Miniaturized blood glucose intelligent monitoring and analysis method based on integrated edge computing

Abstract

The application relates to the cross field of computer technology and biomedical engineering, and discloses a miniaturized blood glucose intelligent monitoring and analysis method integrated with edge computing. The method comprises the following steps: collecting glucose concentration and time sequence multi-source physiological signals through a wearable sensor; performing data segmentation and multi-modal feature alignment based on a circadian rhythm; inputting the features into a lightweight biological model obtained through knowledge distillation to perform blood glucose prediction; and combining a memory feedback mechanism to adaptively adjust the reasoning frequency. The system comprises a sensing unit, a preprocessing unit, a lightweight model execution unit and a risk warning unit. While ensuring prediction accuracy, the application compresses the model to within 150,000 parameters, adapts to 512 KB on-chip memory, and realizes low-power, high-stability end-side continuous blood glucose monitoring and warning.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of computer technology and biomedical engineering, and specifically relates to a miniaturized intelligent blood glucose monitoring and analysis method integrating edge computing. Background Technology

[0002] With the continued growth in demand for chronic disease management, blood glucose monitoring technology is rapidly developing towards intelligence, continuous monitoring, and personalization. Traditional blood glucose monitoring relies on finger-prick blood sampling and discrete measurements, making it difficult to capture dynamic trends in blood glucose levels. While emerging continuous glucose monitoring (CGM) systems can provide time-series data streams, their data analysis largely depends on cloud servers. This centralized architecture faces significant bottlenecks in practical applications: on the one hand, data uploads suffer from network latency and privacy risks; on the other hand, cloud models typically employ highly complex deep neural networks to improve prediction accuracy, resulting in enormous computational and storage overhead, making direct deployment on resource-constrained wearable or portable edge devices impossible.

[0003] Miniaturized intelligent blood glucose monitoring integrating edge computing has become a key path to achieving localized real-time analysis. This approach aims to embed blood glucose trend prediction models into low-power, low-memory edge terminals, reducing communication dependencies and improving response speed while ensuring user data privacy. However, existing deep learning models generally have a large number of parameters and high memory consumption. When deployed to typical edge devices (such as microcontrollers based on ARM Cortex-M series or low-order RISC-V architecture), they are prone to triggering system memory overflow due to model loading exceeding the available RAM capacity, leading to frequent device restarts and complete loss of continuous monitoring capabilities.

[0004] While pruning, knowledge distillation, or 8-bit quantization are commonly used techniques for lightweighting models, these methods often overlook the physiological characteristics of blood glucose time-series data and the strict constraints of edge hardware. Structured pruning lacks a protection mechanism for key physiological feature channels, leading to distorted prediction trends; while standard quantization training does not consider the nonlinear error accumulation effect of fixed-point computing units on edge devices, causing a significant decrease in accuracy under low-bit representation.

[0005] Existing solutions fail to coordinate and optimize model structure, training strategies and hardware deployment processes, making it difficult for lightweight models to run stably and maintain clinically acceptable prediction accuracy under limited memory (usually less than 512KB) and low computing power (<100DMIPS). Summary of the Invention

[0006] This invention provides a miniaturized intelligent blood glucose monitoring and analysis method integrating edge computing, aiming to solve the technical problems of memory overflow, device restart, and inability to achieve continuous monitoring caused by excessive model parameters when deploying blood glucose prediction artificial intelligence models on edge computing devices. This invention constructs a lightweight multi-stage feature distillation architecture, combined with a dynamic sparse activation mechanism and a context-aware data compression strategy based on physiological rhythms. While ensuring the accuracy of blood glucose prediction, it significantly reduces the memory footprint and computational complexity of the model during runtime, enabling the blood glucose prediction model to run stably, continuously, and with low power consumption on resource-constrained edge computing devices.

[0007] This invention provides a miniaturized intelligent blood glucose monitoring and analysis method integrating edge computing, comprising:

[0008] The wearable continuous glucose monitoring sensor collects real-time time-series data on glucose concentration in the user's subcutaneous interstitial fluid.

[0009] Acquire multi-source physiological signals synchronized with the glucose concentration time-series data, including heart rate variability data, skin temperature data, body acceleration data, and ambient temperature and humidity data;

[0010] The glucose concentration time-series data is preprocessed using a segmented sliding window based on the diurnal rhythm cycle to generate a blood glucose fragment sequence with a fixed time span.

