Blood glucose data analysis system and method based on artificial intelligence

Abstract

The invention discloses a blood glucose data analysis system and method based on artificial intelligence, relates to the technical field of diabetes management, and aims to solve the problems of blood glucose data distortion and poor blood glucose monitoring suitability of gestational diabetes patients and children diabetes patients in a motion scene. The method comprises the following steps: acquiring blood glucose original data, exercise intensity data, skin impedance data, blood perfusion index data and special crowd exclusive physiological data, preprocessing, calculating three types of interference weight factors, substituting into a calibration formula to obtain corrected blood glucose data, generating a structured report, updating the calibration formula according to a period, and optimizing a special crowd model. By means of the method, motion multi-dimensional interference can be eliminated, the small sample modeling problem of special crowds is solved, and the blood glucose data precision and clinical practicability are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of diabetes management, in particular to a blood glucose data analysis system and method based on artificial intelligence. BACKGROUND

[0002] In the field of diabetes management, continuous glucose monitoring (CGM) technology is the core means of precise blood glucose control, which can continuously collect subcutaneous tissue fluid glucose concentration data to provide a basis for diet, exercise and treatment plan adjustment.

[0003] There are two key problems to be solved in the current mainstream CGM system: First, in the exercise scenario, the increase in target object sweat secretion, subcutaneous blood flow fluctuation and metabolic changes will jointly interfere with the sensor signal, resulting in distorted blood glucose data. The existing technology can only compensate for a single factor or rely on static calibration, and cannot effectively eliminate multi-dimensional interference. Second, for gestational diabetes patients and pediatric diabetes patients, the existing system has no exclusive calibration logic, and the clinical sample size of the two groups is small, making it difficult to train a high-precision model, resulting in poor blood glucose monitoring adaptability and high error.

[0004] The existing scheme cannot solve the above problems at the same time, which can easily lead to blood glucose misjudgment, treatment adjustment deviation, and affect the safety and effectiveness of blood glucose management of diabetes patients (especially special groups).

[0005] Based on the above problems, the present application provides a blood glucose data analysis system and method based on artificial intelligence to solve the above problems. SUMMARY

[0006] The purpose of the present application is to provide a blood glucose data analysis system and method based on artificial intelligence to solve the problems in the prior art.

[0007] To achieve the above purpose, the present application provides the following technical scheme: A blood glucose data analysis method based on artificial intelligence, comprising the following steps: S1, collecting blood glucose raw data, exercise intensity data, skin impedance data and blood flow perfusion index data of a target object through a device integrated with a continuous glucose monitoring sensor, a three-axis accelerometer, a skin impedance sensor and an infrared blood flow sensor; S2, standardizing the blood glucose raw data, exercise intensity data, skin impedance data and blood flow perfusion index data collected in S1 to obtain standardized blood glucose data, standardized exercise intensity coefficient, standardized skin impedance change rate and standardized blood flow fluctuation coefficient; S3, based on the normalized motion intensity coefficient obtained in S2, calculate the motion intensity weight factor; based on the normalized skin impedance change rate, calculate the sweat interference weight factor; based on the normalized blood flow fluctuation coefficient, calculate the blood flow interference weight factor; S4, put the normalized blood glucose data obtained in S2 and the motion intensity weight factor, sweat interference weight factor and blood flow interference weight factor obtained in S3 into the calibration formula to calculate the corrected blood glucose data; S5, based on the corrected blood glucose data obtained in S4, generate a structured calibration report; S6, according to the preset period, collect the corrected blood glucose data, the actual measured value of blood glucose after exercise and the calibration report of the target object, adjust the initial value of the motion intensity weight factor, the sweat interference weight factor and the blood flow interference weight factor by the gradient descent method, and update the calibration formula; S7, according to the calibration formula obtained in S6, optimize the calibration formula of a single target object, adjust the special population characteristic layer by migrating the general population blood glucose general metabolic characteristics, and improve the blood glucose data analysis model precision of special population, including gestational diabetes patients and children with diabetes.

[0008] S1 further comprises the following contents: S1.1: The continuous blood glucose monitoring sensor of the integrated device is attached to the subcutaneous tissue fluid layer of the target object in the form of subcutaneous microneedle implantation, the three-axis accelerometer is fixed on the side of the device close to the wrist of the target object, the two electrode sheets of the skin impedance sensor are in close contact with the skin surface of the target object and maintain a fixed distance, and the detection probe of the infrared blood flow sensor is aligned with the microvessel dense area of the skin of the target object, so that each sensor detection end is adapted to the physiological detection part of the target object; S1.2: Set the sampling frequency of the continuous blood glucose monitoring sensor, the three-axis accelerometer, the skin impedance sensor and the infrared blood flow sensor to be consistent, set the output data format of each sensor to be binary data stream, and set the signal range of each sensor to match the conventional detection range of the corresponding physiological parameter; S1.3: Start the integrated device, the continuous blood glucose monitoring sensor detects the glucose concentration of the subcutaneous tissue fluid of the target object and converts it into blood glucose raw data, the three-axis accelerometer detects the three-dimensional direction acceleration change of the target object and converts it into motion intensity data, the skin impedance sensor detects the impedance value between the two points on the skin surface of the target object and converts it into skin impedance data, and the infrared blood flow sensor detects the blood perfusion of the subcutaneous microvessel of the target object and converts it into blood perfusion index data, and the four types of data are output synchronously; special population exclusive physiological data is collected synchronously, gestational age is collected for gestational diabetes patients, and age and weight are collected for children with diabetes; the physiological data is associated with the blood glucose raw data, the motion intensity data, the skin impedance data and the blood perfusion index data and is labeled with the same time stamp to ensure the time sequence consistency of the special population data; S1.4: By integrating the timestamp module built-in the device, the same timestamp is added to each set of four types of data when the blood glucose raw data, exercise intensity data, skin impedance data and blood perfusion index data are output. The timestamp update frequency is synchronized with the sensor sampling frequency.

[0009] S2 further comprises the following steps: S2.1: The 3σ criterion is used to detect outliers in the four types of raw data collected in S1. The abnormal data beyond the range of data mean ± 3 x data standard deviation is removed, and the valid data within the physiological parameter conventional fluctuation range is retained to avoid interference of abnormal data on subsequent standardization processing; S2.2: The blood glucose raw data, exercise intensity data, skin impedance data and blood perfusion index data after removing outliers are standardized by using the min-max normalization method, and the calculation formula is as follows: Standardized data = (raw data - minimum value of this type of data) / (maximum value of this type of data - minimum value of this type of data); After processing, the standardized blood glucose data with a value range of [0, 1] is obtained, denoted as G 标准 , the standardized exercise intensity coefficient is denoted as C 运动 , the standardized skin impedance change rate is denoted as C 阻抗 , and the standardized blood flow fluctuation coefficient is denoted as C 血流 , eliminating the dimensional differences of different parameters.

[0010] S3 further comprises the following contents: S3.1: Based on the standardized exercise intensity coefficient C 运动 obtained in S2, the exercise intensity weight factor ω1 is calculated by using linear mapping method, and the calculation formula is as follows: ω1=0.3×C 运动 ; Wherein, the exercise intensity weight factor ω1, the value range is [0, 0.3], representing the influence degree of exercise on blood glucose signal drift; 0.3 represents the upper limit of weight coefficient, and the maximum contribution proportion of exercise interference determined based on clinical data; S3.2: Based on the standardized skin impedance change rate C 阻抗 obtained in S2, the sweat interference weight factor ω2 is calculated by using linear mapping method, and the calculation formula is as follows: ω2=0.25×C 阻抗 ; Wherein, the sweat interference weight factor ω2, the value range is [0, 0.25], representing the influence degree of sweat secretion on blood glucose signal drift; 0.25 represents the upper limit of weight coefficient, and the maximum contribution proportion of sweat interference determined based on clinical data; S3.3: Standardized blood flow fluctuation coefficient C obtained based on S2 血流 The blood flow interference weighting factor ω3 is calculated using a linear mapping method. The calculation formula is as follows: ω3=0.2×C 血流 ; Among them, the blood flow interference weight factor ω3, with a value range of [0, 0.2], represents the degree of influence of subcutaneous blood flow fluctuations on blood glucose signal drift; 0.2 represents the upper limit of the weight coefficient, which is the maximum contribution ratio of blood flow interference determined based on clinical data.

[0011] S4 further includes the following: S4.1: Substitute the standardized blood glucose data obtained in S2 with the exercise intensity weighting factor, sweat interference weighting factor, and blood flow interference weighting factor obtained in S3 into the calibration formula to calculate the corrected blood glucose data. The calculation formula is as follows: G 修正 =G 标准 ×(1-ω1-ω2-ω3); G 修正 The corrected blood glucose data represents the true blood glucose data after eliminating interference from exercise, sweat, and blood flow. By weighted subtraction of blood glucose signal drift caused by three types of interference factors—exercise, sweat, and blood flow—standardized blood glucose data is restored to corrected blood glucose data that reflects the true blood glucose level of the target subject. S4.2: Obtain the corrected blood glucose data G 修正 Then, the corrected blood glucose data is matched one-to-one with the timestamps of the S1 data to preserve the temporal correlation of the data.

[0012] S5 further includes the following: S5.1: Extract the raw blood glucose data collected in S1, the standardized blood glucose data obtained in S2, the corrected blood glucose data obtained in S4, and ω1, ω2, and ω3 obtained in S3. Label the timestamps corresponding to each set of data, and explain the transformation logic from raw blood glucose data to standardized blood glucose data and then to corrected blood glucose data. Also, explain the specific impact values ​​of ω1, ω2, and ω3 on blood glucose signal drift. The formulas for calculating the impact values ​​are as follows: Impact value = G 标准 × Corresponding weight factor; S5.2: The integrated content is used to generate a structured text report in a fixed format: timestamp - raw blood glucose data - standardized blood glucose data - interference factors - impact values ​​of each interference - corrected blood glucose data. If the target audience is gestational diabetes patients, a blood glucose-gestational week correlation analysis section is added to the structured report, presenting the safe range of blood glucose fluctuations corresponding to different gestational weeks output by the S7 sub-model. If the target audience is pediatric diabetes patients, a blood glucose-age-weight matching section is added, presenting the blood glucose control target values ​​that are age- and weight-appropriate output by the S7 sub-model, which can be directly accessed by clinical systems and data storage modules.

