Noninvasive blood glucose dynamic monitoring method and system for wearable device

Abstract

The invention relates to a noninvasive blood glucose dynamic monitoring method and system for wearable equipment. According to the method, physiological and behavior signals of a user are continuously collected and fused into a unified feature vector, and the vector is synchronously analyzed by utilizing a cascaded neural network model subjected to personalized training, so that a real-time blood glucose value is predicted; more importantly, the instantaneous influence of a specific behavior on blood glucose is quantified, the anti-fact blood glucose state is calculated if the behavior does not exist, then a process contribution value representing the health effort degree of a user is calculated according to the attribution information, and finally a monitoring report fusing a numerical result and behavior attribution is generated. According to the scheme, interpretable attribution of blood glucose fluctuation is achieved, a monitoring focus is extended from a single blood glucose result to a behavior process causing the result, and an accurate data basis is provided for management and excitation of a fair, anti-speculation and health encouraging process.

Description

Technical Field

[0001] This invention belongs to the field of intelligent health monitoring, and in particular relates to a non-invasive dynamic blood glucose monitoring method and system for wearable devices. Background Technology

[0002] Existing health management platforms, especially those dealing with chronic diseases such as diabetes, generally rely on continuous or periodic monitoring of users' physiological indicators. Among these, non-invasive blood glucose monitoring technology has become a research hotspot due to its convenience. The mainstream method uses wearable devices to collect single physiological signals such as photoplethysmography (PPG) waves and then uses algorithmic models to estimate blood glucose levels.

[0003] However, this type of technology has significant limitations. On the one hand, its monitoring accuracy is easily affected by factors such as users' daily activities and environmental changes, and its stability needs improvement. On the other hand, and more fundamentally, its output is only an isolated blood glucose value, forming a "data black box." This limitation produces serious negative effects when combined with incentive mechanisms: platforms usually implement rewards or punishments based on these final blood glucose readings, forming a purely results-oriented model. This model implicitly encourages users' short-term speculative behavior, such as obtaining a "good-looking" reading through unhealthy extreme methods (such as excessive dieting) before key measurement points, which completely violates the long-term purpose of health management. At the same time, it cannot fairly distinguish whether blood glucose fluctuations are due to users' controllable health behaviors (such as regular exercise) or uncontrollable factors (such as illness, stress, or drug reactions). This "one-size-fits-all" evaluation method is not only unscientific, but also discourages users, especially new users or those in poor condition, leading them to abandon continuous management. Therefore, the core contradiction in the current technological ecosystem lies in the huge gap between the limited information provided by monitoring methods (only numerical results) and the rich decision-making basis required for health management (understanding the causes of results). This gap makes any incentive mechanism built on it potentially deviate from the right direction and even lead to harmful behaviors.

[0004] Therefore, there is an urgent need for a monitoring method that can penetrate the "black box," provide reliable blood glucose estimates, and explain the causal relationship between blood glucose changes and users' specific behaviors, thereby laying the technological foundation for building a truly effective, fair, and long-term health-friendly intelligent management platform. Summary of the Invention

[0005] Therefore, it is necessary to provide a non-invasive dynamic blood glucose monitoring method and system for wearable devices to address the aforementioned technical problems.

[0006] In a first aspect, this application provides a method for noninvasive dynamic blood glucose monitoring in wearable devices, comprising:

[0007] S1. Collect the user's multimodal physiological signals and inertial signals through wearable devices, perform behavior recognition on the inertial signals, and obtain the user's behavior sequence;

[0008] S2. Extract blood glucose-related features from multimodal physiological signals and construct physiological feature vectors based on these features; extract behavioral context information from user behavior sequences and construct behavioral feature vectors based on this context information; align and concatenate the physiological feature vectors and behavioral feature vectors in time sequence to generate a fused feature vector.

[0009] S3. Input the real-time generated fusion feature vector into the pre-trained cascaded neural network model and output the real-time blood glucose estimate, the behavior influence coefficient reflecting the degree of influence of a specific behavior on the instantaneous change in blood glucose, and the counterfactual blood glucose estimate assuming that no specific behavior has occurred.

[0010] The cascaded neural network model is trained based on the user's historical data and fine-tuned with the user's personalized data. The historical data includes historical multimodal physiological signals and historical inertial signals. The cascaded neural network model includes a shared base layer, a blood glucose prediction branch, and a behavioral attribution branch. The shared base layer is used to receive and store the fused feature vector. The blood glucose prediction branch is connected to the output of the shared base layer and is used to output the real-time blood glucose estimate. The behavioral attribution branch is connected to the output of the shared base layer and is used to output the influence coefficient and the counterfactual blood glucose estimate.

[0011] S4. Quantify the process contribution value of user health behavior based on the behavior impact coefficient and counterfactual blood glucose estimate; generate a monitoring report with behavior attribution results based on the process contribution value and real-time blood glucose estimate; wherein, the process contribution value represents the degree of positive effort that the user makes in performing specific health behaviors on their own blood glucose status.

[0012] Secondly, this application also provides a non-invasive dynamic blood glucose monitoring system for wearable devices, for implementing the method described in the first aspect, the system comprising:

[0013] The multi-source signal sensing and behavior analysis module is used to collect the user's multimodal physiological signals and inertial signals, perform behavior recognition on the inertial signals, and obtain the user's behavior sequence.

[0014] The feature fusion generation module is used to extract blood glucose-related features from multimodal physiological signals and construct physiological feature vectors based on blood glucose-related features; extract behavioral context information from user behavior sequences and construct behavioral feature vectors based on behavioral context information; and concatenate the physiological feature vectors and behavioral feature vectors in temporal alignment to generate a fused feature vector.

[0015] The cascaded neural network analysis module is used to input the real-time generated fused feature vector into the pre-trained cascaded neural network model and output the real-time blood glucose estimate, the behavior influence coefficient reflecting the degree of influence of a specific behavior on the instantaneous change in blood glucose, and the counterfactual blood glucose estimate assuming that no specific behavior has occurred.

[0016] The cascaded neural network model is trained based on the user's historical data and fine-tuned with the user's personalized data. The historical data includes historical multimodal physiological signals and historical inertial signals. The cascaded neural network model includes a shared base layer, a blood glucose prediction branch, and a behavioral attribution branch. The shared base layer is used to receive and store the fused feature vector. The blood glucose prediction branch is connected to the output of the shared base layer and is used to output the real-time blood glucose estimate. The behavioral attribution branch is connected to the output of the shared base layer and is used to output the influence coefficient and the counterfactual blood glucose estimate.

[0017] The health attribution report generation module is used to quantify the process contribution value of user health behaviors based on the behavior impact coefficient and counterfactual blood glucose estimates; based on the process contribution value and real-time blood glucose estimates, it generates a monitoring report with behavioral attribution results; wherein, the process contribution value represents the degree of positive effort that users exert on their own blood glucose status by performing specific health behaviors.

[0018] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a non-invasive dynamic blood glucose monitoring method for a wearable device as described in the first aspect.

[0019] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a non-invasive dynamic blood glucose monitoring method for a wearable device as described in the first aspect.

[0020] The aforementioned non-invasive dynamic blood glucose monitoring method and system for wearable devices continuously collects users' physiological and behavioral signals and fuses them into a unified feature vector. A personalized, cascaded neural network model is then used to synchronously analyze this vector. This not only predicts real-time blood glucose values ​​but, more importantly, quantifies the instantaneous impact of specific behaviors on blood glucose and calculates the counterfactual blood glucose state without these behaviors. Based on this attribution information, a process contribution value representing the user's health effort is calculated, ultimately generating a monitoring report that integrates numerical results and behavioral attributions. This approach achieves interpretable attribution of blood glucose fluctuations, extending the monitoring focus from a single blood glucose result to the behavioral process leading to that result. This provides a precise data foundation for implementing fair, anti-speculation, and health-encouraging management and incentives. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart illustrating a non-invasive dynamic blood glucose monitoring method for wearable devices provided by the present invention;

[0023] Figure 2 This is a flowchart illustrating the process contribution value of quantifying user health behavior in an optional embodiment of the present invention.

[0024] Figure 3 This is a schematic diagram of a non-invasive dynamic blood glucose monitoring system for wearable devices provided by the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0026] refer to Figure 1 The document presents a flowchart illustrating a non-invasive dynamic blood glucose monitoring method for wearable devices, as provided in this application. The method includes the following steps:

[0027] S1. Collect the user's multimodal physiological signals and inertial signals through wearable devices, perform behavior recognition on the inertial signals, and obtain the user's behavior sequence.

[0028] Specifically, wearable devices can be smartwatches. The selection of multimodal physiological signals revolves around the physiological correlation of blood glucose monitoring, mainly including photoplethysmography (PPG) signals, skin conductance signals, body surface temperature signals, and respiratory rate signals. These signals are collected collaboratively by dedicated sensors integrated into the wearable device. PPG signals are achieved through a dual-wavelength photoelectric sensor. This sensor emits specific wavelengths of light into the skin, receives the light signals reflected by blood and tissue, and converts them into voltage signals. Its design is based on the fact that changes in blood glucose concentration affect the absorption and reflection characteristics of specific wavelengths of light, which are reflected in the morphological changes of the pulse wave. Skin conductance signals are collected through electrode pads in contact with the skin, reflecting the intensity of sympathetic nerve activity, which is indirectly related to metabolic changes caused by blood glucose fluctuations. Body surface temperature signals are collected through a high-precision temperature sensor, capturing subtle fluctuations in body temperature due to metabolic state and environmental changes. Since blood glucose metabolism involves energy release, it affects body surface temperature. The respiratory rate signal does not require an additional sensor. It is obtained by performing frequency domain analysis on the baseline drift characteristics of the photoplethysmography pulse wave signal, separating the respiratory-related frequency components, and then calculating the signal, thus achieving efficient signal acquisition.

[0029] Inertial signals are acquired through a six-axis inertial measurement unit (IMU) built into the wearable device. This unit integrates a three-axis accelerometer and a three-axis gyroscope to capture the linear and angular accelerations of the user's body movements, providing core data support for behavior recognition. To ensure the temporal consistency of multimodal signals, a hardware synchronous triggering mechanism is used in the acquisition process. The device's built-in microcontroller centrally controls the startup and data acquisition process of all sensors. The output data of each sensor is accompanied by a high-precision timestamp generated by the microcontroller. Subsequent timestamp alignment processing eliminates the acquisition delay caused by differences in the response speeds of different sensors. To address potential signal noise and missing data during data acquisition, targeted preprocessing measures are required: For photoplethysmography (PPG) signals, wavelet denoising algorithms are used to remove baseline drift and power frequency interference, and adaptive thresholding is employed to handle abnormal peak values; for electrodermal response (EDS) signals, median filtering is used to remove random noise, followed by smoothing using a moving average method; for inertial signals, a Kalman filter algorithm is used to dynamically correct measurement errors based on preset motion and observation models, improving signal stability; for signal loss caused by brief sensor detachment from the skin, linear interpolation is used to supplement the missing data, and missing data segments exceeding a set duration are marked as invalid to avoid interfering with subsequent analysis.