[0011] The multi-source physiological signals are normalized at the channel level and aligned in the time domain to form a multimodal physiological feature matrix that is strictly synchronized with the blood glucose fragment sequence;

[0012] The multimodal physiological feature matrix is ​​input into a lightweight student model trained by a teacher model. The lightweight student model adopts a hybrid structure of depthwise separable convolution and gated recurrent units, and embeds a dynamic sparse activation module to output a sequence of predicted blood glucose concentration values ​​within a future preset time window.

[0013] During the operation phase of the edge computing device, only the lightweight student model is loaded and executed, while an adaptive inference frequency adjustment mechanism based on runtime memory pressure feedback is enabled to keep the system memory usage below a preset safety threshold.

[0014] As one embodiment of the present invention, the wearable continuous blood glucose monitoring sensor is a miniature implantable sensing probe based on the electrochemical principle of glucose oxidase. Its sampling frequency is set to once every 5 minutes, the output resolution is 0.1 millimoles per liter, the power supply is provided by a miniature flexible thin-film battery, and the data is transmitted to the edge computing device through a near-field communication protocol.

[0015] In one embodiment of the present invention, the acquisition of the multi-source physiological signals is accomplished by a multi-sensor array integrated in the same wearable wristband. The heart rate variability data is obtained by a photoplethysmography pulse wave sensor that collects the raw signal at a sampling rate of 100 Hz and then extracts the low-frequency to high-frequency power ratio through fast Fourier transform. The skin temperature data is output by a digital temperature sensor with an accuracy of 0.5 degrees Celsius. The body acceleration data is collected by a triaxial microelectromechanical system accelerometer at a sampling rate of 50 Hz and converted into a displacement energy index through integration. The environmental temperature and humidity data is updated by an integrated temperature and humidity sensing chip at a frequency of once every 10 seconds.

[0016] As one embodiment of the present invention, the segmented sliding window preprocessing based on the circadian rhythm cycle specifically includes: taking 24 hours as a complete physiological cycle, dividing the continuously collected glucose concentration time series data into two sub-intervals according to the local clock: the daytime activity period and the nighttime sleep period; within each sub-interval, using a sliding window with a length of 6 hours and a step size of 30 minutes to extract blood glucose segments; each blood glucose segment contains 72 continuous sampling points, and is attached with a circadian phase label corresponding to the start time of the segment, wherein the circadian phase label has a value of 0 or 1, representing nighttime or daytime respectively.

[0017] As one embodiment of the present invention, the channel-level normalization and time-domain alignment specifically include: independently calculating the mean and standard deviation of each type of physiological signal in the historical data of the most recent 7 days, and using this to standardize the current collected value; for physiological signals with a sampling frequency higher than that of the glucose sensor, using linear interpolation to downsample them to the same 5-minute interval as the glucose data; for signals with a sampling frequency lower than that of the glucose sensor, using the method of keeping the last valid value for upsampling and filling, to ensure that all physiological signals are strictly aligned with the blood glucose segment in the time dimension.

[0018] In one embodiment of the present invention, the teacher model is a deep temporal convolutional network with more than 2 million parameters. Its input is a complete multimodal physiological feature matrix, and its output is a blood glucose concentration prediction curve for the next 3 hours. The lightweight student model has no more than 150,000 parameters. Its network structure consists of three depthwise separable convolutional blocks connected in series with two gated recurrent unit layers. Each depthwise separable convolutional block includes a one-dimensional convolutional layer with a channel-wise convolutional kernel size of 3, a pointwise convolutional layer, a batch normalization layer, and a modified linear unit activation function.

[0019] The dynamic sparse activation module is set after each gated recurrent unit layer. It dynamically closes some neuron connections based on the L1 norm of the current input features. The closing ratio is controlled by a preset sparsity threshold, which is dynamically adjusted according to the current available memory capacity of the device.

[0020] As one embodiment of the present invention, the training process of the lightweight student model includes the following steps: first, the teacher model is trained on a cloud server using a large-scale historical blood glucose and physiological dataset until convergence;

[0021] Then, with the teacher model parameters fixed, the same dataset is input into both the teacher and student models. The mean squared error loss between the output prediction curves of the two models and the KL divergence loss between the intermediate layer features of the student model and the corresponding layer features of the teacher model are calculated. The two loss functions are jointly optimized, and the student model parameters are updated through the backpropagation algorithm until the root mean square error of the student model on the validation set is less than 0.8 mmol / L.

[0022] As one embodiment of the present invention, the adaptive inference frequency adjustment mechanism specifically includes: embedding a memory monitoring agent in the operating system kernel of the edge computing device to collect the heap memory usage of the model inference process in real time; when the memory usage is detected to exceed 80% of the total available memory, extending the trigger interval of model inference from the default 30 minutes to 45 minutes; when the memory usage drops below 60%, restoring the default inference interval; the adjustment of the inference interval is achieved by modifying the timer parameters in the task scheduler to ensure that the system is always in a memory-safe operating state.

