Noninvasive hyperglycemia detection method based on multi-modal physiological signals

Abstract

The invention discloses a non-invasive hyperglycemia detection method based on a multi-modal physiological signal, and the method comprises the steps: obtaining the multi-modal physiological signal of a person to be detected, the multi-modal physiological signal comprising a BVP signal, an HR signal and an EDA signal; extracting time domain features and frequency domain features of the BVP signal, the HR signal and the EDA signal through three-stage discrete wavelet transform, extracting DNN features of the BVP signal, the HR signal and the EDA signal through a preset CNN-LSTM network, and fusing the time domain features, the frequency domain features and the DNN features by using a deep neural network to obtain a mixed feature vector; and inputting the time domain features, the frequency domain features, the DNN features and the mixed feature vectors of the BVP signal, the HR signal and the EDA signal into a pre-trained integrated learning model, and predicting whether the to-be-detected person is in a high glucose state. The method has the advantages of noninvasiveness, high precision, real-time performance and individuation, is suitable for continuous monitoring and early screening of prediabetic people, and has good clinical prospects and practical application values.

Description

Technical Field

[0001] The present invention relates to the technical field of blood sugar detection, and in particular to a non-invasive hyperglycemia detection method based on multimodal physiological signals. Background Art

[0002] In recent years, the global prevalence of diabetes has continued to rise, becoming one of the chronic diseases that pose a serious threat to public health. According to data from the Chinese Center for Disease Control and Prevention, my country now has over 130 million people with diabetes, ranking first globally. The prevalence of prediabetes is over 35%, and is increasing annually. Prediabetes is a critical stage in the development of diabetes. Prompt identification and effective blood sugar control can significantly reduce the risk of developing type 2 diabetes and its associated cardiovascular diseases, thus possessing significant preventive value and clinical significance.

[0003] However, the common blood glucose monitoring technologies currently available on the market still cannot meet the long-term, high-frequency, and burden-free use needs of people with prediabetes. Existing technologies mainly include: 1) Invasive testing: such as fingertip blood sampling, measuring the glucose concentration in the blood through a traditional blood glucose meter. Although this method has high accuracy, its invasive operation brings obvious discomfort and pain, which limits its application in daily high-frequency monitoring, and is especially not conducive to self-management of people with prediabetes who have frequent blood glucose fluctuations. 2) Minimally invasive testing: such as subcutaneous implantable sensor systems, which can achieve continuous dynamic blood glucose monitoring, but are expensive, have limited lifespan, have biocompatibility issues, and require regular replacement and calibration. The user experience is not good and it is not yet suitable as a universal screening method. 3) Non-invasive testing: such as near-infrared spectroscopy, Raman spectroscopy, electrochemistry, radio frequency and other technical methods, most of which are in the experimental verification stage. Although user acceptance is high, there are problems such as insufficient measurement sensitivity, poor anti-interference ability, and large individual differences, which make it difficult to meet the clinical dual requirements of real-time and accuracy. At present, the vast majority of people with prediabetes basically do not use the above-mentioned invasive or minimally invasive methods to monitor their blood sugar over the long term due to equipment costs, operation complexity and experience differences, making it difficult to detect and intervene in potential high sugar risks in a timely manner.

[0004] The development of smart wearable devices has provided a viable path for low-cost, continuous, and comfortable acquisition of physiological signals. Research has shown that blood sugar fluctuations, by influencing autonomic nervous system function, trigger a series of observable physiological changes, such as: 1) blood volume pulse (BVP); 2) heart rate (HR); and 3) electrodermal activity (EDA). These signals can be collected continuously and noninvasively through devices such as smartwatches and wristbands, offering good user adaptability. Previous studies have attempted to combine PPG and ECG signals with machine learning methods to predict blood sugar status, initially validating the feasibility of this approach. For example, methods such as SVM (support vector machine) and naive Bayesian models are used to model time series features, while deep neural networks such as CNN (convolutional neural network) and LSTM (long short-term memory network) are employed to improve recognition performance. However, these methods still face the following technical bottlenecks: 1) limited information from a single signal source makes it susceptible to external interference; 2) feature engineering relies on manual design and has limited generalizability; 3) models are not optimized for individual differences, resulting in poor generalization of prediction results; and 4) they are difficult to deploy efficiently on edge devices and lack real-time responsiveness.

[0005] Therefore, there is an urgent need for a new blood glucose status detection method that integrates multimodal physiological signals, has high precision, low invasiveness, real-time response, and edge deployment, which is particularly suitable for early identification and intervention management of people at risk of prediabetes. Summary of the Invention

[0006] The present invention aims to solve at least one of the technical problems in the related art to a certain extent. To this end, the first purpose of the present invention is to propose a non-invasive hyperglycemia detection method based on multimodal physiological signals, combining the multimodal physiological signals of BVP, HR, and EDA, using three-level discrete wavelet transform to extract multi-scale features, and using CNN_LSTM deep neural network and integrated learning model to identify and predict high sugar state. The method supports deployment on smart wearable devices, has the characteristics of non-invasive, real-time, personalized and edge-available, and provides high-risk groups with a continuous, convenient and accurate blood sugar health management plan, which has significant social value and clinical prospects.

