Noninvasive blood glucose detection method and system based on LSTM-XGBoost hybrid model

Abstract

The invention discloses a non-invasive blood glucose detection method based on an LSTM-XGBoost fusion model. A dataset of relationships between PPG signals and blood glucose levels is obtained. Designing a blood glucose detection model, wherein the blood glucose detection model sequentially comprises a data preprocessing module, a feature extraction module and an XGBoost regression device; the feature extraction module is realized by adopting an LSTM (Long Short Term Memory) network, and the XGBoost regression device is realized by adopting an XGBoost algorithm. And training the blood glucose detection model by adopting the data set of the relationship between the PPG signal and the blood glucose level. And using the trained blood glucose detection model to predict the blood glucose level based on the PPG signal collected by the wearable device. The accuracy and timeliness of noninvasive blood glucose detection are effectively improved.

Description

Technical Field

[0001] This application relates to a method for optically non-invasive blood glucose level detection. Background Art

[0002] The optically non-invasive blood glucose detection method is a method for detecting (predicting) blood glucose levels without harming the human body. The PPG (photoplethysmography) signal of the user's blood is collected by a sensor on a wearable device (such as a smart watch), and the glucose level in the user's blood is quantitatively evaluated.

[0003] Currently, the optically non-invasive blood glucose detection methods based on PPG signals are mainly divided into the following two types.

[0004] The first type is the method that combines artificial feature engineering and regression models. This method uses traditional feature extraction techniques, such as Mel frequency cepstral coefficients, wavelet analysis, and spectrum analysis, etc., to simplify data through different forms of transformation, which may lead to the loss of important information. In addition, many traditional feature extraction methods are sensitive to noise and are prone to feature distortion in the case of large noise interference. This method usually also requires manual selection of parameters (such as window size, step size, etc.), and improper parameter settings may affect the effect of feature extraction.

[0005] The second type is the method that uses a backpropagation neural network. This method usually uses CNN (convolutional neural network), LSTM (long short-term memory), or a combination of the two to extract features of the original signal. This method gradually extracts complex feature representations through multiple hidden layers, avoiding the loss of important information. Moreover, the neural network structure can automatically adjust its internal state, reducing the influence of human factors on features and having better generalization ability. LSTM can learn the time dependence in data, so it is more robust when dealing with noise. However, this method uses a fully connected layer as a regressor, and there are a large number of floating-point operations in the blood glucose level prediction process, which limits the timeliness of prediction.

[0006] From the above analysis, it can be seen that the existing optically non-invasive blood glucose level detection methods still have deficiencies in accuracy and timeliness. Summary of the Invention

[0007] The technical problem to be solved by this application is: how to improve the accuracy and timeliness of the optically non-invasive blood glucose detection method and provide faster and more accurate blood glucose detection results.

[0008] To solve the above technical problems, the present application proposes a non-invasive blood glucose detection method based on an LSTM-XGBoost fusion model, which includes the following steps. Step S1: Obtain a dataset of the relationship between PPG signals and blood glucose levels. Step S2: Design a blood glucose detection model, which successively includes a data preprocessing module, a feature extraction module, and an XGBoost regressor; the feature extraction module is implemented using an LSTM network, and the XGBoost regressor is implemented using the XGBoost algorithm. Therefore, the blood glucose detection model is an LSTM-XGBoost fusion model. The order of step S1 and step S2 can be either arbitrary and prior, or simultaneous. Step S3: Use the dataset of the relationship between the PPG signal and the blood glucose level to train the blood glucose detection model; the training of the blood glucose detection model is divided into two stages, namely the training of the LSTM network and the training of the XGBoost model. Step S4: Use the trained blood glucose detection model to predict the blood glucose level based on the PPG signal collected by the wearable device.

[0009] Further, in step S1, the dataset has multiple samples, each sample includes a PPG signal and a blood glucose value, and the blood glucose values in all samples are converted to the unit of mmol / L.

