Digitalized nutrition service pushing method and system based on diabetics

Abstract

The invention provides a digital nutrition service pushing method and system based on a diabetic patient, and the method comprises the steps: firstly obtaining the real-time monitoring log data, containing blood glucose level data, insulin injection records and diet intake information, of the diabetic patient, inputting the real-time monitoring log data into a pre-trained nutrition analysis model, and carrying out the multi-dimensional processing; the method comprises the following steps: generating a dynamic nutrition demand characteristic representing a nutrient element demand change trend in a preset time period, matching the dynamic nutrition demand characteristic with service entries in a digital nutrition database, determining a candidate nutrition service set and sorting, and performing spatial correlation analysis on a real-time geographic position of a patient and a position of a service provider to obtain a real-time geographic position of the patient. According to the method and the system, a reachable target nutrition service subset is screened out, finally, the reachable target nutrition service subset is sent to a patient terminal device according to a pushing strategy, service selection feedback is received to trigger a service execution instruction, and personalized and reachable digital nutrition service accurate pushing is realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a method and system for delivering digital nutrition services to diabetic patients. Background Technology

[0002] In the field of health management for patients with diabetes, nutritional support is crucial for controlling the disease and maintaining good physical condition. However, existing nutritional service methods for patients with diabetes have many shortcomings.

[0003] For example, traditional diabetes nutrition services mainly rely on regular face-to-face consultations and guidance from medical staff. This is not only limited by the time and energy of medical staff, making it impossible to track and adjust patients' nutritional status in real time, but also makes it difficult to ensure the consistency and accuracy of nutritional guidance due to differences in the professional level and experience of different medical staff.

[0004] With the development of information technology, some digital nutrition services have emerged. However, most of these services simply provide general nutritional knowledge and standardized dietary advice, failing to fully consider each diabetic patient's unique physical condition, daily dietary intake, and changes in their condition. For example, these solutions do not incorporate key information such as real-time blood glucose levels and insulin injection records to dynamically assess the patient's nutritional needs, resulting in a lack of targeted nutritional services that fail to meet the individualized needs of patients at different times.

[0005] Furthermore, related technologies often focus solely on providing nutritional services, neglecting the patient's geographical location and the service provider's location information. This makes it difficult for patients to conveniently and quickly access many recommended nutritional services, reducing the service's practicality and patient acceptance. Moreover, the inability to optimize subsequent service recommendations based on patients' actual choices and feedback hinders the continuous improvement of nutritional services to better meet patients' needs. Summary of the Invention

[0006] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for pushing digital nutrition services to diabetic patients, the method comprising: Acquire real-time monitoring log data of diabetic patients, including blood glucose level data, insulin injection records, and dietary intake information; The real-time monitoring log data is input into a pre-trained nutrition analysis model for multi-dimensional processing to generate dynamic nutrition demand features. These dynamic nutrition demand features are used to characterize the changing trend of the nutritional element requirements of the diabetic patient within a preset time period. The dynamic nutritional needs characteristics are matched with nutritional service items in a pre-stored digital nutritional database to determine a priority-ranked set of candidate nutritional services. Spatial correlation analysis is performed based on the real-time geographic location information of the diabetic patients and the location information of service providers in the candidate nutrition service set to filter out an accessible subset of target nutrition services. The target nutrition service subset is sent to the terminal device corresponding to the diabetic patient according to a preset push strategy, and the service selection feedback returned by the terminal device is received to trigger the service execution instruction.

[0007] In another aspect, embodiments of the present invention also provide a digital nutrition service delivery system for diabetic patients, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0008] Based on the above, this application embodiment acquires real-time monitoring log data and uses a pre-trained nutritional analysis model for multi-dimensional processing to generate dynamic nutritional demand characteristics that characterize the changing trends of a patient's nutritional element needs within a preset time period. This changes the previous static and singular nutritional assessment model, enabling timely capture of changes in a patient's nutritional needs caused by factors such as blood glucose levels, insulin injections, and dietary intake. Next, by integrating the dynamic nutritional demand characteristics with a digital nutritional database, a priority-ranked set of candidate nutritional services is determined. Then, spatial correlation analysis is performed using the patient's real-time geographic location information and the service provider's location information to filter out an accessible subset of target nutritional services. This not only accurately matches services that meet the patient's nutritional needs from a massive number of nutritional service items but also ensures that these services are geographically accessible to the patient, greatly improving the practicality and accessibility of nutritional services. Finally, after sending the target nutritional service subset to the patient's terminal device according to the preset push strategy, the service selection feedback returned by the terminal device is received to trigger the service execution command. This makes the nutritional service push no longer a one-way information transmission, but can adjust and optimize the service in a timely manner based on the patient's feedback, further improving the patient's acceptance and compliance with nutritional services. This provides diabetic patients with a more scientific, efficient, personalized and continuously optimized digital nutritional service experience, which helps to better manage the patient's nutritional status and assist in the treatment and rehabilitation of diabetes. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the execution flow of the digital nutrition service push method for diabetic patients provided in an embodiment of the present invention.

[0010] Figure 2 This is a schematic diagram of the hardware architecture of a digital nutrition service delivery system for diabetic patients provided in an embodiment of the present invention. Detailed Implementation

[0011] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a method for pushing digital nutrition services to diabetic patients according to an embodiment of the present invention. The following is a detailed description of this method for pushing digital nutrition services to diabetic patients.

[0012] Step S110: Obtain real-time monitoring log data of diabetic patients, including blood glucose level data, insulin injection records, and dietary intake information.

[0013] In this embodiment, we take a diabetic patient named Mr. Li as an example. Mr. Li uses a diabetes management device to monitor his health status every day. The diabetes management device can measure Mr. Li's blood sugar level regularly, such as once every morning after waking up, before and after three meals, and once before going to bed. The blood sugar values ​​obtained from these measurements constitute blood sugar level data.

[0014] In detail, regarding insulin injections, Mr. Li follows his doctor's instructions and needs to inject a certain dose of insulin daily to control his blood sugar. Each time he injects, Mr. Li uses a diabetes management device to record the time and specific dosage of the injection, thus generating an insulin injection record.

[0015] Regarding dietary intake information, Mr. Li can input the types and approximate quantities of food he eats into the diabetes management device before or after each meal. For example, if he ate a slice of whole-wheat bread, an egg, and a glass of milk for breakfast, the device can analyze the nutritional composition of these foods based on its built-in food nutrition database. In this way, the device can obtain Mr. Li's real-time monitoring log data, including blood glucose level data, insulin injection records, and dietary intake information.

[0016] Step S120: Input the real-time monitoring log data into a pre-trained nutrition analysis model for multi-dimensional processing to generate dynamic nutrition demand features. The dynamic nutrition demand features are used to characterize the changing trend of the nutritional element requirements of the diabetic patient within a preset time period.

[0017] Continuing with Mr. Li's example, after his real-time monitoring log data is input into the pre-trained nutritional analysis model, the model performs multi-dimensional processing. First, it segments Mr. Li's blood glucose level data into a time series. Assuming the nutritional analysis model is set to a continuous monitoring period of 4 hours, then from Mr. Li's daily blood glucose monitoring data, the blood glucose data after waking up in the morning to the blood glucose data 4 hours after breakfast constitutes a blood glucose fluctuation segment.

[0018] Then, the injection time and dosage data were extracted from Mr. Li's insulin injection record. For example, Mr. Li injected 10 units of insulin at 9 am. An insulin action time window was constructed. The duration of this insulin action time window was set to about 6 hours based on the type of insulin and Mr. Li's individual metabolic situation. This insulin action time window was then time-domain aligned with the previous blood glucose fluctuation segment.

[0019] The study further analyzes the types of food and the proportions of nutrients in Mr. Li's dietary intake. Assuming that Mr. Li's breakfast of whole-wheat bread is rich in dietary fiber, eggs are rich in protein, and milk is rich in calcium, dietary influencing factors corresponding to the time-domain aligned blood glucose fluctuation segments are generated. Next, the blood glucose fluctuation segments are expanded along the time dimension to calculate the rate of blood glucose change for each monitoring period, such as the speed of blood glucose rise after breakfast, and the trend continuity between adjacent periods, such as whether blood glucose decreases slowly or remains stable from breakfast to lunch, generating time-related coding results. Within the insulin action time window, drug efficacy concentration gradient intervals are divided. Based on the preset drug efficacy decay curve, and considering that Mr. Li injected 10 units of insulin, the drug efficacy coverage intensity of each gradient interval is calculated, generating drug efficacy coverage mapping results.

[0020] Furthermore, dietary influencing factors can be decomposed into nutrient category dimensions. The difference between Mr. Li's current intake ratio and the recommended ratio of a standard diabetic diet model can be compared. For example, is Mr. Li's dietary fiber intake at breakfast higher or lower than the standard recommendation? This generates a quantitative result of nutritional deviation. Then, the time-related encoding results are input into a time attention channel. A sliding time window is used to capture the dependence of blood glucose fluctuations across time periods. Assuming the sliding time window is 2 hours, a time attention weight matrix is ​​generated based on the blood glucose fluctuations within this time window.