[0007] To achieve the above objectives, the first embodiment of the present invention provides a non-invasive hyperglycemia detection method based on multimodal physiological signals, the method comprising:

[0008] S1, obtaining multimodal physiological signals of the person to be tested, the multimodal physiological signals including BVP signal, HR signal and EDA signal;

[0009] S2, preprocessing BVP signal, HR signal and EDA signal;

[0010] S3, extracts the time domain features and frequency domain features of BVP signals, HR signals, and EDA signals through a three-level discrete wavelet transform, extracts the DNN features of BVP signals, HR signals, and EDA signals through a preset CNN-LSTM network, and uses a deep neural network to fuse the time domain features, frequency domain features, and DNN features to obtain a hybrid feature vector;

[0011] S4, the time domain features, frequency domain features, DNN features and mixed feature vectors of BVP signals, HR signals and EDA signals are input into a pre-trained integrated learning model to predict whether the person being tested is in a high blood sugar state. The integrated learning model is integrated by the LightGBM model, XGBoost model and GBDT model.

[0012] In addition, the non-invasive hyperglycemia detection method based on multimodal physiological signals according to the above embodiment of the present invention may also have the following additional technical features:

[0013] According to one embodiment of the present invention, step S1 includes: using a smart wearable device to synchronously collect BVP signals, HR signals and EDA signals of a prediabetic patient, the sampling period is a preset duration, and the collection process is non-invasive; and the BVP signals, HR signals, and EDA signals are time-aligned.

[0014] According to one embodiment of the present invention, step S2 includes: performing band-pass filtering on the BVP signal, the HR signal and the EDA signal, and the frequency range of the band-pass filter is set to 5 Hz to 20 Hz.

[0015] According to one embodiment of the present invention, step S3 includes:

[0016] S31, applying three-level discrete wavelet transform to the processed BVP signal, HR signal, and EDA signal respectively, to decompose the original signal into sub-signals at several scale levels;

[0017] S32, extracting statistical features of the BVP signal, HR signal, and EDA signal from the perspectives of the time domain and the frequency domain to obtain time domain features and frequency domain features;

[0018] S33, extracting the pattern features of BVP signals, HR signals, and EDA signals in local spatial regions through a convolutional neural network, capturing the distance dependency of BVP signals, HR signals, and EDA signals in time series through a long short-term memory network, and concatenating the feature vectors of BVP signals, HR signals, and EDA signals in the feature dimension to obtain a joint feature, and inputting the joint feature into a fully connected deep neural network for dimensionality reduction to obtain a DNN feature;

[0019] S34, concatenating the time domain features, frequency domain features, and DNN features of the BVP signal, HR signal, and EDA signal to obtain a mixed feature vector.

[0020] According to one embodiment of the present invention, in step S31, the original signal is decomposed into sub-signals at several scale levels using the following formula:

[0021] x(t)=cA3(t)+cD3(t)+cD2(t)+cD1(t)

[0022] Among them, cA3(t) is the third-level low-frequency approximation component, cD1(t) is the first-level high-frequency detail component, cD2(t) is the second-level high-frequency detail component, and cD3(t) is the third-level high-frequency detail component.

[0023] According to one embodiment of the present invention, in step S32, the time domain features are used to reflect the energy distribution, fluctuation degree, symmetry and shape change of the signal in the time series, and the time domain features include: maximum value, minimum value, maximum absolute value, average value, peak-to-peak value, absolute average value, root mean square value, root amplitude, standard deviation, kurtosis, skewness, margin index, waveform index, pulse index, and crest factor; the frequency domain features are calculated by performing Fourier transform on the signal, and the frequency domain features include: power spectral density mean, power spectral density standard deviation, power spectral density skewness, power spectral density kurtosis, spectrum peak, main frequency, frequency distribution mean, frequency distribution standard deviation, frequency distribution skewness, and frequency distribution kurtosis.

[0024] According to one embodiment of the present invention, step S33 includes:

[0025] For each physiological signal S∈{BVP,EDA,HR}, CNN is used to analyze the original time series. Perform convolution operation:

[0026] H (c) =ReLU(W c *S t +b c )

[0027] Among them, W c is the convolution kernel, b c is the weight, * represents the convolution operation, D c Channel output;

[0028] LSTM outputs feature H by convolution (c) As input, we get the time series feature vector

[0029] H (l) =LSTM(H (c) )

[0030] After the above CNN-LSTM operation, the features of BVP signal, HR signal and EDA signal are extracted respectively, which are recorded as:

[0031]

[0032] The feature vectors of the three signals are concatenated in the feature dimension to obtain a joint feature representation:

[0033]

[0034] The joint feature representation is fed into a fully connected deep neural network for dimensionality reduction:

[0035]

[0036] Among them, d1 is the output feature dimension.

[0037] According to one embodiment of the present invention, in step S34, the expression of the mixed feature vector is as follows:

[0038]

[0039] Among them, F hybrid is the mixed eigenvector, F stat is the statistical feature extracted by wavelet decomposition and Fourier transform, including the time domain feature F time and frequency domain features F freq , F cnn-lstm It is a 300-dimensional deep feature vector obtained by modeling the BVP signal, EDA signal, and HR signal using the CNN-LSTM network combined with the fully connected neural network DNN. d2 represents the total feature dimension after mixing.

