Pregnancy blood glucose fluctuation prediction and intervention recommendation system based on multi-source time sequence data and blood glucose dynamic prediction method

Abstract

The invention belongs to the technical field of intelligent medical treatment, and provides a pregnancy blood glucose fluctuation prediction and intervention recommendation system based on multi-source time sequence data and a blood glucose dynamic prediction method, and the system comprises a data acquisition layer, a data processing layer, an algorithm calculation layer and an application interaction layer. The data acquisition layer comprises a CGM access module, a glucometer access module, an intelligent wearable device access module and an EHR / EMR access module; the data processing layer comprises a data cleaning module, a time sequence database and a feature engineering module; the algorithm calculation layer comprises a blood glucose prediction module, a user blood glucose portrait generation module and an intervention recommendation module; the application interaction layer is an operation and feedback entrance for doctors and users and comprises a doctor end, a user end, an early warning module and an interaction module. According to the method, the multi-source time series data related to the blood glucose fluctuation are integrated, the dynamic correlation model is established for predicting the blood glucose fluctuation, the monitoring blank is filled, the monitoring frequency is optimized, and the user compliance is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent medical treatment, and particularly relates to a pregnancy blood glucose fluctuation prediction and intervention recommendation system based on multi-source time sequence data and a blood glucose dynamic prediction method. BACKGROUND

[0002] GDM, i.e. gestational diabetes mellitus, is a kind of abnormal glucose tolerance that occurs or is found for the first time during pregnancy. The onset of GDM is closely related to insulin resistance caused by an increase in placental hormone secretion in the middle and late stages of pregnancy. The incidence in some areas of China has exceeded 20%. As a common complication during pregnancy, GDM poses a significant threat to maternal and infant health, can cause complications such as gestational hypertension and polyhydramnios in pregnant women, and can increase the risk of adverse outcomes such as fetal growth restriction, macrosomia and neonatal hypoglycemia. In addition, the probability of patients developing type 2 diabetes after delivery is as high as 30%-50%.

[0003] In clinical diagnosis and treatment, the monitoring and management of GDM mainly rely on fingertip blood glucose monitoring and glycated hemoglobin (HbA1c) detection, but there are still obvious limitations. First, traditional fingertip blood sampling can only provide discrete single-point blood glucose data, which cannot fully reflect key dynamic fluctuation information such as postprandial blood glucose peak delay and asymptomatic nocturnal hypoglycemia throughout the day, resulting in a monitoring blind area and easy missed optimal intervention window. Second, HbA1c reflects the average blood glucose level in the past 2-3 months, which cannot reflect short-term blood glucose fluctuations and has limited guiding significance for this stage of rapid blood glucose change during pregnancy. Third, frequent daily fingertip blood puncture not only brings physical pain and inconvenience to pregnant women, but also causes great psychological burden and anxiety, which seriously reduces the compliance of patients to monitoring and subsequent intervention programs, and becomes a bottleneck for blood glucose management. In addition, existing intervention recommendations are mostly universal programs, which lack dynamic adjustment based on individual real-time metabolic characteristics and behavior patterns, and it is difficult to achieve precise and personalized management. Therefore, a technical solution is needed to realize continuous blood glucose trend prediction and provide personalized intervention recommendations to make up for the shortcomings of existing technologies. SUMMARY

[0004] To solve the problems that the existing GDM monitoring and management methods are fragmented and lagging, and it is difficult to meet the differentiated management needs and the monitoring burden of pregnant women is too large, the application provides a pregnancy blood glucose fluctuation prediction and intervention recommendation system based on multi-source time sequence data and a blood glucose dynamic prediction method.

[0005] In one scheme, the pregnancy blood glucose fluctuation prediction and intervention recommendation system based on multi-source time sequence data comprises a data acquisition layer, a data processing layer, an algorithm calculation layer and an application interaction layer. The data acquisition layer includes a CGM access module, a blood glucose meter access module, a smart wearable device access module, and an EHR / EMR access module. The blood glucose meter access module and the CGM access module collect point and line data of the user's blood glucose, respectively. The smart wearable device access module obtains the corresponding lifestyle data of the user during the blood glucose monitoring period. The EHR / EMR access module imports the user's medical and physiological data, thereby forming the access of the user's multi-source data. The data processing layer includes a data cleaning module, a time-series database, and a feature engineering module. The data cleaning module removes outliers and performs interpolation to complete the multi-source data provided by the data acquisition layer. The time-series database provides storage space for the multi-source data. The feature engineering module extracts time-series features from the multi-source data and correlates the features of the multi-source data to capture their combined effects on blood glucose. The algorithm computation layer includes a blood glucose prediction module, a user blood glucose profile generation module, and an intervention recommendation module. The blood glucose prediction module uses an LSTM-Transformer hybrid model to predict the blood glucose fluctuations of users in a specified future period based on time-series correlation. The user blood glucose profile generation module uses K-Means++ clustering to divide users into groups based on their blood glucose metabolism patterns, forming user profile tags. The intervention recommendation module responds based on the blood glucose fluctuation prediction results and dynamically optimizes the general intervention recommendation scheme by combining it with the user profile tags. The application interaction layer serves as the entry point for doctors and users to perform operations and provide feedback. The application interaction layer includes the doctor's side, the user's side, the early warning module, and the interaction module.

[0006] Furthermore, the system is used in a periodic update and calibration prediction manner. When the system is used for the first time and each time the system is updated, CGM detects the user's blood glucose fluctuations over the past 14 days. Each update supports several rounds of short-term predictions. The CGM data collection interval is once every 5 minutes. When using the finger-prick blood glucose meter for the first time, blood glucose data should be measured 1 to 3 days before the target prediction cycle. In subsequent cycles, data should be calibrated by testing the finger-prick blood glucose meter the day before each prediction cycle. The frequency of finger-prick blood glucose testing is 3 to 4 times a day, corresponding to the user's fasting and post-meal status. The CGM access module mainly provides minute-level continuous blood glucose data and derived features for CGM detection. The collected features include minute-level blood glucose values ​​within the detection period, blood glucose collection timestamps corresponding to the blood glucose values, postprandial peak features, blood glucose change rate, periodic fluctuation features, and abnormal event markers. The blood glucose meter access module mainly provides single-point accurate data at specific time points of finger prick blood glucose meter testing, and calibrates and supplements the CGM data; The user's lifestyle characteristics acquired by the smart wearable device access module include heart rate, exercise data, sleep data, and daily routine data; The user medical and physiological characteristics imported by the EHR / EMR access module include maternal basic physiological data, metabolic function-related indicators, and liver and kidney function indicators.

[0007] Furthermore, the data cleaning module mainly performs data cleaning for CGM and comparison and correction between finger blood and CGM data. The CGM data cleaning method uses linear interpolation and anomaly filtering based on clinical rules. The comparison and correction between finger blood and CGM data is performed through time alignment and the establishment of a linear calibration model. The feature engineering module uses a 1-hour sliding window with a step size of 5 minutes. Within the 1-hour sliding window, it extracts statistical features, trend features, and morphological features of CGM data. At the same time, it transforms continuous data of lifestyle behavior into discrete event markers, including eating events, exercise events, and sleep events. The feature engineering module performs feature association based on time windows for multi-source data. For each event time point, the CGM feature vector is concatenated with the event flag vector, and the concatenated feature vector is directly used as the input of the prediction model.

