Insulin dosage estimation system

Abstract

The invention discloses an insulin dosage estimation system which comprises a blood glucose detector, an insulin pump / pen, diet data acquisition equipment and wearable equipment, the data acquisition module is respectively used for acquiring real-time continuous blood glucose concentration data, insulin historical use data, nutrition intake data, body index data, environmental factor data, psychological state data, special event data, physical activity data and physiological index data of a user; the data storage module is used for storing data, and the processing module is used for mapping the data to a plurality of time grids to obtain blood glucose features, physiological status features, behavior pattern features and environment features; and fusing to obtain a multi-modal fusion feature and predicting an insulin estimated dose. According to the method, the multi-modal fusion features are obtained based on the data, then the insulin dosage is predicted, the insulin dosage considers the multi-modal data influencing the blood glucose concentration of the user, and the accuracy is high from crowd classification, individual difference and environmental influence dimension correction.

Description

Technical Field

[0001] This application relates to the field of medical and health technology, and in particular to an insulin dosage prediction system. Background Technology

[0002] For some groups, insulin is often needed to control blood sugar. Currently, most of these groups can only judge the insulin dosage based on their own experience and cannot accurately estimate the dosage of insulin to be used. Summary of the Invention

[0003] The main objective of this application is to propose an insulin dosage prediction system that aims to solve the problem in the prior art where users cannot accurately predict insulin dosage based on their own experience.

[0004] To achieve the above objectives, this application proposes an insulin dosage prediction system, comprising: A blood glucose meter is used to acquire real-time, continuous blood glucose concentration data of a target user. Insulin pumps / pens are used to record the target user's historical insulin usage data. Dietary data collection equipment is used to collect nutritional intake data of target users, as well as to allow target users to record physical indicator data, environmental factor data, psychological state data, and special event data; Wearable devices are used to collect physical activity data and physiological index data of target users; The data storage module is connected to the blood glucose meter, the insulin pump / pen, the diet data acquisition device, and the wearable device, respectively. The data storage module is used to receive data collected by the blood glucose meter, the insulin pump / pen, the diet data acquisition device, and the wearable device. A processing module, connected to the data storage module, is used to map the data acquired by the data storage module to multiple preset time grids, wherein the multiple time grids have different time granularities; Based on each time grid, the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics are obtained. The user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics are fused to obtain multimodal fusion features; The target insulin dose is determined based on the multimodal fusion features and a preset prediction model. The preset model comprises a base layer, a population classification layer, an individual adaptation layer, and a scenario specialization layer connected in sequence. The base layer determines the user's first estimated insulin dose based on the multimodal fusion features. The population classification layer predicts the user's subgroup classification based on the multimodal fusion features and corrects the first estimated insulin dose based on the subgroup classification to obtain a second estimated insulin dose. The individual adaptation layer determines the user's physiological response pattern based on the multimodal fusion features and corrects the second estimated insulin dose based on the user's physiological response pattern to obtain a third estimated insulin dose. The scenario specialization layer determines the specific scenario in which the user is located based on the multimodal fusion features and corrects the third estimated insulin dose based on the specific scenario to obtain a fourth insulin dose, which is then used as the target insulin dose.

[0005] In this embodiment of the application, the preset multiple time grids include: a low-resolution time grid, a medium-resolution time grid, and a high-resolution time grid, wherein the time granularity of the low-resolution time grid is greater than that of the medium-resolution time grid, and the time granularity of the medium-resolution time grid is greater than that of the high-resolution time grid. The processing module is configured to map the acquired data to multiple preset time grids based on the following method: The acquired data is mapped to the high-resolution time grid, the medium-resolution time grid, and the low-resolution time grid, respectively.

[0006] In this embodiment of the application, the nutritional intake data includes: intake data of protein, carbohydrates and fat; The physical activity data includes: activity type, activity duration, and activity heart rate changes; The physiological indicators include: weight, body temperature, blood pressure, and sleep quality; The physical indicators include: glycated hemoglobin, liver and kidney function, and insulin sensitivity data; The environmental factor data includes: season, ambient temperature, and ambient humidity; The psychological state data includes: stress and emotional state data; The special event data includes: disease data and menstrual data.

[0007] In this embodiment of the application, the blood glucose characteristics include the following sub-characteristics: the mean, standard deviation, rate of change, fluctuation frequency, peak and trough, circadian rhythm, and hypoglycemia risk index of the user's blood glucose; The user's physiological state characteristics include the following sub-characteristics: active insulin level, dynamic characteristics of insulin sensitivity, insulin absorption rate, energy balance assessment data, carbohydrate utilization efficiency, metabolic state markers, inflammatory state, and hormone fluctuation patterns. The user's behavioral pattern characteristics include the following sub-features: meal regularity, carbohydrate preference, eating speed, exercise habits, daily activity intensity distribution, sedentary behavior identification, sleep regularity, sleep quality, and sleep-wake transition characteristics; The environmental characteristics include the following sub-characteristics: time of day, weekday / weekend characteristics, seasonal characteristics, holiday characteristics, social activities, travel status, temperature, humidity, altitude, emotional state, and stress level. The processing module is configured to acquire the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics based on each time grid in the following manner: The user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics are determined based on the various sub-features included in the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics, respectively.

[0008] In this embodiment of the application, the processing module is configured to determine the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics based on the following: The blood glucose feature is determined based on the sub-features included in the blood glucose feature and their corresponding weights; The physiological state features are determined based on the sub-features and their corresponding weights contained in the physiological state features. The behavioral pattern features are determined based on the sub-features and their corresponding weights contained in the behavioral pattern features. The environmental feature is determined based on the various sub-features included in the environmental feature and their corresponding weights.

[0009] In this embodiment of the application, the processing module is further configured to determine the weight of each sub-feature included in the user's blood glucose feature, physiological state feature, and behavioral pattern feature based on the importance of each sub-feature in its respective feature category.

[0010] In this embodiment of the application, the processing module is further configured to fuse the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics in the following manner to obtain the user's multimodal fusion characteristics: The reliability of the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics was determined respectively; Based on the reliability of the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics, the weights of the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics are determined. Based on the weights of the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics, the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics are weighted and fused to obtain a multimodal fusion feature that includes the user's real-time status and contextual information.