[0030] The behavior recognition process is based on preprocessed inertial signals and employs a sliding window segmentation strategy. By appropriately setting the window size and overlap rate, it ensures both real-time behavior recognition and complete capture of the temporal features of the behavior. Multi-dimensional features are extracted from the inertial signals of each data window, covering time-domain, frequency-domain, and time-frequency-domain features. Time-domain features include the mean, variance, kurtosis, skewness, maximum, minimum values, and correlation coefficients between different axes of acceleration and gyroscope data, reflecting the statistical characteristics of the signal. Frequency-domain features are extracted after converting the signal to the frequency domain using Fast Fourier Transform, including peak power spectral density, peak frequency, total energy, and the energy proportion of different frequency bands, capturing the frequency distribution characteristics of the signal. Time-frequency-domain features are extracted using wavelet packet transform to extract the energy entropy and information entropy of each node, characterizing the joint distribution characteristics of the signal in the time and frequency domains.

[0031] The behavior recognition model can employ a CNN-LSTM hybrid neural network architecture. This architecture combines the local feature extraction capability of convolutional neural networks with the temporal dependency capture capability of long short-term memory networks, making it suitable for behavior recognition tasks based on inertial signals. The model input is the extracted multi-dimensional inertial feature vector, and the output is the probability distribution of various behaviors, covering the main categories of users' daily behaviors. Model training consists of two stages: pre-training and fine-tuning. In the pre-training stage, a public behavior dataset is used, divided proportionally into training and validation sets. An appropriate optimizer and loss function are selected, and the model is trained through multiple rounds to master the feature patterns of general behaviors. In the fine-tuning stage, personalized user behavior data is used. When users use the device for the first time, they need to manually annotate their behavior logs for a certain period. After aligning the annotated data with the inertial signals, the parameters of the pre-trained model are adjusted to adapt the model to the individual motion characteristics of the user. During the generation of behavior sequences, the probability distribution of behavior output by the model is processed by a time-series smoothing algorithm to eliminate instantaneous misidentification results. The final output is a behavior sequence sorted by timestamp, which includes the start time, end time and behavior type of each behavior. For identification results with too short a duration, they are merged into the preceding and following behaviors to ensure the rationality and coherence of the behavior sequence.

[0032] S2. Extract blood glucose-related features from multimodal physiological signals and construct physiological feature vectors based on blood glucose-related features; extract behavioral context information from user behavior sequences and construct behavioral feature vectors based on behavioral context information; align and concatenate the physiological feature vectors and behavioral feature vectors in time sequence to generate a fused feature vector.

[0033] Specifically, for the extraction of blood glucose-related features from multimodal physiological signals, targeted solutions need to be designed based on the correlation mechanisms between various signals and blood glucose concentration. After preprocessing, photoplethysmography (PPG) signals contain rich time-domain and frequency-domain features, which are directly related to blood glucose concentration: In terms of time-domain features, pulse wave propagation time is extracted by identifying the systolic peak and diastolic trough of the pulse wave, calculating the relevant time intervals of continuous pulse cycles, and performing statistical processing; pulse amplitude, i.e., the voltage difference between the systolic peak and diastolic trough, is extracted, and relevant statistical features are obtained through multi-cycle statistics; the slopes of the rising and falling edges of the pulse wave are extracted to reflect the rate of change of the pulse wave; the ratio of diastolic to systolic amplitude and the coefficient of variation of the pulse cycle are extracted to comprehensively capture the morphological features of the pulse wave. In terms of frequency domain characteristics, Fourier transform was performed on the photoplethysmography (PPG) signal to extract heart rate variability-related features, including power in different frequency bands, power ratio of frequency bands, and total power spectral density. At the same time, the light absorption coefficient-related features of the dual-wavelength PPG signal were extracted. According to the Lambert-Beer law, there is a positive correlation between the light absorption coefficient and the blood glucose concentration. By calculating the relevant parameters and ratios of the light signals at different wavelengths, features related to blood glucose were obtained.

[0034] Features extracted from the skin conductance response signal include parameters related to skin conductance level and response. Skin conductance level is the DC component of the signal, calculated using a sliding window averaging method. After identifying relevant events using a threshold method, features such as peak value, rise time, and fall time are extracted from the skin conductance response. These features reflect the intensity of sympathetic nerve activity and indirectly correlate with metabolic changes caused by blood glucose fluctuations. Features extracted from the body surface temperature signal include average body temperature, temperature fluctuation amplitude, and temperature change rate, comprehensively capturing the pattern of body temperature changes with metabolic state. Features extracted from the respiratory rate signal include average respiratory rate and respiratory rate variation coefficient, obtained through frequency domain analysis of the photoplethysmography (PPG) signal. Changes in respiratory rate are related to the body's metabolic state and thus indirectly correlated with blood glucose levels.

[0035] All extracted blood glucose-related features need to be normalized to eliminate the influence of different units on subsequent model training. The normalization method used is Min-Max normalization, and its formula is as follows:

[0036]

[0037] in, Represents the original feature value. This represents the minimum value of the feature in the historical dataset. This represents the maximum value of the feature in the historical dataset. This represents the normalized feature value. This formula maps the value range of all features to the same interval, ensuring that different features have equal weight in model training. The normalized features are arranged in a preset order to construct a physiological feature vector. This vector comprehensively represents the user's physiological state and key information related to blood glucose. Each vector is accompanied by a corresponding timestamp, consistent with the timestamp at the time of signal acquisition, ensuring time-series traceability.

[0038] The construction of behavioral feature vectors is based on the extraction of contextual information from user behavior sequences. First, the behavior sequence is structurally parsed, extracting multi-dimensional contextual information from each behavioral element to comprehensively characterize the relationship between behavior and blood glucose changes. The extracted contextual information includes: the current behavior type, serving as the core identifier of the behavior; the behavior duration, quantifying the time span of the behavior; the behavior occurrence time, divided into multiple time periods according to temporal patterns to reflect the temporal background of the behavior; the preceding behavior type, i.e., the behavior before the current behavior, used to characterize the temporal correlation of the behavior; the time interval between the preceding and current behaviors, quantifying the degree of temporal correlation between the two behaviors; subsequent behavior prediction, predicting the next possible behavior type based on the temporal patterns of the behavior sequence to characterize the continuity of the behavior; behavior intensity, quantifying the intensity of motion-type behaviors based on inertial signal-related features, while non-motion-type behaviors are labeled as specific categories; and the association between the behavior and key blood glucose periods, i.e., whether the behavior occurs during a specific period closely related to blood glucose changes, comprehensively capturing the contextual information of the behavior.

[0039] The aforementioned contextual information is quantified, transforming non-numerical information into numerical features that the model can process. Behavior type is converted into a vector form using one-hot encoding; behavior duration is normalized, mapping it to a specific interval based on the extreme values ​​of behavior duration in historical data; behavior occurrence time is converted into a vector using one-hot encoding; preceding behavior type is also one-hot encoded, including cases with "no" preceding behavior; preceding behavior time interval is normalized, mapping it based on the extreme values ​​of intervals in historical data; subsequent behavior prediction uses probabilistic encoding to generate probability vectors for various behaviors; behavior intensity is one-hot encoded to distinguish between exercise intensity levels and non-exercise types; the association between behavior and key blood glucose time periods uses multi-label encoding, with each key time period represented by a specific numerical value indicating whether an association exists. The quantified features are concatenated in a preset order to construct behavior feature vectors. Each behavior feature vector corresponds to a behavior in the behavior sequence, accompanied by the start and end timestamps of that behavior, ensuring the temporal integrity of the behavior information.

[0040] The core of time-series alignment and stitching is to achieve a one-to-one correspondence between physiological feature vectors and behavioral feature vectors in the time dimension. Since the generation frequencies of physiological and behavioral feature vectors differ, time-series alignment needs to be achieved through methods such as interpolation. In practice, the timestamps of each physiological feature vector are traversed, and the target behavior containing that timestamp in the behavioral sequence is searched to obtain the corresponding behavioral feature vector. If the timestamp falls within the interval between two behaviors, linear interpolation is used to generate the behavioral feature vector for that interval, expressed by the following formula:

[0041]

[0042] in, Represents timestamp The corresponding behavioral feature vector, The feature vector representing the previous behavior. The feature vector representing the next behavior. Represents the end timestamp of the previous action. This represents the start timestamp of the next action. If the timestamp is before or after the start of the action sequence, the nearest neighbor padding method is used, filling it with the feature vector of the first or last action. After alignment, the physiological feature vector corresponding to each timestamp is concatenated with the action feature vector in sequence to generate a fused feature vector. Each fused feature vector is accompanied by a corresponding timestamp to ensure temporal consistency. Finally, a sequence of fused feature vectors ordered by timestamps is formed, providing the data foundation for the input of subsequent cascaded neural network models.

[0043] S3. Input the real-time generated fusion feature vector into the pre-trained cascaded neural network model, and output the real-time blood glucose estimate, the behavior influence coefficient reflecting the degree of influence of a specific behavior on the instantaneous change in blood glucose, and the counterfactual blood glucose estimate assuming that no specific behavior has occurred.

[0044] The cascaded neural network model is trained based on the user's historical data and fine-tuned with the user's personalized data. The historical data includes historical multimodal physiological signals and historical inertial signals. The cascaded neural network model includes a shared base layer, a blood glucose prediction branch, and a behavioral attribution branch. The shared base layer is used to receive and store the fused feature vector. The blood glucose prediction branch is connected to the output of the shared base layer and is used to output the real-time blood glucose estimate. The behavioral attribution branch is connected to the output of the shared base layer and is used to output the influence coefficient and the counterfactual blood glucose estimate.

[0045] Specifically, the cascaded neural network model adopts an architecture design of "shared base layer + dual task branches". The shared base layer is responsible for extracting common key features from the fused features, while the dual task branches focus on blood glucose prediction and behavioral attribution tasks respectively, realizing feature reuse and task collaboration, improving the efficiency of model operation while ensuring the performance of the two tasks.