[0008] Furthermore, the LSTM-Transformer model architecture of the blood glucose prediction module includes an input layer, an LSTM layer, a Transformer encoder layer, and an output layer, and the LSTM-Transformer model is trained using autoregression. The input layer uses a fixed-length historical window of 24 hours, 48 ​​hours, or 72 hours as the input length, and directly uses the time-series feature sequence processed by the feature engineering module through feature association as the input data; The LSTM layer receives continuous feature sequences and captures the associated patterns between blood glucose fluctuations and event markers within a local time step through a gating mechanism. The Transformer encoder layer takes the hidden state sequence at each time point output by the LSTM layer as the input of the Transformer encoder, uses the Transformer's self-attention mechanism to calculate the relationship score between any two time points in the sequence, and integrates global information of the entire input length window. The output layer passes the final hidden state output by the Transformer through a fully connected network; and outputs a tensor with the blood glucose value sequence at time points of the target prediction period.

[0009] Furthermore, the user blood glucose profile generation module uses K-Means++ clustering to group users with similar blood glucose metabolism patterns into one category and assigns clinically meaningful labels to each group. The number of clusters is determined by combining the elbow rule with silhouette coefficient analysis, and the exploration range of k is 2-10; the execution logic of K-Means++ clustering is as follows: The first step is to randomly select the first centroid. The second step is to calculate the shortest distance between each data point and the existing centroids. The larger the distance value, the greater the probability of it being selected as the next centroid. The third step is to repeat the second step until k centroids are selected. The fourth step is to use the standard K-Means algorithm to iterate until the centroids are stable. After clustering is completed, the centroid of each cluster is analyzed, and clinical labels are assigned to each cluster based on its representative characteristics. The label types include postprandial high glucose, fasting high glucose, dawn phenomenon, low glycemic risk, GI sensitive, and exercise high response.

[0010] Furthermore, the blood glucose fluctuation prediction results are divided into three categories based on the degree of abnormality: severe abnormality, moderate abnormality, and mild abnormality. Severe abnormalities require guiding users to seek emergency medical treatment, while moderate and mild abnormalities are supported by intervention recommendation modules that generate intervention recommendations. The intervention plan for moderate abnormalities needs to be submitted to the doctor for review. The general intervention recommendation plan of the intervention recommendation module is based on the "IF-THEN" rule base of evidence-based medicine. When the intervention recommendation module is executed, it matches a rough, multi-option general plan framework from the medical knowledge base based on abnormal blood glucose fluctuation data. Then, it filters, strengthens or specifies the items in the general plan according to the user's tag type to optimize the plan. Solution optimization follows the priority order of effectiveness over security over compliance.

[0011] Furthermore, to achieve the above objectives, this application also provides a method for dynamic blood glucose prediction. This method is used to implement the aforementioned gestational blood glucose fluctuation prediction and intervention recommendation system based on multi-source time-series data. The method specifically includes the following steps: S1. Selective screening for GDM risk factors before 15 weeks of gestation: Users are selectively screened by fasting blood glucose, glycated hemoglobin or random blood glucose to exclude users who already had diabetes before pregnancy, and non-diabetic users are divided into normal pregnant women and high-risk pregnant women. S2, Weeks 15-23 of pregnancy, mid-pregnancy risk warning period: For ordinary pregnant women using the system for the first time to predict blood glucose fluctuations, the single prediction cycle is 3 days. Routine finger prick blood tests are performed 2 days a week, once a day. System updates are performed at 15 weeks and 20 weeks of pregnancy. During system updates, CGM tests are performed on the user's historical blood glucose fluctuations for 14 days. For high-risk pregnant women using the system for the first time to predict blood glucose fluctuations, the single prediction cycle is 3 days. Routine finger prick blood tests are performed 2 days a week, twice a day. System updates are performed at 15 weeks, 19 weeks and 23 weeks of pregnancy. During system updates, CGM tests are performed on the user's blood glucose fluctuations over the past 14 days. S3, 24-28 weeks of gestation, mid-pregnancy screening and diagnosis period: At 24 weeks of gestation, all pregnant women underwent an OGTT screening, which further categorized users into ordinary pregnant women, high-risk pregnant women without a confirmed diagnosis, and pregnant women diagnosed with GDM. For ordinary pregnant women, the single-round prediction cycle of the system is 2 days, and routine finger prick blood tests are performed 2 days a week, twice a day. The system is updated at 24 weeks and 27 weeks of pregnancy, and the CGM test is performed on the user's blood glucose fluctuations over the past 14 days at the time of the system update. For pregnant women who have not been diagnosed with high-risk status, the single-round prediction cycle of the system is 2 days. Routine finger prick blood tests are performed 2 days a week, 3 to 4 times a day. The system is updated at 26 weeks of pregnancy, and the CGM test is performed on the user's blood glucose fluctuations over the past 14 days at the time of the system update. Pregnant women diagnosed with GDM should stop using the system for prediction and follow routine monitoring methods in accordance with treatment requirements. S4. Weeks 29-36 of pregnancy, peak resistance intervention period in late pregnancy: For ordinary pregnant women, the single-round prediction cycle of the system is 2 days, and routine finger prick blood tests are performed 3 to 4 days a week, 3 times a day. The system is updated at 30 weeks, 33 weeks and 36 weeks of pregnancy. When the system is updated, CGM tests are performed on the user's blood glucose fluctuations over the past 14 days. For pregnant women who have not been diagnosed with high-risk status, the single-round prediction cycle of the system is 1 to 2 days. Routine finger prick blood tests are performed 4 days a week, 3 to 4 times a day. The system is updated at 29 weeks, 32 weeks and 35 weeks of pregnancy. At the time of the system update, CGM tests are performed on the user's blood glucose fluctuations over the past 14 days. S5. Week 37 of pregnancy - delivery, labor preparation period: For both normal pregnant women and pregnant women who have not been diagnosed with high risk, the single-round prediction cycle of the system is 1 day; the system is updated every 1.5 weeks, and CGM detects the user's blood glucose fluctuations over the past 7 days when the system is updated.

[0012] Furthermore, the frequency of routine finger-prick blood glucose tests at each stage of pregnancy is dynamically adjusted based on the error of the prediction cycle of each system. Specifically, four tests per day correspond to the user's fasting and three post-meal blood glucose levels; three tests per day correspond to the user's fasting and two post-meal blood glucose levels; two tests per day correspond to the user's fasting and one post-meal blood glucose level; and one test per day corresponds to the user's fasting blood glucose level. The specific adjustment rules are as follows: For pregnancies between 15 and 23 weeks, if the systematic prediction error is less than 0.3 mmol / L, routine finger-prick blood tests should be reduced by one day per week; if the systematic prediction error is greater than 0.8 mmol / L, routine finger-prick blood tests should be increased by one day per week. During 24-28 weeks of gestation, if the systematic prediction error is less than 0.5 mmol / L, routine finger prick blood tests should be reduced to twice a day; if the systematic prediction error is greater than 1.2 mmol / L, routine finger prick blood tests should be increased to four times a day. For pregnancies between 29 and 36 weeks, if the systematic prediction error is less than 0.4 mmol / L, routine finger prick blood tests should be increased to 4 times a day on one day of the week, and maintained at 3 times a day on the other testing days; if the systematic prediction error is greater than 1.0 mmol / L, routine finger prick blood tests should be maintained at 4 times a day on the testing days. During the 37th week of pregnancy to delivery, the system prediction error was greater than 0.8 mmol / L, so a bedtime blood glucose test was added.

