A multi-mode fusion blood glucose monitoring method based on ECG and PPG

Abstract

The present application provides a multi-mode fusion blood glucose monitoring method based on ECG and PPG, the method comprising: obtaining ECG signals and PPG signals of a user; fusing ECG signals and PPG signals at data level, wherein the data level fusion comprises storing ECG blocks and PPG blocks in a first data pool, and storing ECG segments and PPG segments in a second data pool; extracting time series statistical features from ECG blocks and PPG blocks, and extracting spatial morphological features from ECG segments and PPG segments; determining a time series statistical feature matrix according to time series statistical features, determining a spatial morphological feature matrix according to spatial morphological features, fusing time series statistical features and spatial morphological features to obtain a feature fusion matrix; using the time series statistical feature matrix, the spatial morphological feature matrix and the feature fusion matrix as inputs of several blood glucose prediction algorithms to obtain blood glucose prediction results. The scheme provides feasibility for improving the accuracy of non-invasive blood glucose monitoring.

Description

Technical Field

[0001] The present invention belongs to the field of medical health technology, and in particular relates to a multi-mode fusion blood glucose monitoring method based on ECG and PPG. Background Art

[0002] Diabetes is a chronic non-communicable disease with metabolic disorders characterized by long-term hyperglycemia. Active blood glucose monitoring and personalized blood glucose regulation based on blood glucose levels are one of the important means to effectively reduce diabetes and delay complications. Traditional single fingertip blood sampling monitoring will cause greater pain and infection risks to patients, and continuous monitoring cannot be achieved. The accuracy of continuous blood glucose monitoring methods based on implantable sensors gradually decreases over time, and long-term wearing may cause allergic reactions such as redness, swelling, and blistering, which limits the compliance of diabetic patients. Therefore, studying a non-invasive and continuous blood glucose monitoring method has important social value in terms of active health and chronic disease management.

[0003] Currently, real-time monitoring of characteristic changes of ECG (electrocardiogram) and PPG (photoplethysmography) is used to achieve non-invasive monitoring of blood glucose concentration. The principle of non-invasive blood glucose monitoring based on ECG and PPG is that changes in blood glucose concentration will stimulate the human sympathetic nervous system. Under the activation of such stress stimulation, the sympathetic nervous system will trigger the release of adrenaline or acetylcholine hormones from the adrenal glands, thereby causing changes in cardiac activity, resulting in corresponding changes in ECG and PPG. Blood glucose concentration can be monitored by measuring the characteristic changes of ECG and PPG in real time.

[0004] In the related art, the accuracy of the extracted ECG and PPG features is low when performing non-invasive blood glucose monitoring. Summary of the invention

[0005] The purpose of the embodiments of this specification is to provide a multi-mode fusion blood glucose monitoring method based on ECG and PPG.

[0006] To solve the above technical problems, the embodiments of the present application are implemented in the following ways:

[0007] The present application provides a multi-mode fusion blood glucose monitoring method based on ECG and PPG, the method comprising:

[0008] Obtain the user's ECG signal and PPG signal;

[0009] Performing data-level fusion on the ECG signal and the PPG signal, wherein the data-level fusion includes storing the ECG block and the PPG block in a first data pool, and storing the ECG segment and the PPG segment in a second data pool, wherein the ECG block and the ECG segment are both obtained by intercepting the ECG signal, and the PPG block and the PPG segment are both obtained by intercepting the PPG signal;

[0010] Extract time series statistical features from ECG blocks and PPG blocks, and extract spatial morphological features from ECG segments and PPG segments; determine a time series statistical feature matrix based on the time series statistical features, determine a spatial morphological feature matrix based on the spatial morphological features, and fuse the time series statistical features and the spatial morphological features to obtain a feature fusion matrix;

[0011] The time series statistical feature matrix, spatial morphological feature matrix and feature fusion matrix are used as inputs of several blood glucose prediction algorithms to obtain blood glucose prediction results.

[0012] In one of the embodiments, the signal-to-noise ratios of the ECG signal and the PPG signal are both greater than a preset threshold.

[0013] In one embodiment, extracting an ECG block from an ECG signal and extracting a PPG block from a PPG signal comprises:

[0014] Determine the length of ECG signal and PPG signal;

[0015] According to the length of the ECG signal and the PPG signal, the R point of the ECG signal and the PPG signal are extracted by using the secondary difference method. When the R point of the ECG signal and the P point of the PPG signal reach a preset number, the ECG signal and the PPG signal are stopped.

