Multi-modal fusion noninvasive blood glucose dynamic monitoring device

Abstract

The invention discloses a multi-modal fusion noninvasive blood glucose dynamic monitoring device, and belongs to the technical field of blood glucose detection. The whole device is in a conventional watch form, can be worn on the wrist of a human body, and comprises a shell and a watchband, the interior of the shell is divided into two areas which are respectively used for accommodating the sweat biochemical detection module, the circuit board and various components. Sweat on the skin surface is efficiently and noninvasively collected through the sweat biochemical detection module, and the glucose concentration and PH in the collected sweat are detected in real time with high sensitivity through the multi-parameter electrochemical sensor. Fusion analysis is carried out on four physiological signal data of sweat glucose concentration, PPG signal, sweat pH and body temperature, and finally the corresponding blood glucose concentration and related parameter values are accurately calculated. The problems that in an existing non-invasive detection technology, a single signal source is low in reliability and prone to being interfered are solved, the accuracy and stability of non-invasive blood glucose detection are improved, and non-invasive, continuous and high-precision blood glucose detection is achieved.

Description

Technical Field

[0001] This invention belongs to the field of blood glucose detection technology, specifically relating to a multimodal fusion non-invasive dynamic blood glucose monitoring device. It is a multimodal fusion non-invasive blood glucose monitoring system that combines physiological signals such as sweat glucose concentration detection and photoplethysmography (PPG) with external environmental signals such as pH and temperature for acquisition and analysis. Background Technology

[0002] Diabetes is a chronic metabolic disease characterized by high blood sugar. The continuous abnormal fluctuations in blood sugar levels can damage the patient's organs and tissues, potentially causing irreversible organic damage such as cardiovascular disease, retinopathy, neuropathy, and kidney failure, seriously endangering personal health and affecting the social economy.

[0003] Effective blood glucose monitoring plays a crucial role in early disease detection, disease control, and treatment of complications. Current clinical treatment protocols are primarily based on a comprehensive management approach encompassing blood glucose monitoring, insulin injection, and dietary nutrition regulation. Traditional blood glucose testing methods involve collecting blood samples from the patient's fingertip. This method has significant drawbacks: each test requires pricking the skin, causing pain for the user. Long-term testing can lead to skin damage and infection risks, resulting in poor user compliance and hindering high-frequency, routine monitoring. Therefore, stable, reliable, and non-invasive blood glucose monitoring methods are particularly important for long-term personalized monitoring and prevention of diabetes.

[0004] To address the limitations of invasive testing, non-invasive blood glucose monitoring technology has become a research hotspot. Existing non-invasive testing technologies mainly include optical detection and body fluid detection. Optical detection relies on the absorption characteristics of blood glucose at specific wavelengths of light, but its accuracy is easily affected by individual differences such as skin thickness, skin color, and blood perfusion volume, resulting in significant fluctuations in accuracy. Body fluid detection methods based on sweat indirectly reflect blood glucose levels by analyzing the glucose concentration in sweat, but when used alone, their accuracy is difficult to guarantee due to factors such as sweat secretion volume, collection efficiency, and individual metabolic differences. Furthermore, human blood glucose levels are also related to factors such as blood circulation status reflected by PPG signals and environmental temperature and humidity. Detection methods using a single signal source cannot comprehensively reflect the patterns of blood glucose changes, leading to insufficient reliability of the test results. Summary of the Invention

[0005] To address the aforementioned issues, this invention presents a non-invasive blood glucose monitoring device based on multimodal signal fusion, which is a wearable non-invasive blood glucose sensing and monitoring device based on multi-signal detection.

[0006] The technical solution of the present invention is as follows:

[0007] A multimodal fusion non-invasive dynamic blood glucose monitoring device, the overall shape of which is a conventional watch, can be worn on the human wrist, including a shell 30 and a strap 15.

[0008] The outer casing 30 is divided into two regions: region 1 (26) and region 2 (27). Region 1 (26) is used to house the sweat biochemical detection module 1. Region 2 (27) is used to house the circuit board and various components, including the battery, power circuit module 6, main control circuit module 7, reset circuit module 29, PPG sensor 2, and temperature sensor 5.

