Blood glucose dynamic early warning method, device and equipment based on data enhancement

Abstract

The application discloses a blood glucose dynamic early warning method, device and equipment based on data enhancement; the method comprises the following steps: acquiring a continuous blood glucose monitoring time sequence and a physiological intervention event time sequence of a target individual in a historical period; implementing data enhancement by using a Lorenz chaotic noise and constructing a training data set to train a blood glucose prediction model; inputting multi-modal physiological time sequence data of a current time window into the trained model to obtain a predicted blood glucose trajectory of a prediction time window; and generating and sending blood glucose early warning information based on the predicted blood glucose trajectory; since the application enhances data based on the Lorenz chaotic noise, promotes the model to learn stable metabolic evolution characteristics, and fuses blood glucose and irregular intervention events to deduce future trajectories, the application can reduce the interference of extreme value noise and irregular sampling on the prediction result, and improve the stability of blood glucose prediction and early warning in a complex monitoring environment.

Description

Technical Field

[0001] This application relates to the field of data-enhanced dynamic blood glucose early warning, and more particularly to a data-enhanced dynamic blood glucose early warning method, device, and equipment. Background Technology

[0002] In the process of modern medicine's transformation towards intelligence and digitalization, the refined management of diabetes highly relies on the real-time and accurate physiological data provided by underlying sensing networks (such as continuous glucose monitoring devices). This time-series data provides the basic data for closed-loop artificial pancreas and intelligent early warning, serving not only as a key basis for optimizing intervention strategies and improving the time within the target range, but also as the core foundation for ensuring patient safety and preventing glucose metabolism disorders. At the microscopic level, minimally invasive sensors deployed on the patient's body surface capture the glucose concentration and rate of change in subcutaneous interstitial fluid in real time, providing decision support for dynamically adjusting diet and insulin infusion strategies. At the macroscopic level, massive amounts of physiological data, through in-depth mining and correlation analysis, can reveal hidden individualized metabolic patterns, fully leveraging the advantages of medical auxiliary systems and providing stronger support for medical resource allocation and intervention decisions.

[0003] However, the complexity and uncontrolled nature of real-world wearing environments pose challenges to the continuous acquisition of high-quality, continuous blood glucose data. During sleep or high-intensity activity, the probes of precision sensors are susceptible to impact from localized limb pressure and sweat, leading to data noise or even physical blockage, resulting in hard data gaps or extreme spikes (such as pressure-induced pseudohypoglycemia). Daily behaviors and various interventions (such as eating, exercise, and medication) not only significantly disrupt the stability of basal metabolism, but the randomness and irregularity of their occurrence also cause a non-uniform distribution of the data stream over time. These interfering factors result in clinical-grade blood glucose data generally exhibiting characteristics of non-stationarity, strong noise, and irregular sampling. The negative impact of impaired data integrity and accuracy on blood glucose early warning monitoring systems is multi-dimensional. In daily monitoring, distortion and missing key time-series data can cause state observers to fail, making it difficult for the prediction system to reconstruct the patient's true glucose metabolism trajectory, thereby masking early warning signs. This information "blind spot" not only causes frequent false alarms leading to "alarm fatigue," but may also delay the best intervention window for fragile fluctuations and acute hypoglycemia, inducing serious chain physiological harm and posing a serious threat to the patient's life, health, and medication safety. Summary of the Invention

[0004] The main objective of this application is to provide a data-enhanced dynamic blood glucose early warning method, device, and equipment, which aims to solve the technical problem that existing technologies are unable to stably and accurately predict future blood glucose trajectories and provide dynamic early warnings when continuous blood glucose monitoring data contains complex abnormal noise, individual metabolic lag, and irregular sampling.

[0005] To achieve the above objectives, this application provides a data-enhanced dynamic blood glucose early warning method, the method comprising the following steps: Acquire the first multimodal physiological time series data of the target individual for historical periods, including continuous blood glucose monitoring time series and physiological intervention event time series collected in historical periods; Lorentz chaotic noise is generated based on the Lorentz chaotic system, and the first multimodal physiological time series data is augmented based on the Lorentz chaotic noise. A training dataset is constructed based on the data-augmented first multimodal physiological time-series data. A pre-constructed blood glucose prediction model is trained based on the training dataset to obtain the trained blood glucose prediction model. The second multimodal physiological time-series data collected within the current time window of the target individual is input into the trained blood glucose prediction model to obtain the predicted blood glucose trajectory of the prediction time window; Based on the predicted blood glucose trajectory, a blood glucose warning message is generated and sent to the warning terminal associated with the target individual. The warning terminal includes wearable smart terminals and mobile terminals.

[0006] Furthermore, to achieve the above objectives, this application also proposes a data-enhanced dynamic blood glucose warning device that applies the data-enhanced dynamic blood glucose warning method described above, wherein the data-enhanced dynamic blood glucose warning device comprises: The historical data acquisition module is used to acquire the first multimodal physiological time series data of the target individual during historical periods. The first multimodal physiological time series data includes the continuous blood glucose monitoring time series and the physiological intervention event time series collected during historical periods. The data augmentation module is used to generate Lorentz chaotic noise based on the Lorentz chaotic system, and to perform data augmentation on the first multimodal physiological time series data based on the Lorentz chaotic noise. The model training module is used to construct a training dataset based on the data-augmented first multimodal physiological time-series data, train a pre-constructed blood glucose prediction model based on the training dataset, and obtain the trained blood glucose prediction model. The blood glucose trajectory prediction module is used to input the second multimodal physiological time series data collected within the current time window of the target individual into the trained blood glucose prediction model to obtain the predicted blood glucose trajectory of the prediction time window. The blood glucose dynamic early warning module is used to generate blood glucose early warning information based on the predicted blood glucose trajectory, and send the blood glucose early warning information to the early warning terminal associated with the target individual. The early warning terminal includes wearable smart terminals and mobile terminals.

[0007] Furthermore, to achieve the above objectives, this application also proposes a data-enhanced dynamic blood glucose warning device, the device comprising: a memory, a processor, and a data-enhanced dynamic blood glucose warning program stored in the memory, the processor being used to run the data-enhanced dynamic blood glucose warning program, the computer program being configured to implement the steps of the data-enhanced dynamic blood glucose warning method as described above.

[0008] This application acquires continuous blood glucose monitoring time series and physiological intervention event time series, preserving the correlation between blood glucose changes and dietary, medication, and exercise factors under a unified temporal semantics, providing multidimensional input for modeling individual metabolic states. It expands the complex anomaly monitoring patterns in the training data by generating nonlinear, non-periodic, and time-correlated perturbation sequences using a Lorenz chaotic system, reducing the model's dependence on isolated spikes and local extrema. A training dataset is constructed using data augmentation samples and real blood glucose trajectory labels, enabling the model to learn mapping relationships to the real blood glucose evolution trajectory even under perturbation inputs. The trained blood glucose prediction model extracts temporal features and extrapolates continuous states from current multimodal physiological time series data, allowing irregular intervention events, metabolic lag features, and blood glucose change trends to jointly participate in future trajectory calculations. By analyzing and predicting blood glucose trajectories and sending early warning information to associated terminals, it transforms single-point post-event judgment into early warning for prediction time windows. Therefore, this invention, while preserving real dynamic blood glucose changes, suppresses the transmission of abnormal measurement perturbations to prediction results, improving the stability and accuracy of blood glucose prediction and dynamic early warning under complex wearing environments and irregular sampling conditions. Attached Figure Description

[0009] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0010] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart illustrating the first embodiment of the data-enhanced dynamic blood glucose early warning method of this application; Figure 2 This is a flowchart illustrating the second embodiment of the data-enhanced dynamic blood glucose early warning method of this application; Figure 3 This is a flowchart illustrating the third embodiment of the data-enhanced dynamic blood glucose early warning method of this application. Figure 4 This is a schematic diagram of the module structure of the data-enhanced dynamic blood glucose warning device according to an embodiment of this application; Figure 5 This is a schematic diagram of the hardware operating environment involved in the data-enhanced dynamic blood glucose early warning method in the embodiments of this application.

[0012] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0013] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0014] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0015] It should be noted that the executing entity of this application embodiment can be an electronic device with data processing, network communication, and program execution functions, such as a server, cloud computing node, edge computing device, medical data processing platform, or mobile computing terminal. Continuous blood glucose monitoring devices, wearable smart terminals, and mobile terminals can serve as sources of physiological data or terminals for receiving early warning information. The following uses a data-enhanced dynamic blood glucose early warning device (hereinafter referred to as the early warning device) as an example to illustrate this embodiment and the following embodiments.

[0016] The information and data involved in this application may include continuous blood glucose monitoring data, dietary intake records, drug intervention records, exercise records, data acquisition device information, sampling time information, individualized blood glucose thresholds, and early warning terminal information of the target individual. All of the above information and data were obtained with the authorization of the target individual or relevant data providers and are used for the construction of multimodal physiological time-series data, training of blood glucose prediction models, prediction of future blood glucose trajectories, and dynamic early warning.

[0017] In some embodiments, the early warning device may also use methods such as authentication, access control, data desensitization, and encrypted transmission to process the aforementioned information and data, so as to restrict the scope of data access and ensure that the collection, transmission, storage, and use of relevant information and data comply with applicable laws and regulations and data security requirements.