[0010] Further, in step S2, the data preprocessing module is used to preprocess the PPG signal collected by the wearable device, including standard score processing, downsampling, and removing motion artifacts. The formula for standard score processing is: where x is the original value, μ is the mean value, σ is the standard deviation, and z is the value after standardization. Downsampling is to downsample the collected PPG signal to be equal to the sampling frequency of the heart rate sensor in the wearable device, and perform band-pass filtering on the downsampled data from 0.5 Hz to 15 Hz. The PPG signal of the wearable device set is the sum of the clean PPG signal and the motion artifact. The motion artifact is approximately calculated through the acceleration data collected by the triaxial accelerometer on the wearable device, and the RLS filter is used to remove the motion artifact.

[0011] Further, in step S2, the PPG signal includes three channels of red light, green light, and blue light, and only the red light channel of the PPG signal is used for blood glucose detection.

[0012] Further, in step S2, the feature extraction module includes two LSTM layers; each LSTM layer is composed of a series of LSTM cells, and the number of LSTM cells included in each LSTM layer is equal to the dimension of the output at each time step, that is, the dimension of the hidden state. The hidden state of an LSTM cell at time step t is h t; The output of an LSTM cell is a 1×T-sized vector [h1, h2, h3, ……, h T , where T represents the number of time steps. The hidden state hh t of an LSTM layer at time step t is the overall hidden state of all the LSTM cells included in this LSTM layer at time step t; the output of an LSTM layer is a T×P-sized matrix obtained by concatenating the hidden states of this LSTM layer at all time steps, where T represents the number of time steps and P represents the number of LSTM cells included in this LSTM layer, and each column represents the output of an LSTM cell in this LSTM layer. The output of the second LSTM layer is a T×P2-sized matrix, where P2 represents the number of LSTM cells included in the second LSTM layer. Flattening this matrix by rows gives the LSTM features F extracted by the two LSTM layers.

[0013] Further, in step S2, the feature extraction module further includes one or more manual feature calculation units, which are respectively used to calculate the zero crossing rate ZCR, autocorrelation ACR, Teager-Kaiser energy KTE, power spectral density PSD, and fast Fourier transform FFT features. The ZCR feature is used to represent the number of sign changes of the PPG signal. The ACR feature is used to represent the periodic components of the PPG signal. The KTE feature is used to represent the energy characteristics of the PPG signal. The PSD feature is used to characterize the energy distribution of different frequency components in the PPG signal. The FFT feature is used to characterize the kurtosis and skewness of the frequency distribution of the PPG signal. Concatenating the LSTM features F extracted by the two LSTM layers and the manual features obtained by one or more manual feature calculation units together gives the feature vector F’ output by the feature extraction model.

[0014] Further, in step S2, during the training process, there are also successively a first fully connected layer, an activation function, and a second fully connected layer after the feature extraction model and before the XGBoost regressor; in the trained blood glucose detection model, there is no structure of "first fully connected layer - activation function - second fully connected layer", and the output of the feature extraction model is directly fed into the XGBoost regressor.

[0015] Further, in step S2, the XGBoost regressor receives the feature vector F' input by the feature extraction module and outputs the predicted blood glucose value; the XGBoost regressor is a boosting tree model composed of multiple decision trees; the decision tree is a binary tree composed of multiple nodes, including a root node, internal nodes, and leaf nodes; the root node is the start of the prediction process, and it determines which child node to enter based on the input features and splitting conditions; traverse the internal nodes, layer by layer downward according to the feature values and splitting conditions until reaching the leaf node; once reaching the leaf node, the predicted value corresponding to this node can be obtained; accumulate the prediction results of the final leaf nodes of each decision tree, and finally obtain the blood glucose prediction value.