[0009] The sweat biochemical detection module 1 adopts a multi-layer stacked structure, consisting of, from top to bottom: an encapsulation layer 14, an electrode layer 13, a microfluidic sweat collection layer 12, a sweat sample introduction layer 10, and a skin adhesion layer 9. The microfluidic sweat collection layer 12 is formed by etching a central liquid storage cavity 22 and five sweat sample inlets 21 and one sweat sample outlet 25 around the central liquid storage cavity 22 on the microfluidic chip. The five sweat sample inlets 21 are arranged at equal intervals to form a semi-circular arc structure, and the positions of the five sweat sample inlets 21 and one sweat sample outlet 25 together form a complete circular structure. The five sweat sample inlets 21 and one sweat sample outlet 25 are connected to the central liquid storage cavity 22 through an etched channel 23. The sweat sampling layer 10 has its lower surface tightly attached to the skin adhesive layer 9, and its upper surface tightly attached to the lower surface of the microfluidic sweat collection layer 12. Both the sweat sampling layer 10 and the skin adhesive layer 9 are provided with through holes corresponding to 5 sweat sampling ports 21 and 1 sweat sampling port 25. The through holes on the sweat sampling layer 10 corresponding to the sweat sampling ports 21 are covered with a porous filter membrane 11, which is used to filter large molecular impurities such as dander and lipids in the sweat. The electrode layer 13 is a multi-parameter electrochemical sensor 28, which includes a reference electrode 18, a counter electrode 17, a glucose working electrode 19, a pH working electrode 20, and a multi-parameter electrochemical sensor interface 8. The glucose working electrode 19 and the pH working electrode 20 are located between the reference electrode 18 and the counter electrode 17, and the reference electrode 18 and the counter electrode 17 partially enclose the glucose working electrode 19 and the pH working electrode 20. The positions of the reference electrode 18, the counter electrode 17, the glucose working electrode 19, and the pH working electrode 20 are directly opposite the central liquid storage cavity 22 of the microfluidic sweat collection layer 12, and are connected to the main control circuit module 7 through the multi-parameter electrochemical sensor interface 8. The encapsulation layer 14 is the outermost protective cover plate, which is provided with through holes corresponding to the sweat sample outlet 25.

[0010] The power supply circuit module 6 is connected to the battery, reset circuit module 29, main control circuit module 7, multi-parameter electrochemical sensor 28, PPG sensor 2, and temperature sensor 5, providing a stable operating voltage for each functional module. The main control circuit module 7 is connected to the reset circuit module 29, multi-parameter electrochemical sensor 28, PPG sensor 2, and temperature sensor 5, controlling the operating logic of each functional module. It is used to coordinate the working timing of each module, receive and process sensor data, execute multi-modal signal fusion algorithms, and transmit the processing results to external devices through the built-in wireless communication unit. The reset circuit module 29 is responsible for ensuring the normal initialization of the main control circuit module 7 and protecting system safety.

[0011] The skin adhesive layer 9 is made of medical-grade biocompatible adhesive material to ensure a comfortable and close fit to the skin.

[0012] The microfluidic sweat collection layer 12 is made of flexible polymer materials such as polydimethylsiloxane (PDMS) and cyclic olefin polymer (COP).

[0013] The electrode layer 13 is made of flexible substrate materials such as polyimide (PI) and polyethylene terephthalate (PET).

[0014] The inner wall of the channel 23 is symmetrically and evenly spaced with several triangular micropillars 24. The angle between the triangular micropillars 24 and the inner wall is 45°. Utilizing a strong capillary effect, sweat flows directionally and rapidly from each sweat inlet 21 to the central reservoir 22, and then flows directionally from the central reservoir 22 to the sweat outlet 25. The channel 23 has a reduced cross-section, meaning the radial dimension of the inlet is larger than the radial dimension of the outlet. The radial dimension of the sweat outlet 25 is larger than the radial dimension of the sweat inlet 21, and the radial dimension of the channel 23 connected to the sweat outlet 25 is larger than the radial dimension of the channel 23 connected to the sweat inlet 21, which facilitates continuous sweat flow.

[0015] The temperature sensor 5 is located at the bottom of the housing 30 and is used to measure the surface temperature of human skin, providing temperature drift compensation for the measured value of glucose concentration in sweat.

[0016] The PPG sensor 2 includes at least one light-emitting diode 3 and a photodetector 4 for detecting human photoplethysmography (PPG) signals.

[0017] The end of the watch strap 15 is provided with a magnetic component 16, which makes it easy for users to wear and fix it.