[0021] Then, the pharmacodynamic coverage mapping results are input into the pharmacodynamic attention channel. The importance coefficients of different gradient intervals are adjusted based on the temporal overlap between the pharmacodynamic coverage intensity and blood glucose fluctuation segments, generating a pharmacodynamic attention weight matrix. The quantification results of nutritional deviation are input into the nutritional attention channel. Nutritional compensation priorities are determined based on the deviation direction and magnitude of nutrient category dimensions, generating a nutritional attention weight matrix. Subsequently, cross-channel interactive verification is performed on these three weight matrices. The weight values ​​of corresponding time periods in each matrix are fused through weighted summation to generate a multi-channel joint attention coefficient. Based on this multi-channel joint attention coefficient, blood glucose fluctuation segments, insulin action time windows, and dietary influencing factors are dynamically weighted and fused to generate a cross-modal feature vector. The cross-modal feature vector undergoes hierarchical compression and dimension alignment to extract core feature groups reflecting blood glucose regulation needs and nutritional imbalance compensation needs. Finally, the core feature groups are input into a fully connected network for nonlinear transformation to generate dynamic nutritional demand features containing the intensity and temporal distribution patterns of demand for various nutritional elements (such as carbohydrates, proteins, and fats) within a future preset time period (e.g., the next week).

[0022] Step S130: Match the dynamic nutritional demand characteristics with the nutritional service items in the pre-stored digital nutritional database to determine the priority-ranked candidate nutritional service set.

[0023] In this embodiment, the pre-stored digital nutrition database contains numerous nutrition service entries. Taking one such entry as an example, this service is a customized nutrition package delivery service provided by a professional diabetes nutrition service organization. The nutritional composition list of this service includes providing a specific amount of high-quality protein, an appropriate amount of low-carbohydrate, and abundant dietary fiber daily. The service type is identified as regular delivery, and the applicable conditions include a blood glucose threshold range of 3.9-7.2 mmol / L, an insulin sensitivity coefficient within a certain range, and a dietary restriction list that prohibits high-sugar and high-fat foods.

[0024] In detail, the demand intensity curves for each nutrient element within a preset time period (one week) can be extracted from Mr. Li's dynamic nutritional needs. For example, in the coming week, due to increased physical activity, Mr. Li's demand intensity for protein will be higher at certain times, while his demand for carbohydrates will be relatively lower. The distribution pattern of nutrient supply in the nutritional service items can be analyzed to generate a nutritional service supply time series diagram. For example, this nutritional package provides protein-rich foods every morning from Monday to Friday, while the carbohydrate content is lower at noon and in the evening.

[0025] Then, the timestamp sequence of the demand intensity curve is aligned with the effective time range of the nutritional service supply time series diagram to determine the set of overlapping time windows and the corresponding demand intensity and supply values. Within the set of overlapping time windows, for the nutrient element protein, the absolute difference between its demand intensity and supply values ​​in the corresponding time window is calculated to generate a single-window difference metric. It is then checked whether the supply value covers the preset fluctuation tolerance range of the demand intensity value. If it does, a single-window coverage compensation coefficient is generated. The single-window difference metric and the single-window coverage compensation coefficient are weighted and summed to generate a single-window matching index. The single-window matching index of the same nutrient element in all overlapping time windows is integrated and accumulated to generate a global matching score for that nutrient element.

[0026] Based on this, the global matching scores of all nutritional elements (protein, carbohydrates, dietary fiber, etc.) are weighted and fused according to a preset weight ratio to generate an initial matching score. If the list of nutritional elements contains redundant nutritional elements not included in Mr. Li's dynamic nutritional needs, such as a certain specific vitamin, the initial matching score is adjusted by deducting from the number of redundant nutritional elements and a preset penalty coefficient. The supply values ​​of each nutritional element in the nutritional service supply time series diagram are checked to see if they meet the future demand forecast values ​​of the demand intensity curve. If they do, the matching score after deduction is compensated for gain based on the forecast coverage duration and a preset reward coefficient. The compensated matching score is then limited to a preset standardized range to generate the final matching score.

[0027] Then, the current blood glucose level is extracted from Mr. Li's real-time monitoring log data, assumed to be 6.5 mmol / L, and determined to be within the blood glucose threshold range of 3.9-7.2 mmol / L. Based on this, Mr. Li's real-time insulin sensitivity coefficient can be derived using the dosage data in his insulin injection record and a preset sensitivity coefficient calculation model. This coefficient is then compared with the insulin sensitivity coefficient specified in the service application conditions, and the result is found to be within the allowable error range. Next, the food types in Mr. Li's dietary intake information are compared with the dietary restrictions list, and no conflicts are found. Since the matching score of this nutritional service item exceeds the preset threshold and passes the condition compliance verification, it is added to the initial candidate set.

[0028] Understandably, by performing the above operations on all nutrition service entries in the digital nutrition database, the nutrition service entries in the initial candidate set are sorted in descending order according to the matching score, and the entries with the same score are sorted a second time based on the urgency of the service type identifier, thus generating a priority-sorted candidate nutrition service set.

[0029] Step S140: Based on the real-time geographical location information of the diabetic patient and the location information of the service providers in the candidate nutrition service set, a spatial correlation analysis is performed to filter out an accessible subset of target nutrition services.

[0030] For example, the diabetes management device used by Mr. Li has a positioning function, which can obtain his real-time geographical location information, such as his current location at home with latitude and longitude coordinates (X1, Y1). Therefore, the location information of the service provider corresponding to each nutritional service item can be extracted from the candidate nutritional service set.

[0031] Taking the previously mentioned diabetes nutrition service provider as an example, its location information includes latitude and longitude coordinates (X2, Y2) and a service coverage radius of 10 kilometers. Based on Mr. Li's real-time geographical location information (X1, Y1) and the service provider's latitude and longitude coordinates (X2, Y2), the straight-line distance is calculated. Assuming the distance calculation formula yields a distance of 8 kilometers, since 8 kilometers is less than the 10-kilometer service coverage radius, this nutrition service item is marked as an accessible service.

[0032] Based on the above scheme, each item in the candidate nutrition service set can be evaluated, and a subset of target nutrition services can be generated by prioritizing the accessible services. For each accessible service, real-time traffic data can be added to estimate the service arrival time. For example, the map service interface can be called to obtain the route planning results between Mr. Li's home (X1, Y1) and the nutrition service provider (X2, Y2). The route planning results show that the estimated travel time is 30 minutes if driving and 45 minutes if taking public transportation. Based on the frequently used travel preference data stored in Mr. Li's terminal device, it is found that Mr. Li prefers to drive, so 30 minutes is selected as the basic arrival time. Then, based on real-time traffic flow data, if the current traffic flow is high, the basic arrival time is dynamically adjusted. Assuming the adjusted service arrival time is 35 minutes, this adjusted service arrival time is then bound to the corresponding accessible service.

[0033] Step S150: Send the target nutrition service subset to the terminal device corresponding to the diabetic patient according to a preset push strategy, and receive the service selection feedback returned by the terminal device to trigger the service execution instruction.

[0034] For example, Mr. Li's terminal device is his smartphone. Therefore, a push notification time sequence can be generated based on the priority and estimated arrival time of each target nutrition service within the target nutrition service subset. For instance, if there are three services in the target nutrition service subset, ranked from highest to lowest priority: nutrition meal delivery service, diabetes nutrition consultation service, and diabetes exercise guidance service, with estimated arrival times of 35 minutes, 2 hours, and 4 hours respectively, the effective push period is selected based on the time window overlapping with Mr. Li's terminal device activity. Assuming Mr. Li uses his phone most frequently between 9:00 AM and 6:00 PM, then during this time period, notification messages containing service details and arrival times are sent to Mr. Li's phone in descending order of priority. The notification message for the nutrition meal delivery service is sent first, including information such as the types of food included, nutritional components, and estimated delivery time of 35 minutes. If no service selection feedback is received from Mr. Li within a preset waiting time (e.g., 10 minutes), a backup push channel is automatically triggered.

[0035] Furthermore, it can also detect whether Mr. Li's phone's network connection status is active. If it is active and no feedback is received, the notification message is resent via the SMS service channel. If the network connection status is inactive, the push operation is delayed until the next valid push period (e.g., between 2 PM and 3 PM) and the number of push failures is recorded.

[0036] When the number of failed push notifications exceeds a preset limit (e.g., 3 times), a service interruption alert is sent to a preset emergency contact (e.g., Mr. Li's family). After seeing the notification, Mr. Li selects the nutrition package delivery service and provides the service identifier and execution time requirement on his phone, such as requesting delivery within 30 minutes. The system can then parse the service identifier and execution time requirement from the service selection feedback, extract the corresponding service provider's interface address from the target nutrition service subset based on the service identifier, encapsulate the execution time requirement and the service provider's interface address into a service request message, and send it to the service provider's server via an application programming interface. Upon receiving the request, the service provider's server returns a service confirmation signal, generating a confirmation message containing service details and an execution timestamp, which is then sent to Mr. Li's phone.

[0037] Based on the above steps, this embodiment of the application acquires real-time monitoring log data and uses a pre-trained nutritional analysis model for multi-dimensional processing to generate dynamic nutritional demand characteristics that characterize the changing trends of a patient's nutritional element needs within a preset time period. This changes the previous static and singular nutritional assessment model, enabling timely capture of changes in a patient's nutritional needs caused by factors such as blood glucose levels, insulin injections, and dietary intake. Next, by integrating the dynamic nutritional demand characteristics with a digital nutritional database, a priority-ranked set of candidate nutritional services is determined. Then, spatial correlation analysis is performed using the patient's real-time geographic location information and the service provider's location information to filter out an accessible subset of target nutritional services. This not only accurately matches services that meet the patient's nutritional needs from a massive number of nutritional service items but also ensures that these services are geographically accessible to the patient, greatly improving the practicality and accessibility of nutritional services. Finally, after sending the target nutritional service subset to the patient's terminal device according to the preset push strategy, the service selection feedback returned by the terminal device is received to trigger the service execution command. This makes the nutritional service push no longer a one-way information transmission, but can adjust and optimize the service in a timely manner based on the patient's feedback, further improving the patient's acceptance and compliance with nutritional services. This provides diabetic patients with a more scientific, efficient, personalized and continuously optimized digital nutritional service experience, which helps to better manage the patient's nutritional status and assist in the treatment and rehabilitation of diabetes.