[0011] In this embodiment of the application, the processing module is further configured to: The reliability of the blood glucose feature is determined based on the signal quality, data freshness, and data missing rate of the real-time continuous blood glucose concentration data. The reliability of the physiological state characteristics is determined based on the accuracy and timeliness of the historical insulin usage data. The reliability of the behavioral pattern characteristics is determined based on the timeliness and completeness of nutritional intake data. The reliability of the environmental characteristics is determined based on the accuracy, timeliness, and completeness of the environmental factor data.

[0012] In this embodiment of the application, the processing module is further configured to determine the weights of the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics based on the following formula:

[0013] Where BG represents blood glucose characteristics, PS represents physiological state characteristics, BH represents behavioral pattern characteristics, and ENV represents environmental characteristics, k∈[BG, PS, BH, ENV] This represents the sum of the reliability of each feature. Represents weight, This represents the reliability of the value of k. .

[0014] In this embodiment of the application, the processing module is further configured to determine the weights of the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics based on the following formula:

[0015] Where k∈[BG, PS, BH, ENV], BG represents blood glucose characteristics, PS represents physiological state characteristics, BH represents behavioral pattern characteristics, and ENV represents environmental characteristics. Represents a priori importance, It is a regulatory factor. Represents weight.

[0016] In this embodiment of the application, the processing module is further configured to determine multimodal fusion features based on the following method:

[0017] in, Represents multimodal fusion features, Represents the weight of blood glucose characteristics. Represents blood glucose characteristics, Represents the weight of physiological state characteristics. Represents physiological characteristics. Represents the weight of behavioral state features. Represents behavioral state characteristics. Represents the weight of environmental characteristics. It represents environmental characteristics.

[0018] The insulin dosage prediction system in this application embodiment acquires the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics based on various data that can affect the user's blood glucose concentration, and fuses them to obtain multimodal fusion features. Then, it predicts the target insulin dose based on a preset prediction model. The target insulin dose not only takes into account the multimodal data that can affect the user's blood glucose concentration, but also corrects from the dimensions of population classification, individual differences, and environmental influences, resulting in high accuracy. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0020] Figure 1 This is a block diagram of an insulin dosage prediction system according to an embodiment of this application; The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] like Figure 1 As shown in the embodiment of this application, an insulin dosage prediction system 100 is proposed, comprising: Blood glucose meter 110 is used to acquire real-time continuous blood glucose concentration data of the target user; Insulin pump / pen 120, used to record the target user's historical insulin usage data; Dietary data acquisition device 130 is used to collect nutritional intake data of target users, as well as to allow target users to record physical indicator data, environmental factor data, psychological state data, and special event data. Wearable device 140 is used to collect physical activity data and physiological index data of the target user; The data storage module 150 is connected to the blood glucose meter 110, the insulin pump / pen 120, the diet data acquisition device 130, and the wearable device 140, respectively. The data storage module 150 is used to receive the data collected by the blood glucose meter 110, the insulin pump / pen 120, the diet data acquisition device 130, and the wearable device 140. The processing module 160 is connected to the data storage module 150 and is used to map the data acquired by the data storage module 150 to multiple preset time grids, wherein the multiple time grids have different time granularities. Based on each time grid, the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics are obtained. The user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics are fused to obtain multimodal fusion features; The target insulin dose is determined based on the multimodal fusion features and a preset prediction model. The preset model comprises a base layer, a population classification layer, an individual adaptation layer, and a scenario specialization layer connected in sequence. The base layer determines the user's first estimated insulin dose based on the multimodal fusion features. The population classification layer predicts the user's subgroup classification based on the multimodal fusion features and corrects the first estimated insulin dose based on the subgroup classification to obtain a second estimated insulin dose. The individual adaptation layer determines the user's physiological response pattern based on the multimodal fusion features and corrects the second estimated insulin dose based on the user's physiological response pattern to obtain a third estimated insulin dose. The scenario specialization layer determines the specific scenario in which the user is located based on the multimodal fusion features and corrects the third estimated insulin dose based on the specific scenario to obtain a fourth insulin dose, which is then used as the target insulin dose.

[0022] In this embodiment of the application, the user's real-time continuous blood glucose concentration data can be continuously collected in real time based on a continuous glucose monitor.

[0023] In this embodiment, the historical insulin usage data includes usage time, usage method, and usage dose, which can be recorded based on the smart insulin pump / pen 120.

[0024] In this embodiment, nutritional intake data may include intake data for protein, carbohydrates, and fat. Specifically, it may include the intake time and amount of each of the three components, which can be obtained through the dietary data acquisition device 130. In this embodiment, the dietary data acquisition device 130 can be an app installed on a mobile phone, tablet, or other terminal. This app has a built-in food database. Users can determine the amount of protein, carbohydrates, and fat contained in the food by taking a picture of the food and comparing it with the built-in food database. Alternatively, users can actively record the data through manual or voice input methods.

[0025] In this embodiment, the physical indicators include: glycated hemoglobin, liver and kidney function, and insulin sensitivity data. These data can be obtained through laboratory tests. Liver and kidney function includes liver function and kidney function. Liver function includes: alanine aminotransferase (ALT) (U / L), aspartate aminotransferase (AST) (U / L), total bilirubin (μmol / L), albumin (g / L), and alkaline phosphatase (U / L). Kidney function includes serum creatinine (μmol / L), glomerular filtration rate (mL / min / 1.73m²), and blood urea nitrogen (mmol / L). Insulin sensitivity data includes fasting insulin level (μIU / mL), HOMA-IR (fasting insulin (μIU / mL) × fasting blood glucose (mmol / L) / 22.5), Matsuda index (derived from oral glucose tolerance test, no unit), and high insulin positive glucose clamp test (gold standard, mg / kg / min). Once the target user's body indicator data is acquired, it can be recorded and stored in the corresponding app based on the diet data collection device 130.

[0026] In this embodiment, the environmental factor data includes: season, ambient temperature, and ambient humidity. The season can be determined based on the climate and month of the user's location, while the ambient temperature and humidity can be obtained through monitoring with a thermometer and a hygrometer. Once the environmental factor data is determined, it can also be stored based on records from the app corresponding to the dietary data collection device 130.

[0027] In this embodiment, the psychological state data includes stress and emotional state data. The stress and emotional state data can be determined through a questionnaire. Once the psychological state data is determined, it can be recorded and stored using the app corresponding to the dietary data collection device 130.

[0028] In this embodiment, the special event data includes: disease data and menstrual data. The disease data and menstrual data can be tagged by the user. For example, the type and time of the disease, as well as the start and end times of menstruation, can be tagged. Once the special time data is determined, it can be stored based on the app records corresponding to the dietary data collection device 130.