[0018] The above-mentioned multimodal fusion non-invasive dynamic blood glucose monitoring device operates as follows: The device is worn on the wrist via a strap 15. The sweat biochemical detection module 1 is firmly attached to the skin surface through a skin adhesive layer 9. Sweat secreted by the skin, driven by a strong capillary effect, is filtered through a porous filter membrane 11 and enters the sweat inlet 21. It is then drawn into the directional flow channel 23 with triangular micropillars 24 and collected in the central reservoir 22. A multi-parameter electrochemical sensor 28 performs biochemical detection on the sweat in the central reservoir 22, obtaining signals related to glucose concentration and pH value. The detected sweat flows out through the through-hole on the encapsulation layer 14 corresponding to the sweat outlet 25. Simultaneously, the PPG sensor 2 and temperature sensor 5 synchronously collect corresponding physiological signals. The main control circuit module 7 synchronously acquires, preprocesses, and fuses multimodal signals from the multi-parameter electrochemical sensor 28, PPG sensor 2, and temperature sensor 5 to calculate a more accurate and interference-resistant blood glucose value, which is then wirelessly uploaded to the smart terminal for display and recording.

[0019] The multimodal signal fusion algorithm model includes a preprocessing module, a lightweight Transformer encoder, and an output module. The lightweight Transformer encoder backbone network consists of multiple identical stacked layers. Each layer includes a grouped query attention mechanism and a deep separable feedforward network, and outputs layer features through residual connections and layer normalization. After processing through N layers of lightweight Transformer, a deeply fused multimodal temporal feature sequence is output.

[0020] The beneficial effects of this invention are:

[0021] This invention provides a non-invasive, real-time device for multimodal blood glucose detection, including sweat glucose concentration detection, PPG signal detection, pH concentration detection, and body temperature detection. By fusing complementary information, it overcomes the limitations of single signals, thereby improving accuracy, stability, and anti-interference capabilities. Through non-invasive blood glucose detection and embedded technology, this device not only reduces the operational burden on medical staff and improves patients' quality of life, but also provides continuous blood glucose monitoring data in daily life, facilitating better assessment of patients' blood glucose levels by medical staff. This invention contributes to the early prevention of diabetes. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall invention.

[0023] Figure 2 This is a top view of the entire invention.

[0024] Figure 3 A schematic diagram of the sweat biochemical detection module design.

[0025] Figure 4 This is a schematic diagram of a multi-parameter electrochemical sensor.

[0026] Figure 5 This is a schematic diagram of a microfluidic sweat collection layer.

[0027] Figure 6 This is a schematic diagram of the hardware system framework connection of the present invention.

[0028] Figure 7 This is a schematic diagram of the circuit framework connection of the present invention.

[0029] In the diagram: 1. Sweat biochemical detection module; 2. PPG sensor; 3. LED; 4. Photodetector; 5. Temperature sensor; 6. Power circuit module; 7. Main control circuit module; 8. Multi-parameter electrochemical sensor interface; 9. Skin adhesive layer; 10. Sweat sample introduction layer; 11. Porous filter membrane; 12. Microfluidic sweat collection layer; 13. Electrode layer; 14. Encapsulation layer; 15. Strap; 16. Magnetic component; 17. Counter electrode; 18. Reference electrode; 19. Glucose working electrode; 20. pH working electrode; 21. Sweat sample inlet; 22. Central reservoir; 23. Microfluidic channel; 24. Triangular microcolumn; 25. Sweat sample outlet; 26. Region 1; 27. Region 2; 28. Multi-parameter electrochemical sensor; 29. ​​Reset circuit module; 30. Housing. Detailed Implementation

[0030] The technical solution of the present invention will be further described below with reference to specific embodiments and accompanying drawings.

[0031] like Figure 1 , 2 As shown in Figure 7, this invention provides a multimodal fusion non-invasive dynamic blood glucose monitoring device. The device is shaped like a conventional watch, with a magnetic component 16 at the end of the strap 15 for easy wearing on the wrist. The outer casing 30 is divided into two areas: area one 26 houses the sweat biochemical detection module 11, and area two 27, the multi-source sensor detection module, houses the circuit board and various components, namely the power circuit module 6, the main control circuit module 7, the reset circuit module 29, the PPG sensor 2, and the temperature sensor 5. The two areas are connected to the multi-parameter electrochemical sensor interface 8 through the main control circuit module 7.

[0032] like Figure 6 and 7 As shown, this invention uses a rechargeable lithium battery as the main power source. The power circuit module 6 is responsible for voltage conversion and stabilization, and provides a stable voltage for all power-consuming modules in the system. It includes: the main control circuit module 7, the reset circuit module 29, the multi-parameter electrochemical sensor 28, the PPG sensor 2, and the temperature sensor 5.