[0013] Furthermore, when the average error of the system for two consecutive prediction cycles is greater than or equal to 1.0 mmol / L, the user's blood glucose profile is corrected. When the average error of the system over three consecutive prediction cycles is greater than or equal to 1.2 mmol / L, the parameters of the prediction model are updated and the model is iterated.

[0014] Furthermore, the system provides differentiated intervention recommendations for different abnormal blood glucose events that occur in the prediction results of different pregnancy stages; the definition of abnormal blood glucose is based on the proportion of blood glucose approaching the threshold and the magnitude of exceeding the threshold, and intervention or verification actions are triggered when either of the two conditions is met. The abnormality criteria and triggering standards for each stage of pregnancy are as follows: During the 15-23 week period of pregnancy, the predicted data is close to the target threshold for more than 30% of the time, or the predicted data exceeds the threshold by more than 1.0 mmol / L. During the 24-28 week period of pregnancy, the predicted data is close to the target threshold for more than 15% of the time, or the predicted data exceeds the threshold by more than 0.8 mmol / L. During the 29-36 week period of pregnancy, the predicted data is close to the target threshold for more than 10% of the time, or the predicted data exceeds the threshold by more than 0.5 mmol / L. From 37 weeks of gestation to delivery, the predicted data is close to the target threshold for more than 5% of the time, or the predicted data exceeds the threshold by more than 0.3 mmol / L.

[0015] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application constructs a multi-source time-series data association network of "blood glucose-diet-exercise-physiological indicators" by integrating CGM blood glucose fluctuation data, finger-prick blood glucose data, lifestyle data, and medical physiological data. This breaks down the information fragmentation barriers of traditional single-point detection, enabling continuous and comprehensive capture of the complete blood glucose fluctuation curve, including postprandial peaks and nighttime troughs, achieving full-cycle blood glucose insight. The system predicts and guides periodic testing, dynamically assesses user risk, and develops differentiated monitoring strategies for users at different stages of pregnancy, reducing unnecessary finger-prick blood punctures, alleviating the physical and mental stress of pregnant women, and improving their willingness and compliance with long-term monitoring management. At the same time, this application dynamically optimizes the general intervention plan based on different user blood glucose profiles, outputting personalized intervention plans that match the individual user's metabolism and lifestyle habits, enhancing the accuracy of intervention. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the system modules of a pregnancy blood glucose fluctuation prediction and intervention recommendation system based on multi-source time series data in one embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0019] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0020] To address the issues of fragmented and outdated existing GDM monitoring and management methods, which fail to meet differentiated management needs and place an excessive monitoring burden on pregnant women, this application provides a gestational blood glucose fluctuation prediction and intervention recommendation system based on multi-source time-series data. Please refer to [link to relevant documentation]. Figure 1 The system includes a data acquisition layer, a data processing layer, an algorithm calculation layer, and an application interaction layer.

[0021] In this embodiment, the data acquisition layer includes a CGM access module, a blood glucose meter access module, a smart wearable device access module, and an EHR / EMR access module; the blood glucose meter access module and the CGM access module collect point and line data of the user's blood glucose, respectively; the smart wearable device access module obtains the corresponding lifestyle data of the user during the blood glucose monitoring period; the EHR / EMR access module imports the user's medical and physiological data, thereby forming the access of the user's multi-source data.

[0022] CGM, or Continuous Glucose Monitoring, continuously monitors the glucose concentration in subcutaneous interstitial fluid using a glucose sensor, providing 24 / 7 blood glucose information. A finger-prick blood glucose meter is a portable medical device that quickly measures blood glucose levels by pricking the fingertip to collect capillary blood. Smart wearable devices are primarily smart fitness trackers, which now possess multi-dimensional physiological data collection capabilities, providing relevant lifestyle data to the system. EMR refers to a user's electronic medical record, recording medical information such as medical history, examination results, and treatment plans; while EHR is a more comprehensive patient health information system, based primarily on electronic medical records, but also including shared medical data such as laboratory results and impact reports.

[0023] In this embodiment, the data processing layer includes a data cleaning module, a time-series database, and a feature engineering module. The data cleaning module removes outliers and performs interpolation to complete the multi-source data provided by the data acquisition layer. The time-series database provides storage space for the multi-source data. The feature engineering module extracts time-series features from the multi-source data and associates the multi-source data features to capture their combined effects on blood glucose.

[0024] In this embodiment, the algorithm computation layer includes a blood glucose prediction module, a user blood glucose profile generation module, and an intervention recommendation module. The blood glucose prediction module uses an LSTM-Transformer hybrid model to predict the blood glucose fluctuations of users in a specified future period based on temporal correlation. The user blood glucose profile generation module uses K-Means++ clustering to divide users into groups based on their blood glucose metabolism patterns, forming user profile tags. The intervention recommendation module responds based on the blood glucose fluctuation prediction results and dynamically optimizes the general intervention recommendation scheme by combining it with the user profile tags.

[0025] In this embodiment, the application interaction layer serves as the entry point for doctors and users to operate and provide feedback. The application interaction layer includes a doctor's end, a user's end, an early warning module, and an interaction module. Specifically, the doctor's end and the user's end can be a web platform or an app; the doctor's end mainly involves blood glucose trend charts, profile tabs, and intervention plan review pages; the user's end mainly involves daily blood glucose curves, intervention plans, and intervention execution records. The early warning module and the feedback module assist the doctor's end and the user's end in interacting based on system predictions and intervention execution status.

[0026] Therefore, this application constructs a multi-source time-series data association network of "blood glucose-diet-exercise-physiological indicators" by integrating CGM blood glucose fluctuation data, finger-prick blood glucose data, lifestyle data, and medical physiological data. This breaks down the information fragmentation barriers of traditional single-point detection, enabling continuous and comprehensive capture of the complete blood glucose fluctuation curve, including postprandial peaks and nighttime troughs. This fundamentally fills the monitoring gap and provides an unprecedented data foundation for precise intervention. Simultaneously, this application, through high-precision short-term blood glucose prediction, can dynamically assess user risk and intelligently determine differentiated finger-prick blood glucose testing frequencies for users at different stages of pregnancy and risk levels. During periods of stable blood glucose, unnecessary punctures can be appropriately reduced, thereby greatly alleviating the psychological and physical stress of pregnant women while ensuring medical safety, and improving their willingness and adherence to long-term monitoring and management. Furthermore, this application, by constructing a "user blood glucose profile" and combining it with future blood glucose fluctuation predictions for intervention recommendations, enables deep personalization of intervention programs. It dynamically optimizes general intervention programs based on different user blood glucose profiles, outputting personalized intervention programs that match individual user metabolism and lifestyle habits, improving the effectiveness of interventions and user experience.

[0027] In one embodiment, since the metabolic characteristics of pregnant women are relatively stable at each stage of pregnancy and do not change drastically in the short term, the system is used in a periodic update and calibration prediction manner. After each update, the system can be used for subsequent multiple rounds of prediction while maintaining prediction accuracy. When the system is first used and each time it is updated, CGM detects the user's blood glucose fluctuations over a historical 14 days. Each update supports several subsequent rounds of short-term predictions; the CGM data collection interval is once every 5 minutes.

