A high-precision blood glucose monitoring method and system based on multi-modal data fusion

Abstract

The application relates to a high-precision blood glucose monitoring method and system based on multi-modal data fusion, which comprises the following steps: collecting electrochemical, bioimpedance, physiological, environmental and behavioral multi-source signals synchronously and adding time stamps to form a multi-modal data set. The electrochemical signal is used as a time sequence reference to align and preprocess, and a standardized data matrix is constructed. Cross features of the environment and behavior are extracted and fused to form an initial feature set. A fusion feature vector is constructed, and real-time temperature and personal metabolic history data are combined for dynamic compensation to generate calibration features. The calibration features are input into a pre-trained CNN-LSTM model to output blood glucose prediction values and confidence scores. When the prediction value exceeds the safety range or the confidence is lower than the threshold, a safety control instruction is triggered. Finally, the system adaptively updates the model parameters according to the instruction execution result and user feedback, thereby improving the overall accuracy and reliability of blood glucose monitoring.

Description

Technical Field

[0001] This application relates to the fields of medical electronics and biosensing technology, and in particular to a high-precision blood glucose monitoring method and system based on multimodal data fusion. Background Technology

[0002] With the continued rise in global diabetes prevalence, blood glucose monitoring technology has become a core component of diabetes management. Traditional finger-prick blood sampling devices and continuous glucose monitoring systems achieve quantitative analysis of blood glucose levels through technologies such as electrochemical sensing and near-infrared spectroscopy. Recent research has further explored non-invasive alternatives such as bioimpedance and sweat glucose detection. At the technology integration level, wearable devices are gradually integrating multiple sensors, including ECG, motion, and environmental sensors, providing richer contextual information for blood glucose trend analysis. Simultaneously, artificial intelligence algorithms are beginning to be applied to blood glucose prediction models, improving the predictability of clinical decisions through time-series data analysis. These developments have laid the foundation for dynamic blood glucose management, driving the evolution of medical devices from single-measurement devices to comprehensive health management platforms.

[0003] However, existing technologies still have some limitations. On the one hand, the integration efficiency of multi-source heterogeneous data is low, and the temporal asynchrony of different sensing modules makes it difficult to effectively correlate physiological signals. For example, changes in blood flow caused by exercise can simultaneously interfere with electrochemical signals and impedance measurements, but traditional systems lack cross-modal collaborative correction mechanisms. On the other hand, single-modal data acquisition is susceptible to environmental interference and individual differences. Problems such as the interference of temperature fluctuations on enzyme reaction sensitivity and the attenuation of optical signals by skin thickness have not yet been systematically solved. In addition, existing algorithm models usually rely on static calibration, making it difficult to adapt to the dynamic changes in the patient's metabolic state, especially in some complex scenarios where prediction bias is significant. Summary of the Invention

[0004] To address the aforementioned technical issues, this application provides a high-precision blood glucose monitoring method and system based on multimodal data fusion.

[0005] Firstly, this application provides a high-precision blood glucose monitoring method based on multimodal data fusion, employing the following technical solution: A high-precision blood glucose monitoring method based on multimodal data fusion, the high-precision blood glucose monitoring method comprising: By synchronously collecting users' electrochemical current signals, bioimpedance spectrum signals, physiological parameter signals, environmental parameter signals, and user behavior signals through multiple source sensors, a multi-source data set with timestamps is generated. Based on the sampling timing of the electrochemical current signal, the multi-source data set is time-aligned and preprocessed to generate a standardized data matrix; The time-domain features of electrochemical signals, the frequency-domain features of bioimpedance signals, and the nonlinear features of physiological parameter signals are extracted from the standardized data matrix, and the environmental parameter signals and user behavior signals are combined to generate environmental-behavior cross features, thus obtaining an initial feature vector set. Principal component analysis is performed on the initial feature vector set to reduce dimensionality, and a target feature subset is selected by a feature selection algorithm to generate a fused feature vector. Based on real-time ambient temperature data and pre-stored user historical metabolic data, the fused feature vector is dynamically compensated to generate a calibration feature vector. The calibration feature vector is input into the pre-trained blood glucose prediction model, which outputs the predicted blood glucose value and the corresponding confidence score. When the predicted blood glucose value exceeds the preset safety range or the confidence score is lower than the preset confidence threshold, a safety control command is triggered. The parameters of the blood glucose prediction model are updated based on the execution results of the safety control instructions and user feedback data.

[0006] By employing the aforementioned technical solutions, the interaction between external factors and metabolism is quantified based on environmental-behavioral cross-features. Individual differences and temperature interference are eliminated through dynamic compensation, and the risk of false alarms is reduced through confidence-driven decision-making, ultimately achieving clinical-grade accuracy in blood glucose prediction. Simultaneously, the system's self-updating capability allows it to adapt to long-term physiological changes in users, improving the safety and adherence of diabetes management and providing reliable technical support for precision medicine.

[0007] Optionally, the step of performing time alignment and preprocessing on the multi-source data set based on the sampling time sequence of the electrochemical current signal to generate a standardized data matrix includes: Retrieve a multi-source data set with timestamps; Using the sampling time point of the electrochemical current signal as the reference time series, time alignment is performed on the sampling points of other signals in the multi-source data set to generate a time-aligned data set; The time-aligned dataset is locally filtered using a sliding window mechanism, multi-scale noise removal is performed to eliminate high-frequency interference, missing data points are detected and interpolation methods are used to fill in the complete data sequence, resulting in a preprocessed dataset. Perform standardization operations on the preprocessed dataset to generate a standardized data matrix.

[0008] By adopting the above technical solution, a high-quality multimodal data fusion foundation was constructed, which transformed the originally heterogeneous and noisy multi-source physiological signals into a synchronous and comparable standard data matrix, providing reliable data support for subsequent feature extraction and model building. This not only improved the accuracy and stability of blood glucose monitoring, but also enabled physiological information from different modalities to effectively complement each other, giving full play to the technical advantages of multimodal fusion.

[0009] Optionally, the steps of performing principal component analysis to reduce the dimensionality of the initial feature vector set and filtering the target feature subset using a feature selection algorithm to generate the fused feature vector include: Calculate the covariance matrix of the initial eigenvector set, perform eigenvalue decomposition, and arrange the eigenvectors in descending order of eigenvalues; Feature vectors whose cumulative contribution rate exceeds a preset contribution rate threshold are selected as principal component basis vectors. The initial feature vectors are then projected onto the low-dimensional space spanned by the principal component basis vectors to obtain the low-dimensional feature space. In the low-dimensional feature space, a subset of target features that meet preset conditions and are correlated with blood glucose concentration are selected using a feature selection algorithm; The features in the target feature subset are weighted and combined according to preset weight coefficients to output a fused feature vector.

[0010] By adopting the above technical solution, an efficient conversion from multimodal raw features to optimized fused feature vectors is achieved. Through a two-level optimization mechanism of dimensionality reduction and feature selection, the quality and efficiency of the feature set are improved while retaining key blood glucose-related information, enabling the subsequent prediction model to focus on the most relevant feature patterns, thereby significantly improving the accuracy and reliability of blood glucose monitoring.

[0011] Optionally, the step of dynamically compensating the fused feature vector based on real-time ambient temperature data and pre-stored user historical metabolic data to generate a calibration feature vector includes: Acquire fused feature vectors, real-time ambient temperature data, and user historical metabolic data; Calculate the temperature sensitivity compensation coefficient based on the real-time ambient temperature data. The individual metabolic parameters from the user's historical metabolic data are used to generate a metabolic adaptive scaling factor; The fused feature vector and the temperature sensitivity compensation coefficient are subjected to a first operation to generate a temperature compensation feature vector; The temperature compensation feature vector is combined with the metabolic adaptive scaling factor to generate a calibration feature vector.

[0012] By adopting the above technical solution, a dual calibration mechanism based on temperature sensitivity compensation and metabolic adaptive scaling is used to improve the environmental robustness and individual adaptability of multimodal feature vectors. This calibration method not only improves the accuracy of blood glucose prediction models in different usage environments and among different user groups, but more importantly, it establishes an intelligent calibration system that can adapt to environmental changes and individual differences, providing key technical support for realizing full-scenario, personalized blood glucose monitoring.

[0013] Optionally, the step of inputting the calibration feature vector into the pre-trained blood glucose prediction model and outputting the predicted blood glucose value and the corresponding confidence score includes: The calibration feature vector is input into the pre-trained blood glucose prediction model, and spatial dimension features are extracted through feature space decomposition. Perform time-series analysis on the spatial dimensional features to generate predicted blood glucose concentration values; The uncertainty measure of features is calculated based on the hidden layer state within the blood glucose prediction model. A confidence score is generated based on the aforementioned feature uncertainty measure; Output the predicted blood glucose value and the corresponding confidence score.