[0038] In one possible implementation, step S120 includes: step S121, performing time series segmentation on the blood glucose level data to obtain blood glucose fluctuation segments for multiple consecutive monitoring periods.

[0039] For example, Mr. Li's blood glucose monitoring device records his blood glucose values ​​at fixed intervals, assuming each continuous monitoring period is 3 hours. The first period is from 6:00 AM to 9:00 AM, the second period is from 9:00 AM to 12:00 PM, and so on. This divides the day's blood glucose level data into multiple segments of blood glucose fluctuation. For example, Mr. Li's blood glucose level is 5.8 mmol / L at 6:00 AM and 6.2 mmol / L at 9:00 AM. This fluctuation within the timeframe can be considered a single segment of blood glucose fluctuation.

[0040] Step S122: Extract the injection time point and dose data from the insulin injection record, construct the insulin action time window, and align the insulin action time window with the blood glucose fluctuation segment in the time domain.

[0041] For example, Mr. Li injected 8 units of insulin at 10:00 AM. Based on the characteristics of this insulin, its effective time window is set to 4 hours. This effective time window is then time-domain aligned with the blood glucose fluctuation segment that includes the 10:00 AM injection time. For instance, the blood glucose fluctuation segment mentioned earlier from 9:00 AM to 12:00 PM includes the 10:00 AM injection time, thus achieving time-domain alignment.

[0042] Step S123: Analyze the food types and nutrient ratios in the dietary intake information to generate dietary influencing factors corresponding to the blood glucose fluctuation segments aligned with the time domain.

[0043] For example, Mr. Li ate lunch at 11:00 AM, consisting of 100 grams of brown rice, 200 grams of vegetables, and 50 grams of chicken. Brown rice primarily contains carbohydrates, vegetables are rich in vitamins and dietary fiber, and chicken contains protein and other nutrients. Based on the nutritional proportions of these foods—for example, brown rice contains 70% carbohydrates, vegetables contain a variety of vitamins, and chicken contains 20% protein—and combined with Mr. Li's portion sizes, a dietary impact factor corresponding to the blood glucose fluctuation segment aligned with the 9:00 AM to 12:00 PM timeframe is generated. This dietary impact factor reflects the potential impact of food intake on blood glucose fluctuations.

[0044] Step S124: Expand the blood glucose fluctuation segment in time dimension, extract the rate of blood glucose change in each monitoring period and the trend continuity between adjacent periods, and generate time correlation coding results.

[0045] For example, during the period from 9:00 AM to 12:00 PM, Mr. Li's blood glucose level rose from 6.0 mmol / L at 9:00 AM to 6.5 mmol / L at 12:00 PM. The rate of change of blood glucose during this period was calculated as a constant value. Next, adjacent time periods were examined. For example, if the trend was continuous with the previous time period (6:00 AM to 9:00 AM), and if the blood glucose level was stable in the previous time period but rose in this time period, this trend change was recorded. All this information was combined to generate a time-related coding result.

[0046] Step S125: Divide the drug efficacy concentration gradient intervals within the insulin action time window, calculate the drug efficacy coverage intensity of each gradient interval based on the injection dose data and the preset drug efficacy decay curve, and generate the drug efficacy coverage mapping result.

[0047] For example, for Mr. Li's injection of 8 units of insulin at 10:00 AM, within its 4-hour effective time window, the concentration is divided into several gradient intervals, such as 0-1 hours as the high concentration interval, 1-2 hours as the medium concentration interval, and 2-4 hours as the low concentration interval. Based on the preset efficacy decay curve, since 8 units of insulin were injected, the efficacy coverage intensity is calculated to be higher in the high concentration interval, followed by the medium concentration interval, and lower in the low concentration interval, thus generating the efficacy coverage range mapping result.

[0048] Step S126: Decompose the dietary influencing factors into nutrient category dimensions, compare the difference between the current intake ratio and the recommended ratio of the standard diabetic diet model, and generate a quantitative result of nutritional deviation.

[0049] For example, the proportions of carbohydrates, protein, and vitamins in Mr. Li's lunch are compared with those recommended by a standard diabetic diet model. Assuming the standard model recommends carbohydrates make up 40% of total calories at lunch, while Mr. Li's actual brown rice intake contains 50% carbohydrates, this difference is calculated. This comparison is performed for all nutrient categories, and the magnitude and direction of the difference are used to generate a quantification of nutritional bias.

[0050] Step S127: Input the time correlation encoding result into the time attention channel, capture the blood glucose fluctuation dependence across time periods through a sliding time window, and generate a time attention weight matrix.

[0051] For example, setting the sliding time window to 2 hours, looking at Mr. Li's blood sugar fluctuations throughout the day, when the sliding time window is between 9:00 and 11:00, the blood sugar fluctuations during this period are dependent on the blood sugar fluctuations of the preceding 2-hour period (7:00-9:00) and the following 2-hour period (11:00-13:00). Based on the strength of this dependency, for example, if the blood sugar fluctuations between 9:00 and 11:00 are highly correlated with those between 11:00 and 13:00, a higher weight will be assigned to the corresponding position in the time attention weight matrix, thus generating the time attention weight matrix.

[0052] Step S128: Input the drug efficacy coverage mapping result into the drug efficacy attention channel, adjust the importance coefficients of different gradient intervals based on the temporal overlap between drug efficacy coverage intensity and blood glucose fluctuation segments, and generate a drug efficacy attention weight matrix.

[0053] For example, in the mapping results of Mr. Li's insulin efficacy coverage, the interval with high efficacy coverage intensity (such as the high concentration interval of 0-1 hour) has a high degree of temporal overlap with the blood glucose fluctuation segment (9 o'clock-12 o'clock), which will increase the importance coefficient of this interval in the efficacy attention weight matrix, while the interval with low overlap will decrease the coefficient, thus generating the efficacy attention weight matrix.

[0054] Step S129: Input the nutrient deviation quantification result into the nutrient attention channel, determine the nutrient compensation priority based on the deviation direction and magnitude of the nutrient category dimension, and generate a nutrient attention weight matrix.

[0055] Taking Mr. Li's lunch as an example, his carbohydrate intake ratio is higher than the standard recommended ratio. The deviation is positive and the magnitude is large. Therefore, in the nutrition attention weight matrix, the priority of nutritional compensation for carbohydrates is low, while the priority of compensation for insufficient nutrients (such as certain vitamins) is high. This is how the nutrition attention weight matrix is ​​generated.

[0056] Step S1210: Perform cross-channel interactive verification on the time attention weight matrix, drug efficacy attention weight matrix and nutrition attention weight matrix, and fuse the weight values ​​of the corresponding time periods in each matrix by weighted summation to generate multi-channel joint attention coefficients.

[0057] Suppose that in a certain time period, the weight of that time period in the time attention weight matrix is ​​0.3, in the efficacy attention weight matrix it is 0.4, and in the nutrition attention weight matrix it is 0.3. Following a preset weighted summation rule (such as setting different weights based on the importance of each matrix), the multi-channel joint attention coefficient for that time period is calculated. This calculation is performed for all time periods to obtain the complete multi-channel joint attention coefficient.

[0058] Step S1211: Dynamically weighted and fused the blood glucose fluctuation segment, insulin action time window, and dietary influencing factors according to the multi-channel joint attention coefficient to generate a cross-modal feature vector.

[0059] For example, for the period from 9 am to 12 pm, the multi-channel joint attention coefficient is assumed to be 0.5. Based on this coefficient, the blood glucose fluctuation segments, insulin action time window and dietary influencing factors of this period are weighted and fused. The fused result together with similar results from other periods constitutes a cross-modal feature vector.

[0060] Step S1212: Perform hierarchical compression and dimension alignment on the cross-modal feature vector to extract the core feature group that reflects the needs of blood glucose regulation and nutritional imbalance compensation.

[0061] Cross-modal feature vectors contain information from multiple dimensions and levels. By using hierarchical compression, some unimportant information is removed, and information from different dimensions is aligned. For example, information related to blood sugar regulation and information related to nutritional imbalance compensation are organized on the same dimension, and core feature groups reflecting the needs of blood sugar regulation and nutritional imbalance compensation are extracted, such as the need to reduce carbohydrate intake and supplement specific vitamins when blood sugar is too high.

[0062] Step S1213: Input the core feature group into a fully connected network for nonlinear transformation to generate dynamic nutritional demand features that include the intensity and time-series distribution of the demand for each nutrient element within a future preset time period.

[0063] For example, based on the information in the core feature group, the fully connected network performs a nonlinear transformation to calculate the dynamic nutritional demand characteristics of Mr. Li in the coming week, such as the low demand intensity of carbohydrates from Monday to Wednesday due to low activity levels, while the demand intensity of protein remains stable, and the increased demand intensity of both carbohydrates and protein from Thursday to Friday when activity levels increase.

[0064] In one possible implementation, the pre-trained nutritional analysis model is trained through the following steps: Step S210, collecting a monitoring log dataset of historical diabetic patients, the monitoring log dataset including historical blood glucose data, historical insulin use records, and historical dietary records.

[0065] In this embodiment, monitoring log datasets from numerous patients with a history of diabetes (including but not limited to patients like Mr. Li) can be collected. Taking Mr. Li as an example, his historical blood glucose data is recorded by daily blood glucose monitoring devices, covering blood glucose measurements from different dates and times. For instance, his historical blood glucose data consists of multiple daily measurements taken over the past month, such as fasting blood glucose in the morning, blood glucose before and after meals, and blood glucose before bedtime. His historical insulin usage records include the timestamp (accurate to the minute) and dosage (units) of each insulin injection, such as the specific dosage (e.g., 10 units, 8 units, etc.) he injected at certain times in the past (e.g., 8 am, 7 pm, etc.). His historical diet records detail the types and approximate quantities of food consumed at each meal, such as an egg, two slices of bread, and a glass of milk for breakfast, and brown rice, vegetables, and meat for lunch. This information comprehensively constitutes Mr. Li's historical diet record. By combining similar data from numerous patients like Mr. Li, a monitoring log dataset for model training is formed.