[0029] In this embodiment, physical activity data includes: activity type, activity duration, and activity heart rate changes; physiological indicator data includes: weight, body temperature, blood pressure, and sleep quality. Wearable devices can be smartwatches, wristbands, or other similar devices used to acquire the target user's physical activity data, as well as body temperature, sleep quality, and blood pressure data; additionally, a smart scale can be used to acquire the target user's weight.

[0030] In the embodiments of this application, the data storage module 150 can be, for example, a memory with read and write functions. The data storage module 150 can be wirelessly connected to the blood glucose meter 110, insulin pump / pen 120, diet data acquisition device 130, and wearable device 140. Whenever the blood glucose meter 110, insulin pump / pen 120, diet data acquisition device 130, and wearable device 140 have new data, it can be synchronously sent to the data storage module 150 and stored in the data storage module 150.

[0031] In this embodiment, the processing module 160 is electrically connected to the data storage module 150. The processing module 160 can read various data of the target user stored in the data storage module 150 and predict the insulin dosage based on the target user's data. The connection mentioned in this application can be a wired connection or a wireless communication connection.

[0032] In this embodiment of the application, the preset multiple time grids include: a low-resolution time grid, a medium-resolution time grid, and a high-resolution time grid, wherein the time granularity of the low-resolution time grid is greater than that of the medium-resolution time grid, and the time granularity of the medium-resolution time grid is greater than that of the high-resolution time grid.

[0033] In this embodiment of the application, the monitoring frequency of real-time continuous blood glucose concentration data can be once every 1-5 minutes; Body temperature and blood pressure can be monitored every 30 minutes or every 60 minutes; weight can be monitored every half day, every day, or every few days; sleep quality can be monitored once a day; stress and mood data can be monitored once a day or every few days; menstrual data can be monitored once a month; and season, ambient temperature, and ambient humidity can be monitored once a day.

[0034] In the embodiments of this application, the monitoring frequency of glycated hemoglobin, liver and kidney function, and insulin sensitivity data can be once every few weeks, once a month, or once every few months.

[0035] In the embodiments of this application, disease data, historical insulin dosing data, protein intake data, carbohydrate intake data, fat intake data, activity type, activity duration, and activity heart rate changes are all data that are generated only at specific times, and therefore are all discrete data that can be recorded in real time.

[0036] In this embodiment of the application, the processing module 160 can set the time granularity of each time grid in the low-resolution time grid to be greater than one hour, the time granularity of each time grid in the medium-resolution time grid to be 15-30 minutes, and the time granularity of each time grid in the high-resolution time grid to be 1-5 minutes, based on the monitoring frequency of each data.

[0037] In this embodiment of the application, after determining the time granularity of the low-resolution time grid, the medium-resolution time grid, and the high-resolution time grid, the processing module 160 can construct the corresponding time grid and map the acquired data in the three time grids.

[0038] In this embodiment of the application, the blood glucose characteristics include the following sub-characteristics: the mean, standard deviation, rate of change, fluctuation frequency, peak and trough, circadian rhythm, and hypoglycemia risk index of the user's blood glucose.

[0039] In this embodiment of the application, the physiological state characteristics of the user include the following sub-characteristics: active insulin level, dynamic characteristics of insulin sensitivity, insulin absorption rate, energy balance assessment data, carbohydrate utilization efficiency, metabolic state markers, inflammatory state, and hormone fluctuation patterns.

[0040] The user's behavioral pattern characteristics include the following sub-features: meal regularity, carbohydrate preference, eating speed, exercise habits, daily activity intensity distribution, sedentary behavior identification, sleep regularity, sleep quality, and sleep-wake transition characteristics; The environmental characteristics include the following sub-characteristics: time of day, weekday / weekend characteristics, seasonal characteristics, holiday characteristics, social activities, travel status, temperature, humidity, altitude, emotional state, and stress level.

[0041] In this embodiment of the application, the processing module 160 extracts each sub-feature of the blood glucose feature in the following way: Mean and standard deviation: These statistics can be calculated based on real-time continuous blood glucose concentration data in a high-resolution time grid within a specific time window (such as 1 hour, 4 hours, or 24 hours).

[0042] Rate of change: It can be calculated by time difference based on real-time continuous blood glucose concentration data in a high-resolution time grid, such as the ratio of the difference in blood glucose concentration between two time points to the time interval between the two time points.

[0043] Fluctuation frequency: Based on the spectrum analysis model, such as the Fast Fourier Transform (FFT) or wavelet transform method, the high-resolution time grid is input into the spectrum analysis model, the time series of real-time continuous blood glucose concentration data in the high-resolution time grid is converted to the frequency domain, the intensity of different frequency components is analyzed, and the fluctuation frequency is obtained. The spectrum analysis model can be integrated into the processing module 160.

[0044] Peak and valley: Peak detection can be calculated based on real-time continuous blood glucose concentration data in a high-resolution time grid using peak detection models, such as sliding window extreme value detection or derivative-based methods. Peak detection models can be integrated into processing module 160.

[0045] In addition, the peak detection model can be predefined, such as predefining the criteria for peaks / valleys, such as the point where the local maximum / minimum value differs from the neighboring value by more than a threshold as a peak / valley. The peak detection model can be integrated into the processing module 160.

[0046] Circadian rhythm: It can be determined based on real-time continuous blood glucose concentration data in medium- or low-resolution time grids using time series decomposition models, such as seasonal trend decomposition (STL) or cosine fitting models. The time series decomposition model can be integrated into the processing module 160.

[0047] Additionally, the time series decomposition model can be pre-trained using real-time continuous blood glucose concentration data from multiple days of target users to extract cyclical patterns over a 24-hour period. For different target users, the time series decomposition model needs to be trained based on real-time continuous blood glucose concentration data from each user to learn their unique diurnal metabolic patterns.

[0048] Low Blood Glucose Risk Index: In this embodiment of the application, the Low Blood Glucose Risk Index (LBGI) can be obtained by analyzing real-time continuous blood glucose concentration data in a high-resolution time grid using blood glucose analysis software (such as EasyGV or GlyCulator), which can be integrated into the processing module 160.

[0049] In this embodiment of the application, the processing module 160 extracts the various sub-features included in the physiological state features in the following manner: Active insulin levels: These can be determined using a pharmacokinetic model (such as the Pk model) based on a low- or medium-resolution time grid. The pharmacokinetic model can be integrated into the processing module 160. Based on historical insulin usage data, the pharmacokinetic model can estimate the amount of active insulin in the target user's body after insulin administration.