[0040] According to one embodiment of the present invention, step S4 includes: stat ,F dnn ,F hybird}As the input of the integrated learning model, it is input into 9 independent classifiers formed by the XGBoost model, LightGBM model and GBDT model respectively, and each classifier outputs the predicted probability value of high sugar; the 9 predicted probability values ​​are fused and the comprehensive predicted probability is calculated by weighted average; if the comprehensive predicted probability is greater than or equal to the preset threshold, the person to be tested is judged to be in a high sugar state; if the comprehensive predicted probability is less than the preset threshold, the person to be tested is judged to be in a normal state.

[0041] According to one embodiment of the present invention, training an integrated learning model includes: collecting BVP signals, HR signals and EDA signals of multiple subjects, collecting glycated hemoglobin corresponding to the multiple subjects, and assigning "normal" or "high sugar" labels to the multiple subjects according to the glycated hemoglobin; preprocessing the BVP signals, HR signals and EDA signals of the multiple subjects, and dividing the signal data of the multiple subjects into a training set and a test set; extracting the time domain features and frequency domain features of the BVP signals, HR signals and EDA signals of the multiple subjects through a three-level discrete wavelet transform, and using a preset CNN to calculate the time domain features and frequency domain features of the BVP signals, HR signals and EDA signals of the multiple subjects. The -LSTM model extracts the DNN features of the BVP signals, HR signals, and EDA signals of multiple subjects, and uses a deep neural network to fuse the time domain features, frequency domain features, and DNN features of multiple subjects to obtain mixed feature vectors of multiple subjects; the time domain features, frequency domain features, DNN features, and mixed feature vectors of the BVP signals, HR signals, and EDA signals of the training set and test set are input into the ensemble learning model, and the ensemble learning model is trained using a five-fold cross-validation strategy; after the ensemble learning model is trained, it is evaluated based on the prediction accuracy of the subjects.

[0042] Compared with the prior art, the advantages of the present invention are:

[0043] 1) Continuous blood glucose monitoring can be achieved using wearable devices without blood sampling or implantation;

[0044] 2) Make full use of the complementarity of BVP, HR, and EDA signals to improve model stability;

[0045] 3) Improve prediction capabilities through collaborative modeling of deep neural networks and ensemble learning models;

[0046] 4) It helps in pre-diabetes screening, intelligent management of chronic diseases and public health services, thus reducing the incidence of chronic diseases. Therefore, it has relatively strong practical application value.

[0047] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 Flowchart of a non-invasive hyperglycemia detection method based on multimodal physiological signals according to an embodiment of the present invention;

[0049] Figure 2 2. FIG is a diagram showing the result of a BVP signal subjected to bandpass filtering according to an embodiment of the present invention;

[0050] Figure 3is a structural diagram of a three-level discrete wavelet transform according to an embodiment of the present invention;

[0051] Figure 4 is a flow chart of feature extraction according to one embodiment of the present invention;

[0052] Figure 5 The figure is a flowchart of model training and prediction according to one embodiment of the present invention. DETAILED DESCRIPTION

[0053] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0054] The following describes a non-invasive hyperglycemia detection method based on multimodal physiological signals according to an embodiment of the present invention with reference to the accompanying drawings.

[0055] like Figure 1 As shown, the non-invasive hyperglycemia detection method based on multimodal physiological signals according to an embodiment of the present invention may include the following steps:

[0056] S1, obtaining multimodal physiological signals of the person to be tested, where the multimodal physiological signals include BVP signal, HR signal and EDA signal.

[0057] According to one embodiment of the present invention, step S1 includes: using a smart wearable device to synchronously collect the BVP signal, HR signal and EDA signal of the person to be tested, the sampling period is a preset duration, and the collection process is non-invasive; and the BVP signal, HR signal, and EDA signal are time-aligned.

[0058] Specifically, the multimodal physiological signals of the person being tested for glucose levels are first collected through a smart wearable device. The wearable device used in this invention must support at least three types of physiological signal sensors: blood volume pulse (BVP), heart rate (HR), and electrodermal activity (EDA). The sampling frequency must be ≥1Hz to ensure sensitivity to short-term high glucose levels.

[0059] Furthermore, the original BVP, HR, and EDA signals were time-aligned, with a 5-minute sampling window. Each sample must contain complete data of the three types of signals for 5 consecutive minutes.

[0060] S2, preprocessing the BVP signal, HR signal and EDA signal.

[0061] According to one embodiment of the present invention, step S2 includes: performing band-pass filtering on the BVP signal, the HR signal and the EDA signal, and the frequency range of the band-pass filter is set to 5 Hz to 20 Hz.

[0062] Specifically, to effectively eliminate low-frequency drift and high-frequency noise present in the original signal, the present invention uses a bandpass filter to uniformly filter the three types of physiological signals. The filter settings are as follows: the frequency range is set to 5Hz to 20Hz; this range covers most physiological rhythm information and is the interval of the main frequencies of BVP, EDA, and HR. The filtered signal retains key physiological change trends, which helps improve the accuracy of subsequent feature extraction and model discrimination.

