Miniaturized blood glucose intelligent monitoring and analysis method integrating edge calculation

Abstract

The invention relates to the crossing field of computer technology and biomedical engineering, and discloses a miniaturized intelligent blood glucose monitoring and analyzing method integrating edge computing. The method comprises the steps that glucose concentration and time sequence multi-source physiological signals are collected through a wearable sensor; performing data segmentation and multi-modal feature alignment based on the circadian rhythm; inputting the features into a lightweight student model obtained through knowledge distillation for blood glucose prediction; and the reasoning frequency is adaptively adjusted in combination with a memory feedback mechanism. The system comprises a sensing unit, a preprocessing unit, a lightweight model execution unit and a risk early warning unit. While the prediction precision is guaranteed, the model is compressed to be within 150 thousand parameters, a 512 KB on-chip memory is adapted, and low-power-consumption and high-stability end-side continuous blood glucose monitoring and early warning are achieved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of computer technology and biomedical engineering, and specifically relates to a miniaturized blood glucose intelligent monitoring and analysis method integrated with edge computing. BACKGROUND

[0002] With the continuous growth of chronic disease management needs, blood glucose monitoring technology is rapidly developing towards intelligence, continuity and personalization. Traditional blood glucose monitoring relies on fingertip blood sampling and discrete measurement, which is difficult to capture the trend of blood glucose dynamic changes. Although the emerging continuous glucose monitoring (CGM) system can provide time series data streams, its data analysis relies on cloud servers for processing. This centralized architecture faces significant bottlenecks in practical applications: on the one hand, data uploading has network delay and privacy leakage risks; on the other hand, cloud models usually use high complexity deep neural networks to improve prediction accuracy, resulting in huge computing and storage overheads, which cannot be directly deployed on resource-constrained wearable or portable edge devices.

[0003] Miniaturized blood glucose intelligent monitoring integrated with edge computing has become a key path to realize local real-time analysis. This direction aims to embed blood glucose trend prediction models into low-power and small-memory edge terminals, while ensuring user data privacy, reducing communication dependence and improving response speed. However, existing deep learning models generally have large parameter quantities and high memory occupancy. When deployed to typical edge devices (such as microcontrollers based on ARM Cortex-M series or low-order RISC-V architecture), the model loading often exceeds the available RAM capacity, triggering system memory overflow and causing frequent device restarts, completely losing continuous monitoring capabilities.

[0004] In terms of model lightweight, although general pruning, knowledge distillation or 8-bit quantization methods are used, these methods often ignore the physiological characteristics of blood glucose time series data and the strict constraints of edge hardware. Structured pruning lacks protection mechanisms for key physiological feature channels, leading to distorted prediction trends. Standard quantization training does not consider the nonlinear error accumulation effect of edge device fixed-point operation units, causing significant accuracy decline in low-bit representation.

[0005] Existing solutions fail to optimize model structure, training strategy and hardware deployment process simultaneously, making it difficult for lightweight models to run stably under limited memory (usually less than 512KB) and low computing power (<100 DMIPS) conditions while maintaining clinically acceptable prediction accuracy. SUMMARY

[0006] The application provides a small-sized blood glucose intelligent monitoring and analysis method integrated with edge computing, aiming to solve the technical problem that the edge computing device cannot realize continuous monitoring due to memory overflow, triggering device restart and excessive model parameters when deploying a blood glucose prediction artificial intelligence model. By constructing a light-weight multi-stage feature distillation architecture, combining a dynamic sparse activation mechanism and a context-aware data compression strategy based on physiological rhythms, the application significantly reduces the model runtime memory occupation and computational complexity while ensuring the accuracy of blood glucose prediction, enabling the blood glucose prediction model to run stably, continuously and with low power consumption on resource-constrained edge computing devices.

[0007] The application provides a small-sized blood glucose intelligent monitoring and analysis method integrated with edge computing, which comprises:

[0008] Real-time collection of subcutaneous interstitial fluid glucose concentration time series data of a user through a wearable continuous blood glucose monitoring sensor;

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

[0010] Segmented sliding window preprocessing of the glucose concentration time series data based on circadian rhythm cycles to generate a blood glucose segment sequence with a fixed time span;

[0011] Channel-level normalization and time domain alignment of the multi-source physiological signals to form a multi-modal physiological feature matrix strictly synchronized with the blood glucose segment sequence;

[0012] Inputting the multi-modal physiological feature matrix into a light-weight student model guided by a teacher model, the light-weight student model adopts a hybrid structure of depth separable convolution and gated recurrent unit and embeds a dynamic sparse activation module for outputting a blood glucose concentration prediction value sequence within a future preset time window;

[0013] In the running phase of the edge computing device, only the light-weight student model is loaded and executed, and an adaptive inference frequency adjustment mechanism based on runtime memory pressure feedback is enabled to maintain the system memory occupation below a preset safety threshold.