[0050] In addition, the pharmacokinetic model can be trained in advance. During training, the parameters of the pharmacokinetic model, such as insulin absorption rate and half-life, can be set based on the average value of the population. Then, based on the Bayesian method, the actual blood glucose response data (glucose increase data and blood glucose decrease data) of the user can be adjusted in a personalized way.

[0051] Dynamic characteristics of insulin sensitivity: Insulin sensitivity can be extrapolated using an insulin sensitivity model based on a low-resolution or medium-resolution time grid. This model can be integrated into the processing module 160. The insulin sensitivity model can employ an autoregressive model, a Kalman filter, or a neural network. During extrapolation, the insulin sensitivity model uses insulin administration data, nutritional intake data, and blood glucose concentration change data to obtain the dynamic characteristics of the target user's insulin sensitivity.

[0052] In addition, the insulin sensitivity model can be pre-trained. During training, based on the target user's historical insulin administration data, nutritional intake data, and blood glucose concentration change data, the model predicts the magnitude of the effect of a unit of insulin on blood glucose. Furthermore, the target user's historical insulin administration data, nutritional intake data, and blood glucose concentration change data during training can include data from different times and activity states, thereby learning the target user's insulin sensitivity under different time periods and activity states.

[0053] Insulin absorption rate: The insulin absorption rate can be determined based on a low-resolution or medium-resolution time grid and a nonlinear mixed effects model (NLME), which can be integrated into the processing module 160.

[0054] The nonlinear mixed-effects model can be pre-trained based on population and individual data to learn the factors affecting insulin absorption rate. Population data may include age, sex, race, BMI, blood glucose type, and insulin usage type (rapid-acting / long-acting); individual data may include subcutaneous fat thickness, local blood flow, insulin application site (abdomen, thigh, buttocks), and post-use exercise data.

[0055] Energy balance assessment data: This can be calculated based on low-resolution or medium-resolution time grids and physical energy balance models, which can be integrated into processing module 160. Energy balance models can include, for example, the Harris-Benedict equation, the Mifflin-St Jeor equation, the WHO / FAO / UNU energy requirement model, and dynamic energy balance models (Differential Equation Models). Using any of these models, based on nutritional intake records, physical activity data, and basal metabolic rate, the target user's energy balance assessment data can be calculated.

[0056] In addition, the parameters of the energy balance model can be calibrated based on the target user's historical weight change data. For example, the energy allocation coefficient in the energy balance model can record the target user's weight change (ΔW) and energy deficit (ΔE) over a period of time (e.g., 4 weeks), and record the percentage change in body fat (which can be determined through a DEXA model or skinfold thickness measurement). ΔW * body fat change percentage / ΔE is the calibrated energy allocation coefficient.

[0057] Carbon utilization efficiency: can be calculated based on a low-resolution or medium-resolution time grid and a machine learning model. The machine learning module can be integrated into the processing module 160, and the machine learning model can adopt random forest or gradient boosting tree algorithms.

[0058] In addition, the machine learning model can be pre-trained. During training, it can be trained based on the target user's carbohydrate intake records and the corresponding blood glucose response curves to minimize the error between the machine learning model's predicted blood glucose response and the actual blood glucose response.

[0059] Metabolic state labels can be calculated based on low-resolution or medium-resolution time grids and latent variable models (LVMs), which can be integrated into processing module 160.

[0060] In addition, the latent variable model can be trained in advance. During training, based on the target user's multi-source physiological index data, such as resting heart rate, heart rate variability (HRV), and body temperature, an unsupervised learning method can be used to identify the target user's potential metabolic state patterns. Alternatively, supervised learning can be performed using multi-source physiological index data with labels (known metabolic status).

[0061] Inflammation status: This can be calculated based on low-resolution or medium-resolution time grids and Bayesian networks or decision tree models, which can be integrated into processing module 160. Specifically, the Bayesian network or decision tree model can predict the inflammation status of the target user based on their resting heart rate, body temperature, sleep quality, self-reported symptoms, etc.

[0062] In addition, Bayesian networks or decision tree models can be pre-trained using labeled inflammatory state data (such as C-reactive protein).

[0063] Hormone fluctuation patterns can be calculated based on low-resolution or medium-resolution time grids and periodic time series models, such as seasonal ARIMA (SARIMA) models and periodic neural network models. Periodic time series models can be integrated into processing module 160.

[0064] In addition, during training, periodic patterns are identified based on user-recorded periodic information (such as a woman's menstrual cycle) and related physiological indicators, and a mapping relationship is established with user-recorded events (such as menstrual cycle phases).

[0065] The processing module 160 extracts the various sub-features included in the behavioral pattern features in the following ways: Meal patterns can be determined using time series clustering algorithms or regularity measurement algorithms based on low-resolution or medium-resolution time grids. These algorithms can be integrated into processing module 160. For example, based on time records in nutritional intake data, the standard deviation or information entropy of meal times can be calculated to assess the degree of regularity.

[0066] Carbohydrate preference value: Statistical analysis can be performed based on low-resolution or medium-resolution time grids, or the distribution pattern of carbohydrate proportion in nutrient intake data can be determined based on classification algorithm analysis. The classification algorithm can be integrated into the processing module 160.

[0067] Feeding speed: Feeding stages and speed can be identified using a time series segmentation model based on a low-resolution or medium-resolution time grid. The time series segmentation model can be integrated into the processing module 160.

[0068] Exercise habits: Based on low-resolution or medium-resolution time grids and activity recognition models, regular activity patterns of target users can be extracted. The activity recognition model can be integrated into the processing module 160.

[0069] Daily activity intensity distribution: This can be determined based on a low-resolution or medium-resolution time grid and an energy expenditure estimation model, which can be integrated into the processing module 160. The energy expenditure estimation model can be pre-trained based on calibrated heart rate and energy expenditure data. During computation, the energy expenditure estimation model maps the target user's physical activity data to different intensity levels, such as light, moderate, and high intensity, and calculates the time percentage for each intensity level.

[0070] Sedentary behavior recognition: The target user's sedentary state can be identified based on a low-resolution or medium-resolution time grid and an activity state classifier, and the duration can be calculated. The activity state classifier can be integrated into the processing module 160.

[0071] Sleep patterns: Based on low-resolution time grids and time series models, the variability of sleep onset and wake-up times of target users can be analyzed. The time series model can be integrated into sub-processing module 160.