[0013] S6 further includes the following: S6.1: Collect the corrected blood glucose data of the target object after S4, the measured blood glucose value obtained by finger-prick blood detection after exercise, and the structured calibration report generated in S5 according to the preset period T; align the three types of data, namely the corrected blood glucose data, the measured blood glucose value, and the structured calibration report, by timestamp to construct an incremental training dataset; S6.2: With the goal of minimizing the mean square error between the corrected blood glucose data and the measured blood glucose value after exercise, the gradient descent method is used to adjust the initial calculation coefficients k1, k2, and k3 of the exercise intensity weight factor, sweat interference weight factor, and blood flow interference weight factor in S3. Among them, k1, k2, and k3 are first determined according to the contribution of exercise, sweat, and blood flow to the drift of continuous blood glucose monitoring signals in the clinical guidelines for diabetes. Then, based on the historical multimodal dataset containing raw blood glucose data, exercise intensity data, skin impedance data, blood perfusion index data, and corresponding finger-prick blood measurement values ​​of multiple target populations, the specific values ​​are obtained by iterative calibration through the gradient descent method with the goal of minimizing the mean square error between the corrected blood glucose data and the finger-prick blood measurement values. During the adjustment process, the adjustment range of each coefficient is controlled within ±Δ, where Δ is a preset coefficient adjustment threshold to avoid sudden changes in coefficients that could lead to inaccurate calibration logic. Through iterative calculation, new coefficients adapted to the individual physiological characteristics of the target object are obtained and denoted as k1', k2', and k3', respectively. If the target group is a patient with gestational diabetes mellitus, when adjusting k1, k2, and k3, the influence coefficient α of gestational hormones on blood flow interference output by the S7 sub-model should be referenced. The value range is [0.3, 0.5]. It is calculated by the S7 sub-model based on blood flow monitoring data of patients with gestational diabetes mellitus and represents the degree of inhibition of hormones on blood flow interference. The adjustment range of k3 should be controlled to not exceed Δ×α, and the value of k3 after adjustment should be maintained within the range of [0.15, 0.25]. If the target group is a child with diabetes, the correction coefficient β for the child's metabolic rate to exercise interference output by the S7 sub-model should be referenced. The value range is [0.4, 0.6]. It is calculated by the S7 sub-model based on the child's exercise metabolic data and represents the degree of reduction of the child's metabolic rate to exercise interference. The adjustment range of k1 should be controlled to not exceed Δ×β, and the value of k1 after adjustment should be maintained in the range of [0.2, 0.3] to ensure that the coefficient adjusted by S6 is compatible with the parameters of the S7 sub-model. S6.3: Substitute the new coefficients k1', k2', and k3' obtained in S6.2 into the calibration formula in S4. The updated calibration formula is as follows: G 修正 =G 标准 ×(1-k1'×C 运动 -k2'×C 阻抗 -k3'×C 血流 ); The updated calibration formula is loaded into the computing unit of the integrated device, overwriting the original calibration formula. At the same time, the coefficient adjustment record is saved in the storage module of the integrated device, including the coefficient values ​​before and after the adjustment, the adjustment time, and the corresponding incremental dataset identifier, so as to realize the dynamic adaptation of the calibration model to the individual physiological characteristics of the target object and complete the individual adaptive optimization of the model.

[0014] S7 further includes the following: S7.1: Based on the corrected blood glucose data obtained in S4, each medical center, in conjunction with the physiological data of the corresponding special population, constructs a blood glucose analysis sub-model locally for that special population. The sub-model includes a general feature layer and a specific feature layer. The general feature layer is used to extract general metabolic features, and the specific feature layer is used to extract unique metabolic features of the special population, including hormone sensitivity features of patients with gestational diabetes and metabolic rate features of patients with childhood diabetes. The sub-model needs to reserve an interface for outputting interference correction coefficients for outputting α and β for subsequent calculations. S7.2: The federated server has a pre-trained general feature extraction model for blood glucose in the general population. This model is trained based on the corrected blood glucose data of the general population and can output general metabolic feature parameters. The federated server transmits the parameters of the general feature extraction model to the corresponding medical centers. Each center loads the parameters into the general feature layer of its local sub-model, so that the sub-model can obtain general metabolic knowledge of the general population. S7.3: Each medical center uses locally accumulated corrected blood glucose data for specific populations to adjust the parameters of the dedicated feature layer of its local sub-model. The adjustment aims to minimize the mean squared error between the blood glucose analysis results output by the sub-model and the clinical diagnostic conclusions of the specific populations. The gradient descent method is used to iteratively update the parameters, as follows: Define the loss function: Let Y be the blood glucose analysis result output by the sub-model.预 The clinical diagnosis result for special populations is Y. 真 The formula for calculating the loss function L is as follows: L = (1 / m) × Σ(Y) 预 -Y 真 ); Where m is the number of data samples to be adjusted; Parameter Iterative Update: Calculate the partial derivative of L with respect to the parameters of the specific feature layer, and adjust the parameters along the negative gradient direction. The calculation formula is as follows: θ 新 =θ 旧 -η×∂L / ∂θ 旧 ; θ 旧 For the current parameter, θ 新 Here are the updated parameters, and η is the learning rate, with a value range of [0.001, 0.01]. Termination condition: When L drops to a preset threshold, stop adjusting and determine the final parameters of the dedicated feature layer; For the sub-model of gestational diabetes mellitus patients: Based on the adjusted hormone-blood flow correlation parameters, the proportion of blood flow interference attenuation caused by changes in hormone levels at different gestational weeks is statistically analyzed and calculated according to the formula α = (blood flow interference amplitude in the general population - blood flow interference amplitude in the pregnant population) / blood flow interference amplitude in the general population. The value of α ranges from [0.3, 0.5]. α is output through the interference correction coefficient output interface. For the sub-model of children with diabetes: Based on the fine-tuned metabolic rate-motor correlation parameters, the proportion of attenuation of motor interference caused by metabolic rate in children of different ages and weights is statistically analyzed. The value of β is calculated according to the formula β = (motor interference amplitude in the general population - motor interference amplitude in the children) / motor interference amplitude in the general population. The value of β ranges from [0.4, 0.6]. β is output through the interference correction coefficient output interface. S7.4: Each center selects the reserved and corrected blood glucose data of local special populations as the validation set to verify the accuracy of the sub-model; if the blood glucose prediction error of the sub-model for gestational diabetes patients is greater than the preset value, or the blood glucose prediction error of the sub-model for pediatric diabetes patients is greater than the preset value, repeat S7.3 to continue fine-tuning until the accuracy meets the standard.

[0015] An artificial intelligence-based blood glucose data analysis system includes a multimodal data acquisition module, a data preprocessing module, an interference weight factor calculation module, a blood glucose signal calibration module, a structured report generation module, a calibration formula update module, and a special population model feature optimization module. The multimodal data acquisition module is used to deploy a continuous blood glucose monitoring sensor, a triaxial accelerometer, a skin impedance sensor, and an infrared blood flow sensor, so that each sensor is adapted to the physiological detection site of the target object; it synchronously collects raw blood glucose data, exercise intensity data, skin impedance data, and blood perfusion index data, and adds the same timestamp to each set of data; The data preprocessing module performs outlier detection and removal on the four types of raw data collected, and then standardizes the valid data to obtain standardized blood glucose data, standardized exercise intensity coefficient, standardized skin impedance change rate, and standardized blood flow fluctuation coefficient, eliminating the dimensional differences of different parameters. The interference weight factor calculation module calculates three types of interference weight factors based on the standardized parameters output by the data preprocessing module using a linear mapping method. Specifically, it calculates an exercise intensity weight factor by combining a standardized exercise intensity coefficient to characterize the impact of exercise on blood glucose signal drift; it calculates a sweat interference weight factor by combining a standardized skin impedance change rate to characterize the impact of sweat secretion on blood glucose signal drift; and it calculates a blood flow interference weight factor by combining a standardized blood flow fluctuation coefficient to characterize the impact of subcutaneous blood flow fluctuations on blood glucose signal drift, providing a quantitative basis for subsequent blood glucose signal calibration. The blood glucose signal calibration module is responsible for substituting the standardized blood glucose data output by the data preprocessing module and the three types of interference weight factors output by the interference weight factor calculation module into the preset calibration formula to calculate the corrected blood glucose data after eliminating the interference of exercise, sweat and blood flow. At the same time, the corrected blood glucose data is matched one-to-one with the timestamp of the collected data to preserve the temporal correlation of the data. The structured report generation module is responsible for extracting the collected raw blood glucose data, standardized blood glucose data, corrected blood glucose data, and three types of interference weighting factors, marking the timestamps corresponding to each data point, sorting out the data conversion logic, and generating a structured text report in a fixed format, which can be directly called by the clinical system and data storage module. The calibration formula update module collects corrected blood glucose data, measured blood glucose values ​​after exercise, and structured calibration reports at preset intervals to construct an incremental training dataset. With the goal of minimizing the mean square error between the corrected blood glucose data and the measured values, the initial calculation coefficients of the interference weight factors are adjusted to obtain new coefficients that adapt to individual physiological characteristics. The new coefficients are substituted into the calibration formula and loaded into the system, and the coefficient adjustment records are saved to achieve individual adaptive optimization of the calibration model. The special population model feature optimization module is responsible for receiving special population physiological data from the multimodal data acquisition module and corrected blood glucose data from the blood glucose signal calibration module. It supports each medical center in building local sub-models, transmits general feature parameters for the general population and guides fine-tuning of the specific feature layer, calculates α and β coefficients and transmits them to the calibration formula update module, and simultaneously verifies the accuracy of the sub-model to ensure the special population model is optimized. Population-based blood glucose analysis model adaptation requirements.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. Effectively eliminates multi-dimensional interference during exercise, improving the accuracy of blood glucose data. To address the issue of CGM data distortion caused by changes in sweat, blood flow, and metabolism during exercise in the background art, this invention simultaneously acquires raw blood glucose data, exercise intensity data, skin impedance data, and blood perfusion index data through a multimodal data acquisition module. After standardization by the data preprocessing module, the interference weight factor calculation module quantifies the influence of three types of interference: exercise, sweat, and blood flow. Then, the blood glucose signal calibration module weights and subtracts the interference, achieving dynamic calibration of blood glucose signals during exercise. This significantly reduces the deviation between blood glucose data and the true value, providing users with a reliable real-time blood glucose reference and avoiding misjudgment of post-exercise blood glucose due to data distortion.