[0046] The shared base layer is designed to address the characteristics of temporal fusion feature vectors. It employs a structure combining one-dimensional convolutional layers and fully connected layers to fully extract temporal dependencies and nonlinear correlations from the fusion features. The core function of the one-dimensional convolutional layer is to capture local temporal features within the fusion features. Feature extraction is progressively deepened through multiple convolutional operations. Each convolutional layer is followed by a batch normalization layer and a dropout layer. The batch normalization layer standardizes the input data, accelerating model training convergence and improving model stability. The dropout layer randomly discards some neurons to prevent overfitting during training. A fully connected layer follows the convolutional layers to further extract higher-order nonlinear features from the fusion features, forming a dense feature vector. This feature vector contains key information related to blood glucose prediction and behavioral attribution, providing a unified, high-quality feature input for the two subsequent task branches. The parameters of the shared base layer are updated together during the training of both task branches, ensuring that the extracted general features can simultaneously meet the needs of both blood glucose prediction and behavioral attribution tasks, achieving efficient feature utilization.

[0047] The blood glucose prediction branch, as a regression task branch, focuses on learning the mapping relationship between dense feature vectors and blood glucose values, outputting real-time blood glucose estimates. This branch employs a fully connected layer structure. The first fully connected layer further extracts specific features related to blood glucose prediction, introducing a non-linear transformation through an activation function to enhance the model's ability to fit complex mapping relationships. This layer is followed by batch normalization and dropout layers to ensure the model's generalization ability. The second fully connected layer uses a linear activation function. Since blood glucose values ​​are continuous variables, a linear activation function ensures that the output value is not limited in range and can accurately match the actual fluctuation range of blood glucose values. The loss function for this branch is the mean squared error loss function, expressed as:

[0048]

[0049] in, The mean squared error loss value representing the blood glucose prediction branch. Represents the number of training samples. Representing the The actual blood glucose value of each sample The representative model predicted the first The model calculates the blood glucose values ​​of each sample. By minimizing this loss function, the model can progressively learn the precise mapping relationship between the dense feature vector and the true blood glucose value, thereby improving the accuracy of blood glucose prediction.

[0050] The behavior attribution branch comprises two parallel subtasks, outputting a behavior impact coefficient and a counterfactual blood glucose estimate, respectively. Its design is based on the latent outcome framework in causal inference, learning the blood glucose differences between when the behavior occurs and when it doesn't, thus quantifying the causal relationship between behavior and blood glucose changes. The behavior impact coefficient sub-branch employs a fully connected layer structure, using an activation function to restrict the output range to a specific interval. This interval is designed to clearly distinguish between the positive and negative impacts of behavior on blood glucose, and the absolute value of the output characterizes the degree of influence of the behavior on blood glucose. The loss function for this sub-task is the mean squared error loss function, labeled with a benchmark value of behavior impact calculated based on the actual blood glucose value and behavior logs. This benchmark value is obtained by matching blood glucose data with and without the target behavior under similar physiological states, used to quantify the degree of influence of behavior on blood glucose in real-world scenarios. The counterfactual blood glucose estimate sub-branch has the same structure as the blood glucose prediction branch, employing a fully connected layer and a linear activation function. It outputs a blood glucose estimate assuming the current behavior did not occur. Its loss function is also the mean squared error loss function, labeled with the actual blood glucose value when the behavior did not occur. The total loss function of the behavioral attribution branch is the weighted sum of the loss functions of the two subtasks, and its formula is as follows:

[0051]

[0052] in, The total loss value representing the behavioral attribution branch. The mean squared error loss value of the subtask representing the influence coefficient of the representative behavior. This represents the mean squared error loss value for the counterfactual blood glucose estimation subtask. By jointly minimizing this total loss function, the model can accurately quantify the impact of behavior on blood glucose, while also accurately estimating blood glucose levels in counterfactual scenarios.

[0053] The model training process is divided into two stages: pre-training and personalized fine-tuning. The pre-training stage aims to improve the model's generalization ability, enabling it to adapt to the common characteristics of different users. The personalized fine-tuning stage optimizes model parameters to improve the model's adaptability to specific users based on individual differences. The pre-training stage is based on a large-scale historical dataset, which covers long-term monitoring data from multiple users, including multimodal physiological signals, inertial signals, corresponding invasive blood glucose measurement labels, and behavioral logs. The dataset is divided by user and cross-validation is used to ensure that the pre-trained model does not depend on specific user data and has good generalization ability. The training process adopts a multi-task joint training strategy. The total loss function is the weighted sum of the loss functions of the blood glucose prediction branch and the behavior attribution branch, and its formula is expressed as:

[0054]

[0055] in, This represents the total loss value during model training. The weights representing the blood glucose prediction branch loss function, The weights represent the loss function of the behavioral attribution branch. These weights are determined through hyperparameter tuning to highlight the core role of blood glucose prediction while ensuring the accuracy of behavioral attribution. During training, an appropriate optimizer and learning rate scheduling strategy are selected. The model gradually converges through multiple training rounds. When the total loss function on the validation set fails to decrease for several consecutive training cycles, an early stopping strategy is used to terminate training, and the pre-trained model parameters are saved.

[0056] In the personalized fine-tuning phase, model parameters are adjusted using a small amount of personalized data to quickly adapt the model to the individual characteristics of new users, taking into account their individual differences. After initial device use, new users are required to wear the device continuously for a certain period. During this time, the device collects the user's multimodal physiological and inertial signals, and the user is required to complete a certain number of invasive blood glucose measurements and accurately record behavioral logs, constructing a personalized dataset covering the user's typical behaviors and physiological states. During fine-tuning, pre-trained model parameters are loaded, and some low-level parameters of the shared base layer are frozen to preserve the model's general feature extraction capabilities. Only the parameters of the later layers of the shared base layer and all parameters of the two task branches are updated. The optimizer used in the fine-tuning phase is the same as that used in the pre-training phase, with a small learning rate to avoid excessive parameter updates that could destroy the general features learned in the pre-training phase. The training cycle is set to a small number of rounds, achieving rapid adaptation of the model to the individual characteristics of users through fine-tuning with a small amount of data. The fine-tuned model is stored in the local edge computing module of the wearable device, supporting real-time inference. The fused feature vectors are input into the model frame by frame in chronological order to ensure the real-time nature of the output results. The output real-time blood glucose estimates, behavioral impact coefficients, and counterfactual blood glucose estimates are all accompanied by corresponding timestamps, maintaining temporal consistency with the input feature vectors and providing data support for the calculation of contribution values ​​in subsequent processes.

[0057] S4. Quantify the process contribution value of user health behavior based on the behavior impact coefficient and counterfactual blood glucose estimate; generate a monitoring report with behavior attribution results based on the process contribution value and real-time blood glucose estimate; wherein, the process contribution value represents the degree of positive effort that the user makes in performing specific health behaviors on their own blood glucose status.

[0058] Specifically, quantifying the process contribution value first requires clarifying the definition criteria for healthy behaviors. Based on clinical guidelines and relevant research findings in diabetes health management, healthy behaviors mainly include regular exercise, timely medication, healthy diet, and sufficient sleep. These behaviors have been clinically validated to have a positive effect on blood glucose control. Unhealthy behaviors include sedentary lifestyles, overeating, and staying up late, which may have a negative impact on blood glucose control. Therefore, only the process contribution value of healthy behaviors is calculated, while the process contribution value of unhealthy behaviors is uniformly set to a specific benchmark value to avoid inappropriate assessments due to unhealthy behaviors. The quantification logic is based on the contribution assessment framework in causal inference, comprehensively considering the direction and magnitude of the behavioral influence coefficient, the difference between real-time blood glucose estimates and counterfactual blood glucose estimates, and the clinical standards for blood glucose target ranges. Through multi-step calculations, the degree of positive effort is accurately quantified.

[0059] The first step is to determine the target blood glucose range for the current monitoring period. Based on clinical guidelines, different target ranges are defined for different time periods, covering key periods such as fasting, specific postprandial periods, and bedtime. The target ranges for different time periods are set based on the physiological patterns of blood glucose metabolism and clinical health standards. This target range is stored in the wearable device's configuration file, allowing users to make personalized adjustments under the guidance of a doctor. The adjustment process requires specific authorization verification to ensure the scientific validity and seriousness of the target range.

[0060] The second step is to calculate the blood glucose deviation, which includes real-time blood glucose deviation and counterfactual blood glucose deviation. Real-time blood glucose deviation characterizes the degree to which the current blood glucose value deviates from the target range for the corresponding time period. If the real-time blood glucose estimate is within the target range, the real-time blood glucose deviation is zero; if the real-time blood glucose estimate is higher than the upper limit of the target range, the real-time blood glucose deviation is the difference between the real-time blood glucose estimate and the upper limit of the target range; if the real-time blood glucose estimate is lower than the lower limit of the target range, the real-time blood glucose deviation is the difference between the lower limit of the target range and the real-time blood glucose estimate. Counterfactual blood glucose deviation characterizes the degree to which the blood glucose value deviates from the target range if the current healthy behavior had not occurred. Its calculation logic is the same as real-time blood glucose deviation, only the real-time blood glucose estimate is replaced with the counterfactual blood glucose estimate.

[0061] The third step is to calculate the effectiveness coefficient of the behavioral impact, which is expressed by the following formula:

[0062]

[0063] in, The effectiveness coefficient representing the influence of behavior. This represents a counterfactual blood glucose bias. This represents the real-time blood glucose deviation. The value represents the minimum value, used to avoid the calculation anomaly of a zero denominator when the counterfactual blood glucose deviation is zero. The effectiveness coefficient ranges from zero to one. When the counterfactual blood glucose deviation is greater than the real-time blood glucose deviation, the effectiveness coefficient is positive, indicating that the health behavior effectively reduces the blood glucose deviation; when the counterfactual blood glucose deviation is less than or equal to the real-time blood glucose deviation, the effectiveness coefficient is zero, indicating that the health behavior has no effect on improving the blood glucose deviation, possibly due to insufficient behavior intensity, inappropriate timing, or other factors.

[0064] The fourth step is to calculate the contribution value of the process, and the formula is as follows:

[0065]

[0066] in, Representative process contribution value, The effectiveness coefficient representing the influence of behavior. Representative behavior influence coefficient, The absolute value of the behavior impact coefficient characterizes the intensity of the behavior's effect on blood glucose. Multiplying by 100% converts the process contribution value into a percentage, making it easier for users to understand and perceive. The process contribution value is calculated cumulatively over time windows. Each health behavior is divided into multiple equal time windows based on its duration, and the contribution value is calculated independently for each time window. The final cumulative process contribution value is the arithmetic mean of the contribution values ​​from all time windows. This avoids interference from data fluctuations at a single time point, ensuring the stability and objectivity of the assessment. For the same health behavior occurring continuously, the cumulative contribution value is the average of the contribution values ​​from all corresponding time windows. For the same type of health behavior occurring intermittently, the cumulative contribution value for each segment of the behavior is calculated separately, allowing users to understand the effectiveness of the behavior at different times.