[0018] Based on this, the first embodiment of this application provides a data-enhanced dynamic blood glucose early warning method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the data-enhanced dynamic blood glucose early warning method of this application.

[0019] In this embodiment, the data-enhanced dynamic blood glucose early warning method includes: Step S10: Obtain the first multimodal physiological time series data of the target individual for the historical period.

[0020] It should be noted that the first multimodal physiological time series data refers to a time series dataset containing multiple physiological and behavioral monitoring data of the target individual over a period of time, including continuous blood glucose monitoring time series and physiological intervention event time series collected in historical periods.

[0021] Continuous blood glucose monitoring time series refers to the sequence of glucose concentration values ​​in subcutaneous interstitial fluid collected by a continuous blood glucose monitoring device at a fixed sampling period (e.g., every 5 minutes). Physiological intervention event time series refers to the time series recording the occurrence and related parameters of external intervention behaviors that affect glucose metabolism, such as eating, medication, and exercise. For example, dietary intake records may include the time and estimated grams of carbohydrate intake, medication intervention records may include the time and dosage of insulin injection, and exercise records may include heart rate and step count data.

[0022] In some embodiments, the warning device can establish a data connection with a continuous blood glucose monitoring device, wearable smart terminal (such as a smartwatch), and mobile terminal (such as a smartphone) worn by the target individual via a wireless communication interface (such as Bluetooth or WiFi) and send a data synchronization request.

[0023] The early warning device, in response to data synchronization permission, acquires continuous blood glucose monitoring data from a continuous glucose monitoring device for a historical period (e.g., the past 30 days), obtaining a chronologically ordered sequence of glucose concentration values. Simultaneously, it acquires physiological intervention event records from wearable smart terminals and mobile terminals within the same historical period, including but not limited to dietary intake records, medication intervention records, and exercise records. Using the sampling time of the continuous blood glucose monitoring data as the baseline time axis, the trigger time of each event in the physiological intervention event records is mapped to the nearest sampling window on the time axis. For missing sampling points on the time axis due to device transmission failures, linear interpolation or spline interpolation is used to complete the data.

[0024] Finally, the completed continuous blood glucose monitoring time series and the aligned physiological intervention event time series are aggregated in multiple dimensions along the time step dimension, feature variable dimension, and modality category dimension to obtain the first multimodal physiological time series data.

[0025] Step S20: Generate Lorentz chaotic noise based on the Lorentz chaotic system, and perform data augmentation on the first multimodal physiological time series data based on the Lorentz chaotic noise.

[0026] It should be noted that a Lorentz chaotic system refers to a dynamic system composed of multiple coupled nonlinear state variables that can generate deterministic chaotic trajectories. Lorentz chaotic noise refers to the noise perturbation sequence formed after sampling and scaling the state variable sequence of a Lorentz chaotic system.

[0027] It should be understood that this embodiment introduces a data augmentation mechanism based on Lorentz noise. By utilizing the "heavy-tailed" characteristic of the Lorentz distribution, noise is injected into the sequence during the training phase, forcing the model to learn and adapt to the interference of extreme outliers. Compared to traditional predictors that rely on stationary data for training, it effectively solves the problem of model collapse caused by extreme value spikes due to sensor physical pressure or aging, significantly enhancing the robustness of the prediction model in real and complex monitoring environments.

[0028] In some embodiments, the early warning device sets the state control parameters, initial state, and numerical solution step size of the Lorentz chaotic system. It obtains a sequence of chaotic state variables through numerical solution, and then performs transient removal, centering, normalization, and amplitude scaling on the selected state variable sequence to form Lorentz chaotic noise corresponding to the continuous blood glucose monitoring time series. The device superimposes the Lorentz chaotic noise onto at least one continuous sensor data channel of the first multimodal physiological time series data according to the set injection location and injection probability, generating data augmentation samples. The injection process only affects the training data copy; the original data continues to be used to construct supervision labels or original training samples.

[0029] Step S30: Construct a training dataset based on the data-augmented first multimodal physiological time-series data, train a pre-constructed blood glucose prediction model based on the training dataset, and obtain the trained blood glucose prediction model.

[0030] It should be noted that the training dataset refers to the data set consisting of model input samples and corresponding supervision labels. The model input samples may include data-enhanced historical blood glucose data, dietary data, medication intervention data, exercise data, and corresponding time information. The actual blood glucose trajectory labels refer to the sequence of blood glucose values ​​actually collected by continuous glucose monitoring devices after the input time window.

[0031] It should be noted that the blood glucose prediction model consists of multiple sub-network modules, including a segmented attention feature extraction module, a neural network constant differential equation continuous-time evolution module, and a decoding module.

[0032] The pre-built blood glucose prediction model has initialized model parameters before training. The parameters are updated through forward propagation, loss calculation and backward gradient propagation during the training process. The trained blood glucose prediction model has knowledge of blood glucose metabolism evolution learned from the data.

[0033] In some embodiments, the early warning device divides the data-enhanced first multimodal physiological time-series data into samples using a sliding time window (the window length is, for example, historical data from the past 8 hours, and the sliding step size is, for example, 30 minutes). The training input sequence for each sample is the multidimensional physiological features within the window, and the real blood glucose trajectory label is the real blood glucose sequence for a future period of time after the window (e.g., the next 60 minutes).

[0034] The constructed training dataset is input into the pre-built blood glucose prediction model, which consists of three parts: a segmented attention feature extraction sub-network for capturing the evolution trend of metabolic bands, a continuous-time evolution sub-network for modeling the evolution of physiological states in the continuous time domain using neural ordinary differential equations, and a fully connected decoding sub-network for mapping the hidden state tensor to the blood glucose prediction value.

[0035] During training, the parameters are updated using a stochastic gradient descent optimizer based on the prediction loss function. Training stops and the trained blood glucose prediction model is output when the number of training iterations reaches a preset upper limit or the prediction error on the validation set no longer decreases.

[0036] Step S40: Input the second multimodal physiological time series data collected within the current time window of the target individual into the trained blood glucose prediction model to obtain the predicted blood glucose trajectory of the prediction time window.

[0037] It should be noted that the second multimodal physiological time series data refers to the real-time data collected for the online prediction and inference stage. Its data structure and modality category are the same as those of the first multimodal physiological time series data, but the time range corresponds to the recent historical window before the current moment.

[0038] A prediction time window refers to the time interval from the current moment to a predetermined future moment (e.g., the next 60 minutes). A predicted blood glucose trajectory is a continuous trajectory composed of blood glucose concentration values ​​at a series of predicted time points within the prediction time window, rather than a single isolated value point.

[0039] In its implementation, the early warning device acquires real-time second-modal physiological time-series data of the target individual within the current time window, and constructs an input tensor using the same time alignment and aggregation method as in the training phase. This input tensor is then fed into the trained blood glucose prediction model. The blood glucose prediction model first segments the input sequence into segments and performs historical similarity band retrieval based on dynamic time warping through a segmented attention feature extraction subnetwork to generate a current context hidden state tensor. Then, the current context hidden state tensor, physiological intervention event data, and corresponding time information are input into a continuous-time evolution subnetwork of neural frequent differential equations. The hidden state is integrally derived in the continuous time domain to obtain the predicted hidden state trajectory covering the prediction time window. Finally, the predicted hidden state trajectory is input into a decoding subnetwork, which outputs the predicted blood glucose values ​​at each prediction time point within the prediction time window, and these values ​​are then strung together to form the predicted blood glucose trajectory.

[0040] Step S50: Generate blood glucose warning information based on the predicted blood glucose trajectory, and send the blood glucose warning information to the warning terminal associated with the target individual. The warning terminal includes wearable smart terminals and mobile terminals.

[0041] It should be noted that blood glucose warning information refers to a notification message generated based on predicted blood glucose trajectory and individualized safety threshold determination, alerting the patient or their caregiver to the potential risk of future high or low blood glucose levels. This message may include the risk type, the expected occurrence time, and recommended measures. The warning terminal refers to an electronic device carried by the patient or capable of receiving messages in a timely manner, such as wearable smart devices like smartwatches and mobile devices like smartphones.

[0042] Understandably, this step transforms continuous numerical prediction results into timely warning signals, enabling patients to receive advance notice before their blood glucose levels actually deviate from the safe range, thus gaining a critical decision-making window for taking preventative measures (such as supplementing carbohydrates or adjusting insulin infusions).

[0043] In practice, the early warning device acquires the preset individualized blood glucose thresholds corresponding to the target individual, including the preset upper limit threshold (e.g., 10.0 mmol / L) and the preset lower limit threshold (e.g., 3.9 mmol / L).

[0044] The system iterates through the predicted blood glucose values ​​at each predicted time point in the predicted blood glucose trajectory, comparing each value with a preset upper and lower blood glucose threshold, and recording the predicted time when a threshold breach occurs. When it detects that the predicted blood glucose value will breach the preset upper blood glucose threshold within a future predicted time window, a high blood glucose warning is generated; when it detects that the value will breach the preset lower blood glucose threshold, a low blood glucose warning is generated.