[0063] For example, the signal changes before and after filtering of BVP physiological signal are as follows: Figure 2 As shown. Among them:

[0064] The horizontal axis (X-axis) "Time (Seconds)" represents the acquisition time in seconds (s), reflecting the dynamic process of the signal changing over time.

[0065] The vertical axis (Y-axis) "Amplitude" represents the amplitude of the signal. The unit is a digital dimension, which corresponds to the relative value of the sampling output of the smart wearable device and reflects the instantaneous intensity of the physiological signal.

[0066] Figure 2 In the figure, the blue "Raw Signal" represents the original blood volume pulse (BVP) signal, which contains a large amount of unprocessed noise and interference information; the yellow "Filtered Signal" represents the signal processed through a bandpass filter set in the frequency range of 5Hz to 20Hz. This significantly suppresses low-frequency drift and high-frequency noise, preserving key physiological fluctuation trends. As can be seen, the filtered signal curve is smoother and the characteristic waveform is clearer, laying a good data foundation for subsequent wavelet feature extraction and deep feature learning.

[0067] S3, extracts the time domain features and frequency domain features of BVP signals, HR signals and EDA signals through three-level discrete wavelet transform, extracts the DNN features of BVP signals, HR signals and EDA signals through the preset CNN-LSTM network, and uses the deep neural network to fuse the time domain features, frequency domain features and DNN features to obtain a mixed feature vector.

[0068] According to one embodiment of the present invention, step S3 includes:

[0069] S31 , applying three-level discrete wavelet transform to the processed BVP signal, HR signal, and EDA signal respectively, to decompose the original signal into sub-signals at several scale levels.

[0070] According to one embodiment of the present invention, in step S31, the original signal is decomposed into sub-signals at several scale levels using the following formula:

[0071] x(t)=cA3(t)+cD3(t)+cD2(t)+cD1(t)

[0072] Among them, cA3(t) is the third-level low-frequency approximation component, cD1(t) is the first-level high-frequency detail component, cD2(t) is the second-level high-frequency detail component, and cD3(t) is the third-level high-frequency detail component.

[0073] Specifically, the wavelet transform process is as follows Figure 3 As shown. In the figure, "Raw Signal" refers to the original collected physiological signal input; the original signal is decomposed into multiple scales through DWT, and different frequency components are extracted in turn; in each level of decomposition, "Hd↓2" represents the high-frequency detail component of the signal after being processed by a high-pass filter (High-pass filter) and down-sampled (Downsampling), and "Ld↓2" represents the low-frequency approximate component of the signal after being processed by a low-pass filter (Low-pass filter) and down-sampled. Through step-by-step decomposition, the high-frequency coefficients (cD1, cD2, cD3) and low-frequency coefficients (cA3) of each level are obtained. The figure also illustrates the waveform details of each scale signal after restoration by interpolation (↑2Hr), which intuitively shows the physiological information contained in different frequency bands. Further, as Figure 4 As shown in the upper part, after completing the basic bandpass filtering processing, the present invention uniformly preprocesses the original multimodal signals such as BVP, HR, EDA, etc. collected. "Raw signals partitioned into consecutive five-minute intervals" means that the continuously collected raw signals are divided into 5-minute time windows to ensure the consistency and comparability of feature extraction. In each five-minute sample segment, wavelet decomposition is applied to the BVP, HR, and EDA signals respectively. Through wavelet decomposition, the original signal is split into sub-components at different frequency levels, where:

[0074] "EDA = {EDA_raw, EDA_d1, EDA_d2, EDA_d3}" indicates that the skin electrodermal activity signal (EDA) is decomposed into the original component (EDA_raw) and the first-level (EDA_d1), second-level (EDA_d2), and third-level (EDA_d3) high-frequency detail components;

[0075] "BVP = {BVP_raw, BVP_d1, BVP_d2, BVP_d3}" represents the multi-scale components of the blood volume pulse signal (BVP);

[0076] “HR={HR_raw, HR_d1, HR_d2, HR_d3}” represents the multi-scale components of the heart rate signal (HR).

[0077] After wavelet decomposition, some characteristic components can be restored through a wavelet reconstruction process, further enhancing signal stability and integrity. Wavelet multiscale decomposition and reconstruction not only extracts local signal variations but also preserves the overall structure, laying a solid foundation for subsequent time and frequency domain feature extraction.

[0078] S32 , performing statistical feature extraction on the BVP signal, HR signal, and EDA signal from both the time domain and frequency domain perspectives to obtain time domain features and frequency domain features.

[0079] Specifically, after completing the wavelet decomposition of the signal, the statistical features of the BVP, HR, and EDA signals of each sample are extracted from the time domain and frequency domain. The specific structure is as follows: Figure 3 This step aims to convert the original continuous time series signal into a feature vector that can be processed by the machine learning algorithm.

[0080] First, in terms of time-domain features extraction, for each 5-minute sample signal segment x(t), the following time-domain statistical features are calculated: maximum value max(x), minimum value min(x), maximum absolute value (peak value) max(|x|), average value Peak-to-peak value max(x)-min(x), absolute average value RMS value Square root amplitude Standard deviation Kurtosis Skewness Margin index Waveform index rms / |μ|, pulse index max(|x|) / |μ|, crest factor max(|x|) / rms.

