Diabetes early subtype recognition and risk prediction system based on continuous blood glucose monitoring
By identifying metabolic subtypes through DTW and K-shape clustering, and combining glucose-insulin kinetics models and LSTM networks, the accuracy and interpretability issues in early diabetes subtype identification and risk prediction are solved, enabling high-precision and personalized blood glucose management and intervention.
Patent Information
- Application Number
- CN202610015363.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-07
- Publication Date
- 2026-04-07
AI Technical Summary
In the early identification of diabetes subtypes and risk prediction, existing technologies struggle to capture blood glucose fluctuations using static indicators, lack explanations based on physiological mechanisms, and single models cannot adequately account for both physiological mechanisms and short-term behavioral fluctuations, resulting in insufficient accuracy in identification and prediction.
The DTW algorithm is used to extract typical daily curves, and K-shape clustering is used to identify metabolic subtypes. A glucose-insulin dynamics model is constructed, and the residuals are learned through an LSTM network. The model output is dynamically fused to achieve personalized blood glucose prediction.
It improves the personalization and accuracy of blood glucose management, enhances the clinical reliability and real-time robustness of predictions, and provides physiological explanations and individualized intervention recommendations.
Smart Images
Figure CN121812167A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical testing technology, particularly the field of dynamic blood glucose monitoring, identification and analysis, and relates to a system for early subtype identification and risk prediction of diabetes based on continuous blood glucose monitoring. Background Technology
[0002] Diabetes is a global public health problem. Traditional diagnosis and detection mainly rely on static biochemical indicators such as fasting plasma glucose (FPG) and glycated hemoglobin (HbA1c). In recent years, the widespread application of continuous glucose monitoring (CGM) technology has enabled researchers to obtain high temporal resolution dynamic glucose data, which is crucial for in-depth analysis of the duration of postprandial hyperglycemia, glycemic variability, and individual heterogeneity.
[0003] Existing research mainly employs the following methods: 1) Pattern recognition methods: using time-series clustering algorithms (such as K-means, spectral clustering, etc.) to identify different blood glucose fluctuation patterns or metabolic subtypes. 2) Predictive modeling methods: using statistical modeling methods (such as ARIMA, etc.) or machine learning and deep learning algorithms (such as support vector machines, random forests, LSTM, etc.) to predict blood glucose levels. 3) Mechanism explanation methods: constructing physiological dynamic models (such as Minimal Model, OGTT Minimal Model, etc.) to characterize the glucose-insulin regulation mechanism using mathematical equations, extracting physiologically significant parameters such as insulin sensitivity, β-cell function, and glucose clearance rate. Currently, the main advantages of these existing technologies in the field of blood glucose detection and analysis include: strong detection continuity: CGM technology can achieve dynamic recording, providing high temporal resolution data for studying glucose metabolism processes. Improved diagnostic accuracy: abnormal metabolic states can be identified through dynamic blood glucose indicators (such as target range time percentage, blood glucose fluctuation coefficient, etc.). A certain degree of mechanistic rationality: dynamic models can explain the influence of relevant factors on blood glucose regulation from a physiological mechanism perspective.
[0004] Despite the progress and applications of current research in dynamic blood glucose analysis, the following limitations and shortcomings still exist, affecting the accurate identification of individual blood glucose subtypes and the high-precision prediction of blood glucose levels, especially for high-risk groups in the early stages of diabetes: 1) Static indicators are insufficient to fully capture the characteristics of blood glucose fluctuations and their risks. Static indicators such as FPG and HbA1c cannot reflect the dynamic changes in blood glucose and individual heterogeneity, and are not sensitive enough to early metabolic abnormalities, making it difficult to detect hidden blood glucose metabolism abnormalities in the early stages of diabetes.
[0005] 2) Lack of reliable physiological mechanism explanation. Although black box models such as deep learning can improve prediction accuracy to some extent, their internal features are difficult to interpret, which is not conducive to clinical application and physiological interpretation. They are prone to overfitting, resulting in poor extrapolation of predictions and difficulty in reflecting individual blood glucose fluctuation patterns.
