with saltiness, sweetness, and sourness as the axes. For example, the taste vector for the recipe "Tomato and Egg Soup" is (Saltiness 3, Sweetness 2, Sourness 3, Spiciness 1, Umami 4, Bitterness 1), corresponding to the coordinate point (3, 2, 3) in 3D space.

[0066] A database of 200 condiments was created, recording the weight of each condiment's contribution to each flavor dimension. For example, salt contributes a weight of 0.8 to saltiness and 0 to all other dimensions; sugar contributes a weight of 0.9 to sweetness. The condiment weights were normalized to ensure that the sum of the weights for each dimension was 1. The recipe and weights for a composite condiment called "doubanjiang" are: chili peppers (spiciness weight 0.6), salt (saltiness weight 0.3), and fermented beans (umami weight 0.1).

[0067] For example, let's take customer C002's "Mapo Tofu" lunch on day 10. The ingredients used were 300g tofu, 20g fermented black bean paste (doubanjiang), and 3g salt. The calculation process is as follows: Doubanjiang contribution: Spiciness = 20 × 0.6 = 12, Saltiness = 20 × 0.3 = 6, Umami = 20 × 0.1 = 2; Salt contribution: Saltiness = 3 × 0.8 = 2.4; Total taste characteristics: Spiciness = 12, Saltiness = 6 + 2.4 = 8.4, Sweetness = 0, Sourness = 0, Umami = 2, Bitterness = 0. The taste perception threshold is set to 3, and dimensions below the threshold are set to 0. The final taste vector is (Spiciness 4, Saltiness 3, Sweetness 0, Sourness 0, Umami 1, Bitterness 0).

[0068] The basic data of customer C002 (age 35, male, BMI 24), health data (blood pressure 120 / 80 mmHg, body fat percentage 20%), the identified main dietary pattern "high-protein balanced" (coded as 01), and the taste feature vector (4, 3, 0, 0, 1, 0) are integrated to form the final input vector: [35, 1, 24, 120, 80, 20, 01, 4, 3, 0, 0, 1, 0], which is used for subsequent personalized nutritional needs prediction model training and recommendation.

[0069] Step S300: A personalized nutritional needs prediction model is established using the Transformer architecture. The data set is divided for training. The daily nutrient intake range is predicted based on the input vector as a nutritional constraint. Candidate recipes are retrieved based on the nutritional constraint. The cosine similarity between the taste characteristics of the candidate recipes and the customer's taste characteristics is calculated to perform taste matching and obtain recommended recipes.

[0070] Specifically, in the input and embedding layers, for the input vector, numerical features are directly input, categorical features are converted to continuous vectors through a trainable embedding matrix, all features are merged and positional encoding is added to form the final input vector; in the encoder layer, the number of heads and the dimensions of each head are set, the query, key, and value matrices are calculated, and the attention score is calculated; for the two fully connected layers, the activation function is GELU; residual connections are added after each sublayer, and layer normalization and stabilization training are performed; in the output layer, the time dimension is compressed through global average pooling, and the fully connected layer regresses to predict the nutrient intake range;

[0071] The dataset was stratified by health goals and BMI levels to ensure consistent distribution across subsets. During training, Huber Loss was used to balance the influence of outliers, the AdamW optimizer was used, and a cosine annealing strategy was used for the learning rate. Dropout was added to the encoder layer to prevent overfitting, the weight matrix was initialized using Xavier, and layer normalization was added before the activation function. Model training and validation were performed to predict the daily nutrient intake range as a nutritional constraint.

[0072] In one embodiment, categorical feature embedding is performed, converting gender (0→1), BMI level (2→5-dimensional vector), and dietary pattern (1→8-dimensional vector) into continuous vectors using a trainable embedding matrix. For time series data, specifically daily input vectors, sine / cosine positional encoding is added.

