Diet management system for diabetics

Abstract

The invention discloses a diet management system for diabetic patients, which relates to the technical field of intelligent diet management and is technically characterized by comprising a data acquisition module, a nutrition analysis and suggestion generation module, a blood glucose prediction module, a feedback learning module, a personalized strategy recommendation module, an excitation and verification module and a communication module, the system automatically collects health and diet data of a user through a camera, a body fat scale, a glucometer and wearable equipment, identifies food types and estimates components by using a convolutional neural network, and realizes automatic calculation of energy and nutritional ingredients. The food-blood glucose response model constructed on the basis of individual historical data can predict postprandial blood glucose changes, sensitivity coefficients are dynamically updated through a regression correction mechanism, individualized blood glucose prediction is achieved, the system further generates intake suggestions through a component back-calculation method, and the intake suggestions and the intake suggestions are given to a user. And the result is visually displayed on the terminal in the form of a blood glucose curve and an energy distribution diagram, and real-time diet guidance is provided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent diet management, in particular to a diabetes patient diet management system. BACKGROUND

[0002] Diabetes is a common chronic metabolic disease, and long-term high blood sugar can cause various complications and continuously affect the quality of life of patients. Clinical practice generally believes that diet control is one of the most critical links in diabetes management. A reasonable dietary structure can help patients stabilize blood sugar levels, reduce the burden on the islets, and delay disease progression. However, in reality, many patients have difficulty in scientific diet management. On the one hand, ordinary patients often have difficulty in accurately assessing daily energy and nutrient intake, and lack intuitive understanding of the glycemic index of food. On the other hand, there are large metabolic differences between individuals, and the same food has a significant difference in the influence on blood sugar of different patients, and traditional unified dietary recommendations cannot accurately adapt to each individual.

[0003] There are already some diet record applications or diabetes auxiliary management software on the market, but most of them still rely on manual input of food names and quantities by users, and the data lacks objectivity and real-time. Some systems try to identify food categories in the dish through image recognition technology, but the recognition accuracy is limited, and the food volume or mass cannot be reliably estimated. Existing solutions usually only stay at the level of nutrition component statistics or calorie calculation, and have not yet been able to establish individualized prediction models combined with user's historical blood sugar changes, exercise amount, sleep status, etc. In addition, current dietary recommendations are mostly in static form, cannot be dynamically adjusted with blood sugar status and physiological indicators, and also lack feedback and behavior incentive mechanisms for user execution. After long-term use, users often gradually give up recording because of the lack of interaction and sense of achievement, and the system cannot therefore be continuously optimized.

[0004] With the development of wearable devices, mobile health terminals, and artificial intelligence technology, the conditions for obtaining and analyzing personal health data have been greatly improved. Blood glucose meters, body fat scales, smart bracelets, and other devices can provide high-frequency physiological data continuously, and deep learning and time series analysis models also make it more possible to predict blood sugar fluctuations.

[0005] Therefore, there is an urgent need for a diabetes patient diet management system to solve the above problems. SUMMARY

[0006] The present application aims to solve the technical problems raised in the background art, and provides a diabetes patient diet management system, which comprises a data acquisition module, a nutrition analysis and suggestion generation module, a blood glucose prediction module, a feedback learning module, a personalized strategy recommendation module, an incentive and closed-loop verification module, and a model self-learning module; wherein the data acquisition module is used to acquire health data and dietary intake data of a user, and the data acquisition module comprises:

[0007] The user index information acquisition unit is used to acquire one or more of the user's blood glucose level, blood lipid level, liver and kidney function index, body weight, body fat rate, target body weight, and blood glucose control target;

[0008] The user nutrition intake information acquisition unit is used to acquire dietary information through the image shooting function or manual input mode of the user terminal;

[0009] The nutrition analysis and suggestion generation module is used to comprehensively analyze the output of the data acquisition module and generate a nutrition plan containing the proportion of three major nutrients, recommended combination, glycemic index, and personalized intake suggestion;

[0010] The blood glucose prediction module is used to predict the change of blood glucose after the intake of a specific food based on the "food-blood glucose" response prediction model established for the user, and generate a personalized medication or dietary intervention reminder based on the prediction result, wherein the blood glucose prediction module outputs a quantitative personal blood glucose response index;

[0011] The feedback learning module is used to acquire and analyze multi-dimensional feedback data of the user to continuously optimize the prediction model and personalized suggestions, and the feedback learning module comprises:

[0012] The dietary feedback acquisition unit is used to acquire the pre-meal image, food category, portion information, and exercise and energy consumption data of the user;

[0013] The physiological feedback acquisition unit is used to automatically acquire blood glucose meter measurement data and construct a "time-blood glucose" curve, or record the user's post-meal subjective feeling;