[0014] By employing the above technical solutions, a high-precision and self-evaluating blood glucose prediction system was achieved. Combining deep learning models with rigorous uncertainty quantification techniques, it provides both accurate blood glucose concentration predictions and the ability to objectively assess the reliability of the prediction results. This dual-output mechanism enhances the system's practicality and security, enabling it to adaptively adjust decision-making strategies based on prediction confidence in complex real-world application scenarios, avoiding blind reliance on potentially unreliable predictions.

[0015] Optionally, the step of updating the parameters of the blood glucose prediction model based on the execution result of the safety control command and user feedback data includes: Obtain the execution results of safety control commands, user feedback data, and venous blood calibration measurement values; An incremental training sample set is constructed based on the execution results and user feedback data. The actual blood glucose labels of the incremental training sample set are labeled according to the venous blood calibration measurement values; Extract the current parameter state of the blood glucose prediction model to generate an initial parameter set; Based on the labeled incremental training sample set, the initial parameter set is optimized with the objective function of minimizing prediction error and risk penalty; The optimized parameter set is then updated in the blood glucose prediction model.

[0016] By adopting the above technical solutions, the blood glucose prediction system has achieved self-optimization and continuous improvement capabilities, enabling the system to learn and progress continuously from practical use. This dynamic optimization mechanism not only improves the model's individual adaptability and predictive accuracy, but more importantly, ensures the safety of clinical applications through the design of a risk-aware objective function. This allows the system to increasingly accurately match users' unique physiological characteristics and lifestyles over long-term use, providing a key methodological foundation for achieving truly personalized, safe, and reliable intelligent blood glucose monitoring, and promoting the evolution of blood glucose management from static models to dynamic learning systems.

[0017] Optionally, after performing principal component analysis to reduce the dimensionality of the initial feature vector set and filtering the target feature subset using a feature selection algorithm to generate the fused feature vector, the method further includes: The system acquires user motion state recognition results in real time. When high-intensity motion is detected, the collaborative correction mechanism is activated to generate motion interference evaluation factors. Based on the motion interference assessment factor, the time-domain feature component of the electrochemical signal and the frequency-domain feature component of the bioimpedance signal in the fused feature vector are coupled and compensated to generate a motion interference compensated feature vector. Based on the temperature and humidity data in the environmental parameter signals, calculate the temperature and humidity combined interference index; The individual metabolic correction parameters in the pre-stored user historical metabolic data are called to adaptively scale the motion interference compensated feature vector to generate a personalized scaled feature vector. The personalized scaling feature vector is fused with the temperature and humidity combined interference index through multivariate regression to generate an environmental adaptive feature vector. The step of dynamically compensating the fused feature vector by outputting the environment adaptive feature vector.

[0018] By adopting the above technical solutions, the three types of interfering factors—exercise, environment, and metabolism—are transformed from being treated independently to being analyzed in a coordinated manner. More refined signal compensation is achieved by quantifying their interaction effects. Particularly in addressing individual differences, the system intelligently corrects signal variations in different users exhibiting the same physiological state through metabolic parameter embedding and personalized regression weights. This improves monitoring robustness in complex scenarios, maintaining clinical-grade accuracy under conditions of high-intensity exercise, extreme environments, and metabolic fluctuations, providing reliable technical support for the full life-cycle management of diabetic patients.

[0019] Secondly, this application provides a high-precision blood glucose monitoring system based on multimodal data fusion, which adopts the following technical solution: A high-precision blood glucose monitoring system based on multimodal data fusion, the monitoring system comprising: The multi-source data acquisition module is used to simultaneously acquire the user's electrochemical current signal, bioimpedance spectrum signal, physiological parameter signal, environmental parameter signal and user behavior signal through multi-source sensors, and generate a multi-source data set with timestamps; The data processing module is used to perform time alignment and preprocessing on the multi-source data set based on the sampling time sequence of the electrochemical current signal to generate a standardized data matrix. The multimodal feature extraction module is used to extract the time-domain features of electrochemical signals, the frequency-domain features of bioimpedance signals, and the nonlinear features of physiological parameter signals from the standardized data matrix, and to jointly generate environment-behavior cross features by combining the environmental parameter signals and user behavior signals to obtain an initial feature vector set. The feature dimensionality reduction and selection module is used to perform principal component analysis to reduce the dimensionality of the initial feature vector set, and to filter the target feature subset through a feature selection algorithm to generate a fused feature vector. The dynamic feature compensation module is used to dynamically compensate the fused feature vector based on real-time ambient temperature data and pre-stored user historical metabolic data to generate a calibration feature vector. The blood glucose prediction module is used to input the calibration feature vector into the pre-trained blood glucose prediction model and output the blood glucose prediction value and the corresponding confidence score. The safety control trigger module is used to trigger a safety control command when the blood glucose prediction value exceeds the preset safety range or the confidence score is lower than the preset confidence threshold. The model parameter update module is used to update the parameters of the blood glucose prediction model based on the execution results of the safety control instructions and user feedback data.

[0020] Thirdly, this application provides a computer device, which adopts the following technical solution: A computer device includes a memory, a processor, and a computer program stored in the memory, the processor executing the computer program to perform the steps of the method as described in the first aspect.

[0021] Fourthly, this application provides a computer-readable storage medium, which adopts the following technical solution: A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as in any of the methods in the first aspect. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the first process of a high-precision blood glucose monitoring method based on multimodal data fusion according to one embodiment of this application.

[0023] Figure 2This is a schematic diagram of the second process of a high-precision blood glucose monitoring method based on multimodal data fusion, according to one embodiment of this application.

[0024] Figure 3 This is a schematic diagram of the third process of a high-precision blood glucose monitoring method based on multimodal data fusion according to one embodiment of this application.

[0025] Figure 4 This is a schematic diagram of the fourth process of a high-precision blood glucose monitoring method based on multimodal data fusion according to one embodiment of this application.

[0026] Figure 5 This is a schematic diagram of the fifth step of a high-precision blood glucose monitoring method based on multimodal data fusion, according to one embodiment of this application.

[0027] Figure 6 This is a schematic diagram of the sixth process of a high-precision blood glucose monitoring method based on multimodal data fusion according to one embodiment of this application.

[0028] Figure 7 This is a schematic diagram of the seventh process of a high-precision blood glucose monitoring method based on multimodal data fusion according to one embodiment of this application. Detailed Implementation

[0029] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figures 1-7 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0030] This application discloses a high-precision blood glucose monitoring method based on multimodal data fusion.

[0031] Reference Figure 1 A high-precision blood glucose monitoring method based on multimodal data fusion, specifically including: Step S101: Simultaneously collect the user's electrochemical current signal, bioimpedance spectrum signal, physiological parameter signal, environmental parameter signal and user behavior signal through multi-source sensors to generate a multi-source data set with timestamps. Among these, electrochemical current signals are the core of blood glucose detection. They directly reflect blood glucose concentration based on the current changes generated by the glucose oxidase reaction, but are easily interfered with by factors such as temperature and blood oxygen. Bioimpedance spectroscopy signals indirectly assess the glucose diffusion in extracellular fluid by measuring the electrical impedance characteristics of tissues at different frequencies. Its multi-band scanning (e.g., 1kHz-1MHz) can distinguish the conductivity differences between intracellular and extracellular fluids, thus providing complementary metabolic information. Physiological parameter signals (e.g., heart rate variability, skin conductance) can capture the stress state of the autonomic nervous system, such as blood glucose fluctuations caused by stress or exercise. Environmental parameter signals (e.g., temperature, humidity, air pressure) are used to quantify the impact of external conditions on sensor accuracy and human metabolic rate. User behavior signals (e.g., exercise intensity, dietary image recognition) are related to energy consumption and carbohydrate intake, directly correlated with blood glucose change trends.

[0032] It is important to note that the key to synchronous data acquisition lies in adding precise timestamps to all signals. This ensures the temporal consistency of multi-source data in subsequent analysis and avoids causal misjudgments due to acquisition time discrepancies. For example, changes in blood glucose after exercise must be matched with simultaneous increases in heart rate and changes in skin conductance to accurately distinguish between physiological responses and sensor noise.