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

[0015] As an embodiment of the present application, the acquisition of the multi-source physiological signals is completed by a multi-sensor array integrated in the same wearable wristband, wherein the heart rate variability data is obtained by a photoelectric plethysmogram sensor, the raw signal is sampled at a rate of 100 Hz, and the low-frequency and high-frequency power ratio is extracted by fast Fourier transform; the skin temperature data is output by a digital temperature sensor with an accuracy of 0.5 degrees Celsius; the body motion acceleration data is collected by a three-axis micro-electro-mechanical system accelerometer at a sampling rate of 50 Hz, and is converted into a displacement energy index by integral operation; and the ambient temperature and humidity data are updated by an integrated temperature and humidity sensor chip at a frequency of once every 10 seconds.

[0016] As an embodiment of the present application, the segmented sliding window preprocessing based on the circadian rhythm cycle specifically comprises: dividing the continuously collected glucose concentration time series data into two subintervals of a daytime active period and a nighttime sleep period according to a local clock with 24 hours as a complete physiological cycle; in each subinterval, a sliding window with a length of 6 hours and a step of 30 minutes is used to intercept a blood glucose segment; each blood glucose segment contains 72 consecutive sampling points, and is additionally attached with a circadian phase label corresponding to the starting time of the segment, the circadian phase label takes a value of 0 or 1, representing nighttime or daytime respectively.

[0017] As an embodiment of the present application, the channel-level normalization and time-domain alignment specifically comprises: calculating the mean and standard deviation of each type of physiological signal in the last 7 days of historical data to standardize the current collected value; 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, up-sampling is performed by retaining the last valid value to ensure that all physiological signals are strictly aligned with the blood glucose segment in the time dimension.

[0018] As an embodiment of the present application, the teacher model is a deep time series convolutional network with more than 2 million parameters, the input of which is a complete multi-modal physiological feature matrix, and the output is a blood glucose concentration prediction curve within the next 3 hours; the parameter quantity of the lightweight student model does not exceed 150,000, and the network structure thereof is composed of three depth separable convolution blocks connected in series with two gate recurrent unit layers, each depth separable convolution block contains a one-dimensional convolution layer with a channel-wise convolution kernel size of three, a point-by-point convolution layer, a batch normalization layer and a rectified linear unit activation function;

[0019] The dynamic sparse activation module is arranged after each gate recurrent unit layer, which dynamically turns off part of the neuron connections according to the L1 norm of the current input features, and the closing ratio is controlled by a preset sparsity threshold, and the sparsity threshold is dynamically adjusted according to the current available memory capacity of the device.

[0020] As an embodiment of the present application, the training process of the lightweight student model comprises the following steps: first, training the teacher model on the cloud server using a large-scale historical blood glucose and physiological data set until convergence;

[0021] Then fix the teacher model parameters, input the same data set into the teacher model and the student model, calculate the mean square error loss between the output prediction curves of the two, and the KL divergence loss between the intermediate layer features of the student model and the corresponding layer features of the teacher model; jointly optimize the above two loss functions, update the student model parameters through the back propagation algorithm, until the prediction root mean square error of the student model on the validation set is less than 0.8 mmol / L.

[0022] As an embodiment of the present application, the adaptive inference frequency adjustment mechanism specifically comprises: embedding a memory monitoring agent in the operating system kernel of the edge computing device, and collecting the heap memory usage of the model inference process in real time; when it is detected that the memory usage exceeds 80% of the total available memory, the trigger interval of model inference is extended from the default 30 minutes to 45 minutes; when the memory usage falls below 60%, the default inference interval is restored; the adjustment of the inference interval is realized by modifying the timer parameters in the task scheduler, ensuring that the system is always in a memory-safe running state.

[0023] As an embodiment of the present application, the edge computing device is a special neural network acceleration coprocessor integrated into the main control chip of a smart watch, with an on-chip static random access memory capacity of 512 KB and a main frequency of 400 MHz, supporting eight-bit integer fixed-point operations, and the model weights and activation values are stored and calculated in eight-bit integer format, and the inference process does not rely on external dynamic random access memory.