[0072] Sleep quality score: This can be determined based on a low-resolution time grid and a multi-parameter fusion model, which can be integrated into the processing module 160. The multi-parameter fusion model is typically a decision tree or a neural network. The multi-parameter fusion model can identify the target user's sleep duration, number of interruptions, proportion of deep sleep, and proportion of rapid eye movement (REM) sleep. Based on these metrics, the target user's sleep quality score is determined.

[0073] In addition, multi-parameter fusion models can be pre-trained based on polysomnography (PSG) as the gold standard, or based on expert ratings or user self-assessments as alternative labels.

[0074] Sleep-wake transition characteristics: can be determined based on a low-resolution temporal grid and a Hidden Markov Model (HMM), which can be integrated into the processing module 160.

[0075] Additionally, labeled sleep stage data can be used to train HMM models to identify features of sleep state transitions, such as transition speed and transition completeness.

[0076] In this embodiment of the application, the processing module 160 extracts the various sub-features included in the environmental features in the following manner: Time of day, weekday / weekend features, seasonal features, and holiday features: can be obtained directly from the low-resolution time grid and the clock and calendar API built into the processing module 160.

[0077] Social activities: Based on low-resolution time grids, location data, and calendar event classification models, possible social scenarios of target users can be inferred. The calendar event classification model can be integrated into the processing module 160.

[0078] Travel status: This can be determined based on a low-resolution time grid and a location data anomaly detection model, which can be integrated into the processing module 160. During training, a regular activity area is established based on the target user's historical location data, and significantly deviating location data is detected. Furthermore, the location data anomaly detection model can be, for example, a Gaussian model, where the location data of the regular activity area follows a Gaussian distribution. The probability density of the current location is calculated, and low-probability points are considered anomalies.

[0079] Temperature, humidity, and altitude can be read directly based on a low-resolution time grid.

[0080] Emotional state: This can be determined based on a low-resolution time grid and either a text-based sentiment analysis model or a speech-based emotion recognition model. The sentiment analysis model or emotion recognition model can be integrated into the processing module 160. Training is based on labeled emotion datasets, such as labeled text and speech samples.

[0081] Stress level characteristics: These can be determined based on a low-resolution time grid and a heart rate variability (HRV) model, which can be integrated into the processing module 160. The HRV model can determine the target user's stress level based on HRV parameters, skin conductance response, and respiratory pattern. During training, labeled stress level data can be used to train the HRV model to recognize physiological signal patterns at different stress levels.

[0082] In physiology, factors affecting a user's blood sugar have drastically different speeds and durations of action over time. For example, the immediate response after insulin injection and the drop in blood sugar during exercise are rapid processes, with a time granularity generally on the order of minutes. Changes in postprandial blood sugar and the entire process of insulin action are medium-duration processes, with a time granularity generally on the order of hours. Changes in insulin sensitivity, the cumulative effect of stress, and the influence of the female menstrual cycle are slow processes, with a time granularity generally on the order of days / weeks.

[0083] In existing technologies, all data is mapped to a single-granularity time grid. For low-frequency data (such as manually recorded meals), excessive interpolation is required, introducing spurious accuracy. For high-frequency data (such as real-time continuous blood glucose concentration), aggregation is required, resulting in the loss of detailed data information. For specific physiological processes, the representation is insufficient. For example, blood glucose control is a complex combination of processes at multiple time scales, which cannot be fully captured by a single time scale. Insulin action has two phases: rapid onset and slow decline, which cannot be accurately represented by a single time granularity. Carbohydrate absorption has different time-dynamic characteristics (rapid, medium, and slow), which cannot be accurately represented by a single time granularity. The effects of exercise have immediate and delayed effects, which cannot be accurately represented by a single time granularity.

[0084] In this embodiment, multiple time grids with different time granularities are constructed. When extracting each sub-feature, it is extracted from the time grid of the corresponding granularity according to the expression characteristics of different sub-features at the time granularity, which can improve the accuracy of each sub-feature.

[0085] After obtaining each sub-feature, in this embodiment of the application, the processing module 160 can also determine the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics based on the following methods: The blood glucose feature is determined based on the sub-features included in the blood glucose feature and their corresponding weights; The physiological state features are determined based on the sub-features and their corresponding weights contained in the physiological state features. The behavioral pattern features are determined based on the sub-features and their corresponding weights contained in the behavioral pattern features. The environmental feature is determined based on the various sub-features included in the environmental feature and their corresponding weights.

[0086] The processing module 160 is configured to determine the weights of each sub-feature included in the blood glucose feature, physiological state feature, behavioral pattern feature, and environmental feature through the following methods 1) and 2): Method 1) In this embodiment of the application, the processing module 160 can assign weights to each sub-feature based on experience.

[0087] For example, blood glucose characteristics include sub-characteristics such as mean, standard deviation, rate of change, fluctuation frequency, peak and trough, circadian rhythm, and hypoglycemia risk index. Weights can be assigned to each sub-characteristic based on experience, and the sum of each sub-characteristic is 1.

[0088] Physiological state characteristics include: active insulin level, dynamic characteristics of insulin sensitivity, insulin absorption rate, energy balance assessment data, carbohydrate utilization efficiency, metabolic state markers, inflammatory state, and hormone fluctuation patterns. Each sub-characteristic can be assigned a weight based on the physiological state, with each sub-characteristic having a weight of 1.

[0089] Behavioral pattern features include: meal regularity, carbohydrate preference, eating speed, exercise habits, daily activity intensity distribution, sedentary behavior recognition, sleep regularity, sleep quality, and sleep-wake transition features. Each sub-feature can be assigned a weight based on experience, with each sub-feature having a weight of 1.

[0090] Environmental features include sub-features: time of day, weekday / weekend features, seasonal features, holiday features, social activities, travel status, temperature, humidity, altitude, emotional state, and stress level features. The impact of each sub-feature on insulin can be judged based on experience, and each sub-feature can be assigned a weight, with each sub-feature having a weight of 1.

[0091] Method 2) In this embodiment of the application, the processing module 160 can determine the weight of each sub-feature based on the importance of each sub-feature in its respective feature category.

[0092] For example, by using feature importance methods such as decision trees or random forests, or by using coefficients based on Lasso regression, the contribution of each sub-feature to insulin demand prediction can be determined, and weights can be assigned to each sub-feature based on its contribution to insulin demand prediction.

[0093] The weights of each sub-feature can be determined using methods 1) and 2) above. Based on the weights of each sub-feature, the features of each type can be determined.