[0067] The monitoring report needs to integrate process contribution values, real-time blood glucose estimates, behavioral impact coefficients, counterfactual blood glucose estimates, and behavioral attribution results. It should be presented using a combination of textual descriptions and data charts, balancing comprehensiveness, logical clarity, and ease of understanding, while strictly adhering to the rigor requirements of medical and health data. The core content of the report includes several parts. The first part is a summary of core monitoring data, listing key data within the current monitoring period in chronological order. This includes real-time blood glucose estimates, corresponding behavioral types, behavioral durations, behavioral impact coefficients, counterfactual blood glucose estimates, and process contribution values ​​for each time point. The data presentation must ensure accuracy and standardization, allowing users to quickly access core information.

[0068] The second part is behavioral attribution analysis. For each healthy behavior, combining the process contribution value and blood glucose change trend, it provides clear attribution conclusions and explains in detail the specific impact of the behavior on blood glucose. For healthy behaviors with high process contribution values, its positive effect on blood glucose control is clearly affirmed, and the meaning of the behavior impact coefficient, the difference between the counterfactual blood glucose estimate and the real-time blood glucose estimate, and the degree of positive effort reflected by the process contribution value are explained. For healthy behaviors with low process contribution values, possible reasons are objectively analyzed, including factors such as behavior intensity, timing of implementation, and individual suitability, to avoid generating negative emotions in users. For unhealthy behaviors, their potential impact on blood glucose is clearly pointed out, and targeted prompts are given based on the behavior impact coefficient and blood glucose change trend to guide users to adjust their behavioral habits.

[0069] The third part consists of trend analysis and personalized recommendations. Based on recent historical monitoring data, it displays the changing trends of process contribution values ​​and real-time blood glucose estimates, visually presenting the changing patterns of users' health behavior efforts and blood glucose control effects through charts. Trend analysis conclusions must provide clear judgments based on data changes, including the changing trend of the average contribution value of health behaviors and the improvement in blood glucose control effects, allowing users to clearly perceive the progress of their health management. Personalized recommendations must closely integrate the attribution analysis results and trend analysis conclusions, providing specific and actionable guidance tailored to the user's behavioral characteristics and current blood glucose control status. This includes recommended health behavior types, implementation periods, and intensity control, while also providing improvement directions for any shortcomings identified by the user, helping them optimize their health management plan and improve the effectiveness of long-term health management.

[0070] The monitoring reports support both local storage and cloud synchronization. Local storage uses a standardized document format and stores the reports in the wearable device's storage module, retaining historical reports for a certain period for easy access. Cloud synchronization uses the device's built-in wireless communication module to upload report data to the health management platform in a standardized format, allowing users to view historical reports and trend analyses via mobile applications or computer platforms. Data charts in the reports are in vector format to ensure clear display on different devices, and the text descriptions use concise and standardized medical and health terminology, balancing professionalism with ease of understanding. This ensures users can accurately understand the correlation between their blood sugar changes and their behavior, providing clear direction and motivation for long-term health management.

[0071] The aforementioned non-invasive dynamic blood glucose monitoring method for wearable devices continuously collects users' physiological and behavioral signals and fuses them into a unified feature vector. A personalized, cascaded neural network model is then used to synchronously analyze this vector. This not only predicts real-time blood glucose values ​​but, more importantly, quantifies the instantaneous impact of specific behaviors on blood glucose and calculates the counterfactual blood glucose state without these behaviors. Based on this attribution information, a process contribution value representing the user's health effort is calculated, ultimately generating a monitoring report that integrates numerical results and behavioral attributions. This approach achieves interpretable attribution of blood glucose fluctuations, extending the monitoring focus from a single blood glucose result to the behavioral process leading to that result. It provides a precise data foundation for implementing fair, anti-speculation, and health-encouraging management and incentives.

[0072] In one optional embodiment, the calculation process of the influence coefficient includes the following steps:

[0073] S11. Construct a fusion feature vector sequence based on the fusion feature vectors of each time period.

[0074] Specifically, the effect of behavior on blood glucose persists for a certain period of time along with physiological metabolic processes. The sequence length needs to be determined by statistically analyzing the decay period of the behavioral impact to ensure that the sequence covers historical periods that are meaningful for assessment at the current moment. This avoids losing key information due to excessive length, while also preventing data redundancy and decreased computational efficiency due to excessive length. The formula for constructing the fused feature vector sequence is as follows:

[0075]

[0076] in, Represents the fused feature vector sequence. Representative moment The fused feature vector, The range of values ​​is to , Representing the current moment, this formula explicitly states that the sequence is arranged in ascending order of timestamps, selecting consecutive times preceding the current moment. A sequence of fused feature vectors.

[0077] Insufficient historical time points in the initial monitoring stage At each time point, a forward padding strategy is adopted, using the earliest effective fusion feature vector to fill the missing positions. Since the user's physiological state and behavioral patterns are relatively stable in the early stages of monitoring, early vectors can approximate the initial state characteristics, avoiding the impact of insufficient sequence length on subsequent calculations. If the fusion feature vector at a certain historical time point is invalid data (such as a missing signal marker), a replacement vector is generated through linear interpolation of adjacent effective vectors. The interpolation process follows the feature distribution law of adjacent vectors, ensuring that the replacement vector conforms to the temporal feature change trend and maintains sequence coherence. After sequence construction, dimensionality verification is required to ensure that the dimension of each vector is consistent. Finally, a two-dimensional tensor of T×D dimensions (D is the dimension of a single fusion feature vector) is output. This tensor structure reflects the temporal relationship and adapts to the batch processing requirements of the subsequent attention mechanism, which is a key design to ensure the effectiveness of subsequent calculations.

[0078] S12. Input the fused feature vector sequence into the behavior attribution branch. Based on the attention mechanism network in the behavior attribution branch, calculate the attention weight distribution vector according to the correlation between the behavior context features at different time points in the fused feature vector sequence and the physiological state features at the current time. The magnitude of the value of each dimension in the attention weight distribution vector represents the importance weight of the behavior context features at the corresponding historical time point in estimating the behavior at the current time.

[0079] Specifically, firstly, two types of feature subsets are separated from the fused feature vector sequence: behavioral context feature subsets. Physiological state feature subvectors The physiological feature subvector of the last element in the current physiological state feature reference sequence. This feature separation step is a prerequisite for attention computation.

[0080] The additive attention mechanism calculates correlation in three steps: first, it maps two classes of features to the same high-dimensional space using two independent fully connected layers; second, it generates intermediate correlation scores through nonlinear transformations; and third, it maps the correlation scores to scalar scores. The formula is as follows:

[0081]

[0082] in, The correlation score between time τ and t represents the time interval t. and These are the weight matrices mapping behavioral context features to physiological state features, respectively. and For the corresponding bias term, For the weight row vectors of a single-output dimension fully connected layer, H is the bias term for this layer, and H is the intermediate feature dimension (determined through hyperparameter tuning). This formula effectively solves the problem of quantifying the correlation of features with different dimensions.

[0083] The attention weight distribution vector is obtained by normalizing all relevance scores using the Softmax function, as shown in the formula:

[0084]

[0085] in, The attention weights at time τ are represented by a vector whose sum is 1, forming the attention weight distribution vector. The weights directly correspond to the importance of historical actions to the current situation; recent key actions have significantly higher weights than earlier, less influential actions, allowing the model to automatically focus on core factors. All parameters of the attention mechanism network ( , All of these (e.g., behavioral attribution branch) are optimized collaboratively during model training. The optimization objective is to minimize the loss function of the behavioral attribution branch, ensuring that the weight allocation accurately reflects the true correlation between behavior and blood glucose changes.

[0086] S13. Based on the attention weight distribution vector, the behavior-related features in the fused feature vector sequence are weighted and summed to obtain the context-aware feature representation vector; the context-aware feature representation vector is then subjected to a nonlinear transformation to obtain a high-dimensional feature vector.

[0087] Specifically, firstly, a subset of behavior-related features is selected from the fused feature vector using a feature selection mechanism to form a behavior-related feature vector. The selection criteria are based on the Pearson correlation coefficient between features and behavior type and intensity, retaining features whose absolute values ​​are greater than a preset threshold (such as physiological signal features directly affected by behavior), ensuring... It accurately reflects the traces of how behavior affects physiological state.

[0088] The weighted summation formula based on attention weights is:

[0089]

[0090] in, Represents a context-aware feature representation vector. Represents attention weight, The formula represents the behavior-related feature vector, which aggregates key historical information. Features with high weights dominate the summation, making the context-aware feature representation vector condense historical behavioral information that has a key impact on current blood glucose changes.

[0091] Because the effect of behavior on blood glucose is non-linear (e.g., the relationship between exercise intensity and blood glucose reduction is not directly proportional), a two-layer multilayer perceptron (MLP) transformation is required. The input to the first MLP layer is a context-aware feature representation vector. The output dimension is ( for The second MLP layer takes the output features of the first layer as input, and its output dimension is [dimension dimension missing]. The activation function chosen is LeakyReLU, which alleviates the vanishing gradient problem in the negative interval and is more suitable for capturing weak nonlinear correlations. The activation function is also LeakyReLU. By increasing dimensionality and nonlinear transformation, the interaction relationships between features are explored in depth. A BatchNorm layer and a Dropout layer are added between the two MLP layers. The BatchNorm layer standardizes the input features, accelerating model convergence and improving stability. The dropout layer can be set to a dropout rate of 0.2, randomly discarding some neurons to avoid overfitting, ultimately generating a high-dimensional feature vector. All parameters of the nonlinear transformation (weight matrix and bias term of the MLP, scaling factor and offset of the BatchNorm layer) are optimized collaboratively during model training to ensure accurate capture of the nonlinear laws of behavioral influence.

[0092] S14. Perform a linear mapping on the high-dimensional feature vector to output a behavior influence coefficient vector containing the behavior influence coefficient; where the behavior influence coefficient vector is... K is the total number of preset behavior categories. The behavior influence coefficient is used to characterize the instantaneous quantitative impact estimate of the k-th type of behavior on the rate of change of blood glucose at time t. This indicates that the k-th type of behavior tends to lower blood sugar at time t. This indicates that the k-th type of behavior tends to increase blood sugar at time t. The size of the value represents the strength of the influence. .