[0028] When using the finger-prick blood glucose meter for the first time, it measures blood glucose data for 1 to 3 days before the target prediction cycle. In subsequent cycles, the data is calibrated by testing with the finger-prick blood glucose meter the day before each prediction cycle. The finger-prick blood glucose test frequency is 3 to 4 times a day, corresponding to the user's fasting and post-meal status.

[0029] In this embodiment, the CGM access module primarily provides minute-level continuous blood glucose data and derived features from CGM detection. The collected features include minute-level blood glucose values ​​within the detection period, the corresponding blood glucose collection timestamps, postprandial peak characteristics, blood glucose change rate, periodic fluctuation characteristics, and abnormal event markers. Real-time blood glucose baseline data is used to construct blood glucose time-series sequences and cross-modal data associations. Blood glucose fluctuation derived parameters can extract trend and amplitude features to support the construction of user blood glucose profiles, and the fluctuation rate can also correct the prediction curve. The blood glucose meter access module primarily provides single-point accurate data at specific time points from finger-prick blood glucose meter detection, using standard point blood glucose values ​​to correct CGM drift. During calibration testing, blood glucose values ​​are compared with model prediction values, and the mean absolute error is calculated for response judgment.

[0030] In this embodiment, the user's lifestyle characteristics acquired by the smart wearable device access module include heart rate, exercise data, sleep data, and daily routine data. Moderate exercise intensity can promote muscle glucose uptake and lower postprandial blood glucose, while high-intensity exercise may lead to stress-induced hyperglycemia; these can be interpreted in conjunction with blood glucose data to explain corresponding blood glucose fluctuations. In sleep data, insufficient sleep or poor sleep quality reduces insulin sensitivity and increases fasting blood glucose; sleep interruptions lead to increased nocturnal glucocorticoid secretion, exacerbating blood glucose fluctuations. Irregular daily routines disrupt the biological clock and affect insulin secretion rhythms. Therefore, by temporally correlating various types of lifestyle data with blood glucose data, behavioral events contributing to blood glucose fluctuations can be explained.

[0031] In this embodiment, the user's medical physiological characteristics imported by the EHR / EMR access module include maternal basic physiological data, metabolic function-related indicators, and liver and kidney function indicators. Maternal basic physiological data includes age, pre-pregnancy BMI, blood pressure, etc., used to construct the user's individual baseline characteristics. Metabolic function-related indicators include fasting insulin (assessing insulin secretion capacity), insulin resistance index (quantifying insulin resistance), and four lipid profiles. Liver and kidney function indicators include creatinine, blood urea nitrogen, alanine aminotransferase (ALT), and aspartate aminotransferase (AST). Kidney insufficiency affects insulin excretion, leading to elevated blood glucose, while abnormal liver function can cause gluconeogenesis disorders.

[0032] In one embodiment, based on the foregoing embodiments, the data cleaning module mainly performs data cleaning for CGM and comparison and correction of finger blood and CGM data; the CGM data cleaning method adopts linear interpolation and anomaly filtering based on clinical rules, and the comparison and correction of finger blood and CGM data is performed by time alignment and establishing a linear calibration model.

[0033] Specifically, the missing data handling logic for CGM data is as follows: for CGM data with a 5-minute interval, if there are fewer than 3 consecutive missing points, linear interpolation is performed; if there are more than 3 missing points, the data segment is marked as invalid and skipped in subsequent feature extraction. The outlier handling logic for CGM data uses an absolute threshold method. Based on clinical knowledge, extreme points with blood glucose levels less than 2.2 mmol / L or greater than 13.9 mmol / L are directly removed or marked, and replaced with interpolated values. The 15-minute window for 3 missing points is considered short-term and does not affect the trend; furthermore, CGM itself may have transient interruptions. Blood glucose levels of 2.2 mmol / L and 13.9 mmol / L are critical values ​​for emergency intervention during pregnancy; single data points exceeding these ranges are highly likely due to equipment error.

[0034] The execution logic for comparing and correcting finger-prick blood and CGM data can be as follows: Pair daily finger-prick blood data with the CGM value closest to the current time point; using the finger-prick blood data as the baseline and the CGM data as input, establish a simple linear regression model for each day; use this model to calibrate the CGM data for the day as a whole. On short timescales, the systematic errors of CGM and finger-prick blood typically exhibit a good linear relationship, making the linear regression model simple and effective.

[0035] In this embodiment, the feature engineering module uses a 1-hour sliding window with a step size of 5 minutes. The 1-hour window size can match the key periods of postprandial blood glucose fluctuations, which aligns with the time nodes of clinical monitoring indicators, namely 1 hour and 2 hours postprandial. At the same time, the time granularity of the 1-hour window can achieve a balance between feature richness and computational efficiency, ensuring the effectiveness of the features; the 5-minute step size is synchronized with the original data, ensuring the continuity of the features.

[0036] Furthermore, the feature engineering module extracts statistical, trend, and morphological features of CGM data within a 1-hour sliding window, while simultaneously transforming continuous lifestyle data into discrete event markers, including eating events, exercise events, and sleep events. Statistical features include mean, standard deviation, minimum, maximum, and coefficient of variation; trend features include the slope of linear regression within the window; and morphological features include whether it represents a local extremum. The event marking logic for lifestyle data can be as follows: based on user-reported or inferred meal times, a 2-hour "post-meal status" marker is generated, i.e., an eating event; when the heart rate consistently exceeds the resting heart rate by 20 bpm for more than 10 minutes, an exercise event is marked, and its duration and average intensity are recorded; sleep events can be marked based on sleep schedules.

[0037] In this embodiment, the feature engineering module performs feature association based on time window event association analysis of multi-source data. For each event time point, the CGM feature vector and the event flag vector are concatenated, and the concatenated feature vector is directly used as the input of the prediction model.

[0038] In one embodiment, based on the foregoing embodiments, the LSTM-Transformer model architecture of the blood glucose prediction module includes an input layer, an LSTM layer, a Transformer encoder layer, and an output layer, and the LSTM-Transformer model undergoes autoregressive training. Autoregressive training means that during training, the model uses its previous prediction result as the input for the next step to simulate the inference process and improve the stability of long-term predictions.

[0039] In this embodiment, the input layer uses a fixed-length historical window of 24 hours, 48 ​​hours, or 72 hours as the input length, and directly uses the time-series feature sequence processed by the feature engineering module through feature association as the input data.

[0040] The LSTM layer receives continuous feature sequences and uses a gating mechanism to capture the accompanying patterns between blood glucose fluctuations and event markers within local time steps, determining which data to retain and which to forget. The number of layers can be set to two to increase depth and enhance the model's expressive power, capturing more complex patterns. Two layers meet functional requirements while preventing overfitting. The number of hidden units can be 128-dimensional to balance model capacity and computational cost. Dropout can be set to 0.2 to introduce slight dropout between layers, preventing overlearning on the training data.

[0041] The Transformer encoder layer takes the hidden state sequence at each time point output by the LSTM layer as input to the Transformer encoder, uses the Transformer's self-attention mechanism to calculate the relationship score between any two time points in the sequence, and integrates global information from the entire input length window to provide rich holistic data for prediction.