[0006] 3) Poor ability to identify prediabetes subtypes. Blood glucose fluctuates dynamically on a minute-by-minute basis, exhibiting different response curves under the stimulation of diet, activity, and emotions, reflecting an individual's ability to regulate blood glucose. However, existing clustering methods mostly rely on static indicators, such as average blood glucose and blood glucose variability index, and do not accurately identify the shape and temporal characteristics of blood glucose curves from a dynamic perspective. This leads to insufficient stability in subtype classification, and the identified blood glucose fluctuation patterns lack effective correlation with physiological mechanisms (such as insulin sensitivity), making it difficult to characterize temporal alignment issues.
[0007] In summary, existing technologies have not yet formed a unified analytical framework that combines dynamic identification of prediabetes subtypes, physiological interpretability, and high predictive performance, thereby enabling personalized risk stratification and assessment of prediabetes high-risk populations. Summary of the Invention
[0008] In view of this, the purpose of the present invention is to provide a system for identifying and predicting the early subtypes of diabetes based on continuous glucose monitoring, which can identify the subtypes of prediabetes and provide a highly accurate and interpretable predictive analysis method for different subtypes, thereby realizing individualized dynamic pattern recognition and hybrid prediction based on continuous glucose monitoring (CGM) data.
[0009] To achieve the above objectives, the present invention provides the following technical solution: A system for early identification and risk prediction of diabetes based on continuous blood glucose monitoring, comprising a data preprocessing module, a subtype identification module, a physiological modeling module, a residual prediction module, and a fusion prediction module; The data preprocessing module is used to interpolate and standardize continuous blood glucose monitoring data; The subtype identification module is used to extract typical daily curves and perform cluster analysis to output metabolic subtype labels; The physiological modeling module is used to construct and fit a glucose-insulin dynamics model and output personalized physiological parameters. The residual prediction module is used to train the LSTM network and predict the residuals of the dynamic model; The fusion prediction module is used to dynamically fuse the outputs of two types of models to generate the final blood glucose prediction result.
[0010] Furthermore, the subtype identification module integrates dynamic time warping algorithm and K-shape clustering algorithm, and supports adaptive selection of the number of clusters.
[0011] Furthermore, the system also includes a user interface for displaying blood glucose prediction curves, subtype classification results, physiological parameter reports, and intervention suggestions.
[0012] Furthermore, the system performs the following steps: S1. Data acquisition and preprocessing: acquire continuous blood glucose monitoring time series data, and perform missing value imputation and standardization. S2. Dynamic feature extraction and clustering: Extract typical daily blood glucose curves of individuals based on dynamic time warping algorithm, and identify metabolic subtypes through K-shape clustering algorithm; S3. Physiological modeling and parameter fitting: Based on the glucose-insulin interaction kinetic model, personalized physiological parameters are fitted by combining individual historical data. S4. Residual learning and nonlinear compensation: Long short-term memory network is used to train and predict the prediction residuals of the dynamic model. S5. Dynamic fusion and prediction output: Based on the output of the dynamic fusion model and the residual prediction model of the real-time prediction error, the final blood glucose prediction result is obtained.
[0013] Furthermore, in step S1, the sampling interval for continuous blood glucose monitoring data is 1-5 minutes, missing values are filled using spline interpolation or Kalman filtering, and the standardization process uses the Z-score normalization method.
[0014] Furthermore, step S2 mainly includes two parts: typical daily extraction and metabolic subtype classification, specifically including: Typical Daily Extraction: The Dynamic Time Warping (DTW) algorithm is used to extract the most representative 24-hour blood glucose pattern for each individual, i.e., the typical daily curve. The DTW algorithm is used to calculate blood glucose curves for the same subject on different dates. and Temporal similarity between ;Typical day The day identified as the center day is the one with the smallest mean DTW distance from all other valid daily curves; time series similarity. The calculation method is as follows: (8) The method for determining a typical daily curve is as follows: (9) in Represents the physical condition of those participating in continuous blood glucose testing. Representative of individuals The Daily blood sugar levels The number of days used to obtain blood glucose data; Metabolic subtype classification: based on extracted, individual-representative 24-hour blood glucose patterns. K-shape clustering algorithm was used to identify metabolic subtypes corresponding to stages of diabetes development; two normalized daily blood glucose curves of typical individuals were used. and The SBD between them is defined as follows: (10) in, It is a lag quantity Normalized cross-correlation at the location, It is a time series of Norm.