[0045] The warning information can include the predicted breakthrough time, the predicted rate of blood glucose change, and the risk level. The device pushes the blood glucose warning information to the target individual's associated smartwatch and smartphone via mobile communication network or Bluetooth, and the warning terminal's application alerts the target individual via sound, vibration, or pop-up window.

[0046] This embodiment establishes a common temporal information basis for blood glucose changes and external intervention factors through multimodal physiological data acquisition, expands complex abnormal monitoring samples through chaotic noise enhancement, enables the blood glucose prediction model to extract stable metabolic evolution features from perturbation inputs through supervised training, generates blood glucose prediction results for multiple future moments through online trajectory extrapolation, and transforms the prediction results into outputtable risk information through terminal early warning, thereby improving the stability of blood glucose prediction and early warning in complex monitoring environments.

[0047] Based on the first embodiment described above, in the second embodiment of this application, the same or similar content as the first embodiment can be referred to the above description, and will not be repeated hereafter.

[0048] Based on this, please refer to Figure 2 , Figure 2 This is a flowchart illustrating the second embodiment of the data-enhanced dynamic blood glucose warning method of this application. In this embodiment, step S10 further includes: Step S101: Send a data collection authorization request to the continuous blood glucose monitoring device, wearable smart terminal and mobile terminal associated with the target individual.

[0049] It should be noted that continuous glucose monitoring devices refer to wearable medical devices that measure interstitial fluid glucose concentration in real time using a subcutaneous sensor probe. A data acquisition authorization request refers to a communication handshake signal initiated by the device to the data source device in accordance with data privacy protection regulations, requesting access to physiological data within a specified time range.

[0050] In some embodiments, the warning device scans for continuous glucose monitoring devices, smartwatches, and smartphones associated with the target individual via wireless communication protocols (such as Bluetooth Low Energy). When a paired device is detected, a data collection authorization request message is sent to each device, containing the requested data time span and a list of data categories. By establishing a data authorization link that complies with privacy protection standards, the compliance and integrity of multi-source physiological data collection are ensured.

[0051] Step S102: In response to the data acquisition authorization request, a data acquisition permission feedback is obtained, and continuous blood glucose monitoring data for historical periods is acquired from the continuous blood glucose monitoring device to obtain a continuous blood glucose monitoring time series.

[0052] In practice, after receiving data acquisition permission feedback from each device, the early warning device establishes a data stream transmission channel with the continuous blood glucose monitoring device and reads each sensor record in the historical period in ascending order of timestamp. The recorded content includes the sampling timestamp and glucose concentration value, which are combined into a continuous blood glucose monitoring time series.

[0053] Step S103: Obtain historical records of physiological intervention events from the wearable smart terminal and the mobile terminal.

[0054] It should be noted that physiological intervention event records refer to the attribute descriptions of external behavioral events that affect glucose metabolism, including dietary intake records, drug intervention records, and exercise records.

[0055] Dietary intake records may include carbohydrate intake and timing; medication intervention records may include insulin injection dosage and timing; exercise records may include exercise type, duration, and average heart rate.

[0056] In some embodiments, the early warning device obtains exercise records and heart rate data from a health application via an application programming interface (API), and obtains manually recorded dietary intake records and drug intervention records from a smartphone, and organizes the records into a physiological intervention event record set according to the event timestamp.

[0057] Step S104: Construct a reference time axis based on the sampling times in the continuous blood glucose monitoring time series, and map the physiological intervention event records to the sampling windows corresponding to the reference time axis according to the corresponding event occurrence times to obtain the aligned physiological intervention event time series.

[0058] It should be noted that the baseline time axis refers to a uniform time grid constructed with the sampling times of the continuous glucose monitoring device as the scale. The sampling window refers to the time interval between two adjacent sampling times.

[0059] In some embodiments, the early warning device compares the timestamps of each event in the physiological intervention event record with the sampling windows on the reference time axis. If the event occurrence time falls within a certain sampling window, the event is marked as occurring at the sampling time corresponding to that sampling window. For cases where there are multiple events within the same sampling window, the parameters of each event are retained separately. After alignment, the time series of physiological intervention events at each sampling time on the reference time axis is obtained.

[0060] Step S105: Identify missing sampling points in the continuous blood glucose monitoring time series based on the reference time axis, and perform temporal interpolation on the missing sampling points to obtain the completed continuous blood glucose monitoring time series.

[0061] It should be noted that missing sampling points refer to sampling moments that should exist on the reference time axis but are actually missing due to equipment transmission interruption or temporary sensor detachment.

[0062] In some embodiments, the early warning device traverses each sampling moment on the reference time axis. When it detects that the blood glucose value of a certain sampling moment is missing, it takes the blood glucose values ​​of a preset number (e.g., 2) of valid sampling moments before and after the missing sampling moment, calculates the missing moment's completion value using linear interpolation, and fills it in to obtain the completed continuous blood glucose monitoring time series.

[0063] Step S106: Aggregate the completed continuous blood glucose monitoring time series and the aligned physiological intervention event time series along the time step dimension, feature variable dimension and modality category dimension to obtain the first multimodal physiological time series data.

[0064] It should be noted that the feature variable dimension refers to the dimension occupied by different physiological features (such as blood glucose level and heart rate) at the same sampling time; the modality category dimension refers to the dimension occupied by different data acquisition modalities (such as continuous monitoring modality and event recording modality).

[0065] In its implementation, to accurately capture the complex physiological changes in glucose in the human body, the early warning device first acquires the real-time time series of subcutaneous interstitial fluid glucose concentration for the target individual through continuous glucose monitoring. Simultaneously, wearable sensors and mobile terminals are used to synchronously collect multimodal intervention data reflecting patient behavior and physiological health. This data includes, but is not limited to: the timing and estimated grams of carbohydrate intake (dietary behavior), the timing and dosage of insulin injection (pharmaceutical intervention), and heart rate and steps (exercise expenditure) that characterize physical activity.

[0066] To address the issues of inconsistent data frequencies and highly random event triggering (such as sparse and irregular discrete events related to diet and medication) during actual data collection, preliminary spatiotemporal alignment is required. Using the sampling period of data collected by the continuous glucose monitoring device as the baseline time axis, discrete physiological and behavioral intervention events are mapped to the nearest sampling window based on their occurrence time. For short-term missing data points caused by device transmission failures, conventional temporal interpolation algorithms are used for preliminary filling to ensure the continuity of the time axis.

[0067] Considering that physiological data in modern high-precision monitoring scenarios are often not simple matrices, but rather tensors containing more dimensional information, after data alignment, historical blood glucose concentrations, time-encoded features, and sparse dietary and drug intervention records are deeply aggregated along the time step dimension, feature variable dimension, and modality category dimension to construct a high-order multidimensional physiological feature tensor. This tensor-based dataset construction method replaces the traditional two-dimensional matrix concatenation, not only accommodating sensor signals at different sampling rates without loss, but also more completely preserving the complex spatial and topological relationships between variables, thus providing high-quality initial state input for subsequent steps such as network training and continuous-time evolution of neural network differential equations.

[0068] Based on the above embodiments, please refer to Figure 3 , Figure 3 This is a flowchart illustrating the third embodiment of the data-enhanced dynamic blood glucose early warning method of this application. In the third embodiment of this application, step S20 further includes: Step S201: Based on the difference between each sampling time of the continuous blood glucose monitoring time series in the first multimodal physiological time series and the reference blood glucose trajectory, generate blood glucose monitoring perturbation samples.

[0069] It should be noted that the reference blood glucose trajectory refers to a blood glucose sequence obtained through reference methods (such as venous blood testing) or high-quality sensor data during a stable wearing period, and is considered a relative true value. Blood glucose monitoring perturbation samples refer to a sequence consisting of the deviation between the sensor measurement value and the reference value at each sampling time, used to analyze the statistical characteristics of sensor perturbation in actual wearing environments.

[0070] Understandably, during actual continuous blood glucose monitoring measurements, due to sensor probe aging, nighttime body pressure, or sweat wetting caused by strenuous exercise, the acquired signals frequently exhibit extreme spikes and peak distortions. Such data exhibits a characteristic where the probability of extreme large / small values ​​of random variables is much higher than that of a normal distribution ("heavy-tailed" characteristic). Traditional data augmentation methods based on Gaussian white noise cannot realistically simulate this extreme jump interference. This embodiment aims to construct a noise resistance mechanism for the model by introducing Lorentz noise with "heavy-tailed" characteristics.

[0071] Step S202: Perform feature analysis on the blood glucose monitoring perturbation sample to obtain the target blood glucose perturbation feature, and set the system parameters of the Lorentz chaotic system based on the target blood glucose perturbation feature.

[0072] It should be noted that the target blood glucose perturbation characteristics refer to key parameters that reflect the statistical distribution of perturbation samples, such as perturbation amplitude variance, perturbation frequency bandwidth, and extreme value occurrence rate.

[0073] The system parameters include state response speed parameters, disturbance excitation intensity parameters, feedback decay parameters, initial state parameters, and numerical solution step size.

[0074] Among them, the state response speed parameter controls the response speed of the linear coupling between state variables in the Lorentz system; the disturbance excitation intensity parameter controls the strength of the system's nonlinear drive and determines whether the system is in a chaotic state; the feedback decay parameter controls the decay characteristics of the third state variable and affects the duration of large amplitude fluctuations in the noise sequence; the initial state parameter is the starting condition for numerical solution; and the numerical solution step size is the time step of numerical integration.