[0014] The behavior feedback acquisition unit is used to record the user's adherence to the system suggestions and post-meal behavior;

[0015] The hidden relationship prediction unit is used to identify the potential correlation between eating time, exercise, sleep, and emotional factors and blood glucose fluctuation;

[0016] The personalized strategy recommendation module is used to dynamically generate a personalized diet strategy based on the comprehensive output of the data acquisition module, the nutrition analysis and suggestion generation module, the blood glucose prediction module, and the feedback learning module;

[0017] The personalized strategy recommendation module comprises:

[0018] a dietary recommendation dynamic adjustment unit configured to provide food replacement and intake component optimization recommendations according to real-time blood glucose and food glycemic index;

[0019] a personalized rule generation unit configured to generate food consumption rules and personalized menus identified by colors;

[0020] an incentive and closed-loop verification module configured to generate a visual report based on user dietary execution, blood glucose control indicators, and physical indicator data, and to improve user compliance through an incentive badge mechanism;

[0021] a model self-learning module configured to perform incremental learning or retraining on prediction models in the system based on data output by the feedback learning module, to achieve individualized continuous optimization.

[0022] Further, the user indicator information acquisition unit is further configured to acquire one or more of the user's age, gender, medical history, and medication information.

[0023] Further, the user nutrition intake information acquisition unit adopts a food recognition and component estimation model based on deep convolution feature extraction;

[0024] The model includes:

[0025] a feature extraction network configured to extract multi-level visual features from food images;

[0026] a food recognition and segmentation sub-module configured to output food category recognition results and the contour areas of each food in the image;

[0027] a component estimation sub-module configured to perform spatial scale conversion in combination with a plate calibration object or known size to estimate the volume or mass of each food;

[0028] a model training unit configured to perform supervised training using a labeled food image dataset, with a loss function that combines classification loss, segmentation loss, and regression loss;

[0029] The segmentation model adopts an "encoder-decoder" structure, and the component estimation sub-module can implement three-dimensional reconstruction in combination with depth information or dual-view images.

[0030] Further, the blood glucose prediction module adopts a machine learning or time series prediction model trained based on user historical data;

[0031] The data preprocessing unit standardizes and extracts features from input historical data, and the input features include food categories, intake amounts, eating time stamps, pre-meal blood glucose values, exercise data, and medication records.

[0032] The model architecture selection includes but is not limited to gradient boosting decision tree, long short-term memory network or gated recurrent unit network;

[0033] The personalized training unit independently creates a model instance for each user and fine-tunes the parameters;

[0034] The prediction output unit outputs the future blood glucose change curve and triggers an intervention reminder when the threshold is exceeded.

[0035] Further, the hidden relationship prediction unit uses association rule mining or neural network model to identify potential nonlinear factors affecting blood glucose fluctuations.

[0036] Further, the dietary suggestion dynamic adjustment unit outputs the component optimization suggestion based on the real-time blood glucose level and the food glycemic index to calculate the user's edible gram range.

[0037] Further, the personalized rule generation unit generates food consumption suggestion rules, in which the green light symbol represents safe food, the yellow light symbol represents food that needs to be controlled in quantity, and the red light symbol represents food that should be avoided.

[0038] Further, the incentive and closed-loop verification module generates a visual feedback report including a weekly diet execution report, a blood glucose control trend report, and a body weight change report.

[0039] Further, the incentive badge mechanism awards virtual or physical rewards when the user continuously achieves the preset blood glucose control goal.

[0040] Further, the system further comprises a communication module for realizing data synchronization and model updating between the user terminal and the cloud server, and a data security and privacy protection module for encryption and permission control during data transmission and cloud storage.

[0041] Compared with the prior art, the present application has the following beneficial effects:

[0042] The system of the present application is a diabetes diet management system that can automatically identify food and perform quantitative nutritional analysis. Unlike traditional methods that rely on manual input or pre-set recipes, the present system uses a convolutional neural network combined with a spatial calibration algorithm to automatically identify food types and quantities from a dish image. Through volume conversion and density correction methods, the system can directly output food quality data and automatically match with database nutrition parameters to achieve high-precision energy and nutrient content calculation. This combination of visual recognition and physical estimation standardizes, quantifies and verifies the diet data collection process, significantly improving the objectivity and repeatability of diet monitoring.