[0066] Step S220: Perform outlier detection and interpolation on the historical blood glucose data to generate a standardized blood glucose sequence, and encode the timestamps and doses in the historical insulin usage records to generate an insulin action intensity matrix.

[0067] For example, suppose that in a patient's historical blood glucose data, due to equipment malfunction or measurement error, an extremely high or extremely low blood glucose value appears (such as 30 mmol / L or 1.0 mmol / L, significantly deviating from the normal range). This value is identified using an outlier detection algorithm. For these outliers, interpolation is used. Based on the trend of surrounding normal blood glucose values, an appropriate interpolation method (such as linear interpolation) is used to replace the outlier, thereby generating a standardized blood glucose sequence.

[0068] The timestamps and doses in historical insulin usage records are encoded. Taking Mr. Li's historical insulin usage record as an example, each injection timestamp (e.g., 8:00 AM) is converted into a numerical form according to a certain encoding rule, and the dose (e.g., 10 units) is also encoded accordingly. Assuming that based on factors such as time sequence and dose size, this encoded information is constructed into an insulin action intensity matrix, where the elements reflect information related to insulin action intensity at different time points.

[0069] Step S230: Extract the food component vector of each meal from the historical diet record, and map the food component vector into a nutrient intake distribution matrix according to the preset nutrient conversion rules.

[0070] For example, for Mr. Li's brown rice, vegetables, and meat in a lunch, the main components of brown rice (such as the proportion of carbohydrates and protein), the content of various vitamins and dietary fiber in vegetables, and the protein and fat content in meat are converted into vector form based on the food nutrition database, thus obtaining the food component vector for each meal.

[0071] According to the preset nutrient conversion rules, food component vectors are mapped into nutrient intake distribution matrices. Following the established conversion rules, the nutrient information from the food component vectors of each meal is integrated into a matrix. The rows of this matrix represent different nutrient categories (such as carbohydrates, proteins, vitamins, etc.), and the columns represent different time windows (such as breakfast, lunch, dinner, etc.). The elements in the matrix represent the intake distribution of various nutrients within the corresponding time period.

[0072] Step S240: Construct an initial neural network model by using the standardized blood glucose sequence, insulin action intensity matrix, and nutrient intake distribution matrix as input features. Perform temporal modeling on the input features through temporal convolutional layers and gated recurrent units to generate intermediate feature vectors.

[0073] Step S250: An adaptive weight allocation module is introduced into the initial neural network model to dynamically adjust the weights of each feature channel according to the dimensional correlation of the intermediate feature vector, and a contrastive learning strategy is used to optimize the similarity between the intermediate feature vector and the preset nutritional requirement label until the initial neural network model converges.

[0074] For example, if some dimensions of the intermediate feature vector are strongly correlated with blood glucose regulation, while others are more strongly correlated with nutritional intake balance, the weights of the corresponding feature channels can be dynamically increased or decreased based on the magnitude of this correlation. A contrastive learning strategy is employed to optimize the similarity between the intermediate feature vector and preset nutritional requirement labels. These preset nutritional requirement labels represent the ideal nutritional needs of diabetic patients under different conditions, determined based on extensive medical research and clinical experience. By continuously adjusting the parameters of the initial neural network model, the similarity between the intermediate feature vector and these preset nutritional requirement labels is gradually increased until the initial neural network model converges, meaning that the performance indicators (such as prediction accuracy) of the initial neural network model no longer significantly improve with training.

[0075] For example, in one possible implementation, step S240 includes: performing time-division slicing on the standardized blood glucose sequence, dividing the standardized blood glucose sequence into multiple blood glucose time-series segments of equal length according to fixed time intervals. Applying multiple parallel one-dimensional convolutional kernels to each blood glucose time-series segment for local feature extraction, each one-dimensional convolutional kernel sliding along the time axis and capturing blood glucose fluctuation patterns at different scales, generating multi-scale blood glucose convolutional features. Concatenating the multi-scale blood glucose convolutional features along the channel dimension to form a blood glucose multi-scale fusion feature tensor. Aligning the insulin action intensity matrix along the time axis to the starting time point of the blood glucose time-series segment, extracting the insulin dose distribution vector within the time range corresponding to each blood glucose time-series segment, and performing a nonlinear transformation on the insulin dose distribution vector, mapping the dose values ​​to a high-dimensional space through a fully connected layer to generate an insulin dose embedding vector. Dividing the nutrient intake distribution matrix into nutrient intake sub-matrices corresponding to the blood glucose time-series segments according to time windows, and performing a weighted summation across nutrient categories on each nutrient intake sub-matrices to generate a time-segment nutrient intake summary vector. The blood glucose multi-scale fusion feature tensor, insulin dose embedding vector, and time-segment nutrient intake summary vector are concatenated by time step to generate a multimodal temporal fusion feature vector. This multimodal temporal fusion feature vector is then input into a bidirectional gated recurrent unit. The hidden state of each time step is calculated along the forward time direction to capture historical dependencies, and along the reverse time direction to capture future dependencies. The forward and reverse hidden states of the same time step are added element-wise to generate a time-step fusion hidden state. All time-step fusion hidden states are then arranged in their original chronological order to form a preliminary temporal feature sequence. A temporal self-attention mechanism is applied to this preliminary temporal feature sequence to calculate the association weight between each time-step feature and the features of other time steps, generating a temporal attention weight distribution. The preliminary temporal feature sequence is then dynamically weighted and summed according to the temporal attention weight distribution to generate a temporally enhanced feature sequence. The time-enhanced feature sequence is input into a multilayer perceptron for cross-time-step feature interaction. A nonlinear activation function is used to learn higher-order correlation patterns between time steps, generating a time-interaction feature vector. This vector is then subjected to hierarchical normalization to generate a standardized temporal feature vector. This standardized vector is input into a densely connected layer for feature compression, generating a dimensionality-reduced compact temporal feature. The compact temporal feature is then smoothed over time, and a moving average algorithm is used to eliminate short-term noise interference, generating a denoised temporal feature sequence. This denoised temporal feature sequence is then residually connected to the original multimodal temporal fusion feature vector to supplement detail information and prevent gradient vanishing, generating the final enhanced temporal feature.A global max pooling operation is performed on the final time-series features to extract the salient feature peaks across the entire time range, generating a peak feature vector. A global average pooling operation is then performed on the final time-series features to extract the statistical distribution characteristics across the entire time range, generating a statistical feature vector. The peak feature vector and the statistical feature vector are horizontally concatenated to generate a comprehensive time-series representation vector. This comprehensive time-series representation vector is then input into a multi-layer fully connected network for higher-order feature transformation to generate the intermediate feature vector.

[0076] For example, the step of performing time-dimensional slicing on the standardized blood glucose sequence, dividing the standardized blood glucose sequence into multiple blood glucose time-series segments of equal length according to fixed time intervals, includes: determining the number of consecutive monitoring points contained in each blood glucose time-series segment according to a preset segment length parameter; traversing the standardized blood glucose sequence in a sliding window manner, and extracting consecutive subsequences as blood glucose time-series segments according to the segment length parameter; when the number of remaining monitoring points is insufficient to fill a complete segment, padding with zeros to the end of the segment length parameter; and adding a timestamp to each blood glucose time-series segment to record its position information in the original sequence. For example, the step of applying multiple parallel one-dimensional convolutional kernels to each blood glucose time-series segment for local feature extraction, where each one-dimensional convolutional kernel slides along the time axis and captures blood glucose fluctuation patterns at different scales to generate multi-scale blood glucose convolutional features, includes: configuring multiple one-dimensional convolutional layers with different kernel widths, each convolutional layer corresponding to feature extraction at a specific time scale; inputting the same blood glucose time-series segment into each one-dimensional convolutional layer, and extracting local blood glucose change features within different time windows through convolution operations. Independent max pooling is performed on the output of each convolutional layer to preserve salient feature responses at each scale. The pooled outputs of each convolutional layer are then stacked along the channel dimension to form a set of blood glucose convolutional features containing multi-scale information.

[0077] For example, the step of inputting the multimodal temporal fusion feature vector into a bidirectional gated recurrent unit, calculating the hidden state at each time step along the forward time direction to capture historical dependencies, and calculating the hidden state at each time step along the reverse time direction to capture future dependencies, includes: inputting the multimodal temporal fusion feature vector into a forward gated recurrent unit in chronological order, and updating the hidden state at each time step sequentially, wherein the calculation of the current hidden state depends on the hidden state of the previous time step. Inputting the multimodal temporal fusion feature vector into a backward gated recurrent unit in reverse chronological order, and updating the hidden state at each time step sequentially, wherein the calculation of the current hidden state depends on the hidden state of the next time step. In the forward gated recurrent unit, a combination of the sigmoid and tanh functions is used to calculate the update gate, reset gate, and candidate hidden states. In the backward gated recurrent unit, the same gating mechanism is used, but the feature vector is processed in reverse chronological order.

[0078] For example, applying a temporal self-attention mechanism to the initial temporal feature sequence to calculate the association weights between each time step feature and the features of other time steps, and generating a temporal attention weight distribution, includes: generating a query vector matrix, a key vector matrix, and a value vector matrix from the initial temporal feature sequence through three independent linear transformation layers; calculating the dot product between the transpose of the query vector matrix and the key vector matrix to generate an original attention score matrix; scaling the original attention score matrix and applying a softmax function to generate a normalized attention weight matrix; multiplying the normalized attention weight matrix by the value vector matrix to generate a weighted context feature vector; and adding the context feature vector to the original initial temporal feature sequence to form an enhanced feature representation with global temporal dependencies.