[0042] The output layer passes the final hidden state output by the Transformer through a fully connected network and outputs a tensor sequence of blood glucose values ​​at specific points in time during the target prediction period. Specifically, the probability of abnormal blood glucose levels within the prediction period can also be used as an auxiliary output.

[0043] For example, taking a user at 20 weeks of pregnancy as an example, the prediction period is the next 72 hours. First, the cleaned and feature-extracted data of the user in the most recent 24 hours is selected as the model input, including feature vectors of 288 time points. Each feature vector contains CGM mean, CGM standard deviation, post-meal marker, exercise marker, sleep marker, etc. For example, at lunchtime the previous day, the CGM mean in the features is higher, and the post-meal marker is 1; when walking in the evening, the exercise marker is 1, and the CGM slope is negative. During model inference, the LSTM layer first processes the sequence of these 288 time steps and identifies that "post-meal marker = 1" is usually accompanied by an increasing CGM slope, while "exercise marker = 1" is often followed by a decreasing CGM slope. The Transformer layer analyzes the entire 24-hour sequence. Its attention mechanism finds that when the sleep quality corresponding to the "sleep marker" is poor, the CGM mean of the next morning is usually higher, thus giving higher weights to these long-range dependencies across time. The final model generates a blood glucose prediction curve for the next 72 hours. The curve shows that based on the user's regular meal and exercise habits, blood glucose will show regular peaks after meals, all within the control range, and blood glucose will be stable at night.

[0044] In one embodiment, the user blood glucose profile generation module uses K-Means++ clustering to group users with similar blood glucose metabolism patterns into one category and assigns a clinically meaningful label to each group. The features used for clustering may include the area under the blood glucose curve (AUC), standard deviation of blood glucose fluctuation, average blood glucose, time of postprandial blood glucose peak, nighttime blood glucose trough, and insulin resistance index. The AUC, standard deviation of blood glucose fluctuation, and average blood glucose can be obtained directly from CGM data or calculated. The time of postprandial peak can be determined for each meal by locking CGM data from 0-3 hours post-meal, finding the highest blood glucose point within that period, calculating the time difference between that point and the meal time, and taking the average over multiple days as the user's feature. The nighttime blood glucose trough is determined during the sleep period provided by the user's sleep data, finding the lowest blood glucose value, and taking the average over multiple days or the quantile of the lowest value. The insulin resistance index is calculated using the insulin resistance index assessed by a homeostasis model.

[0045] In this embodiment, the number of clusters is determined by combining the elbow rule with silhouette coefficient analysis. The exploration range of k is 2-10. The execution logic for determining the number of clusters is as follows: the value of k is increased from 2 to 10, clustering is performed respectively, and the sum of squares within each cluster corresponding to each k value is calculated; the relationship between k value and SSE is plotted, and the point where the decrease in SSE suddenly slows down is selected as the candidate k value; at the same time, the average silhouette coefficient of each k value is calculated, and the closer it is to 1, the better the clustering effect; finally, a k value with low SSE, high silhouette coefficient, and clinical interpretability is selected.

[0046] In this embodiment, the execution logic of K-Means++ clustering is as follows: First, randomly select the first centroid; second, for each data point, calculate its shortest distance to existing centroids; the larger the distance value, the greater the probability of being selected as the next centroid; third, repeat the second step until k centroids are selected; fourth, iterate using the standard K-Means algorithm until the centroids are stable. After clustering is completed, analyze the centroids of each cluster and assign clinical labels to them based on the representative characteristics of each cluster.

[0047] In this embodiment, the label type may include postprandial high glucose type, fasting high glucose type, dawn phenomenon type, hypoglycemia risk type, GI sensitive type, and exercise high response type.

[0048] For example, the fasting blood glucose of the postprandial hyperglycemic group may be normal, but blood glucose spikes after eating and falls slowly. They are very sensitive to carbohydrates, especially high-GI foods, and require strict meal separation and postprandial exercise management. Their determination can be based on the fact that the predicted post-breakfast blood glucose peak and the average daily hyperglycemic duration of users in the cluster are significantly higher than those in other clusters.

[0049] Even without eating, the fasting hyperglycemic group has a high basal blood glucose level. Intervention should focus on overall lifestyle and possible basal insulin therapy. The diagnosis can be based on high average blood glucose and insulin resistance index, but the predicted post-breakfast blood glucose peak is relatively not prominent, and the nighttime blood glucose trough is also high.

[0050] In the dawn phenomenon group, blood sugar levels begin to rise naturally in the early morning, leading to elevated fasting blood sugar levels in the morning. This is related to the circadian rhythm of hormones such as cortisol. A small dose of insulin before bed or adjusting the time of dinner may help. The determination of this group can be based on the additional calculation of the steep rise in blood sugar in the morning, in addition to the fasting hyperglycemia group.

[0051] People at risk of hypoglycemia have unstable blood sugar levels and often experience hypoglycemia before meals or at night. This may be due to irregular eating habits or inappropriate medication dosages, requiring adjustments to their diet and monitoring frequency. Their condition can be determined based on low nighttime blood sugar troughs, which predict a higher frequency of future hypoglycemic events and a potentially larger blood sugar standard deviation.

[0052] People with high glycemic index (GI) experience a much faster and higher blood sugar spike than others after eating the same food. They react strongly to refined rice and flour, sugary drinks, etc., and must strictly choose low-GI foods. Their condition is similar to that of people with high postprandial blood sugar, but they pay more attention to the average time to the peak blood sugar level after a meal.

[0053] The high-response exercise group shows a significant blood sugar reduction effect after moderate-intensity exercise, making them an ideal group for lifestyle intervention. They should be encouraged to establish regular exercise habits. To determine their suitability, a derived feature needs to be created – the rate of blood sugar reduction after exercise. This feature value is significantly positive for this group of users, and the predicted average daily duration of hyperglycemia is relatively low.

[0054] In particular, additional label types such as blood sugar stable type, stress-related type, and sleep-related type can be added based on the correlation or derivative analysis between feature vectors.

[0055] Furthermore, the blood glucose fluctuation prediction results are categorized into three levels of abnormality: severe, moderate, and mild. Severe abnormalities, such as predicted blood glucose values ​​reaching or exceeding clinically set absolute danger thresholds, require the user to seek emergency medical attention. Moderate abnormalities, such as predicted blood glucose levels indicating complex situations requiring professional judgment and involving medication use, are addressed by the intervention recommendation module, which generates an intervention draft and submits it to the doctor for review before providing feedback to the user. Mild abnormalities are predicted short-term blood glucose fluctuations that can be corrected through lifestyle interventions; the intervention recommendation module automatically generates an intervention recommendation plan and provides it to the user.

[0056] In this embodiment, the general intervention recommendation scheme of the intervention recommendation module is based on the "IF-THEN" rule base preset of evidence-based medicine. When the intervention recommendation module is executed, it matches a rough, multi-option general scheme framework from the medical knowledge base based on abnormal blood glucose fluctuation data; then, it filters, strengthens, or specifies the items in the general scheme according to the user's tag type to optimize the scheme.

[0057] The optimization of the plan follows the priority principle of effectiveness before safety before compliance. The intervention recommendation module will first present the most effective and safest intervention options, and then put the options with higher compliance relatively forward based on user profiles and historical preferences to encourage users to implement the intervention.