[0023] As one embodiment of the present invention, the edge computing device is a dedicated neural network acceleration coprocessor integrated into the main control chip of a smartwatch. Its on-chip static random access memory has a capacity of 512 kilobytes, a main frequency of 400 MHz, supports eight-bit integer fixed-point arithmetic, and the model weights and activation values ​​are stored and calculated in eight-bit integer format. The inference process does not depend on external dynamic random access memory.

[0024] As one embodiment of the present invention, after outputting the predicted blood glucose concentration value sequence within a future preset time window, a risk warning generation step is also included: comparing the predicted value sequence with preset low blood glucose threshold of 3.9 mmol / L and high blood glucose threshold of 10.0 mmol / L point by point; if there are more than 3 consecutive data points in the predicted sequence that are lower than the low blood glucose threshold, a first-level low blood glucose warning signal is generated.

[0025] If five or more consecutive data points exceed the hyperglycemia threshold, a level two hyperglycemia warning signal is generated. The warning signal is communicated to the user via a combination of a vibration motor and a screen icon, and is also synchronized to the associated mobile terminal application via Bluetooth Low Energy protocol.

[0026] This invention provides a miniaturized intelligent blood glucose monitoring and analysis system integrating edge computing, comprising:

[0027] A continuous blood glucose monitoring sensor unit is used to collect real-time time-series data on glucose concentration in the user's subcutaneous interstitial fluid.

[0028] A multi-source physiological signal acquisition unit is used to acquire heart rate variability data, skin temperature data, body acceleration data, and ambient temperature and humidity data synchronized with the glucose concentration time-series data.

[0029] The data preprocessing unit is used to perform segmented sliding window preprocessing on the glucose concentration time series data based on the diurnal rhythm cycle, and to perform channel-level normalization and time-domain alignment on the multi-source physiological signals to form a multimodal physiological feature matrix.

[0030] A lightweight blood glucose prediction model execution unit is deployed in an edge computing device to receive the multimodal physiological feature matrix and output a sequence of predicted blood glucose concentration values ​​within a future preset time window. The lightweight blood glucose prediction model execution unit has a built-in dynamic sparse activation module and an adaptive inference frequency adjustment module.

[0031] The risk warning generation unit is used to generate a graded warning signal and output it to the user based on the comparison result between the predicted blood glucose concentration value sequence and a preset threshold.

[0032] As one embodiment of the present invention, the model run by the lightweight blood glucose prediction model execution unit is a student model trained by the knowledge distillation of the teacher model. Its network structure is composed of alternating stacks of depthwise separable convolutional blocks and gated recurrent unit layers. The model parameters are stored in the read-only memory of the edge computing device in an eight-bit integer format. During the inference process, all intermediate activation values ​​are calculated and temporarily stored in the on-chip static random access memory.

[0033] As one embodiment of the present invention, the adaptive inference frequency adjustment module is deeply integrated with the operating system kernel of the edge computing device. It monitors the memory usage of the model inference process in real time and dynamically adjusts the scheduling cycle of the model inference task according to the preset memory safety threshold to ensure that the peak memory usage of the system is always less than 400 kilobytes.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0035] 1. This invention constructs a lightweight, multi-stage feature distillation architecture based on knowledge distillation, which compresses the original blood glucose prediction model with a large number of parameters to less than 150,000 parameters, enabling it to be fully loaded and run on an edge computing device with only 512 kilobytes of on-chip memory, thus completely avoiding the device restart problem caused by memory overflow.

[0036] 2. The introduced dynamic sparse activation module shuts down redundant neuron connections in real time based on the activity of input features, further reducing runtime memory peaks and computational energy consumption. Combined with a context-aware data preprocessing strategy based on circadian rhythms, key physiological rhythm information of blood glucose fluctuations is effectively preserved. While significantly compressing the model, the root mean square error of prediction is controlled within 0.8 mmol / L, meeting the accuracy requirements for clinical use.

[0037] 3. The adaptive inference frequency adjustment mechanism achieves dynamic balancing of inference load through operating system-level memory monitoring, ensuring the stability and reliability of the system during long-term continuous operation. The overall solution realizes a paradigm shift in intelligent blood glucose monitoring from cloud-dependent to edge-autonomous, providing diabetic patients with truly wearable, uninterrupted, and low-power personalized blood glucose management capabilities. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the overall technical architecture of the miniaturized intelligent blood glucose monitoring and analysis method integrating edge computing proposed in this invention.