[0075] It should be understood that this step matches the generation parameters of Lorentz chaotic noise with the perturbation characteristics of the actual device worn by the target individual, so that the subsequently generated artificial noise is statistically close to the noise distribution of the real sensor.

[0076] In some embodiments, the early warning device statistically analyzes the mean and standard deviation of blood glucose monitoring disturbance samples, sets the disturbance excitation intensity parameter based on the standard deviation to ensure that the amplitude range of the generated chaotic noise is consistent with the amplitude range of the actual disturbance, and statistically analyzes the frequency of extreme values ​​(more than three times the standard deviation) in the disturbance samples, adjusting the feedback attenuation parameter accordingly to match the heavy-tailedness of the generated noise to this frequency. The preset state response speed parameter is 10.0, the numerical solution step size is 0.01, and the initial state parameters are randomly generated within a preset range.

[0077] Step S203: Numerically solve the Lorentz chaotic system based on the system parameters to obtain the main disturbance state variable sequence, the disturbance driving state variable sequence, and the feedback state variable sequence.

[0078] It should be noted that the sequence of main disturbance state variables corresponds to the discrete-time sequence of the first state variable in the Lorentz system, the sequence of disturbance-driven state variables corresponds to the discrete-time sequence of the second state variable, and the sequence of feedback state variables corresponds to the discrete-time sequence of the third state variable.

[0079] Step S204: Delete the state variables corresponding to the preset transient time period from the main disturbance state variable sequence, and extract the remaining state variables to obtain the basic chaotic noise sequence.

[0080] It should be noted that the preset transient period refers to the transitional stage in which the system evolves from the initial state and has not yet entered a stable chaotic attractor state. The state sequence within this period does not have typical chaotic statistical characteristics.

[0081] It should be understood that this step preserves the state sequence of the system after it enters steady-state chaos by discarding data from the transient period, ensuring that the noise has stable heavy-tailed statistical distribution characteristics. The early warning device can delete the first preset number (e.g., 1000) of the sampling points of the main disturbance state variable sequence and directly extract the remaining state variable sequence as the basic chaotic noise sequence.

[0082] In this implementation, to accurately simulate the heavy-tailed distortion caused by physical disturbances in a real monitoring environment, Lorentz noise is introduced. Compared to traditional random noise, Lorentz noise exhibits nonlinearity, self-similarity, and complex frequency distribution characteristics. Specifically, the noise is generated using a Lorentz chaotic system, and its dynamic evolution is described by the following set of ordinary differential equations: in, , , The system's state variables, x, y, and z, evolve coupled over time t, jointly describing the instantaneous state of the nonlinear system. Numerical solutions to the above equations yield the time series of x, y, and z. The sequence of the principal disturbance state variable x is used as the core source of the basic Lorentz noise signal; y and z, as coupled driving variables, influence the dynamics of x through formulas (1) and (3), resulting in a nonlinear, self-similar, and complex frequency distribution characteristic containing fluctuations in different frequency bands. It is the coordinated evolution of x, y, and z that distinguishes Lorentz noise from simple Gaussian white noise, enabling a more realistic simulation of the occasional large jumps (i.e., "heavy-tailed") in the sensor signal of a continuous glucose monitoring device when subjected to physical pressure, probe aging, or sweat wetting.

[0083] , and The hyperparameter σ controls the state of the chaotic system; it controls the speed at which the state variable x follows y, determining the strength of the linear coupling between variables. The value of σ directly affects the "jitter" frequency of the noise sequence. A larger σ results in a faster convergence of x to the changing trend of y, leading to more rapid changes in the generated noise signal waveform; a smaller σ results in relatively smooth noise fluctuations. By adjusting σ, the frequency characteristics of the generated noise can be matched to the interference patterns in different scenarios of real blood glucose monitoring signals, such as simulating rapid fluctuations caused by strenuous exercise or slow drifts caused by sleep apnea. In a classic implementation, σ can be set to 10.

[0084] ρ is a control parameter for the nonlinear driving force of the system, determining whether the Lorentz system enters a chaotic state. When ρ exceeds a critical value, the system exhibits extreme sensitivity to initial conditions and long-term unpredictable oscillatory behavior. The value of ρ determines the probability and amplitude of large spikes in the noise sequence. The larger ρ is, the wider the range of chaotic attractors, and the greater the oscillation amplitude of x, y, and z, which is reflected in the generated noise sequence as more frequent and stronger "extreme spikes." This directly simulates abnormal signals such as pseudo-hypoglycemia and pseudo-hypoglycemia generated when the sensor is subjected to external pressure or aging. By setting a ρ value that matches the statistical characteristics of real monitoring disturbances, the augmented training dataset can be made closer to the actual wearing environment of the target individual.

[0085] β controls the decay rate of the state variable z, affecting the stability of the system's motion in the z direction. β influences the evolution of z through formula (3), and then indirectly affects the fluctuation pattern of x through formulas (2) and (1). β determines the duration and decay mode of large fluctuations in the noise sequence: the larger β is, the faster z decays, the shorter the duration of large oscillations in the chaotic trajectory, and the frequency and amplitude of extreme values ​​in the heavy-tailed distribution are suppressed; the smaller β makes extreme fluctuations last longer. This parameter is the key lever for regulating the degree of noise "heavy tail" and can be used to match the persistence characteristics of abnormal sensor signals under different wearing environments (for example, pressure-induced pseudohypoglycemia may last for tens of minutes, while the spikes caused by sweat are mostly brief peaks).

[0086] The early warning equipment extracts the state variable sequence from the nonlinear system by numerically solving it, and uses it as the basis for the Lorentz noise signal.

[0087] Step S205: Center, normalize, and scale the noise amplitude of the basic chaotic noise sequence to obtain Lorentz chaotic noise.

[0088] In practice, the early warning equipment calculates the mean and standard deviation of the basic chaotic noise sequence, subtracts the mean from the sequence value and divides it by the standard deviation to obtain the standardized noise sequence, and then multiplies the standardized noise sequence by a preset noise amplitude to obtain the Lorentz chaotic noise.

[0089] Step S206: Generate a noise injection mask corresponding to the continuous blood glucose monitoring time series based on a preset noise injection probability.

[0090] It should be noted that the preset noise injection probability refers to the probability value used to independently determine whether to inject noise at each sampling time. For example, it can be set to 0.3, which means that on average, about 30% of the sampling points are injected with noise. The noise injection mask is a binary sequence of the same length as the continuous blood glucose monitoring time series. A value of 1 indicates that noise is injected at that sampling time, and a value of 0 indicates that the original value is retained.

[0091] Step S207: Based on the noise injection mask, the Lorentz chaotic noise is superimposed onto the continuous blood glucose monitoring time series in the first multimodal physiological time series data to complete the data augmentation of the first multimodal physiological time series data.

[0092] In the specific implementation, after obtaining the generated Lorentz noise sequence, it is used as an artificially simulated extreme physical interference and repeatedly superimposed onto the previously constructed lossless original multidimensional physiological feature tensor according to a set ratio and random probability. Through this tensor-level noise injection, a complex scenario of multiple sensors facing extreme noise at the same time step can be simulated, thereby generating an enhanced feature tensor containing a large number of virtual "pseudo-hypoglycemia / hyperglycemia" spikes. Subsequently, the enhanced tensor set with extreme noise is fused and concatenated with the original tensor set to form a robust training dataset with a significantly expanded scale.

[0093] This embodiment guides the parameter setting of the Lorentz chaotic system by statistically analyzing noise characteristics from actual monitored disturbances, making the generated artificial noise approximate real sensor noise in terms of amplitude, heavy tails, and frequency characteristics. Through transient rejection and amplitude scaling, the noise sequence can be directly superimposed on blood glucose data. By controlling the noise injection density through random masking, a rich training set with both noisy and noise-free samples is constructed, providing a high-fidelity and diverse data foundation for establishing robustness during the model training phase.

[0094] Based on the above embodiments, in the flowchart of the fourth embodiment of the data-enhanced dynamic blood glucose early warning method of this application, step S30 further includes: Step S301: Based on the sliding time window, the data-enhanced first multimodal physiological time-series data and the corresponding first multimodal physiological time-series data are divided into samples to obtain a training dataset, which includes training input sequences and corresponding real blood glucose trajectory labels.

[0095] In some embodiments, the early warning device can set the sliding time window length to the past 8 hours, the prediction time window length to the future 60 minutes, and the sliding step size to 30 minutes. Starting from the beginning of the data, the augmented multidimensional physiological features within the current window are extracted as the training input sequence, and the continuous blood glucose monitoring time series within 60 minutes after the end of the window are extracted as the real blood glucose trajectory label. The window is slid forward 30 minutes along the time axis, and the above extraction operation is repeated until the end of the data. All samples combined constitute the training dataset.