[0016] The distances between adjacent R points and the distances between adjacent P points are calculated respectively. When the distances between adjacent R points and the distances between adjacent P points meet the requirements, for the ECG signal, 0.3×HR×SP data points are taken to the left of the first R point and 0.7×HR×SP data points are taken to the right of the last R point to intercept the ECG signal, and the intercepted part is saved as an ECG block; for the PPG signal, 0.3×HR×SP data points are taken to the left of the first P point and 0.7×HR×SP data points are taken to the right of the last P point to intercept the PPG signal, and the intercepted part is saved as a PPG block; wherein HR is the average heart rate of the user, and SP is the sampling rate of the wearable device; the wearable device is used to collect ECG signals and PPG signals.

[0017] In one embodiment, extracting an ECG segment from an ECG signal and extracting a PPG segment from a PPG signal comprises:

[0018] For the ECG signal, for each R point, 0.3×HR×SP data points are taken to the left and 0.7×HR×SP data points are taken to the right to intercept the ECG signal, and the intercepted part is saved as an ECG segment;

[0019] For the PPG signal, for each point P, 0.3×HR×SP data points are taken to the left and 0.7×HR×SP data points are taken to the right to intercept the PPG signal, and the intercepted part is saved as a PPG segment.

[0020] In one of the embodiments, when the distance between adjacent R points and the distance between adjacent P points do not meet the requirements, the ECG signal and the PPG signal are reacquired.

[0021] In one embodiment, extracting time series statistical features from the ECG block and the PPG block includes:

[0022] Perform discrete wavelet transform on the data block, decompose the signal at different levels, and obtain the signal detail coefficients corresponding to different levels; the data block includes ECG block and PPG block;

[0023] According to the wavelet signal reconstruction principle, signals with different detail coefficients are reconstructed respectively to obtain a first reconstructed signal;

[0024] The reconstructed signal is fused with the data block to obtain a data block matrix; the data block matrix includes an ECG block matrix and a PPG block matrix;

[0025] The time series statistical characteristics of each element in the ECG block matrix and the PPG block matrix are calculated respectively. The time series characteristics include at least the kurtosis, skewness, signal mobility, signal complexity, fractal dimension, power spectrum entropy, Kolmogorov entropy and Shannon entropy of each element.

[0026] In one embodiment, determining a time series statistical feature matrix according to the time series statistical feature includes:

[0027] Based on the feature selection method, the time series statistical features are screened to obtain effective features; among which, the effective features meet the following requirements: the time series statistical features rank in the top 50% of all features in terms of feature importance and the time series statistical features are the intersection of three feature selection algorithms, including: univariate feature selection algorithm, recursive feature elimination algorithm, and L1-based feature selection algorithm;

[0028] Generate a time series statistical feature matrix based on effective features.

[0029] In one embodiment, extracting spatial morphological features from ECG segments and PPG segments includes:

[0030] The ECG segment is used as the input of the first residual network model, and the PPG segment is used as the input of the second residual network model to extract spatial morphological features;

[0031] Among them, the first residual network model and the second residual network model both include two weight layers, and the input is directly connected to the output layer using a jumper.

[0032] In one embodiment, determining a spatial morphological feature matrix according to the spatial morphological feature includes:

[0033] The spatial morphological features are used as the input of the deep neural network, and multiple hidden layers are used to compress and reduce the spatial morphological features, and the spatial morphological feature matrix is ​​generated in the output layer of the deep neural network.

[0034] In one embodiment, the time series statistical feature matrix, the spatial morphological feature matrix and the feature fusion matrix are used as inputs of several blood glucose prediction algorithms to obtain blood glucose prediction results, including:

[0035] Using the time series statistical feature matrix, the spatial morphological feature matrix and the feature fusion matrix as inputs of several blood glucose prediction algorithms, several different blood glucose prediction models are obtained;

[0036] Obtaining a number of historical blood sugar prediction values ​​and current blood sugar prediction values ​​of a number of different blood sugar prediction models;

[0037] Taking several historical blood sugar prediction values ​​and current blood sugar prediction values ​​as inputs of the fuzzy integral algorithm, a corresponding blood sugar prediction value is obtained;

[0038] According to the predicted blood glucose value, the prediction error of each blood glucose prediction model is calculated;

[0039] Determine the weight of each blood glucose prediction model according to the size of the prediction error;

[0040] According to the weight of each blood sugar prediction model, the output value of each blood sugar prediction model is fused to obtain the blood sugar prediction result.