[0016] Further, in step S3, the dataset of the relationship between the PPG signal and the blood glucose level is divided into a training dataset and a test dataset. First, use the training dataset to train the LSTM model, and the LSTM model is used to extract the features of the PPG signal; set the loss function to mean squared error and minimize it through the RMSProp algorithm for root mean square propagation. Subsequently, the XGBoost model is trained using the feature vector F' output by the feature extraction module; set the loss function to mean squared error and simplify it by performing a second-order Taylor expansion of the loss function. Finally, test the trained LSTM-XGBoost hybrid model on the test dataset, and use mean squared error, coefficient of determination, and Pearson correlation coefficient to evaluate the performance of the trained LSTM-XGBoost hybrid model; adjust the parameters of the LSTM-XGBoost hybrid model according to the test results.

[0017] This application also proposes a non-invasive blood glucose detection system based on an LSTM-XGBoost fusion model, including a dataset acquisition unit, a blood glucose detection model construction unit, a blood glucose detection model training unit, and a blood glucose detection model application unit. The dataset acquisition unit is used to acquire the dataset of the relationship between the PPG signal and the blood glucose level. The blood glucose detection model construction unit is used to design a blood glucose detection model, and the blood glucose detection model sequentially includes a data preprocessing module, a feature extraction module, and an XGBoost regressor; the feature extraction module is implemented using an LSTM network, the XGBoost regressor is implemented using the XGBoost algorithm, and the blood glucose detection model is an LSTM-XGBoost fusion model. The blood glucose detection model training unit is used to train the blood glucose detection model using the dataset of the relationship between the PPG signal and the blood glucose level; the training of the blood glucose detection model is divided into two stages, namely the training of the LSTM network and the training of the XGBoost model. The blood glucose detection model application unit is used to predict the blood glucose level based on the PPG signal collected by the wearable device using the trained blood glucose detection model.

[0018] The technical effects achieved by this application are as follows: preprocess the measured PPG signal, including removing motion artifacts; use the LSTM model to extract the feature F of the preprocessed PPG signal, and combine it with the manual features of the PPG signal. The sum F' of the two fully explores various information hidden in the PPG signal; use the XGBoost algorithm to construct a regressor to predict blood glucose levels, effectively improving the accuracy and timeliness of non-invasive blood glucose detection. Description of the Drawings

[0019] Figure 1 It is a schematic flow chart of the non-invasive blood glucose detection method based on the LSTM-XGBoost fusion model proposed by this application.

[0020] Figure 2 It is a schematic structural diagram of the non-invasive blood glucose detection system based on the LSTM-XGBoost fusion model proposed by this application.

[0021] Explanation of the reference numerals in the figure: data set acquisition unit 1, blood glucose detection model construction unit 2, blood glucose detection model training unit 3, blood glucose detection model application unit 4. Detailed Implementation Manner

[0022] Please refer to Figure 1 , the non-invasive blood glucose detection method based on the LSTM-XGBoost fusion model proposed by this application includes the following steps. The non-invasive blood glucose detection method is essentially a non-invasive blood glucose prediction method based on PPG signals.

[0023] Step S1: Obtain a data set on the relationship between PPG signals and blood glucose levels. For example, collect a publicly available data set on the relationship between PPG signals and blood glucose levels, denoted as where N represents the number of samples included in the publicly available data set, x (i) represents the PPG signal of the i-th sample, and y (i) represents the blood glucose value of the i-th sample. The detection device usually displays blood glucose levels in mmol / L (millimoles per liter), so convert the blood glucose values in the publicly available data set to mmol / L to meet the detection requirements.

[0024] Step S2: Design a blood glucose detection model, which successively includes a data preprocessing module, a feature extraction module, and an XGBoost regressor. The feature extraction module is implemented using an LSTM network. LSTM is a type of recurrent neural network (RNN). The XGBoost regressor is implemented using the XGBoost (extreme gradient boosting) algorithm. XGBoost is a machine learning algorithm based on gradient-boosted decision trees. Therefore, the blood glucose detection model is an LSTM-XGBoost fusion model.