[0094] For example, blood glucose characteristics can be obtained by weighting based on the feature values ​​and weights of each of its sub-features; physiological state characteristics can be obtained by weighting based on the feature values ​​and weights of each of its sub-features; behavioral pattern characteristics can be obtained by weighting based on the feature values ​​and weights of each of its sub-features; and environmental characteristics can be obtained by weighting based on the feature values ​​and weights of each of its sub-features.

[0095] After determining the blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics, the processing module 160 can fuse the blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics to obtain multimodal fusion characteristics.

[0096] In this embodiment of the application, the processing module 160 is configured to determine multimodal fusion features based on the following method: The reliability of the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics was determined respectively; Based on the reliability of the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics, the weights of the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics are determined. Based on the weights of the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics, the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics are weighted and fused to obtain a multimodal fusion feature that includes the user's real-time status and contextual information.

[0097] In this embodiment, the reliability of blood glucose characteristics can be determined based on the signal quality, data freshness, and data missing rate of the real-time continuous blood glucose concentration data.

[0098] In the embodiments of this application, the reliability of physiological state characteristics can be determined based on the accuracy and timeliness of historical insulin usage data.

[0099] In the embodiments of this application, the reliability of behavioral pattern characteristics can be determined based on the timeliness and completeness of nutritional intake data. In the embodiments of this application, the reliability of environmental characteristics can be determined based on the accuracy, timeliness, and completeness of the environmental factor data.

[0100] In this embodiment, the processing module 160 is configured to determine the weights of the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics based on the following formula (1): (1) Here, BG represents blood glucose characteristics, PS represents physiological state characteristics, BH represents behavioral pattern characteristics, and ENV represents environmental characteristics. This represents the sum of the reliability of each feature. The weights representing feature k Represents the reliability of feature k, k∈[BG, PS, BH, ENV].

[0101] In this embodiment, besides determining the weights of the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics based on their reliability, the weights of each characteristic can also be determined by combining prior importance. For example, in some cases, even if the current reliability of a certain characteristic is slightly low, if it is important for insulin dose prediction (i.e., its prior importance is high), it still needs to be assigned a basic weight. The magnitude of the basic weight can be determined based on experience.

[0102] Specifically, in this embodiment of the application, the processing module 160 can determine the weights of the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics based on the prior importance of the following formula (2): (2) Where BG represents blood glucose characteristics, PS represents physiological state characteristics, BH represents behavioral pattern characteristics, and ENV represents environmental characteristics. The prior importance of feature k is represented. It is a regulatory factor. The weights represent the features k, where k∈[BG, PS, BH, ENV].

[0103] After determining the weights of blood glucose features, physiological state features, behavioral pattern features, and environmental features, the various features can be fused to obtain multimodal fusion features.

[0104] In this embodiment of the application, the processing module 160 can obtain the multimodal fusion features based on the following formula (3): (3) in, Represents multimodal fusion features, Represents the weight of blood glucose characteristics. Represents blood glucose characteristics, Represents the weight of physiological state characteristics. Represents physiological characteristics. Represents the weight of behavioral state features. Represents behavioral state characteristics. Represents the weight of environmental characteristics. It represents environmental characteristics.

[0105] In this embodiment, the base layer of the preset prediction model can be constructed based on a temporal deep learning neural network, such as a temporal long short-term memory network (LSTM) or a gated recurrent unit (GRU). The base layer is used to capture the basic patterns of dynamic changes in the user's blood glucose based on the fused features. Additionally, an attention mechanism can be integrated into the temporal deep learning neural network to identify which features among the input fused features have a greater impact on the current blood glucose level.

[0106] In this embodiment of the application, the base layer can also be trained based on the multimodal fusion features of different groups of people, wherein the multimodal fusion features of different groups of people can all be obtained based on the above method.

[0107] The base layer, trained with multimodal fusion features from different populations, can learn universal physiological patterns across different populations, providing a foundation for dynamically adjusting insulin prediction doses based on population classification layers, individual adaptation layers, and scenario specialization layers.

[0108] In this embodiment of the application, the population classification layer of the preset prediction model can use a classification algorithm to divide users into different population subgroups. The classification algorithm can use methods such as vector machines and decision trees. The population subgroups can include young type 1 diabetic users, carbohydrate-sensitive users, elderly type 2 diabetic users, users with insulin resistance, gestational diabetic users, and large populations with postprandial blood glucose fluctuations.

[0109] In addition, the population classification layer can be trained in advance. During training, the population classification layer can be trained based on the multimodal fusion features of young type 1 diabetic users, carbohydrate-sensitive users, elderly type 2 diabetic users, users with insulin resistance, gestational diabetic users, and large populations with postprandial blood glucose fluctuations. This allows the population classification layer to learn the unique characteristics of the multimodal fusion features of different population subgroups.

[0110] In this embodiment, there may be systematic differences in insulin sensitivity and response patterns to specific foods among users in different population subgroups. Based on the population classification layer in this embodiment, the user's population subgroup can be accurately located, the user's insulin sensitivity and response pattern to specific foods can be determined, and then the insulin dose can be precisely adjusted according to the user's insulin sensitivity and response pattern to specific foods.

[0111] In this embodiment, the individual adaptation layer of the preset prediction model employs a reinforcement learning (RL) algorithm to learn the unique response patterns of each user's blood glucose concentration and insulin. The reinforcement learning-based individual adaptation layer can use the rewards for maintaining the user's blood glucose concentration within the target range and the penalties for exceeding the target range as feedback signals, thereby dynamically adjusting the predicted insulin dose.

[0112] In addition, the individual adaptation layer can employ causal inference neural networks to identify the main causal relationships of blood glucose concentration fluctuations for each user. For example, some users' blood glucose concentrations may react strongly to emotions and stress, while others may react more strongly to the type of food they ingest.

[0113] During training, it can be based on the user's personal historical blood glucose concentration data, historical insulin dosing data, and physical activity data.

[0114] In addition, the individual adaptation layer can also transfer knowledge from the basic layer and the population classification layer to the user's individual adaptation layer based on transfer learning or meta-learning methods, and then further fine-tune the insulin prediction dose based on the user's own data.

[0115] In this embodiment, the scene specialization layer of the preset prediction model can simultaneously predict the user's blood glucose concentration trend and insulin prediction demand based on a multi-task learning algorithm. Additionally, an attention mechanism can be employed to dynamically adjust the attention given to different features in specific scenarios.