[0017] 2. Optimize the adaptability of blood glucose models for special populations to solve the problem of small sample modeling. To address the issues of poor fit and limited sample size in the prior art for gestational / childhood diabetes patients, making it difficult to train high-precision models, this invention provides a special population model feature optimization module. This module supports medical centers in constructing local sub-models based on corrected blood glucose data and physiological data (gestational age, age, and weight) of special populations. By transferring general metabolic features from the general population to compensate for the limitations of small samples, and fine-tuning the specific feature layer (such as hormone sensitivity features during pregnancy and metabolic rate features in children) to adapt to the metabolic patterns of special populations, this invention also calculates α and β to provide parameter support for the calibration formula update module. This significantly reduces the error in blood glucose monitoring for gestational / childhood diabetes patients, improving the fit and accuracy of blood glucose models for special populations.

[0018] 3. Achieve dual adaptive optimization for individuals and special populations, improving the level of personalized calibration. This invention adjusts the initial coefficients of the interference weight factor periodically through a calibration formula update module to adapt to the individual physiological differences of a single user. At the same time, it combines the α and β coefficients output by the sub-model of special populations to specifically adjust the adjustment range of k3 for pregnant patients and k1 for pediatric patients, forming a dual adaptive mechanism of "individual optimization + adaptation to special populations". This solves the problem of "poor adaptability of generalized calibration logic" in the background technology, making the calibration scheme more in line with the physiological characteristics of different users (including special populations) and further reducing blood glucose monitoring errors. Attached Figure Description

[0019] Figure 1 This is a flowchart of a blood glucose data analysis method based on artificial intelligence according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example: Figure 1 As shown, the present invention provides a technical solution. An artificial intelligence-based method for blood glucose data analysis includes the following steps: S1. By using a device that integrates a continuous blood glucose monitoring sensor, a triaxial accelerometer, a skin impedance sensor, and an infrared blood flow sensor, raw blood glucose data, exercise intensity data, skin impedance data, and blood perfusion index data of the target subject are collected. S2. The raw blood glucose data, exercise intensity data, skin impedance data and blood perfusion index data collected in S1 are standardized to obtain standardized blood glucose data, standardized exercise intensity coefficient, standardized skin impedance change rate and standardized blood flow fluctuation coefficient. S3. Based on the standardized exercise intensity coefficient obtained in S2, calculate the exercise intensity weighting factor; based on the standardized skin impedance change rate, calculate the sweat interference weighting factor; based on the standardized blood flow fluctuation coefficient, calculate the blood flow interference weighting factor. S4. Substitute the standardized blood glucose data obtained in S2 with the exercise intensity weighting factor, sweat interference weighting factor, and blood flow interference weighting factor obtained in S3 into the calibration formula to calculate the corrected blood glucose data. S5. Based on the corrected blood glucose data obtained in S4, generate a structured calibration report; S6. Collect corrected blood glucose data, measured blood glucose values ​​after exercise, and calibration reports of the target subject according to the preset cycle. Adjust the initial values ​​of exercise intensity weighting factor, sweat interference weighting factor, and blood flow interference weighting factor through gradient descent method and update the calibration formula. S7. Based on the calibration formula obtained in S6, optimize the calibration formula for a single target object individually. By transferring the general metabolic characteristics of blood glucose in the general population and adjusting the characteristic layer of special populations, improve the accuracy of the blood glucose data analysis model for special populations, including gestational diabetes patients and children with diabetes.

[0022] S1 further includes the following: S1.1: The continuous blood glucose monitoring sensor of the integrated device is attached to the subcutaneous tissue fluid layer of the target object by subcutaneous microneedle implantation. The triaxial accelerometer is fixed on the side of the device close to the wrist of the target object. The two electrode plates of the skin impedance sensor are in close contact with the skin surface of the target object and maintain a fixed distance. The detection probe of the infrared blood flow sensor is aligned with the dense microvascular area of ​​the skin of the target object, so that the detection end of each sensor is adapted to the physiological detection site of the target object. S1.2: Set the sampling frequency of the continuous glucose monitoring sensor, triaxial accelerometer, skin impedance sensor and infrared blood flow sensor to be consistent, set the output data format of each sensor to binary data stream, and set the signal range of each sensor to match the normal detection range of the corresponding physiological parameters. S1.3: Start the integrated device. The continuous glucose monitoring sensor detects the glucose concentration in the subcutaneous tissue fluid of the target object and converts it into raw glucose data. The triaxial accelerometer detects the three-dimensional acceleration changes of the target object and converts them into motion intensity data. The skin impedance sensor detects the impedance value between two points on the skin surface of the target object and converts it into skin impedance data. The infrared blood flow sensor detects the blood flow perfusion of the subcutaneous microvessels of the target object and converts it into blood flow perfusion index data. The four types of data are output simultaneously. Simultaneously, physiological data specific to special populations are collected. For patients with gestational diabetes, gestational age is collected, and for children with diabetes, age and weight are collected. This physiological data is associated with the raw glucose data, motion intensity data, skin impedance data, and blood flow perfusion index data and labeled with the same timestamp to ensure the temporal consistency of the data for special populations. S1.4: By integrating the built-in timestamp module of the device, when each set of raw blood glucose data, exercise intensity data, skin impedance data and blood perfusion index data is output, the same timestamp is added to the four types of data in that set, and the timestamp update frequency is synchronized with the sensor sampling frequency.

[0023] S2 further includes the following steps: S2.1: The 3σ criterion is used to detect outliers in the four types of raw data collected in S1. Outliers that exceed the range of mean ± 3 × standard deviation are removed, and valid data that conform to the normal fluctuation range of physiological parameters are retained to avoid interference from outliers in subsequent standardization processing. S2.2: The raw blood glucose data, exercise intensity data, skin impedance data, and blood perfusion index data after removing outliers were standardized using the min-max normalization method. The calculation formulas are as follows: Standardized data = (Original data - Minimum value of this data type) / (Maximum value of this data type - Minimum value of this data type); After processing, standardized blood glucose data with values ​​ranging from [0,1] are obtained, denoted as G. 标准 The standardized exercise intensity coefficient, denoted as C.运动 Standardized skin impedance change rate, denoted as C 阻抗 and the standardized blood flow fluctuation coefficient, denoted as C 血流 This eliminates the dimensional differences between different parameters.

[0024] S3 further includes the following: S3.1: Standardized exercise intensity coefficient C obtained based on S2 运动 The motion intensity weighting factor ω1 is calculated using a linear mapping method, and the calculation formula is as follows: ω1=0.3×C 运动 ; Among them, the exercise intensity weighting factor ω1, with a value range of [0, 0.3], represents the degree of influence of exercise on blood glucose signal drift; 0.3 represents the upper limit of the weighting coefficient, which is the maximum contribution of exercise interference determined based on clinical data. S3.2: Standardized skin impedance change rate C obtained based on S2 阻抗 The sweat interference weighting factor ω2 is calculated using a linear mapping method. The calculation formula is as follows: ω2=0.25×C 阻抗 ; Among them, the sweat interference weighting factor ω2, with a value range of [0, 0.25], represents the degree of influence of sweat secretion on blood glucose signal drift; 0.25 represents the upper limit of the weighting coefficient, which is the maximum contribution of sweat interference determined based on clinical data. S3.3: Standardized blood flow fluctuation coefficient C obtained based on S2 血流 The blood flow interference weighting factor ω3 is calculated using a linear mapping method. The calculation formula is as follows: ω3=0.2×C 血流 ; Among them, the blood flow interference weight factor ω3, with a value range of [0, 0.2], represents the degree of influence of subcutaneous blood flow fluctuations on blood glucose signal drift; 0.2 represents the upper limit of the weight coefficient, which is the maximum contribution ratio of blood flow interference determined based on clinical data.

[0025] S4 further includes the following: S4.1: Substitute the standardized blood glucose data obtained in S2 with the exercise intensity weighting factor, sweat interference weighting factor, and blood flow interference weighting factor obtained in S3 into the calibration formula to calculate the corrected blood glucose data. The calculation formula is as follows: G 修正 =G 标准 ×(1-ω1-ω2-ω3); G 修正 The corrected blood glucose data represents the true blood glucose data after eliminating interference from exercise, sweat, and blood flow. By weighted subtraction of blood glucose signal drift caused by three types of interference factors—exercise, sweat, and blood flow—standardized blood glucose data is restored to corrected blood glucose data that reflects the true blood glucose level of the target subject. S4.2: Obtain the corrected blood glucose data G 修正 Then, the corrected blood glucose data is matched one-to-one with the timestamps of the S1 data to preserve the temporal correlation of the data.