[0096] Step S302: Train the pre-built blood glucose prediction model based on the training dataset. The blood glucose prediction model includes a segmented attention feature extraction module, a neural constant differential equation continuous-time evolution module, and a decoding module. In its implementation, the early warning device utilizes the augmented training dataset to drive subsequent forward propagation and backward parameter updates of the network based on attention mechanisms and neural network differential equations. In a training environment containing Lorentz noise, the model is forced to search for the true patterns of data evolution amidst extreme local perturbations, thus abandoning overfitting to isolated transition points and instead relying on the overall metabolic trends (band characteristics) of the context for deduction. This data augmentation strategy endows the prediction system with strong anti-interference immunity at the underlying data input level, ensuring that it can maintain stable and high-precision early warning output even when facing complex wearing environments of real patients in the future.

[0097] The training process for the blood glucose prediction model includes: Step S3021: Input the training input sequence into the segmented attention feature extraction module to obtain the context hidden state tensor; Step S3022: Input the context hidden state tensor, the physiological intervention event data in the training input sequence and the corresponding time information into the continuous-time evolution module of the neural constant differential equation to obtain the training prediction hidden state trajectory; Step S3023: Input the training predicted hidden state trajectory into the decoding module to obtain the training predicted blood glucose trajectory; Step S3024: Calculate the prediction loss function value based on the difference information between the training predicted blood glucose trajectory and the actual blood glucose trajectory label; Step S3025: Iteratively update the model parameters of the blood glucose prediction model based on the predicted loss function value. When the iteratively updated blood glucose prediction model meets the iteration stopping condition, output the trained blood glucose prediction model.

[0098] It should be noted that the segmented attention feature extraction module refers to a neural network submodule that can segment the input time series into time segments, encode features within each time segment, and fuse features from historically similar bands through an attention mechanism. The neural network constant-time evolution module refers to a submodule that uses neural network parameterization functions to fit the rate of change of the hidden state and deduce the trajectory of the hidden state in the continuous time domain through numerical integration. The decoding module refers to a submodule that nonlinearly maps the high-dimensional hidden state tensor to a predicted blood glucose value through a fully connected layer.

[0099] Furthermore, in order to accurately extract dynamic metabolic features, the segmented attention feature extraction module is configured to perform the following steps: Step S30211: Divide the training input sequence along the time axis into target prediction bands and multiple historical candidate metabolic bands; Step S30212: Perform feature encoding on the target predicted band and each historical candidate metabolic band to obtain the target band feature sequence and multiple historical band feature tensors; Step S30213: Based on the dynamic time warping algorithm, perform time alignment and distance calculation on the target band feature sequence and the feature tensors of each historical band respectively to obtain multiple dynamic warping distances; Step S30214: Filter a preset number of historical band feature tensors according to the order of the dynamic regularization distance from smallest to largest to obtain a set of historical similar band feature tensors; Step S30215: Perform exponential mapping and normalization on the dynamic regularization distance corresponding to the set of historical similar band feature tensors to obtain the dynamic attention weights corresponding to each historical similar band feature tensor. Step S30216: Based on the dynamic attention weights, the set of historical similar band feature tensors is weighted and fused to obtain the contextual hidden state tensor.

[0100] It is important to note that time series modeling and prediction techniques are currently crucial for solving the problem of continuous online blood glucose prediction. Traditional time series prediction methods are diverse, but each has significant limitations. Methods based on statistics or shallow machine learning, such as simple autoregressive imputation and prediction, while computationally simple and easy to implement, completely ignore the inherent nonlinear relationships and time lag effects between physiological variables. Deep learning-based methods, such as those using generative adversarial networks or long short-term memory networks, can learn complex patterns and time dependencies in data by constructing complex neural network models, achieving good prediction results on controlled datasets. However, these deep learning models are often structurally rigid and lack resistance mechanisms when dealing with data containing real-world extreme noise, making them prone to prediction collapse. Transformer models based on attention mechanisms, considering the global correlation of long-term series, have become one of the mainstream methods in the current prediction field. However, existing transformer models, when processing complex metabolic data, often only focus on the correlation mapping between discrete time points, ignoring the inherent band evolution and time lag effects of the human metabolic absorption process, making it difficult to capture the continuity of physiological characteristics. Meanwhile, traditional methods are mostly based on the assumption of static discrete time steps, ignoring the continuous physiological correlations of variables such as eating and insulin injection over time in the real world, and are difficult to overcome irregular sampling intervals. This lack of continuous dynamic metabolic perception and personalized feature matching mechanisms makes it difficult for models to effectively distinguish between local mutations that conform to physiological laws and random hardware noise, resulting in the prediction results still needing breakthroughs in terms of accuracy and robustness.

[0101] Understandably, the segmented attention feature extraction module receives the multidimensional physiological feature tensor containing Lorentz noise from the above steps and inputs it into the reconstructed feature extraction network. Addressing the limitation of traditional transformer models that rely solely on the correlation between discrete time points, making them ill-suited for data like blood glucose data that requires consideration of evolution over a period, the following mechanism is used to extract and fuse dynamic metabolic features: 1. Segmented Mapping and Local Time Segment Construction of Continuous Time Series: In the feature encoding stage, the system no longer processes isolated discrete blood glucose sampling points one by one. Instead, it introduces a time segmentation window to divide the continuous multidimensional physiological feature tensor into multiple time series data segments with fixed time spans along the time axis. In a physical and physiological sense, each time series data segment is no longer a numerical value lacking context, but a complete representation of a local physiological metabolic wave (e.g., the trajectory of the rising phase of carbohydrate absorption after a meal, or the trajectory of the rapid falling phase after insulin injection). The model then flattens and linearly maps the features within these data segments to generate query vectors representing the current local trend and rate of change, as well as key and value vectors representing historical features.

[0102] 2. Retrieval of Historical Metabolic Trajectory Based on Dynamic Time Warping: Due to the significant time lag and nonlinear changes in the metabolic response to diet or medication by different individuals or the same individual under different physiological states (e.g., today's postprandial glucose peak may be 15 minutes later than yesterday), using Euclidean distance with strictly aligned time scales to retrieve historical features would lead to serious misjudgments of physiological states. Therefore, this embodiment innovatively introduces a dynamic time warping algorithm. Using the target prediction band extracted at the current time step, a similarity search is performed in the patient's historical feature database. The dynamic time warping algorithm overcomes the time shift of physiological responses by performing nonlinear elastic stretching and compression on the time axis, accurately calculating the dynamic similarity distance between the current band and historical bands, and filtering out… The historical metabolic sequence segment with the highest dynamic similarity. Specifically, the target predicted band feature sequence truncated by the current sliding window is defined as... (in (where the time step length is the current band), and a candidate metabolic band sequence in the historical database is defined as... (in The time step length of the historical band, due to the metabolic lag effect, and They are often not equal.

[0103] To find the optimal alignment path between these two physiological bands, a... Distance matrix Elements in the matrix Indicates the current band number Physiological state at any moment Compared with historical bands Physiological state at any moment The Euclidean distance between them.

[0104] Subsequently, a continuous regular path is found in the distance matrix using a dynamic programming algorithm. , where the elements on the path Alignment mapping representing time points. The system solves for the true metabolic trajectory similarity by minimizing the cumulative distance along this path, and its optimization objective is defined as: In this formula, the regularized path In a physical sense, this represents the elastic time alignment of the human metabolic rate at the physiological response level. For example, it elastically maps the steep waveform of a rapid postprandial blood glucose rise to a historically elongated, smoother blood glucose rise waveform due to slower digestion. The minimum value calculated using the above formula... This is the similarity distance of pure physiological evolutionary trends after eliminating the interference of time offset. The system uses this distance to dynamically assign attention weights to historical sequences; the smaller the distance of a historical segment, the more exponentially it increases its contribution weight to the current predicted hidden state.

[0105] 3. Dynamic allocation of attention weights and fusion of hidden states: After obtaining... After finding the most matching historical similar sequence segment, the model does not use the traditional attention allocation method of calculating the inner product. Instead, it directly uses the similarity distance calculated by the dynamic time planning algorithm to allocate attention weights to the historical encoded information.

[0106] First, the current predicted band With the Similar historical bands Minimized regular distance between Converted into physiological similarity score To prevent gradient vanishing due to excessive distance, a hyperparameter adjustment mechanism is used. Exponential mapping function: Then, the softmax function was used to analyze this. The similarity scores of the candidate bands are normalized according to their probability distribution, and the probability distribution of the i-th band is calculated. The dynamic attention weights of each historical band in the current prediction task : Considering that actual monitoring integrates multi-dimensional signals such as blood glucose concentration, dietary intake, and drug dosage, the extracted features are often not simple matrices, but rather tensors containing more dimensional information. Therefore, we define... For the first The model outputs the context hidden state tensor for the current time step through a weighted summation of tensor-level features corresponding to each historical metabolic band. : because Determined entirely by the similarity of continuous physiological bands, this mechanism forces the network to automatically filter out interference nodes that differ from the current trend during each inference, accurately extract and fuse the patient's historically extremely similar personalized metabolic tensor features, thereby providing a highly faithful and variable-correlation-rich dynamic initial state for subsequent steps of continuous physiological inference.