[0033] The main control circuit module 7 of this invention connects to and controls the operating logic of the reset circuit module 29, the multi-parameter electrochemical sensor 28, the PPG sensor 2, and the temperature sensor 5 via circuitry. Its functions include sending commands to each sensor, starting detection at regular intervals or on demand, receiving and initially processing the raw data collected by each sensor, performing multimodal signal fusion calculations, and transmitting the processed data results to an external smart terminal or cloud server via a built-in wireless communication module.

[0034] The reset circuit module 29 described in this invention is connected to the main control circuit module 7, ensuring that the main control circuit module 7 can be initialized normally under power-on or abnormal conditions, and performing a reset operation when the system is malfunctioning, thereby ensuring system stability and safety.

[0035] The temperature sensor 5 of this invention is attached to the inner wall of the outer casing 30, close to the skin, and is used to measure the temperature signal of the human skin surface in real time. This signal is used to compensate for possible deviations in the measured value of sweat glucose concentration due to temperature changes.

[0036] The PPG sensor 2 of this invention includes at least one light-emitting diode (LED) 3 and a photodetector 4 (PD). The LED 3 emits light of a specific wavelength to illuminate superficial blood vessels in the skin, and the photodetector 4 receives the light signal after absorption and scattering by the blood in the tissue, converting it into an electrical signal, namely a photoplethysmography (PPG) signal. This signal contains information such as heart rate and blood oxygen saturation, and can be used as an indirect physiological parameter related to blood glucose. It can be fused with sweat glucose concentration for analysis to improve the accuracy of blood glucose estimation.

[0037] like Figure 3 As shown, the sweat biochemical detection module 1 of the present invention adopts a multi-layer stacked structure, which consists of the following layers from top to bottom: encapsulation layer 14, electrode layer 13, microfluidic sweat collection layer 12, sweat sample injection layer 10, and skin adhesion layer 9.

[0038] like Figure 5 As shown, the microfluidic sweat collection layer 12 is made of flexible polymer materials such as polydimethylsiloxane (PDMS) and cyclic olefin polymer (COP). Micrometer-scale fluid channels 23 are etched inside the microfluidic sweat collection layer 12, connecting five sweat inlets 21 and a central reservoir 22, ultimately leading to a sweat outlet 25. The five sweat inlets 21 are arranged in an equally spaced semi-circular arc shape. An array of triangular micropillars 24 with an included angle of 45° is designed on the sidewall of the channels 23. Utilizing a strong capillary effect, the triangular micropillars 24 can guide sweat from each sweat inlet 21 to flow directionally and rapidly into the central reservoir 22, preventing sweat backflow. The central reservoir 22 is located directly below the detection area of ​​the multi-parameter electrochemical sensor 28.

[0039] The radial dimension of the sweat outlet 25 is designed to be slightly larger than that of the sweat inlet 21, which is conducive to sweat evaporation and further promotes the directional flow of sweat into the sweat outlet 25. The radial dimension of the flow channel 23 is designed to be gradually narrowing from the sweat inlet 21 to the sweat outlet 25, which is conducive to the continuous outflow of sweat and the expulsion of air, and maintains sweat renewal.

[0040] The skin adhesive layer 9 uses medical-grade biocompatible adhesive material to ensure that the patch can be comfortably and firmly adhered to the skin of the wrist. This layer is designed with through holes corresponding to five sweat inlets 21 and one sweat outlet 25.

[0041] The sweat sample introduction layer 10 is closely attached to the skin adhesive layer 9. This layer is designed with through holes corresponding to five sweat sample inlets 21 and one sweat sample outlet 25. A porous filter membrane 11 is covered at the through hole corresponding to each sweat sample inlet 21 to filter out large molecular impurities such as dander and lipids when sweat enters, preventing microchannel blockage and electrode contamination.

[0042] The electrode layer 13 is made of a flexible substrate material, on which a multi-parameter electrochemical sensor 28 is fabricated using screen printing or sputtering processes. This sensor includes a reference electrode 18, a counter electrode 17, a glucose working electrode 19, and a pH working electrode 20, used to monitor the pH value of sweat and provide compensation parameters for glucose concentration measurement. The multi-parameter electrochemical sensor 28 is connected to the main control circuit board in region two 27 via interface 8.