[0033] Step S102: Based on the sampling time sequence of the electrochemical current signal, perform time alignment and preprocessing on the multi-source data set to generate a standardized data matrix; Electrochemical signals are used as the reference because they have the highest sampling rate (e.g., 100Hz) and can provide the finest time resolution. Other signals (e.g., physiological parameters at 1Hz, environmental data at 0.1Hz) need to be aligned to a unified time axis through spline interpolation and other methods to eliminate time drift caused by device response delay or differences in acquisition frequency.

[0034] Specifically, the preprocessing includes median filtering to remove impulse noise, wavelet thresholding (such as the db4 wavelet basis) to separate the signal from motion artifacts, and baseline drift correction to eliminate the effects of long-term sensor drift. For example, bioimpedance signals are easily affected by changes in electrode contact. By weakening high-frequency noise in the frequency domain through wavelet transform, the slow trend of glucose concentration in tissue fluid can be reflected more accurately.

[0035] Step S103: Extract the time-domain features of electrochemical signals, the frequency-domain features of bioimpedance signals, and the nonlinear features of physiological parameter signals from the standardized data matrix, and combine the environmental parameter signals and user behavior signals to generate environmental-behavior cross features, thus obtaining an initial feature vector set. Among them, the time-domain characteristics of electrochemical signals (such as mean, variance, and slope) directly reflect the instantaneous rate of change and stability of blood glucose concentration; the frequency-domain characteristics of bioimpedance signals (such as power spectral density and phase angle) can identify changes in the dielectric properties of tissue fluid at different frequencies, and these properties are related to the polarity of glucose molecules; the nonlinear characteristics of physiological parameters (such as sample entropy and fractal dimension) are used to quantify the complexity and stress state of the autonomic nervous system. For example, a decrease in the entropy value of heart rate variability may predict autonomic nervous system disorder before hypoglycemia.

[0036] In this application embodiment, the generation of environment-behavior cross-features is a key innovation. It quantifies the interaction effect between environment and behavior through mathematical operations (such as the product of temperature and exercise intensity). For example, when exercising in a high-temperature environment, the increase in body temperature and the acceleration of blood flow will synergistically affect the glucose diffusion rate, while traditional methods only process temperature or exercise data separately. This cross-feature couples the external environment with the human body's active behavior, more accurately simulating metabolic dynamics in real-world scenarios.

[0037] Finally, all features together constitute the initial feature vector set, which essentially maps multi-source signals into a comprehensive set of indicators that can simultaneously describe direct blood glucose measurements, indirect physiological states, and external interference factors.

[0038] Step S104: Perform principal component analysis to reduce the dimensionality of the initial feature vector set, and use a feature selection algorithm to filter the target feature subset to generate a fused feature vector; Principal Component Analysis (PCA) projects initial features (e.g., 500-dimensional) into a lower-dimensional space (e.g., 50-dimensional) through orthogonal transformation, retaining 95% of the variance contribution rate. Mathematically, it extracts the direction of maximum variance from the data as a new feature, eliminating redundant information while maximizing representation of the original data distribution. For example, the time-domain characteristics of electrochemical signals and the frequency-domain characteristics of bioimpedance may be highly correlated (both influenced by blood flow), and PCA can merge them into a few composite components.

[0039] Subsequently, feature selection algorithms (such as ReliefF) filter features based on the strength of their association with blood glucose levels: ReliefF assigns weights to features by calculating their discriminative power among neighboring samples, prioritizing features that can significantly separate high and low blood glucose levels (e.g., environmental-behavioral cross-features may have higher weights). The physical meaning of this process is to focus on the core indicators most sensitive to blood glucose changes, eliminate noisy or irrelevant features, and generate a fused feature vector that retains the complementarity of multimodal data while reducing computational complexity, providing optimized input for subsequent models.

[0040] Step S105: Based on real-time ambient temperature data and pre-stored user historical metabolic data, dynamically compensate the fused feature vector to generate a calibration feature vector. The underlying logic of this step is to achieve real-time adaptive calibration to address individual differences and environmental fluctuations, thereby improving monitoring accuracy. Environmental temperature compensation is based on the Arrhenius equation, whose physical principle is that temperature affects the activation energy of electrochemical reactions: as temperature increases, the enzyme reaction rate accelerates, leading to an enhanced current signal. The sensitivity coefficient needs to be adjusted exponentially to avoid misinterpreting temperature effects as changes in blood glucose. Individualized correction relies on historical metabolic data (such as insulin sensitivity and basal metabolic rate), fine-tuning feature weights through transfer learning: for example, for users with slow metabolisms, bioimpedance features may be more predictive than electrochemical features, and dynamic compensation will correspondingly increase their weight.

[0041] Understandably, the essence of this compensation mechanism is to establish a feedback loop for personalized blood glucose prediction. It corrects the physical meaning of the feature vector bidirectionally through real-time data (temperature) and historical baselines (metabolic profiles), ensuring that the fused features accurately correspond to the true blood glucose value under different environments and individual conditions. The calibrated feature vector not only contains the measured value but also incorporates environmental and individual regulatory information, thus better reflecting physiological reality.

[0042] Step S106: Input the calibration feature vector into the pre-trained blood glucose prediction model and output the blood glucose prediction value and the corresponding confidence score. Among them, deep learning models are used to learn the complex mapping relationship of blood glucose changes from multimodal features and quantify the uncertainty of prediction.

[0043] In this embodiment, the blood glucose prediction model typically employs a CNN-LSTM hybrid architecture: the CNN convolutional layers extract spatial features (such as the local correlation between electrochemical signals and impedance signals) through local perception and weight sharing, while the LSTM layers utilize gating mechanisms to memorize long-term dependencies (such as the temporal patterns of postprandial blood glucose rise and post-exercise blood glucose decline). The confidence score is calculated based on the standard deviation of the prediction residuals, and its mathematical essence is to evaluate the stability of the model's historical performance under the current input features: if the current feature combination is highly consistent with the distribution of the training data, the confidence is high; conversely, if the features are abnormal (such as sudden strenuous exercise), the confidence decreases. For example, when rare values ​​appear in the environmental-behavioral cross-features, the model may output a high blood glucose prediction but with low confidence, indicating that the result needs verification.

[0044] Step S107: When the predicted blood glucose value exceeds the preset safety range or the confidence score is lower than the preset confidence threshold, a safety control command is triggered. In the embodiments of this application, the preset safety range (e.g., 3.9-13.9 mmol / L) can be based on clinical guidelines. Exceeding the range may cause hypoglycemic coma or hyperglycemic ketoacidosis. The confidence threshold (e.g., 0.8) is used to prevent the risk of model uncertainty. For example, if the signal quality is poor and the prediction is unreliable, a warning should be issued even if the value is within the safety range.

[0045] Furthermore, the execution results of safety control commands record the system's automatic intervention behavior when abnormalities are detected, such as insulin infusion suspension or alarm triggering. These behaviors themselves indirectly verify the reliability of the prediction results. The triggering mechanism of safety control commands is essentially a multi-condition decision-making logic: it simultaneously considers the absolute risk of blood glucose levels and the relative reliability of the prediction model. For example, when the confidence level is low, the system may prioritize re-detection rather than directly triggering an alarm to avoid false intervention. This design couples physiological indicators with the model state, achieving a leap from passive monitoring to proactive management, and is particularly suitable for scenarios such as closed-loop insulin infusion systems, where false alarms may lead to treatment accidents.

[0046] Step S108: Update the parameters of the blood glucose prediction model based on the execution results of the safety control instructions and user feedback data.

[0047] Specifically, the parameter update mechanism can be based on reinforcement learning (such as the PPO algorithm), whose objective function balances the blood glucose target achievement rate with the risk of hypoglycemia. For example, after the system triggers hypoglycemia warnings multiple times, PPO adjusts the model parameters to reduce dependence on certain sensitive features (such as post-exercise heart rate). User feedback data (such as manually entered blood glucose values ​​and dietary confirmations) provides supervisory information to correct model biases. If user feedback does not match the prediction, the system fine-tunes the network weights through incremental learning, enabling the model to adapt to metabolic changes more quickly (such as the new normal after medication adjustment). The essence of this process is to build an adaptive iterative optimization loop, which transforms each decision result into a learning sample, enabling the model to not only make static predictions but also evolve dynamically, ultimately achieving continuous optimization of personalized blood glucose monitoring.

[0048] In the above implementation, the interaction between external factors and metabolism is quantified based on environmental-behavioral cross-features. Individual differences and temperature interference are eliminated through dynamic compensation, and the risk of false alarms is reduced through confidence-driven decision-making, ultimately achieving clinical-grade accuracy in blood glucose prediction. Simultaneously, the system's self-updating capability allows it to adapt to long-term physiological changes in users, improving the safety and adherence of diabetes management and providing reliable technical support for precision medicine.