[0079] For example, the step of smoothing the compact time-series features over time and using a moving average algorithm to eliminate short-term noise interference to generate a denoised time-series feature sequence includes: setting a sliding window size and sliding the window along the time axis to the compact time-series features; calculating the arithmetic mean of the features within each window and replacing the feature value at the center of the window with this average; performing edge compensation on the processed feature sequence and using mirror filling to handle window out-of-bounds situations at both ends of the sequence; and repeating the moving average operation multiple times until the fluctuation amplitude of the feature sequence is lower than a preset threshold.

[0080] In this embodiment, the previously generated standardized blood glucose sequence, insulin action intensity matrix, and nutrient intake distribution matrix can be used as input features. These input features carry important information about the diabetic patient's blood glucose fluctuations, insulin action, and nutrient intake over a historical period.

[0081] Next, the standardized blood glucose sequence is sliced ​​in the time dimension. Based on a preset segment length parameter, assuming each segment consists of four consecutive monitoring points, taking a patient's standardized blood glucose sequence as an example, starting from the first monitoring point, a sliding window is used to extract four consecutive monitoring points at a time as a blood glucose time sequence segment. When the number of remaining monitoring points is insufficient to fill a complete segment (e.g., only two monitoring points remain), zeros are padded to four monitoring points. A timestamp is added to each blood glucose time sequence segment to record its position in the original sequence.

[0082] Next, multiple parallel one-dimensional convolutional kernels are applied to each blood glucose time series segment for local feature extraction. Multiple one-dimensional convolutional layers with different kernel widths (e.g., three kernels with widths of 3, 5, and 7) are configured, each corresponding to feature extraction at a specific time scale. The same blood glucose time series segment is input into these three one-dimensional convolutional layers, and local blood glucose change features within different time windows are extracted through convolution operations (e.g., a convolutional kernel with a width of 3 extracts local blood glucose change features within three consecutive monitoring point time windows). Independent max-pooling is performed on the output of each convolutional layer; for example, in the output of a convolutional layer with a width of 3, the maximum value is retained as the salient feature response at that scale. The pooled outputs of each convolutional layer are stacked along the channel dimension to form a blood glucose convolutional feature set containing multi-scale information.

[0083] Next, the insulin action intensity matrix is ​​aligned along the time axis to the starting time point of the blood glucose time series segment. Taking Mr. Li's insulin action intensity matrix as an example, assuming the blood glucose time series segment starts at 8:00 AM, the insulin dose distribution information corresponding to 8:00 AM and later is extracted from the insulin action intensity matrix. For the insulin dose distribution vector within the time range corresponding to each blood glucose time series segment, such as the insulin dose distribution vector corresponding to the blood glucose time series segment from 8:00 AM to 12:00 PM, a nonlinear transformation is performed on it. The dose values ​​(such as 8 units, 10 units, etc.) are mapped to a high-dimensional space through a fully connected layer to generate an insulin dose embedding vector.

[0084] Next, the nutrient intake distribution matrix is ​​divided into nutrient intake sub-matrices corresponding to blood glucose time series segments according to time windows. For example, the nutrient intake distribution matrix is ​​divided according to the same time division method as the blood glucose time series segments (such as time windows corresponding to 4 monitoring points), resulting in nutrient intake sub-matrices corresponding to each blood glucose time series segment. A weighted summation is then performed across nutrient categories for each nutrient intake sub-matrix. For example, for nutrient categories such as carbohydrates, proteins, and vitamins, different weights are assigned according to their importance in diabetes nutrition, and a time-segment nutrient intake summary vector is calculated.

[0085] Next, the blood glucose multi-scale fusion feature tensor, insulin dose embedding vector, and time-segment nutrient intake summary vector are concatenated at each time step to generate a multimodal temporal fusion feature vector. For example, at a certain time step, the elements of the corresponding blood glucose multi-scale fusion feature tensor, the value of the insulin dose embedding vector, and the value of the time-segment nutrient intake summary vector are concatenated in sequence to form a multimodal temporal fusion feature vector.

[0086] Next, the multimodal temporal fusion feature vector is input into the bidirectional gated recurrent unit. The hidden state at each time step is calculated along the forward temporal direction to capture historical dependencies. The multimodal temporal fusion feature vector is then input into the forward gated recurrent unit in chronological order. For each time step, the calculation of the current hidden state depends on the hidden state of the previous time step. The update gate, reset gate, and candidate hidden states are calculated using a combination of the sigmoid and tanh functions. For example, when calculating the hidden state at time step t, it is determined based on the hidden state of the previous time step (t-1), the currently input multimodal temporal fusion feature vector, and the values ​​of the update and reset gates calculated using the sigmoid and tanh functions. The hidden state at each time step is then calculated along the reverse temporal direction to capture future dependencies. The multimodal temporal fusion feature vector is then input into the backward gated recurrent unit in reverse temporal order, using the same gating mechanism as the forward gated recurrent unit but processing the feature vector in reverse time step order.

[0087] Next, the forward and reverse hidden states at the same time step are added element-wise to generate a time-step fused hidden state. For example, for the t-th time step, the corresponding elements of the forward hidden state obtained from the forward-gated loop unit and the reverse hidden state obtained from the reverse-gated loop unit are added to obtain the time-step fused hidden state. All time-step fused hidden states are arranged in their original time order to form a preliminary temporal feature sequence.

[0088] Next, a temporal self-attention mechanism is applied to the initial temporal feature sequence. The initial temporal feature sequence is passed through three independent linear transformation layers to generate a query vector matrix, a key vector matrix, and a value vector matrix. The dot product between the transpose of the query vector matrix and the key vector matrix is ​​calculated to generate the original attention score matrix. The original attention score matrix is ​​scaled (e.g., divided by the square root of the feature dimension) and a softmax function is applied to generate a normalized attention weight matrix. The normalized attention weight matrix is ​​multiplied by the value vector matrix to generate a weighted context feature vector. The context feature vector is added to the original initial temporal feature sequence to form an enhanced feature representation with global temporal dependencies.

[0089] Next, the preliminary temporal feature sequence is dynamically weighted and summed according to the temporal attention weight distribution to generate a temporally enhanced feature sequence. Then, the features at each time step in the preliminary temporal feature sequence are weighted and summed according to the weight values ​​in the calculated temporal attention weight distribution to obtain the temporally enhanced feature sequence.

[0090] Next, the temporally augmented feature sequence is input into a multilayer perceptron for cross-time step feature interaction. A high-order correlation pattern between time steps is learned using a non-linear activation function (such as ReLU), generating a temporally interactive feature vector. This vector is then subjected to hierarchical normalization to generate a standardized temporal feature vector. This standardized feature vector is then input into a densely connected layer for feature compression, generating a dimensionality-reduced, compact temporal feature.

[0091] Next, the compact temporal features are smoothed along the time dimension. A sliding window size of 3 (for example) is set, and the compact temporal features are slid along the time axis. Within each window, the arithmetic mean of the features is calculated, and this mean is used to replace the feature value at the center of the window. Edge compensation is performed on the processed feature sequence, using mirror filling to handle window overflows at both ends of the sequence. For example, at the beginning and end of the sequence, the feature values ​​closest to the edge are mirrored to fill the gaps. This sliding average operation is repeated multiple times until the fluctuation amplitude of the feature sequence is below a preset threshold, generating a denoised temporal feature sequence. The denoised temporal feature sequence is then residually concatenated with the original multimodal temporal fusion feature vector to supplement detail information and prevent gradient vanishing, generating the enhanced final temporal features.

[0092] Next, a global max pooling operation is performed on the final temporal features to extract the salient feature peaks across the entire time range, generating peak feature vectors. For example, the maximum value in each feature dimension is found across the entire time range, and these maximum values ​​form the peak feature vector. Simultaneously, a global average pooling operation is performed on the final temporal features to extract the statistical distribution characteristics across the entire time range, generating statistical feature vectors. The peak feature vectors and statistical feature vectors are then concatenated horizontally to generate a comprehensive temporal representation vector. This comprehensive temporal representation vector is then input into a multi-layer fully connected network for higher-order feature transformation, generating intermediate feature vectors.

[0093] In one possible implementation, step S130 includes: step S131, reading multiple nutrition service entries from the digital nutrition database, each nutrition service entry containing a service type identifier, a list of nutritional elements, and applicable service conditions.

[0094] Considering Mr. Li, the diabetic patient mentioned earlier, his dynamic nutritional needs characteristics have been generated through the previous process, and the pre-stored digital nutrition database contains numerous nutritional service items for diabetic patients.

[0095] Taking one of the nutrition services as an example, this is a comprehensive nutrition service provided by a specific nutrition institution. Its service type is identified as "customized nutrition package delivery and health consultation". The list of nutritional elements includes providing a certain amount of high-quality protein (such as 50-60 grams), low carbohydrates (such as 100-120 grams), and abundant dietary fiber (such as 25-30 grams) every day. The service is applicable under the following conditions: blood glucose threshold range of 3.9-7.2 mmol / L, insulin sensitivity coefficient requirement between 0.8 and 1.2, and dietary restrictions that prohibit high-sugar and high-fat foods (such as candy, fried foods, etc.).

[0096] Step S132: For each nutrition service item, calculate the matching score between the list of nutrient elements and the trend of nutrient element demand changes in the dynamic nutrition demand characteristics, and verify the compliance of the service application conditions with the real-time monitoring log data of the diabetic patient.