[0015] Furthermore, the K-shape clustering algorithm uses shape-related distance as a similarity measure, with the number of clusters K ranging from 2 to 5. The optimal number of clusters is determined by the silhouette coefficient, Calinski-Harabasz index, or Davies-Bouldin index.
[0016] Furthermore, step S3 specifically includes: establishing a glucose-insulin interaction kinetic model, which is based on the Bergman minimal model form: (11) In the model, G(t) represents blood glucose concentration, X(t) represents the dynamic effect of insulin, I(t) represents insulin concentration, and D(t) represents dietary glucose intake. For individualized physiological parameters, This is the insulin-independent glucose clearance rate (range [0.017, 0.100]). This is the rate of insulin action and metabolism (value range [0.050, 0.200]). This is the sensitivity to the dynamic action of insulin (value range [0.1, 12.0] × 10⁻¹⁰). -12 The model parameters are obtained using individual historical data through least squares fitting or Bayesian estimation methods; the calculation method for D(t) is as follows: (12) Where N represents the total number of meals, This indicates the amount of nutrients consumed at each meal. It is the carbohydrate and protein absorption coefficient that inhibits fat absorption. It is the gamma probability density function, and the specific calculation method is as follows: (13) Its shape parameters area The scale parameter is s. For the first The timing of each meal.
[0017] Further, in step S4, a Long Short-Term Memory (LSTM) neural network is constructed to specifically learn the predicted residual sequence. The LSTM network includes an input layer, two hidden LSTM layers, and an output layer. The input consists of the blood glucose, insulin, and prediction error sequences from previous time steps, and the output is the predicted residual value. During training, the loss function is the mean squared error (MSE) form, and the preferred optimization algorithm is the Adam adaptive optimizer, with a learning rate range of [range missing]. The residual prediction model obtained through training achieves nonlinear compensation for the error of the dynamic model, thereby improving the overall prediction accuracy.
[0018] Furthermore, in step S5, the prediction results of the glucose-insulin kinetic model and the LSTM residual learning model are combined, and the fusion weights are dynamically adjusted based on the real-time prediction error to achieve an adaptive balance between physiological mechanisms and nonlinear perturbations; at time... Blood glucose prediction values output by the hybrid prediction model The expression is (14) in: The blood glucose value is predicted by the glucose-insulin kinetic model; The residual compensation value is obtained from the LSTM residual prediction model; These are dynamic weighting coefficients used to adaptively weight the predictions from two models.
[0019] The beneficial effects of this invention are as follows: The beneficial effects of this invention are mainly reflected in the following aspects: 1) This solution can dynamically measure blood glucose fluctuations and identify prediabetes subtypes based on these fluctuations. Compared to previous patents, by using DTW to extract typical days and K-shape clustering, it can better achieve accurate identification of individualized metabolic subtypes using a small amount of blood glucose monitoring data. This solves the problem of insufficient blood glucose dynamic feature identification capability in existing technologies and improves the personalization and accuracy of blood glucose management.
[0020] 2) This invention combines physiological dynamics modeling (S3) and deep residual learning (S4) modules, effectively solving the dilemma in medicine where good predictions lack physiological basis and physiological models have low prediction accuracy. It realizes a unified hybrid analysis system with high accuracy and interpretability, solving the problems of poor model interpretability and limited prediction performance in the prior art, and increasing the clinical reliability and guidance value of the results.
[0021] 3) The dynamic adaptive fusion mechanism (S5) of this invention adjusts the model fusion weights based on the real-time prediction error. This system achieves a comprehensive and accurate capture of physiological mechanisms and nonlinear fluctuations, solving the problem that single models struggle to simultaneously account for physiological mechanisms and short-term behavioral fluctuations, thus improving the real-time robustness and accuracy of predictions. Simultaneously, the system can quantify the effects of individual physiological characteristics, addressing the difficulty of traditional prediction models in providing physiological explanations, and enhancing physicians' decision support capabilities for developing personalized intervention plans for high-risk prediabetes individuals based on their metabolic status.
[0022] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 The overall technical approach for hybrid modeling of glucose metabolism subtype identification and glucose prediction; Figure 2 This is a roadmap for CGM-based subtype identification technology. Figure 3 A diagram illustrating the construction mechanism of a glucose-insulin hybrid model based on residual fusion and dynamic weights. Detailed Implementation
[0024] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.