[0043] The encapsulation layer 14 is the uppermost protective layer, and it has through holes corresponding to the sweat sample outlet 25 to facilitate the smooth discharge of sweat after testing.

[0044] The multi-parameter electrochemical sensor 28 described in this invention is crucial for sweat biochemical monitoring. This sensor is integrated with a microfluidic chip for sweat detection and pH concentration detection. Figure 4 As shown, the multi-parameter electrochemical sensor 28 includes a reference electrode 18, a counter electrode 17, a glucose working electrode 19, and a pH working electrode 20.

[0045] In use, the user wears the device on their wrist, with the sweat biochemical detection module 11 adhering tightly to the skin via the skin adhesive layer 9. Sweat, continuously secreted by the skin, is driven by capillary force generated by the triangular micropillars 24, filtered through the porous filter membrane 11, and then enters through five inlets 21. It is then rapidly and directionally drawn into the microfluidic channel 23 and collected in the central reservoir 22. The sweat in the central reservoir 22 comes into contact with the multi-parameter electrochemical sensor 28, undergoing an electrochemical reaction. The glucose working electrode 19 generates a current signal related to glucose concentration, while the pH working electrode 20 and temperature sensor 5 provide compensation parameters. The main control circuit module 7 simultaneously acquires the four-electrode signals, PPG signal, and temperature signal. The detected sweat is discharged from the system through the sweat outlet 25, achieving continuous sweat renewal.

[0046] The data fusion analysis described in this invention involves four physiological signals. Among them, sweat glucose concentration is the most directly relevant signal, as it correlates with blood glucose concentration, providing direct biochemical evidence. PPG signal provides indirect information about human metabolism and cardiovascular status. Blood glucose fluctuations affect blood viscosity and vascular function, and these changes are reflected in heart rate, heart rate variability, and PPG signal waveform characteristics. Sweat pH provides background information about the metabolic environment. The body's acid-base balance is related to metabolic state and serves as an auxiliary indicator for assessing the risk of abnormal blood glucose levels. Body temperature provides background information on metabolic rate and the external environment, affecting metabolic rate and sweat secretion rate, indirectly affecting the measured value of sweat glucose. Calibrating body temperature can improve the reliability of sweat glucose measurement.

[0047] The main control circuit module 7 uses a multimodal signal fusion algorithm model to fuse four physiological signal data collected: sweat glucose concentration, PPG signal, sweat pH, and body temperature, and finally calculates a relatively accurate and stable blood glucose value. The results can be wirelessly transmitted to the user's mobile APP for display, recording, and alerts.

[0048] The multimodal signal fusion algorithm model includes a preprocessing module, a lightweight Transformer encoder, and an output module.

[0049] The preprocessing module's workflow is as follows:

[0050] The first step is symbol definition and unified time base construction. The core task is to establish a globally unified time base while classifying heterogeneous input signals. The input consists of a heterogeneous sensor dataset, including high-frequency raw PPG signal data, pre-processed PPG manual feature data synchronized to the reference clock, and three types of sparsely and asynchronously acquired biochemical signal data: sweat glucose, pH value, and body temperature. The PPG manual feature data includes heart rate and heart rate variability. All subsequent processing revolves around this time base to ensure the consistency of multimodal information across the time dimension.

[0051] For high-frequency raw PPG signals, a raw waveform time window is extracted from the vicinity of each master clock time. This window data is first processed by a depthwise separable convolution module. Depthwise convolution and pointwise convolution operations are performed sequentially. First, a lightweight depthwise convolution is performed along the channel dimension to capture local waveform patterns. Then, pointwise convolution enables information exchange between channels, significantly reducing computational parameters. The feature map output from the convolution then enters a static query attention aggregation module. An important score is assigned to all time-point features of the convolution output using a learnable static query vector. Finally, a fixed-dimensional, highly compact raw signal feature vector is obtained by weighted summation of the features across the entire time window.

[0052] For the synchronized low-frequency PPG handcrafted feature signal, a nonlinear transformation is performed through a fully connected layer containing the GeLU activation function, projecting it into a semantic embedding space compatible with other modal features of the system.