[0049] Reference Figure 2 As one implementation of step S102, the step of time alignment and preprocessing of the multi-source data set to generate a standardized data matrix, based on the sampling timing of the electrochemical current signal, includes: Step S201: Obtain a multi-source data set with timestamps; The precise recording of timestamps ensures the traceability of these heterogeneous signals across time. Essentially, it assigns a unique time coordinate to each data point, forming a multi-dimensional data stream with strict temporal relationships, thus establishing a unified temporal reference system for subsequent collaborative analysis. For example, when a rapid rise in blood sugar is detected, it is necessary to simultaneously check whether there are dietary behavior signals at the same time, thereby distinguishing between normal postprandial blood sugar fluctuations and abnormal metabolic events.

[0050] Step S202: Using the sampling time point of the electrochemical current signal as the reference time series, perform time alignment operation on the sampling points of other signals in the multi-source data set to generate a time-aligned data set; Electrochemical current signals were chosen as the reference time series because they have the highest sampling frequency and time resolution, providing the finest time scale reference. For signals with lower sampling rates (such as physiological parameters typically at 1 Hz and environmental parameters at 0.1 Hz), they need to be resampled onto the reference time axis using interpolation algorithms. The essence of this alignment operation is to construct a unified time coordinate system, enabling signals with different physical meanings and different sampling frequencies to establish a correspondence at the same point in time.

[0051] For example, due to limitations in the measurement principle, the sampling point of bioimpedance signals may be hundreds of milliseconds later than that of electrochemical signals. Spline interpolation algorithms can reconstruct the estimated value at the electrochemical sampling time, ensuring strict time synchronization between the two signals during analysis. This temporal alignment not only eliminates acquisition delays between devices but, more importantly, provides time constraints for subsequent research on causal relationships between multimodal signals. For instance, it allows for accurate analysis of how many seconds after the start of exercise will blood glucose characteristic changes occur. This temporal relationship is crucial for establishing accurate blood glucose prediction models.

[0052] Step S203: Apply a sliding window mechanism to perform local filtering on the time-aligned dataset, perform multi-scale noise removal to eliminate high-frequency interference, detect missing data points and use interpolation methods to fill in the complete data sequence to obtain the preprocessed dataset; Among these methods, multi-level signal processing techniques are used to eliminate noise interference and abnormal fluctuations in the data, thereby improving signal quality. Specifically, data filtering employs a sliding window mechanism, which calculates local statistics by moving a window across the time series, effectively smoothing random fluctuations while preserving trend changes. Noise removal is based on the principle of multi-scale analysis, employing corresponding strategies for different types of noise. For example, wavelet transform can separate high-frequency noise components in the signal, while median filtering can effectively suppress impulse interference. Missing data processing reconstructs data continuity through interpolation methods, the core of which is to infer reasonable values ​​for missing points based on the variation patterns of known data points.

[0053] The physical significance of these preprocessing operations lies in restoring the true physiological signal characteristics masked by various interferences. For example, changes in electrode contact caused by movement can produce baseline drift in bioimpedance signals. Adaptive filtering can correct for this slowly varying interference, while environmental electromagnetic noise manifests as high-frequency spikes, which need to be suppressed by wavelet threshold denoising. The preprocessed data is not only smoother and more stable, but more importantly, it eliminates spurious changes caused by non-physiological factors, providing a reliable foundation for subsequent feature extraction.

[0054] Step S204: Perform standardization processing on the preprocessed dataset to generate a standardized data matrix.

[0055] Specifically, standardization is based on the statistical characteristics of each signal dimension. It calculates the mean and standard deviation of each signal, and then applies Z-score standardization to transform the original data into a standard distribution with a mean of 0 and a standard deviation of 1. The mathematical essence of this process is to map signals with different physical dimensions and numerical ranges to a unified numerical space, enabling subsequent machine learning algorithms to treat each feature fairly and preventing certain large numerical features from dominating model training. For example, typical electrochemical current signals are in the nanoampere range, while bioimpedance values ​​can reach the kiloohm range. Without standardization, the model will naturally focus more on the larger impedance signals and ignore the weaker current signals.

[0056] In the above implementation, a high-quality multimodal data fusion foundation is constructed, which transforms the originally heterogeneous and noisy multi-source physiological signals into a synchronous and comparable standard data matrix, providing reliable data support for subsequent feature extraction and model building. This not only improves the accuracy and stability of blood glucose monitoring, but also enables physiological information from different modalities to effectively complement each other, giving full play to the technical advantages of multimodal fusion.

[0057] Reference Figure 3 As one implementation of step S104, the steps of performing principal component analysis to reduce the dimensionality of the initial feature vector set and filtering the target feature subset using a feature selection algorithm to generate the fused feature vector include: Step S301: Calculate the covariance matrix of the initial eigenvector set, perform eigenvalue decomposition, and arrange the eigenvectors in descending order of eigenvalues. Specifically, the covariance matrix of the entire initial set of eigenvectors is calculated. Each element of this matrix quantifies the covariance relationship between any two feature dimensions, and the diagonal elements represent the variance of each feature itself. Then, eigenvalue decomposition is performed on the covariance matrix. This is a crucial matrix decomposition operation, the mathematical result of which is a set of eigenvalues ​​and their corresponding eigenvectors. The magnitude of the eigenvalue represents the variance (i.e., information content) of the data along the corresponding eigenvector direction; the eigenvectors define a new set of orthogonal (i.e., completely uncorrelated) coordinate axis directions, called "principal component" directions. The eigenvectors are arranged in descending order of eigenvalue, which is equivalent to sorting them according to the amount of original information carried by the new direction, with the direction carrying the most information first.

[0058] Step S302: Select the feature vector whose cumulative contribution rate exceeds the preset contribution rate threshold as the principal component basis vector, and project the initial feature vector onto the low-dimensional space spanned by the principal component basis vector to obtain the low-dimensional feature space. The cumulative contribution rate is defined as the proportion of the sum of the first m largest eigenvalues ​​to the sum of all eigenvalues. It quantifies the percentage of total variation in the original data that can be explained by the first m principal components. A preset contribution rate threshold (e.g., 95%) is set, and then the contribution rate is accumulated starting from the principal component with the most information (the eigenvector corresponding to the largest eigenvalue) until the cumulative contribution rate of the selected principal components exceeds the threshold. These selected eigenvectors constitute the principal component basis vectors.

[0059] Next, each initial feature vector is projected onto a low-dimensional space spanned by the selected principal component basis vectors. The mathematical operation of projection is the dot product of the vectors and the basis vectors (or matrix multiplication), transforming the original high-dimensional feature representation of each sample onto a new, unrelated set of low-dimensional coordinate axes ordered by importance. After this transformation, the output is a low-dimensional feature space with dimensions far lower than the original feature space.

[0060] By discarding principal component directions with low variance contribution rates (usually corresponding to noise or weak, redundant information) through orthogonal transformation, the data is compressed from high dimension to low dimension while preserving the main distribution structure (variance) of the original data to the maximum extent. This significantly reduces the complexity of the data and the computational burden of subsequent processing, and eliminates multicollinearity among features.

[0061] Step S303: In the low-dimensional feature space, a subset of target features that meet the preset conditions for correlation with blood glucose concentration is selected by a feature selection algorithm; Feature selection algorithms (such as ReliefF) assign importance weights to features by evaluating their discriminative ability in distinguishing samples with different blood glucose levels. Their core mechanism examines the clustering of feature values ​​among similar samples and their segregation among dissimilar samples. For example, an ideal blood glucose-related feature should result in similar feature values ​​for samples with the same blood glucose level, while showing significant differences in feature values ​​for samples with different blood glucose levels. This evaluation considers not only the linear correlation between features and blood glucose values ​​but also captures nonlinear relationships; for instance, nonlinear features of certain physiological parameters may exhibit better discriminative ability within specific blood glucose ranges. The feature selection process essentially seeks a balance between feature effectiveness and redundancy, retaining features that contribute most to blood glucose prediction while eliminating noisy features that, although containing information, have low correlation with blood glucose levels.

[0062] Step S304: The features in the target feature subset are weighted and combined according to preset weight coefficients to output a fused feature vector.