[0025] like Figure 1 — Figure 3 As shown, Figure 1 A comprehensive technical roadmap for hybrid modeling of glucose metabolism subtype identification and glucose prediction. Figure 2 This is a roadmap for CGM-based subtype identification technology. Figure 3 This diagram illustrates the mechanism for constructing a glucose-insulin hybrid model based on residual fusion and dynamic weights. This invention provides a system for early diabetes subtype identification and risk prediction based on continuous glucose monitoring. The system includes a data preprocessing module, a subtype identification module, a physiological modeling module, a residual prediction module, and a fusion prediction module. The system executes the following steps: S1: Data acquisition and preprocessing; S2: Dynamic feature extraction and clustering; S3: Physiological modeling and parameter fitting; S4: Residual learning and nonlinear compensation; S5: Dynamic fusion and prediction output.
[0026] Specifically: In step S1, continuous glucose monitoring (CGM) time-series data are collected, and imputation processing is performed on missing values in the data to construct a continuous and complete data sequence. Subsequently, standardization processing is used to eliminate the dimensional differences between different indicators to ensure data consistency and comparability.
[0027] In step S2, the Dynamic Time Warping (DTW) algorithm is used to extract the typical day that best reflects the inherent blood glucose change pattern for different individuals' blood glucose time series data. Based on the extracted typical day features, the K-shape clustering algorithm is used to identify blood glucose dynamic subtypes with significant differences.
[0028] In step S3, a glucose-insulin interaction kinetic model is constructed, and personalized kinetic parameters that fit the individual's physiological characteristics are obtained by combining individual blood glucose and dietary monitoring data.
[0029] In step S4, a long short-term memory (LSTM) network is trained to perform specialized learning and accurate prediction of the prediction residuals of the glucose-insulin kinetic model, thereby compensating for the nonlinear error of the model.
[0030] In step S5, based on the real-time prediction error of the model, the fusion weights of the kinetic model and the LSTM residual prediction model are dynamically and adaptively adjusted to finally output accurate blood glucose prediction results and corresponding risk level prompts.
[0031] In this embodiment, the important parameters of the module are shown in the following table: In this embodiment, the technical solution specifically includes: 1. Data Acquisition and Preprocessing In this step, individual blood glucose time-series data are acquired from a continuous glucose monitoring device. The data records changes in blood glucose concentration in time series format, with a sampling interval of 3 minutes (this part is supported and guaranteed by mature and reliable equipment). For missing and outlier values in the original data, interpolation or sliding window methods are used for data repair and smoothing to ensure the continuity and integrity of the obtained sequence. Preferably, spline interpolation or Kalman filtering algorithms can be used for time series smoothing. Subsequently, the corrected blood glucose sequence is standardized using the Z-score normalization formula to eliminate dimensional differences between individuals, ensuring the stability and comparability of subsequent analyses.
[0032] 2. Dynamic Feature Extraction and Clustering The core of this step lies in eliminating data noise and phase bias through dynamic feature extraction to obtain the most representative blood glucose curves for each individual, enabling accurate subsequent clustering. This step includes two parts: extraction of typical daily data and classification of metabolic subtypes.
[0033] 2.1 Typical Daily Extraction In this embodiment, the Dynamic Time Warping (DTW) algorithm is used to extract the most representative 24-hour blood glucose pattern for each individual, i.e., the typical daily curve. The DTW algorithm is used to calculate blood glucose curves for the same subject on different dates. and Temporal similarity between DTW effectively eliminates curve shifts caused by factors such as irregular meal times by minimizing local phase mismatches by allowing time series to be warped and aligned along the time axis. To limit unrealistic warping, the Sakoe-Chiba window is preferred. Typical day The day identified as the center day is the one with the smallest mean DTW distance from all other day curves among all valid day curves. This curve preserves the typical diurnal structure of the subjects while suppressing the influence of anomalous days. (Temporal similarity) The calculation method is as follows: (15) The method for determining a typical daily curve is as follows: (16) in Represents the physical condition of those participating in continuous blood glucose testing. Representative of individuals The Daily blood sugar levels The number of days used to obtain blood glucose data.