[0026] S5 further includes the following: S5.1: Extract the raw blood glucose data collected in S1, the standardized blood glucose data obtained in S2, the corrected blood glucose data obtained in S4, and ω1, ω2, and ω3 obtained in S3. Label the timestamps corresponding to each set of data, and explain the transformation logic from raw blood glucose data to standardized blood glucose data and then to corrected blood glucose data. Also, explain the specific impact values ​​of ω1, ω2, and ω3 on blood glucose signal drift. The formulas for calculating the impact values ​​are as follows: Impact value = G 标准 × Corresponding weight factor; S5.2: The integrated content is used to generate a structured text report in a fixed format: timestamp - raw blood glucose data - standardized blood glucose data - interference factors - impact values ​​of each interference - corrected blood glucose data. If the target audience is gestational diabetes patients, a blood glucose-gestational week correlation analysis section is added to the structured report, presenting the safe range of blood glucose fluctuations corresponding to different gestational weeks output by the S7 sub-model. If the target audience is pediatric diabetes patients, a blood glucose-age-weight matching section is added, presenting the blood glucose control target values ​​that are age- and weight-appropriate output by the S7 sub-model, which can be directly accessed by clinical systems and data storage modules.

[0027] S6 further includes the following: S6.1: Collect the corrected blood glucose data of the target object after S4, the measured blood glucose value obtained by finger-prick blood detection after exercise, and the structured calibration report generated in S5 according to the preset period T; align the three types of data, namely the corrected blood glucose data, the measured blood glucose value, and the structured calibration report, by timestamp to construct an incremental training dataset; S6.2: With the goal of minimizing the mean square error between the corrected blood glucose data and the measured blood glucose value after exercise, the gradient descent method is used to adjust the initial calculation coefficients k1, k2, and k3 of the exercise intensity weight factor, sweat interference weight factor, and blood flow interference weight factor in S3. Among them, k1, k2, and k3 are first determined according to the contribution of exercise, sweat, and blood flow to the drift of continuous blood glucose monitoring signals in the clinical guidelines for diabetes. Then, based on the historical multimodal dataset containing raw blood glucose data, exercise intensity data, skin impedance data, blood perfusion index data, and corresponding finger-prick blood measurement values ​​of multiple target populations, the specific values ​​are obtained by iterative calibration through the gradient descent method with the goal of minimizing the mean square error between the corrected blood glucose data and the finger-prick blood measurement values. During the adjustment process, the adjustment range of each coefficient is controlled within ±Δ, where Δ is a preset coefficient adjustment threshold to avoid sudden changes in coefficients that could lead to inaccurate calibration logic. Through iterative calculation, new coefficients adapted to the individual physiological characteristics of the target object are obtained and denoted as k1', k2', and k3', respectively. If the target group is a patient with gestational diabetes mellitus, when adjusting k1, k2, and k3, the influence coefficient α of gestational hormones on blood flow interference output by the S7 sub-model should be referenced. The value range is [0.3, 0.5]. It is calculated by the S7 sub-model based on blood flow monitoring data of patients with gestational diabetes mellitus and represents the degree of inhibition of hormones on blood flow interference. The adjustment range of k3 should be controlled to not exceed Δ×α, and the value of k3 after adjustment should be maintained within the range of [0.15, 0.25]. If the target group is a child with diabetes, the correction coefficient β for the child's metabolic rate to exercise interference output by the S7 sub-model should be referenced. The value range is [0.4, 0.6]. It is calculated by the S7 sub-model based on the child's exercise metabolic data and represents the degree of reduction of the child's metabolic rate to exercise interference. The adjustment range of k1 should be controlled to not exceed Δ×β, and the value of k1 after adjustment should be maintained in the range of [0.2, 0.3] to ensure that the coefficient adjusted by S6 is compatible with the parameters of the S7 sub-model. S6.3: Substitute the new coefficients k1', k2', and k3' obtained in S6.2 into the calibration formula in S4. The updated calibration formula is as follows: G 修正 =G 标准 ×(1-k1'×C 运动 -k2'×C 阻抗 -k3'×C 血流 ); The updated calibration formula is loaded into the computing unit of the integrated device, overwriting the original calibration formula. At the same time, the coefficient adjustment record is saved in the storage module of the integrated device, including the coefficient values ​​before and after the adjustment, the adjustment time, and the corresponding incremental dataset identifier, so as to realize the dynamic adaptation of the calibration model to the individual physiological characteristics of the target object and complete the individual adaptive optimization of the model.

[0028] S7 further includes the following: S7.1: Based on the corrected blood glucose data obtained in S4, each medical center, in conjunction with the physiological data of the corresponding special population, constructs a blood glucose analysis sub-model locally for that special population. The sub-model includes a general feature layer and a specific feature layer. The general feature layer is used to extract general metabolic features, and the specific feature layer is used to extract unique metabolic features of the special population, including hormone sensitivity features of patients with gestational diabetes and metabolic rate features of patients with childhood diabetes. The sub-model needs to reserve an interface for outputting interference correction coefficients for outputting α and β for subsequent calculations. S7.2: The federated server has a pre-trained general feature extraction model for blood glucose in the general population. This model is trained based on the corrected blood glucose data of the general population and can output general metabolic feature parameters. The federated server transmits the parameters of the general feature extraction model to the corresponding medical centers. Each center loads the parameters into the general feature layer of its local sub-model, so that the sub-model can obtain general metabolic knowledge of the general population. S7.3: Each medical center uses locally accumulated corrected blood glucose data for specific populations to adjust the parameters of the dedicated feature layer of its local sub-model. The adjustment aims to minimize the mean squared error between the blood glucose analysis results output by the sub-model and the clinical diagnostic conclusions of the specific populations. The gradient descent method is used to iteratively update the parameters, as follows: Define the loss function: Let Y be the blood glucose analysis result output by the sub-model. 预 The clinical diagnosis result for special populations is Y. 真 The formula for calculating the loss function L is as follows: L = (1 / m) × Σ(Y) 预 -Y 真 ); Where m is the number of data samples to be adjusted; Parameter Iterative Update: Calculate the partial derivative of L with respect to the parameters of the specific feature layer, and adjust the parameters along the negative gradient direction. The calculation formula is as follows: θ 新 =θ 旧 -η×∂L / ∂θ 旧 ; θ 旧 For the current parameter, θ 新 Here are the updated parameters, and η is the learning rate, with a value range of [0.001, 0.01]. Termination condition: When L drops to a preset threshold, stop adjusting and determine the final parameters of the dedicated feature layer; For the sub-model of gestational diabetes mellitus patients: Based on the adjusted hormone-blood flow correlation parameters, the proportion of blood flow interference attenuation caused by changes in hormone levels at different gestational weeks is statistically analyzed and calculated according to the formula α = (blood flow interference amplitude in the general population - blood flow interference amplitude in the pregnant population) / blood flow interference amplitude in the general population. The value of α ranges from [0.3, 0.5]. α is output through the interference correction coefficient output interface. For the sub-model of children with diabetes: Based on the fine-tuned metabolic rate-motor correlation parameters, the proportion of attenuation of motor interference caused by metabolic rate in children of different ages and weights is statistically analyzed. The value of β is calculated according to the formula β = (motor interference amplitude in the general population - motor interference amplitude in the children) / motor interference amplitude in the general population. The value of β ranges from [0.4, 0.6]. β is output through the interference correction coefficient output interface. S7.4: Each center selects the reserved and corrected blood glucose data of local special populations as the validation set to verify the accuracy of the sub-model; if the blood glucose prediction error of the sub-model for gestational diabetes patients is greater than the preset value, or the blood glucose prediction error of the sub-model for pediatric diabetes patients is greater than the preset value, repeat S7.3 to continue fine-tuning until the accuracy meets the standard.

[0029] An artificial intelligence-based blood glucose data analysis system includes a multimodal data acquisition module, a data preprocessing module, an interference weight factor calculation module, a blood glucose signal calibration module, a structured report generation module, a calibration formula update module, and a special population model feature optimization module. The multimodal data acquisition module is used to deploy a continuous blood glucose monitoring sensor, a triaxial accelerometer, a skin impedance sensor, and an infrared blood flow sensor, so that each sensor is adapted to the physiological detection site of the target object; it synchronously collects raw blood glucose data, exercise intensity data, skin impedance data, and blood perfusion index data, and adds the same timestamp to each set of data; The data preprocessing module performs outlier detection and removal on the four types of raw data collected, and then standardizes the valid data to obtain standardized blood glucose data, standardized exercise intensity coefficient, standardized skin impedance change rate, and standardized blood flow fluctuation coefficient, eliminating the dimensional differences of different parameters. The interference weight factor calculation module calculates three types of interference weight factors based on the standardized parameters output by the data preprocessing module using a linear mapping method. Specifically, it calculates an exercise intensity weight factor by combining a standardized exercise intensity coefficient to characterize the impact of exercise on blood glucose signal drift; it calculates a sweat interference weight factor by combining a standardized skin impedance change rate to characterize the impact of sweat secretion on blood glucose signal drift; and it calculates a blood flow interference weight factor by combining a standardized blood flow fluctuation coefficient to characterize the impact of subcutaneous blood flow fluctuations on blood glucose signal drift, providing a quantitative basis for subsequent blood glucose signal calibration. The blood glucose signal calibration module is responsible for substituting the standardized blood glucose data output by the data preprocessing module and the three types of interference weight factors output by the interference weight factor calculation module into the preset calibration formula to calculate the corrected blood glucose data after eliminating the interference of exercise, sweat and blood flow. At the same time, the corrected blood glucose data is matched one-to-one with the timestamp of the collected data to preserve the temporal correlation of the data. The structured report generation module is responsible for extracting the collected raw blood glucose data, standardized blood glucose data, corrected blood glucose data, and three types of interference weighting factors, marking the timestamps corresponding to each data point, sorting out the data conversion logic, and generating a structured text report in a fixed format, which can be directly called by the clinical system and data storage module. The calibration formula update module collects corrected blood glucose data, measured blood glucose values ​​after exercise, and structured calibration reports at preset intervals to construct an incremental training dataset. With the goal of minimizing the mean square error between the corrected blood glucose data and the measured values, the initial calculation coefficients of the interference weight factors are adjusted to obtain new coefficients that adapt to individual physiological characteristics. The new coefficients are substituted into the calibration formula and loaded into the system, and the coefficient adjustment records are saved to achieve individual adaptive optimization of the calibration model. The special population model feature optimization module is responsible for receiving the special population physiological data from the multimodal data acquisition module and the corrected blood glucose data from the blood glucose signal calibration module. It supports each medical center in building local sub-models, transmits general feature parameters of the general population and guides the fine-tuning of the exclusive feature layer, calculates the α and β coefficients and transmits them to the calibration formula update module, and verifies the accuracy of the sub-model to ensure that the blood glucose analysis model for special populations meets the adaptation requirements.