[0058] As one implementation, this application also provides a method for dynamic blood glucose prediction. This method is used to implement a pregnancy blood glucose fluctuation prediction and intervention recommendation system based on multi-source time-series data as described in the foregoing embodiments. The method specifically includes the following steps: S1. Selective screening for GDM risk factors before 15 weeks of gestation: Users undergo selective screening using fasting blood glucose, glycated hemoglobin, or random blood glucose to exclude users who already had diabetes before pregnancy, and to classify non-diabetic users into normal pregnant women and high-risk pregnant women.

[0059] S2, Weeks 15-23 of pregnancy, mid-pregnancy risk warning period: During the risk warning period, the metabolic characteristics of pregnant women are relatively stable. The main approach is to monitor blood glucose response profiles to identify potential GDM risks in advance.

[0060] For the first time a pregnant woman uses the system to predict blood glucose fluctuations, the prediction cycle is 3 days. Routine finger prick blood tests are performed 2 days a week, once a day. System updates are performed at 15 weeks and 20 weeks of pregnancy. During system updates, CGM tests are performed on the user's blood glucose fluctuations over the past 14 days.

[0061] For high-risk pregnant women using the system for the first time to predict blood glucose fluctuations, the single prediction cycle is 3 days. Routine finger prick blood tests are performed 2 days a week, twice a day. System updates are performed at 15 weeks, 19 weeks and 23 weeks of pregnancy. During system updates, CGM tests are performed on the user's blood glucose fluctuations over the past 14 days.

[0062] S3, 24-28 weeks of gestation, mid-pregnancy screening and diagnosis period: Insulin resistance first appears between 24 and 28 weeks of gestation, with blood glucose fluctuations becoming more gradual. At 24 weeks of gestation, all pregnant women undergo an oral glucose tolerance test (OGTT) to further categorize them into normal pregnant women, high-risk pregnant women without a confirmed diagnosis, and pregnant women diagnosed with GDM.

[0063] For the average pregnant woman, the single-round prediction cycle of the system is 2 days, and routine finger-prick blood tests are performed 2 days a week, twice a day. The system is updated at 24 weeks and 27 weeks of pregnancy, and the CGM test is performed on the user's blood glucose fluctuations over the past 14 days at the time of the system update.

[0064] For pregnant women who have not been diagnosed with high risk, the single-round prediction cycle of the system is 2 days. Routine finger prick blood tests are performed 2 days a week, 3 to 4 times a day. The system is updated at 26 weeks of pregnancy, and CGM tests are performed on the user's blood glucose fluctuations over the past 14 days at the time of the system update.

[0065] Pregnant women diagnosed with GDM should stop using the system for prediction and follow routine monitoring methods as required by the diagnosis and treatment plan. After diagnosis, blood should be drawn every 2-4 weeks to check glycated hemoglobin.

[0066] S4. Weeks 29-36 of pregnancy, peak resistance intervention period in late pregnancy: During the 29th to 36th week of pregnancy, placental lactogen increases, insulin resistance is strongest, and fetal growth accelerates, requiring strict monitoring and control of blood sugar fluctuations.

[0067] For the average pregnant woman, the single-round prediction cycle of the system is 2 days. Routine finger prick blood tests are performed 3 to 4 days a week, 3 times a day. The system is updated at 30 weeks, 33 weeks and 36 weeks of pregnancy. When the system is updated, CGM tests are performed on the user's blood glucose fluctuations over the past 14 days.

[0068] For pregnant women who have not been diagnosed with high risk, the single-round prediction cycle of the system is 1 to 2 days. Routine finger prick blood tests are performed 4 days a week, 3 to 4 times a day. The system is updated at 29 weeks, 32 weeks and 35 weeks of pregnancy. At the time of the system update, CGM tests are performed on the user's blood glucose fluctuations over the past 14 days.

[0069] S5. Week 37 of pregnancy - delivery, labor preparation period: The core of the labor preparation period is to ensure a safe delivery and reduce the risk of neonatal hypoglycemia. For both normal pregnant women and those not diagnosed with high-risk pregnancies, the single-round prediction cycle of the system is 1 day; the system is updated every 1.5 weeks, and during the update, CGM (Continuous Glucose Monitoring) measures the user's blood glucose fluctuations over the past 7 days.

[0070] Specifically, the frequency of routine finger-prick blood glucose tests at each stage of pregnancy is dynamically adjusted based on the error of the prediction cycle by each system. Specifically, four tests per day correspond to the user's fasting and three post-meal blood glucose levels; three tests per day correspond to the user's fasting and two post-meal blood glucose levels; two tests per day correspond to the user's fasting and one post-meal blood glucose level; and one test per day corresponds to the user's fasting blood glucose level. The specific adjustment rules are as follows: For pregnancies between 15 and 23 weeks, if the systematic prediction error is less than 0.3 mmol / L, routine finger-prick blood tests should be reduced by one day per week; if the systematic prediction error is greater than 0.8 mmol / L, routine finger-prick blood tests should be increased by one day per week. During 24-28 weeks of gestation, if the systematic prediction error is less than 0.5 mmol / L, routine finger prick blood tests should be reduced to twice a day; if the systematic prediction error is greater than 1.2 mmol / L, routine finger prick blood tests should be increased to four times a day. For pregnancies between 29 and 36 weeks, if the systematic prediction error is less than 0.4 mmol / L, routine finger prick blood tests should be increased to 4 times a day on one day of the week, and maintained at 3 times a day on the other testing days; if the systematic prediction error is greater than 1.0 mmol / L, routine finger prick blood tests should be maintained at 4 times a day on the testing days. During the 37th week of pregnancy to delivery, the system prediction error was greater than 0.8 mmol / L, so a bedtime blood glucose test was added.

[0071] Furthermore, when the average error over two consecutive prediction cycles is greater than or equal to 1.0 mmol / L, the user's blood glucose profile is corrected. A user's blood glucose profile is a condensation of individual characteristics; only when the error is persistent and consistent in direction does it indicate a mismatch between the profile and the current individual characteristics. An average error greater than or equal to 1.0 mmol / L exceeds clinically acceptable error; without correction, subsequent predictions will continue to deviate, increasing the risk of complications.

[0072] When the average error over three consecutive prediction cycles is greater than or equal to 1.2 mmol / L, the prediction model parameters are updated and the model is iterated. Updating the parameters of the LSTM-Transformer prediction model requires retraining some network layers; strictly setting the duration of three consecutive prediction cycles can avoid system instability caused by frequent iterations. A threshold of 1.2 mmol / L is needed to achieve more than 20% of the blood glucose control target; without model iteration, many user intervention plans may fail.

[0073] Furthermore, the system provides differentiated intervention recommendations based on the different abnormal blood glucose events predicted by the system at different stages of pregnancy. Abnormal blood glucose levels are defined based on the percentage of time blood glucose levels approaching or exceeding the threshold during the prediction period, and the magnitude of any single predicted blood glucose level exceeding the threshold; either condition triggers an intervention or verification action.

[0074] Specifically, the abnormality criteria and triggering standards for each stage of pregnancy are as follows: Between 15 and 23 weeks of gestation, if the predicted data stays close to the target threshold for more than 30% of the time, or if the predicted data exceeds the threshold by more than 1.0 mmol / L, the risk is low in the first and second trimesters, and the criteria for defining the proportion are lenient, so excessive intervention should be avoided. Verification can be done by adding one finger prick blood test, or dietary intervention can be adjusted.