[0053] For the three sparse biochemical signals—blood glucose, pH, and body temperature—a time-series alignment method based on learnable decay was employed. For each target time point, the algorithm finds all available historical observations for that modality in the vicinity of that time. Then, using a learnable decay parameter, a weight is calculated for each observation, which depends on the time difference between the observation and the target time. Observations that are closer in time receive a larger weight. Finally, the feature estimate for the target time is the weighted average of all neighboring observations. This estimate is concatenated with time freshness (i.e., the difference between the target node and the most recent acquisition time), and a linear transformation is performed to generate a feature vector sequence, yielding the feature vector sequences for each of the three biochemical signals.

[0054] After completing the independent feature extraction and alignment for each modality, multimodal fusion is performed. The input consists of feature vectors from the original PPG signal, PPG handcrafted features, sweat glucose, pH value, and body temperature output at the same time. First, these five vectors are concatenated along the channel dimension to form a comprehensive feature vector containing semantic information from the original PPG signal, PPG handcrafted features, sweat glucose, pH value, and body temperature. Then, a linear projection layer maps this comprehensive vector to the model's unified latent space dimension. Simultaneously, to endow the model with the ability to perceive temporal order, learnable positional codes are added to each position in the sequence, ultimately transforming it into a complete sequence containing multimodal information, which serves as the input to subsequent networks.

[0055] The lightweight Transformer encoder described above is as follows:

[0056] The fused sequence is then fed into a lightweight Transformer encoder backbone. This backbone consists of multiple identical stacked layers, each containing two core sub-modules. The first sub-module is a grouped query attention mechanism, which transforms the input into query, key, and value vectors. It reduces computation through an asymmetric design where the number of query heads exceeds the number of key / value heads. After attention score calculation, Softmax normalization, and weighted summation, residual connections and layer normalization stabilize the feature distribution. Unlike standard multi-head attention, this mechanism significantly reduces the number of heads in the key and value vectors while maintaining the same number of heads in the query vector. This drastically reduces the number of parameters and computation for key-value pairs while maintaining representational capabilities close to standard attention. The second sub-module is a depthwise separable feedforward network, which replaces the traditional fully connected feedforward network with a depthwise separable convolution structure. It uses a lightweight structure of pointwise convolution → depthwise convolution → GeLU function activation → pointwise convolution to perform efficient nonlinear transformations between layers and channels. Residual connections and layer normalization then output layer features. After processing by N layers of lightweight Transformer, the output is a deeply fused multimodal temporal feature sequence, which is directly fed into the output layer. This sequence integrates global and local information from five modalities, providing high-quality feature support for blood glucose prediction.

[0057] The output layer is specifically as follows:

[0058] The feature sequence, processed by the lightweight Transformer encoder, enters the global regression output layer. First, global average pooling is performed along the time dimension, compressing the features from all time steps into a single global state vector, which summarizes the overall physiological state within the entire observation window. This global state vector is finally mapped to the final blood glucose prediction value through a simple linear regression layer.

[0059] A multi-parameter non-invasive dynamic blood glucose monitoring device was constructed to acquire four types of temporal multimodal signals in parallel: sweat glucose concentration, photoplethysmography (PPG) signal, sweat pH value, and body surface temperature. Simultaneously, clinical standard reference blood glucose values ​​were recorded as model label data. Strict control of time synchronization accuracy was required during the acquisition process to ensure precise alignment of the timestamps of the multimodal signals with the reference blood glucose values.

[0060] A total of 20-30 subjects were recruited, covering different metabolic states such as normal metabolism and diabetes, while taking into account demographic characteristics such as age and gender to ensure the representativeness and generalization ability of the dataset. Each subject was required to provide 50-100 valid paired samples (including complete multimodal signal fragments and corresponding reference blood glucose labels), with the total dataset size controlled at 1000-3000 high-quality samples. A subject-independent partitioning strategy was adopted, with all paired data from 70% of the subjects used as the model training set and all data from the remaining 30% used as the validation set, avoiding model overfitting caused by cross-subject data. Based on the training set data, the multimodal signal fusion algorithm model was subjected to supervised training. By adjusting the model structure, optimizing the feature fusion strategy, and iteratively tuning parameters, the model's prediction accuracy and stability for blood glucose values ​​were continuously improved. The real-time synchronous signals collected by the aforementioned non-invasive blood glucose monitoring device were preprocessed according to a standard format and input into the optimized model. The model then outputs accurate blood glucose prediction results, achieving efficient non-invasive blood glucose detection.

[0061] In summary, this invention achieves efficient and targeted collection and detection of sweat through an innovative microfluidic structure, and by combining multimodal signals such as sweat glucose concentration, PPG, temperature, and pH, it effectively overcomes the limitations of single-signal detection, providing a comfortable, convenient, and accurate non-invasive dynamic blood glucose monitoring solution.