[0043] In the scheme of the present application, a "food-blood glucose response" model based on individualized data is established in blood glucose prediction. The system uses a time series learning algorithm (such as long short-term memory network LSTM or gradient boosting tree GBDT) to model the fusion of user historical diet, blood glucose, exercise and medication information, and introduces a dynamic regression correction mechanism of blood glucose sensitivity coefficient a, so that the model can adjust itself according to the individual's metabolic characteristics. Compared with the existing model based on population average data, the model can continuously self-learn, and has significant improvement in prediction accuracy and stability. The system can not only predict the blood glucose rise of specific food, but also give quantifiable intake adjustment suggestions when abnormal trends are detected, realizing individualized dynamic blood glucose management.

[0044] In the scheme of the present application, the blood glucose prediction result is further combined with the diet strategy recommendation module to form an intelligent control method that can be intervened in real time. The system automatically obtains the edible safe component through the reverse calculation formula: M allow =(BG target -BG now ) / [α×(GI / 100)], and feeds back the result to the terminal interface; the user can obtain the quantifiable guidance before meals, instead of post-treatment prompt, and the system simultaneously adjusts the parameters according to the historical prediction deviation to realize continuous self-optimization. This quantitative control method changes the diet guidance from static suggestion to dynamic decision, solves the problem of "data dispersion and feedback lag" in traditional diabetes management, and has substantial progress in blood glucose prediction accuracy and clinical guidance significance. BRIEF DESCRIPTION OF DRAWINGS

[0045] Fig. 1 is the system block diagram of the diabetes patient diet management system in the embodiment of the present application;

[0046] Fig. 2 is the flow chart of the diabetes patient diet management system in the embodiment of the present application;

[0047] Fig. 3 is the time-blood glucose change curve diagram in the embodiment 2 of the present application. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical scheme and advantages of the present application more clear, the following combines embodiments and drawings to further explain the present application. Figs. 1-3 It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0049] Embodiment 1: as Figs. 1-2As shown, the diabetes patient diet management system provided by the embodiment of the present application, around the actual problems of the diabetes patient daily diet difficult to quantify, blood glucose fluctuation affected by many factors, large individual differences, insufficient compliance, etc., builds a continuous workflow with data collection, nutrition analysis, blood glucose prediction, feedback learning, personalized strategy recommendation, incentive and closed-loop verification, model self-learning and communication synchronization as the core link, makes the data flow and logical feedback between each module consistent, so as to realize the dynamic optimization and continuous individualized regulation of the system, and the data interface is used to connect between each functional module, so that the system can automatically update the nutrition suggestion and blood glucose prediction result after obtaining new user data, and return the updated result to the user terminal for display and incentive.

[0050] The following will be described in conjunction with a typical use scene of a type 2 diabetes patient in the present embodiment, but those skilled in the art can understand that the system is also applicable to other people who need to control postprandial blood glucose.

[0051] In the present embodiment, first, the data acquisition module is used to uniformly acquire the basic health data and diet intake data of the user. When the user uses it for the first time, the personal basic information is entered in the user terminal (mobile APP, small program or wearable device supporting application), including age, gender, height, weight, body fat rate, past medical history, current medication information and target fasting blood glucose value and 2-hour postprandial blood glucose target value given by the doctor. In order to improve the degree of automation, the user index information acquisition unit in the system can be connected with external devices such as home blood glucose meter, smart body fat scale, wristwatch type heart rate meter, sports bracelet, etc., and the data of blood glucose level, weight, heart rate, exercise steps, energy consumption, etc. are automatically uploaded to the cloud server of the system according to the time stamp. For users who need to monitor liver and kidney function and blood lipid level for a long time, after the user completes the hospital physical examination or outpatient examination, the test results can be entered into the system by manual input or photographing and identifying the test sheet. The system will also write this part of data into the personal health record of the user, and use it as the basis for determining whether to limit fat and whether to limit protein in the subsequent diet suggestion.

[0052] On the acquisition of dietary data, the user nutrition intake information collection unit supports two ways: one is manual input mode, that is, the user selects the breakfast, lunch or dinner time period, checks the food in the food library, and inputs the approximate amount; the other is image shooting mode, that is, the user takes a photo of the plate before eating and uploads it to the mobile phone, and the system automatically identifies the food type and amount. In order to improve the recognition accuracy, the embodiment adopts a food recognition and quantity estimation model based on deep convolution feature extraction. The model uses a pre-trained convolutional neural network (such as ResNet, MobileNet, etc.) as a feature extraction network to perform multi-level feature coding on the input food image and obtain a high-dimensional visual feature vector; then the feature is sent to a food recognition and segmentation submodule, which uses a semantic segmentation network with an encoder-decoder structure to segment the pixel regions belonging to different foods in the image and outputs the food category label corresponding to each region through a classification head. In order to realize quantity estimation, the user can place a calibration object (such as a standard card or a small bowl with a fixed specification) of known size next to the plate when taking a photo, and the system calculates the ratio of the food pixel area to the calibration object pixel area in the quantity estimation submodule, thereby obtaining the actual area occupied by the food, and then combines the known depth of the plate or the system's default average thickness model to calculate the volume, and the volume estimation formula in the quantity estimation submodule can be as follows:

[0053] V=A x h;

[0054] Where V is the estimated food volume, unit mL; A is the horizontal projection area of the food region in the image converted to the actual scale, unit cm 2 ; h is the average thickness or height of the reference plate or this type of food preset in the system, unit cm. Since the user may use different units of area and height in actual use at home, the system will unify the dimensions and convert them to mL in the background to ensure the consistency of subsequent nutritional analysis and calculation. If the user has taken a dual-view image or the terminal has a depth camera, the depth image and the segmentation result can be used for three-dimensional reconstruction, thereby improving the estimation accuracy of h.

[0055] After obtaining the volume V of the food in the quantity estimation submodule, the system converts the volume to mass according to the preset density parameter p i of this type of food in the food composition database, and the conversion formula is:

[0056] m i =V x p i ;

[0057] Where m i is the actual intake mass of the i-th food, unit g; V is the volume of the food obtained by image segmentation and calibration, unit mL or cm 3 ; pi is the density or volume mass coefficient of the food, in g / mL. The system will unify the unit to g in actual calculation, and then multiply the m i into the incoming nutrition analysis and suggestion generation module for subsequent calculation of energy, macronutrients, and glycemic load.

[0058] When the volume V or mass m i of the food is obtained, the system matches the food with the built-in food composition database, which establishes fields such as energy, protein, fat, carbohydrate, dietary fiber, sodium, and typical glycemic index GI per 100g for each food.

[0059] The nutrition analysis and suggestion generation module sums up the nutritional components of all the foods consumed by the user this time to obtain the proportion of macronutrients in this meal. Assuming there are n foods in this meal, the system obtains the actual intake mass m i of each food i (i = 1, 2,..., n) (unit g), and looks up the nutritional components per 100g of the food, with the carbohydrate content denoted as C i (g / 100g), the protein content as P i (g / 100g), and the fat content as F i (g / 100g). Then the total intake of macronutrients in this meal can be represented as:

[0060]

[0061] where CHO meal is the total carbohydrate intake in this meal, in g; PRO meal is the total protein intake in this meal, in g; FAT meal is the total fat intake in this meal, in g; m i is the actual intake mass of the i-th food, in g; C i , P i , and F i are the standard contents of carbohydrates, proteins, and fats, respectively. After obtaining the absolute amounts of macronutrients, the system converts them into energy proportions. Assuming that carbohydrates and proteins provide 4kcal per gram and fats provide 9kcal per gram, the total energy is:

[0062] E meal = 4 × CHO meal + 4 × PRO meal + 9 × FAT meal ;

[0063] where E mealFor the total energy of the meal, unit kcal; then the system calculates the proportion of energy of the three major nutrients, which is used to determine whether it meets the recommended proportion range of diabetic diet.

[0064] In order to better evaluate the postprandial glucose load, the nutrition analysis and suggestion generation module will also calculate the glucose load GL of the meal, the formula is written as:

[0065]

[0066] Among them, GL meal is the glucose load of the meal; GI i is the glycemic index of the i-th food; is used to convert the GI in percentage to a coefficient of 0-1, the system will calculate GL meal and compare it with the "single meal glucose load upper limit" set by the user when setting up the file, when GL meal exceeds the upper limit, the system will generate specific recommendations such as "replace high GI staple food", "reduce the amount of food" or "add coarse fiber food in the same meal" in the personalized strategy recommendation module.

[0067] After completing the nutrition analysis, the blood glucose prediction module begins to predict the change of blood glucose after the user ingests the meal. Since the islet function, drug type, exercise amount, and sleep condition of different users are different, the system establishes an independent "food-blood glucose response" prediction model for each user. In order to make it clear, this embodiment uses a time series model for description, the input vector of the prediction model can be constructed as:

[0068] x t =[BG t-1 ,GL meal ,E meal ,ACT t ,MED t ,TIME t ];

[0069] Among them, x t is the model input vector at time point t; BG t-1 is the blood glucose value measured by the user before the meal, unit mmol / L; GL meal is the glucose load of the meal; E meal is the total energy of the meal; ACT t is the exercise amount or activity amount index of the user in the time period (such as the number of steps or calories consumed); MED t is the drug feature quantity of the user in the period (which can be whether to inject insulin, drug dose or oral hypoglycemic drug type code); TIME tmeal time features (e.g. embedding representation of early / mid / late or specific time stamp). When recurrent neural network (e.g. LSTM) is adopted, the model takes input sequence of multiple consecutive time steps as sample, memorizes past blood glucose change pattern through hidden layer, and finally outputs predicted blood glucose value at future time points:

[0070]

[0071] wherein, is the predicted blood glucose value at the kth time interval after the current time point, unit mmol / L; f(·) is the trained time series prediction model; L is the length of historical window for prediction, representing the length of data that the model backtracks at one time; k is the prediction step, used to obtain blood glucose level at 30 minutes, 60 minutes or 120 minutes after meal. In this embodiment, the system can output at least the predicted blood glucose value at 2 hours after meal and compare it with the postprandial blood glucose control target BG target set by the user, when the system will generate medication reminder or additional exercise suggestion. The meanings of each letter in the above formula have been given in the foregoing. When gradient boosting tree (GBDT) is adopted by the model, the input features remain unchanged, and only the training method is different.

[0072] In order to enable the model to constantly adapt to the real blood glucose performance of the user, the system is provided with a feedback learning module. After each meal, if the blood glucose meter is worn or connected, the system can automatically pull the actual blood glucose values of the user at 30 minutes, 60 minutes and 120 minutes after meal, and construct a "time-blood glucose" measured curve: At the same time, the user can also actively fill in the subjective feeling after this meal on the terminal, such as whether fatigue, palpitation, obvious hunger and other symptoms occur. This part of data will also be recorded in the physiological feedback collection unit.

[0073] The behavior feedback collection unit is responsible for recording the execution of the user on the system suggestion. For example, the system suggests that the rice of this meal should be reduced to 80g, but the user actually eats 120g, then the unit will record a behavior label of "compliance = partial execution". When the system prompts the user to take a brisk walk for 15 minutes at 30 minutes after dinner, and the terminal of the user monitors that the user has indeed completed the exercise, then it will record "compliance = execution". These behavior labels are very important for subsequent association analysis, because the blood glucose changes caused by the same diet plan under the conditions of "complete execution" and "non-execution" are different.

[0074] In order to discover the hidden factors affecting blood glucose fluctuation from complex daily life data, the embodiment sets a hidden relationship prediction unit in the feedback learning module, which can use association rule mining or a simple feedforward neural network to identify patterns such as "dinner too late + insufficient sleep duration + elevated fasting blood glucose the next morning", and in one implementation, the system records meal time as T e , records the amount of exercise on the day as A d , records sleep duration as S d , and records the maximum blood glucose fluctuation amplitude on the day as ΔBG d , a multivariate relationship can be fitted as follows:

[0075] ΔBG d = g(T e , A d , S d , ST d );

[0076] where ΔBG d is the maximum fluctuation value of blood glucose on the day, in mmol / L; g(·) is a nonlinear function learned from historical data; ST a is the emotional or stress score on the day (which can be obtained by user self-evaluation). When the system finds that a certain combination of features often leads to ΔBG d exceeding the threshold, it will automatically add rules such as "move dinner time forward", "increase light exercise at night", or "reduce high-Gl food on high-stress days" to the subsequent personalized strategy recommendations.

[0077] The personalized strategy recommendation module is the part that generates a diet plan that the user can actually see and execute based on all the data and analysis results described above. On the one hand, it calculates the edible grams based on real-time blood glucose levels and the glycemic index of the food to be eaten, and on the other hand, it dynamically adjusts according to the user's long-term goals. For portion optimization, the system can give the following calculation method: assuming the user's current blood glucose is BG now , the system estimates that the user's postprandial blood glucose will rise by ΔBG exp without taking additional interventions, and the user's postprandial blood glucose target is BG target , then the system will limit the intake of the main meal M allow so that:

[0078] BG now + ΔBG food (M allow ) ≤ BG target ;

[0079] where ΔBG food (M allow ) is the blood glucose fluctuation value when the food intake is M allowThe predicted increment of blood glucose caused by the food, which can be derived from the aforementioned "food-blood glucose response" model. If it is approximated as being proportional to the mass of the food, it can be written as:

[0080]

[0081] where a is the blood glucose sensitivity coefficient estimated by the system according to the user's personal pancreatic function, with a unit of (mmol / L) / g; M allow is the mass of the food allowed to be ingested, with a unit of g; GI is the glycemic index of the food. To ensure that the final blood glucose does not exceed the target blood glucose, the following can be solved:

[0082]

[0083] The calculated M allow is returned to the user, which is expressed as "the recommended mass of rice for this meal is not more than 80g" or "the recommended mass of noodles for this meal is half a portion". When the user uploads a food quantity that exceeds this value, the system will give a replacement suggestion, such as using low-GI coarse grains and small miscellaneous grains instead of refined rice and noodles. The parameter a is dynamically calculated by the system based on the user's personal historical diet-blood glucose data through regression analysis, representing the sensitivity coefficient of unit food mass to blood glucose rise, with a value range of 0.005-0.02 (mmol / L) / g. The system will use the population average as the initial value when it is first used, and a will be gradually corrected according to the user's actual blood glucose response in the continuous optimization process of the model self-learning module, so that the prediction result is more in line with individual differences.