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

[0025] If there are more than 5 consecutive data points above the high blood glucose threshold, a second-level high blood glucose warning signal is generated; the warning signal is prompted to the user through a dual mode of vibration motor and screen icon, and is synchronized to the associated mobile terminal application program through the Bluetooth low energy protocol.

[0026] The present application provides a small-sized blood glucose intelligent monitoring and analysis system integrated with edge computing, which comprises:

[0027] A continuous blood glucose monitoring sensor unit is used to collect real-time subcutaneous interstitial fluid glucose concentration time series data of a user.

[0028] a multi-source physiological signal acquisition unit for acquiring heart rate variability data, skin temperature data, body motion acceleration data, and environmental temperature and humidity data synchronized with the glucose concentration time series data;

[0029] a data preprocessing unit for segmenting and sliding window preprocessing the glucose concentration time series data based on circadian rhythm cycles, and channel-level normalization and time domain alignment of the multi-source physiological signals to form a multi-modal physiological feature matrix;

[0030] a lightweight blood glucose prediction model execution unit deployed in an edge computing device for receiving the multi-modal physiological feature matrix and outputting a blood glucose concentration prediction value sequence within a future preset time window, the lightweight blood glucose prediction model execution unit having a dynamic sparse activation module and an adaptive inference frequency adjustment module built-in;

[0031] a risk warning generation unit for generating a graded warning signal according to the comparison result of the blood glucose concentration prediction value sequence and a preset threshold and outputting to the user.

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

[0033] As an embodiment of the present application, the adaptive inference frequency adjustment module is deeply integrated with the operating system kernel of the edge computing device, which monitors the memory occupation of the model inference process in real time, and dynamically adjusts the scheduling period of the model inference task according to the preset memory safety threshold, ensuring that the peak memory occupation of the system is always below 400 kilobytes.

[0034] Compared with the prior art, the present application has the following advantages:

[0035] 1. The present application compresses the originally large parameter blood glucose prediction model to within 150,000 parameters by constructing a lightweight multi-stage feature distillation architecture based on knowledge distillation, so that it can be completely loaded and run on an edge computing device with only 512 kilobytes of on-chip memory, completely avoiding the problem of device restart 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 6The application provides a small-sized blood glucose intelligent monitoring and analysis method based on integrated edge computing, aiming to solve the technical problem that the edge computing device cannot realize continuous monitoring due to memory overflow and device restart caused by excessive model parameters when deploying a blood glucose prediction artificial intelligence model. The method constructs a lightweight multi-stage feature distillation architecture, combines a dynamic sparse activation mechanism and a context-aware data compression strategy based on physiological rhythms, significantly reduces the model runtime memory occupation and computational complexity under the premise of ensuring the accuracy of blood glucose prediction, and enables the blood glucose prediction model to run stably, continuously and with low power consumption on resource-constrained edge computing devices.

[0045] The method comprises the following steps:

[0046] S1, collecting subcutaneous interstitial fluid glucose concentration time series data of a user in real time through a wearable continuous glucose monitoring sensor;

[0047] S2, acquiring 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 motion acceleration data and environmental temperature and humidity data;

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

[0049] S4, performing channel-level normalization and time domain alignment on the multi-source physiological signals to form a multi-modal physiological feature matrix strictly synchronized with the blood glucose segment sequence;

[0050] S5, inputting the multi-modal physiological feature matrix into a lightweight student model guided by a teacher model, the lightweight student model adopting a hybrid structure of depth separable convolution and gated recurrent unit and embedding a dynamic sparse activation module, for outputting a blood glucose concentration prediction value sequence within a preset time window in the future;

[0051] S6, in the running stage of the edge computing device, only loading and executing the lightweight student model, while enabling an adaptive inference frequency adjustment mechanism based on runtime memory pressure feedback to maintain the system memory occupation below a preset safety threshold.

[0052] In step S1, the subcutaneous interstitial fluid glucose concentration time series data of a user is collected in real time through a wearable continuous glucose monitoring sensor. The wearable continuous glucose monitoring sensor is a miniature implantable sensing probe based on glucose oxidase electrochemical principle, with a sampling frequency of 1 time per 5 minutes and an output resolution of 0.1 millimoles per liter. The sensor is powered by a miniature flexible thin film battery, and data is transmitted to the edge computing device through near field communication protocol.