[0093] Specifically, this step transforms high-dimensional feature vectors into behavioral influence coefficient vectors with clear physical meaning through linear mapping, enabling precise output of the instantaneous quantitative impact of various behaviors on the rate of change in blood glucose at the current moment. The linear mapping is implemented through a fully connected network with a single output layer, and its formula is:

[0094]

[0095] in, Represents the vector of behavioral influence coefficients. Represents the weight matrix (dimension: , High-dimensional feature vectors (dimensions) The bias vector is represented by a dimension of K×1, where K represents the total number of preset behavior categories, covering all major daily behaviors of users.

[0096] weight matrix Each row vector corresponds to a class of behavior, and its element values ​​represent the contribution of each dimension of the high-dimensional feature vector to the influence coefficient of that class of behavior; bias vector The baseline values ​​used to adjust the influence coefficients of various behaviors compensate for fixed influencing factors (such as the inherent metabolic effects of behaviors) that are not fully captured in the high-dimensional feature vectors. The learning process for these parameters is synchronized with the overall model training. The pre-training phase is based on a large-scale historical dataset, minimizing the mean square error between the predicted and true baseline values ​​of the behavior influence coefficients. The process involves learning the general patterns of how various behaviors affect blood glucose; in the personalized fine-tuning stage, parameters are further optimized by combining user-specific data to adapt to individual metabolic differences among users (such as differences in blood glucose response to the same behavior among different users); the true baseline value is calculated by matching the differences in blood glucose change rates with and without the target behavior under similar physiological states.

[0097] The quantification standard for the behavioral influence coefficient vector is as follows: This indicates that type k behavior tends to lower blood sugar. This indicates a tendency to raise blood sugar, with the absolute value representing the intensity of the effect. Normalization is then applied to... The value range is controlled within the interval [-1, 1]. The closer the absolute value is to 1, the stronger the influence; the closer it is to 0, the weaker the influence. Each All data includes timestamps, maintaining chronological consistency with real-time and counterfactual blood glucose estimates. This provides accurate data support for subsequent contribution value calculations and behavioral attribution analysis, enabling users to clearly understand the specific impact of various behaviors on their blood glucose levels at different times.

[0098] In one optional embodiment, the calculation process for the counterfactual blood glucose estimate includes the following steps:

[0099] S21. Input the fused feature vector into the behavior attribution branch. In the behavior attribution branch, according to the preset counterfactual intervention instructions, replace the feature dimension values ​​in the fused feature vector corresponding to the specified target behavior category with neutral benchmark values ​​calculated based on the user's historical behavior baseline to obtain the counterfactual feature vector.

[0100] Specifically, counterfactual intervention instructions are pre-defined structured instructions that include the target behavior category identifier, the type of intervention operation, and the definition of the intervention scope, ensuring that the intervention only applies to the feature dimensions directly related to the target behavior and does not affect the effectiveness of other features.

[0101] The target behavior is selected from the fused feature vectors by using a pre-defined feature dimension index table. This index table clearly defines the position of the feature dimension corresponding to each type of behavior, covering dimensions such as one-hot encoding, duration, intensity, occurrence time, and correlation with key blood glucose periods of the target behavior, ensuring the accuracy of the selection.

[0102] The neutral baseline is calculated based on the user's historical behavior baseline, and its core is to obtain the typical feature values ​​when there is no targeted behavior. Constructing the historical behavior baseline requires filtering sample data of past user behavior without a targeted intent, removing outliers, and then using corresponding statistical methods according to feature type: for continuous features, the arithmetic mean is calculated using the following formula:

[0103]

[0104] in, As a neutral benchmark value, For the number of valid samples, For the first The feature values ​​of each sample; for discrete coded features, the mode is taken as the benchmark value.

[0105] The feature replacement process locates the target dimension according to the index table, replaces the original value with a neutral baseline value, and keeps the vector dimension and other features unchanged, which conforms to the "single-variable intervention" principle of counterfactual reasoning. After replacement, the integrity of the vector and the rationality of the values ​​are verified to ensure that the counterfactual feature vector can accurately simulate the scenario where "the target behavior did not occur".

[0106] S22. Input the counterfactual feature vector into the counterfactual inference subnetwork embedded in the behavior attribution branch. Based on the counterfactual inference subnetwork, perform multi-layer nonlinear transformation on the counterfactual feature vector to obtain the high-dimensional counterfactual feature representation.

[0107] Specifically, this step uses a counterfactual inference subnetwork embedded in the behavioral attribution branch to perform multi-layer nonlinear transformations on the counterfactual feature vectors, extracting high-dimensional counterfactual feature representations. The subnetwork structure is optimized for the characteristics of counterfactual features and complements the shared base layer.

[0108] The sub-network adopts a hybrid architecture of "convolutional feature extraction + fully connected feature enhancement". After inputting the counterfactual feature vector, it is first mapped to the feature space inside the behavior attribution branch through a dimension adaptation layer to ensure compatibility. The convolutional feature extraction module contains multiple one-dimensional convolutional layers to capture the temporal dependencies in the counterfactual features. Each layer is followed by a batch normalization layer and a dropout layer to improve the model's stability and generalization ability. The fully connected feature enhancement module gradually improves the level of feature abstraction through a multilayer perceptron. The activation function is selected to alleviate the gradient vanishing problem and deeply mine the feature interaction relationships.

[0109] Subnetwork training requires generating counterfactual samples, obtained by performing the same counterfactual intervention based on historical data. The loss function uses the mean squared error between the counterfactual blood glucose estimate and the true counterfactual label. During training, the subnetwork parameters are updated collaboratively with other model parameters. The pre-training phase learns general mapping rules, and the personalized fine-tuning phase adapts to individual user characteristics to ensure accurate capture of feature association patterns in "targetless behavior" scenarios.

[0110] S23. Perform regression mapping on the counterfactual high-dimensional feature representation and output the counterfactual blood glucose estimate.

[0111] Specifically, the regression mapping employs a two-layer fully connected structure of "feature compression + precise regression". The first layer is the feature compression layer, which filters and integrates high-dimensional features, strengthens the expression of key features, and suppresses redundant information. The second layer is the precise regression layer, which uses a linear activation function to ensure that the output value matches the actual fluctuation range of blood glucose, avoiding output distortion. The loss function formula for the regression mapping is:

[0112]

[0113] in, The loss value. The number of training samples. For the first Counterfactual blood glucose estimates for each sample. The corresponding true counterfactual label.

[0114] True counterfactual labels are obtained by matching samples with similar physiological states. For the original sample containing the target behavior, samples with high physiological similarity and no target behavior are selected from historical data, and their true blood glucose values ​​are used as labels; if no perfectly matching samples are found, the K-nearest neighbor algorithm is used to select similar samples, and the mean is calculated as the label.

[0115] During training, the regression layer parameters are optimized through backpropagation. After output, a numerical verification mechanism is added to ensure rationality. The counterfactual blood glucose estimate is accompanied by a timestamp to maintain temporal consistency with the real-time blood glucose estimate, providing a reliable reference for the calculation of contribution values ​​in subsequent processes.

[0116] refer to Figure 2 In one optional embodiment, the process contribution value of user health behavior is quantified based on the behavioral impact coefficient and the counterfactual blood glucose estimate, including the following steps:

[0117] S31. Extract a specific health behavior that occurs consecutively from the behavior influence coefficient vector. Corresponding coefficient subsequence ;in, Indicates time, , and Specific health behaviors The start and end times.

[0118] Specifically, the core of this step is to accurately extract the coefficient subsequence corresponding to a single, continuous occurrence of a specific health behavior A from the behavior influence coefficient vector. The key lies in defining the behavioral boundaries and time intervals for a "single, continuous occurrence," ensuring that the coefficient subsequence is perfectly aligned with the target behavior's temporal sequence. First, based on the behavior sequence identification results, the continuous occurrence period of the specific health behavior A needs to be determined. This involves filtering out continuous behavioral segments of type A without temporal interruption using the behavior type identifier and timestamp information in the behavior sequence, and clarifying their start time. With end time This forms a complete duration range of behavior. This behavioral boundary definition needs to exclude situations such as behavioral interruption and switching, and only retain single-type, continuously executed behavioral fragments to ensure the integrity of the behavior corresponding to the coefficient subsequence.

[0119] Subsequently, based on time intervals Extract the coefficient values ​​at corresponding times from the behavioral influence coefficient vector to construct a coefficient subsequence. The extraction process is achieved through timestamp matching, that is, it iterates through the coefficients of τ at all times in the behavior influence coefficient vector, and filters out those τ belonging to... Furthermore, the coefficient elements corresponding to behavior type A are arranged in ascending order of timestamp to form a subsequence. It is necessary to ensure that each extracted coefficient element... Each value strictly corresponds to the execution state of a specific health behavior A at time τ, avoiding the mixing of coefficients across behaviors and time periods, ensuring the specificity and purity of the coefficient subsequence, and providing accurate behavioral impact intensity data support for the subsequent calculation of the contribution value of the original process.

[0120] S32. Calculate a specific health behavior by integrating the absolute value of the coefficient subsequence over the duration of the behavior. Original process contribution value Among them, the contribution value of the original process The calculation formula in discrete time is:

[0121]

[0122] in, Indicates the sampling time interval. Indicates at time Time corresponds to specific health behaviors The absolute value of the behavioral influence coefficient.

[0123] Specifically, the core of this step is to quantify the original process contribution value of a specific health behavior A by discretely integrating the absolute values ​​of the coefficient subsequence over the duration of the behavior. Essentially, this accumulates the instantaneous impact intensity during the behavior's execution, reflecting the degree of sustained effort. The core formula for calculating the original process contribution value includes... The value of τ is the contribution of the original process, and its range is the duration interval of the specific health behavior A. It covers all sampling moments of behavior execution; Let τ be the absolute value of the behavioral influence coefficient of behavior A at time τ, which represents the instantaneous intensity of the effect of behavior on blood glucose at that time. The sampling time interval is directly related to the acquisition frequency of the multimodal signal, ensuring the timing consistency between integral calculation and signal acquisition.

[0124] The core logic of this discrete integral formula is to transform the integral over continuous time into a summation over discrete time intervals. By multiplying the instantaneous impact intensity at each sampling moment by the time interval, the total cumulative impact intensity of the behavior over the entire process is obtained. The physical meaning of the original process contribution value lies in simply measuring the "process effort" of the behavior, which is only related to the impact intensity and duration of the behavior itself, and does not involve the actual effect of blood glucose changes. This quantitative dimension supplements the shortcomings of evaluation that only focuses on the result, and lays the foundation for subsequent comprehensive evaluation that combines actual benefits.

[0125] S33, Obtaining information on specific health behaviors The sequence of counterfactual blood glucose estimates, which is composed of counterfactual blood glucose estimates, and the sequence of real-time blood glucose estimates, which is constructed from real-time blood glucose estimates, within the affected time window.