Claims

1. A multimodal fusion non-invasive dynamic blood glucose monitoring device, characterized in that, The device is in the shape of a regular watch and can be worn on the wrist, including a case (30) and a strap (15). The outer shell (30) is divided into two regions: Region 1 (26) and Region 2 (27). Region 1 (26) is used to house the sweat biochemical detection module (1). Region 2 (27) is used to house the circuit board and various components, including the battery, power circuit module (6), main control circuit module (7), reset circuit module (29), PPG sensor (2) and temperature sensor (5). The sweat biochemical detection module (1) adopts a multi-layer stacked structure, which consists of the following layers from top to bottom: encapsulation layer (14), electrode layer (13), microfluidic sweat collection layer (12), sweat sample introduction layer (10), and skin adhesion layer (9). The microfluidic sweat collection layer (12) is formed by etching a central liquid storage cavity (22) on the microfluidic chip, as well as five sweat sample inlets (21) and one sweat sample outlet (25) around the central liquid storage cavity (22). The five sweat sample inlets (21) are arranged at equal intervals to form a semi-circular arc structure. The locations of the sweat sampling ports (25) together form a complete circular structure. The five sweat sampling ports (21) and one sweat sampling port (25) are connected to the central liquid storage cavity (22) through an etched channel (23). The lower surface of the sweat sampling layer (10) is in close contact with the skin adhesive layer (9), and the upper surface is in close contact with the lower surface of the microfluidic sweat collection layer (12). Both the sweat sampling layer (10) and the skin adhesive layer (9) are provided with through holes corresponding to the five sweat sampling ports (21) and one sweat sampling port (25). (10) A porous filter membrane (11) is covered at the through hole corresponding to the sweat inlet (21). The porous filter membrane (11) is used to filter large molecular impurities such as dander and lipids in the sweat. The electrode layer (13) is a multi-parameter electrochemical sensor (28). The multi-parameter electrochemical sensor (28) includes a reference electrode (18), a counter electrode (17), a glucose working electrode (19), a pH working electrode (20), and a multi-parameter electrochemical sensor interface (8). The glucose working electrode (19) and the pH working electrode (20) are located at the reference electrode (18). Between the reference electrode (18) and the counter electrode (17), the reference electrode (18) and the counter electrode (17) partially enclose the glucose working electrode (19) and the pH working electrode (20). The positions of the reference electrode (18), the counter electrode (17), the glucose working electrode (19) and the pH working electrode (20) are directly opposite the central liquid storage cavity (22) of the microfluidic sweat collection layer (12), and are connected to the main control circuit module (7) through the multi-parameter electrochemical sensor interface (8). The encapsulation layer (14) is the outermost protective cover plate, and is provided with through holes corresponding to the sweat sample outlet (25). The power supply circuit module (6) is connected to the battery, reset circuit module (29), main control circuit module (7), multi-parameter electrochemical sensor (28), PPG sensor (2) and temperature sensor (5) to provide stable operating voltage for each functional module; the main control circuit module (7) is connected to the reset circuit module (29), multi-parameter electrochemical sensor (28), PPG sensor (2) and temperature sensor (5) to control the operating logic of each functional module, to coordinate the working sequence of each module, to receive and process sensor data, to execute multi-modal signal fusion algorithm, and to transmit the processing results to external devices through the built-in wireless communication unit; the reset circuit module (29) is responsible for ensuring the normal initialization of the main control circuit module (7) and protecting system safety.

2. The multimodal fusion non-invasive dynamic blood glucose monitoring device according to claim 1, characterized in that, The skin adhesive layer (9) is made of medical-grade biocompatible adhesive material; the microfluidic sweat collection layer (12) is made of polydimethylsiloxane or cyclic olefin polymer; the electrode layer (13) is made of polyimide or polyethylene terephthalate.

3. A multimodal fusion non-invasive dynamic blood glucose monitoring device according to claim 1 or 2, characterized in that, The inner wall of the channel (23) is symmetrically provided with several triangular micropillars (24) at equal intervals. The angle between the triangular micropillars (24) and the inner wall is 45°. Utilizing the strong capillary effect, sweat flows directionally and rapidly from each sweat inlet (21) to the central reservoir (22), and then flows directionally from the central reservoir (22) to the sweat outlet (25). The cross-section of the channel (23) is reduced, that is, the radial dimension of the inlet is greater than the radial dimension of the outlet. The radial dimension of the sweat outlet (25) is greater than the radial dimension of the sweat inlet (21), and the radial dimension of the channel (23) connected to the sweat outlet (25) is greater than the radial dimension of the channel (23) connected to the sweat inlet (21), which is conducive to the continuous flow of sweat.