[0075] Between 24 and 28 weeks of gestation, if the predicted data remains close to the target threshold for more than 15% of the time, or if the predicted data exceeds the threshold by more than 0.8 mmol / L, this is a critical period for GDM diagnosis. Strict criteria are used to ensure precise intervention. Verification can be done by adding two finger-prick blood tests on the same day or one finger-prick blood test on each of two consecutive days. Intervention primarily focuses on dietary and exercise guidance; insulin assessment may be initiated for high-risk individuals.

[0076] Between 29 and 36 weeks of gestation, if the predicted blood glucose level is close to the target threshold for more than 10% of the time, or exceeds the threshold by more than 0.5 mmol / L, this stage carries a high risk of adverse glucose spikes and has the strictest abnormality criteria to avoid the impact of high glucose on the fetus. If an abnormality is identified, an additional finger-prick blood test should be performed immediately for verification. Intervention should primarily involve diet or insulin.

[0077] From 37 weeks of gestation to delivery, if the predicted blood glucose level is close to the target threshold for more than 5% of the time, or exceeds the threshold by more than 0.3 mmol / L, the delivery preparation stage prioritizes delivery safety, maintaining a strict abnormality definition ratio to address unforeseen risks. If an abnormality is defined, a finger prick blood test can be performed every 2 hours for three consecutive times to confirm. If pre-delivery blood glucose is high, intravenous insulin intervention or adjustments to the delivery plan may be considered.

[0078] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A system for predicting and recommending intervention for gestational glucose fluctuation based on multi-source time series data, characterized in that, The system comprises a data acquisition layer, a data processing layer, an algorithm calculation layer and an application interaction layer; The data acquisition layer comprises a CGM access module, a blood glucose meter access module, a smart wearable device access module and an EHR / EMR access module; the blood glucose meter access module and the CGM access module collect point-line data of blood glucose of a user respectively; the smart wearable device access module obtains corresponding life behavior data of the user within a blood glucose monitoring period; and the EHR / EMR access module imports medical physiological data of the user, thereby forming access of multi-source data of the user; The data processing layer comprises a data cleaning module, a time series database and a feature engineering module; the data cleaning module performs outlier rejection and interpolation completion on multi-source data provided by the data acquisition layer; the time series database provides storage space for the multi-source data; and the feature engineering module extracts time sequence features from the multi-source data and associates multi-source data features to capture their joint influence on blood glucose; The algorithm calculation layer comprises a blood glucose prediction module, a user blood glucose portrait generation module and an intervention recommendation module; the blood glucose prediction module predicts blood glucose fluctuation of a user in a specified future period based on time sequence association through an LSTM-Transformer hybrid model; the user blood glucose portrait generation module divides users into groups based on blood glucose metabolism mode through K-Means++ clustering, thereby forming a portrait label of the user; and the intervention recommendation module responds according to the blood glucose fluctuation prediction result, and dynamically optimizes a general intervention recommendation scheme in combination with the portrait label of the user; The application interaction layer is an operation and feedback entry for doctors and users, and comprises a doctor end, a user end, a warning module and an interaction module.

2. The multi-source time series data-based blood glucose fluctuation prediction and intervention recommendation system for pregnancy according to claim 1, characterized in that The system uses a periodic updating and calibration prediction mode; when the system is used for the first time and is updated each time, the CGM detects blood glucose fluctuation of the user for 14 days; each update supports short-term prediction for several rounds; wherein the CGM collection interval is 1 time / 5 minutes; The finger blood glucose meter detects blood glucose data 1 to 3 days before the target prediction period when the system is used for the first time; and in subsequent periods, data calibration is performed through the finger blood glucose meter detection the day before each prediction; the frequency of single-day detection of the finger blood glucose meter is 3 to 4 times, corresponding to the fasting and post-meal state of the user; The CGM access module mainly provides minute-level continuous blood glucose data and derived features detected by the CGM, and the collected features involve minute-level blood glucose values, blood glucose collection time stamps corresponding to the blood glucose values, postprandial peak features, blood glucose change rates, periodic fluctuation features and abnormal event markers; The blood glucose meter access module mainly provides single-point accurate data of specific time points detected by the finger blood glucose meter, and calibrates and supplements the data of the CGM; The user life behavior features obtained by the smart wearable device access module involve heart rate, exercise data, sleep data and work and rest data; The user medical physiological characteristics imported by the EHR / EMR access module relate to maternal basic physiological data, metabolic function related indexes, and liver and kidney function indexes.

3. The multi-source time series data-based gestational blood glucose fluctuation prediction and intervention recommendation system according to claim 2, characterized in that, The data cleaning module mainly performs data cleaning on CGM and comparison and correction of finger blood and CGM data; the method of CGM data cleaning adopts linear interpolation and abnormal filtering based on clinical rules, and the comparison and correction of finger blood and CGM data is performed through time alignment and establishment of a linear calibration model; The feature engineering module adopts a 1-hour sliding window with a step of 5 minutes; within the 1-hour sliding window, statistical features, trend features and morphological features of CGM data are extracted, and continuous life behavior data is converted into discrete event markers, including eating events, exercise events and sleep events; The feature engineering module performs feature correlation and event correlation analysis of multi-source data based on a time window; for each event time point, the CGM feature vector and the event marker vector are spliced, and the spliced feature vector is directly used as the input of the prediction model.

4. The multi-source time series data-based gestational blood glucose fluctuation prediction and intervention recommendation system according to claim 3, characterized in that, The LSTM-Transformer model architecture of the blood glucose prediction module includes an input layer, an LSTM layer, a Transformer encoder layer and an output layer, and the LSTM-Transformer model is trained in an autoregressive manner; The input layer uses a fixed-length historical window of 24 hours, 48 hours or 72 hours as the input length, and directly uses the time series feature sequence processed by the feature correlation of the feature engineering module as the input data; The LSTM layer receives continuous feature sequences and captures the accompanying patterns between blood glucose fluctuations and event markers within a local time step through a gate control mechanism; The Transformer encoder layer takes the hidden state sequence of each time point output by the LSTM layer as the input of the Transformer encoder, calculates the relationship score between any two time points in the sequence using the self-attention mechanism of the Transformer, and integrates the global information of the entire input length window; The output layer passes the final hidden state output by the Transformer through a fully connected network; and outputs the blood glucose value sequence at the target prediction period as a tensor.

5. The multi-source time series data-based gestational blood glucose fluctuation prediction and intervention recommendation system according to claim 4, characterized in that, The user blood glucose profile generation module classifies users with similar blood glucose metabolism patterns into a class through K-Means++ clustering, and assigns a clinically meaningful label to each class of population; The elbow rule combined with the silhouette coefficient analysis is used to determine the number of clusters, and the exploration range of k is 2-10; the execution logic of K-Means++ clustering is as follows: The first step is to randomly select the first centroid; the second step is to calculate the shortest distance between each data point and the existing centroid, and the greater the distance value, the greater the probability of being selected as the next centroid; the third step is to repeat the second step until k centroids are selected; the fourth step is to use the standard K-Means algorithm for iteration until the centroid is stable; After clustering, analyze the centroid of each cluster, and assign a clinical label to each cluster according to its representative features; the label types include postprandial high sugar type, fasting high sugar type, dawn phenomenon type, hypoglycemia risk type, GI sensitivity type, and exercise high response type.