[0053] The sensor probe is implanted in the subcutaneous tissue of the user's upper arm. Its working principle is that glucose and glucose oxidase fixed on the sensing membrane undergo specific catalytic reaction to generate hydrogen peroxide, which is oxidized on the working electrode to produce a current signal. The current intensity 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 output in digital form.

[0054] To ensure biocompatibility and signal stability during long-term wear, an anti-protein adsorption coating is applied to the surface of the sensing membrane, and a temperature compensation circuit is built in to eliminate the influence of environmental temperature fluctuations on the electrochemical reaction rate. After each sampling is completed, the original current signal is digitally filtered to remove high-frequency noise, and converted into glucose concentration values based on the pre-calibrated sensitivity coefficient to form a time-stamped and time-aligned data stream.

[0055] In step S2, multi-source physiological signals synchronized with the glucose concentration time series data are obtained. The multi-source physiological signals include heart rate variability data, skin temperature data, body motion acceleration data, and environmental temperature and humidity data. These signals are collected by a multi-sensor array integrated in the same wearable wristband. Among them, the heart rate variability data is collected by a photoplethysmogram sensor at a sampling rate of 100 Hz. The original photoelectric signal is filtered by a bandpass filter to remove baseline drift and motion artifacts, and then the frequency domain features are extracted by fast Fourier transform to calculate the ratio of low frequency power (0.04 to 0.15 Hz) to high frequency power (0.15 to 0.4 Hz) as the heart rate variability index. The skin temperature data is output by a digital temperature sensor with an accuracy of 0.5 degrees Celsius. The sensor is in close contact with the user's wrist skin and updates the reading every 10 seconds. The body motion acceleration data is collected by a three-axis micro-electro-mechanical system accelerometer at a sampling rate of 50 Hz. The acceleration components in three orthogonal directions are then squared and summed to obtain the combined acceleration, which is then time-integrated to convert it into a displacement energy index, which is used to represent the user's activity intensity.

[0056] The environmental temperature and humidity data are updated every 10 seconds by an integrated temperature and humidity sensor chip. The chip is located on the outside of the wristband and is directly exposed to the environment. All sensors are uniformly scheduled by the same master microcontroller to ensure that their sampling start times are synchronized with the global system clock, providing a basis for subsequent time alignment.

[0057] In step S3, the glucose concentration time series data is segmented and sliding window preprocessed based on circadian cycle to generate a blood glucose segment sequence with a fixed time span. This preprocessing takes 24 hours as a complete physiological cycle, and the continuously collected glucose concentration time series data is divided into two subintervals of daytime active period (06:00 to 22:00) and nighttime sleep period (22:00 to 06:00 the next day) according to the local clock. In each subinterval, a sliding window with a length of 6 hours and a step of 30 minutes is used to intercept the blood glucose segment. Each blood glucose segment contains 72 consecutive sampling points (because the sampling interval is 5 minutes, 6 hours corresponds to 72 points), and the circadian phase label corresponding to the starting time of the segment is attached. The circadian phase label takes the value of 0 or 1, representing nighttime or daytime, respectively. This label is embedded as an additional context feature into the subsequent model input, guiding the model to distinguish blood glucose fluctuation patterns under different physiological states.

[0058] The sliding window operation is performed in real time in the data buffer. Whenever a new sampling point arrives, the buffer is updated and checked to see if the window generation condition is met. If it is met, all glucose values and their timestamps within the current window are extracted, and a complete blood glucose segment instance is formed after the phase label is attached. This process effectively converts an infinite time series stream into fixed-length supervised learning samples while preserving key circadian rhythm information, avoiding the decline in model generalization ability due to the neglect of physiological periodicity.

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

[0060] This normalization strategy takes into account individual differences and long-term trend drift, improving the model's cross-user adaptability. Second, time-domain alignment is performed: for physiological signals with a higher sampling frequency than the glucose sensor (such as heart rate variability and body motion acceleration), linear interpolation is used to downsample them to the same 5-minute interval as the glucose data; for signals with a lower sampling frequency than the glucose sensor (such as environmental temperature and humidity, 1 sample every 10 seconds, which is actually higher than the glucose sampling rate, here the last valid value is kept to fill the upsampled value.