[0126] Specifically, the core of this step is to obtain the counterfactual blood glucose estimation sequence and the real-time blood glucose estimation sequence within the time window in which a specific health behavior A is affected. The key is to clarify the definition of the "affected time window," which is the complete period during which behavior A affects blood glucose, distinct from the duration interval of the behavior itself. The determination of the affected time window is based on the physiological and metabolic patterns of the impact of behavior on blood glucose, covering the period during and after the behavior, ensuring a complete capture of the entire process of the behavior's impact on blood glucose, and avoiding the omission of subsequent delayed effects by only considering the duration of the behavior.

[0127] Both the counterfactual blood glucose estimation series and the real-time blood glucose estimation series were constructed using the affected time window as the boundary, arranging all blood glucose estimates within the corresponding time period in ascending order of timestamps. The counterfactual blood glucose estimation series consists of the counterfactual blood glucose estimates at each moment within the time window, representing the blood glucose change trend when "behavior A did not occur." The real-time blood glucose estimation series consists of the real-time blood glucose estimates for the same period, representing the actual blood glucose change trend when "behavior A occurred." The construction of both series must strictly ensure a one-to-one correspondence of timestamps to guarantee the accuracy of subsequent area under the curve (AUC) difference calculations, providing paired time-series data to verify the actual blood glucose improvement benefits of behavior A.

[0128] S34. Calculate the difference between the area under the curve (AUC) of the counterfactual blood glucose estimation sequence and the real-time blood glucose estimation sequence within the affected time window. Normalize the AUC difference to obtain the benefit verification contribution value. .

[0129] Specifically, the core of this step is to calculate and normalize the difference between the area under the curve (AUC) of the counterfactual and real-time blood glucose estimation sequences to obtain a benefit verification contribution value. This value is used to verify the actual blood glucose improvement effect corresponding to the original process contribution value, thus achieving a correlation verification between "process effort" and "actual benefit." First, the AUC of the two sequences within the affected time window is calculated. In discrete time, the trapezoidal method is used to calculate the AUC, i.e., by summing the areas of the trapezoids formed by adjacent blood glucose estimates and time intervals, the AUC value of each sequence is obtained, denoted as follows: (Counterfactual sequence) and (Real-time sequence).

[0130] Difference in area under curve The calculation is as follows The sign and magnitude of the difference directly reflect the actual impact of behavior A on blood sugar: a positive difference indicates that behavior A lowers the actual blood sugar level than the counterfactual scenario, producing a positive improvement benefit; the larger the difference, the more significant the positive benefit. Subsequently, [further details are needed]. Normalization is performed, mapping the result to the [0,1] interval, to obtain the benefit verification contribution value. The normalization process uses the preset maximum possible benefit difference as a reference benchmark to ensure that the benefit verification contribution values ​​of different behaviors and different users are comparable. Its core function is to provide actual effect verification for the original process contribution value and avoid one-sided evaluation based solely on the intensity of behavior.

[0131] S35, Contribution to the original process Contribution value verified by benefits Weighted fusion is performed to generate the final process contribution value used for incentive evaluation. The formula for weighted fusion is:

[0132]

[0133] in, Indicates specific health behaviors The final process contribution value, This represents the preset reference maximum value used when normalizing the original process contribution metric. The preset weighting coefficients, and .

[0134] Specifically, the core of this step is to generate a final process contribution value that balances process effort and actual benefits by weighted fusion of the original process contribution value and the benefit verification contribution value, thus providing a scientific basis for incentive evaluation. The core calculation formula includes… The contribution value to the final process of a specific health behavior A; The contribution value to the normalized original process. The preset reference maximum value is used to eliminate the differences in the units of the original contribution value caused by different behavior types and durations, so that the normalization result is in the range [0,1]; η is the weighting coefficient, with a value range of This is used to adjust the proportion of process effort and actual benefits in the final score.

[0135] The weighting coefficient η is set in conjunction with health management goals. If the focus is on encouraging users to continuously perform healthy behaviors, the η value can be appropriately increased; if the focus is on emphasizing the actual blood glucose improvement effect of the behaviors, the η value can be decreased to ensure the flexibility and pertinence of the assessment system. The settings are based on the statistical results of raw process contribution values ​​from large-scale user historical data, selecting the representative maximum value as a reference benchmark to ensure the rationality of normalization. Final process contribution value. It integrates the sustained effort intensity of behavior execution (original contribution value) with the actual blood glucose improvement benefit brought about by the behavior (benefit verification contribution value), which avoids the problem of "ineffective effort" caused by simply evaluating the behavior intensity, and prevents short-term speculative behavior caused by evaluating only blood glucose results, providing core quantitative indicators for building a fair and effective incentive mechanism.

[0136] In one optional embodiment, a monitoring report with behavioral attribution results is generated based on the process contribution value and real-time blood glucose estimate, including the following steps:

[0137] S41. By summing up the process contribution values ​​corresponding to all positive health behaviors within the preset reporting period, the total health effort score for the preset reporting period is obtained.

[0138] Specifically, the labeling of positive health behaviors requires a dual assessment of the behavior's essential attributes and its actual impact direction; not all health behaviors are directly included in the cumulative assessment. First, based on a pre-defined health behavior category system, behaviors that are controllable by the user and have clinically validated positive potential for blood sugar control (such as regular exercise, healthy eating, and timely medication) are selected, excluding neutral behaviors with no clear positive effect. Second, a secondary verification is performed based on the direction of the behavior's impact coefficient, only including behaviors with the impact coefficient... Positive behaviors are defined as positive health behaviors, ensuring that all behaviors included in the scoring actually have a positive impact on lowering or inhibiting blood sugar levels, thus avoiding the inclusion of "ineffective efforts" in the scoring. The labeling process is achieved through a dual matching of behavior type identifiers and coefficient symbols, forming a dedicated dataset of positive health behaviors, providing an accurate source for score calculation.

[0139] The definition of the preset reporting period needs to balance the regularity of health management with the timeliness of user feedback. The period length is determined through analysis of health management practices and user behavior habits, and supports multiple period configurations (such as daily, weekly, and monthly periods). The weekly period is used as the core reporting period by default—this avoids excessive fluctuations in scores due to insufficient data in the daily period, and prevents the monthly period from being too long and making it difficult to quickly reflect the effects of behavior adjustments. The time boundaries of each reporting period are clearly defined. For example, the weekly period starts at a fixed time each week (e.g., Monday 00:00) and ends at a fixed time (e.g., Sunday 24:00), ensuring the uniformity and continuity of period division and facilitating cross-period comparative analysis for users.

[0140] The summation process must adhere to strict data processing standards to ensure the accuracy of the integral calculation. First, the final process contribution values ​​of positive health behaviors within the reporting period are cleaned to remove extreme outliers caused by signal anomalies or model errors (values ​​exceeding the range of Q1-1.5IQR to Q3+1.5IQR are identified and removed using the interquartile range). Then, all the cleaned final process contribution values ​​are summed sequentially according to the chronological order of the behaviors, using the following formula:

[0141]

[0142] TotalScore represents the total health effort score, and M represents the number of effective positive health behaviors during the reporting period. The final contribution value for the i-th positive health behavior. During the accumulation process, the contribution details of each behavior (including behavior type, occurrence time, and contribution value) should be retained to facilitate subsequent user tracking of the score composition.

[0143] The core value of the Total Health Effort Score lies in quantifying a user's "process-oriented effort." Its significance extends beyond the score value for a single period; it also supports cross-period trend analysis. By comparing the total scores across different reporting periods, users can intuitively perceive the degree of persistence in their health behaviors and the potential for improvement. Simultaneously, the score can be compared with preset effort target thresholds (personally set based on the user's health status and management needs) to generate feedback on effort achievement, providing core quantitative evidence for the implementation of incentive mechanisms and helping users build motivation for continuous health management.

[0144] S42. Based on the probability of fluctuation type output by analyzing blood glucose fluctuation events through the behavioral attribution branch, perform root cause classification on each blood glucose fluctuation event within the preset reporting period to obtain the classification statistics of each blood glucose fluctuation event belonging to different preset cause categories; wherein, the probability of fluctuation type is generated by the behavioral attribution branch based on the fused feature vector.

[0145] Specifically, the definition of blood glucose fluctuation events needs to avoid omitting significant fluctuations and misjudging normal fluctuations, employing a dual-threshold method of "relative rate of change + absolute rate of change." First, the relative and absolute rates of change of blood glucose estimates at adjacent time points are calculated. The relative rate of change is the ratio of the difference in blood glucose levels at adjacent time points to the blood glucose value at the previous time point, while the absolute rate of change is the absolute value of the difference in blood glucose levels at adjacent time points. Then, dual thresholds (relative and absolute thresholds) are set. When either threshold exceeds a preset range, it is determined as a blood glucose fluctuation event. The setting of the dual thresholds needs to be based on the statistical distribution of large-scale blood glucose data, taking into account the differences among users with different baseline blood glucose levels. For example, the absolute threshold for the high blood glucose range can be appropriately relaxed, while it should be tightened for the low blood glucose range, ensuring the adaptability and accuracy of the fluctuation event definition. Furthermore, to avoid frequent triggering of fluctuation judgments within a short period, a minimum interval time for fluctuation events is set. Consecutive exceedances of the threshold under the same fluctuation trend are merged into a single fluctuation event, ensuring the independence and integrity of fluctuation events.

[0146] The pre-defined categorization of causes should comprehensively cover the main influencing factors of blood glucose fluctuations, while also considering the practicality and interpretability of the classification. It should be divided into multiple dimensions based on the controllability and source of the influencing factors. Specifically, this includes user-controllable health behaviors (such as dietary behavior, exercise behavior, and medication behavior), uncontrollable physiological states (such as hormonal fluctuations, sleep quality, and changes in basal metabolism), environmental factors (such as temperature changes and stress), and measurement errors (such as abnormal sensor contact). Each major category can be further subdivided into subcategories (e.g., dietary behavior can be divided into high-sugar diet, high-fat diet, and binge eating), forming a clear and comprehensive root cause classification system. This classification system must be consistent with the output dimensions of the behavioral attribution branch to ensure a one-to-one correspondence between the probability of fluctuation types and the cause categories.

[0147] The probability of fluctuation type is generated by the behavioral attribution branch based on the fused feature vector. Its core is to learn the association patterns between fused features (physiological features + behavioral features) and various fluctuation causes through model learning, outputting a probability distribution vector for each fluctuation event belonging to various preset causes. For example, if the probability distribution vector for a fluctuation event is [dietary behavior: 0.75, exercise behavior: 0.1, physiological state: 0.1, other: 0.05], then the probability values ​​of each cause directly reflect the likelihood of it causing the fluctuation. The root cause determination adopts the "maximum probability principle," that is, selecting the cause category with the largest value in the probability distribution vector as the root cause of the fluctuation event. If the maximum probability value is lower than a preset confidence threshold (e.g., 0.5), it is determined as "unknown cause," avoiding forced classification when evidence is insufficient.