[0097] Step S133: If the matching score exceeds the preset threshold and passes the condition compliance verification, the corresponding nutrition service item is added to the initial candidate set.

[0098] Step S134: Sort the nutrition service items in the initial candidate set in descending order according to the matching score, and sort the items with the same score in a second order based on the urgency of the service type identifier, so as to generate the priority-sorted candidate nutrition service set.

[0099] In one possible implementation, step S132 includes: step S1321, extracting the demand intensity curve of each nutrient element within a preset time period from the dynamic nutritional demand characteristics, wherein the demand intensity curve includes demand intensity values ​​under multiple time windows and corresponding timestamp sequences.

[0100] Step S1322: Analyze the supply distribution pattern of each nutrient element in the nutrient element composition list to generate a nutrient service supply time sequence diagram that includes supply quantity values ​​and effective time ranges.

[0101] Step S1323: Align the timestamp sequence of the demand intensity curve with the effective time range of the nutrition service supply time series diagram to determine the set of overlapping time windows and the corresponding demand intensity value and supply quantity value.

[0102] Within the set of overlapping time windows, the following operations are performed for each nutrient element: Calculate the absolute difference between the demand intensity value and the supply value of the nutrient element in the corresponding time window to generate a single-window difference metric. Detect whether the supply value covers a preset fluctuation tolerance range of the demand intensity value; if so, generate a single-window coverage compensation coefficient. Calculate a weighted sum of the single-window difference metric and the single-window coverage compensation coefficient to generate a single-window matching index. Integrate and accumulate the single-window matching indices of the same nutrient element across all overlapping time windows to generate a global matching score for the nutrient element. Calculate and merge the global matching scores of all nutrient elements according to a preset weight ratio to generate an initial matching score. Identify whether there are redundant nutrient elements in the nutrient element composition list that are not included in the dynamic nutrient demand characteristics; if so, deduct and adjust the initial matching score according to the number of types of redundant nutrient elements and a preset penalty coefficient. Detect whether the supply value of each nutrient element in the nutrient service supply time series diagram meets the future demand forecast value of the demand intensity curve; if so, compensate the deducted matching score with a gain based on the predicted coverage duration and a preset reward coefficient. The gain-compensated matching score is limited to a preset standardized range to generate the final matching score.

[0103] For example, extract the demand intensity curves for each nutrient element within a preset time period (e.g., the next week) from Mr. Li's dynamic nutritional needs. Assume that due to increased physical activity recently, Mr. Li's protein demand intensity is higher from Monday to Wednesday, approximately 70 grams per day, and slightly lower from Thursday to Sunday, approximately 50 grams per day, with each demand intensity value corresponding to a specific timestamp sequence. Analyze the supply distribution pattern of each nutrient element in this nutritional service item to generate a nutritional service supply time series diagram. For this nutritional service, 55 grams of protein are provided daily from Monday to Friday, and 50 grams of protein are provided on Saturday and Sunday, with the effective time range for each supply clearly defined. Align the timestamp sequences of the demand intensity curves with the effective time ranges of the nutritional service supply time series diagram to determine the overlapping time window set and the corresponding demand intensity and supply values. Within the overlapping time window set, for the nutrient element protein, calculate the absolute difference between its demand intensity value and supply value in the corresponding time window to generate a single-window difference metric. For example, in the Monday time window, Mr. Li needs 70 grams of protein, while the service supply is 55 grams, with an absolute difference of 15 grams. This is the single-window difference metric. The system checks whether the supply value covers the preset fluctuation tolerance range of the demand intensity value. If it does, a single-window coverage compensation coefficient is generated. Assuming Mr. Li's protein demand fluctuation tolerance range is ±10 grams, since 55 grams falls within the range of 60-80 grams (70 grams ± 10 grams), a corresponding single-window coverage compensation coefficient is generated. The single-window difference metric and the single-window coverage compensation coefficient are weighted and summed to generate a single-window matching index. The single-window matching indices for the same nutrient element (protein) across all overlapping time windows are integrated and accumulated to generate a global matching score for that nutrient element. This process is performed on all nutrient elements (such as protein, carbohydrates, dietary fiber, etc.), and the global matching scores of all nutrient elements are weighted and fused according to a preset weight ratio (e.g., set based on the importance of nutritional needs for diabetic patients) to generate an initial matching score. The system identifies redundant nutrients not included in Mr. Li's dynamic nutritional needs from the nutrient composition list. For example, it might include a specific micronutrient not covered by Mr. Li's dynamic nutritional needs. The initial matching score is adjusted by deducting from the number of redundant nutrients (1) and a preset penalty coefficient (e.g., 0.1). The system checks if the supply values ​​of each nutrient in the nutritional service supply time series graph meet the future demand forecasts from the demand intensity curve. If so, it compensates for the deducted matching score by applying a gain based on the predicted coverage period and a preset reward coefficient (e.g., a predicted coverage period of 5 days and a reward coefficient of 0.05). The compensated matching score is then limited to a preset standardized range (e.g., 0-1) to generate the final matching score.

[0104] Furthermore, the service application conditions include a blood glucose threshold range, an insulin sensitivity coefficient, and a dietary restriction list. The condition compliance verification includes the following steps: extracting the current blood glucose value from the real-time monitoring log data of the diabetic patient and determining whether the current blood glucose value is within the blood glucose threshold range; deriving the real-time insulin sensitivity coefficient of the diabetic patient based on the dosage data in the insulin injection record and a preset sensitivity coefficient calculation model, and comparing the difference with the insulin sensitivity coefficient in the service application conditions to obtain a difference comparison result; comparing the food types in the dietary intake information with the dietary restriction list to detect any conflicts; if the current blood glucose value is within the blood glucose threshold range, the difference comparison result is within the allowable error range, and there are no conflicts between the food types in the dietary intake information and the dietary restriction list, then the condition compliance verification is deemed successful.

[0105] For example, extracting Mr. Li's current blood glucose level from his real-time monitoring log data, assuming it's 6.5 mmol / L, we determine if it falls within the blood glucose threshold range of 3.9-7.2 mmol / L. Clearly, 6.5 mmol / L is within this range. Based on the dosage data in Mr. Li's insulin injection record and a pre-defined sensitivity coefficient calculation model, we derive Mr. Li's real-time insulin sensitivity coefficient. Assuming Mr. Li injected 8 units of insulin, the calculation model yields a real-time insulin sensitivity coefficient of 1.0. This is compared with the insulin sensitivity coefficient (0.8-1.2) in the service application conditions, and the difference is found to be within the allowable error range. We then compare the food types in Mr. Li's dietary intake information with the dietary restrictions list. Mr. Li's recent food intake consisted of brown rice, vegetables, and chicken, and no conflict was found with the dietary restrictions list (high-sugar, high-fat foods). Since the current blood glucose level is within the blood glucose threshold range, the difference comparison result is within the allowable error range, and there are no conflicts between the food types in the dietary intake information and the dietary restrictions list, this nutritional service item is deemed to have passed the condition compliance verification.

[0106] If the matching score of a nutrition service item exceeds a preset threshold and passes the condition compliance verification, it is added to the initial candidate set. This process is repeated for all nutrition service items in the digital nutrition database.

[0107] Finally, the nutritional service items in the initial candidate set are sorted in descending order based on their matching scores. For example, if there are three nutritional service items in the initial candidate set with matching scores of 0.8, 0.7, and 0.6, they are sorted from highest to lowest score. For items with the same matching score (assuming there are two items with a score of 0.7), a secondary sort is performed based on the urgency of the service type identifier. If one service type is identified as "emergency nutritional intervention" and the other as "routine nutritional maintenance," then "emergency nutritional intervention" is ranked first, thus generating a priority-ranked set of candidate nutritional services.

[0108] In one possible implementation, step S140 includes: step S141, extracting the service provider location information corresponding to each nutrition service entry from the candidate nutrition service set, wherein the service provider location information includes latitude and longitude coordinates and service coverage radius.

[0109] Step S142: Calculate the straight-line distance based on the real-time geographical location information of the diabetic patient and the latitude and longitude coordinates, and determine whether the straight-line distance is less than or equal to the service coverage radius.

[0110] Step S143: If the straight-line distance is less than or equal to the service coverage radius, then mark the corresponding nutrition service item as an accessible service.

[0111] For example, one of the nutrition services is provided by a diabetes nutrition center located in a city. The service provider's location information includes latitude and longitude coordinates (X1, Y1), and its service coverage radius is 15 kilometers. Mr. Li's diabetes management device can obtain his real-time geographic location information. Assuming Mr. Li is currently at home, with latitude and longitude coordinates of (X2, Y2), the straight-line distance is calculated using a specific distance calculation formula (such as a planar distance formula based on the Pythagorean theorem or a more precise geodetic distance formula). The calculated straight-line distance between Mr. Li's location and the nutrition service provider is 10 kilometers.

[0112] Since a service coverage radius of 10 kilometers is less than 15 kilometers, this nutrition service item is marked as an accessible service. This operation is performed on each nutrition service item in the candidate nutrition service set. For example, if another nutrition service provider has latitude and longitude coordinates of (X3, Y3) and a service coverage radius of 10 kilometers, and the calculated straight-line distance to Mr. Li is 12 kilometers, this service will not be marked as an accessible service because 12 kilometers is greater than 10 kilometers.

[0113] Assuming that three nutritional services are marked as reachable after the distance assessment, they are sorted according to their priority in the candidate nutritional service set (e.g., priority determined by factors such as matching score) to form a subset of target nutritional services. Then, real-time traffic data is added to each reachable service to estimate service arrival time.

[0114] Step S144: Generate the target nutrition service subset according to the priority of the accessible services, and add real-time traffic data to each accessible service to estimate the service arrival time.