[0025] The data preprocessing module is used to preprocess the PPG signals collected by the wearable device, including one or more of z-score processing, downsampling, and removing motion artifacts. The PPG signal contains three channels: red, green, and blue. The waveform of the red channel is more stable and has less noise. Therefore, the red channel (single channel) of the PPG signal is preferably used for blood glucose detection (prediction).

[0026] Z-score processing is used to remove the influence of data units on the prediction results. The formula for z-score processing is: where x is the original value, μ is the mean, σ is the standard deviation, and z is the value after standardization.

[0027] Downsampling is to downsample the collected PPG signal to the same sampling frequency as the heart rate sensor in the wearable device, for example, a frequency of 50 Hz. Preferably, the downsampled data is subjected to a band-pass filtering process from 0.5 Hz to 15 Hz to ensure the elimination of interference from irrelevant components such as respiratory activity and cardiac activity.

[0028] Wearable devices (such as smartwatches and smart bracelets) are usually worn on the user's wrist. When the user's hand moves, there are motion artifacts in the PPG signals collected by the wearable device, which will interfere with blood glucose level detection. The collected PPG signal is the sum of a clean PPG signal and motion artifacts. The motion artifacts are approximately calculated using the acceleration data collected by the three-axis accelerometer on the wearable device, and an RLS filter (recursive least squares adaptive filter) is used to remove the motion artifacts.

[0029] The PPG signal of the red channel after z-score processing, downsampling, and removing motion artifacts output by the data preprocessing module is used as the PPG signal for the subsequent modules.

[0030] The feature extraction module includes two LSTM layers. Each LSTM layer consists of a series of LSTM cells, and the number of LSTM cells contained in each LSTM layer is equal to the dimension of the output at each time step, that is, the dimension of the hidden state. A time step is a concept in LSTM, and a time step is the position of an element in the sequence.

[0031] The hidden state of an LSTM cell at time step t is h t . The output of an LSTM cell is a 1×T-sized vector [h1, h2, h3, ……, h T obtained by concatenating the hidden states of the LSTM cell at all time steps, where T represents the number of time steps.

[0032] The hidden state hh of an LSTM layer at time step t t is the overall hidden state of all the LSTM cells contained in the LSTM layer at time step t. The output of an LSTM layer is a T×P-sized matrix obtained by concatenating the hidden states of the LSTM layer at all time steps, where T represents the number of time steps and P represents the number of LSTM cells contained in the LSTM layer, and each column represents the output of an LSTM cell in the LSTM layer.

[0033] The output of the second LSTM layer is, for example, a T×P2-sized matrix, where P2 represents the number of LSTM cells contained in the second LSTM layer. Flattening this matrix by rows gives the LSTM feature F extracted by the two LSTM layers. Flattening refers to the process of converting a multi-dimensional structure (such as a nested list, multi-dimensional array, or matrix) into a one-dimensional linear structure.

[0034] To take into account the characteristics of physiological signals, the feature extraction module adds one or more handcrafted features to the LSTM feature F extracted by the two LSTM layers, including ZCR (zero-crossing rate), ACR (autocorrelation), KTE (Kaiser-Teager energy), PSD (Power spectral density), and the results of FFT (fast Fourier transform). For this purpose, the feature extraction module also includes one or more corresponding handcrafted feature calculation units.

[0035] The ZCR feature is used to represent the number of sign changes in the PPG signal, and the number of sign changes refers to the number of times the signs (positive, negative) of adjacent elements in a sequence change. Among them, T represents the number of time steps. 1 R<0 represents the indicator function. R represents all x t x t-1 constitutes the set, 1 R<0 (x t x t-1 ) returns 1 when x t x t-1 < 0, and returns 0 otherwise. x t is the preprocessed PPG signal value at the current sampling moment, x t-1 is the preprocessed PPG signal value at the previous sampling moment.