[0063] Different features contribute differently to blood glucose prediction, so appropriate weighting coefficients need to be assigned according to their importance. For example, the time-domain features of electrochemical signals, as direct measurement indicators, may receive higher weights, while environmental-behavioral cross-features, as indirect correction terms, may have relatively lower weights. However, this weighting is not fixed but can be adaptively adjusted according to individual physiological characteristics and usage environment.

[0064] Next, the weighted combination is mathematically represented as a linear weighted summation process, where the value of each feature is multiplied by its corresponding weight coefficient, and then summed (or concatenated into a new vector). The final output "fusion feature vector" has a dimension equal to the number of features (k) in the target feature subset, and the value of each dimension reflects the original features after weight modulation.

[0065] The above implementation achieves efficient conversion from multimodal raw features to optimized fused feature vectors. Through a two-level optimization mechanism of dimensionality reduction and feature selection, the quality and efficiency of the feature set are improved while retaining key blood glucose-related information, enabling the subsequent prediction model to focus on the most relevant feature patterns, thereby significantly improving the accuracy and reliability of blood glucose monitoring.

[0066] Reference Figure 4 As one implementation of step S105, the step of dynamically compensating the fused feature vector based on real-time ambient temperature data and pre-stored user historical metabolic data to generate a calibration feature vector includes: Step S401: Obtain the fused feature vector, real-time ambient temperature data, and user historical metabolic data; The introduction of real-time ambient temperature data is based on the physical characteristics of the electrochemical biosensor. Because the catalytic activity of glucose oxidase is significantly temperature-dependent, its reaction rate changes non-linearly with temperature, and this temperature effect directly alters the baseline value of the electrochemical current signal. User historical metabolic data records individual physiological parameters, such as the insulin sensitivity coefficient reflecting the body's response to insulin, and the carbohydrate metabolism ratio characterizing the efficiency of blood glucose conversion after food intake. These parameters collectively define the unique glucose metabolism kinetics of an individual.

[0067] Step S402: Calculate the temperature sensitivity compensation coefficient based on real-time ambient temperature data; In this embodiment, the temperature sensitivity compensation coefficient is calculated based on the temperature-reaction rate relationship described by the Arrhenius equation in enzyme kinetics. Its physical essence is that ambient temperature affects catalytic efficiency by altering the molecular kinetic energy of glucose oxidase. When the temperature deviates from standard detection conditions, enzyme activity undergoes a systematic shift, leading to deviations in the current signal generated at the same blood glucose concentration.

[0068] Understandably, the process of generating the compensation coefficient essentially involves establishing a correction mapping between the signal output under the current ambient temperature and the ideal detection temperature. For example, when the ambient temperature rises, the enzyme reaction rate increases, which strengthens the current signal. In this case, a compensation coefficient less than 1 is needed to counteract this temperature-induced signal amplification effect. This compensation is not a simple linear adjustment but requires consideration of the comprehensive impact of temperature on enzyme activity, diffusion rate, and electrode interface characteristics, thereby ensuring an accurate signal reference can be obtained under different ambient temperatures.

[0069] Step S403: Call the individual metabolic parameters in the user's historical metabolic data to generate a metabolic adaptive scaling factor; Among them, the metabolic adaptive scaling factor is a personalized regulatory parameter established based on long-term user monitoring data. It reflects the influence pattern of a specific individual's glucose metabolism characteristics on the signal-glucose relationship. For example, individuals with higher insulin sensitivity may have a faster ability to regulate blood glucose fluctuations, which can lead to a systematic difference in the correspondence between electrochemical signals and blood glucose values ​​compared to insulin-resistant individuals.

[0070] Understandably, the process of generating scaling factors essentially involves learning the correlation patterns between individual physiological characteristics and multimodal signal responses. By quantifying the signal expression characteristics under different metabolic states, the general feature vector is adjusted to a numerical range that matches the individual's physiological characteristics. This personalized scaling not only considers basic metabolic parameters but may also include time-varying metabolic adaptability, enabling the system to adapt to the long-term evolution of the user's metabolic state.

[0071] Step S404: Perform a first operation on the fused feature vector and the temperature sensitivity compensation coefficient to generate a temperature compensation feature vector; The purpose of this step is to eliminate the systematic impact of ambient temperature fluctuations on the eigenvectors. The first operation typically employs element-wise multiplication, where each dimension of the eigenvector is operated on separately with a temperature compensation coefficient. The mathematical essence of this operation is to scale the eigenvalues ​​proportionally.

[0072] Understandably, temperature changes affect different features to varying degrees: electrochemical features may be most sensitive to temperature and require significant compensation, while some physiological parameters may be relatively insensitive to temperature, requiring less compensation. This dimensionally differentiated compensation accurately corrects the systematic bias of temperature on multimodal features while preserving the relative relationships between features. The temperature-compensated eigenvectors eliminate the primary interfering factor of ambient temperature, allowing the feature values ​​to more accurately reflect physiological states rather than the influence of environmental conditions.

[0073] Step S405: Perform a second operation on the temperature compensation feature vector and the metabolic adaptation scaling factor to generate a calibration feature vector.

[0074] The purpose of this step is to further adapt to individual metabolic characteristics based on temperature compensation. The second operation typically includes two parts: scaling and bias adjustment. The scaling operation is used to adjust the overall amplitude of the feature vector to match the individual's metabolic level, while the bias adjustment is used to correct for differences in feature baselines caused by different metabolic types. For example, individuals with low insulin sensitivity may have high basal blood glucose levels, which will cause all feature values ​​to exhibit a systematic shift, requiring correction through the bias term.

[0075] It is understandable that individuals in different metabolic states may exhibit different physiological signal patterns at the same blood glucose level. The metabolic scaling factor adjusts the numerical distribution of the feature vector so that the same blood glucose value corresponds to similar feature expressions in different individuals. The resulting calibrated feature vector eliminates environmental interference and adapts to individual differences, becoming a standardized feature representation that accurately reflects blood glucose status.

[0076] In the above embodiments, a dual calibration mechanism based on temperature sensitivity compensation and metabolic adaptive scaling improves the environmental robustness and individual adaptability of multimodal feature vectors. This calibration method not only enhances the accuracy of blood glucose prediction models in different usage environments and among different user groups, but more importantly, it establishes an intelligent calibration system that can adapt to environmental changes and individual differences, providing key technical support for achieving full-scenario, personalized blood glucose monitoring.

[0077] Reference Figure 5As one implementation of step S106, the step of inputting the calibration feature vector into the pre-trained blood glucose prediction model and outputting the predicted blood glucose value and the corresponding confidence score includes: Step S501: Input the calibration feature vector into the pre-trained blood glucose prediction model and extract the spatial dimension features through feature space decomposition operation; The feature space decomposition operation includes: performing convolution kernel sliding operations on the input feature vector, generating high-dimensional feature maps through non-linear activation functions, and performing spatial pooling operations to reduce the feature dimension.

[0078] Specifically, feature space decomposition identifies specific physiological state patterns expressed by different feature combinations through the sliding operation of convolutional kernels along the feature dimensions. For example, certain electrochemical features and bioimpedance features in specific frequency bands may form spatial correlation patterns, which are strongly correlated with the postprandial blood glucose rise phase. The introduction of nonlinear activation functions enables the model to learn complex interactions between features, while spatial pooling preserves salient features while reducing data dimensionality.

[0079] In some embodiments, the blood glucose prediction model may employ a CNN-LSTM hybrid architecture, wherein the CNN convolutional layers use 3×3 convolutional kernels to extract spatial features, and the LSTM layers use 128 hidden units to capture temporal dependencies.

[0080] Step S502: Perform time series analysis on the spatial dimension features to generate predicted blood glucose concentration values; Specifically, the logical principle of this step is based on the continuous temporal nature of blood glucose changes, utilizing the memory function of recurrent neural networks to build a dynamic prediction model. The temporal analysis operation selectively retains historical information through gating mechanisms (such as the forget gate and input gate in LSTM) and establishes the temporal dependency between current features and past states. For example, current exercise characteristics need to be combined with dietary records over a past period to accurately predict blood glucose trends.

[0081] Understandably, this time-series modeling can distinguish the rhythm of different physiological processes' effects on blood glucose: the rise in blood glucose caused by food absorption is relatively slow, while the drop in blood glucose caused by exercise may appear more quickly. By analyzing the evolution of features over time, the model can not only make judgments based on the current state but also predict based on trends, thereby estimating blood glucose concentration values ​​more accurately.

[0082] Step S503: Calculate the feature uncertainty measure based on the hidden layer state inside the blood glucose prediction model; The calculation of feature uncertainty measure includes: extracting the activation variance of neurons in the hidden layer of the model, calculating the Jacobian matrix norm of the prediction result with respect to the input features, and linearly combining the activation variance and the Jacobian matrix norm into an uncertainty index.