[0034] 2.2 Metabolic Subtype Classification Based on extracted, individual-representative 24-hour blood glucose patterns This invention further employs the K-shape clustering algorithm to identify metabolic subtypes corresponding to stages of diabetes development. This method effectively handles temporal misalignment between meals, amplitude differences caused by baseline variations, and noise caused by sensor variations. This clustering method uses shape-related distance as the metric, rather than the traditional Euclidean distance. Two normalized daily blood glucose curves for typical individuals are shown. and The SBD between them is defined as follows: (17) in, It is a lag quantity Normalized cross-correlation at the location, It is a time series of Norm. SBD allows matching subjects whose overall waveforms are similar despite the presence of early or delayed peaks, demonstrating robustness to shape features.
[0035] The K-shape algorithm used in this invention exhibits robust performance in time series clustering. The cluster centers, through iterative updates, are determined as the first principal eigenvector of the cross-correlation matrix among member sequences, ensuring that the cluster centers possess a rotation-invariant average shape. The K-shape algorithm sets a convergence tolerance of [value missing]. In this embodiment, the number of clusters K was tested in the range of 2 to 5, and clustering effectiveness indicators such as silhouette coefficient, Calinski-Harabasz index and Davies-Bouldin index were used to evaluate the results in order to determine an optimal number of clusters that balances statistical optimality and physiological interpretability.
[0036] 3. Physiological modeling and parameter fitting In this step, a glucose-insulin interaction kinetic model is established to reveal the individual's metabolic mechanisms. This model is based on the Bergman minimal model form: (18) In the model, G(t) represents blood glucose concentration, X(t) represents the dynamic effect of insulin, I(t) represents insulin concentration, and D(t) represents dietary glucose intake. For individualized physiological parameters, This is the insulin-independent glucose clearance rate (range [0.017, 0.100]). This is the rate of insulin action and metabolism (value range [0.050, 0.200]). This is the sensitivity to the dynamic action of insulin (value range [0.1, 12.0] × 10⁻¹⁰). -12 The model parameters are obtained using individual historical data through least squares fitting or Bayesian estimation. Preferably, individual diet, exercise records, and CGM data can be combined for joint fitting to improve the model's individual adaptability. The calculation method for D(t) is as follows: (19) Where N represents the total number of meals, This indicates the amount of nutrients consumed at each meal. It is the carbohydrate and protein absorption coefficient that inhibits fat absorption. It is the gamma probability density function, and the specific calculation method is as follows: (20) Its shape parameter is a, and its scale parameter is s. For the first The timing of the next meal was determined. Model parameter estimation employed a nonlinear least squares method (Levenberg-Marquardt algorithm), with the objective function being "minimizing the RMSE between the predicted ODE and the actual CGM data," and the convergence criterion being iteration error < .
[0037] 4. Residual learning and nonlinear compensation In this step, to compensate for the shortcomings of the physiological dynamics model in representing high-dimensional nonlinear features, a Long Short-Term Memory (LSTM) neural network is constructed to specifically learn the predicted residual sequence. The LSTM network includes an input layer, two LSTM hidden layers, and an output layer. The input consists of the previous time-series blood glucose, insulin, and prediction error sequences, and the output is the predicted residual value. During training, the loss function is in the form of mean squared error (MSE), and the preferred optimization algorithm is the Adam adaptive optimizer, with a learning rate range of [range missing]. The residual prediction model obtained through training can achieve nonlinear compensation for the errors of the dynamic model, thereby improving the overall prediction accuracy.
[0038] 5. Dynamic Fusion and Predictive Output In this step, to further improve the real-time performance and robustness of the prediction results, a dynamic hybrid fusion strategy based on prediction error feedback is proposed. This strategy combines the prediction results of the glucose-insulin kinetic model and the LSTM residual learning model, and dynamically adjusts the fusion weights according to the real-time prediction error to achieve an adaptive balance between physiological mechanisms and nonlinear perturbations. At time t... Blood glucose prediction values output by the hybrid prediction model The expression is (twenty one) in: The blood glucose value is predicted by the glucose-insulin kinetic model; The residual compensation value is obtained from the LSTM residual prediction model; These are dynamic weighting coefficients used to adaptively weight the predictions from two models.