[0041] It can be seen from the technical solution provided in the above embodiments of this specification that this solution: simultaneously extracts the time series statistical features and spatial morphological features of ECG signals and PPG signals, the extracted features are more comprehensive, and the present application also realizes the fusion of ECG and PPG signals at three levels: data level, feature level, and decision level, which provides feasibility for improving the accuracy of non-invasive blood glucose monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of this specification or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0043] Figure 1 This is a block diagram of the implementation of the multi-mode fusion blood glucose monitoring method based on ECG and PPG provided in this application;

[0044] Figure 2 A schematic diagram of the process of the multi-mode fusion blood glucose monitoring method based on ECG and PPG provided in this application;

[0045] Figure 3 (a) is a schematic diagram of obtaining an ECG block provided by the present application; Figure 3 (b) is a schematic diagram of obtaining a PPG block provided by the present application; Figure 3 (c) is a schematic diagram of obtaining ECG segments provided by the present application; Figure 3 (d) is a schematic diagram of obtaining the PPG fragment provided by the present application;

[0046] Figure 4 A schematic diagram of a framework for feature-level fusion of ECG signals and PPG signals provided in this application;

[0047] Figure 5 Schematic diagram of the framework for ECG and PPG decision-level fusion provided in this application. DETAILED DESCRIPTION

[0048] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.

[0049] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0050] It will be apparent to those skilled in the art that various modifications and variations may be made to the specific embodiments of the present application description without departing from the scope or spirit of the present application. Other embodiments derived from the present application description will be apparent to those skilled in the art. The present application description and examples are exemplary only.

[0051] The words “include,” “including,” “have,” “contain,” etc. used in this document are open-ended terms, meaning including but not limited to.

[0052] Unless otherwise specified, "parts" in this application are all calculated by mass.

[0053] At present, the mainstream method is to monitor the characteristic changes of ECG and PPG in real time to achieve non-invasive monitoring of blood glucose concentration. In some technologies, the collected characteristics are single, resulting in low monitoring accuracy.

[0054] In the related art, in a possible design mode, a non-invasive blood glucose detection method specifically includes the following steps: synchronous monitoring step: collecting synchronously monitored PPG signals and ECG signals; obtaining pulse wave transmission time step: obtaining pulse wave transmission time according to the rising point of the PPG signal and the peak value of the ECG signal; signal processing step: processing the PPG signal and the ECG signal, removing noise, and obtaining processed PPG signals and ECG signals; feature extraction step: performing segmentation processing and extracting signal features according to the processed PPG signal and ECG signal, and obtaining PPG signal features and ECG signal features. The present invention also relates to a non-invasive blood glucose detection system. The present invention realizes accurate measurement of blood glucose by taking pulse wave transmission time, PPG signal features, and ECG signal features as core parameters, and can eliminate the influence of external factors such as skin color, stratum corneum, and vascular wall thickness. In this design mode, the features of single-cycle ECG and PPG are mainly obtained from the time domain and frequency domain, ignoring other features, and the features obtained are single.

[0055] In another possible implementation, a non-invasive blood glucose meter includes: a non-invasive blood glucose meter body, a signal acquisition processing unit and a power management unit; the non-invasive blood glucose meter body includes: a shell and a detection unit; the detection unit is arranged on the shell, and is used to accommodate the measured organism; the power management unit is arranged in the shell, and is used to power the signal acquisition processing unit; the signal acquisition processing unit is arranged in the shell, and is used to detect and process the vital sign parameters and environmental characteristic parameters of the measured organism, and generate and output blood glucose test result data; wherein the vital sign parameters include the photoplethysmography (PPG) signal of the measured organism, the contact conduction temperature of the measured organism, and the thermal radiation temperature of the measured organism; the environmental characteristic parameters include: the temperature and humidity around the measured organism, and the temperature and humidity of the environment in which the non-invasive blood glucose meter is located. In this design, there is no mention of the fusion of ECG and PPG, and there is also a lack of feature extraction of PPG.