[0036] The ACR feature is used to represent the periodic component of the PPG signal and has two forms: time domain and frequency domain. The ACR feature in the time domain Among them, τ represents the delay step, x t is the preprocessed PPG signal value at the current sampling moment, x t+τ is x t a version after a delay of τ. The value range of R ss (τ) is between -1 and 1. -1 indicates complete negative correlation, 0 indicates no correlation, and 1 indicates complete positive correlation. The ACR feature in the frequency domain R ss (ω) = |S(ω)| 2 . Among them, ω represents the angular frequency, and S(ω) represents the Fourier transform.

[0037] The KTE feature is used to represent the energy characteristics of the PPG signal, and the calculation formula is: Φ[x t = x t x t - x t+1 x t-1 . Among them, x t is the preprocessed PPG signal value at the current sampling moment, x t-1 is the preprocessed PPG signal value at the previous sampling moment, x t+1 is the preprocessed PPG signal value at the next sampling moment. The PPG signal is a sequence of length T, and Φ[x t is a sequence of length T - 2 because there is no x T+1 and x0. Calculate the average value of Φ[x t variance s 2 , kurtosis, and skewness to characterize the KTE feature. Among them, N is the length of Φ[x t , which is equal to T - 2. Φ i is an element in the sequence Φ[x t . Φ[xt Kurtosis of Φ[x t Skewness of

[0038] The power spectral density is used to characterize the energy distribution of different frequency components in the PPG signal. Preferably, the Welch method with faster calculation speed is adopted to calculate the PSD.

[0039] FFT is a classical spectrum analysis method. The FFT features take the kurtosis and skewness of the frequency distribution.

[0040] Finally, the LSTM features F extracted by the two LSTM layers and the manual features obtained by one or more manual feature calculation units are concatenated together to obtain the feature vector F' output by the feature extraction model.

[0041] Preferably, during the training process, before the XGBoost regressor after the feature extraction model, there are also successively a first fully connected layer, an activation function (ReLU), and a second fully connected layer. The first fully connected layer receives the feature vector of the previous layer as input and outputs a new feature vector, where each input feature value is connected to each output feature value. The activation function introduces non-linearity, and the formula is relu(x) = max(0, x). Among them, the max() function represents taking the maximum value between 0 and x. The second fully connected layer receives the feature vector of the previous layer as input and outputs a number. The structure of "the first fully connected layer - activation function - the second fully connected layer" is a regressor, and the disadvantage is that there are a large number of floating-point operations. It is only used to assist the training of the LSTM model and is discarded after the training ends. In the trained blood glucose detection model, there is no structure of "the first fully connected layer - activation function - the second fully connected layer", and the output of the feature extraction model is directly sent to the XGBoost regressor.

[0042] The XGBoost regressor receives the feature vector F' passed in by the feature extraction module and outputs the predicted blood glucose value. The XGBoost regressor is a boosting tree model composed of multiple decision trees. The decision tree is a binary tree composed of multiple nodes, including the root node, internal nodes, and leaf nodes. The root node is the start of the prediction process. According to the input features and splitting conditions, it judges which child node to enter; traverse the internal nodes, and layer by layer downward according to the feature values and splitting conditions until reaching the leaf node; once reaching the leaf node, the predicted value corresponding to that node will be obtained. The prediction results of the final leaf nodes of each decision tree are accumulated, and finally the blood glucose prediction value is obtained.

[0043] The order of step S1 and step S2 is not strictly restricted, either S1 is in front arbitrarily, or they are carried out simultaneously.

[0044] Step S3: Train the blood glucose detection model using the dataset of the relationship between the PPG signal and the blood glucose level. The training of the blood glucose detection model is divided into two stages, namely, the training of the LSTM network (which is a main component of the feature extraction module) and the training of the XGBoost model (i.e., the XGBoost regressor). For example, divide the dataset into a training dataset and a test dataset in a ratio of 8:2.

[0045] First, use the training dataset to train the LSTM model, which is used to extract the features of the PPG signal. For example, set the loss function to mean squared error. The loss function measures the difference between the predicted value and the true value and is minimized by the RMSProp (root mean square prop) algorithm. The RMSProp optimization algorithm uses the exponential weighted moving average of the element-wise squares of the mini-batch stochastic gradient to adjust the learning rate, reducing the oscillation and instability during the training process, enabling faster convergence, and having better generalization ability.