[0084] The personalized rule generation unit will identify the long-term stable dietary restrictions and recommendations with colors to generate a "food signal light" exclusive to the user, where green light represents food that can be eaten without special restrictions; yellow light represents food that needs to control the intake amount or control the frequency of eating, and the system will mark the corresponding M allow or "≤2 times per week" together; red light represents food that is recommended to avoid, such as high-sugar drinks and sugary desserts, etc. In this way, the user can quickly understand what they can eat and how much they should eat even without looking at complex nutritional numbers.

[0085] In order to make the whole system more and more accurate, the embodiment also provides a model self-learning module. The module periodically extracts the latest triple samples of "recommended diet-actual implementation-actual blood glucose" from the feedback learning module, adds them to the user's exclusive prediction model as incremental data, and directly trains the new samples for incremental training for the scene using gradient boosting trees. For the scene using LSTM, the data of the recent period of time can be retrained in a sliding window manner. After updating the model parameters, the model self-learning module numbers the new model version and saves it to the cloud, so as to use the model with higher accuracy in the next prediction. In this way, even if the user's living habits change (for example, starting to take evening walks), the system can automatically realize that "the same dinner will not cause such high blood glucose" after a few days, so as to reduce the blood glucose increment estimation for the user.

[0086] In order to ensure the real-time flow of data between the user terminal and the cloud server, the system also provides a communication module. The communication module supports secure transmission protocols such as HTTPS, is responsible for synchronizing the images collected by the user terminal, the manually entered diet data, and the blood glucose data uploaded by the external device to the cloud analysis service in real time, and is also responsible for pushing the nutritional recommendations, blood glucose prediction results, personalized menus, and visual reports calculated by the cloud back to the user terminal. When the user takes pictures or enters data in a network-free environment, the communication module will temporarily cache the data locally, and automatically upload it after the network is restored, ensuring the integrity of the data. In order to protect the safety and privacy of user data, the communication module of the embodiment uses TLS1.3 protocol for end-to-end encryption during data transmission, and uses AES-256 to encrypt and store user health data and image files during cloud storage. At the same time, an access token and permission grading mechanism are set up, and only authorized users and system services can access the corresponding data. All data is desensitized during transmission and storage to prevent user identity information from being reverse analyzed.

[0087] Embodiment 2: In this embodiment 2, a 52-year-old male patient with type 2 diabetes is taken as an example. The patient is 172 cm tall, weighs 76 kg, has a body fat rate of 25.4%, and the fasting blood glucose control target is 7.0 mmol / L. The postprandial 2-hour blood glucose control target is 10.0 mmol / L. The system automatically collects health data through the Bluetooth-connected body fat scale, smart bracelet, and home blood glucose meter when the patient first uses it, forming an initial health profile containing blood glucose level, heart rate, body weight, body fat rate, exercise steps, and sleep duration. The user sets the daily target energy intake to 1900kcal in the system, and the system allocates the proportion of the three major nutrients as follows: carbohydrates 50%, proteins 20%, and fats 30%.

[0088] In a dinner scenario, the patient opens the system mobile application, takes a picture of the dinner plate through the phone camera, and the image contains rice, braised pork, stir-fried green vegetables, and apple slices. The system's convolutional neural network model first performs feature extraction and semantic segmentation, identifies the four food regions, and calculates their respective projection areas A i The system calculates the volume of each food based on the known height of the dinner plate h = 3.5 cm:

[0089] V i = A i × h;

[0090] where V i is the volume of the i-th food, unit mL; A i is the identified actual area, unit cm; h is the average height of the dinner plate. The system obtains the volume of rice V1 = 210 mL, braised pork V2 = 90 mL, green vegetables and apple V4 = 60 mL.

[0091] The component estimation sub-module then converts the volume to mass based on the food density parameters p i in the database:

[0092] m i = V i × p i ;

[0093] where p i is the corresponding food density (g / mL). Thus, the mass of rice m1 = 216 g, braised pork m2 = 85.5 g, green vegetables m3 = 60 g, and apple m4 = 39 g.