[0081] These time-domain statistical characteristics reflect the energy distribution, fluctuation degree, symmetry and shape changes of the signal in the time series, and are an important basis for modeling the changing laws of physiological signals.

[0082] Secondly, in terms of frequency-domain features extraction, the following statistical features are calculated by Fourier transforming the signal: power spectral density mean Power spectral density standard deviation Power spectral density skewness Power spectral density kurtosis Spectral peak max(P(f)), main frequency f max , the mean of the frequency distribution Frequency distribution standard deviation Frequency distribution skewness Frequency distribution kurtosis

[0083] These frequency-domain features not only reflect the signal's concentration and energy distribution within the frequency distribution but also reveal frequency drift and variation patterns caused by abnormal autonomic nervous system function during hyperglycemia. After extracting statistical features from the time and frequency domains, all features are integrated into unified time-frequency features, providing a high-dimensional, rich input feature space for subsequent training of hyperglycemia detection models.

[0084] S33, extracts the pattern features of BVP signals, HR signals, and EDA signals in local spatial regions through a convolutional neural network, captures the distance dependence of BVP signals, HR signals, and EDA signals in time series through a long short-term memory network, and concatenates the feature vectors of BVP signals, HR signals, and EDA signals in the feature dimension to obtain joint features, and inputs the joint features into a fully connected deep neural network for dimensionality reduction processing to obtain DNN features.

[0085] According to one embodiment of the present invention, step S33 includes:

[0086] For each physiological signal S∈{BVP,EDA,HR}, CNN is used to analyze the original time series. Perform convolution operation:

[0087] H (c) =ReLU(W c *S t +b c )

[0088] Among them, W c is the convolution kernel, b c is the weight, * represents the convolution operation, D c Channel output;

[0089] LSTM outputs feature H by convolution(c) As input, we get the time series feature vector

[0090] H (l) =LSTM(H (c) )

[0091] After the above CNN-LSTM operation, the features of BVP signal, HR signal and EDA signal are extracted respectively, which are recorded as:

[0092]

[0093] The feature vectors of the three signals are concatenated in the feature dimension to obtain a joint feature representation:

[0094]

[0095] The joint feature representation is fed into a fully connected deep neural network for dimensionality reduction:

[0096]

[0097] Where d1 is the output feature dimension. In the embodiment of the present invention, d1 is set to 300, and the DNN features are input into the subsequent classifier module as the final deep feature representation. Through the above-mentioned deep feature extraction process, the high-order variation patterns of physiological signals in terms of spatial local features and temporal dynamic characteristics can be effectively captured, thereby improving the discrimination ability of the high-sugar state recognition model.

[0098] S34, concatenating the time domain features, frequency domain features, and DNN features of the BVP signal, HR signal, and EDA signal to obtain a mixed feature vector.

[0099] According to one embodiment of the present invention, in step S34, the expression of the mixed feature vector is as follows:

[0100]

[0101] Among them, F hybrid is the mixed eigenvector, F stat is the statistical feature extracted by wavelet decomposition and Fourier transform, including the time domain feature F time and frequency domain features F freq , F cnn-lstm It is a 300-dimensional deep feature vector obtained by modeling the BVP signal, EDA signal, and HR signal using the CNN-LSTM network combined with the fully connected neural network DNN. d2 represents the total feature dimension after mixing.

[0102] Specifically, after completing the time domain and frequency domain feature extraction based on wavelet transform and Fourier transform and the deep representation learning based on CNN-LSTM network, the present invention further introduces a hybrid learning structure to effectively integrate traditional statistical features with deep model extraction features to construct a high-dimensional feature vector with better representation capabilities, such as Figure 4 shown.

[0103] The following features were obtained in the previous steps:

[0104] F stat : Statistical features extracted by wavelet decomposition and Fourier transform, including time domain features F time and frequency domain features F freq ;

[0105] F dnn : The 300-dimensional deep feature vector is obtained by modeling BVP, EDA, and HR signals using the CNN-LSTM network combined with the fully connected neural network DNN.

[0106] To unify the representation, these features are concatenated to form hybrid features:

[0107]

[0108] Where d2 represents the total feature dimension after mixing. Finally, the mixed feature vector F hybrid The input is a two-layer, fully connected deep neural network (DNN) that undergoes further nonlinear feature mapping and dimensionality reduction, extracting a more compact and discriminative feature representation, providing high-quality input for subsequent classifier modules. This hybrid feature extraction process effectively integrates information from physiological signals at multiple scales, including timing, frequency, and deep dynamic changes, significantly improving the model's comprehensive expressive power and robustness in identifying high-sugar states.

[0109] S4, the time domain features, frequency domain features, DNN features and mixed feature vectors of BVP signals, HR signals and EDA signals are input into a pre-trained integrated learning model to predict whether the person being tested is in a high blood sugar state. The integrated learning model is integrated by the LightGBM model, XGBoost model and GBDT model.