[0039] Figure 2 This is a schematic diagram illustrating the core principle framework of the lightweight multi-stage feature distillation architecture and dynamic sparse activation mechanism in this invention.

[0040] Figure 3 This is a flowchart illustrating the logical flow of blood glucose data segmentation and sliding window preprocessing based on circadian rhythms and the channel-level normalization alignment of multi-source physiological signals in this invention.

[0041] Figure 4 This is a diagram of the two-stage model construction and execution framework of teacher-student model knowledge distillation training and lightweight student model deployment in this invention;

[0042] Figure 5 This is a diagram illustrating the operational control logic framework of the edge computing device memory monitoring and adaptive inference frequency adjustment mechanism in this invention.

[0043] Figure 6 This is a schematic diagram of the end-to-end system interaction and data flow for terminal-side multimodal data acquisition, predictive reasoning, and hierarchical risk warning generation in this invention. Detailed Implementation

[0044] Please refer to Figures 1 to 6This invention provides a miniaturized intelligent blood glucose monitoring and analysis method integrating edge computing, aiming to solve the technical problems of memory overflow, device restart, and inability to achieve continuous monitoring caused by excessively large model parameters when deploying blood glucose prediction artificial intelligence models on edge computing devices. This method constructs a lightweight multi-stage feature distillation architecture, combined with a dynamic sparse activation mechanism and a context-aware data compression strategy based on physiological rhythms. While ensuring the accuracy of blood glucose prediction, it significantly reduces the memory footprint and computational complexity of the model during runtime, enabling the blood glucose prediction model to run stably, continuously, and with low power consumption on resource-constrained edge computing devices.

[0045] The method includes the following steps:

[0046] S1 collects real-time time-series data of glucose concentration in the user's subcutaneous interstitial fluid through a wearable continuous blood glucose monitoring sensor;

[0047] S2, acquire multi-source physiological signals synchronized with the glucose concentration time-series data, the multi-source physiological signals including heart rate variability data, skin temperature data, body acceleration data, and ambient temperature and humidity data;

[0048] S3, perform segmented sliding window preprocessing on the glucose concentration time series data based on the diurnal rhythm cycle to generate a blood glucose segment sequence with a fixed time span;

[0049] S4, perform channel-level normalization and time-domain alignment on the multi-source physiological signals to form a multimodal physiological feature matrix that is strictly synchronized with the blood glucose fragment sequence;

[0050] S5, the multimodal physiological feature matrix is ​​input into a lightweight student model trained by the teacher model. The lightweight student model adopts a hybrid structure of depthwise separable convolution and gated recurrent unit and embeds a dynamic sparse activation module to output a sequence of predicted blood glucose concentration values ​​within a future preset time window.

[0051] S6, during the operation phase of the edge computing device, only the lightweight student model is loaded and executed, while an adaptive inference frequency adjustment mechanism based on runtime memory pressure feedback is enabled to maintain system memory usage below a preset safety threshold.

[0052] In step S1, a wearable continuous glucose monitoring sensor is used to collect real-time time-series data on the glucose concentration in the user's subcutaneous interstitial fluid. The wearable continuous glucose monitoring sensor is a miniature implantable sensor probe based on the electrochemical principle of glucose oxidase, with a sampling frequency set to once every 5 minutes and an output resolution of 0.1 millimoles per liter. The sensor is powered by a miniature flexible thin-film battery, and the data is transmitted to an edge computing device via a near-field communication protocol.

[0053] The sensor probe is implanted subcutaneously in the user's upper arm. Its working principle involves a specific catalytic reaction between glucose and glucose oxidase immobilized on the sensing membrane, producing hydrogen peroxide. This hydrogen peroxide is then oxidized at the working electrode, generating a current signal. The intensity of this current is linearly related to the glucose concentration in the interstitial fluid. The signal is processed by a preamplifier and an analog-to-digital converter, and then output in digital form.

[0054] To ensure biocompatibility and signal stability during long-term wear, the sensing membrane surface is coated with an anti-protein adsorption coating, and a built-in temperature compensation circuit is incorporated to eliminate the impact of ambient temperature fluctuations on the electrochemical reaction rate. After each sampling, the raw current signal is digitally filtered to remove high-frequency noise and converted into a glucose concentration value according to a pre-calibrated sensitivity coefficient, forming a time-series data stream with precisely aligned timestamps.