[0073] A four-head attention mechanism is set up, with each head dimension of 8 (a total of 4 × 8 = 32). The query matrix Q, key matrix K, and value matrix V are all 32 × 32. The attention score of one head is calculated to be -0.077. After Softmax normalization, the attention weight of that head is obtained and used to weight the sum of the value vector.

[0074] Loss function: Huber Loss (δ = 1.0) to mitigate the influence of outliers. Optimizer: AdamW (β1 = 0.9, β2 = 0.999, weight decay 0.01). Learning rate strategy: Cosine annealing, initial learning rate 5e-4, cycle 20 epochs, minimum learning rate 1e-5. Regularization: Encoder layer Dropout = 0.1, weights initialized with Xavier uniform distribution.

[0075] After training, the input features are: BMI = 33.2, body fat percentage 32%, diet pattern "high sugar and high fat", and taste sweetness 3. Model output: protein intake range predicted to be 55-75g / day (actual recommended value: protein requirement for women in the fat loss period is approximately 1.2-1.6g / kg body weight, i.e. 85kg x 1.2 = 102g, but the range is narrowed after the model is adjusted based on individual metabolic data, possibly because the client's renal function indicators are implicit in the input vector). Training set MAE = 3.2g, validation set MAE = 4.1g (better than the traditional linear regression model MAE = 6.5g). R 2 =0.89, indicating that the model explains 89% of the variation in protein demand.

[0076] According to the model prediction, the daily nutrient constraints of customer C003 are: energy: 1800-2000kcal, protein: 55-75g, fat: ≤50g (fat loss target), carbohydrates: 200-250g, sodium: ≤2000mg.

[0077] Recipes that meet the nutritional constraints were screened from a recipe library (1,000 dishes), resulting in 200 candidate recipes. Customer C003's taste vector is [Saltiness 4, Sweetness 3, Sourness 2, Spiciness 2, Umami 3, Bitterness 1] (normalized to [0, 1]). For example, for recipe A: Scrambled Eggs with Tomatoes, the taste vector is [Saltiness 0.6, Sweetness 0.3, Sourness 0.2, Spiciness 0, Umami 0.5, Bitterness 0], with a similarity of 0.91. Sorted in descending order of similarity, the top three recommended recipes are: Scrambled Eggs with Tomatoes (similarity 0.91, 18g protein per serving, meeting the nutritional constraints), Stir-fried Vegetables (similarity 0.88, 5g protein per serving, requires a protein source), and Chicken Breast Salad (similarity 0.85, 25g protein per serving, 8g fat per serving, meeting fat loss requirements).

[0078] After customer C003 followed the recommended recipe for 4 weeks, his body fat percentage dropped by 1.5% and his protein intake rate increased from 60% to 85%, verifying the effectiveness of the model's predictions.

[0079] Step S400: Based on the recipe selection records of similar customers, the recommended recipes are optimized; using the Q-Learning algorithm, the recommendation strategy is dynamically adjusted according to the customer's rating of the recommended recipes to obtain the optimal recipe; when the customer consumes the optimal recipe, the Pearson correlation coefficient is used to calculate the changes in health data before and after the dietary intervention to quantify the intervention effect; based on the intervention effect, the nutrients required by the customer are adjusted and the optimal recipe is adjusted accordingly.

[0080] Specifically, based on health data, we extract BMI, body fat percentage, blood pressure, and average sleep duration in the past week as health features, dietary patterns and taste characteristics as dietary features, and the time, scene, and season of the customer's meal as contextual features. We then perform bucketing on continuous features and one-hot encoding on categorical features to form discrete feature vectors.

[0081] Each candidate recipe is mapped to a unique action ID, and the customer's rating of the recommended recipe is mapped to a standardized reward ranging from -1 to 1. Delayed rewards are set based on the next morning's weight change, blood pressure fluctuations, and improvement in nutrient compliance. A total reward is set based on the standardized rewards and the delayed rewards. Failure to adopt a recommended recipe or triggering a dietary risk rule is included in the penalty mechanism.