[0094] The nutrition analysis and suggestion generation module uses the nutrition content data C i , P i , F i (respectively representing carbohydrate, protein, and fat content) per 100 g of food in the database to calculate the total amount of the three major nutrients in this meal:

[0095]

[0096] Further, the total energy of the meal is calculated:

[0097] E meal = 4 × (CHO meal + PRO meal ) + 9 × FAT meal ;

[0098] where the energy unit is kcal; the calculation result shows that the energy of the meal is about 720 kcal, with carbohydrates accounting for 53%, proteins accounting for 18%, and fats accounting for 29%, which meets the recommended proportion.

[0099] To evaluate the potential glycemic impact of the meal, the system calculates the postprandial glycemic load:

[0100]

[0101] where GI i is the glycemic index of each food. The total glycemic load GL meal of the meal is calculated to be 43, slightly above the system-set upper limit of 40.

[0102] The glycemic prediction module then predicts the postprandial glycemic values based on the user-individualized "food-glycemic response" model, with the model input feature vector:

[0103] x t = [BG t-1 , GL meal , E meal , ACT t , MED t , TIME t ];

[0104] where BG t-1 is the preprandial glycemic value (7.3 mmol / L), ACT t is the exercise indicator (in steps or calories), MED t is the medication code, and TIME t is the meal time feature. The model output is the predicted glycemic curve for the next 2 hours:

[0105]

[0106] where L is the length of the historical data window, and k is the prediction step size. For the current scenario, the system adopts an individual glycemic sensitivity coefficient a = 0.012 (mmol / L) / g, with rice as the main carbohydrate source and a glycemic index GI = 81, resulting in a predicted 2-hour postprandial glycemic value of:

[0107]

[0108] The predicted value is below the control target of 10.0 mmol / L, and the system determines that the meal glycemic risk is acceptable, providing the suggestion "walking for 15 minutes 30 minutes after the meal can further optimize the glycemic curve."

[0109] If the prediction result exceeds the target threshold, the system will call the back-calculation formula to calculate the consumable grams:

[0110]

[0111] where M allow represents the maximum mass (g) of the food allowed to be consumed, and BG targetTarget blood glucose (mmol / L) for user, BG now Current blood glucose, the system will automatically prompt the user "rice recommended intake not more than 80g" or give alternative food program, after meals, the system records the actual blood glucose data by blood glucose meter: 30 minutes 8.9mmol / L, 60 minutes 9.6mmol / L, 120 minutes 9.2mmol / L.

[0112] The feedback learning module will form a time-blood glucose curve {(0,7.3),(30,8.9),(60,9.6),(120,9.2)} with the measured points, and compare it with the predicted curve, as shown in the following figure. Fig. 2

[0113] The system updates the blood glucose sensitivity coefficient to α=0.0118(mmol / L) / g according to the error back correction parameter, and realizes model adaptive optimization.

[0114] In the process of using for many days, the system will further explore the correlation between "eating time-exercise amount-sleep duration-emotional state" and blood glucose fluctuation, and establish an experience model:

[0115] ΔBG d =g(T e ,A d ,S d ,ST d );

[0116] Wherein, ΔBG d is the daily blood glucose fluctuation amplitude; T e is the eating time; A d is the exercise amount; S d is the sleep duration; ST d is the emotional stress score.

[0117] The model is used to generate longer-term lifestyle intervention suggestions, and the incentive and closed-loop verification module generates a "stable sugar control" badge according to the number of consecutive compliance days, and visualizes the blood glucose trend, compliance rate and body weight change in the weekly report. The communication module is responsible for real-time transmission of data, uses TLS1.3 security protocol and AES-256 encryption algorithm to ensure privacy and security, all data are desensitized before uploading, and the results are synchronized back to the user terminal after the cloud completes the calculation, the blood glucose curve, energy distribution and personalized suggestions are displayed.

[0118] The above only describes the preferred embodiments of the present application and does not limit the present application, any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.​