6. The multi-source time series data-based gestational blood glucose fluctuation prediction and intervention recommendation system according to claim 5, characterized in that, The blood glucose fluctuation prediction result is divided into three categories of severe abnormality, moderate abnormality and mild abnormality according to the degree of abnormality; the severe abnormality needs to guide the user to seek medical treatment urgently, and the moderate abnormality and the mild abnormality generate an intervention recommendation scheme by the intervention recommendation module, wherein the intervention scheme of the moderate abnormality needs to be submitted to the doctor end for review; The general intervention recommendation scheme of the intervention recommendation module is preset based on the evidence-based medicine "IF-THEN" rule base; when the intervention recommendation module is executed, a rough, multiple-item general scheme framework is matched from the medical knowledge base according to the abnormal data of blood glucose fluctuation; then, the items in the general scheme are filtered, strengthened or specified for scheme optimization according to the label type of the user; The scheme optimization follows the priority ordering principle of effectiveness first, safety second and compliance third.

7. A blood glucose dynamic prediction method characterized by, The method for implementing the multi-source time series data-based gestational blood glucose fluctuation prediction and intervention recommendation system according to any one of claims 1 to 6 comprises the following processes: S1, selective screening of GDM risk factors before 15 weeks of pregnancy: Users perform selective screening through fasting blood glucose, glycosylated hemoglobin or random blood glucose, exclude users who have had diabetes before pregnancy, and divide non-diabetic users into ordinary pregnant women and high-risk pregnant women; S2, risk warning period in the second trimester of pregnancy (15-23 weeks of pregnancy): Ordinary pregnant women use the system for the first time to predict blood glucose fluctuation, with a single round prediction period of 3 days, regular finger blood test 2 days a week, and 1 time a day; system update is performed at 15 weeks of pregnancy and 20 weeks of pregnancy, and the system updates the blood glucose fluctuation of the user history of 14 days detected by CGM; High-risk pregnant women use the system for the first time to predict blood glucose fluctuation, with a single round prediction period of 3 days, regular finger blood test 2 days a week, and 2 times a day; system update is performed at 15 weeks of pregnancy, 19 weeks of pregnancy and 23 weeks of pregnancy, and the system updates the blood glucose fluctuation of the user history of 14 days detected by CGM; S3, screening and diagnosis period in the second trimester of pregnancy (24-28 weeks of pregnancy): At 24 weeks of pregnancy, all pregnant women perform OGTT screening to further divide the users into ordinary pregnant women, high-risk pregnant women who are not diagnosed, and GDM diagnosed pregnant women; Ordinary pregnant women use the system with a single round prediction period of 2 days, regular finger blood test 2 days a week, and 2 times a day; system update is performed at 24 weeks of pregnancy and 27 weeks of pregnancy, and the system updates the blood glucose fluctuation of the user history of 14 days detected by CGM; The single round prediction cycle of the system for pregnant women with undiagnosed high risk is 2 days, and the conventional finger blood test is performed 3 to 4 times a day for 2 days a week; the system is updated at 26 weeks of pregnancy, and the CGM detects the blood glucose fluctuation of the user history for 14 days at the time of system update; The pregnant women with diagnosed GDM stop using the system for prediction, and follow the conventional monitoring mode according to the requirements of diagnosis and treatment; S4, 29-36 weeks of pregnancy, intervention period of resistance peak in late pregnancy: The single round prediction cycle of the system for ordinary pregnant women is 2 days, and the conventional finger blood test is performed 3 times a day for 3 to 4 days a week; the system is updated at 30 weeks of pregnancy, 33 weeks of pregnancy and 36 weeks of pregnancy, and the CGM detects the blood glucose fluctuation of the user history for 14 days at the time of system update; The single round prediction cycle of the system for pregnant women with undiagnosed high risk is 1 to 2 days, and the conventional finger blood test is performed 3 to 4 times a day for 4 days a week; the system is updated at 29 weeks of pregnancy, 32 weeks of pregnancy and 35 weeks of pregnancy, and the CGM detects the blood glucose fluctuation of the user history for 14 days at the time of system update; S5, 37 weeks of pregnancy-delivery, preparation period for delivery: The single round prediction cycle of the system for ordinary pregnant women and pregnant women with undiagnosed high risk is 1 day; the system is updated every 1.5 weeks, and the CGM detects the blood glucose fluctuation of the user history for 7 days at the time of system update.

8. The blood glucose dynamics prediction method according to claim 7, characterized by, The frequency of conventional finger blood test in each stage of pregnancy is dynamically adjusted according to the error of each system prediction cycle, wherein 4 times of detection a day corresponds to the blood glucose of the user after fasting and three meals, 3 times of detection a day corresponds to the blood glucose of the user after fasting and two meals, 2 times of detection a day corresponds to the blood glucose of the user after fasting and one meal, and 1 time of detection a day corresponds to the fasting blood glucose of the user; the specific adjustment rules are as follows: During 15-23 weeks of pregnancy, if the system prediction error is less than 0.3 mmol / L, the conventional finger blood test is reduced by 1 day a week; if the system prediction error is greater than 0.8 mmol / L, the conventional finger blood test is increased by 1 day a week; During 24-28 weeks of pregnancy, if the system prediction error is less than 0.5 mmol / L, the conventional finger blood test is reduced to 2 times a day; if the system prediction error is greater than 1.2 mmol / L, the conventional finger blood test is increased to 4 times a day; During 29-36 weeks of pregnancy, if the system prediction error is less than 0.4 mmol / L, the conventional finger blood test is increased to 4 times a day on 1 day a week, and the remaining detection days are maintained at 3 times a day; if the system prediction error is greater than 1.0 mmol / L, the conventional finger blood test is maintained at 4 times a day on the detection day; During 37 weeks of pregnancy-delivery, if the system prediction error is greater than 0.8 mmol / L, the pre-sleep blood glucose is added.

9. The blood glucose dynamic prediction method according to claim 8, characterized in that, when the cycle average error of the system for 2 consecutive prediction cycles is greater than or equal to 1.0 mmol / L, the user's blood glucose profile is corrected; when the cycle average error of the system for 3 consecutive prediction cycles is greater than or equal to 1.2 mmol / L, the parameters of the prediction model are updated and the model is iterated.

10. The blood glucose dynamic prediction method according to claim 7, characterized in that, The system makes differentiated intervention recommendations for different blood glucose abnormal events that occur in the prediction results of different pregnancy stages; the blood glucose abnormality is determined based on the threshold proximity ratio and the threshold amplitude, and either one triggers intervention or verification action; The abnormality definition ratio and trigger standard for each pregnancy stage are as follows: For 15-23 weeks of pregnancy, the duration of predicted data close to the stage target threshold is more than 30%, or the amplitude of predicted data exceeding the threshold is greater than 1.0 mmol / L; For 24-28 weeks of pregnancy, the duration of predicted data close to the stage target threshold is more than 15%, or the amplitude of predicted data exceeding the threshold is greater than 0.8 mmol / L; For 29-36 weeks of pregnancy, the duration of predicted data close to the stage target threshold is more than 10%, or the amplitude of predicted data exceeding the threshold is greater than 0.5 mmol / L; For 37 weeks of pregnancy to delivery, the duration of predicted data close to the stage target threshold is more than 5%, or the amplitude of predicted data exceeding the threshold is greater than 0.3 mmol / L.

Images