[0055] In step S2, multi-source physiological signals synchronized with the glucose concentration time-series data are acquired. These multi-source physiological signals include heart rate variability data, skin temperature data, body acceleration data, and ambient temperature and humidity data. These signals are acquired through a multi-sensor array integrated into the same wearable wristband. Specifically, heart rate variability data is acquired by a photoplethysmography (PPG) sensor at a sampling rate of 100 Hz, collecting raw photoelectric signals. After bandpass filtering to remove baseline drift and motion artifacts, the signal is processed by Fast Fourier Transform (FFT) to extract frequency domain features, and the ratio of low-frequency power (0.04 to 0.15 Hz) to high-frequency power (0.15 to 0.4 Hz) is calculated as the heart rate variability index. Skin temperature data is output by a digital temperature sensor with an accuracy of 0.5 degrees Celsius. This sensor is in close contact with the user's wrist skin and updates its reading every 10 seconds. The body acceleration data is collected by a triaxial microelectromechanical system accelerometer at a sampling rate of 50 Hz, which collects acceleration components in three orthogonal directions. Then, the summation of the squares of the signals of each axis is performed and the square root is taken to obtain the resultant acceleration. The resultant acceleration sequence is then integrated over time and converted into a displacement energy index to characterize the intensity of user activity.

[0056] Ambient temperature and humidity data are updated every 10 seconds by an integrated temperature and humidity sensor chip located on the outside of the wristband, directly exposed to the ambient air. All sensors are uniformly scheduled by the same main microcontroller to ensure that their sampling start time is synchronized with the system's global clock, providing a foundation for subsequent time-domain alignment.

[0057] In step S3, the glucose concentration time-series data undergoes segmented sliding window preprocessing based on the circadian rhythm cycle to generate a blood glucose segment sequence with a fixed time span. This preprocessing uses a 24-hour period as a complete physiological cycle, dividing the continuously collected glucose concentration time-series data into two sub-intervals according to the local clock: the daytime activity period (06:00 to 22:00) and the nighttime sleep period (22:00 to 06:00 the next day). Within each sub-interval, a sliding window with a length of 6 hours and a step size of 30 minutes is used to extract blood glucose segments. Each blood glucose segment contains 72 consecutive sampling points (72 points for 6 hours due to a 5-minute sampling interval), and is accompanied by a circadian phase label corresponding to the segment's starting time. The circadian phase label has a value of 0 or 1, representing nighttime or daytime, respectively. This label is embedded as additional contextual features into the subsequent model input to guide the model in distinguishing blood glucose fluctuation patterns under different physiological states.

[0058] The sliding window operation is performed in real time within the data buffer. Whenever a new sampling point arrives, the buffer is updated and the window generation conditions are checked. If met, all glucose values ​​and their timestamps within the current window are extracted, and phase labels are appended to form a complete blood glucose segment instance. This process effectively transforms an infinitely long time series into fixed-length supervised learning samples while preserving crucial circadian rhythm information, avoiding a decline in model generalization ability due to ignoring physiological periodicity.

[0059] In step S4, the multi-source physiological signals are normalized at the channel level and aligned in the time domain to form a multimodal physiological feature matrix that is strictly synchronized with the blood glucose fragment sequence. First, the mean and standard deviation of each type of physiological signal are independently calculated in the historical data of the most recent 7 days to standardize the current collected value. For example, the current value of heart rate variability data is subtracted from the mean of the user in the same time period (such as at night) over the past 7 days, and then divided by its standard deviation to obtain the normalized dimensionless index.

[0060] This normalization strategy considers individual differences and long-term trend drift, improving the model's cross-user adaptability. Secondly, temporal alignment is performed: for physiological signals with sampling frequencies higher than those of the glucose sensor (such as heart rate variability and body acceleration), linear interpolation is used to downsample them to the same 5-minute interval as the glucose data; for signals with sampling frequencies lower than those of the glucose sensor (such as ambient temperature and humidity, once every 10 seconds, which is actually higher than the glucose sampling rate; this is corrected here by assuming the existence of even lower frequency signals, but according to the material description, all signals are not lower than the glucose sampling rate, so only downsampling is needed in actual processing), upsampling padding is performed by retaining the last valid value.

[0061] Specifically, using the timestamp sequence of blood glucose fragments as the baseline time axis, interpolation or padding is performed on each type of physiological signal at its corresponding time point to ensure that all signals are defined at 72 time points. Finally, the normalized and aligned physiological signals are stacked according to time points to form a 72-row, five-row multimodal physiological feature matrix, where the five rows correspond to glucose concentration, heart rate variability, skin temperature, body acceleration energy index, and ambient temperature and humidity, respectively. This matrix serves as the model input, fully preserving the synchronicity of multi-source signals in the time dimension and the independent statistical characteristics between channels.