[0083] Specifically, the hidden layer states reflect the model's understanding and encoding of input features, and the variance of their activation values ​​characterizes the model's familiarity with the current input pattern. When the input features are highly consistent with the distribution of the training data, the activation patterns of neurons are usually relatively stable; however, when encountering rare or unusual feature combinations, the activation patterns will fluctuate significantly. The Jacobian matrix norm measures the sensitivity of the model's output to the input features, that is, the degree of prediction fluctuation that may be caused by small changes in features.

[0084] Step S504: Generate a confidence score based on the feature uncertainty measure; The confidence score transforms the uncertainty measure into a standardized value between 0 and 1 using a non-linear mapping function. This transformation process takes into account the different reliability requirements of various application scenarios. A high confidence score indicates that the current input features fall within the core region of the model's training data distribution, and the model's response to these features is stable and predictable; a low confidence score suggests that the current prediction may be affected by rare factors and requires caution.

[0085] Step S505: Output the predicted blood glucose value and the corresponding confidence score.

[0086] The blood glucose prediction provides a specific estimate of blood glucose concentration, while the confidence score indicates the reliability of this estimate. This dual-output design allows subsequent applications to take different action strategies based on the confidence level: high-confidence predictions can be used directly for treatment decisions, medium-confidence predictions may need to be verified in conjunction with other information, and low-confidence predictions can trigger a remeasurement or manual confirmation process.

[0087] The above implementation achieves a high-precision and self-evaluating blood glucose prediction system. By combining a deep learning model with rigorous uncertainty quantification techniques, it provides both accurate blood glucose concentration predictions and the ability to objectively assess the reliability of the prediction results. This dual-output mechanism enhances the system's practicality and security, enabling it to adaptively adjust its decision-making strategy based on prediction confidence in complex real-world application scenarios, avoiding blind reliance on potentially unreliable prediction results.

[0088] Reference Figure 6 As one implementation of step S108, the step of updating the parameters of the blood glucose prediction model based on the execution result of the safety control command and user feedback data includes: Step S601: Obtain the execution result of the safety control command, user feedback data, and venous blood calibration measurement value; Among them, the execution results of safety control commands record the system's automatic intervention behavior when abnormal situations are detected, such as insulin infusion suspension or alarm triggering. These behaviors themselves are indirect verification of the reliability of the prediction results. User feedback data includes subjective information such as symptom descriptions and dietary records, providing clinical-level status annotations. For example, the "dizziness" symptoms reported by users may be related to hypoglycemic events. Venous blood calibration measurements, as the medical gold standard for blood glucose detection, provide an absolutely accurate reference benchmark for model prediction.

[0089] Step S602: Construct an incremental training sample set based on the execution results and user feedback data; The process of constructing the incremental training sample set involves in-depth data mining of system anomaly detection points and user-initiated feedback moments. By analyzing the triggering logic and timestamp information of security control commands, key time points where model predictions may be biased are located. Correlating user feedback data involves overlaying expert clinical judgments onto these time points; for example, matching user-reported "post-meal discomfort" with sensor data for the corresponding time period.

[0090] Understandably, the innovation of this sample construction method lies in its focus not only on collecting ordinary continuous monitoring data, but also on specific scenarios where the model performs poorly or users have doubts, enabling newly trained models to specifically improve their weaknesses. By using time window segmentation technology, continuous data streams are transformed into independent samples, preserving the contextual information of events while meeting the input data format requirements of machine learning models.

[0091] Step S603: Label the incremental training sample set with the true blood glucose labels based on the venous blood calibration measurement values; Among these, venous blood calibration measurements have the highest medical authority, but their collection frequency is limited and they are usually discrete time points. The annotation process can use interpolation algorithms to establish a continuous blood glucose curve between venous blood measurements, and make reasonable estimates of blood glucose values ​​at times not directly calibrated based on physiological principles.

[0092] For example, curve fitting based on glucose metabolism kinetics is performed between two adjacent venous blood measurement points to ensure that the labeled values ​​conform to both medical facts and physiological laws. This refined labeling strategy provides the model with high-quality supervisory information, enabling the model to learn real blood glucose change patterns rather than superficial correlations with sensor signals.

[0093] Step S604: Extract the current parameter state of the blood glucose prediction model and generate an initial parameter set; The current parameter state reflects the optimized state the model has reached after training on historical data, including the weight configurations of each layer such as feature extraction and time series analysis. Extracting these parameters as the starting point for optimization essentially acknowledges that the existing model already possesses a considerable degree of predictive ability, and that the next round of optimization should be an improvement upon this foundation rather than a complete retraining. This strategy maintains the continuity of the model, preventing performance fluctuations caused by significant parameter adjustments, while also ensuring that new knowledge can be gradually integrated into the existing model.

[0094] It's important to note that the generation of parameter sets needs to consider the characteristics of different parts of the model. For example, convolutional layer parameters typically represent general feature extractors and should be relatively conservatively varied; while fully connected layer parameters are directly related to the specific prediction task and can be given more room for adjustment. This differentiated parameter handling helps to achieve targeted optimization while maintaining model stability.

[0095] Step S605: Based on the labeled incremental training sample set, optimize the initial parameter set with the objective function of minimizing prediction error and risk penalty; Minimizing prediction error ensures that the model output is as close as possible to the actual blood glucose value, which is the basic accuracy requirement of the model. The risk penalty term sets additional constraints for hypoglycemia events that are of most concern in clinical practice. When the predicted value is close to or below the safety threshold, an exponentially increasing penalty is applied, forcing the model to be extra cautious in the sensitive range.

[0096] In this embodiment, the dual-objective design recognizes that blood glucose prediction differs from typical regression problems. The clinical consequences of prediction bias are asymmetrical across different intervals, and the risk of overestimating blood glucose is far lower than the risk of severe hypoglycemia caused by underestimating it. During optimization, constraint algorithms can be used to control the magnitude of parameter changes, ensuring that the model does not forget previously learned patterns while absorbing new knowledge, thus maintaining its predictive ability on historical data. This robust optimization strategy allows model updates to adapt to changes in individual characteristics without losing generality due to overfitting to new samples.

[0097] Step S606: Update the optimized parameter set to the blood glucose prediction model.

[0098] It's important to note that parameter updates are not simple replacements but require multi-stage verification: First, the generalization ability of the new model is evaluated on an independent test set to ensure it maintains good performance on unseen data; second, specific safety tests are conducted, focusing on the prediction accuracy in hypoglycemia regions; finally, simulation validation ensures the model's compatibility with other system modules. This cautious update mechanism reflects the rigor expected of medical devices, ensuring that every parameter adjustment is confirmed as a safety improvement rather than introducing new risks. After the update is complete, the system records version information and change logs, establishing a complete model evolution trajectory to provide a basis for subsequent problem tracing and performance analysis.

[0099] The above implementation achieves the self-optimization and continuous improvement capabilities of the blood glucose prediction system, enabling the system to learn and improve continuously from practical use. This dynamic optimization mechanism not only enhances the model's individual adaptability and predictive accuracy, but more importantly, ensures the safety of clinical applications through the design of a risk-aware objective function. This allows the system to increasingly accurately match users' unique physiological characteristics and lifestyles over long-term use, providing a key methodological foundation for achieving truly personalized, safe, and reliable intelligent blood glucose monitoring, and promoting the evolution of blood glucose management from static models to dynamic learning systems.

[0100] Reference Figure 7 As a further implementation of the high-precision blood glucose monitoring method, after step S104, which involves performing principal component analysis to reduce the dimensionality of the initial feature vector set and filtering the target feature subset using a feature selection algorithm to generate the fused feature vector, the method further includes: Step S701: Acquire the user's motion state recognition result in real time. When a high-intensity motion state is detected, activate the collaborative correction mechanism to generate a motion interference evaluation factor. In this embodiment of the application, the criteria for determining high-intensity exercise are a three-dimensional acceleration vector and a sustained intensity of more than 2.5g for 10 seconds; the exercise interference assessment factor is generated through the following process: extracting heart rate variability (SDNN) data from 0 to 90 seconds after the start of exercise; calculating the skin conductance rise slope during the same period; and generating the product factor of the SDNN change rate and the skin conductance slope as the assessment factor.