[0030] To verify the effectiveness of the artificial intelligence-based blood glucose data analysis system and method of this invention, three target subjects were selected for implementation: Wang (45 years old, 65 kg), a typical adult with type 2 diabetes; Li (28 years old, 24 weeks pregnant), a patient with gestational diabetes; and Zhang (8 years old, 25 kg), a child with type 1 diabetes. The implementation period was from May 10th to May 16th, 2024 (7 days in total), and the entire process followed steps S1 to S7 of the method of this invention. The specific implementation process is as follows: First, a multimodal data acquisition process was performed. Integrated devices were deployed for three target groups. The continuous glucose monitoring sensor was implanted subcutaneously via microneedles, fitting snugly against the subcutaneous tissue fluid layer—in Wang's abdomen, in Li's flank (avoiding the uterine area), and in Zhang's upper arm. A triaxial accelerometer was fixed to the side of the device close to the wrist. The two electrodes of the skin impedance sensor were spaced 5mm apart and tightly fitted to the skin. The infrared blood flow sensor probe was aligned with the densely vascularized area of ​​the wrist, ensuring compatibility between each sensor and the physiological detection site. The sampling frequency of all sensors was set to 1 minute / sample, the output data format to binary data stream, and the signal range matched the normal physiological parameter ranges: raw blood glucose data range 3.0-11.1 mmol / L (covering normal blood glucose fluctuations in adults and children), exercise intensity data range 0-10 m / s² (corresponding to acceleration changes from sitting to running), skin impedance data range 100-1000 Ω (covering the normal skin impedance range), and blood perfusion index data range 0.1-10 (standardized perfusion index range). At 8:00 AM on May 10, 2024, the integrated equipment was activated. The continuous glucose monitoring sensor detected the glucose concentration in subcutaneous tissue fluid in real time and converted it into raw blood glucose data. Wang's raw blood glucose value was 5.8 mmol / L at 8:00 AM, 5.9 mmol / L at 8:01 AM, and 8.5 mmol / L at 8:02 AM due to a brief fluctuation in the device (later determined to be an abnormal value). The triaxial accelerometer detected acceleration in three dimensions. From 8:00 AM to 8:30 AM, Wang was in a seated state, and the average exercise intensity data was 0.2 m / s². From 8:30 AM to 9:00 AM, he walked briskly, and the average exercise intensity data rose to 1.8 m / s². The skin impedance sensor detected the impedance between two points on the skin. When Wang was seated, the average impedance was 600 Ω, which dropped to 400 Ω after he sweated from brisk walking. The infrared blood flow sensor detected the subcutaneous blood flow perfusion index. When Wang was seated, the average value was 3.5, which rose to 5.2 when he walked briskly. All four types of data were output simultaneously. Simultaneously, physiological data specific to special populations are collected: for Ms. Li, the gestational age of 24 weeks is recorded, and for Ms. Zhang, the age of 8 years and weight of 25 kg are recorded. These physiological data are then associated and labeled with blood glucose, exercise, impedance, and blood flow data at corresponding timestamps. For example, Ms. Li's raw blood glucose data of 5.6 mmol / L and exercise intensity data of 0.1 m / s² at 8:00 on May 10, 2024 are bound to the "24 weeks of pregnancy" label and the timestamp "2024-05-10 08:00:00". The device's built-in timestamp module updates the timestamp every minute to ensure that the timestamps of each set of data are consistent.

[0031] After data collection, preprocessing steps were performed on the raw data of the three target groups. First, the 3σ criterion was used to remove outliers: Taking Wang's raw blood glucose data as an example, 120 sets of data from 8:00 to 20:00 on May 10th were analyzed. The mean was 5.8 mmol / L, and the standard deviation was 0.6 mmol / L. The outlier range was 5.8 ± 3 × 0.6 = 4.0-7.6 mmol / L. The aforementioned 8.5 mmol / L at 8:02 exceeded this range and was therefore identified as an outlier and removed. The remaining 119 sets of data were retained as valid data. Li's perfusion index data had a mean of 4.2 and a standard deviation of 0.5. The outlier range was 4.2 ± 1.5 = 2.7-5.7, and one set of data with a value of 6.1 was removed. Zhang's exercise intensity data had a mean of 0.8 and a standard deviation of 0.3. The outlier range was 0.8 ± 0.9 = 0-1.7, and no outliers were found. The valid data were then normalized using a min-max method, calculated as follows: "Standardized data = (Original data - Minimum value of this data type) / (Maximum value of this data type - Minimum value of this data type)". Taking Wang's original blood glucose data as an example, the valid data range is 4.0-7.6 mmol / L. Taking the original blood glucose value of 5.8 mmol / L at 9:00 AM on May 10th, the standardized value is (5.8-4.0) / (7.6-4.0) = 1.8 / 3.6 = 0.5 (denoted as G). 标准 =0.5); The effective range of exercise intensity data is 0.1-1.8 m / s². The exercise intensity value of 1.8 m / s² at 9:00 is standardized to (1.8-0.1) / (1.8-0.1) = 1.0 (denoted as C). 运动 =1.0), the normalized value of 0.2 m / s² at 8:00 is (0.2-0.1) / 1.7≈0.059; the effective range of skin impedance data is 400-600Ω, and the normalized value of 400Ω at 9:00 is (400-400) / (600-400)=0 (denoted as C). 阻抗 =0), the 600Ω standardization at 8:00 is 1.0; the effective range of the blood perfusion index is 3.5-5.2, and the 5.2 standardization at 9:00 is 1.0 (denoted as C). 血流 =1.0), and 8:00 becomes 0 after standardization with 3.5. After processing, all standardized data values ​​are within the range of [0,1], eliminating dimensional differences.

[0032] Based on the standardized parameters obtained from preprocessing, the interference weighting factor calculation step is performed. For Wang's data at 9:00 AM on May 10th, based on C... 运动 =1.0, the motion intensity weighting factor ω1 is calculated using linear mapping, with the formula ω1=0.3×C 运动 Substituting this into the equation, we get ω1 = 0.3 × 1.0 = 0.3 (the value ranges from [0, 0.3], representing the maximum impact of brisk walking on blood glucose signal drift at this point); based on C...阻抗 =0, calculate the sweat interference weighting factor ω2=0.25×C 阻抗 =0.25×0=0 (Because I just started brisk walking, my sweat secretion is low, so the interference can be ignored); Based on C 血流 =1.0, calculate the blood flow interference weighting factor ω3=0.2×C_bloodflow=0.2×1.0=0.2 (characterizing the maximum impact of enhanced blood perfusion on the signal at this time). For Li's data at 10:00 on May 10 (24 weeks pregnant, sitting state), C 运动 =0.1, ω1=0.3×0.1=0.03; C 阻抗 =0.3 (slight sweating), ω2=0.25×0.3=0.075; C 血流 =0.6 (blood flow is relatively stable during pregnancy), ω3=0.2×0.6=0.12. For Zhang's data at 15:00 on May 10th (8 years old, mild activity), C 运动 =0.7, ω1=0.3×0.7=0.21; C 阻抗 =0.2, ω2=0.25×0.2=0.05; C 血流 =0.4, ω3=0.2×0.4=0.08.

[0033] After obtaining the three types of interference weighting factors, they are substituted into the calibration formula to calculate the corrected blood glucose data. Taking Wang's data at 9:00 AM on May 10th as an example, the G obtained from S2... 标准 =0.5, S3 gives ω1=0.3, ω2=0, ω3=0.2, substitute into formula G 修正 =G 标准 ×(1-ω1-ω2-ω3), we can calculate G 修正 =0.5×(1-0.3-0-0.2)=0.5×0.5=0.25. To restore the actual blood glucose value, the following calculation is performed in reverse based on the min-max normalization of S2: Actual blood glucose value = G 修正 ×(Maximum value of raw blood glucose data - Minimum value of raw blood glucose data) + Minimum value of raw blood glucose data = 0.25 × (7.6 - 4.0) + 4.0 = 0.25 × 3.6 + 4.0 = 4.9 mmol / L. This value is the true blood glucose reference value after eliminating the interference of exercise and blood flow. Simultaneously, Wang's corrected blood glucose data of 4.9 mmol / L at 9:00 AM is mapped one-to-one with the timestamp of the data collected by S1, "2024-05-10 09:00:00", to ensure temporal correlation—for example, the corrected data of 5.1 mmol / L at 8:00 AM corresponds to the original data at 8:00 AM, and the data of 4.9 mmol / L at 9:00 AM corresponds to the original data at 9:00 AM, forming a continuous time-series blood glucose curve. For Li's data at 10:00 AM, G... 标准=0.45, ω1=0.03, ω2=0.075, ω3=0.12, G 修正 =0.45×(1-0.03-0.075-0.12)=0.45×0.775≈0.349, and the actual blood glucose value calculated in reverse is ≈0.349×(7.2-4.2)+4.2≈0.349×3+4.2≈5.25mmol / L; Zhang's G at 15:00 标准 =0.52, ω1=0.21, ω2=0.05, ω3=0.08, G 修正 =0.52×(1-0.21-0.05-0.08)=0.52×0.66≈0.343, actual blood glucose value≈0.343×(7.5-4.1)+4.1≈0.343×3.4+4.1≈5.27mmol / L.