[0082] Initialize the Q-table and use the ε-greedy algorithm to randomly select candidate recipes and choose the action with the largest Q-value. When the Q-values ​​are the same, select from high to low according to the following priorities: small historical score standard deviation, short cooking time, and low ingredient repetition rate. Update the Q-value. Record the status, action, user score, and next-day health data of each recommendation, filter out invalid records that have not been rated by the user and have no health changes, perform real-time updates and offline batch training. Select the action A* with the highest Q-value as the optimal recipe.

[0083] Furthermore, in the Pearson correlation coefficient, the independent variable X is set as the actual daily nutrient intake during the intervention period, the dependent variable Y is set as the change in health data before and after the intervention, the null hypothesis H0 is set as no significant linear correlation between X and Y, and the alternative hypothesis H1 is set as a significant linear correlation between X and Y; the correlation coefficient ρ is calculated using the following formula:

[0084]

[0085] Where n is the number of data points in the dataset, X i is the nutrient intake on day i, represents the mean of X, Y i is the health data on day i, is the mean of Y;

[0086] Perform a significance test and calculate the statistic t. The formula is:

[0087]

[0088] Look up the table to determine whether t exceeds the critical value and test significance;

[0089] According to the absolute value of the correlation coefficient ρ, the intervention effect is divided into different levels to clarify the degree of correlation between dietary intervention and health data.

[0090] Furthermore, based on the intervention effect, an upward adjustment strategy is implemented for positively correlated indicators, and a downward adjustment strategy is implemented for negatively correlated indicators. The adjustment range is controlled, and user feedback of discomfort triggers a rollback operation; the adjusted nutrient range is used as a new constraint condition, the personalized nutritional demand prediction model is updated, and the nutritional constraints are regenerated; for recipes that meet the adjusted nutritional range, a reward value is given in the Q-Learning algorithm according to the correlation strength; the adjustment effect is verified to obtain the optimal adjusted recipe.

[0091] In a specific embodiment, extracting features includes:

[0092] Health characteristics: BMI = 24.5 (bucketed into the "normal-overweight" range, coded 1), body fat percentage 22% (bucketed into the "medium" range, coded 2), blood pressure 120 / 80 mmHg (normal, coded 0), average sleep time in the past week 6.5 hours (bucketed into "6-7 hours", coded 1).

[0093] Dietary characteristics: Dietary pattern: "high protein and low carbohydrate" (unique hot encoding: [0,1,0], there are 3 types of patterns in total), taste characteristics: saltiness 4, sweetness 2 (normalized to [0,1] is 0.8 and 0.4).

[0094] Situational characteristics: meal time: dinner (coded 2, 0 = breakfast, 1 = lunch, 2 = dinner), scene: home (coded 0), season: summer (coded 1, 0 = spring / autumn, 1 = summer, 2 = winter).

[0095] The final discrete feature vector is: [1(BMI), 2(body fat percentage), 0(blood pressure), 1(sleep), 0, 1, 0(dietary pattern), 0.8, 0.4(taste), 2, 0, 1(scenario)].

[0096] The current candidate set consists of 3 dishes:

[0097] Action A1: Pan-fried chicken breast salad (30g protein, 10g fat, 20g carbohydrates);

[0098] Action A2: Tomato and Beef Soup (25g protein, 8g fat, 30g carbohydrates);

[0099] Action A3: Stir-Fried Seasonal Vegetables with Brown Rice (15g protein, 5g fat, 40g carbohydrates).

[0100] If the customer rates A1 4 stars → Standardized reward +0.8; rates A2 3 stars → +0.3; rates A3 2 stars → -0.2. If the weight drops by 0.2kg the next morning → +0.5, and the nutrient compliance rate increases by 10% → +0.3. Total reward = Standardized reward + Delayed reward (e.g., if choosing A1 results in weight loss + increased compliance rate, total reward = 0.8 + 0.5 + 0.3 = 1.6).