[0056] In another possible implementation, a non-invasive blood glucose detection method based on comprehensive analysis in the time and frequency domains includes: 1) acquiring fingertip photoelectric volumetric pulse wave signals while performing OGTT experiment collection to detect blood glucose concentration; 2) using wavelet threshold method and curve fitting method to denoise the photoelectric volumetric pulse wave signals; 3) using cluster analysis to extract representative waveforms; 4) using time domain analysis to extract characteristic parameters with physiological significance; 5) using fast Fourier transform to perform spectrum analysis on the signal to advance frequency domain characteristic parameters; 6) extracting time domain and frequency domain characteristic parameters related to blood glucose through correlation analysis; 7) using the acquired characteristic parameters and the real blood glucose value to establish a genetic algorithm to optimize the BP neural network model; 8) using Parkers error grid analysis to predict the error between the blood glucose value and the actual blood glucose value, of which 80.3% of the predicted results are in the clinically accurate area of ​​area A and 19.7% are in the clinically acceptable area of ​​area B. In this design, features are mainly extracted from the time domain and frequency domain, and other features are ignored.

[0057] Based on the defects of the above-mentioned technologies, the present application proposes a multi-modal fusion blood glucose monitoring method of ECG and PPG. The method extracts the temporal statistical characteristics and spatial morphological characteristics of ECG signals and PPG signals, and fuses the ECG signals and PPG signals at the data level, feature level, and decision level, thereby providing users with a non-invasive, fast, and low-cost blood glucose monitoring solution.

[0058] The present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0059] Reference Figure 1 , which shows a block diagram of the implementation of the multi-mode fusion blood glucose monitoring method based on ECG and PPG provided in the embodiment of the present application. Figure 1As shown, it mainly includes four modules, namely: 1) ECG and PPG data acquisition module; 2) ECG and PPG data level fusion module; 3) ECG and PPG feature level fusion module; 4) decision-level fusion module based on weight adaptation. Among them, the ECG and PPG data acquisition module is used to obtain the user's ECG and PPG signals through wearable devices; the ECG and PPG data level fusion module is used to perform data-level fusion on the acquired ECG and PPG signals; the ECG and PPG feature level fusion module is used to extract time series statistical features from ECG blocks and PPG blocks, and extract spatial morphological features from ECG segments and PPG segments, and fuse the time series statistical features with the spatial morphological features; the decision-level fusion module based on weight adaptation is used to realize the fusion of different blood sugar prediction algorithms to obtain blood sugar prediction results.

[0060] Reference Figure 2 , which shows a flow chart of a multi-mode fusion blood glucose monitoring method based on ECG and PPG applicable to an embodiment of the present application.

[0061] like Figure 2 As shown, the multi-mode fusion blood glucose monitoring method based on ECG and PPG may include:

[0062] S210: Obtain the ECG signal and PPG signal of the user.

[0063] It is understandable that the user's ECG signal and PPG signal can be obtained by Figure 1 The ECG and PPG data acquisition modules are completed, and the user's ECG signal and PPG signal can be collected by the wearable device. Among them, the wearable device integrates both the ECG signal collector and the PPG signal collector.

[0064] Among them, the ECG signal collector can adopt a one-channel three-electrode acquisition mode. When the user wears the ECG signal collector, the three electrodes can be placed on the left ankle, right ankle, and left wrist respectively; they can also be placed on other parts of the user's chest, arms, etc.

[0065] When a user wears the PPG signal collector, it can be placed on the user's fingers, wrists, earlobes, etc.

[0066] After wearing, the user starts the wearable device. After activation, the wearable device automatically detects the signal-to-noise ratio of the acquired ECG signal and PPG signal to evaluate the signal quality. If the signal-to-noise ratio is found to be less than the preset threshold (understandably, the preset threshold can be set according to actual needs), the user is prompted to recheck the status of the ECG signal collector and the PPG signal collector. If the signal quality meets the requirements, the ECG signal and the PPG signal are transmitted to the ECG and PPG data level fusion module.

[0067] S220: Perform data-level fusion on the ECG signal and the PPG signal.

[0068] It can be understood that data-level fusion of ECG signals and PPG signals can be performed by Figure 1 The ECG and PPG data level fusion module is completed. The specific steps are as follows:

[0069] 1) Determine the length of the ECG signal and the PPG signal: The signal length is calculated based on the heartbeat cycle. The default value is 20 heartbeat cycles, that is, N=20. Users can also set the signal length in the wearable device as long as N≥1 is satisfied.