[0115] Step S144 includes: Step S1441, calling the map service interface to obtain the path planning result between the real-time geographic location information and the service provider's location information, wherein the path planning result includes the estimated travel time for various modes of transportation.

[0116] Step S1442: Based on the commonly used travel preference data stored in the terminal device of the diabetic patient, select the estimated travel time corresponding to the highest priority mode of transportation as the basic arrival time.

[0117] Step S1443: Dynamically correct the basic arrival time based on real-time traffic flow data to generate a corrected service arrival time, and bind the corrected service arrival time with the corresponding accessible service.

[0118] Taking the first accessible service (nutrition center) mentioned earlier as an example, the route planning results returned by the map service interface show the estimated travel time for various modes of transportation, such as driving (25 minutes), public transportation (35 minutes), and walking (90 minutes). Based on Mr. Li's frequently used travel preference data stored on his device, assuming his device indicates a preference for driving, 25 minutes is chosen as the base arrival time. Next, the base arrival time is dynamically adjusted based on real-time traffic flow data. If the current traffic flow is high, an additional 5 minutes of travel time is calculated based on traffic flow data and pre-set algorithms in existing technology (such as a model of the impact of road congestion on driving speed). The adjusted service arrival time is then 30 minutes, and this adjusted service arrival time is linked to the corresponding accessible service (the nutrition service provided by this nutrition center).

[0119] In one possible implementation, step S150 includes: step S151, generating a push time series based on the priority and estimated arrival time of each target nutrition service in the target nutrition service subset.

[0120] For example, the target nutrition service subset contains three services, ranked from highest to lowest priority: nutrition meal delivery service, diabetes nutrition consultation service, and diabetes exercise guidance service, with estimated arrival times of 30 minutes, 2 hours, and 4 hours, respectively. Based on this information, a preliminary push time sequence is generated, prioritizing the nutrition meal delivery service because it has the highest priority and a relatively short arrival time.

[0121] Step S152: Select a time window that overlaps with the activity period of the terminal device of the diabetic patient as an effective push period in the push time series.

[0122] For example, Mr. Li's activity time period for his terminal device (such as his mobile phone) can be determined by analyzing his daily mobile phone usage patterns. Suppose that the analysis reveals that Mr. Li uses his mobile phone more frequently between 9 am and 6 pm, then a suitable time window can be found within this period to push service messages.

[0123] Step S153: During the effective push period, send notification messages containing service details and arrival time to the terminal device in descending order of priority.

[0124] For example, a notification message for the delivery of a nutrition meal package can be sent first. The message includes information such as the types of food included in the service (e.g., protein-rich chicken, fiber-rich vegetables, etc.), nutritional components (e.g., specific protein, carbohydrate, and vitamin content), and estimated delivery time in 30 minutes.

[0125] Step S154: If no service selection feedback is received within the preset waiting time, the backup push channel is automatically triggered to resend the notification message.

[0126] The triggering conditions for the backup push channel include: detecting whether the network connection status of the terminal device is active.

[0127] If the network connection is active and no feedback is received, the notification message will be resent via the SMS service channel.

[0128] If the network connection is inactive, the push operation will be delayed until the next valid push period, and the number of push failures will be recorded.

[0129] When the number of failed push notifications exceeds a preset limit, a service interruption alert is sent to a preset emergency contact.

[0130] In this embodiment, if no service selection feedback is received from Mr. Li within a preset waiting time (e.g., 10 minutes), a backup push channel is automatically triggered. Then, the network connection status of Mr. Li's phone is checked for activity. If the network connection is active and no feedback is received, the notification message is resent via SMS. Assuming Mr. Li's network connection is normal, but he hasn't checked his messages, the notification message for the nutrition package delivery service will be resent via SMS, with content similar to the message previously sent within the mobile application. If the network connection is inactive, for example, if Mr. Li's phone is in airplane mode or in an area with poor network signal, the push operation will be delayed until the next valid push period (e.g., between 2 PM and 3 PM), and the number of push failures will be recorded. If multiple push failures occur due to network issues, and the number of failures exceeds a preset limit (e.g., 3 times), a service interruption alert is sent to preset emergency contacts (e.g., Mr. Li's family members), informing them that there is a problem with the nutrition package delivery service push notification, which may affect Mr. Li's timely access to the service.

[0131] In one possible implementation, step S150 further includes: step S155, parsing the service identifier and execution time requirement in the service selection feedback.

[0132] After Mr. Li saw the notification message, he selected the nutrition package delivery service on his mobile phone and provided the relevant information. Upon receiving the service selection feedback from the terminal device, the system parsed the service identifier and execution time requirement from the feedback. The service identifier is a code used to uniquely identify the nutrition package delivery service, and the execution time requirement is that Mr. Li hopes for delivery within 25 minutes (assuming he needs to adjust the previously estimated 30-minute delivery time).

[0133] Step S156: Extract the corresponding service provider interface address from the target nutrition service subset according to the service identifier.

[0134] For example, the service provider's interface address is the network address used to communicate with the provider of the nutrition package delivery service (such as a specific catering company or delivery agency). The execution time requirement and the service provider's interface address are encapsulated into a service request message. This message contains Mr. Li's specific requirements for service execution (delivery within 25 minutes) and address information for interacting with the service provider.

[0135] Step S157: Encapsulate the execution time requirement and the service provider's interface address into a service request message, and send it to the service provider's server through the application programming interface.

[0136] Step S158: Receive the service confirmation signal returned by the service provider's server, and generate confirmation information including service details and execution timestamp, and send it to the terminal device.

[0137] In detail, the service request message can be sent to the service provider's server via an application programming interface (API). Upon receiving the request, the service provider's server processes it based on its resources and delivery arrangements. If Mr. Li's request can be met, a service confirmation signal is returned. After receiving the confirmation signal, the system generates a confirmation message containing service details (such as the specific food package contents) and an execution timestamp (such as a confirmed delivery time of 25 minutes later) and sends it to Mr. Li's mobile phone, informing him that his service request has been confirmed and providing specific execution time information.

[0138] Figure 2 The illustration shows exemplary hardware and software components of a digital nutrition service delivery system 100 for diabetic patients, which can implement the ideas of this application, according to some embodiments of this application. For example, a processor 120 can be used in the digital nutrition service delivery system 100 for diabetic patients and to perform the functions in this application.

[0139] The digital nutrition service delivery system 100 for diabetic patients can be a general-purpose server or a special-purpose server; both can be used to implement the digital nutrition service delivery method for diabetic patients described in this application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the load.

[0140] For example, a digital nutrition service delivery system 100 for diabetic patients may include a network port 210 connected to a network, one or more processors 220 for executing program instructions, a communication bus 230, and various forms of storage media 240, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the digital nutrition service delivery system 100 for diabetic patients may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The methods of this application can be implemented according to these program instructions. The digital nutrition service delivery system 100 for diabetic patients also includes an input / output (I / O) interface 250 between the computer and other input / output devices.

[0141] For ease of explanation, only one processor is described in the digital nutrition service delivery system 100 for diabetic patients. However, it should be noted that the digital nutrition service delivery system 100 for diabetic patients in this application may also include multiple processors, and therefore the steps performed by one processor described in this application may also be performed jointly or individually by multiple processors. For example, if the processor of the digital nutrition service delivery system 100 for diabetic patients performs steps A and B, it should be understood that steps A and B may also be performed jointly by two different processors or individually by one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0142] Furthermore, this embodiment of the invention also provides a readable storage medium, wherein computer-executable instructions are preset in the readable storage medium, and when the processor executes the computer-executable instructions, the above-mentioned method for pushing digital nutrition services based on diabetic patients is implemented.

[0143] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for delivering digital nutrition services to diabetic patients, characterized in that, The method includes: Acquire real-time monitoring log data of diabetic patients, including blood glucose level data, insulin injection records, and dietary intake information; The real-time monitoring log data is input into a pre-trained nutrition analysis model for multi-dimensional processing to generate dynamic nutrition demand features. These dynamic nutrition demand features are used to characterize the changing trend of the nutritional element requirements of the diabetic patient within a preset time period. The dynamic nutritional needs characteristics are matched with nutritional service items in a pre-stored digital nutritional database to determine a priority-ranked set of candidate nutritional services. Spatial correlation analysis is performed based on the real-time geographic location information of the diabetic patients and the location information of service providers in the candidate nutrition service set to filter out an accessible subset of target nutrition services. The target nutrition service subset is sent to the terminal device corresponding to the diabetic patient according to a preset push strategy, and the service selection feedback returned by the terminal device is received to trigger the service execution instruction.