[0101] Not adopting the recommended recipe → -0.5; Triggering the risk of sodium exceeding the standard → -1.0.

[0102] Initial Q-table, with an 80% probability of choosing the action with the maximum Q value and a 20% probability of random selection. Update the Q value. When Q(A1) = Q(A2) = 0.85, select A1 according to the historical score standard deviation (standard deviation of A1 is 0.1 < standard deviation of A2 is 0.3).

[0103] Independent variable X: Daily protein intake, Dependent variable Y: Change in body fat percentage (measured daily, cumulative decrease of 1.2% in 14 days, average 0.0857% / day, standard deviation 0.03%). Calculate the correlation coefficient ρ, the formula is as follows:

[0104]

[0105] Conduct a significance test: Degrees of freedom df = 12, α = 0.05, critical value 2.179 → t = 3.72 > 2.179, reject H0, confirm a strong positive correlation (ρ = 0.75, P < 0.01).

[0106] Level: Significantly correlated (|ρ| = 0.75 ≥ 0.5), Direction: Positive correlation, as protein intake increases, body fat percentage decreases.

[0107] Conduct a positive correlation upward adjustment: The lower limit of protein recommendation increases from 55g → 55×(1 + 15%) = 63.25g, and the upper limit increases from 75g → 75×(1 + 10%) = 82.5g. Control the adjustment range: The single adjustment range is 15% (not exceeding the 20% threshold). On the 3rd day after adjustment, the customer feedbacks that the satiety is too strong → Roll back 50% of the adjustment range, with the lower limit 60g and the upper limit 80g.

[0108] The protein range input to the Transformer model was changed to 60-80g, and recommendations were regenerated. Recipes with ≥60g of protein received a bonus of +0.2 (e.g., recipe A1 (30g protein / serving) requires a different protein source, with no bonus; newly added recipe A4 (40g protein / serving) receives a bonus of +0.2). Seven days after the adjustment, protein compliance increased from 70% to 85%, body fat percentage decreased another 0.8%, and the ρ value stabilized at 0.72.

[0109] In a second aspect, the present application provides a customer health data intelligent management system, comprising:

[0110] Health data management and nutrient calculation module: includes: basic data collection unit, health data collection unit and nutrient calculation unit; the basic data collection unit collects basic data through the registration form, the health data collection unit sets the health data management cycle, uses health monitoring equipment to monitor customers and obtain health data; the nutrient calculation unit calculates nutrients based on the customer's daily menu input;

[0111] Input vector generation module: This module includes a dietary pattern recognition unit, a taste feature extraction unit, and an input vector integration unit. The dietary pattern recognition unit establishes a daily nutrient matrix based on nutrients and identifies dietary patterns using a K-means clustering algorithm. The taste feature extraction unit defines a taste space and extracts taste features based on the seasoning usage of daily recipes. The input vector integration unit integrates basic data, health data, and dietary patterns to form an input vector.

[0112] The recommended recipe generation module includes a nutritional constraint calculation unit and a recommended recipe generation unit. The nutritional constraint calculation unit uses the Transformer architecture to establish a personalized nutritional requirement prediction model, divides the data set for training, and predicts the daily nutrient intake range based on the input vector as a nutritional constraint. The recommended recipe generation unit retrieves candidate recipes based on the nutritional constraints, calculates the cosine similarity between the taste characteristics of the candidate recipes and the customer's taste characteristics, performs taste matching, and obtains recommended recipes.