[0070] 2) According to the set signal length, the quadratic difference method is used to extract the R point of the ECG signal and the P point of the PPG signal, such as Figure 3 (a) and Figure 3 As shown in (b), after a total of N R points and N P points are obtained, a command is sent to the ECG and PPG data acquisition modules to suspend data acquisition.

[0071] 3) Calculate the distance between adjacent R points and the distance between adjacent P points respectively and make judgments. If the distance between adjacent R points or the distance between adjacent P points is greater than a set threshold (it can be understood that the threshold can be set according to actual needs), send instructions to the ECG signal collector and the PPG signal collector to re-collect the ECG signal and the PPG signal.

[0072] 4) When the distance between adjacent R points and the distance between adjacent P points meet the requirements, for the acquired ECG signal, take 0.3×HR×SP data points to the left of the first R point, and take 0.7×HR×SP data points to the right of the Nth R point (i.e., the last R point), and save the above intercepted ECG signal as an ECG block and store it in data pool 1 (i.e., the first data pool). Similarly, for the acquired PPG signal, take 0.3×HR×SP data points to the left of the first P point, and take 0.7×HR×SP data points to the right of the Nth P point (i.e., the last P point), and save the above intercepted PPG signal as a PPG block and store it in data pool 1 (i.e., the first data pool). HR is the average heart rate of the user in seconds, and SP is the sampling rate of the wearable device in Hertz.

[0073] 5) For the acquired ECG signal, for each R point, 0.3×HR×SP data points are taken to the left and 0.7×HR×SP data points are taken to the right, and the intercepted ECG signal is saved as an ECG segment, such as Figure 3 As shown in (c) in FIG. 1 . Since the length of the ECG signal is N, a total of N ECG segments are obtained, and the N ECG segments are stored in data pool 2 (i.e., the second data pool). Similarly, for the acquired PPG signal, for each point P, 0.3×HR×SP data points are taken to the left and 0.7×HR×SP data points are taken to the right, and the intercepted PPG signal is saved as a PPG segment, as shown in FIG. Figure 3 As shown in (d) in FIG. 1 , since the length of the PPG signal is N, a total of N PPG segments are obtained, and the N PPG segments are stored in the data pool 2 (ie, the second data pool).

[0074] In the above, the ECG block and the PPG block are stored in the data pool 1, and the ECG segment and the PPG segment are stored in the data pool 2, thereby realizing the data-level fusion of ECG and PPG.

[0075] S230. Extract time series statistical features from the ECG block and the PPG block, and extract spatial morphological features from the ECG segment and the PPG segment; determine the statistical feature matrix based on the time series statistical features, determine the spatial morphological feature matrix based on the spatial morphological features, fuse the time series statistical features and the spatial morphological features to obtain a feature fusion matrix.

[0076] It can be understood that the feature-level fusion of ECG signals and PPG signals can be achieved by Figure 1 The ECG and PPG feature-level fusion modules are completed. For the convenience of description, the ECG block and the PPG block can be collectively referred to as data blocks.

[0077] Extract time series statistical features from ECG block and PPG block, including:

[0078] Perform discrete wavelet transform on the data block, decompose the signal at different levels, and obtain the signal detail coefficients corresponding to different levels; the data block includes ECG block and PPG block;

[0079] According to the wavelet signal reconstruction principle, signals with different detail coefficients are reconstructed respectively to obtain a first reconstructed signal;

[0080] The reconstructed signal is fused with the data block to obtain a data block matrix; the data block matrix includes an ECG block matrix and a PPG block matrix;

[0081] The time series statistical characteristics of each element in the ECG block matrix and the PPG block matrix are calculated respectively. The time series characteristics include at least the kurtosis, skewness, signal mobility, signal complexity, fractal dimension, power spectrum entropy, Kolmogorov entropy and Shannon entropy of each element.