[0116] Furthermore, in this embodiment, the scene specialization layer can include multiple sub-scene layers, such as: scenes before and after meals, scenes before and after exercise, scenes before and after sleep, long-distance travel, weekday scenes, holiday scenes, and sick scenes. Each sub-scene layer can share the underlying feature representation, but has its own specialized part. When using it, attention-based models can be flexibly switched between different scenes, or conditional computation can be used to activate the corresponding sub-scene layer according to the current scene.

[0117] The scenario-specific layer in this application embodiment can accurately model the unique characteristics of users' insulin needs in different scenarios, thereby enabling precise prediction of insulin requirements based on these different scenarios. For example, pre-meal dose prediction may focus more on expected carbohydrate intake, while bedtime dose prediction focuses more on avoiding the risk of nocturnal hypoglycemia, improving the predictive ability to adapt to unconventional or special situations.

[0118] The preset model will now be described in detail with reference to specific embodiments, as follows: Suppose the user is a 40-year-old male with type 1 diabetes who is preparing to eat lunch at 2 pm on Saturday and has the following characteristics: Blood glucose characteristics: Current blood glucose is 7.8 mmol / L, showing an upward trend, with significant fluctuations in blood glucose over the past 24 hours; Physiological characteristics: Active insulin level (IOB) was 0.3 units; aerobic exercise was performed in the morning. Behavioral pattern characteristics: The user is about to consume a mixed meal containing 60g of carbohydrates. The typical carbohydrate intake for this user's lunch is 45-50g. Environmental characteristics: It's the weekend, the weather is warm, and users are in a relaxed mood.

[0119] The user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics are fused to obtain multimodal fusion features, which are then input into a preset model.

[0120] After obtaining the multimodal fusion features, the processing module 160 sends them to the base layer. Based on the multimodal fusion features, the base layer outputs a preliminary blood glucose prediction trajectory and an initial insulin dose (first insulin dose). For example, the base layer predicts that if the user does not inject insulin, their blood glucose will reach approximately 13 mmol / L two hours after a meal; the recommended initial insulin dose (first insulin dose) is 5 units (based on the average insulin / carbohydrate ratio of the general population). The first insulin dose is inferred from the base layer, which is trained on a large amount of user data, and only considers the basic relationships of the input multimodal fusion features, without adjusting for specific populations, individual characteristics, or environments. In addition, the base layer can also perform a preliminary assessment of the importance of each input feature; for example, in this scenario, the user's carbohydrate intake and current blood glucose level are identified as the most critical factors.

[0121] After the base layer, the multimodal fusion features are input to the population classification layer. The population classification layer analyzes the user's characteristics and historical data and classifies them as a type 1 diabetic, moderately insulin sensitive, and good aerobic exercise response population.

[0122] Based on the population classification, the population group adjusts the first estimated insulin dose predicted by the base layer. The typical insulin / carbohydrate ratio of this population group may be 10% lower than that of the general population, and the insulin dose usually needs to be further reduced after exercise. Therefore, the adjusted insulin dose is recommended to be reduced to 4.5 units, that is, the second estimated insulin dose is 4.5 units.

[0123] After the population classification layer, the multimodal fusion features are input into the individual adaptation layer. The individual adaptation layer analyzes the user's historical data and determines that the user's insulin sensitivity is usually about 15% higher on weekends than on weekdays (possibly due to lower stress). For this individual user, a mixed meal of 60g carbohydrates usually requires about 4 units of insulin. After exercise, the user especially needs to reduce the insulin dose by about 20% to avoid hypoglycemia. Based on this, the estimated second insulin dose is further adjusted to 3.8 units, i.e., the estimated third insulin dose is 3.8 units.

[0124] After the individual adaptation layer, the multimodal fusion features are input into the scene specialization layer. The scene specialization layer identifies that the user is in the specific scenario of "eating after exercise" and that the carbohydrate content of the meal is slightly higher than usual, thereby enhancing the importance of "large blood glucose fluctuations in the past 24 hours". Combined with blood glucose prediction based on multi-task learning, it is determined that even if 3.8 units of insulin are used, there is still about a 20% probability that the postprandial blood glucose will exceed 10 mmol / L. However, if the dose is increased further, the risk of hypoglycemia will increase significantly. Considering the special circumstances after exercise, the third insulin dose is kept unchanged, that is, the fourth insulin dose is still 3.8 units, and 3.8 units is output as the target insulin dose.

[0125] In the embodiments of this application, by processing layer by layer, starting from the most basic physiological laws and combining population characteristics, individual differences and specific scenarios, a highly personalized insulin dose that is adapted to the current scenario can be given.

[0126] The insulin dosage prediction system 100 in this embodiment acquires the user's blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics based on various data that can affect the user's blood glucose concentration. These are then fused to obtain multimodal fusion features, which are then used to predict the target insulin dose based on a preset prediction model. This target insulin dose not only considers the multimodal data that can affect the user's blood glucose concentration but also corrects for population classification, individual differences, and environmental influences, resulting in high accuracy. Furthermore, the blood glucose characteristics, physiological state characteristics, behavioral pattern characteristics, and environmental characteristics are obtained based on time grids with different time granularities, further enhancing accuracy.

[0127] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the technical scope disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. An insulin dosage estimation system, comprising: a blood glucose meter configured to obtain real-time continuous blood glucose concentration data of a target user; an insulin pump / pen configured to record insulin historical usage data of the target user; a dietary data collection device configured to collect nutrition intake data of the target user, and to record body index data, environmental factor data, psychological state data, special event data of the target user; a wearable device configured to collect physical activity data, physiological index data of the target user; a data storage module connected to the blood glucose meter, the insulin pump / pen, the dietary data collection device, and the wearable device, respectively, and configured to receive data collected by the blood glucose meter, the insulin pump / pen, the dietary data collection device, and the wearable device; a processing module connected to the data storage module, and configured to map the data obtained by the data storage module to a plurality of preset time grids with different time granularities; based on each time grid, to obtain blood glucose characteristics, physiological state characteristics, behavior pattern characteristics, and environmental characteristics of the target user; to fuse the blood glucose characteristics, the physiological state characteristics, the behavior pattern characteristics, and the environmental characteristics of the target user to obtain multi-modal fusion characteristics; based on the multi-modal fusion characteristics and a preset prediction model, to determine a target insulin dosage; wherein the preset prediction model comprises a base layer, a population classification layer, an individual adaptation layer, and a scenario specialization layer connected in sequence, the base layer is configured to determine a first insulin estimated dosage of the target user based on the multi-modal fusion characteristics, the population classification layer is configured to predict a sub-population classification of the target user based on the multi-modal fusion characteristics, and to correct the first insulin estimated dosage based on the sub-population classification to obtain a second insulin estimated dosage, the individual adaptation layer is configured to determine a physiological response pattern of the target user based on the multi-modal fusion characteristics, and to correct the second insulin estimated dosage based on the physiological response pattern of the target user to obtain a third insulin estimated dosage, and the scenario specialization layer is configured to determine a specific scenario in which the target user is located based on the multi-modal fusion characteristics, and to correct the third insulin estimated dosage based on the specific scenario to obtain a fourth insulin dosage, and to take the fourth insulin dosage as the target insulin dosage.