Claims

1. A dietary management system for diabetic patients, characterized in that, The system includes a data acquisition module, a nutrition analysis and suggestion generation module, a blood glucose prediction module, a feedback learning module, a personalized strategy recommendation module, an incentive and closed-loop verification module, and a model self-learning module; wherein, the data acquisition module is used to collect users' health data and dietary intake data, and the data acquisition module includes: The user indicator information collection unit is used to collect one or more of the following: blood glucose level, blood lipid level, liver and kidney function indicators, weight, body fat percentage, target weight, and blood glucose control target. The user nutrition intake information collection unit is used to obtain dietary information through the image capture function of the user terminal or by manual input. The nutrition analysis and recommendation generation module is used to perform comprehensive analysis based on the output of the data acquisition module to generate a nutrition plan that includes the proportion of the three major nutrients, recommended combinations, glycemic index, and personalized intake recommendations. The blood glucose prediction module is used to predict blood glucose changes after the intake of a specific food based on a "food-blood glucose" response prediction model established for the user, and to generate personalized medication or dietary intervention reminders based on the prediction results. The blood glucose prediction model outputs a quantitative personal blood glucose response index. The feedback learning module is used to collect and analyze multi-dimensional user feedback data to continuously optimize the prediction model and personalized suggestions. The feedback learning module includes: Dietary feedback collection unit: used to acquire user pre-meal images, food types, portion sizes, and exercise and energy consumption data; Physiological feedback acquisition unit: used to automatically acquire blood glucose meter measurement data and construct a "time-blood glucose" curve, or record the user's subjective feelings after a meal; Behavioral feedback collection unit: used to record user compliance with system suggestions and post-meal behavior; Hidden Relationship Prediction Unit: Used to identify potential associations between eating time, exercise, sleep, and emotional factors and blood glucose fluctuations; The personalized strategy recommendation module is used to dynamically generate personalized dietary strategies based on the combined output of the data acquisition module, nutrition analysis and suggestion generation module, blood glucose prediction module and feedback learning module. The personalized strategy recommendation module includes: The dynamic adjustment unit for dietary recommendations is used to provide suggestions on food substitutions and portion optimization based on real-time blood glucose and the glycemic index of foods. The personalized rule generation unit is used to generate color-coded food consumption rules and personalized menus; The incentive and closed-loop verification module is used to generate visual reports based on users' dietary adherence, blood glucose control indicators, and physical indicators, and to improve user compliance through an incentive badge mechanism. The model self-learning module is used to incrementally learn or retrain the prediction model in the system based on the data output by the feedback learning module, so as to achieve individualized continuous optimization.

2. The dietary management system for diabetic patients according to claim 1, characterized in that, The user indicator information collection unit is also used to collect one or more of the user's age, gender, medical history, and medication information.

3. The dietary management system for diabetic patients according to claim 1, characterized in that, The user nutrition intake information collection unit adopts a food identification and portion estimation model based on deep convolution feature extraction. The model includes: Feature extraction networks are used to extract multi-level visual features from food images; The food recognition and segmentation submodule is used to output the food type recognition results and the outline regions of each food in the image; The portion estimation submodule is used to perform spatial scale conversion by combining the plate calibration object or known dimensions to estimate the volume or mass of each food item; The model training unit uses a labeled food image dataset for supervised training, and the loss function combines classification loss, segmentation loss, and regression loss. The segmentation model adopts an "encoder-decoder" structure, and the component estimation submodule can combine depth information or dual-view images to achieve 3D reconstruction.

4. The dietary management system for diabetic patients according to claim 1, characterized in that, The blood glucose prediction module uses a machine learning or time-series prediction model trained based on the user's historical data. The data preprocessing unit standardizes and extracts features from the input historical data. Input features include food type, intake, eating timestamp, pre-meal blood glucose level, exercise data, and medication records. The model architecture may include, but is not limited to, gradient boosting decision trees, long short-term memory networks, or gated recurrent unit networks; The personalized training unit creates a model instance and fine-tunes the parameters for each user independently. The predictive output unit outputs a curve of future blood glucose changes and triggers an intervention alert when the threshold is exceeded.

5. A dietary management system for diabetic patients according to claim 1, characterized in that, The hidden relationship prediction unit uses association rule mining or neural network models to identify potential nonlinear factors affecting blood glucose fluctuations.

6. A dietary management system for diabetic patients according to claim 1, characterized in that, The portion optimization suggestions output by the dynamic adjustment unit of the diet recommendations are based on real-time blood glucose levels and the glycemic index of food to calculate the range of grams that the user can eat.

7. A dietary management system for diabetic patients according to claim 1, characterized in that, In the personalized rule generation unit, the food recommendation rules are indicated by a green light for foods that are safe to eat, a yellow light for foods whose intake needs to be controlled, and a red light for foods that should be avoided.

8. A dietary management system for diabetic patients according to claim 1, characterized in that, The visualization feedback reports generated by the incentive and closed-loop verification module include weekly diet execution reports, blood glucose control trend reports, and weight change reports.

9. A dietary management system for diabetic patients according to claim 1, characterized in that, The incentive medal mechanism awards virtual or physical rewards to users who continuously achieve preset blood glucose control goals.

10. A dietary management system for diabetic patients according to any one of claims 1-9, characterized in that, The system also includes a communication module for data synchronization and model updates between the user terminal and the cloud server, and a data security and privacy protection module for encryption and access control during data transmission and cloud storage.

Images