[0110] According to one embodiment of the present invention, step S4 includes: stat ,F dnn ,F hybird}As the input of the integrated learning model, it is input into 9 independent classifiers formed by the XGBoost model, LightGBM model and GBDT model respectively, and each classifier outputs the predicted probability value of high sugar; the 9 predicted probability values ​​are fused and the comprehensive predicted probability is calculated by weighted average; if the comprehensive predicted probability is greater than or equal to the preset threshold, the person to be tested is judged to be in a high sugar state; if the comprehensive predicted probability is less than the preset threshold, the person to be tested is judged to be in a normal state.

[0111] Specifically, to accurately identify individuals experiencing hyperglycemia, the present invention designed a multi-model integrated training and validation framework after feature extraction. This framework integrates multi-source feature information, uses multiple classifiers for prediction, and averages the fused probabilities, improving the robustness and generalization of predictions.

[0112] The specific process is as follows Figure 5 As shown in the figure, "Input features" include time-frequency features, hybrid features, and deep neural network features, which are derived from the feature extraction modules in the previous steps.

[0113] The "Prediction algorithms" section includes three ensemble learning models: GBDT (Gradient Boosting Decision Tree), XGBoost (eXtreme Gradient Boosting), and LightGBM (Light Gradient Boosting Machine);

[0114] Each type of input feature is fed into three prediction algorithms and trained to generate nine independent sub-classifiers;

[0115] Each sub-classifier outputs a predicted probability value p in the inference phase i , "P" represents the single prediction probability output by each classifier;

[0116] The "Calculate probability" module takes a weighted average of the output probabilities of the nine sub-classifiers (the present invention adopts equal weighted average) to calculate the comprehensive prediction probability

[0117] The "Output prediction" module makes a final judgment based on the comprehensive prediction probability and the set threshold (0.5), and outputs the prediction result of high sugar state or normal state.

[0118] Specifically, the three types of features {F stat ,F dnn ,F hybird As model input, three ensemble learning models are selected: XGBoost, LightGBM, and GBDT. Each model is trained on the three types of features mentioned above, forming a total of 9 independent classifiers:

[0119] Model i,j =Classifier j (F i ),

[0120] i∈{stat,dnn,hybird},

[0121] j∈{XGBoost,LightGBM,GBDT}}

[0122] Among them, Model stat,XGBoost Represents the XGBoost training model on time-frequency features.

[0123] When a sample is input, all models will output the predicted probability p of high sugar respectively. i , a total of 9 probability values ​​are obtained:

[0124] p1,p2,...,p9∈[0,1]

[0125] The final fusion decision adopts weighted averaging (the present invention adopts equal weighting) to calculate the comprehensive prediction probability:

[0126]

[0127] According to the set judgment rules, if If the sample has high sugar content, it is judged to be in a high sugar state; otherwise, it is judged to be in a normal state.

[0128] According to one embodiment of the present invention, training an integrated learning model includes: collecting BVP signals, HR signals and EDA signals of multiple subjects, collecting glycated hemoglobin corresponding to the multiple subjects, and assigning "normal" or "high sugar" labels to the multiple subjects according to the glycated hemoglobin; preprocessing the BVP signals, HR signals and EDA signals of the multiple subjects, and dividing the signal data of the multiple subjects into a training set and a test set; extracting the time domain features and frequency domain features of the BVP signals, HR signals and EDA signals of the multiple subjects through a three-level discrete wavelet transform, and using a preset CNN to calculate the time domain features and frequency domain features of the BVP signals, HR signals and EDA signals of the multiple subjects. The -LSTM model extracts the DNN features of the BVP signals, HR signals, and EDA signals of multiple subjects, and uses a deep neural network to fuse the time domain features, frequency domain features, and DNN features of multiple subjects to obtain mixed feature vectors of multiple subjects; the time domain features, frequency domain features, DNN features, and mixed feature vectors of the BVP signals, HR signals, and EDA signals of the training set and test set are input into the ensemble learning model, and the ensemble learning model is trained using a five-fold cross-validation strategy; after the ensemble learning model is trained, it is evaluated based on the prediction accuracy of the subjects.

[0129] Specifically, the multimodal physiological signals of people with prediabetes are first collected through smart wearable devices. The wearable devices used in this invention must support at least three types of physiological signal sensors: blood volume pulse (BVP), heart rate (HR), and electrodermal activity (EDA). The sampling frequency must be ≥1Hz to ensure sensitivity to short-term high blood sugar states.

[0130] For example, a total of 16 subjects were tested, with their specific genders and glycated hemoglobin (HbA1c) levels shown in Table 1. HbA1c is an important clinical indicator for determining prediabetes. In this embodiment, users are divided into normoglycemic and prediabetic groups based on their HbA1c values, with 5.7% being the dividing line between high and low blood sugar levels for labeling.

[0131] Table 1 Information of the subjects

[0132] ID 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 gender male female male male male male male male female male female female female female male female HbA1c (%) 5.5 5.6 5.9 6.4 5.7 5.8 5.3 5.6 6.1 6.0 6.0 5.6 5.7 5.5 5.5 5.5

[0133] The original BVP, HR, and EDA signals were time-aligned. Specifically, using a 5-minute sampling window, each sample contained complete data for the three signal types for 5 consecutive minutes. Each valid sample was assigned a "normal" or "high glucose" label based on its corresponding blood glucose value compared to a 5.7% threshold.