[0082] like Figure 4 As shown, the specific steps are as follows:

[0083] 1) Perform discrete wavelet transform on the ECG block to perform signal decomposition at different levels. In this application, 7-layer decomposition is used as an example. After wavelet transform, signal detail coefficients of layers 1 to 7 are obtained. According to the principle of wavelet signal reconstruction, the signals of different detail coefficients are reconstructed respectively to obtain reconstructed signals, which are fused with the initial ECG block to obtain the ECG block matrix ECG_block, which contains 8 elements, namely ECG_block={ECG raw ,ECG detail1 ,...,ECG detail7}, where ECG raw For the initial ECG block, ECG detail1 ECG block reconstructed from the first layer detail coefficients based on wavelet transform, ECG detail2 is the ECG block reconstructed based on the second layer detail coefficients of wavelet transform, and so on. Similarly, the PPG block is processed in the same way, which will not be repeated here, and the PPG block matrix PPG_block is obtained, which contains 8 elements, namely PPG_block = {PPG raw ,PPG detail1 ,...,PPG detail7}, where PPG raw is the initial PPG block, PPG detail1 PPG is the PPG block reconstructed based on the first layer detail coefficients of wavelet transform. detail2 The PPG block is reconstructed based on the second layer detail coefficients of wavelet transform, and so on.

[0084] 2) Calculate the time series statistical characteristics of each element in the ECG block matrix ECG_block and the PPG block matrix PPG_block respectively. The time series statistical characteristics include the kurtosis (Kur), skewness (Ske), signal mobility (SM), signal complexity (SC), fractal dimension (FD), power spectrum entropy (PSE), Kolmogorov entropy (KE), Shannon entropy (SE), etc. of each element.

[0085] 3) Based on the feature selection method, the time series statistical features extracted in step 2) are analyzed. If the time series statistical features meet the following two conditions, they are considered to be valid features: a. The feature importance of the time series statistical features ranks in the top 50% among all features; b. The time series statistical features are selected simultaneously by three different feature selection algorithms, including the univariate feature selection algorithm (UFS), the recursive feature elimination algorithm (RFE), and the L1-based feature selection algorithm (L1-FS), that is, the time series statistical features are the intersection of the above three feature selection algorithms.

[0086] 4) By using three different feature selection algorithms to screen the time series statistical features, effective features are obtained, and the time series statistical feature matrix F is generated based on the obtained effective features. temporal and save it.

[0087] 5) Establish two residual network models (i.e., the first residual network model and the second residual network model), the residual network model includes two weight layers, and the input is directly connected to the output layer using a jumper. The ECG segment and the PPG segment are respectively used as the input of the residual network model. In the residual network model, the input ECG segment and the PPG segment are respectively subjected to two weight layers for spatial morphological feature extraction. At the same time, the input ECG segment and the PPG segment are directly connected to the output layer using a jumper. By obtaining the features passing through the weight layer and the residual values ​​of the ECG segment and the PPG segment, the residual network model is controlled by using the residual value feedback, thereby deepening the depth of the network to extract more spatial morphological features.

[0088] 6) After extracting the spatial morphological features of ECG and PPG segments through the residual network model, considering that the number of extracted features is relatively large, in order to reduce the calculation time, the spatial morphological features of ECG and PPG segments are used as the input of the deep neural network, and multiple hidden layers are used to compress and reduce the spatial morphological features of ECG and PPG segments, so as to generate the spatial morphological feature matrix F in the output layer of the deep neural network. morphological and save it.

[0089] 7) The time series statistical feature matrix F temporal and spatial morphological feature matrix F morphological Fusion is performed to obtain the feature fusion matrix Ffusion ={αF temporal ,βF morphological}, where α and β are the time series statistical feature matrices F temporal and spatial morphological feature matrix F morphological The weight of the model can be determined through clinical experience or based on the particle swarm optimization algorithm.

[0090] S240, using the time series statistical feature matrix, the spatial morphological feature matrix and the feature fusion matrix as inputs of several blood glucose prediction algorithms to obtain blood glucose prediction results.

[0091] It is understandable that this step can be Figure 1 The ECG and PPG decision-level fusion module is completed. Figure 5 As shown, the specific steps are as follows:

[0092] 1) The time series statistical feature matrix F obtained by the ECG and PPG feature level fusion module temporal , spatial morphological feature matrix F morphological , feature fusion matrix F fusion ={αF temporal ,βF morphological} are used as inputs of different blood sugar prediction algorithms. In this patent, three machine learning algorithms, namely support vector machine, random forest and naive Bayes, are used as examples of blood sugar prediction algorithms. When the above three features are respectively passed through the three machine learning algorithms, 3×3=9 different blood sugar prediction models can be obtained.