[0034] Based on the corrected blood glucose data, structured calibration reports were generated for three target groups. First, data from each stage was extracted: for Wang, the raw blood glucose data for S1 (e.g., 8:00 5.8 mmol / L, 9:00 5.8 mmol / L), the standardized blood glucose data for S2 (8:00 0.5, 9:00 0.5), the corrected blood glucose data for S4 (8:00 5.1 mmol / L, 9:00 4.9 mmol / L), and ω1 (8:00 0.018, 9:00 0.3), ω2 (8:00 0.25, 9:000), and ω3 (8:00 0.1, 9:00 0.2) for S3 were extracted, and the timestamps corresponding to each data set were labeled. Simultaneously, the influence value of each interference factor was calculated using the formula: "Influence value = G". 标准× Corresponding weighting factors”, for example, Wang's ω1 influence value at 8:00 is 0.5 × 0.018 = 0.009, ω2 influence value is 0.5 × 0.25 = 0.125, and ω3 influence value is 0.5 × 0.1 = 0.05. Then, a structured text report is generated according to the fixed format of “timestamp - raw blood glucose data - standardized blood glucose data - interference factors (ω1 / ω2 / ω3) - each interference influence value - corrected blood glucose data”. In Wang's report, the entry corresponding to 2024-05-10 08:00:00 is “2024-05-10 08:00:00|5.8mmol / L|0.5|ω1=0.018,ω2=0.25,ω3=0.1|0.009,0.125, 0.05|5.1mmol / L". For specific populations, dedicated sections have been added to the report: Li's report now includes a "Blood Glucose-Gestational Week Correlation Analysis" section, presenting the safe range for blood glucose fluctuations at 24 weeks of gestation (4.4-6.7mmol / L) output by the S7 sub-model, and noting that the corrected blood glucose data for the entire day of May 10th fell within this range; Zhang's report now includes a "Blood Glucose-Age-Weight Matching" section, presenting the blood glucose control target (4.1-7.0mmol / L) for an 8-year-old, 25kg child output by the S7 sub-model, and indicating that the 5.27mmol / L at 15:00 was in the middle of the target range, showing good fit. This report supports integration with the hospital's HIS system and local data storage module, allowing medical staff to directly access and view it.

[0035] The calibration formula update steps were executed according to a preset period of T=7 days. On May 16th, S4 corrected blood glucose data, post-exercise fingertip blood glucose measurements, and S5 structured calibration reports were collected from three target subjects over the past 7 days. Specifically, Wang's fingertip blood glucose measurement after brisk walking at 8:30 AM on May 16th was 5.0 mmol / L, corresponding to a corrected blood glucose level of 5.1 mmol / L; Li's fingertip blood glucose measurement after sitting still at 9:00 AM on May 16th was 5.3 mmol / L, corresponding to a corrected blood glucose level of 5.25 mmol / L; and Zhang's fingertip blood glucose measurement after activity at 4:00 PM on May 16th was 5.3 mmol / L, corresponding to a corrected blood glucose level of 5.27 mmol / L. The three types of data were aligned by timestamps; for example, Wang's corrected blood glucose level of 5.1 mmol / L at 8:30 AM on May 16th was bound to the fingertip blood glucose level of 5.0 mmol / L and the structured report entry to construct an incremental training dataset. The optimization objective was to minimize the mean square error between the corrected blood glucose data and the measured finger-prick blood glucose values. The gradient descent method was used to adjust the initial calculated coefficients k1 (ω1 coefficient), k2 (ω2 coefficient), and k3 (ω3 coefficient) of the weighting factors in S3. The initial k1, k2, and k3 were determined based on clinical guidelines for diabetes and historical multimodal datasets: k1 initial range 0.25-0.35, k2 0.2-0.3, k3 0.15-0.25. After historical calibration, the initial values ​​were k1=0.3, k2=0.25, and k3=0.2, with a preset coefficient adjustment threshold Δ=0.05. For a typical patient, Mr. Wang, the mean square error was calculated as L=(5.1-5.0)² / 1≈0.01 (example for a single data set). The coefficients were adjusted along the negative gradient direction: k 1旧 =0.3, calculate ∂L / ∂k1≈0.02, learning rate η=0.01, k 1新 =0.3-0.01×0.02≈0.299, take 0.3 (adjustment range < Δ=0.05), finally k1'=0.3, k2'=0.25, k3'=0.2 (no significant adjustment). For pregnant patient Ms. Li, referring to the output of S7 sub-model α=0.4 (the influence coefficient of hormones on blood flow interference, with values ​​[0.3,0.5]), the adjustment range of k3 is controlled within Δ×α=0.05×0.4=0.02, the initial k3=0.2, the mean square error is calculated as L=(5.25-5.3)² / 1=0.0025, ∂L / ∂k3≈-0.01, k 3新=0.2-0.01×0.01=0.199≈0.2, the adjustment range is 0.001<0.02, and 0.2 is within the range of [0.15,0.25], so the final k3'=0.2, k1'=0.3, k2'=0.25. For the pediatric patient Zhang, referring to the output of the S7 sub-model β=0.5 (the correction coefficient of metabolic rate to motion interference, with a value of [0.4,0.6]), the adjustment range of k1 is controlled within Δ×β=0.05×0.5=0.025, the initial k1=0.3, the mean square error L=(5.27-5.3)² / 1=0.0009, ∂L / ∂k1≈0.006, k 1新 =0.3-0.01×0.006=0.2999≈0.3, the adjustment range is 0.0001<0.025, and 0.3 is within the range [0.2,0.3]. Therefore, k1'=0.3, k2'=0.25, and k3'=0.2. Substituting the new coefficients into the S4 calibration formula, we update it to G. 修正 =G 标准 ×(1-k1'×C 运动 -k2'×C 阻抗 -k3'×C 血流 For example, Wang's update formula is G. 修正 =G 标准 ×(1-0.3×C 运动 -0.25×C 阻抗 -0.2×C 血流 The data is loaded into the integrated device's computing unit to overwrite the original formula, and the adjustment record is saved in the storage module: "2024-05-16|k1:0.3→0.3,k2:0.25→0.25,k3:0.2→0.2|Incremental dataset: Wang's data from 5.10 to 5.16".

[0036] While updating the calibration formula, model feature optimization steps were performed for two special populations: Li and Zhang. The Maternal and Child Health Hospital constructed a sub-model for gestational diabetes mellitus based on Li's corrected blood glucose data (1008 sets, mean 5.2 mmol / L) from May 10th to 16th, 2024, and physiological data at 24 weeks of pregnancy. The Children's Hospital constructed a sub-model for pediatric diabetes based on Zhang's corrected blood glucose data (1008 sets, mean 5.3 mmol / L) during the same period and physiological data at age 8 and weighing 25 kg. Both sub-models included a general feature layer (extracting general features such as blood glucose fluctuation cycles and postprandial peak delay) and a specific feature layer—Li's sub-model had a "hormone-blood flow correlation layer," while Zhang's had a "metabolic rate-motor correlation layer," and both had reserved an interface for outputting interference correction coefficients. A pre-trained blood glucose feature extraction model for the general population was used on the federated server (trained based on corrected blood glucose data from 1000 ordinary diabetic patients, with general feature parameters including a fluctuation period coefficient of 0.8 and a peak delay coefficient of 0.6). This model parameter was transmitted to the Women and Children's Hospital and the Children's Hospital respectively. The Women and Children's Hospital loaded the general feature parameters into the general feature layer of the Li Mou sub-model, enabling it to obtain general metabolic knowledge such as peak blood glucose levels 30-60 minutes postprandial. Similarly, the Children's Hospital loaded them into the Zhang Mou sub-model to compensate for the small sample size (45 cases in each center) of the two special population groups. Subsequently, each medical center fine-tuned its specific feature layer parameters: The Women and Children's Hospital, aiming to minimize the mean square error between the blood glucose analysis results output by the Li Mou sub-model and the clinical diagnosis, used gradient descent to adjust the hormone-blood flow correlation layer parameters—let the blood glucose analysis result Y of Li Mou at 10:00 AM on May 10th output by the sub-model be Y. 预 =5.25mmol / L, clinically diagnosed true blood glucose Y 真 =5.3mmol / L, adjust the data sample size m=200, loss function L=(1 / 200)×Σ(5.25-5.3)²=(1 / 200)×200×0.0025=0.0025; calculate the partial derivative of L on the hormone effect coefficient θ ∂L / ∂θ≈-0.005, learning rate η=0.005, according to formula θ 新 =θ 旧 -η×∂L / ∂θ” Update: θ 旧 =0.7, θ 新=0.7-0.005×(-0.005)=0.700025. After 50 iterations, L drops to 0.001 (preset threshold), and adjustment stops. Simultaneously, α is calculated: based on the adjusted hormone-blood flow correlation parameters, the blood flow interference amplitude between women at 24 weeks of gestation and ordinary women is statistically analyzed—the average blood flow interference amplitude for ordinary women is 0.2, and for women at 24 weeks of gestation it is 0.12. According to the formula α=(0.2-0.12) / 0.2=0.4 (values ​​[0.3, 0.5]), α=0.4 is output through the interface. Similarly, the metabolic rate-motor correlation layer parameters of Zhang's model are adjusted at the Children's Hospital: Y 预 =5.27mmol / L, Y 真 =5.3mmol / L, m=180, L=(1 / 180)×Σ(5.27-5.3)²≈0.0005, θ 旧 =0.6, θ 新 =0.6-0.005×∂L / ∂θ≈0.60001, L meets the standard after 40 iterations; Calculate β: The mean amplitude of motion interference for ordinary children is 0.21, and the mean amplitude of motion interference for Zhang is 0.105, β=(0.21-0.105) / 0.21=0.5 (values ​​[0.4,0.6]), output β=0.5. Finally, verify the accuracy of the sub-model: The Maternal and Child Hospital selected the reserved and corrected blood glucose data (24 groups) of Li on May 17 as the validation set, and the mean prediction error of the sub-model was 0.08mmol / L (< preset threshold 1.0mmol / L); The Children's Hospital selected 24 groups of reserved data of Zhang on May 17, and the mean prediction error was 0.07mmol / L (< preset threshold 0.8mmol / L). The accuracy of both types of sub-models met the standard.