2. The method for delivering digital nutrition services to diabetic patients according to claim 1, characterized in that, The step of inputting the real-time monitoring log data into a pre-trained nutrition analysis model for multi-dimensional processing to generate dynamic nutrition requirement features includes: The blood glucose level data is segmented into time series to obtain blood glucose fluctuation segments in multiple consecutive monitoring periods; Extract the injection time point and dose data from the insulin injection record, construct an insulin action time window, and align the insulin action time window with the blood glucose fluctuation segment in the time domain. The food types and nutrient ratios in the dietary intake information are analyzed to generate dietary influencing factors corresponding to the blood glucose fluctuation segments aligned with the time domain. The blood glucose fluctuation segments are expanded in the time dimension to extract the rate of blood glucose change in each monitoring period and the trend continuity between adjacent periods, generating time correlation coding results; Within the insulin action time window, drug concentration gradient intervals are divided. Based on the injection dose data and the preset drug efficacy decay curve, the drug efficacy coverage intensity of each gradient interval is calculated, and a drug efficacy coverage mapping result is generated. The dietary influencing factors are decomposed into nutrient category dimensions, and the difference between the current intake ratio and the recommended ratio of the standard diabetic diet model is compared to generate a quantitative result of nutritional deviation. The time correlation encoding result is input into the time attention channel, and the blood glucose fluctuation dependence across time periods is captured by a sliding time window to generate a time attention weight matrix; The mapping result of the drug efficacy coverage is input into the drug efficacy attention channel. The importance coefficients of different gradient intervals are adjusted based on the temporal overlap between the drug efficacy coverage intensity and the blood glucose fluctuation segment to generate a drug efficacy attention weight matrix. The nutritional deviation quantification results are input into the nutritional attention channel, and the nutritional compensation priority is determined according to the deviation direction and magnitude of the nutrient category dimension, generating a nutritional attention weight matrix. Cross-channel interactive verification is performed on the time attention weight matrix, drug efficacy attention weight matrix and nutrition attention weight matrix. The weight values ​​of the corresponding time periods in each matrix are fused by weighted summation to generate multi-channel joint attention coefficients. Based on the multi-channel joint attention coefficient, the blood glucose fluctuation segment, insulin action time window and dietary influencing factors are dynamically weighted and fused to generate a cross-modal feature vector; The cross-modal feature vectors are subjected to hierarchical compression and dimension alignment to extract core feature groups that reflect blood glucose regulation needs and nutritional imbalance compensation needs. The core feature set is input into a fully connected network for nonlinear transformation to generate dynamic nutritional demand features that include the intensity and time-series distribution of the demand for each nutrient element within a preset future time period.

3. The method for delivering digital nutrition services to diabetic patients according to claim 2, characterized in that, The pre-trained nutrient analysis model is obtained through the following steps: Collect a historical monitoring log dataset of diabetic patients, which includes historical blood glucose data, historical insulin use records, and historical dietary records; Outlier detection and interpolation are performed on the historical blood glucose data to generate a standardized blood glucose sequence, and the timestamps and doses in the historical insulin usage records are encoded to generate an insulin action intensity matrix. Extract the food component vectors for each meal from the historical dietary records, and map the food component vectors into a nutrient intake distribution matrix according to a preset nutrient conversion rule; An initial neural network model is constructed, using the standardized blood glucose sequence, insulin action intensity matrix, and nutrient intake distribution matrix as input features. Temporal modeling of the input features is performed through temporal convolutional layers and gated recurrent units to generate intermediate feature vectors. An adaptive weight allocation module is introduced into the initial neural network model to dynamically adjust the weights of each feature channel according to the dimensional correlation of the intermediate feature vector, and a contrastive learning strategy is used to optimize the similarity between the intermediate feature vector and the preset nutritional requirement label until the initial neural network model converges.

4. The method for delivering digital nutrition services to diabetic patients according to claim 1, characterized in that, The step of matching the dynamic nutritional needs characteristics with nutritional service items in a pre-stored digital nutritional database to determine a priority-ranked set of candidate nutritional services includes: Multiple nutrition service entries are read from the digital nutrition database. Each nutrition service entry includes a service type identifier, a list of nutritional elements, and applicable conditions for the service. For each nutritional service item, calculate the matching score between the nutritional element composition list and the trend of nutritional element demand changes in the dynamic nutritional demand characteristics, and verify the compliance of the service with the real-time monitoring log data of the diabetic patient. If the matching score exceeds the preset threshold and passes the condition compliance verification, the corresponding nutrition service item is added to the initial candidate set. The nutrition service items in the initial candidate set are sorted in descending order according to the matching score, and the items with the same score are sorted again according to the urgency of the service type identifier to generate the priority-sorted candidate nutrition service set.

5. The method for delivering digital nutrition services to diabetic patients according to claim 4, characterized in that, For each nutritional service item, the matching score between the list of nutritional element composition and the trend of nutritional element demand changes in the dynamic nutritional demand characteristics is calculated, including: The demand intensity curves of each nutrient element within a preset time period are extracted from the dynamic nutritional demand characteristics. The demand intensity curves include demand intensity values ​​and corresponding timestamp sequences under multiple time windows. The supply distribution pattern of each nutrient element in the nutrient element composition list is analyzed to generate a nutrient service supply time sequence diagram that includes the supply quantity value and the effective time range. Align the timestamp sequence of the demand intensity curve with the effective time range of the nutrition service supply time series diagram to determine the set of overlapping time windows and the corresponding demand intensity value and supply value. Within the set of overlapping time windows, perform the following operations for each nutrient element: Calculate the absolute difference between the demand intensity and supply of the nutrient element in the corresponding time window, and generate a single-window difference metric. Detect whether the supply quantity value covers the preset fluctuation tolerance range of the demand intensity value; if it does, generate a single-window coverage compensation coefficient. The single-window difference metric is weighted and summed with the single-window coverage compensation coefficient to generate a single-window matching degree index. The single-window matching degree index of the same nutrient element across all overlapping time windows is integrated and accumulated to generate a global matching degree score for that nutrient element. The global matching scores of all nutrients are weighted and merged according to a preset weight ratio to generate an initial matching score. Identify whether there are redundant nutrient elements in the nutrient element composition list that are not included in the dynamic nutrient requirement characteristics. If so, adjust the initial matching score by deducting from the number of redundant nutrient elements and a preset penalty coefficient. The supply value of each nutrient element in the nutrition service supply time sequence diagram is checked to see if it meets the future demand prediction value of the demand intensity curve. If it does, the matching score after deduction is compensated by gain according to the predicted coverage time and the preset reward coefficient. The gain-compensated matching score is limited to a preset standardized range to generate the final matching score.

6. The method for delivering digital nutrition services to diabetic patients according to claim 4, characterized in that, The service application conditions include blood glucose threshold range, insulin sensitivity coefficient, and dietary restrictions list. The compliance verification of these conditions includes the following steps: Extract the current blood glucose value from the real-time monitoring log data of the diabetic patient, and determine whether the current blood glucose value is within the blood glucose threshold range; Based on the dose data in the insulin injection record and the preset sensitivity coefficient calculation model, the real-time insulin sensitivity coefficient of the diabetic patient is derived, and the difference is compared with the insulin sensitivity coefficient in the service application conditions to obtain the difference comparison result. The types of food in the dietary intake information are compared with the dietary taboo list to detect any conflicts. If the current blood glucose value is within the blood glucose threshold range, the difference comparison result is within the allowable error range, and there are no conflicting items between the food types in the dietary intake information and the dietary taboo list, then the condition compliance verification is deemed to have passed.

7. The method for delivering digital nutrition services to diabetic patients according to claim 1, characterized in that, The spatial correlation analysis based on the real-time geographic location information of the diabetic patient and the location information of service providers in the candidate nutrition service set is used to filter out an accessible subset of target nutrition services, including: Extract the location information of the service provider corresponding to each nutrition service item from the candidate nutrition service set. The location information of the service provider includes latitude and longitude coordinates and service coverage radius. The straight-line distance is calculated based on the real-time geographic location information of the diabetic patient and the latitude and longitude coordinates, and it is determined whether the straight-line distance is less than or equal to the service coverage radius; If the straight-line distance is less than or equal to the service coverage radius, the corresponding nutrition service item will be marked as an accessible service. The target nutrition service subset is generated based on the priority of the accessible services, and real-time traffic data is added to each accessible service to estimate the service arrival time. The step of adding real-time traffic data to each accessible service to estimate service arrival time includes: Call the map service interface to obtain the path planning results between the real-time geographic location information and the service provider's location information. The path planning results include the estimated travel time for various modes of transportation. Based on the frequently used travel preference data stored in the terminal device of the diabetic patient, the estimated travel time corresponding to the highest priority mode of transportation is selected as the basic arrival time; The base arrival time is dynamically corrected based on real-time traffic flow data to generate a corrected service arrival time, and the corrected service arrival time is then bound to the corresponding accessible service.

8. The method for delivering digital nutrition services to diabetic patients according to claim 1, characterized in that, Sending the target nutrition service subset to the terminal device corresponding to the diabetic patient according to a preset push strategy includes: A push time series is generated based on the priority and estimated arrival time of each target nutrition service in the target nutrition service subset. Select the time window that overlaps with the activity period of the terminal device of the diabetic patient in the push time series as the effective push period; During the effective push period, notification messages containing service details and arrival times are sent to the terminal device in descending order of priority; If no service selection feedback is received within the preset waiting time, the backup push channel will be automatically triggered to resend the notification message; The triggering conditions for the backup push channel include: Detect whether the network connection status of the terminal device is active; If the network connection is active and no feedback is received, the notification message is resent via the SMS service channel. If the network connection status is inactive, the push operation will be delayed until the next valid push period, and the number of push failures will be recorded. When the number of failed push notifications exceeds a preset limit, a service interruption alert is sent to a preset emergency contact.

9. The method for delivering digital nutrition services to diabetic patients according to claim 1, characterized in that, The step of receiving service selection feedback returned by the terminal device to trigger a service execution instruction includes: Parse the service identifier and execution time requirement in the service selection feedback; Extract the corresponding service provider interface address from the target nutrition service subset based on the service identifier; The execution time requirement and the service provider's interface address are encapsulated into a service request message and sent to the service provider's server through the application programming interface; The system receives a service confirmation signal returned by the service provider's server and generates confirmation information containing service details and an execution timestamp, which is then sent to the terminal device.

10. A digital nutrition service delivery system for diabetic patients, characterized in that, The digital nutrition service delivery system for diabetic patients includes a processor and a memory, the memory and the processor being connected. The memory is used to store programs, instructions or code, and the processor is used to execute the programs, instructions or code in the memory to implement the digital nutrition service delivery method for diabetic patients as described in any one of claims 1-9.

Images