[0101] Specifically, the rate of decrease in heart rate variability (SDNN) reflects the degree of inhibition of cardiac rhythm regulation by sympathetic nerve excitation, while the slope of the increase in skin conductance characterizes the intensity of sweat gland activity induced by thermoregulation. Multiplying the rates of change of these two parameters essentially constructs a neurohumoral linkage index, capable of capturing the differentiated physiological responses of individuals with different physical fitness levels to the same exercise intensity. For example, due to the decline in autonomic nervous system regulation, older adults experience a more significant decrease in SDNN at the same exercise intensity, while younger individuals may exhibit a more pronounced skin conductance response. This cross-validation mechanism allows the exercise interference assessment factor to eliminate assessment biases caused by individual differences such as age and physical fitness, providing a precise quantitative basis for subsequent signal compensation.

[0102] For example, in a treadmill test, when the acceleration reaches 3g for 15 seconds, the SDNN decreases from 45ms to 28ms (a decrease rate of 38%), and the skin conductance slope increases from 0.12μS / s to 0.35μS / s. Then the motion interference evaluation factor is W_motion=0.38×0.35=0.133.

[0103] Step S702: Based on the motion interference evaluation factor, couple compensation is performed on the time-domain feature components of the electrochemical signal and the frequency-domain feature components of the bioimpedance signal in the fused feature vector to generate a motion interference compensated feature vector. This step involves differentiated compensation for the changes in different physiological signals under motion interference, addressing the shortcomings of traditional methods in independently correcting multiple signals.

[0104] Specifically, under high-intensity exercise, the electrochemical signal will be affected by the increased blood flow in the subcutaneous tissue, which will accelerate glucose diffusion and cause the current measurement value to be falsely high. This blood flow artifact needs to be corrected in reverse by the attenuation coefficient. The specific correction formula is: F'_{ec}=F_{ec} / (1+0.35·W_motion), where W_motion is the motion interference assessment factor.

[0105] The bioimpedance signal is abnormally reduced due to the transmembrane transport of extracellular glucose promoted by exercise. An exponential compensation mechanism is needed to restore its true level. The specific compensation formula is: Z'_{imp}=Z_{imp}·exp(0.28·W_motion), where the coefficients 0.35 and 0.28 are derived from the regression fitting results of large-scale human exercise experiments.

[0106] It should be noted that these two compensations are not performed in isolation, but rather linked together through motion interference assessment factors, because there is a physiological coupling mechanism between changes in blood flow and glucose diffusion rate. This coupled compensation coordinates the correction amplitude of different signals through a unified interference factor, avoiding overcompensation or undercompensation, and enabling electrochemical and impedance signals to quickly return to their corresponding relationships under steady-state physiological conditions after exercise.

[0107] Step S703: Calculate the temperature and humidity combined interference index based on the temperature and humidity data in the environmental parameter signal; Specifically, ambient temperature directly affects the catalytic activity of glucose oxidase through the Arrhenius effect; deviations from the ideal reaction temperature of 37°C cause nonlinear drift in the current signal. Ambient humidity, on the other hand, affects the conduction path and impedance characteristics of electrical signals in skin tissue by altering the hydration level of the stratum corneum.

[0108] In this embodiment, the temperature and humidity combined interference index identifies a synergistic interference mechanism between temperature and humidity: in high-temperature environments, high humidity exacerbates heat conduction efficiency and accelerates enzyme activity decay; while in low-temperature environments, high humidity may cause condensation on the skin surface, altering electrode contact characteristics. Through weighted combined calculation, this index transforms the originally independent environmental parameters into a comprehensive interference index, more accurately reflecting the degree of influence of the combined effect of temperature and humidity on the monitoring system in the real environment, and providing more comprehensive environmental context information for subsequent compensation.

[0109] Step S704: Call the individual metabolic correction parameters in the pre-stored user historical metabolic data, adaptively scale the feature vector after motion interference compensation, and generate a personalized scaled feature vector. Among them, the insulin sensitivity coefficient (ISF) reflects the ability of a unit of insulin to regulate blood glucose, and the carbohydrate ratio (CR) characterizes an individual's metabolic response to dietary carbohydrates. These two parameters together define an individual's unique metabolic characteristics. Scaling the compensated feature vector based on these parameters essentially embeds the individual's physiological metabolic background at the signal level.

[0110] For example, users with low insulin sensitivity have a relatively slow glucose metabolism rate, and the electrochemical signals corresponding to the same blood glucose level may exhibit different response characteristics. ISF-weighted scaling can correct for signal bias caused by this difference in basal metabolism. This adaptive scaling mechanism enables the system to distinguish the signal characteristics of different types of diabetic patients (such as type 1 and type 2) or even different stages of the disease, realizing a shift from a "general-to-the-group" to an "individual-specific" calibration strategy.

[0111] Step S705: The personalized scaling feature vector and the temperature and humidity combined interference index are fused by multiple regression to generate an environmental adaptive feature vector. Specifically, by constructing an input matrix containing personalized scaled feature vectors and a combined temperature and humidity interference index, and loading user-specific regression weight parameters, the system can learn an individual's signal response patterns under specific environmental conditions. The regression weight matrix essentially encodes the interaction patterns between the user's physiological characteristics and environmental factors. For example, due to the thermal insulation effect of subcutaneous fat, obese users may have significantly higher signal sensitivity to ambient temperature than lean users.

[0112] Understandably, this step no longer treats environmental interference as a fixed bias independent of the individual, but rather dynamically correlates it with the user's metabolic characteristics, thereby more accurately predicting the drift patterns and amplitudes of a particular individual's signal under specific environmental conditions. This personalized environment-physiology coupling model provides stronger adaptability for blood glucose monitoring in complex environments.

[0113] Step S706, outputting the environment adaptive feature vector to the step of dynamically compensating the fused feature vector.

[0114] The environment-adaptive feature vector not only contains multimodal signal features fused with motion disturbance correction, individual metabolic scaling, and environmental disturbance, but also carries parameter correlation information from these correction processes. When this vector is input into the dynamic compensation step, it can synergize with the temperature compensation equation and the transfer learning parameter tuning mechanism.

[0115] In this embodiment, the environmental adaptive feature vector already contains correction information for the combined temperature and humidity disturbances. This information can be used to form a complementary verification with the Arrhenius temperature compensation equation, avoiding repeated corrections. At the same time, the individual metabolic features embedded in the feature vector can also provide prior knowledge for the transfer learning module, accelerating the model's adaptation speed to the metabolic patterns of specific users.

[0116] In the above implementation, the three types of interfering factors—exercise, environment, and metabolism—are transformed from being treated independently to being analyzed in a coordinated manner. By quantifying their interaction effects, more refined signal compensation is achieved. Particularly in addressing individual differences, the system intelligently corrects signal variations in different users exhibiting the same physiological state through metabolic parameter embedding and personalized regression weights. This improves monitoring robustness in complex scenarios, maintaining clinical-grade accuracy under conditions of high-intensity exercise, extreme environments, or metabolic fluctuations, providing reliable technical support for the full life-cycle management of diabetic patients.

[0117] This application also discloses a high-precision blood glucose monitoring system based on multimodal data fusion.

[0118] A high-precision blood glucose monitoring system based on multimodal data fusion specifically includes: The multi-source data acquisition module is used to simultaneously acquire the user's electrochemical current signal, bioimpedance spectrum signal, physiological parameter signal, environmental parameter signal and user behavior signal through multi-source sensors, and generate a multi-source data set with timestamps; The data processing module is used to perform time alignment and preprocessing on multi-source data sets based on the sampling time sequence of electrochemical current signals, and generate a standardized data matrix. The multimodal feature extraction module is used to extract the time-domain features of electrochemical signals, the frequency-domain features of bioimpedance signals, and the nonlinear features of physiological parameter signals from the standardized data matrix, and to jointly generate environmental-behavior cross features by combining environmental parameter signals and user behavior signals to obtain an initial feature vector set. The feature dimensionality reduction and selection module is used to perform principal component analysis to reduce the dimensionality of the initial feature vector set, and to filter the target feature subset through a feature selection algorithm to generate a fused feature vector. The dynamic feature compensation module is used to dynamically compensate the fused feature vector based on real-time ambient temperature data and pre-stored user historical metabolic data to generate a calibration feature vector. The blood glucose prediction module is used to input the calibration feature vector into the pre-trained blood glucose prediction model and output the blood glucose prediction value and the corresponding confidence score. The safety control trigger module is used to trigger a safety control command when the blood glucose prediction value exceeds the preset safety range or the confidence score is lower than the preset confidence threshold. The model parameter update module is used to update the parameters of the blood glucose prediction model based on the execution results of safety control commands and user feedback data.