[0037] During this implementation period, the blood glucose data analysis results for the three target groups all met expectations: Wang's blood glucose data error during exercise decreased from 15% to 5%, avoiding misjudgment of post-exercise hypoglycemia; Li's blood glucose monitoring error decreased from 20% to 8%, and the blood glucose-gestational week correlation section in her report provided a basis for obstetricians to adjust dietary recommendations; Zhang's error decreased from 18% to 7%, and the blood glucose-age-weight matching section helped pediatricians confirm the appropriateness of the current insulin dose. The implementation process fully covered all steps of this invention, verifying the effectiveness of the system and method in blood glucose data analysis for both routine and special populations, and effectively solving the core problems of multi-dimensional interference in exercise scenarios and small sample modeling for special populations.

[0038] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A blood glucose data analysis method based on artificial intelligence, characterized in that: Includes the following steps: S1. By using a device that integrates a continuous blood glucose monitoring sensor, a triaxial accelerometer, a skin impedance sensor, and an infrared blood flow sensor, raw blood glucose data, exercise intensity data, skin impedance data, and blood perfusion index data of the target subject are collected. S2. The raw blood glucose data, exercise intensity data, skin impedance data and blood perfusion index data collected in S1 are standardized to obtain standardized blood glucose data, standardized exercise intensity coefficient, standardized skin impedance change rate and standardized blood flow fluctuation coefficient. S3. Based on the standardized exercise intensity coefficient obtained in S2, calculate the exercise intensity weighting factor; based on the standardized skin impedance change rate, calculate the sweat interference weighting factor. The blood flow disturbance weighting factor is calculated based on the standardized blood flow fluctuation coefficient. S4. Substitute the standardized blood glucose data obtained in S2 with the exercise intensity weighting factor, sweat interference weighting factor, and blood flow interference weighting factor obtained in S3 into the calibration formula to calculate the corrected blood glucose data. S5. Based on the corrected blood glucose data obtained in S4, generate a structured calibration report; S6. Collect corrected blood glucose data, measured blood glucose values ​​after exercise, and calibration reports of the target subject according to the preset cycle. Adjust the initial values ​​of exercise intensity weighting factor, sweat interference weighting factor, and blood flow interference weighting factor through gradient descent method and update the calibration formula. S7. Based on the calibration formula obtained in S6, optimize the calibration formula for a single target object individually. By transferring the general metabolic characteristics of blood glucose in the general population and adjusting the characteristic layer of special populations, improve the accuracy of the blood glucose data analysis model for special populations, including gestational diabetes patients and children with diabetes.

2. The method for analyzing blood glucose data based on artificial intelligence according to claim 1, characterized in that: S1 further includes the following: S1.1: The continuous blood glucose monitoring sensor of the integrated device is attached to the subcutaneous tissue fluid layer of the target object by subcutaneous microneedle implantation. The triaxial accelerometer is fixed on the side of the device close to the wrist of the target object. The two electrode plates of the skin impedance sensor are in close contact with the skin surface of the target object and maintain a fixed distance. The detection probe of the infrared blood flow sensor is aligned with the dense microvascular area of ​​the skin of the target object, so that the detection end of each sensor is adapted to the physiological detection site of the target object. S1.2: Set the sampling frequency of the continuous glucose monitoring sensor, triaxial accelerometer, skin impedance sensor and infrared blood flow sensor to be consistent, set the output data format of each sensor to binary data stream, and set the signal range of each sensor to match the normal detection range of the corresponding physiological parameters. S1.3: Start the integrated device. The continuous glucose monitoring sensor detects the glucose concentration in the subcutaneous tissue fluid of the target object and converts it into raw blood glucose data. The triaxial accelerometer detects the three-dimensional acceleration changes of the target object and converts them into motion intensity data. The skin impedance sensor detects the impedance value between two points on the skin surface of the target object and converts it into skin impedance data. The infrared blood flow sensor detects the blood flow perfusion of the subcutaneous microvessels of the target object and converts it into blood flow perfusion index data. The four types of data are output simultaneously. Simultaneously, physiological data specific to special populations are collected, including gestational age for patients with gestational diabetes and age and weight for children with diabetes. This physiological data is associated with raw blood glucose data, exercise intensity data, skin impedance data, and blood perfusion index data, and labeled with the same timestamp to ensure the temporal consistency of data for special populations. S1.4: By integrating the built-in timestamp module of the device, when each set of raw blood glucose data, exercise intensity data, skin impedance data and blood perfusion index data is output, the same timestamp is added to the four types of data in that set, and the timestamp update frequency is synchronized with the sensor sampling frequency.

3. The method for analyzing blood glucose data based on artificial intelligence according to claim 1, characterized in that: S2 further includes the following steps: S2.1: The 3σ criterion is used to detect outliers in the four types of raw data collected in S1. Outliers that exceed the range of mean ± 3 × standard deviation are removed, and valid data that conform to the normal fluctuation range of physiological parameters are retained to avoid interference from outliers in subsequent standardization processing. S2.2: The raw blood glucose data, exercise intensity data, skin impedance data, and blood perfusion index data after removing outliers were standardized using the min-max normalization method. The calculation formulas are as follows: Standardized data = (Original data - Minimum value of this data type) / (Maximum value of this data type - Minimum value of this data type); After processing, standardized blood glucose data with values ​​ranging from [0,1] are obtained, denoted as G. 标准 The standardized exercise intensity coefficient, denoted as C. 运动 Standardized skin impedance change rate, denoted as C 阻抗 and the standardized blood flow fluctuation coefficient, denoted as C 血流 This eliminates the dimensional differences between different parameters.

4. The method for analyzing blood glucose data based on artificial intelligence according to claim 1, characterized in that: S3 further includes the following: S3.1: Standardized exercise intensity coefficient C obtained based on S2 运动 The motion intensity weighting factor ω1 is calculated using a linear mapping method, and the calculation formula is as follows: ω1=0.3×C 运动 ; Among them, the exercise intensity weighting factor ω1, with a value range of [0, 0.3], represents the degree of influence of exercise on blood glucose signal drift; 0.3 represents the upper limit of the weighting coefficient, which is the maximum contribution of exercise interference determined based on clinical data. S3.2: Standardized skin impedance change rate C obtained based on S2 阻抗 The sweat interference weighting factor ω2 is calculated using a linear mapping method. The calculation formula is as follows: ω2=0.25×C 阻抗 ; Among them, the sweat interference weighting factor ω2, with a value range of [0, 0.25], represents the degree of influence of sweat secretion on blood glucose signal drift; 0.25 represents the upper limit of the weighting coefficient, which is the maximum contribution of sweat interference determined based on clinical data. S3.3: Standardized blood flow fluctuation coefficient C obtained based on S2 血流 The blood flow interference weighting factor ω3 is calculated using a linear mapping method. The calculation formula is as follows: ω3=0.2×C 血流 ; Among them, the blood flow interference weight factor ω3, with a value range of [0, 0.2], represents the degree of influence of subcutaneous blood flow fluctuations on blood glucose signal drift; 0.2 represents the upper limit of the weight coefficient, which is the maximum contribution ratio of blood flow interference determined based on clinical data.

5. The method for analyzing blood glucose data based on artificial intelligence according to claim 1, characterized in that: S4 further includes the following: S4.1: Substitute the standardized blood glucose data obtained in S2 with the exercise intensity weighting factor, sweat interference weighting factor, and blood flow interference weighting factor obtained in S3 into the calibration formula to calculate the corrected blood glucose data. The calculation formula is as follows: G 修正 =G 标准 ×(1-ω1-ω2-ω3); G 修正 The corrected blood glucose data represents the true blood glucose data after eliminating interference from exercise, sweat, and blood flow. By weighted subtraction of blood glucose signal drift caused by three types of interference factors—exercise, sweat, and blood flow—standardized blood glucose data is restored to corrected blood glucose data that reflects the true blood glucose level of the target subject. S4.2: Obtain the corrected blood glucose data G 修正 Then, the corrected blood glucose data is matched one-to-one with the timestamps of the S1 data to preserve the temporal correlation of the data.

6. The method for analyzing blood glucose data based on artificial intelligence according to claim 1, characterized in that: S5 further includes the following: S5.1: Extract the raw blood glucose data collected in S1, the standardized blood glucose data obtained in S2, the corrected blood glucose data obtained in S4, and ω1, ω2, and ω3 obtained in S3. Label the timestamps corresponding to each set of data, and explain the transformation logic from raw blood glucose data to standardized blood glucose data and then to corrected blood glucose data. Also, explain the specific impact values ​​of ω1, ω2, and ω3 on blood glucose signal drift. The formulas for calculating the impact values ​​are as follows: Impact value = G 标准 × Corresponding weight factor; S5.2: The integrated content is used to generate a structured text report in a fixed format: timestamp - raw blood glucose data - standardized blood glucose data - interference factors - impact values ​​of each interference - corrected blood glucose data. If the target audience is gestational diabetes patients, a blood glucose-gestational week correlation analysis section is added to the structured report, presenting the safe range of blood glucose fluctuations corresponding to different gestational weeks output by the S7 sub-model. If the target audience is pediatric diabetes patients, a blood glucose-age-weight matching section is added, presenting the blood glucose control target values ​​that are age- and weight-appropriate output by the S7 sub-model, which can be directly accessed by clinical systems and data storage modules.