[0107] Furthermore, to improve the model's robustness in handling irregular events and missing data in real-world scenarios, the aforementioned continuous-time evolution module for the neural ordinary differential equation is configured to perform the following steps: Step S30221: Set the initial hidden state tensor at the current moment based on the context hidden state tensor; Step S30222: Encode the physiological intervention event data and the corresponding time information to obtain an external driving tensor. The external driving tensor includes the time encoding features corresponding to dietary intake data, drug intervention data and exercise data. Step S30223: Input the initial hidden state tensor, the external driving tensor, and the training prediction time points into the metabolic state derivative network to perform continuous physiological hidden state evolution and obtain the hidden state change rate corresponding to each training prediction time point. Step S30224: Use a numerical integrator to perform continuous-time integration on the rate of change of the hidden state corresponding to each training prediction time point to obtain the training prediction hidden state trajectory.

[0108] Understandably, the continuous-time evolution module of the God's frequent differential equations receives the contextual hidden state tensor output by the fusion of the above steps. The data is then input into the continuous-time evolution module for extrapolation. Traditional discrete-time models typically force time discretization, requiring features to be updated at fixed sampling intervals. This mechanism is highly susceptible to causing gaps in continuous physiological information and loss of the true rate of metabolic change when dealing with irregular intervention events (such as random eating or insulin injection). This embodiment achieves high-fidelity extrapolation in the continuous-time domain through the following mechanism: 1. Metabolic Derivative Fitting Based on Neural Frequent Differential Equations: This embodiment introduces a neural frequent differential equation evolution module. This module no longer directly maps the hidden state of the next time step, but instead utilizes neural network parameterization functions. (i.e., metabolic state derivative network) is used to fit the derivative (i.e., rate of change) of the hidden state tensor over time. Its continuous physiological evolution equation is defined as: Where t represents the continuous physiological evolution time; This represents the physiological hidden state tensor corresponding to a continuous time t. Physiological hidden state tensor The instantaneous rate of change relative to continuous time t; This represents a metabolic state derivative network parameterized by model parameters θ; This represents the external driving tensor corresponding to continuous time t. This is an external driving tensor containing irregular intervention events (diet, medication, etc.). In a physical and physiological sense, the network function... The fit is not simply a numerical mapping difference, but rather the instantaneous physiological metabolic rate in which the glycemic effect of carbohydrate absorption and the hypoglycemic effect of insulin metabolism interact during the human metabolic process.

[0109] 2. Continuous integral derivation based on numerical solver: After determining the metabolic derivative network, the system uses the current time... Hidden state tensor As initial conditions, for the future target time The physiological state is deduced. This process calls an ordinary differential numerical solver to perform integral calculations in the continuous time domain: in, Indicates the start time of continuous blood glucose prediction; Represents the initial hidden state tensor; Indicates the target prediction time within the prediction time window; This represents the predicted physiological hidden state tensor corresponding to the target prediction time. Indicates the integration interval The integral time variable within; Indicates the integration time The corresponding intermediate physiological hidden state tensor; Indicates the integration time The corresponding external driving tensor; The metabolic state derivative network represents the integral time... The calculated rate of change in the physiological latent state.

[0110] The above formula The continuous integration process is perfectly equivalent to the cumulative dynamic evolution of metabolism in the human body at the physical level. It gives the model extremely high temporal flexibility, enabling it to adaptively span arbitrary lengths and irregular sampling intervals, accurately fuse and smoothly transition between random mutation events that occur between two measurements.

[0111] 3. Decoding the hidden state tensor and continuous blood glucose prediction output: The hidden state tensor at any future target time is derived using an ordinary differential numerical solver. The system then inputs this high-dimensional tensor into a fully connected decoding network. The decoding network performs nonlinear dimensionality reduction and feature decoding on the tensor, which contains future evolutionary trends, ultimately mapping and outputting specific continuous blood glucose prediction values ​​for the target time. This mechanism not only eliminates errors caused by non-uniform sampling but also effectively extends the time window for high-precision prediction, providing an extremely reliable physiological evolution benchmark for subsequent early warning decisions.

[0112] Based on the above embodiments, in the flowchart of the fifth embodiment of the data-enhanced dynamic blood glucose early warning method of this application, step S40 further includes: Step S401: Based on the continuous blood glucose monitoring sampling time in the second multimodal physiological time series data, perform time alignment on the physiological intervention event data in the second multimodal physiological time series data to obtain aligned second multimodal physiological time series data; Step S402: Input the aligned second multimodal physiological time series data into the trained blood glucose prediction model to obtain the predicted blood glucose trajectory of the prediction time window; Step S403: The trained blood glucose prediction model is configured to perform the following steps: Step S404: Input the aligned second multimodal physiological time series data into the segmented attention feature extraction module to obtain the current context hidden state tensor; Step S405: Perform feature encoding on the physiological intervention event data and corresponding time information in the aligned second multimodal physiological time series data to obtain the current external driving tensor; Step S406: Input the current context hidden state tensor, the current external driving tensor, and the prediction time point within the prediction time window into the continuous-time evolution module of the neural constant differential equation to obtain the predicted hidden state trajectory. Step S407: Input the predicted hidden state trajectory into the decoding module and output the predicted blood glucose trajectory of the prediction time window.

[0113] It should be noted that the current context hidden state tensor refers to the hidden state representation that integrates historical similar metabolic band information, output by the segmented attention feature extraction module based on real-time input data during the inference phase; the current external driving tensor refers to the external driving signal encoded by real-time physiological intervention event data during the inference phase; and the predicted hidden state trajectory refers to the hidden state sequence covering the prediction time window, derived by the continuous-time evolution module of the neural constant differential equation based on the real-time initial hidden state and the external driving tensor during the inference phase.

[0114] It should be understood that this embodiment applies the metabolic evolution knowledge learned during the training phase to real-time data, and completes blood glucose trajectory prediction through three serial steps: segmented attention retrieval of historical similar patterns, continuous derivation of neural constant differential equations, and decoding mapping. This achieves end-to-end online reasoning from historical data to future trajectories.

[0115] In its implementation, the early warning device inputs the aligned second-modal physiological time-series data into the segmented attention feature extraction module of the trained blood glucose prediction model. This module segments the real-time data along the time axis, performs feature encoding and historical similar band retrieval, and weighted fusion to obtain the current context hidden state tensor. Simultaneously, it performs feature encoding on the physiological intervention event data and corresponding time information in the second-modal physiological time-series data to obtain the current external driving tensor. Using the current context hidden state tensor as initial conditions, it, along with the current external driving tensor and the predicted time points within the prediction time window, is input into the continuous-time evolution module of the neural frequent differential equation. This module calculates the rate of change of the hidden state corresponding to each prediction time point through a metabolic state derivative network and uses a numerical integrator to integrally deduce the predicted hidden state trajectory in the continuous time domain. The predicted hidden state trajectory is input into the decoding module, and after nonlinear mapping through a fully connected layer, the predicted blood glucose value for each prediction time point is output, constituting the predicted blood glucose trajectory of the prediction time window.

[0116] Based on the above embodiments, in the flowchart of the sixth embodiment of the data-enhanced dynamic blood glucose early warning method of this application, step S50 further includes: Step S501: Obtain the preset individualized blood glucose threshold corresponding to the target individual, wherein the preset individualized blood glucose threshold includes a preset upper limit blood glucose threshold and a preset lower limit blood glucose threshold; Step S502: Compare the predicted blood glucose values ​​corresponding to each predicted time point in the predicted blood glucose trajectory with the preset upper limit threshold and the preset lower limit threshold respectively to obtain a blood glucose threshold comparison result sequence; Step S503: Based on the blood glucose threshold comparison result sequence, determine the predicted threshold breakthrough type, predicted threshold breakthrough time, predicted threshold breakthrough magnitude, and predicted blood glucose change rate; Step S504: Determine the blood glucose risk type and blood glucose risk level based on the predicted threshold breakthrough type, the predicted threshold breakthrough time, the predicted threshold breakthrough magnitude, and the predicted blood glucose change rate; Step S505: Generate the blood glucose warning information based on the blood glucose risk type, the blood glucose risk level, the predicted threshold breach time, and the predicted blood glucose trajectory, and send the blood glucose warning information to the warning terminal associated with the target individual.

[0117] In practical implementation, to address the lag limitation of traditional static monitoring, which can only issue alarms after a hazard has occurred, this embodiment achieves dynamic proactive early warning and closed-loop control through the following mechanism: 1. Generation of Future Continuous Blood Glucose Dynamic Trajectory: The system does not simply output isolated predicted values ​​at a single moment. Instead, it generates a smooth and high-fidelity future blood glucose evolution trajectory curve based on the continuous-time evolution module of the neural network's ordinary differential equations, producing a continuous predicted hidden state tensor within the target time window. Simultaneously, by incorporating the tensor characteristic variance from the network computation process, the system synchronously outputs the confidence interval of this predicted trajectory, providing a mathematical basis for assessing the tolerance of subsequent medical interventions.

[0118] 2. Multi-level proactive early warning based on personalized thresholds: The system compares the generated future blood glucose dynamic trajectory with the patient's personalized safety thresholds (such as personalized target range time standards for hypoglycemia and postprandial hyperglycemia). If the predicted trajectory shows a mathematical trend of exceeding the hyperglycemia or hypoglycemia threshold within the next 30 to 60 minutes, the system will trigger a proactive early warning mechanism in advance. This mechanism automatically classifies and determines the risk level (such as extremely dangerous hypoglycemia precursor, general postprandial hyperglycemia) based on the predicted rate and magnitude of exceeding the threshold, and sends an early warning signal to the patient or medical staff through a smart terminal.