[0046] Subsequently, the XGBoost model is trained using the feature vector F' output by the feature extraction module. For example, set the loss function to mean squared error and simplify it by performing a second-order Taylor expansion of the loss function. The criterion for deciding whether to split each node is to judge whether the information gain after splitting the node is greater than that in the unsplit case. XGBoost uses the weighted quantile method to achieve the specific partitioning of each feature. To obtain the splitting points worth trying, a function is established to rank the "importance" of the feature values of this feature. According to the ranking results, the feature values worth trying are selected. Based on the selected features and splitting points, the samples are assigned to the left and right child nodes. When the tree reaches the maximum depth, stop building the tree. Based on the trained tree structure, calculate the predicted values tree by tree and accumulate them to obtain the final prediction result.

[0047] Finally, test the trained LSTM-XGBoost hybrid model on the test dataset and use the mean squared error, coefficient of determination (denoted as R 2 ), and Pearson correlation coefficient to evaluate the performance of the trained LSTM-XGBoost hybrid model. Adjust the parameter settings of the LSTM-XGBoost hybrid model according to the test results to ensure that the model has optimal performance.

[0048] Step S4: Use the trained blood glucose detection model to predict the blood glucose level based on the PPG signals collected by the wearable device. By combining the advantages of the LSTM model and the XGBoost algorithm, the blood glucose detection model proposed in this application fully exploits the temporal information of the PPG signals. The prediction process replaces floating-point operations with traversing decision trees, thereby improving the accuracy and speed of blood glucose prediction and helping users better manage their blood glucose levels.

[0049] Please refer to Figure 2 , the non-invasive blood glucose detection system based on the LSTM-XGBoost fusion model proposed in this application includes a dataset acquisition unit 1, a blood glucose detection model construction unit 2, a blood glucose detection model training unit 3, and a blood glucose detection model application unit 4. Figure 2 The device shown corresponds to Figure 1 the method shown.

[0050] The dataset acquisition unit 1 is used to obtain a dataset on the relationship between PPG signals and blood glucose levels.

[0051] The blood glucose detection model construction unit 2 is used to design a blood glucose detection model, which sequentially includes a data preprocessing module, a feature extraction module, and an XGBoost regressor. The feature extraction module is implemented using an LSTM network. The XGBoost regressor is implemented using the XGBoost algorithm, and the blood glucose detection model is an LSTM-XGBoost fusion model.

[0052] The blood glucose detection model training unit 3 is used to train the blood glucose detection model using the dataset on the relationship between PPG signals and blood glucose levels.

[0053] The blood glucose detection model application unit 4 is used to predict the blood glucose level based on the PPG signals collected by the wearable device using the trained blood glucose detection model.

[0054] This application combines the advantages of the LSTM model in maintaining information integrity, strong anti-noise ability, and simpler parameter settings in feature extraction tasks, and the advantage of the XGBoost algorithm in short inference time, effectively improving the accuracy and timeliness of blood glucose detection by wearable devices (real-time blood glucose detection), providing users with more convenient and reliable blood glucose management.

[0055] Compared with the prior art, the main innovations and beneficial effects of this application are described as follows.

[0056] First, the prior art has disclosed technical solutions for predicting future blood glucose levels based on existing blood glucose levels. This application, however, predicts the current blood glucose level in real time based on the measured PPG signals, featuring speed and timeliness.

[0057] Second, considering the characteristic that PPG data is easily interfered by the user's hand movement and generates motion artifacts, the present application introduces the data of the acceleration sensor to remove motion artifacts in the preprocessing module before the LSTM-XGBoost fusion model, which greatly improves the accuracy of blood glucose prediction.