[0119] The high-precision blood glucose monitoring system based on multimodal data fusion according to the embodiments of this application can implement any of the above methods, and the specific working process of each module in the system can refer to the corresponding process in the above method embodiments.

[0120] In the several embodiments provided in this application, it should be understood that the provided methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for example, the division of a certain module is merely a logical functional division, and in actual implementation there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0121] This application also discloses a computer device.

[0122] A computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a high-precision blood glucose monitoring method based on multimodal data fusion as described above.

[0123] This application also discloses a computer-readable storage medium.

[0124] A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above in any of the high-precision blood glucose monitoring methods based on multimodal data fusion.

[0125] The computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device; the program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0126] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A high-precision blood glucose monitoring method based on multimodal data fusion, characterized in that, The high-precision blood glucose monitoring method includes: By synchronously collecting users' electrochemical current signals, bioimpedance spectrum signals, physiological parameter signals, environmental parameter signals, and user behavior signals through multiple source sensors, a multi-source data set with timestamps is generated. Based on the sampling timing of the electrochemical current signal, the multi-source data set is time-aligned and preprocessed to generate a standardized data matrix; The time-domain features of electrochemical signals, the frequency-domain features of bioimpedance signals, and the nonlinear features of physiological parameter signals are extracted from the standardized data matrix, and the environmental parameter signals and user behavior signals are combined to generate environmental-behavior cross features, thus obtaining an initial feature vector set. Principal component analysis is performed on the initial feature vector set to reduce dimensionality, and a target feature subset is selected by a feature selection algorithm to generate a fused feature vector. Based on real-time ambient temperature data and pre-stored user historical metabolic data, the fused feature vector is dynamically compensated to generate a calibration feature vector. The calibration feature vector is input into the pre-trained blood glucose prediction model, which outputs the predicted blood glucose value and the corresponding confidence score. When the predicted blood glucose value exceeds the preset safety range or the confidence score is lower than the preset confidence threshold, a safety control command is triggered. The parameters of the blood glucose prediction model are updated based on the execution results of the safety control instructions and user feedback data.

2. The high-precision blood glucose monitoring method based on multimodal data fusion according to claim 1, characterized in that, The steps of time alignment and preprocessing of the multi-source data set to generate a standardized data matrix, based on the sampling time sequence of the electrochemical current signal, include: Retrieve a multi-source data set with timestamps; Using the sampling time point of the electrochemical current signal as the reference time series, time alignment is performed on the sampling points of other signals in the multi-source data set to generate a time-aligned data set; The time-aligned dataset is locally filtered using a sliding window mechanism, multi-scale noise removal is performed to eliminate high-frequency interference, missing data points are detected and interpolation methods are used to fill in the complete data sequence, resulting in a preprocessed dataset. Perform standardization operations on the preprocessed dataset to generate a standardized data matrix.

3. The high-precision blood glucose monitoring method based on multimodal data fusion according to claim 1, characterized in that, The steps of performing principal component analysis to reduce the dimensionality of the initial feature vector set, and then filtering the target feature subset using a feature selection algorithm to generate a fused feature vector include: Calculate the covariance matrix of the initial eigenvector set, perform eigenvalue decomposition, and arrange the eigenvectors in descending order of eigenvalues; Feature vectors whose cumulative contribution rate exceeds a preset contribution rate threshold are selected as principal component basis vectors. The initial feature vectors are then projected onto the low-dimensional space spanned by the principal component basis vectors to obtain the low-dimensional feature space. In the low-dimensional feature space, a subset of target features that meet preset conditions and are correlated with blood glucose concentration are selected using a feature selection algorithm; The features in the target feature subset are weighted and combined according to preset weight coefficients to output a fused feature vector.

4. The high-precision blood glucose monitoring method based on multimodal data fusion according to claim 1, characterized in that, The steps for dynamically compensating the fused feature vector and generating a calibration feature vector based on real-time ambient temperature data and pre-stored user historical metabolic data include: Acquire fused feature vectors, real-time ambient temperature data, and user historical metabolic data; Calculate the temperature sensitivity compensation coefficient based on the real-time ambient temperature data. The individual metabolic parameters from the user's historical metabolic data are used to generate a metabolic adaptive scaling factor; The fused feature vector and the temperature sensitivity compensation coefficient are subjected to a first operation to generate a temperature compensation feature vector; The temperature compensation feature vector is combined with the metabolic adaptive scaling factor to generate a calibration feature vector.

5. A high-precision blood glucose monitoring method based on multimodal data fusion according to claim 1, characterized in that, The steps of inputting the calibration feature vector into the pre-trained blood glucose prediction model and outputting the predicted blood glucose value and the corresponding confidence score include: The calibration feature vector is input into the pre-trained blood glucose prediction model, and spatial dimension features are extracted through feature space decomposition. Perform time-series analysis on the spatial dimensional features to generate predicted blood glucose concentration values; The uncertainty measure of features is calculated based on the hidden layer state within the blood glucose prediction model. A confidence score is generated based on the aforementioned feature uncertainty measure; Output the predicted blood glucose value and the corresponding confidence score.

6. A high-precision blood glucose monitoring method based on multimodal data fusion according to claim 1, characterized in that, The steps for updating the parameters of the blood glucose prediction model based on the execution results of the safety control commands and user feedback data include: Obtain the execution results of safety control commands, user feedback data, and venous blood calibration measurement values; An incremental training sample set is constructed based on the execution results and user feedback data. The actual blood glucose labels of the incremental training sample set are labeled according to the venous blood calibration measurement values; Extract the current parameter state of the blood glucose prediction model to generate an initial parameter set; Based on the labeled incremental training sample set, the initial parameter set is optimized with the objective function of minimizing prediction error and risk penalty; The optimized parameter set is then updated in the blood glucose prediction model.

7. A high-precision blood glucose monitoring method based on multimodal data fusion according to any one of claims 1 to 6, characterized in that, After performing principal component analysis to reduce the dimensionality of the initial feature vector set and filtering the target feature subset using a feature selection algorithm to generate a fused feature vector, the method further includes: The system acquires user motion state recognition results in real time. When high-intensity motion is detected, the collaborative correction mechanism is activated to generate motion interference evaluation factors. Based on the motion interference assessment factor, the time-domain feature component of the electrochemical signal and the frequency-domain feature component of the bioimpedance signal in the fused feature vector are coupled and compensated to generate a motion interference compensated feature vector. Based on the temperature and humidity data in the environmental parameter signals, calculate the temperature and humidity combined interference index; The individual metabolic correction parameters in the pre-stored user historical metabolic data are called to adaptively scale the motion interference compensated feature vector to generate a personalized scaled feature vector. The personalized scaling feature vector is fused with the temperature and humidity combined interference index through multivariate regression to generate an environmental adaptive feature vector. The step of dynamically compensating the fused feature vector by outputting the environment adaptive feature vector.

8. A high-precision blood glucose monitoring system based on multimodal data fusion, characterized in that, The monitoring system includes: The multi-source data acquisition module is used to simultaneously acquire the user's electrochemical current signal, bioimpedance spectrum signal, physiological parameter signal, environmental parameter signal and user behavior signal through multi-source sensors, and generate a multi-source data set with timestamps; The data processing module is used to perform time alignment and preprocessing on the multi-source data set based on the sampling time sequence of the electrochemical current signal to generate a standardized data matrix. The multimodal feature extraction module is used to extract the time-domain features of electrochemical signals, the frequency-domain features of bioimpedance signals, and the nonlinear features of physiological parameter signals from the standardized data matrix, and to jointly generate environment-behavior cross features by combining the environmental parameter signals and user behavior signals to obtain an initial feature vector set. The feature dimensionality reduction and selection module is used to perform principal component analysis to reduce the dimensionality of the initial feature vector set, and to filter the target feature subset through a feature selection algorithm to generate a fused feature vector. The dynamic feature compensation module is used to dynamically compensate the fused feature vector based on real-time ambient temperature data and pre-stored user historical metabolic data to generate a calibration feature vector. The blood glucose prediction module is used to input the calibration feature vector into the pre-trained blood glucose prediction model and output the blood glucose prediction value and the corresponding confidence score. The safety control trigger module is used to trigger a safety control command when the blood glucose prediction value exceeds the preset safety range or the confidence score is lower than the preset confidence threshold. The model parameter update module is used to update the parameters of the blood glucose prediction model based on the execution results of the safety control instructions and user feedback data.

9. A computer device, characterized in that: The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 7.

Images