2. The insulin usage dose estimation system of claim 1, wherein, the plurality of preset time grids comprises a low-resolution time grid, a medium-resolution time grid, and a high-resolution time grid, the time granularity of the low-resolution time grid is greater than the time granularity of the medium-resolution time grid, and the time granularity of the medium-resolution time grid is greater than the time granularity of the high-resolution time grid; the processing module is configured to map the obtained data to the plurality of preset time grids in the following manner: map the obtained data to the high-resolution time grid, the medium-resolution time grid, and the low-resolution time grid, respectively.

3. The insulin usage dose estimation system of claim 1, wherein, the nutrition intake data comprises protein, carbohydrate, and fat intake data; the physical activity data comprises activity type, activity duration, and activity heart rate variation. The physiological index data includes: weight, body temperature, blood pressure, sleep quality; The physical index data includes: glycated hemoglobin, liver and kidney function, insulin sensitivity data; The environmental factor data includes: season, environmental temperature, environmental humidity; The psychological state data includes: stress and emotional state data; The special event data includes: disease data, menstrual data.

4. The insulin usage dose estimation system of claim 1, wherein, The blood glucose characteristics include the following sub-features: mean, standard deviation, rate of change, fluctuation frequency, peak and valley, circadian rhythm, and low blood sugar risk index of the user's blood glucose; The user's physiological state characteristics include the following sub-features: active insulin amount, insulin sensitivity dynamic characteristics, insulin absorption rate, energy balance evaluation data, carbohydrate utilization efficiency, metabolic state marker, inflammation state, and hormone fluctuation pattern; The user's behavior pattern characteristics include the following sub-features: meal regularity, carbohydrate preference, eating speed, exercise habit, daily activity intensity distribution, sedentary behavior recognition, sleep regularity, sleep quality, and sleep-wake transition characteristics; The environmental characteristics include the following sub-features: time of day, weekday / weekend feature, seasonal feature, holiday feature, social activity, travel state, temperature, humidity, altitude, emotional state, and stress level feature; The processing module is configured to obtain the user's blood glucose characteristics, physiological state characteristics, behavior pattern characteristics, and environmental characteristics based on each time grid based on the following manner: Determine the user's blood glucose characteristics, physiological state characteristics, behavior pattern characteristics, and environmental characteristics based on each sub-feature included in the user's blood glucose characteristics, physiological state characteristics, behavior pattern characteristics, and environmental characteristics, respectively.

5. The insulin dosage estimation system of claim 4, wherein the processing module is configured to determine the user's blood glucose characteristics, physiological state characteristics, behavior pattern characteristics, and environmental characteristics based on the following manner: Determine the blood glucose characteristics based on each sub-feature included in the blood glucose characteristics and the corresponding weight; Determine the physiological state characteristics based on each sub-feature included in the physiological state characteristics and the corresponding weight; Determine the behavior pattern characteristics based on each sub-feature included in the behavior pattern characteristics and the corresponding weight; Determine the environmental characteristics based on each sub-feature included in the environmental characteristics and the corresponding weight.

6. The insulin dosage estimation system of claim 5, wherein the processing module is further configured to determine the weight of each sub-feature included in the user's blood glucose characteristics, physiological state characteristics, and behavior pattern characteristics based on the importance of the sub-feature in the feature classification to which it belongs.

7. The insulin dosage estimation system of claim 1, wherein the processing module is further configured to fuse the user's blood glucose characteristics, physiological state characteristics, behavior pattern characteristics, and environmental characteristics to obtain the user's multi-modal fusion characteristics based on the following manner: Determine the reliability of the user's blood glucose characteristics, physiological state characteristics, behavior pattern characteristics, and environmental characteristics, respectively; determine weights of the blood glucose feature, the physiological state feature, the behavior pattern feature, and the environment feature of the user based on reliabilities of the blood glucose feature, the physiological state feature, the behavior pattern feature, and the environment feature of the user; weight and fuse the blood glucose feature, the physiological state feature, the behavior pattern feature, and the environment feature of the user based on the weights of the blood glucose feature, the physiological state feature, the behavior pattern feature, and the environment feature of the user, to obtain a multi-modal fusion feature containing real-time state and context information of the user. 8.The insulin usage dose estimation system of claim 7, wherein the processing module is further configured to: determine the reliability of the blood glucose feature based on signal quality, data freshness, and data missing rate of the real-time continuous blood glucose concentration data; determine the reliability of the physiological state feature based on accuracy and timeliness of the insulin historical usage data; determine the reliability of the behavior pattern feature based on timeliness and completeness of the nutrition intake data; determine the reliability of the environment feature based on accuracy, timeliness, and completeness of the environment factor data. 9.The insulin usage dose estimation system of claim 8, wherein the processing module is further configured to determine the weights of the blood glucose feature, the physiological state feature, the behavior pattern feature, and the environment feature of the user based on the following formula: wherein, BG represents blood glucose characteristics, PS represents physiological state characteristics, BH represents behavior pattern characteristics, ENV represents environmental characteristics, k e [BG, PS, BH, ENV], represents the sum of reliabilities of each characteristic, represents the weight, represents the reliability of this k, ; or, determine the weights of the blood glucose feature, the physiological state feature, the behavior pattern feature, and the environment feature of the user based on the following formula: where k e [BG, PS, BH, ENV], BG represents blood glucose characteristics, PS represents physiological state characteristics, BH represents behavior pattern characteristics, and ENV represents environmental characteristics, represents a prior importance, is a regulating factor, represents a weight. 10.The insulin usage dose estimation system of claim 5, wherein the processing module is further configured to determine the multi-modal fusion feature based on the following manner: wherein, representing a multi-modal fusion feature, representing a blood glucose feature weight, representing a blood glucose feature, representing a physiological state feature weight, representing a physiological state feature, representing a behavioral state feature weight, representing a behavioral state feature, representing an environmental feature weight, representing an environmental feature.

Images