[0058] Third, in the LSTM-XGBoost fusion model of the present application, in addition to extracting the features of the PPG signal by the LSTM model, some manual features are added. The newly added manual features reflect the characteristics of the PPG signal from all aspects, further improving the features F extracted by the LSTM model and enhancing the accuracy of blood glucose prediction from another aspect.

[0059] The above are only the preferred embodiments of the present application and are not used to limit the present application. For those skilled in the art, various changes and modifications can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A non-invasive blood glucose detection method based on an LSTM-XGBoost fusion model, characterized in that, It includes the following steps; Step S1: Obtain a dataset of the relationship between PPG signals and blood glucose levels; Step S2: Design a blood glucose detection model, which sequentially includes a data preprocessing module, a feature extraction module, and an XGBoost regressor; the feature extraction module is implemented using an LSTM network, and the XGBoost regressor is implemented using the XGBoost algorithm. Therefore, the blood glucose detection model is an LSTM-XGBoost fusion model; The order of Step S1 and Step S2 can be either arbitrary or simultaneous; Step S3: Train the blood glucose detection model using the dataset of the relationship between PPG signals and blood glucose levels; the training of the blood glucose detection model is divided into two stages, namely the training of the LSTM network and the training of the XGBoost model; Step S4: Use the trained blood glucose detection model to predict blood glucose levels based on the PPG signals collected by the wearable device.

2. The non-invasive blood glucose detection method based on the LSTM-XGBoost fusion model according to claim 1, characterized in that, In Step S1, there are multiple samples in the dataset, and each sample includes a PPG signal and a blood glucose value. Convert the blood glucose values in all samples to the unit of mmol / L.

3. The non-invasive blood glucose detection method based on the LSTM-XGBoost fusion model according to claim 1, characterized in that, In Step S2, the data preprocessing module is used to preprocess the PPG signals collected by the wearable device, including standard score processing, downsampling, and removing motion artifacts; The formula for standard score processing is as follows: where x is the original value, μ is the mean, σ is the standard deviation, and z is the value after standardization; Downsampling is to downsample the collected PPG signals to the same sampling frequency as the heart rate sensor in the wearable device, and perform a band-pass filter on the downsampled data from 0.5 Hz to 15 Hz; The PPG signals of the wearable device set are the sum of clean PPG signals and motion artifacts. The motion artifacts are approximately calculated through the acceleration data collected by the triaxial accelerometer on the wearable device, and the RLS filter is used to remove the motion artifacts.

4. The non-invasive blood glucose detection method based on the LSTM-XGBoost fusion model according to claim 3, characterized in that, In Step S2, the PPG signal contains three channels of red light, green light, and blue light. Only the red light channel of the PPG signal is used for blood glucose detection.

5. The non-invasive blood glucose detection method based on the LSTM-XGBoost fusion model according to claim 1, characterized in that, In Step S2, the feature extraction module includes two LSTM layers; each LSTM layer consists of a series of LSTM cells, and the number of LSTM cells included in each LSTM layer is equal to the dimension of the output at each time step, that is, the dimension of the hidden state; The hidden state of an LSTM cell at time step t is h t ; the output of an LSTM cell is a 1×T-sized vector [h1, h2, h3, ……, h T obtained by concatenating the hidden states of the LSTM cell at all time steps, where T represents the number of time steps; The hidden state $h_t$ of an LSTM layer at time step $t$ t is the overall hidden state of all the LSTM cells contained in that LSTM layer at time step $t$; the output of an LSTM layer is a matrix of size $T \times P$ obtained by concatenating the hidden states of that LSTM layer at all time steps, where $T$ represents the number of time steps and $P$ represents the number of LSTM cells contained in that LSTM layer, and each column represents the output of an LSTM cell in that LSTM layer; The output of the second LSTM layer is a matrix of size T×P2, where P2 represents the number of LSTM cells included in the second LSTM layer. Flatten this matrix by rows to obtain the LSTM feature F extracted by the two LSTM layers.