[0061] Specifically, taking the time stamp sequence of the blood glucose segment as the reference time axis, each type of physiological signal is interpolated or filled at its corresponding time point to ensure that all signals are defined at 72 time points. Finally, the normalized and aligned physiological signals of each type are stacked by time point to form a multi-modal physiological feature matrix with dimensions of 72 rows and five rows, where the five rows correspond to glucose concentration, heart rate variability, skin temperature, body motion acceleration energy index, and environmental temperature and humidity. This matrix is used as the model input, fully preserving the synchronization of multi-source signals in the time dimension and the independent statistical properties between channels.

[0062] In step S5, the multi-modal 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 depth separable convolution and gated recurrent unit and embeds a dynamic sparse activation module for outputting a blood glucose concentration prediction value sequence within a preset time window in the future.

[0063] The teacher model is a deep time series convolution network with more than 2 million parameters, and its input is a complete multi-modal physiological feature matrix, and its output is a blood glucose concentration prediction curve within 3 hours in the future. The parameter quantity of the lightweight student model does not exceed 150,000, and its network structure is composed of 3 depth separable convolution blocks connected in series with 2 gated recurrent unit layers.

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

[0065] The point-wise convolution layer fuses channel information through 1x1 convolution. The gated recurrent unit layer is used to capture long-term temporal dependencies, and its update gate and reset gate mechanisms effectively alleviate the gradient vanishing problem. After each gated recurrent unit layer, a dynamic sparse activation module is set.

[0066] This module calculates the L1 norm of the input features of the current layer. If the L1 norm of a certain input channel of a neuron is lower than a preset sparsity threshold, its output is forced to zero in this forward propagation, thereby turning off the activation of the neuron. The sparsity threshold is not fixed but dynamically adjusted according to the current available memory capacity of the device:

[0067] When the memory is tight, the threshold is increased to increase the sparsity and reduce the number of activated neurons; when the memory is abundant, the threshold is reduced to retain more feature information. This mechanism takes effect in real time during the inference stage without the need to retrain the model. The training process of the lightweight student model is completed in the cloud: first, train the teacher model until it converges; then, fix the parameters of the teacher model, input the same data set into the teacher and student models simultaneously, and calculate the mean square error loss between the output prediction curves of the two models:

[0068]

[0069] wherein is the prediction time step, and are the prediction values of the teacher and student models at the step, respectively. Meanwhile, 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] wherein is the probability distribution of the features after softmax normalization. The joint loss function is:

[0072]

[0073] wherein and are the weight coefficients, and the student model parameters are updated by the back propagation algorithm until the root mean square error of the prediction on the validation set is less than 0.8 mmol / L.

[0074] In step S6, in the running 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 special neural network acceleration coprocessor integrated in the main control chip of a smart watch, with an on-chip static random access memory capacity of 512 kilobytes, a main frequency of 400 megahertz, and support for eight-bit integer fixed-point operations.

[0075] Both the model weights and the activation values are stored and calculated in eight-bit integer format, 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 it is detected that the memory usage exceeds 80% of the total available memory (i.e., 410 kilobytes), the trigger interval of the model inference is extended from the default 30 minutes to 45 minutes; when the memory usage falls below 60% (i.e., 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 even in the case of multitasking or background application memory occupation, the blood glucose prediction task can still obtain sufficient memory resources, avoiding process crashes or device restarts due to insufficient memory. The inference task is scheduled as a low-priority background service, which is only executed when the system is idle and the memory is sufficient, minimizing the impact on user interaction experience.

[0077] After outputting the blood glucose concentration prediction value sequence in the future preset time window, a risk warning generation step is further included. The prediction value sequence is compared with a preset low blood glucose threshold 3.9 mmol / L and a high blood glucose threshold 10.0 mmol / L point by point. If there are more than three consecutive data points below the low blood glucose threshold in the prediction sequence, a first-level low blood glucose warning signal is generated; if there are more than five consecutive data points above the high blood glucose threshold, a second-level high blood glucose warning signal is generated.

[0078] The warning signal is prompted to the user in a dual mode of vibration motor and screen icon, and is synchronized to the associated mobile terminal application program through the Bluetooth low power consumption protocol. The warning logic is executed locally at the edge, without the need for networking, ensuring that timely warnings can still be given in areas without network coverage.

[0079] The vibration mode is distinguished according to the warning level: a short single vibration for first-level warning, and three consecutive long vibrations for second-level warning. The screen icon is displayed in the status bar at all times, and the color changes with the blood glucose state: green for normal, yellow for warning, and red for emergency.