[0039] This invention belongs to the field of dynamic blood glucose monitoring, identification, and analysis. The resulting products or technologies can be widely applied in the following areas and occasions to achieve personalized and precise blood glucose management: (1) Identification of high-risk groups and intelligent blood glucose management agent Applications: Hospital endocrinology departments, diabetes clinics, health check-up centers, etc.
[0040] Functions: 1. Assisted diagnosis and subtype identification: Through Dynamic Time Warping (DTW) and K-shape clustering algorithms, metabolic subtypes with physiological differences are identified, helping doctors to make more accurate subtype diagnoses for patients. 2. Personalized parameter assessment: The physiological dynamics model (S3) outputs key physiological parameters such as insulin sensitivity and glucose clearance rate, providing doctors with quantitative evidence and physiological mechanism explanations for adjusting drug dosages or insulin treatment regimens. 3. Risk warning: Provides accurate predictions of hyperglycemia and hypoglycemia, along with risk level indicators, guiding healthcare professionals to intervene in real time.
[0041] (2) Mobile healthcare for individual blood glucose management and risk warning in diabetes patients Application scenarios: Smartphone applications (Apps), cloud-based health management platforms, and supporting systems for wearable CGM devices.
[0042] Functions: 1. High-precision real-time prediction: Provides users with accurate predictions of future blood glucose changes, helping them to anticipate fluctuations and improve the timeliness of self-management. 2. Lifestyle intervention: Based on individual subtype identification results, recommends personalized dietary advice, exercise programs, and lifestyle adjustments to users, improving the personalization and accuracy of blood glucose management. 3. Behavioral optimization: For example, using dynamic blood glucose monitoring (CGM) with a market value of approximately 300 as a data source, users can optimize meal times, exercise intensity, and insulin injection timing based on the prediction results provided by the system, achieving closed-loop blood glucose regulation and management.
[0043] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications should be covered within the scope of the claims of the present invention.
Claims
1. A system for early subtype identification and risk prediction of diabetes based on continuous blood glucose monitoring, characterized in that: The system includes a data preprocessing module, a subtype identification module, a physiological modeling module, a residual prediction module, and a fusion prediction module; The data preprocessing module is used to interpolate and standardize continuous blood glucose monitoring data; The subtype identification module is used to extract typical daily curves and perform cluster analysis to output metabolic subtype labels; The physiological modeling module is used to construct and fit a glucose-insulin dynamics model and output personalized physiological parameters. The residual prediction module is used to train the LSTM network and predict the residuals of the dynamic model; The fusion prediction module is used to dynamically fuse the outputs of two types of models to generate the final blood glucose prediction result.
2. The system for early subtype identification and risk prediction of diabetes based on continuous blood glucose monitoring according to claim 1, characterized in that: The subtype identification module integrates dynamic time warping algorithm and K-shape clustering algorithm, and supports adaptive selection of the number of clusters.
3. The system for early subtype identification and risk prediction of diabetes based on continuous blood glucose monitoring according to claim 2, characterized in that: The system also includes a user interface for displaying blood glucose prediction curves, subtype classification results, physiological parameter reports, and intervention suggestions.
4. The system for early subtype identification and risk prediction of diabetes based on continuous blood glucose monitoring according to claim 3, characterized in that: The system performs the following steps: S1. Data acquisition and preprocessing: acquire continuous blood glucose monitoring time series data, and perform missing value imputation and standardization. S2. Dynamic feature extraction and clustering: Extract typical daily blood glucose curves of individuals based on dynamic time warping algorithm, and identify metabolic subtypes through K-shape clustering algorithm; S3. Physiological modeling and parameter fitting: Based on the glucose-insulin interaction kinetic model, personalized physiological parameters are fitted by combining individual historical data. S4. Residual learning and nonlinear compensation: Long short-term memory network is used to train and predict the prediction residuals of the dynamic model. S5. Dynamic fusion and prediction output: Based on the output of the dynamic fusion model and the residual prediction model of the real-time prediction error, the final blood glucose prediction result is obtained.
5. The system for early subtype identification and risk prediction of diabetes based on continuous blood glucose monitoring according to claim 4, characterized in that: In step S1, the sampling interval for continuous blood glucose monitoring data is 1-5 minutes, missing values are filled using spline interpolation or Kalman filtering, and the standardization process uses the Z-score normalization method.