[0062] In step S5, the multimodal physiological feature matrix is ​​input into a lightweight student model trained under the guidance of a teacher model. The lightweight student model adopts a hybrid structure of depthwise separable convolution and gated recurrent units, and embeds a dynamic sparse activation module to output a sequence of predicted blood glucose concentration values ​​within a preset future time window.

[0063] The teacher model is a deep temporal convolutional network with over 2 million parameters. Its input is a complete multimodal physiological feature matrix, and its output is a blood glucose concentration prediction curve for the next 3 hours. The lightweight student model has no more than 150,000 parameters, and its network structure consists of 3 depthwise separable convolutional blocks connected in series with 2 gated recurrent unit layers.

[0064] Each depthwise separable convolutional block contains a one-dimensional convolutional layer with a channel-wise convolutional kernel size of 3, a pointwise convolutional layer, a batch normalization layer, and a modified linear unit activation function. The channel-wise convolutional layer performs convolution operations independently on each input channel, significantly reducing the number of parameters;

[0065] Pointwise convolutional layers fuse channel information through 1×1 convolutions. Gated recurrent unit (ROU) layers are used to capture long-term temporal dependencies, and their update and reset gate mechanisms effectively alleviate the vanishing gradient problem. A dynamic sparse activation module is set after each GU layer.

[0066] This module calculates the L1 norm of the input features of the current layer. If the L1 norm of the input channel corresponding to a neuron is lower than a preset sparsity threshold, its output is forcibly set to 0 during the current forward propagation, thereby disabling the activation of that neuron. The sparsity threshold is not fixed but dynamically adjusted based on the device's currently available memory capacity.

[0067] When memory is limited, the threshold is increased to increase sparsity and reduce the number of activated neurons; when memory is ample, the threshold is decreased to retain more feature information. This mechanism takes effect in real time during the inference phase, without requiring model retraining. The training process of the lightweight student model is completed in the cloud: first, the teacher model is trained until convergence; then, the teacher model parameters are fixed, and the same dataset is simultaneously input into both the teacher and student models, calculating the mean squared error loss between their output prediction curves.

[0068]

[0069] in To predict the number of time steps, and The teacher and student models were respectively in the first... The predicted values ​​for each step. Simultaneously, the KL divergence loss between the intermediate layer features of the student model and the corresponding layer features of the teacher model is calculated:

[0070]

[0071] in Let be the probability distribution of the features after softmax normalization. The joint loss function is:

[0072]

[0073] in and The weights are used to update the student model parameters through the backpropagation algorithm until the root mean square error of its prediction on the validation set is less than 0.8 mmol / L.

[0074] In step S6, during the operation phase of the edge computing device, only the lightweight student model is loaded and executed, while an adaptive inference frequency adjustment mechanism based on runtime memory pressure feedback is enabled. The edge computing device is a dedicated neural network acceleration coprocessor integrated into the main control chip of a smartwatch. It has an on-chip static random access memory capacity of 512 kilobytes, a main frequency of 400 MHz, and supports eight-bit integer fixed-point arithmetic.

[0075] Model weights and activation values ​​are stored and calculated using 8-bit integers, and the inference process does not rely on external dynamic random access memory. To prevent memory overflow, a memory monitoring agent is embedded in the operating system kernel to collect the heap memory usage of the model inference process in real time. When memory usage exceeds 80% of the total available memory (410 kilobytes), the trigger interval for model inference is extended from the default 30 minutes to 45 minutes; when memory usage falls back to below 60% (307 kilobytes), the default inference interval is restored. The adjustment of the inference interval is achieved by modifying the timer parameters in the task scheduler.

[0076] This mechanism ensures that the blood glucose prediction task receives sufficient memory resources even under conditions of multi-task concurrency or background application memory consumption, preventing process crashes or device restarts due to insufficient memory. Inference tasks are scheduled as low-priority background services, executing only when the system is idle and memory is plentiful, minimizing the impact on the user experience.

[0077] After outputting the predicted blood glucose concentration sequence within a preset future time window, the process also includes a risk warning generation step. The predicted sequence is compared point-by-point with preset low blood glucose thresholds of 3.9 mmol / L and high blood glucose thresholds of 10.0 mmol / L. If three or more consecutive data points in the predicted sequence are below the low blood glucose threshold, a Level 1 low blood glucose warning signal is generated; if five or more consecutive data points are above the high blood glucose threshold, a Level 2 high blood glucose warning signal is generated.

[0078] The warning signal is delivered to the user via both a vibration motor and a screen icon, and is synchronized to the associated mobile terminal application via Bluetooth Low Energy. The warning logic is executed locally at the edge, without requiring a network connection, ensuring timely alerts even in areas without network coverage.