[0113] Optimal recipe generation module: includes: recommended recipe optimization unit, optimal recipe generation unit, intervention effect quantification unit and optimal recipe adjustment unit; among them, the recommended recipe optimization unit optimizes the recommended recipes based on the recipe selection records of similar customers; the optimal recipe generation unit adopts the Q-Learning algorithm to dynamically adjust the recommendation strategy according to the customer's rating of the recommended recipe to obtain the optimal recipe; when the customer consumes the optimal recipe, the intervention effect quantification unit uses the Pearson correlation coefficient to calculate the changes in health data before and after the dietary intervention and quantify the intervention effect; the optimal recipe adjustment unit adjusts the nutrients required by the customer based on the intervention effect and adjusts the optimal recipe accordingly.

[0114] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A method for intelligent management of customer health data, characterized in that: The following steps are involved: Basic data is collected through registration forms, health data management cycles are set, and health monitoring equipment is used to monitor customers and obtain health data; customers enter their daily menus and nutrients are calculated; A daily nutrient matrix was established based on nutrients, and dietary patterns were identified using the K-means clustering algorithm; Define the taste space and extract taste features based on the seasoning usage of daily recipes; integrate basic data, health data, and dietary patterns to form an input vector; A personalized nutritional needs prediction model was established using the Transformer architecture. The dataset was divided for training, and the daily nutrient intake range was predicted based on the input vector as a nutritional constraint. Retrieve candidate recipes based on nutritional constraints, calculate the cosine similarity between the taste characteristics of the candidate recipes and the taste characteristics of the customer, perform taste matching, and obtain recommended recipes; Based on the recipe selection records of similar customers, recommended recipes are optimized; using the Q-Learning algorithm, the recommendation strategy is dynamically adjusted according to the customer's rating of the recommended recipe to obtain the optimal recipe; when the customer consumes the optimal recipe, the Pearson correlation coefficient is used to calculate the changes in health data before and after the dietary intervention to quantify the intervention effect; based on the intervention effect, the nutrients required by the customer are adjusted and the optimal recipe is adjusted accordingly.

2. A method for intelligent management of customer health data according to claim 1, characterized in that: The basic data is collected through the registration form, the health data management cycle is set, and the health monitoring equipment is used to monitor the customer to obtain health data, including: Wearable device data is collected in seconds and synchronized to the cloud every minute; for home device data, users are advised to measure at fixed times and automatically identify valid measurement periods; for physical examination reports and laboratory test data, the data validity period is automatically marked; health data is timestamp aligned and data normalized. For wearable device data, by identifying the valid measurement period, the data within the period is assigned a corresponding timestamp; for physical examination reports and laboratory test data, the data validity period is marked and the timestamp is determined according to the report generation time; health data is arranged in chronological order through a unified time base to achieve timestamp alignment; Z-score normalization is used to standardize health data; the wearable device data includes heart rate, blood oxygen saturation, blood pressure, body temperature, calorie consumption and sleep data.

3. A method for intelligent management of customer health data according to claim 1, characterized in that: The customer enters a daily menu and calculates nutrients, including: Customers input their daily menus in text format, and NLP technology is used to automatically identify ingredient names and weight descriptions, matching them with standard ingredient units in the database. For photos of recipes taken, the YOLO algorithm is used to identify the main ingredients, and OCR technology is used to parse the recipe information on tableware or packaging. The app supports importing historical recipe data from the diet record app, and achieves structured data migration through the API interface to obtain recipe information. Nutrients are calculated for each ingredient, complex dishes are broken down into main ingredients, side dishes, and seasonings, the nutritional contribution of each part is calculated separately, and the nutrient absorption rate is adjusted according to the cooking method.

4. A method for intelligent management of customer health data according to claim 1, characterized in that: The method of establishing a daily nutrient matrix based on nutrients and identifying dietary patterns using a K-means clustering algorithm includes: The elbow rule is used to calculate the moment of inertia, and the K value at the inflection point of the curve is selected; K non-repeated single-day data points are randomly selected as the initial centroids, and the K-means algorithm is used to calculate the Euclidean distance between each single-day data point and the K centroids, and assign them to the closest cluster; the mean vector of all data points in each cluster is calculated as the new centroid; the mean nutrient value of each cluster is calculated, compared with the recommended value in the dietary guidelines, and significant difference indicators are extracted; the frequency of occurrence of each cluster label for each user is counted, and dietary patterns are identified for verification and optimization.