[0080] Corresponding to the above method, the present application also provides a small-sized blood glucose intelligent monitoring and analysis system integrated with edge computing. The system includes a continuous blood glucose monitoring sensing 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 sensing unit is used to collect the subcutaneous interstitial fluid glucose concentration time series data of the user in real time, and its structure is consistent with that of the sensor described in the aforementioned S1 step.

[0081] The multi-source physiological signal acquisition unit is integrated in a wearable wristband, including a photoplethysmogram sensor, a digital temperature sensor, a three-axis micro-electro-mechanical system accelerometer, and a temperature and humidity sensing chip, for synchronously acquiring heart rate variability, skin temperature, body motion acceleration, and environmental temperature and humidity data. The data preprocessing unit is implemented by a microcontroller firmware in the edge computing device, performing sliding window segmentation based on circadian rhythm, channel-level normalization, and time domain alignment operations, and outputting a multi-modal physiological feature matrix.

[0082] The lightweight blood glucose prediction model execution unit is deployed in a dedicated neural network acceleration coprocessor, which internally stores a lightweight personal model obtained through knowledge distillation training. The model parameters are stored in read-only memory in eight-bit integer format, and all intermediate activation values are calculated and temporarily stored in a 512 kilobyte on-chip static random access memory. The unit is built-in with a dynamic sparse activation module, which can shut down redundant connections in real time according to the activity of input features, and integrates an adaptive inference frequency adjustment module, which is deeply coupled with the operating system kernel, monitors memory occupation in real time, and dynamically adjusts the inference task scheduling period to ensure that the memory peak occupation is always below 400 kilobytes.

[0083] The risk warning generating unit is realized by an interrupt service program in the microcontroller, and performs threshold comparison and continuity judgment immediately after receiving the prediction result, triggering corresponding vibration, display and wireless communication operations. The whole system can continuously operate 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 integrated with edge computing, characterized in that, The method comprises the following steps: Real-time acquisition of subcutaneous interstitial fluid glucose concentration time series data of a user through a wearable continuous glucose monitoring sensor; Obtaining multi-source physiological signals synchronized with the glucose concentration time series data; Segmented sliding window preprocessing of the glucose concentration time series data based on circadian cycle to generate a blood glucose segment sequence with a fixed time span; Channel-level normalization and time domain alignment of the multi-source physiological signals to form a multi-modal physiological feature matrix strictly synchronized with the blood glucose segment sequence; Inputting the multi-modal physiological feature matrix into a lightweight student model guided by a teacher model, the lightweight student model adopting a hybrid structure of depth separable convolution and gated recurrent unit and embedding a dynamic sparse activation module for outputting a blood glucose concentration prediction value sequence within a preset time window in the future; In the running stage of the edge computing device, only the lightweight student model is loaded and executed, and a self-adaptive inference frequency adjustment mechanism based on runtime memory pressure feedback is enabled to maintain the system memory occupancy below a preset safety threshold. 2.The miniaturized blood glucose intelligent monitoring and analyzing method with integrated edge computing according to claim 1, wherein, The multi-source physiological signals include heart rate variability data, skin temperature data, body motion acceleration data, and environmental temperature and humidity data. 3.The miniaturized blood glucose intelligent monitoring and analyzing method with integrated edge computing of claim 2, wherein, The segmented sliding window preprocessing of the glucose concentration time series data based on the circadian cycle to generate a blood glucose segment sequence with a fixed time span comprises: Dividing the continuously collected glucose concentration time series data into two subintervals of daytime active period and nighttime sleep period according to the local clock with 24 hours as a complete physiological cycle; In each subinterval, a blood glucose segment is intercepted with a sliding window of 6 hours in length and 30 minutes in step; Each blood glucose segment contains 72 consecutive sampling points and is attached with a circadian phase label corresponding to the starting time of the segment, the circadian phase label taking a value of 0 or 1, representing nighttime or daytime, respectively. 4.The miniaturized blood glucose intelligent monitoring and analyzing method with integrated edge computing of claim 3, wherein, The channel-level normalization and time domain alignment of the multi-source physiological signals to form a multi-modal physiological feature matrix strictly synchronized with the blood glucose segment sequence comprises: Independently calculating the mean and standard deviation of each type of physiological signal in the last seven days of historical data to standardize the current collected value; 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, the last valid value is retained for upsampling to ensure that all physiological signals are strictly aligned with the blood glucose segment in the time dimension.