6. The non-invasive blood glucose detection method based on the LSTM-XGBoost fusion model according to claim 5, characterized in that, In Step S2, the feature extraction module also includes one or more manual feature calculation units, which are respectively used to calculate the zero crossing rate ZCR, autocorrelation ACR, Teager-Kaiser energy KTE, power spectral density PSD, and fast Fourier transform FFT features; The ZCR feature is used to represent the number of sign changes of the PPG signal; The ACR feature is used to represent the periodic components of the PPG signal; The KTE feature is used to represent the energy characteristics of the PPG signal; The PSD feature is used to characterize the energy distribution of different frequency components in the PPG signal; The FFT feature is used to characterize the kurtosis and skewness of the frequency distribution of the PPG signal; The LSTM features F extracted by two LSTM layers and the manual features obtained by one or more manual feature calculation units are concatenated together to obtain the feature vector F' output by the feature extraction model.

7. The non-invasive blood glucose detection method based on the LSTM-XGBoost fusion model according to claim 1, characterized in that, In step S2, during the training process, a first fully connected layer, an activation function, and a second fully connected layer are sequentially included after the feature extraction model and before the XGBoost regressor; in the trained blood glucose detection model, there is no structure of "first fully connected layer - activation function - second fully connected layer", and the output of the feature extraction model is directly fed into the XGBoost regressor.

8. The non-invasive blood glucose detection method based on the LSTM-XGBoost fusion model according to claim 1, characterized in that In step S2, the XGBoost regressor receives the feature vector F' passed in by the feature extraction module and outputs the predicted blood glucose value; the XGBoost regressor is a boosting tree model composed of multiple decision trees; the decision tree is a binary tree composed of multiple nodes, including a root node, internal nodes, and leaf nodes; the root node is the start of the prediction process, and it judges and selects which child node to enter according to the input features and splitting conditions; traverse the internal nodes, layer by layer downward according to the feature values and splitting conditions until reaching the leaf node; once reaching the leaf node, the predicted value corresponding to this node will be obtained; the prediction results of the final leaf nodes of each decision tree are accumulated to finally obtain the blood glucose prediction value.

9. The non-invasive blood glucose detection method based on the LSTM-XGBoost fusion model according to claim 1, wherein In step S3, the dataset of the relationship between the PPG signal and the blood glucose level is divided into a training dataset and a test dataset; First, use the training dataset to train the LSTM model, and the LSTM model is used to extract the features of the PPG signal; Set the loss function to mean squared error and minimize it through the RMSProp algorithm; Subsequently, the XGBoost model is trained using the feature vector F' output by the feature extraction module; set the loss function to mean squared error and simplify it through the second-order Taylor expansion of the loss function; Finally, test the trained LSTM-XGBoost hybrid model on the test dataset, and use mean squared error, coefficient of determination, and Pearson correlation coefficient to evaluate the performance of the trained LSTM-XGBoost hybrid model; adjust the parameters of the LSTM-XGBoost hybrid model according to the test results.

10. A non-invasive blood glucose detection system based on an LSTM-XGBoost fusion model, characterized in that, It includes a dataset acquisition unit, a blood glucose detection model construction unit, a blood glucose detection model training unit, and a blood glucose detection model application unit; The dataset acquisition unit is used to acquire the dataset of the relationship between the PPG signal and the blood glucose level; The blood glucose detection model construction unit is used to design a blood glucose detection model, and the blood glucose detection model sequentially includes a data preprocessing module, a feature extraction module, and an XGBoost regressor; the feature extraction module is implemented using an LSTM network, the XGBoost regressor is implemented using the XGBoost algorithm, and the blood glucose detection model is an LSTM-XGBoost fusion model; The blood glucose detection model training unit is used to train the blood glucose detection model with the dataset of the relationship between the PPG signal and the blood glucose level; the training of the blood glucose detection model is divided into two stages, namely the training of the LSTM network and the training of the XGBoost model; The blood glucose detection model application unit is used to predict the blood glucose level based on the PPG signal collected by the wearable device by using the trained blood glucose detection model.

Images