5. A method for intelligent management of customer health data according to claim 1, characterized in that: The flavor space is defined, and flavor characteristics are extracted based on the seasoning usage of daily recipes, including: We defined basic taste dimensions, including saltiness, sweetness, sourness, spiciness, umami, and bitterness. We used a Likert scale to quantify each dimension, representing each recipe as a multidimensional taste vector. We constructed a 3D taste space, where each recipe corresponds to a point in the space. Different dietary patterns form clustered distributions in the taste space. A database covering all condiments is established, and the taste contribution weights of the condiments are normalized to ensure that the sum of the weights of each dimension is 1. The usage frequency data is updated regularly, and the weights are dynamically adjusted according to newly entered recipes. For each recipe, the taste characteristics are calculated based on the condiments used and their amounts. A taste perception threshold is set, and the dimension value of the taste characteristics below the threshold is set to 0. For compound condiments, the recipe decomposition method is used to calculate the taste contribution of each ingredient separately, and then the total taste characteristics are obtained by superposition.

6. A method for intelligent management of customer health data according to claim 1, characterized in that: The Transformer architecture is used to establish a personalized nutritional needs prediction model, which divides the data set for training and predicts the daily nutrient intake range based on the input vector as a nutritional constraint, including: In the input and embedding layers, for input vectors, numerical features are directly input, categorical features are converted to continuous vectors using a trainable embedding matrix, all features are merged and positional encoding is added to form the final input vector. In the encoder layer, the number of heads and the dimensions of each head are set, the query, key, and value matrices are calculated, and the attention score is calculated. For the two fully connected layers, the activation function is GELU. Residual connections are added after each sublayer, and layer normalization is performed to stabilize training. In the output layer, the time dimension is compressed through global average pooling, and the fully connected layer regresses to predict the nutrient intake range. The dataset was stratified by health goals and BMI levels to ensure consistent distribution across subsets. During training, Huber Loss was used to balance the influence of outliers, the AdamW optimizer was used, and a cosine annealing strategy was used for the learning rate. Dropout was added to the encoder layer to prevent overfitting, the weight matrix was initialized using Xavier, and layer normalization was added before the activation function. Model training and validation were performed to predict the daily nutrient intake range as a nutritional constraint.

7. A method for intelligent management of customer health data according to claim 1, characterized in that: The Q-Learning algorithm is used to dynamically adjust the recommendation strategy based on the customer's rating of the recommended recipe to obtain the optimal recipe, including: Based on health data, we extract BMI, body fat percentage, blood pressure, and average sleep duration in the past week as health features, dietary patterns and taste characteristics as dietary features, and the time, scene, and season of the customer's meal as contextual features. We then perform bucketing on continuous features and one-hot encoding on categorical features to form discrete feature vectors. Each candidate recipe is mapped to a unique action ID, and the customer's rating of the recommended recipe is mapped to a standardized reward ranging from -1 to 1. Delayed rewards are set based on the next morning's weight change, blood pressure fluctuations, and improvement in nutrient compliance. A total reward is set based on the standardized rewards and the delayed rewards. Failure to adopt a recommended recipe or triggering a dietary risk rule is included in the penalty mechanism. Initialize the Q-table and use the ε-greedy algorithm to randomly select candidate recipes and choose the action with the largest Q-value. When the Q-values ​​are the same, select from high to low according to the following priorities: small historical score standard deviation, short cooking time, and low ingredient repetition rate. Update the Q-value. Record the status, action, user score, and next-day health data of each recommendation, filter out invalid records that have not been rated by the user and have no health changes, perform real-time updates and offline batch training. Select the action A* with the highest Q-value as the optimal recipe.