[0079] Vibration patterns are differentiated based on warning levels: Level 1 warning is a short, single vibration, and Level 2 warning is three consecutive long vibrations. Screen icons are permanently displayed in the status bar, with colors changing according to blood sugar levels: green indicates normal, yellow indicates a warning, and red indicates an emergency.

[0080] Corresponding to the above method, this invention also provides a miniaturized intelligent blood glucose monitoring and analysis system integrating edge computing. This system includes a continuous blood glucose monitoring sensor unit, a multi-source physiological signal acquisition unit, a data preprocessing unit, a lightweight blood glucose prediction model execution unit, and a risk warning generation unit. The continuous blood glucose monitoring sensor unit is used to collect real-time time-series data of glucose concentration in the user's subcutaneous interstitial fluid, and its structure is consistent with the sensor described in step S1 above.

[0081] The multi-source physiological signal acquisition unit is integrated into the wearable wristband and includes a photoplethysmography (PPG) sensor, a digital temperature sensor, a triaxial microelectromechanical system (MEMS) accelerometer, and a temperature and humidity sensor chip. This allows for the simultaneous acquisition of heart rate variability, skin temperature, body acceleration, and ambient temperature and humidity data. The data preprocessing unit is implemented by the microcontroller firmware in the edge computing device, performing circadian rhythm-based sliding window segmentation, channel-level normalization, and temporal alignment operations to output a multimodal physiological feature matrix.

[0082] The lightweight blood glucose prediction model execution unit is deployed in a dedicated neural network acceleration coprocessor. Internally, it stores the lightweight student model trained using knowledge distillation. Model parameters are stored in read-only memory (ROM) in 8-bit integer format. All intermediate activation values ​​are calculated and temporarily stored in a 512-kilobyte on-chip static random access memory (SRAM). This unit incorporates a dynamic sparse activation module that disables redundant connections in real time based on the activity of input features. It also integrates an adaptive inference frequency adjustment module, which is deeply coupled with the operating system kernel. This module monitors memory usage in real time and dynamically adjusts the inference task scheduling cycle to ensure that peak memory usage remains below 400 kilobytes.

[0083] The risk warning generation unit is implemented by an interrupt service routine in the microcontroller. After receiving the prediction result, it immediately performs threshold comparison and continuity judgment, triggering corresponding vibration, display, and wireless communication operations. The entire system can operate continuously for more than 7 days without external power intervention, meeting the daily management needs of diabetic patients.