[0119] 3. Feedforward Compensation and Intelligent Linkage Between the Closed-Loop Insulin Pump: Furthermore, in scenarios involving artificial pancreas or intelligent drug delivery, the system outputs the future blood glucose evolution trajectory and prediction error tensor as prior environment state input to the model predictive control or reinforcement learning algorithm module. Since this predicted trajectory incorporates irregular multimodal events such as dietary intake and exercise expenditure in advance, the control system can calculate the optimal basal insulin rate adjustment or large-dose booster strategy before actual drastic fluctuations in blood glucose occur, dynamically fine-tuning the insulin pump's infusion state. This mechanism fundamentally achieves interference perception and feedforward compensation, ensuring the safety of closed-loop control while maximizing the maintenance of the patient's actual blood glucose within a healthy and stable range.

[0120] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the data-enhanced dynamic blood glucose early warning method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0121] This application also provides a data-enhanced dynamic blood glucose warning device, please refer to... Figure 4 , Figure 4 This is a schematic diagram of the module structure of a data-enhanced dynamic blood glucose warning device, which includes: The historical data acquisition module 10 is used to acquire the first multimodal physiological time series data of the target individual during historical periods. The first multimodal physiological time series data includes the continuous blood glucose monitoring time series and the physiological intervention event time series collected during historical periods. The data augmentation module 20 is used to generate Lorentz chaotic noise based on the Lorentz chaotic system, and to perform data augmentation on the first multimodal physiological time series data based on the Lorentz chaotic noise. The model training module 30 is used to construct a training dataset based on the data-augmented first multimodal physiological time-series data, train a pre-constructed blood glucose prediction model based on the training dataset, and obtain the trained blood glucose prediction model. The blood glucose trajectory prediction module 40 is used to input the second multimodal physiological time series data collected within the current time window of the target individual into the trained blood glucose prediction model to obtain the predicted blood glucose trajectory of the prediction time window. The blood glucose dynamic early warning module 50 is used to generate blood glucose early warning information based on the predicted blood glucose trajectory and send the blood glucose early warning information to the early warning terminal associated with the target individual. The early warning terminal includes a wearable smart terminal and a mobile terminal.

[0122] The data-enhanced dynamic blood glucose warning device provided in this application employs the data-enhanced dynamic blood glucose warning method described in the above embodiments, and can solve the technical problems to be addressed by the aforementioned data-enhanced dynamic blood glucose warning method. Compared with the prior art, the beneficial effects of the data-enhanced dynamic blood glucose warning device provided in this application are the same as those of the data-enhanced dynamic blood glucose warning method provided in the above embodiments, and other technical features in the data-enhanced dynamic blood glucose warning device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0123] This application provides a data-enhanced dynamic blood glucose warning device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to execute the data-enhanced dynamic blood glucose warning method in the first embodiment described above.

[0124] The following is for reference. Figure 5The diagram illustrates a structural schematic suitable for implementing the data-enhanced dynamic blood glucose warning device according to the embodiments of this application. The data-enhanced dynamic blood glucose warning device in the embodiments of this application can be deployed in wearable smart terminals such as smartphones, tablets, and smartwatches, portable medical data processing terminals, or remote medical servers. Figure 5 The data-enhanced dynamic blood glucose warning device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0125] like Figure 5 As shown, the data-enhanced dynamic blood glucose warning device may include a processing unit 1001 (e.g., a central processing unit, graphics processing unit, neural network processor, etc.), which can perform various appropriate actions and processes according to a program stored in the read-only memory ROM 1002 or a program loaded from the storage device 1003 into the random access memory RAM 1004. The RAM 1004 also stores various programs and data required for the operation of the data-enhanced dynamic blood glucose warning device, such as blood glucose prediction model parameters, individualized blood glucose thresholds, historical multimodal physiological time-series data, etc. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output interface (I / O) 1006 is also connected to the bus.

[0126] Typically, the following systems can be connected to I / O interface 1006: input device 1007, including, for example, a touchscreen, a microphone, and a wireless communication module (e.g., Bluetooth module, WiFi module) for receiving physiological data sent by continuous glucose monitoring devices and wearable smart terminals; output device 1008, including, for example, a liquid crystal display (LCD), a speaker, and a vibrator for issuing warning alerts; storage device 1003, including, for example, flash memory and solid-state drive; and communication device 1009. Communication device 1009 allows the data-enhanced dynamic glucose monitoring device to communicate wirelessly or wiredly with other devices to exchange data, such as receiving manually recorded dietary and medication information from patients via mobile communication networks and pushing glucose warning information to warning terminals. Although the figure shows a data-enhanced dynamic glucose monitoring device with various systems, it should be understood that it is not required to implement or possess all the systems shown; alternatively, more or fewer systems may be implemented.

[0127] The data-enhanced dynamic blood glucose warning device provided in this application employs the data-enhanced dynamic blood glucose warning method described in the above embodiments, and can solve the technical problems to be addressed by the aforementioned data-enhanced dynamic blood glucose warning method. Compared with the prior art, the beneficial effects of the data-enhanced dynamic blood glucose warning device provided in this application are the same as those of the data-enhanced dynamic blood glucose warning method provided in the above embodiments, and other technical features of this data-enhanced dynamic blood glucose warning device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0128] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0129] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0130] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the data-enhanced dynamic blood glucose warning method in the above embodiments.

[0131] The computer-readable storage medium provided in this application can be a tangible medium capable of storing computer programs, such as a hard disk, random access memory, read-only memory, flash memory, optical storage device, or magnetic storage device. When the computer program is executed by a processor, it implements the above-described data-enhanced dynamic blood glucose warning method.

[0132] The aforementioned computer-readable storage medium may be included in a data-enhanced dynamic blood glucose warning device; or it may exist independently and not be assembled into a data-enhanced dynamic blood glucose warning device.

[0133] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the data-enhanced dynamic blood glucose warning device, cause the data-enhanced dynamic blood glucose warning device to perform the data-enhanced dynamic blood glucose warning method as described in the embodiments above.

[0134] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including LAN (Local Area Network) or WAN (Wide Area Network)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0135] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0136] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the aforementioned data-enhanced dynamic blood glucose warning method, thereby solving the technical problems to be addressed by the aforementioned data-enhanced dynamic blood glucose warning method. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as the beneficial effects of the data-enhanced dynamic blood glucose warning method provided in the above embodiments, and will not be repeated here.

[0137] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A data-enhanced dynamic blood glucose early warning method, characterized in that, The data-enhanced dynamic blood glucose early warning method includes: Acquire the first multimodal physiological time series data of the target individual for historical periods, including continuous blood glucose monitoring time series and physiological intervention event time series collected in historical periods; Lorentz chaotic noise is generated based on the Lorentz chaotic system, and the first multimodal physiological time series data is augmented based on the Lorentz chaotic noise. A training dataset is constructed based on the data-augmented first multimodal physiological time-series data. A pre-constructed blood glucose prediction model is trained based on the training dataset to obtain the trained blood glucose prediction model. The second multimodal physiological time-series data collected within the current time window of the target individual is input into the trained blood glucose prediction model to obtain the predicted blood glucose trajectory of the prediction time window; Based on the predicted blood glucose trajectory, a blood glucose warning message is generated and sent to the warning terminal associated with the target individual. The warning terminal includes wearable smart terminals and mobile terminals.

2. The data-enhanced dynamic blood glucose early warning method as described in claim 1, characterized in that, The acquisition of the first multimodal physiological time-series data of the target individual for historical periods includes: Send data collection authorization requests to the continuous glucose monitoring devices, wearable smart terminals, and mobile terminals associated with the target individual; In response to the data acquisition authorization request, a data acquisition permission feedback is obtained, and continuous blood glucose monitoring data for historical periods is acquired from the continuous blood glucose monitoring device to obtain a continuous blood glucose monitoring time series; The wearable smart terminal and the mobile terminal acquire historical records of physiological intervention events, including dietary intake records, drug intervention records, and exercise records. A baseline time axis is constructed based on the sampling times in the continuous blood glucose monitoring time series, and the physiological intervention event records are mapped to the sampling windows corresponding to the baseline time axis according to the corresponding event occurrence times to obtain the aligned physiological intervention event time series. Based on the reference time axis, the missing sampling points in the continuous blood glucose monitoring time series are identified, and the missing sampling points are interpolated temporally to obtain the completed continuous blood glucose monitoring time series. The first multimodal physiological time series data is obtained by aggregating the completed continuous blood glucose monitoring time series and the aligned physiological intervention event time series along the time step dimension, feature variable dimension, and modality category dimension.