[0148] The generation of categorized statistical results is based on the root cause determination of all fluctuation events, calculating indicators such as the number, proportion, and frequency of fluctuation events corresponding to each type of cause. Specific statistical dimensions include: the total number of fluctuation events for each type of cause and their percentage of total fluctuation events; the distribution of each type of cause at different time periods (e.g., fasting period, postprandial period); and the top 3 rankings of high-frequency fluctuation causes. The statistical results are presented in a structured table format, clearly showing the main driving factors of the user's blood glucose fluctuations. For example, if the fluctuation rate corresponding to "high-sugar diet" reaches 40%, it suggests that the user needs to focus on adjusting their diet; if the fluctuation rate related to "physiological state" is high, it is necessary to pay attention to the impact of uncontrollable factors on blood glucose, providing a clear direction for optimizing subsequent health management strategies, helping users distinguish between controllable efforts and uncontrollable factors, and focusing on key behavioral adjustments that can produce actual improvement.

[0149] S43. Based on the final process contribution value corresponding to each health behavior within the reporting period, select the top N health behaviors with the highest contribution values; perform plotting processing on the real-time blood glucose estimation sequence and counterfactual blood glucose estimation sequence corresponding to each of the top N health behaviors to generate a blood glucose change comparison chart.

[0150] Specifically, the selection of high-contribution health behaviors needs to balance contribution and representativeness. First, all positive health behaviors within the reporting period are sorted in descending order of their final process contribution value, and the top N behaviors are selected for analysis. The value of N adopts a "user-configurable + default optimization" mode. The default value is set based on report readability and information representativeness, ensuring focus on core contributing behaviors while avoiding overly complex comparison charts due to excessively large N. Users can adjust the size of N according to their needs to meet personalized analysis requirements. If multiple behaviors have the same final process contribution value and are competing for the top N, the principle of "behavior type diversity priority" is adopted, prioritizing different types of behaviors to ensure that the comparison chart can display the effects of various high-value behaviors and avoid the repetitive presentation of a single behavior type. Furthermore, behaviors with excessively short durations or insufficient data completeness must be excluded during the selection process to ensure the reliability and representativeness of the selected behavior's effect data.

[0151] The construction of the blood glucose change comparison chart aims to "intuitively present the differences," employing either a single-row chart or a multi-row combined chart format (combined charts are used when N≤3, and single-row charts are used when N>3). The core elements of the chart are designed as follows: the horizontal axis represents time, marking the complete interval of the affected time window, and clearly indicating the start time of the behavior. With end time The chart also marks key time points (such as 1 hour or 2 hours after the behavior ends, corresponding to the key periods of blood glucose impact); the vertical axis represents the blood glucose estimate, using a uniform blood glucose concentration unit, and the scale range covers the main range of the user's historical blood glucose fluctuations to ensure that the trends of the two sequences can be fully presented; two curves are plotted in the chart at the same time, the real-time blood glucose estimate sequence curve (labeled "actual blood glucose", using a striking color such as blue) and the counterfactual blood glucose estimate sequence curve (labeled "blood glucose if this behavior did not occur", using a contrasting color such as gray), and the two curves are superimposed on the same coordinate system for easy and intuitive comparison of differences.

[0152] To enhance the readability and information density of the comparison chart, key auxiliary elements can be added. Below the curve, label the area under the curve (AUC) values ​​and the difference between the two sequences to quantify the improvement in blood glucose caused by the behavior; use shaded areas to mark the duration intervals of the behavior. Clearly define the temporal relationship between behavior execution and blood glucose changes; label the peak and trough positions of the curve with specific values ​​and corresponding times to highlight key blood glucose change nodes; add legends, axis labels, and chart titles (such as "Comparison of Blood Glucose Improvement Effects of High-Contribution Behavior X") to ensure that users can quickly understand the meaning of the charts.

[0153] Preprocessing of the plotting data is crucial to ensuring the accuracy of the comparison chart. The real-time blood glucose estimation series and the counterfactual blood glucose estimation series are smoothed (using the moving average method) to eliminate curve jitter caused by data noise and highlight the core trend. Outliers in the series (such as extreme values ​​caused by temporary sensor failures) are corrected (using adjacent value interpolation) to ensure the continuity of the curve. The timestamps of the two series are fully aligned, and each time point has a corresponding blood glucose estimate to avoid comparison interruptions due to missing data.

[0154] The core value of blood glucose change comparison charts lies in transforming the abstract "process contribution value" into an intuitive "effect difference." Users can quickly perceive the actual improvement in blood glucose caused by high-contribution behaviors through the difference between the two curves. For example, if the "actual blood glucose" curve is consistently lower than the "hypothetical blood glucose without this behavior" curve, and the AUC difference is significant, it indicates that the behavior effectively lowered blood glucose; if the curves maintain the difference after the behavior ends, it indicates that the effect of the behavior is persistent. This visualization method helps users establish a direct link between "behavior and effect," enhances their understanding of the value of health behaviors, and motivates users to continuously adhere to high-contribution behaviors.

[0155] S44. Integrate the total health effort score, categorized statistical results, and blood glucose change comparison chart to generate a monitoring report.

[0156] Specifically, the monitoring report is integrated through five core modules. The first module consists of a report cover and a core summary. The cover clearly states basic information such as the reporting period, user identification, and generation time. The core summary concisely summarizes the report's core conclusions, including the total health effort score, changes compared to historical periods (e.g., "20% improvement compared to last week"), the main causes of blood sugar fluctuations (e.g., "High-sugar diet is the main cause of fluctuations"), and high-contribution behavioral highlights (e.g., "Post-meal exercise has a significant effect on improving blood sugar"). This allows users to quickly grasp the core information and obtain key feedback without reading the entire report.

[0157] The second module details the total health effort score, presenting the specific numerical value of the total score, calculation details (score composition categorized by behavior type, such as "exercise behavior contributes 50 points, diet behavior contributes 30 points"), cross-period trend comparison charts (such as a line graph showing score changes over the past 4 weeks), and effort achievement status (comparison with preset target thresholds, such as "achievement rate 85%)." This module provides a detailed explanation of the score composition and changes, helping users understand the distribution of their efforts and room for improvement, while reinforcing the motivation for continued effort through trend comparison.

[0158] The third module is the root cause analysis of blood glucose fluctuations. It presents the categorized statistical results in a structured table (the number, percentage, and time distribution of fluctuations for each cause), along with a pie chart to show the percentage distribution of each cause. Key findings are explained in text—for example, "Dietary-related fluctuations account for the highest percentage (40%), with high-sugar diets accounting for 25%, and it is recommended to reduce refined sugar intake"; "Post-exercise blood glucose fluctuations account for 10%, mostly benign fluctuations, and there is no need to worry excessively." This module helps users clarify the "controllable and uncontrollable" boundaries of blood glucose fluctuations, focusing on key causes that can be improved through behavioral adjustments, and avoiding blind adjustments.

[0159] The fourth module visualizes the effects of high-contribution behaviors, presenting blood glucose change comparison charts generated by S43. Each chart is accompanied by brief explanatory text explaining the behavior's contribution ranking, the magnitude of improvement in blood glucose (e.g., "AUC difference of X is equivalent to an average blood glucose reduction of Y"), and key details of the behavior's execution (e.g., "Exercise for 30 minutes 1 hour after a meal"). This module uses a "charts + text" format to allow users to intuitively understand the specific value of high-contribution behaviors, clarifying "which behaviors are effective and why," and providing replicable reference examples for subsequent behavior optimization.

[0160] The fifth module is a summary and personalized recommendations. Based on the information from the first four modules, it extracts core conclusions (such as "Users are continuously improving their health efforts, but their dietary structure still needs optimization") and generates targeted personalized recommendations. Recommendations should be specific and actionable, avoiding vague statements. For example, regarding high-sugar diets, it is recommended to "reduce the intake of sugary drinks and replace them with plain water; ensure that whole grains account for at least 50% of the staple food at each meal"; regarding high-contribution behaviors, it is recommended to "maintain the habit of exercising one hour after meals, and appropriately increase the exercise duration to 40 minutes"; regarding uncontrollable factors, it is recommended to "pay attention to sleep quality, maintain 7-8 hours of sleep per day, and reduce the impact of staying up late on blood sugar." The recommendations section should prioritize them, labeling them with tags such as "Key Optimization," "Continue to Maintain," and "Observe and Monitor" to help users clarify their action priorities.

[0161] The integration of monitoring reports must ensure the consistency of information time sequence and the accuracy of data. All data strictly correspond to the same preset reporting period, and timestamp verification is used to ensure that no cross-period information is mixed in. Data in key stages such as integral calculation, root cause classification, and chart drawing are verified twice to avoid calculation errors or classification biases. The presentation of the reports balances professionalism and readability, with professional terms (such as AUC and area under the curve) accompanied by easy-to-understand explanations, and key information (such as high-contribution behaviors and main causes of fluctuations) highlighted and bolded to facilitate users to quickly grasp the core content.

[0162] The core value of the monitoring report lies in providing users with a closed-loop service of "quantitative feedback - cause analysis - effect verification - action guidance". It not only lets users know the "effort level" and "actual effect" of their own health management, but also clarifies "how to optimize". At the same time, it provides structured data support for communication between users and health managers and doctors, helps to continuously optimize personalized health management plans, and ultimately achieves long-term improvement in blood sugar control.

[0163] The aforementioned non-invasive dynamic blood glucose monitoring method for wearable devices continuously collects users' physiological and behavioral signals and fuses them into a unified feature vector. A personalized, cascaded neural network model is then used to synchronously analyze this vector. This not only predicts real-time blood glucose values ​​but, more importantly, quantifies the instantaneous impact of specific behaviors on blood glucose and calculates the counterfactual blood glucose state without these behaviors. Based on this attribution information, a process contribution value representing the user's health effort is calculated, ultimately generating a monitoring report that integrates numerical results and behavioral attributions. This approach achieves interpretable attribution of blood glucose fluctuations, extending the monitoring focus from a single blood glucose result to the behavioral process leading to that result. It provides a precise data foundation for implementing fair, anti-speculation, and health-encouraging management and incentives.