4. A multimodal fusion non-invasive dynamic blood glucose monitoring device according to claim 1 or 2, characterized in that, The temperature sensor (5) is located at the bottom of the housing (30) and is used to measure the surface temperature of human skin.

5. A multimodal fusion non-invasive dynamic blood glucose monitoring device according to claim 1 or 2, characterized in that, The PPG sensor (2) includes at least one light-emitting diode (3) and a photodetector (4) for detecting human photoplethysmography (PPG) signals.

6. A multimodal fusion non-invasive dynamic blood glucose monitoring device according to claim 1 or 2, characterized in that, The end of the watch strap (15) is provided with a magnetic component (16) for easy wearing and fixation by the user.

7. A multimodal fusion non-invasive dynamic blood glucose monitoring device according to claim 1 or 2, characterized in that, The main control circuit module (7) uses a multimodal signal fusion algorithm to fuse four physiological signal data collected: sweat glucose concentration, PPG signal, sweat pH, and body temperature, and finally calculates a relatively accurate and stable blood glucose value. The results can be wirelessly transmitted to the user's mobile app for display, recording, and alerts.

8. The multimodal fusion non-invasive dynamic blood glucose monitoring device according to claim 7, characterized in that, The multimodal signal fusion algorithm model includes a preprocessing module, a lightweight Transformer encoder, and an output module; The preprocessing module's workflow is as follows: The input is a heterogeneous sensor dataset including raw PPG signal data, PPG manual feature data that has been preliminarily processed and synchronized to the reference clock, and three types of biochemical signal data: sweat glucose, pH value, and body temperature. Among them, the PPG manual feature data includes heart rate and heart rate variability. For the original PPG signal, at each master clock time, a raw waveform time window is extracted from its vicinity. This window data is first processed by a depthwise separable convolution module, and the feature map output by the convolution is then fed into a static query attention aggregation module. An important score is given to the features at all time points of the convolution output through a learnable static query vector. Finally, by weighted summing of the features of the entire time window, a fixed-dimensional, highly compact original signal feature vector is obtained. For synchronized low-frequency PPG handcrafted features, a nonlinear transformation is performed through a fully connected layer containing the GeLU activation function, projecting them into a semantic embedding space compatible with other modal features of the system; For blood glucose, pH, and body temperature, a time-series alignment method based on learnable decay is adopted. For each target time, all available historical observation points in the vicinity of that time and for that mode are found. Then, a learnable decay parameter is used to calculate a weight for each observation, which depends on the time difference between the observation and the target time. Finally, the feature estimate of the target time is the weighted average of all neighboring observations. This estimate is concatenated with the time freshness and linearly transformed to generate a feature vector sequence, resulting in feature vector sequences for the three biochemical signals. After completing the independent feature extraction and alignment of each modality, the multimodal fusion is performed. The input consists of feature vectors output at the same time point from the original PPG signal, PPG handcrafted features, sweat glucose, pH value, and body temperature. First, these five vectors are concatenated along the channel dimension to form a comprehensive feature vector containing semantic information of the original PPG signal, PPG handcrafted features, sweat glucose, pH value, and body temperature. Then, a linear projection layer maps this comprehensive vector to the model's unified latent space dimension. At the same time, in order to give the model the ability to perceive temporal order, a learnable positional code is added to each position in the sequence, ultimately transforming it into a complete sequence of multimodal information, which serves as the input to the subsequent temporal fusion backbone network. The lightweight Transformer encoder described above is as follows: The fused sequence is then fed into a lightweight Transformer encoder backbone network; the backbone network consists of multiple identical stacked layers, each layer containing a grouped query attention mechanism and a deep separable feedforward network, and then outputs layer features through residual connections and layer normalization; after N layers of lightweight Transformer processing, a deep fused multimodal temporal feature sequence is output. The output layer is specifically as follows: The feature sequence processed by the lightweight Transformer encoder enters the global regression output layer. First, global average pooling is performed in the time dimension to compress the features of all time steps into a single global state vector. Then, it is passed through a linear regression layer to be mapped to the final blood glucose prediction value.

Images