3. The data-enhanced dynamic blood glucose early warning method as described in claim 1, characterized in that, The process of generating Lorentz chaotic noise based on a Lorentz chaotic system and performing data augmentation on the first multimodal physiological time-series data based on the Lorentz chaotic noise includes: Based on the difference between each sampling time of the continuous blood glucose monitoring time series in the first multimodal physiological time series and the reference blood glucose trajectory, a blood glucose monitoring perturbation sample is generated. Feature analysis is performed on the blood glucose monitoring perturbation samples to obtain target blood glucose perturbation features, and system parameters of the Lorentz chaotic system are set based on the target blood glucose perturbation features. The system parameters include state response speed parameters, perturbation excitation intensity parameters, feedback decay parameters, initial state parameters, and numerical solution step size. Numerical solutions are obtained for the Lorentz chaotic system based on the system parameters, yielding the main disturbance state variable sequence, the disturbance-driven state variable sequence, and the feedback state variable sequence. Delete the state variables corresponding to the preset transient time period from the main disturbance state variable sequence, and extract the remaining state variables to obtain the basic chaotic noise sequence; The basic chaotic noise sequence is centered, normalized, and its noise amplitude is scaled to obtain Lorentz chaotic noise; A noise injection mask corresponding to the continuous blood glucose monitoring time series is generated based on a preset noise injection probability. Based on the noise injection mask, the Lorentz chaotic noise is superimposed onto the continuous blood glucose monitoring time series in the first multimodal physiological time series data to complete the data augmentation of the first multimodal physiological time series data.

4. The data-enhanced dynamic blood glucose early warning method as described in claim 1, characterized in that, The process involves constructing a training dataset based on data-augmented first multimodal physiological time-series data, training a pre-built blood glucose prediction model based on the training dataset, and obtaining the trained blood glucose prediction model, including: Based on a sliding time window, the data-enhanced first multimodal physiological time-series data and the corresponding first multimodal physiological time-series data are divided into samples to obtain a training dataset, which includes training input sequences and corresponding real blood glucose trajectory labels. The pre-built blood glucose prediction model is trained based on the training dataset. The blood glucose prediction model includes a segmented attention feature extraction module, a continuous-time evolution module of neural ordinary differential equations, and a decoding module. The training process of the blood glucose prediction model includes: The training input sequence is input into the segmented attention feature extraction module to obtain the context hidden state tensor; The contextual hidden state tensor, the physiological intervention event data in the training input sequence and the corresponding time information are input into the continuous-time evolution module of the neural constant differential equation to obtain the training prediction hidden state trajectory. The training predicted hidden state trajectory is input into the decoding module to obtain the training predicted blood glucose trajectory; The prediction loss function value is calculated based on the difference information between the training-predicted blood glucose trajectory and the actual blood glucose trajectory label. The model parameters of the blood glucose prediction model are iteratively updated based on the predicted loss function value. When the iteratively updated blood glucose prediction model meets the iteration stopping condition, the trained blood glucose prediction model is output.

5. The data-enhanced dynamic blood glucose early warning method as described in claim 4, characterized in that, The segmented attention feature extraction module is configured to perform the following steps: The training input sequence is divided along the time axis into a target prediction band and multiple historical candidate metabolic bands; The target predicted band and each historical candidate metabolic band are respectively feature encoded to obtain the target band feature sequence and multiple historical band feature tensors; Based on the dynamic time warping algorithm, the target band feature sequence and each historical band feature tensor are time aligned and distances are calculated to obtain multiple dynamic warping distances. Based on the dynamic regularization distance, a preset number of historical band feature tensors are selected in ascending order to obtain a set of historical similar band feature tensors. The dynamic regularization distance corresponding to the set of historical similar band feature tensors is subjected to exponential mapping and normalization to obtain the dynamic attention weights corresponding to each historical similar band feature tensor. The set of historical similar band feature tensors is weighted and fused based on the dynamic attention weights to obtain the contextual hidden state tensor.

6. The data-enhanced dynamic blood glucose early warning method as described in claim 4, characterized in that, The continuous-time evolution module for the neural ordinary differential equation is configured to perform the following steps: The initial hidden state tensor at the current moment is set based on the context hidden state tensor; The physiological intervention event data and the corresponding time information are feature-encoded to obtain an external driving tensor, which includes the time-encoded features corresponding to dietary intake data, drug intervention data and exercise data. The initial hidden state tensor, the external driving tensor, and the training prediction time points are input into the metabolic state derivative network to perform continuous physiological hidden state evolution, obtaining the hidden state change rate corresponding to each training prediction time point. The continuous physiological hidden state evolution process is described by the following formula: Where t represents the continuous physiological evolution time; This represents the physiological hidden state tensor corresponding to a continuous time t. Physiological hidden state tensor The instantaneous rate of change relative to continuous time t; This represents a metabolic state derivative network parameterized by model parameters θ; This represents the external driving tensor corresponding to continuous time t; By using a numerical integrator to continuously integrate the rate of change of the hidden state at each training prediction time point, the training prediction hidden state trajectory is obtained, as shown in the following formula: in, Indicates the start time of continuous blood glucose prediction; Represents the initial hidden state tensor; Indicates the target prediction time within the prediction time window; This represents the predicted physiological hidden state tensor corresponding to the target prediction time. Indicates the integration interval The integral time variable within; Indicates the integration time The corresponding intermediate physiological hidden state tensor; Indicates the integration time The corresponding external driving tensor; The metabolic state derivative network represents the integral time... The calculated rate of change in the physiological latent state.

7. The data-enhanced dynamic blood glucose early warning method as described in any one of claims 1 to 6, characterized in that, The step of inputting the second multimodal physiological time-series data collected within the current time window of the target individual into the trained blood glucose prediction model to obtain the predicted blood glucose trajectory for the prediction time window includes: Based on the continuous blood glucose monitoring sampling time in the second multimodal physiological time series data, the physiological intervention event data in the second multimodal physiological time series data are time-aligned to obtain aligned second multimodal physiological time series data; The aligned second multimodal physiological time series data is input into the trained blood glucose prediction model to obtain the predicted blood glucose trajectory within the prediction time window. The trained blood glucose prediction model is configured to perform the following steps: The aligned second multimodal physiological time-series data is input into the segmented attention feature extraction module to obtain the current context hidden state tensor; The physiological intervention event data and corresponding time information in the aligned second multimodal physiological time series data are feature-encoded to obtain the current external driving tensor; The current context hidden state tensor, the current external driving tensor, and the prediction time point within the prediction time window are input into the continuous-time evolution module of the neural ordinary differential equation to obtain the predicted hidden state trajectory. The predicted hidden state trajectory is input into the decoding module, and the predicted blood glucose trajectory of the prediction time window is output.

8. The data-enhanced dynamic blood glucose early warning method as described in claim 1, characterized in that, The step of generating blood glucose warning information based on the predicted blood glucose trajectory and sending the blood glucose warning information to the warning terminal associated with the target individual includes: Obtain the preset individualized blood glucose threshold corresponding to the target individual, wherein the preset individualized blood glucose threshold includes a preset upper limit threshold and a preset lower limit threshold; The predicted blood glucose values ​​corresponding to each predicted time point in the predicted blood glucose trajectory are compared with the preset upper blood glucose threshold and the preset lower blood glucose threshold, respectively, to obtain a blood glucose threshold comparison result sequence; Based on the blood glucose threshold comparison result sequence, the predicted threshold breakthrough type, predicted threshold breakthrough time, predicted threshold breakthrough magnitude, and predicted blood glucose change rate are determined. The blood glucose risk type and blood glucose risk level are determined based on the predicted threshold breakthrough type, the predicted threshold breakthrough time, the predicted threshold breakthrough magnitude, and the predicted blood glucose change rate. The blood glucose warning information is generated based on the blood glucose risk type, the blood glucose risk level, the predicted threshold breach time, and the predicted blood glucose trajectory, and the blood glucose warning information is sent to the warning terminal associated with the target individual.

9. A data-enhanced dynamic blood glucose warning device applying the data-enhanced dynamic blood glucose warning method according to any one of claims 1 to 8, characterized in that, The device includes: The historical data acquisition module is used to acquire the first multimodal physiological time series data of the target individual during historical periods. The first multimodal physiological time series data includes the continuous blood glucose monitoring time series and the physiological intervention event time series collected during historical periods. The data augmentation module is used to generate Lorentz chaotic noise based on the Lorentz chaotic system, and to perform data augmentation on the first multimodal physiological time series data based on the Lorentz chaotic noise. The model training module is used to construct a training dataset based on the data-augmented first multimodal physiological time-series data, train a pre-constructed blood glucose prediction model based on the training dataset, and obtain the trained blood glucose prediction model. The blood glucose trajectory prediction module is used to input the second multimodal physiological time series data collected within the current time window of the target individual into the trained blood glucose prediction model to obtain the predicted blood glucose trajectory of the prediction time window. The blood glucose dynamic early warning module is used to generate blood glucose early warning information based on the predicted blood glucose trajectory, and send the blood glucose early warning information to the early warning terminal associated with the target individual. The early warning terminal includes wearable smart terminals and mobile terminals.

10. A data-enhanced dynamic blood glucose early warning device, characterized in that, The data-enhanced dynamic blood glucose warning device includes: a memory, a processor, and a data-enhanced dynamic blood glucose warning program stored in the memory. The processor is used to run the data-enhanced dynamic blood glucose warning program, which is configured to implement the data-enhanced dynamic blood glucose warning method as described in any one of claims 1 to 8.

Images