[0164] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0165] Based on the same inventive concept, this application also provides a system for implementing the aforementioned non-invasive dynamic blood glucose monitoring method for wearable devices. The solution provided by this system is similar to the implementation described in the above method. Therefore, the specific limitations of one or more embodiments of the non-invasive dynamic blood glucose monitoring system for wearable devices provided below can be found in the limitations of the non-invasive dynamic blood glucose monitoring method for wearable devices described above, and will not be repeated here.

[0166] In one exemplary embodiment, such as Figure 3As shown, a non-invasive dynamic blood glucose monitoring system 30 for wearable devices is provided to implement the methods in the above-described method embodiments. The system includes:

[0167] The multi-source signal sensing and behavior analysis module 31 is used to collect the user's multimodal physiological signals and inertial signals, perform behavior recognition on the inertial signals, and obtain the user's behavior sequence.

[0168] The feature fusion generation module 32 is used to extract blood glucose-related features from multimodal physiological signals and construct physiological feature vectors based on blood glucose-related features; extract behavioral context information from user behavior sequences and construct behavioral feature vectors based on behavioral context information; and concatenate the physiological feature vectors and behavioral feature vectors in temporal alignment to generate a fused feature vector.

[0169] The cascaded neural network analysis module 33 is used to input the real-time generated fused feature vector into the pre-trained cascaded neural network model and output the real-time blood glucose estimate, the behavior influence coefficient reflecting the degree of influence of a specific behavior on the instantaneous change of blood glucose, and the counterfactual blood glucose estimate assuming that no specific behavior has occurred.

[0170] The cascaded neural network model is trained based on the user's historical data and fine-tuned with the user's personalized data. The historical data includes historical multimodal physiological signals and historical inertial signals. The cascaded neural network model includes a shared base layer, a blood glucose prediction branch, and a behavioral attribution branch. The shared base layer is used to receive and store the fused feature vector. The blood glucose prediction branch is connected to the output of the shared base layer and is used to output the real-time blood glucose estimate. The behavioral attribution branch is connected to the output of the shared base layer and is used to output the influence coefficient and the counterfactual blood glucose estimate.

[0171] The health attribution report generation module 34 is used to quantify the process contribution value of user health behavior based on the behavior impact coefficient and the counterfactual blood glucose estimate; and to generate a monitoring report with behavior attribution results based on the process contribution value and the real-time blood glucose estimate; wherein, the process contribution value represents the degree of positive effort that the user makes in performing specific health behaviors on their own blood glucose status.

[0172] Embodiments of this application also provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the aforementioned method embodiments.

[0173] Embodiments of this application also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0174] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0175] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for noninvasive dynamic blood glucose monitoring in wearable devices, characterized in that, The method includes: S1. Collect the user's multimodal physiological signals and inertial signals through wearable devices, perform behavior recognition on the inertial signals, and obtain the user's behavior sequence; S2. Extract blood glucose-related features from the multimodal physiological signals and construct a physiological feature vector based on the blood glucose-related features; extract behavioral context information from the user behavior sequence and construct a behavioral feature vector based on the behavioral context information; concatenate the physiological feature vector and the behavioral feature vector in chronological alignment to generate a fused feature vector; S3. Input the fused feature vector generated in real time into the pre-trained cascaded neural network model, and output the real-time blood glucose estimate, the behavior influence coefficient reflecting the degree of influence of a specific behavior on the instantaneous change of blood glucose, and the counterfactual blood glucose estimate assuming that no specific behavior has occurred. The cascaded neural network model is trained based on the user's historical data and fine-tuned with the user's personalized data. The historical data includes historical multimodal physiological signals and historical inertial signals. The cascaded neural network model includes a shared base layer, a blood glucose prediction branch, and a behavioral attribution branch. The shared base layer is used to receive and store the fused feature vector. The blood glucose prediction branch is connected to the output of the shared base layer and is used to output the real-time blood glucose estimate. The behavioral attribution branch is connected to the output of the shared base layer and is used to output the influence coefficient and the counterfactual blood glucose estimate. S4. Quantify the process contribution value of user health behavior based on the behavior influence coefficient and the counterfactual blood glucose estimate; generate a monitoring report with behavior attribution results based on the process contribution value and the real-time blood glucose estimate; wherein, the process contribution value characterizes the degree of positive effort that the user makes in performing specific health behaviors on their own blood glucose status.

2. The method according to claim 1, characterized in that, The calculation process for the influence coefficient includes: S11. Construct a fusion feature vector sequence based on the fusion feature vectors of each time period; S12. Input the fused feature vector sequence into the behavior attribution branch. Based on the attention mechanism network in the behavior attribution branch, calculate the attention weight distribution vector according to the correlation between the behavior context features at different time points in the fused feature vector sequence and the physiological state features at the current time. The value of each dimension in the attention weight distribution vector represents the importance weight of the behavior context features at the corresponding historical time point in estimating the behavior at the current time. S13. Based on the attention weight distribution vector, perform a weighted summation of the behavior-related features in the fused feature vector sequence to obtain a context-aware feature representation vector; perform a nonlinear transformation on the context-aware feature representation vector to obtain a high-dimensional feature vector; S14. Perform a linear mapping on the high-dimensional feature vector to output a behavior influence coefficient vector containing the behavior influence coefficient; wherein, the behavior influence coefficient vector is... K is the total number of preset behavior categories. The behavior influence coefficient is used to characterize the instantaneous quantitative impact estimate of the k-th type of behavior on the rate of change of blood glucose at time t. This indicates that the k-th type of behavior tends to lower blood sugar at time t. This indicates that the k-th type of behavior tends to increase blood sugar at time t. The size of the value represents the strength of the influence. .

3. The method according to claim 1, characterized in that, The calculation process for the counterfactual blood glucose estimate includes: S21. Input the fused feature vector into the behavior attribution branch. In the behavior attribution branch, according to the preset counterfactual intervention instruction, replace the feature dimension value corresponding to the specified target behavior category in the fused feature vector with a neutral benchmark value calculated based on the user's historical behavior baseline to obtain the counterfactual feature vector. S22. Input the counterfactual feature vector into the counterfactual inference subnetwork embedded in the behavior attribution branch, and perform multi-layer nonlinear transformation on the counterfactual feature vector based on the counterfactual inference subnetwork to obtain a high-dimensional counterfactual feature representation. S23. Perform regression mapping on the counterfactual high-dimensional feature representation and output the counterfactual blood glucose estimate.

4. The method according to claim 2, characterized in that, The process of quantifying the contribution value of user health behavior based on the behavioral impact coefficient and the counterfactual blood glucose estimate includes: S31. Extract a specific health behavior that occurs consecutively from the behavior influence coefficient vector. Corresponding coefficient subsequence ;in, Indicates time, , and Each of the specific health behaviors The start and end times; S32. Calculate the specific health behavior by integrating the absolute value of the coefficient subsequence over the duration of the behavior. Original process contribution value Wherein, the original process contribution value The calculation formula in discrete time is: in, Indicates the sampling time interval. Indicates at time Time corresponds to the specific health behavior The absolute value of the influence coefficient of the aforementioned behavior; S33, Obtain the specific health behavior The sequence of counterfactual blood glucose estimates composed of the counterfactual blood glucose estimates within the affected time window and the sequence of real-time blood glucose estimates constructed from the real-time blood glucose estimates; S34. Calculate the area under the curve difference between the counterfactual blood glucose estimation sequence and the real-time blood glucose estimation sequence within the affected time window, and normalize the area under the curve difference to obtain the benefit verification contribution value. ; S35, Contribution value to the original process Contribution value verified by the aforementioned benefit Weighted fusion is performed to generate the final process contribution value used for incentive evaluation. The formula for calculating the weighted fusion is as follows: in, Indicates the specific health behavior The final process contribution value, This represents the preset reference maximum value used when normalizing the contribution metric of the original process. The preset weighting coefficients, and .

5. The method according to claim 4, characterized in that, The generation of a monitoring report with behavioral attribution results based on the process contribution value and the real-time blood glucose estimate includes: S41. By summing up the process contribution values ​​corresponding to all positive health behaviors within the preset reporting period, the total health effort score for the preset reporting period is obtained. S42. Based on the fluctuation type probability output by the behavioral attribution branch in analyzing blood glucose fluctuation events, root cause classification is performed on each blood glucose fluctuation event within the preset reporting period to obtain the classification statistics of each blood glucose fluctuation event belonging to different preset cause categories; wherein, the fluctuation type probability is generated by the behavioral attribution branch based on the fusion feature vector; S43. Based on the final process contribution value corresponding to each health behavior within the reporting period, select the top N health behaviors with the highest contribution values; perform plotting processing on the real-time blood glucose estimation sequence and the counterfactual blood glucose estimation sequence corresponding to each of the top N health behaviors to generate a blood glucose change comparison chart. S44. Integrate the total health effort score, the classification statistics results, and the blood glucose change comparison chart to generate the monitoring report.

6. A non-invasive dynamic blood glucose monitoring system for wearable devices, used to implement the method according to any one of claims 1 to 5, characterized in that, The device includes: The multi-source signal sensing and behavior analysis module is used to collect the user's multimodal physiological signals and inertial signals, perform behavior recognition on the inertial signals, and obtain the user's behavior sequence. The feature fusion generation module is used to extract blood glucose-related features from the multimodal physiological signals and construct a physiological feature vector based on the blood glucose-related features; extract behavioral context information from the user behavior sequence and construct a behavioral feature vector based on the behavioral context information; and concatenate the physiological feature vector and the behavioral feature vector in a time-series alignment to generate a fused feature vector. The cascaded neural network analysis module is used to input the fused feature vector generated in real time into the pre-trained cascaded neural network model and output the real-time blood glucose estimate, the behavior influence coefficient reflecting the degree of influence of a specific behavior on the instantaneous change in blood glucose, and the counterfactual blood glucose estimate assuming that no specific behavior has occurred. The cascaded neural network model is trained based on the user's historical data and fine-tuned with the user's personalized data. The historical data includes historical multimodal physiological signals and historical inertial signals. The cascaded neural network model includes a shared base layer, a blood glucose prediction branch, and a behavioral attribution branch. The shared base layer is used to receive and store the fused feature vector. The blood glucose prediction branch is connected to the output of the shared base layer and is used to output the real-time blood glucose estimate. The behavioral attribution branch is connected to the output of the shared base layer and is used to output the influence coefficient and the counterfactual blood glucose estimate. The health attribution report generation module is used to quantify the process contribution value of a user's health behavior based on the behavior impact coefficient and the counterfactual blood glucose estimate; and to generate a monitoring report with behavior attribution results based on the process contribution value and the real-time blood glucose estimate; wherein the process contribution value characterizes the degree of positive effort a user makes in improving their blood glucose status by performing specific health behaviors.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 5.

Images