8. A method for intelligent management of customer health data according to claim 1, characterized in that: When the customer consumes the optimal recipe, the Pearson correlation coefficient is used to calculate the changes in health data before and after the dietary intervention to quantify the intervention effect, including: In the Pearson correlation coefficient, the independent variable X is set as the actual daily nutrient intake during the intervention period, the dependent variable Y is set as the change in health data before and after the intervention, the null hypothesis H0 is set as no significant linear correlation between X and Y, and the alternative hypothesis H1 is set as a significant linear correlation between X and Y. The correlation coefficient ρ is calculated using the following formula: Where n is the number of data points in the dataset, X i is the nutrient intake on day i, represents the mean of X, Y i is the health data on day i, is the mean of Y; Perform a significance test and calculate the statistic t. The formula is: Look up the table to determine whether t exceeds the critical value and test significance; According to the absolute value of the correlation coefficient ρ, the intervention effect is divided into different levels to clarify the degree of correlation between dietary intervention and health data.

9. A method for intelligent management of customer health data according to claim 1, characterized in that: The method of adjusting the nutrients required by the client based on the intervention effect and adjusting the optimal recipe accordingly includes: Based on the intervention effect, an upward adjustment strategy is implemented for positively correlated indicators, and a downward adjustment strategy is implemented for negatively correlated indicators. The adjustment range is controlled, and user feedback of discomfort triggers a rollback operation; the adjusted nutrient range is used as a new constraint condition, the personalized nutritional demand prediction model is updated, and the nutritional constraints are regenerated; for recipes that meet the adjusted nutritional range, a reward value is given in the Q-Learning algorithm according to the correlation strength; the adjustment effect is verified to obtain the optimal adjusted recipe.

10. A customer health data intelligent management system, using a customer health data intelligent management method according to any one of claims 1 to 9, characterized in that: include: Health data management and nutrient calculation module: includes: basic data collection unit, health data collection unit and nutrient calculation unit; the basic data collection unit collects basic data through the registration form, the health data collection unit sets the health data management cycle, uses health monitoring equipment to monitor customers and obtain health data; the nutrient calculation unit calculates nutrients based on the customer's daily menu input; Input vector generation module: This module includes a dietary pattern recognition unit, a taste feature extraction unit, and an input vector integration unit. The dietary pattern recognition unit establishes a daily nutrient matrix based on nutrients and identifies dietary patterns using a K-means clustering algorithm. The taste feature extraction unit defines a taste space and extracts taste features based on the seasoning usage of daily recipes. The input vector integration unit integrates basic data, health data, and dietary patterns to form an input vector. The recommended recipe generation module includes a nutritional constraint calculation unit and a recommended recipe generation unit. The nutritional constraint calculation unit uses the Transformer architecture to establish a personalized nutritional requirement prediction model, divides the data set for training, and predicts the daily nutrient intake range based on the input vector as a nutritional constraint. The recommended recipe generation unit retrieves candidate recipes based on the nutritional constraints, calculates the cosine similarity between the taste characteristics of the candidate recipes and the customer's taste characteristics, performs taste matching, and obtains recommended recipes. Optimal recipe generation module: includes: recommended recipe optimization unit, optimal recipe generation unit, intervention effect quantification unit and optimal recipe adjustment unit; among them, the recommended recipe optimization unit optimizes the recommended recipes based on the recipe selection records of similar customers; the optimal recipe generation unit adopts the Q-Learning algorithm to dynamically adjust the recommendation strategy according to the customer's rating of the recommended recipe to obtain the optimal recipe; when the customer consumes the optimal recipe, the intervention effect quantification unit uses the Pearson correlation coefficient to calculate the changes in health data before and after the dietary intervention and quantify the intervention effect; the optimal recipe adjustment unit adjusts the nutrients required by the customer based on the intervention effect and adjusts the optimal recipe accordingly.

Images