6. The system for early subtype identification and risk prediction of diabetes based on continuous blood glucose monitoring according to claim 5, characterized in that: Step S2 mainly includes two parts: typical daily extraction and metabolic subtype classification, specifically including: Typical Daily Extraction: The Dynamic Time Warping (DTW) algorithm is used to extract the most representative 24-hour blood glucose pattern for each individual, i.e., the typical daily curve. The DTW algorithm is used to calculate blood glucose curves for the same subject on different dates. and Temporal similarity between ;Typical day The day identified as the center day is the one with the smallest mean DTW distance from all other valid daily curves; time series similarity. The calculation method is as follows: (1) The method for determining a typical daily curve is as follows: (2) in Represents the physical condition of those participating in continuous blood glucose testing. Representative of individuals The Daily blood sugar levels The number of days used to obtain blood glucose data; Metabolic subtype classification: based on extracted, individual-representative 24-hour blood glucose patterns. The K-shape clustering algorithm was used to identify metabolic subtypes corresponding to the stages of diabetes development; two normalized typical daily blood glucose curves for individuals were used. and The SBD between them is defined as follows: (3) in, It is a lag quantity Normalized cross-correlation at the location, It is a time series of Norm.
7. A system for early subtype identification and risk prediction of diabetes based on continuous blood glucose monitoring according to claim 6, characterized in that: The K-shape clustering algorithm uses shape-related distance as a similarity measure, with the number of clusters K ranging from 2 to 5. The optimal number of clusters is determined by the silhouette coefficient, Calinski-Harabasz index, or Davies-Bouldin index.
8. A system for early subtype identification and risk prediction of diabetes based on continuous blood glucose monitoring according to claim 7, characterized in that: Step S3 specifically includes: establishing a glucose-insulin interaction kinetic model, which is based on the Bergman minimal model form. (4) In the model, G(t) represents blood glucose concentration, X(t) represents the dynamic effect of insulin, I(t) represents insulin concentration, and D(t) represents dietary glucose intake. For individualized physiological parameters, This is the insulin-independent glucose clearance rate, with values ranging from [0.017, 0.100]. This represents the metabolic rate of insulin action, with values ranging from [0.050, 0.200]. This represents the sensitivity to the dynamic action of insulin, with values ranging from [0.1, 12.0] × 10⁻⁶. -12 The model parameters were obtained using individual historical data through least squares fitting or Bayesian estimation methods; the calculation method for D(t) is as follows: (5) Where N represents the total number of meals, This indicates the amount of nutrients consumed at each meal. It is the carbohydrate and protein absorption coefficient that inhibits fat absorption. It is the gamma probability density function, and the specific calculation method is as follows: (6) Its shape parameter is a, and its scale parameter is s. For the first The timing of each meal.
9. A system for early subtype identification and risk prediction of diabetes based on continuous blood glucose monitoring according to claim 8, characterized in that: In step S4, a Long Short-Term Memory (LSTM) neural network is constructed to specifically learn the predicted residual sequence. The LSTM network includes an input layer, two hidden LSTM layers, and an output layer. The input consists of the previous blood glucose, insulin, and prediction error sequences, and the output is the predicted residual value. During training, the loss function is the mean squared error (MSE) form, and the preferred optimization algorithm is the Adam adaptive optimizer, with a learning rate range of [range missing]. The residual prediction model obtained through training achieves nonlinear compensation for the error of the dynamic model, thereby improving the overall prediction accuracy.
10. A system for early subtype identification and risk prediction of diabetes based on continuous blood glucose monitoring according to claim 9, characterized in that: In step S5, the prediction results of the glucose-insulin kinetic model and the LSTM residual learning model are combined, and the fusion weights are dynamically adjusted according to the real-time prediction error to achieve an adaptive balance between physiological mechanisms and nonlinear perturbations. At any moment Table of blood glucose prediction values output by the hybrid prediction model The expression is (7) in: The blood glucose value is predicted by the glucose-insulin kinetic model; The residual compensation value is obtained from the LSTM residual prediction model; These are dynamic weighting coefficients used to adaptively weight the predictions from two models.