[0134] Furthermore, after completing the acquisition of the original physiological signals, the present invention performs unified preprocessing operations on the three types of signals: BVP (blood volume pulse), HR (heart rate) and EDA (skin electrodermal activity) to ensure the quality, comparability and effectiveness of the subsequent model training of the data. First, the health status of each subject is preliminarily divided according to his or her glycosylated hemoglobin (HbA1c) value. Those with HbA1c higher than 5.7% are classified as "prediabetic people", and those with HbA1c of 5.7% and below are classified as "normal blood sugar people" for sample labeling for subsequent model training. Subsequently, 12 of the total 16 subjects were used for the model training set, and 4 were used for the test set. In order to effectively eliminate the low-frequency drift and high-frequency noise present in the original signal, the present invention uses a bandpass filter to perform unified filtering processing on the three types of physiological signals. The filter settings are as follows: the frequency range is set to 5 Hz to 20 Hz; this range covers most physiological rhythm information and is the interval where the main frequencies of BVP, EDA, and HR are located; the filtered signal retains key physiological change trends, which helps to improve the accuracy of subsequent feature extraction and model discrimination.

[0135] Furthermore, the time domain features and frequency domain features of the BVP signals, HR signals and EDA signals of multiple subjects are extracted through the three-level discrete wavelet transform, and the DNN features of the BVP signals, HR signals and EDA signals of multiple subjects are extracted through the preset CNN-LSTM model, and the time domain features, frequency domain features and DNN features of multiple subjects are fused using a deep neural network to obtain a mixed feature vector of multiple subjects. Please refer to the description of the aforementioned embodiment for details, which will not be repeated here.

[0136] The ensemble learning model was trained using a 50-fold cross-validation strategy, using data from 12 subjects as the training set and 4 users as the test set for independent evaluation. This prevented data leakage and ensured model generalization. Accuracy was used as the evaluation metric to verify the model's performance across different users and days. Table 2 shows the model's prediction accuracy for Subject 3 over the 9 days.

[0137] Table 2

[0138] Number of days 1 2 3 4 5 6 7 8 9 Accuracy 0.9714 0.8773 0.7950 0.9043 0.7630 0.8987 0.8671 0.8483 0.8846

[0139] As shown in Table 2, the prediction accuracy for subject 3 remained high over nine consecutive days, ranging from a low of 76.3% to a high of 97.1%. The overall accuracy fluctuated slightly, with a mean of approximately 0.87. This demonstrates that the proposed fusion feature and multi-model voting mechanism exhibits good stability and generalization capabilities in the task of continuous monitoring of individual users, can reliably identify high glucose levels in people with prediabetes, and has potential for application in preclinical intervention and chronic disease management.

[0140] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0141] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0142] In the present invention, unless otherwise specified or limited, the terms "installed," "connected," "connect," "fixed," etc. should be understood in a broad sense. For example, they can refer to fixed connection, detachable connection, or integration; mechanical connection, electrical connection; direct connection, or indirect connection through an intermediate medium; internal communication between two components, or interaction between two components, unless otherwise specified. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0143] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A non-invasive hyperglycemia detection method based on multimodal physiological signals, characterized in that: The method comprises: S1, obtaining a multimodal physiological signal of a person to be measured, wherein the multimodal physiological signal includes a BVP signal, a HR signal, and an EDA signal; S2, preprocessing BVP signal, HR signal and EDA signal; S3, extracts the time domain features and frequency domain features of BVP signals, HR signals, and EDA signals through a three-level discrete wavelet transform, extracts the DNN features of BVP signals, HR signals, and EDA signals through a preset CNN-LSTM network, and uses a deep neural network to fuse the time domain features, frequency domain features, and DNN features to obtain a hybrid feature vector; S4, the time domain features, frequency domain features, DNN features and mixed feature vectors of the BVP signal, HR signal and EDA signal are input into a pre-trained integrated learning model to predict whether the person being tested is in a high blood sugar state, wherein the integrated learning model is integrated by the LightGBM model, the XGBoost model and the GBDT model.

2. The non-invasive hyperglycemia detection method based on multimodal physiological signals according to claim 1, characterized in that: Step S1 includes: Smart wearable devices are used to synchronously collect BVP, HR, and EDA signals from prediabetic patients. The sampling period is a preset duration and the collection process is non-invasive. The BVP signal, HR signal, and EDA signal are time-aligned.

3. The non-invasive hyperglycemia detection method based on multimodal physiological signals according to claim 1, characterized in that: Step S2 includes: The BVP signal, HR signal and EDA signal are subjected to band-pass filtering, and the frequency range of the band-pass filter is set to 5 Hz to 20 Hz.