Claims

1. A miniaturized intelligent blood glucose monitoring and analysis method integrating edge computing, characterized in that, include: The wearable continuous glucose monitoring sensor collects real-time time-series data on glucose concentration in the user's subcutaneous interstitial fluid. Acquire multi-source physiological signals synchronized with the glucose concentration time-series data; The glucose concentration time-series data is preprocessed using a segmented sliding window based on the diurnal rhythm cycle to generate a blood glucose fragment sequence with a fixed time span. The multi-source physiological signals are normalized at the channel level and aligned in the time domain to form a multimodal physiological feature matrix that is strictly synchronized with the blood glucose fragment sequence; The multimodal physiological feature matrix is ​​input into a lightweight student model trained by a teacher model. The lightweight student model adopts a hybrid structure of depthwise separable convolution and gated recurrent units, and embeds a dynamic sparse activation module to output a sequence of predicted blood glucose concentration values ​​within a future preset time window. During the operation phase of the edge computing device, only the lightweight student model is loaded and executed, while an adaptive inference frequency adjustment mechanism based on runtime memory pressure feedback is enabled to keep the system memory usage below a preset safety threshold. The multi-source physiological signals include heart rate variability data, skin temperature data, body acceleration data, and ambient temperature and humidity data. The glucose concentration time-series data is preprocessed using a segmented sliding window based on the circadian rhythm cycle to generate blood glucose segment sequences with fixed time spans, including: Using a 24-hour period as a complete physiological cycle, the continuously collected glucose concentration time-series data were divided into two sub-intervals: the daytime activity period and the nighttime sleep period, according to the local clock. Within each sub-interval, a sliding window with a length of 6 hours and a step size of 30 minutes is used to extract blood glucose segments; Each blood glucose segment contains 72 consecutive sampling points and is accompanied by a day / night phase label corresponding to the start time of the segment. The day / night phase label has a value of 0 or 1, representing night or day, respectively. The multi-source physiological signals are normalized at the channel level and aligned in the time domain to form a multimodal physiological feature matrix that is strictly synchronized with the blood glucose fragment sequence, including: For each type of physiological signal, the mean and standard deviation of the historical data over the most recent seven days are calculated independently to standardize the current collected values. For physiological signals with a sampling frequency higher than that of the glucose sensor, linear interpolation is used to downsample them to the same 5-minute interval as the glucose data. For signals with a sampling frequency lower than that of the glucose sensor, upsampling is performed by retaining the last valid value to ensure that all physiological signals are strictly aligned with blood glucose segments in the time dimension. The multimodal physiological feature matrix is ​​input into a lightweight student model trained under the guidance of a teacher model. This lightweight student model employs a hybrid structure of depthwise separable convolutions and gated recurrent units, and embeds a dynamic sparse activation module to output a sequence of predicted blood glucose concentration values ​​within a preset future time window, including: The lightweight student model's network structure consists of three depthwise separable convolutional blocks connected in series with two gated recurrent unit layers. Each depthwise separable convolutional block contains a one-dimensional convolutional layer with a channel-wise convolutional kernel size of 3, a pointwise convolutional layer, a batch normalized layer, and a modified linear unit activation function; A dynamic sparse activation module is set after each gated recurrent unit layer. It dynamically shuts down some neuron connections based on the L1 norm of the current input features. The shut-down ratio is controlled by a preset sparsity threshold, which is dynamically adjusted according to the current available memory capacity of the device. The wearable continuous blood glucose monitoring sensor is a miniature implantable sensing probe based on the electrochemical principle of glucose oxidase. Its sampling frequency is set to once every 5 minutes, the output resolution is 0.1 millimoles per liter, the power supply is provided by a miniature flexible thin film battery, and the data is transmitted to the edge computing device through a near field communication protocol. The acquisition of the multi-source physiological signals is accomplished through a multi-sensor array integrated in the same wearable wristband. The heart rate variability data is obtained by a photoplethysmography pulse wave sensor that collects the raw signal at a sampling rate of 100 Hz and then extracts the low-frequency to high-frequency power ratio through fast Fourier transform. The skin temperature data is output by a digital temperature sensor with an accuracy of 0.5 degrees Celsius. The body acceleration data is collected by a triaxial microelectromechanical system accelerometer at a sampling rate of 50 Hz and converted into a displacement energy index through integration. The environmental temperature and humidity data is updated by an integrated temperature and humidity sensor chip at a frequency of once every 10 seconds.

2. The miniaturized intelligent blood glucose monitoring and analysis method integrating edge computing according to claim 1, characterized in that, The lightweight student model is a student model trained through knowledge distillation of the teacher model, and its training process includes: A deep temporal convolutional network with over 2 million parameters was trained on a cloud server using a large-scale historical blood glucose and physiological dataset until convergence. With the teacher model parameters fixed, the same dataset is input into the teacher model and the student model. The mean squared error loss between the output prediction curves of the two models and the KL divergence loss between the intermediate layer features of the student model and the corresponding layer features of the teacher model are calculated. The two loss functions mentioned above are jointly optimized, and the student model parameters are updated through the backpropagation algorithm until the root mean square error of the student model's prediction on the validation set is less than 0.8 mmol / L.

3. The miniaturized intelligent blood glucose monitoring and analysis method integrating edge computing according to claim 2, characterized in that, During the operation phase of the edge computing device, only the lightweight student model is loaded and executed, while an adaptive inference frequency adjustment mechanism based on runtime memory pressure feedback is enabled to maintain system memory usage below a preset safety threshold, including: Embed a memory monitoring agent in the operating system kernel of the edge computing device to collect the heap memory usage of the model inference process in real time; When memory usage exceeds 80% of the total available memory, the trigger interval for model inference will be extended from the default 30 minutes to 45 minutes. When memory usage drops below 60%, the default inference interval is restored. The inference interval can be adjusted by modifying the timer parameters in the task scheduler.

4. The miniaturized intelligent blood glucose monitoring and analysis method integrating edge computing according to claim 1, characterized in that, After outputting the predicted blood glucose concentration sequence within a preset future time window, the process also includes a risk warning generation step: The predicted value sequence was compared point by point with the preset low blood sugar threshold of 3.9 mmol / L and high blood sugar threshold of 10.0 mmol / L. If there are more than three consecutive data points below the hypoglycemia threshold in the predicted sequence, a level one hypoglycemia warning signal will be generated. If there are 5 or more consecutive data points that are higher than the hyperglycemia threshold, a level 2 hyperglycemia warning signal will be generated. The warning signal is delivered to the user via both a vibration motor and a screen icon, and is synchronized to the associated mobile terminal application via Bluetooth Low Energy protocol.

Images