5. The method of claim 4, wherein the edge computing-integrated miniaturized blood glucose intelligent monitoring and analysis method is characterized by, The lightweight student model guided by the teacher model adopts a hybrid structure of depth separable convolution and gated recurrent unit and embeds a dynamic sparse activation module for outputting a blood glucose concentration prediction value sequence within a preset time window in the future, comprising: The network structure of the lightweight student model is composed of three depth separable convolution blocks connected in series with two layers of gated recurrent units; Each depth separable convolution block contains a one-dimensional convolution layer with a channel-wise kernel size of 3, a point-wise convolution layer, a batch normalization layer, and a rectified linear unit activation function. A dynamic sparse activation module is arranged after each gate recurrent unit layer, which dynamically turns off part of neuron connections according to the L1 norm of the current input features, and the proportion of turning off is controlled by a preset sparsity threshold, which is dynamically adjusted according to the current available memory capacity of the device.

6. The integrated edge computing miniaturized blood glucose intelligent monitoring and analysis method according to claim 5, characterized in that, The lightweight student model is a student model trained by knowledge distillation of a teacher model, and the training process includes: A deep time series convolutional network with more than 2 million parameters is trained on a cloud server using a large-scale historical blood glucose and physiological data set as a teacher model until convergence; The parameters of the teacher model are fixed, the same data set is input into the teacher model and the student model, and the mean square error loss between the output prediction curves of the two is calculated, as well as the KL divergence loss between the intermediate layer features of the student model and the corresponding layer features of the teacher model; The above two loss functions are jointly optimized, and the student model parameters are updated by the back propagation algorithm until the prediction root mean square error of the student model on the validation set is less than 0.8 mmol / L.

7. The integrated edge computing miniaturized blood glucose intelligent monitoring and analysis method according to claim 6, characterized in that, In the running phase of the edge computing device, only the lightweight student model is loaded and executed, and a self-adaptive inference frequency adjustment mechanism based on runtime memory pressure feedback is enabled to maintain the system memory occupancy below a preset safety threshold, including: A memory monitoring agent is embedded 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, the model inference trigger interval is extended from the default 30 minutes to 45 minutes; When the memory usage falls below 60%, the default inference interval is restored; The adjustment of the inference interval is realized by modifying the timer parameters in the task scheduler. 8.The miniaturized blood glucose intelligent monitoring and analyzing method with integrated edge computing of claim 7, wherein, The wearable continuous glucose monitoring sensor is a miniature implantable sensing probe based on glucose oxidase electrochemical principle, with a sampling frequency of every 5 minutes, an output resolution of 0.1 mmol / L, and power supplied by a miniature flexible thin film battery. Data is transmitted to the edge computing device through the near field communication protocol. 9.The miniaturized blood glucose intelligent monitoring and analyzing method with integrated edge computing of claim 8, wherein, The acquisition of the multi-source physiological signals is completed by a multi-sensor array integrated in the same wearable wristband. The heart rate variability data is obtained by collecting the original signal at a sampling rate of 100 Hz from the photoplethysmogram sensor and then extracting the low-frequency and high-frequency power ratio through fast Fourier transform. The skin temperature data is output by the digital temperature sensor with an accuracy of 0.5°C. The body motion acceleration data is collected by a three-axis micro-electro-mechanical system accelerometer at a sampling rate of 50 Hz and converted into displacement energy index through integration operation. The environmental temperature and humidity data are updated by the integrated temperature and humidity sensor chip at a frequency of once every 10 seconds.

10. The miniaturized blood glucose intelligent monitoring and analysis method with integrated edge computing according to claim 9, wherein, After outputting the blood glucose concentration prediction value sequence within the future preset time window, a risk warning generation step is further included: The prediction value sequence is compared point by point with the preset low blood glucose threshold of 3.9 mmol / L and the high blood glucose threshold of 10.0 mmol / L; If there are more than 3 consecutive data points below the low blood glucose threshold in the prediction sequence, a first-level hypoglycemia warning signal is generated; If there are more than 5 consecutive data points above the high blood glucose threshold, a second-level hyperglycemia warning signal is generated; The early warning signal prompts the user in a dual mode of vibrating motor and screen icon, and is synchronized to the associated mobile terminal application program through the Bluetooth low power consumption protocol.

Images