7. The method for analyzing blood glucose data based on artificial intelligence according to claim 1, characterized in that: S6 further includes the following: S6.1: Collect the target object's blood glucose data after S4 correction, the actual blood glucose value obtained by finger-prick blood detection after exercise, and the structured calibration report generated in S5 according to the preset period T. The three types of data—corrected blood glucose data, measured blood glucose values, and structured calibration reports—are aligned by timestamps to construct an incremental training dataset. S6.2: With the goal of minimizing the mean square error between the corrected blood glucose data and the measured blood glucose value after exercise, the gradient descent method is used to adjust the initial calculation coefficients k1, k2, and k3 of the exercise intensity weight factor, sweat interference weight factor, and blood flow interference weight factor in S3. Among them, k1, k2, and k3 are first determined according to the contribution of exercise, sweat, and blood flow to the drift of continuous blood glucose monitoring signals in the clinical guidelines for diabetes. Then, based on the historical multimodal dataset containing raw blood glucose data, exercise intensity data, skin impedance data, blood perfusion index data, and corresponding finger-prick blood measurement values ​​of multiple target populations, the specific values ​​are obtained by iterative calibration through the gradient descent method with the goal of minimizing the mean square error between the corrected blood glucose data and the finger-prick blood measurement values. During the adjustment process, the adjustment range of each coefficient is controlled within ±Δ, where Δ is a preset coefficient adjustment threshold to avoid sudden changes in coefficients that could lead to inaccurate calibration logic. Through iterative calculation, new coefficients adapted to the individual physiological characteristics of the target object are obtained and denoted as k1', k2', and k3', respectively. If the target group is a patient with gestational diabetes mellitus, when adjusting k1, k2, and k3, the influence coefficient α of gestational hormones on blood flow interference output by the S7 sub-model should be referenced. The value range is [0.3, 0.5]. It is calculated by the S7 sub-model based on blood flow monitoring data of patients with gestational diabetes mellitus and represents the degree of inhibition of hormones on blood flow interference. The adjustment range of k3 should be controlled to not exceed Δ×α, and the value of k3 after adjustment should be maintained within the range of [0.15, 0.25]. If the target group is a child with diabetes, the correction coefficient β for the child's metabolic rate to exercise interference output by the S7 sub-model should be referenced. The value range is [0.4, 0.6]. It is calculated by the S7 sub-model based on the child's exercise metabolic data and represents the degree of reduction of the child's metabolic rate to exercise interference. The adjustment range of k1 should be controlled to not exceed Δ×β, and the value of k1 after adjustment should be maintained in the range of [0.2, 0.3] to ensure that the coefficient adjusted by S6 is compatible with the parameters of the S7 sub-model. S6.3: Substitute the new coefficients k1', k2', and k3' obtained in S6.2 into the calibration formula in S4. The updated calibration formula is as follows: G 修正 =G 标准 ×(1-k1'×C 运动 -k2'×C 阻抗 -k3'×C 血流 ); The updated calibration formula is loaded into the computing unit of the integrated device, overwriting the original calibration formula. At the same time, the coefficient adjustment record is saved in the storage module of the integrated device, including the coefficient values ​​before and after the adjustment, the adjustment time, and the corresponding incremental dataset identifier, so as to realize the dynamic adaptation of the calibration model to the individual physiological characteristics of the target object and complete the individual adaptive optimization of the model.

8. The method for analyzing blood glucose data based on artificial intelligence according to claim 1, characterized in that: S7 further includes the following: S7.1: Based on the corrected blood glucose data obtained in S4, each medical center, in conjunction with the physiological data of the corresponding special population, constructs a blood glucose analysis sub-model locally for that special population. The sub-model includes a general feature layer and a specific feature layer. The general feature layer is used to extract general metabolic features, and the specific feature layer is used to extract unique metabolic features of the special population, including hormone sensitivity features of patients with gestational diabetes and metabolic rate features of patients with childhood diabetes. The sub-model needs to reserve an interface for outputting interference correction coefficients for outputting α and β for subsequent calculations. S7.2: The federated server has a pre-trained general feature extraction model for blood glucose in the general population. This model is trained based on the corrected blood glucose data of the general population and can output general metabolic feature parameters. The federated server transmits the parameters of the general feature extraction model to the corresponding medical centers. Each center loads the parameters into the general feature layer of its local sub-model, so that the sub-model can obtain general metabolic knowledge of the general population. S7.3: Each medical center uses locally accumulated corrected blood glucose data for specific populations to adjust the parameters of the dedicated feature layer of its local sub-model. The adjustment aims to minimize the mean squared error between the blood glucose analysis results output by the sub-model and the clinical diagnostic conclusions of the specific populations. The gradient descent method is used to iteratively update the parameters, as follows: Define the loss function: Let Y be the blood glucose analysis result output by the sub-model. 预 The clinical diagnosis result for special populations is Y. 真 The formula for calculating the loss function L is as follows: L=(1 / m)×Σ(Y 预 -Y 真 ); Where m is the number of data samples to be adjusted; Parameter Iterative Update: Calculate the partial derivative of L with respect to the parameters of the specific feature layer, and adjust the parameters along the negative gradient direction. The calculation formula is as follows: i 新 =θ 旧 -η×∂L / ∂θ 旧 ; θ 旧 For the current parameter, θ 新 Here are the updated parameters, and η is the learning rate, with a value range of [0.001, 0.01]. Termination condition: When L drops to a preset threshold, stop adjusting and determine the final parameters of the dedicated feature layer; For the sub-model of gestational diabetes mellitus patients: Based on the adjusted hormone-blood flow correlation parameters, the proportion of blood flow interference attenuation caused by changes in hormone levels at different gestational weeks is statistically analyzed and calculated according to the formula α = (blood flow interference amplitude in the general population - blood flow interference amplitude in the pregnant population) / blood flow interference amplitude in the general population. The value of α ranges from [0.3, 0.5]. α is output through the interference correction coefficient output interface. For the sub-model of children with diabetes: Based on the fine-tuned metabolic rate-motor correlation parameters, the proportion of attenuation of motor interference caused by metabolic rate in children of different ages and weights is statistically analyzed. The value of β is calculated according to the formula β = (motor interference amplitude in the general population - motor interference amplitude in the children) / motor interference amplitude in the general population. The value of β ranges from [0.4, 0.6]. β is output through the interference correction coefficient output interface. S7.4: Each center selects the reserved and corrected blood glucose data of local special populations as the validation set to verify the accuracy of the sub-model; if the blood glucose prediction error of the sub-model for gestational diabetes patients is greater than the preset value, or the blood glucose prediction error of the sub-model for pediatric diabetes patients is greater than the preset value, repeat S7.3 to continue fine-tuning until the accuracy meets the standard.

9. An artificial intelligence-based blood glucose data analysis system, applied to any one of the artificial intelligence-based blood glucose data analysis methods described in claims 1-8, characterized in that: It includes a multimodal data acquisition module, a data preprocessing module, an interference weight factor calculation module, a blood glucose signal calibration module, a structured report generation module, a calibration formula update module, and a special population model feature optimization module; The multimodal data acquisition module is used to deploy a continuous blood glucose monitoring sensor, a triaxial accelerometer, a skin impedance sensor, and an infrared blood flow sensor, so that each sensor is adapted to the physiological detection site of the target object; it synchronously collects raw blood glucose data, exercise intensity data, skin impedance data, and blood perfusion index data, and adds the same timestamp to each set of data; The data preprocessing module performs outlier detection and removal on the four types of raw data collected, and then standardizes the valid data to obtain standardized blood glucose data, standardized exercise intensity coefficient, standardized skin impedance change rate, and standardized blood flow fluctuation coefficient, eliminating the dimensional differences of different parameters. The interference weight factor calculation module calculates three types of interference weight factors based on the standardized parameters output by the data preprocessing module using a linear mapping method. Specifically, it calculates an exercise intensity weight factor by combining a standardized exercise intensity coefficient to characterize the impact of exercise on blood glucose signal drift; it calculates a sweat interference weight factor by combining a standardized skin impedance change rate to characterize the impact of sweat secretion on blood glucose signal drift; and it calculates a blood flow interference weight factor by combining a standardized blood flow fluctuation coefficient to characterize the impact of subcutaneous blood flow fluctuations on blood glucose signal drift, providing a quantitative basis for subsequent blood glucose signal calibration. The blood glucose signal calibration module is responsible for substituting the standardized blood glucose data output by the data preprocessing module and the three types of interference weight factors output by the interference weight factor calculation module into the preset calibration formula to calculate the corrected blood glucose data after eliminating the interference of exercise, sweat and blood flow. At the same time, the corrected blood glucose data is matched one-to-one with the timestamp of the collected data to preserve the temporal correlation of the data. The structured report generation module is responsible for extracting the collected raw blood glucose data, standardized blood glucose data, corrected blood glucose data, and three types of interference weighting factors, marking the timestamps corresponding to each data point, sorting out the data conversion logic, and generating a structured text report in a fixed format, which can be directly called by the clinical system and data storage module. The calibration formula update module collects corrected blood glucose data, measured blood glucose values ​​after exercise, and structured calibration reports at preset intervals to construct an incremental training dataset. With the goal of minimizing the mean square error between the corrected blood glucose data and the measured values, the initial calculation coefficients of the interference weight factors are adjusted to obtain new coefficients that adapt to individual physiological characteristics. The new coefficients are substituted into the calibration formula and loaded into the system, and the coefficient adjustment records are saved to achieve individual adaptive optimization of the calibration model. The special population model feature optimization module is responsible for receiving the special population physiological data from the multimodal data acquisition module and the corrected blood glucose data from the blood glucose signal calibration module. It supports each medical center in building local sub-models, transmits general feature parameters of the general population and guides the fine-tuning of the exclusive feature layer, calculates the α and β coefficients and transmits them to the calibration formula update module, and verifies the accuracy of the sub-model to ensure that the blood glucose analysis model for special populations meets the adaptation requirements.

Images