4. The non-invasive hyperglycemia detection method based on multimodal physiological signals according to claim 1, characterized in that: Step S3 includes: S31, applying three-level discrete wavelet transform to the processed BVP signal, HR signal, and EDA signal respectively, to decompose the original signal into sub-signals at several scale levels; S32, extracting statistical features of the BVP signal, HR signal, and EDA signal from the perspectives of the time domain and the frequency domain to obtain time domain features and frequency domain features; S33, extracting the pattern features of BVP signals, HR signals, and EDA signals in local spatial regions through a convolutional neural network, capturing the distance dependency of BVP signals, HR signals, and EDA signals in time series through a long short-term memory network, and concatenating the feature vectors of BVP signals, HR signals, and EDA signals in the feature dimension to obtain a joint feature, and inputting the joint feature into a fully connected deep neural network for dimensionality reduction to obtain a DNN feature; S34, concatenating the time domain features, frequency domain features, and DNN features of the BVP signal, HR signal, and EDA signal to obtain a mixed feature vector.

5. The non-invasive hyperglycemia detection method based on multimodal physiological signals according to claim 4, characterized in that: In step S31, the original signal is decomposed into sub-signals at several scale levels using the following formula: x(t)=cA3(t)+cD3(t)+cD2(t)+cD1(t) Among them, cA3(t) is the third-level low-frequency approximation component, cD1(t) is the first-level high-frequency detail component, cD2(t) is the second-level high-frequency detail component, and cD3(t) is the third-level high-frequency detail component.

6. The non-invasive hyperglycemia detection method based on multimodal physiological signals according to claim 4, characterized in that: In step S32, the time domain features are used to reflect the energy distribution, fluctuation degree, symmetry and shape change of the signal in the time series. The time domain features include: maximum value, minimum value, maximum absolute value, average value, peak-to-peak value, absolute average value, root mean square value, root mean square amplitude, standard deviation, kurtosis, skewness, margin index, waveform index, pulse index, and crest factor; The frequency domain features are calculated by performing Fourier transform on the signal. The frequency domain features include: power spectral density mean, power spectral density standard deviation, power spectral density skewness, power spectral density kurtosis, spectrum peak, main frequency, frequency distribution mean, frequency distribution standard deviation, frequency distribution skewness, and frequency distribution kurtosis.

7. The non-invasive hyperglycemia detection method based on multimodal physiological signals according to claim 4, characterized in that: Step S33 includes: For each physiological signal S∈{BVP,EDA,HR}, CNN is used to analyze the original time series. Perform convolution operation: H (c) =ReLU(W c *S t +b c ) Among them, W c is the convolution kernel, b c is the weight, * represents the convolution operation, Channel output; LSTM outputs feature H by convolution (c) As input, we get the time series feature vector H (l) =LSTM(H (c) ) After the above CNN-LSTM operation, the features of BVP signal, HR signal and EDA signal are extracted respectively, which are recorded as: H BVP ,H EDA , The feature vectors of the three signals are concatenated in the feature dimension to obtain a joint feature representation: The joint feature representation is fed into a fully connected deep neural network for dimensionality reduction: Among them, d1 is the output feature dimension.

8. The non-invasive hyperglycemia detection method based on multimodal physiological signals according to claim 7, characterized in that: In step S34, the expression of the mixed feature vector is as follows: Among them, F hybrid is the mixed eigenvector, F stat is the statistical feature extracted by wavelet decomposition and Fourier transform, including the time domain feature F time and frequency domain features F freq , F cnn-lstm It is a 300-dimensional deep feature vector obtained by modeling the BVP signal, EDA signal, and HR signal using the CNN-LSTM network combined with the fully connected neural network DNN. d2 represents the total feature dimension after mixing.

9. The non-invasive hyperglycemia detection method based on multimodal physiological signals according to claim 8, characterized in that: Step S4 includes: The three types of features {F stat ,F dnn ,F hybird As the input of the ensemble learning model, the data are input into 9 independent classifiers formed by the XGBoost model, the LightGBM model, and the GBDT model, and each classifier outputs a predicted probability value of high sugar; The 9 prediction probability values ​​are fused and judged and the comprehensive prediction probability is calculated by weighted average method; If the comprehensive predicted probability is greater than or equal to the preset threshold, the person being tested is determined to be in a high sugar state; If the comprehensive predicted probability is less than the preset threshold, the person to be tested is determined to be in a normal state.

10. The non-invasive hyperglycemia detection method based on multimodal physiological signals according to claim 9, characterized in that: Training the integrated learning model includes: collecting BVP signals, HR signals, and EDA signals from multiple subjects, collecting corresponding glycated hemoglobin levels from the multiple subjects, and assigning "normal" or "high glucose" labels to the multiple subjects based on the glycated hemoglobin levels; Preprocessing the BVP signals, HR signals and EDA signals of multiple subjects, and dividing the signal data of the multiple subjects into a training set and a test set; The time domain features and frequency domain features of the BVP signals, HR signals, and EDA signals of multiple subjects were extracted through a three-level discrete wavelet transform. The DNN features of the BVP signals, HR signals, and EDA signals of multiple subjects were extracted through a preset CNN-LSTM model. The time domain features, frequency domain features, and DNN features of multiple subjects were fused using a deep neural network to obtain a mixed feature vector of multiple subjects. Inputting the time domain features, frequency domain features, DNN features and mixed feature vectors of the BVP signal, HR signal and EDA signal of the training set and the test set into the ensemble learning model, and training the ensemble learning model using a five-fold cross-validation strategy; After the training of the integrated learning model is completed, the integrated learning model is evaluated by the prediction accuracy of the subjects.

Images