[0093] 2) Based on the above 9 different blood sugar prediction models, taking into account the continuity of blood sugar changes, obtain M historical blood sugar prediction values ​​(M≥1, the default value of M is M=7) and the current blood sugar prediction value of each blood sugar prediction model. The M historical blood sugar prediction values ​​and the current blood sugar prediction value are used as inputs of the fuzzy integral algorithm, and the results of the above 9 different blood sugar prediction models are analyzed according to the fuzzy integral operation to obtain 9 different blood sugar prediction values.

[0094] 3) According to the predicted blood glucose value, the prediction errors of different blood glucose prediction models are calculated, and the weight of each blood glucose prediction model is automatically determined according to the size of the prediction error, as shown in the formula:

[0095]

[0096] where ω i is the weight of the i-th blood glucose prediction model, R i is the prediction error of the i-th blood glucose prediction model, is the sum of the prediction errors of 9 different blood glucose prediction models, and r trepresents the blood glucose measurement value of the training set at time t, r' t represents the predicted blood glucose value of the training set at time t, T is the total number of blood glucose used for calculation, and T=M+1.

[0097] 4) According to the weight of each blood glucose prediction model obtained, the output value of each blood glucose prediction model is integrated to finally obtain the blood glucose prediction result BG fusion , as shown in the formula:

[0098]

[0099] Among them BG fusion To integrate the blood glucose prediction results of different blood glucose prediction models, ω i is the weight of the i-th blood glucose prediction model, r” i is the prediction result of the i-th blood glucose prediction model in the test set.

[0100] In the embodiments of the present application, the time series statistical features and spatial morphological features of the ECG signal and the PPG signal are extracted at the same time. Compared with other technologies, the extracted features are more comprehensive. In addition, the present application also realizes the fusion of ECG and PPG signals at three levels: data level, feature level, and decision level, which provides feasibility for improving the accuracy of non-invasive blood glucose monitoring.

[0101] It should be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of further restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0102] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

Claims

1. A multi-mode fusion blood glucose monitoring method based on ECG and PPG, characterized in that: The method comprises: Obtain the user's ECG signal and PPG signal; Performing data-level fusion on the ECG signal and the PPG signal, wherein the data-level fusion includes storing the ECG block and the PPG block in a first data pool, and storing the ECG segment and the PPG segment in a second data pool, wherein the ECG block and the ECG segment are both obtained by intercepting the ECG signal, including: Determine the length of the ECG signal and the PPG signal; according to the length of the ECG signal and the PPG signal, use the secondary difference method to extract the R point of the ECG signal and the PPG signal, and when the R point of the ECG signal and the PPG signal P point reach a preset number, stop acquiring the ECG signal and the PPG signal; respectively calculate the distance between adjacent R points and the distance between adjacent P points, and when the distance between adjacent R points and the distance between adjacent P points meet the requirements, for the ECG signal, start from the first R point to the left Taking 0.3×HR×SP data points and taking 0.7×HR×SP data points to the right of the last R point to intercept the ECG signal, and saving the intercepted part as the ECG block; for the PPG signal, taking 0.3×HR×SP data points to the left of the first P point and taking 0.7×HR×SP data points to the right of the last P point to intercept the PPG signal, and saving the intercepted part as the PPG block; wherein HR is the average heartbeat cycle of the user, and SP is the sampling rate of the wearable device; the wearable device is used to collect the ECG signal and the PPG signal; The PPG block and the PPG segment are both obtained by intercepting the PPG signal, including: for each R point of the ECG signal, 0.3×HR×SP data points are taken to the left and 0.7×HR×SP data points are taken to the right to intercept the ECG signal, and the intercepted part is saved as an ECG segment; For each point P of the PPG signal, 0.3×HR×SP data points are taken to the left and 0.7×HR×SP data points are taken to the right to intercept the PPG signal, and the intercepted parts are saved as PPG segments; Extracting time series statistical features from the ECG block and the PPG block, and extracting spatial morphological features from the ECG segment and the PPG segment; determining a time series statistical feature matrix according to the time series statistical features, determining a spatial morphological feature matrix according to the spatial morphological features, and fusing the time series statistical features and the spatial morphological features to obtain a feature fusion matrix; Determine several blood glucose prediction algorithms, use the time series statistical feature matrix as the input of these blood glucose prediction algorithms, use the spatial morphological feature matrix as the input of these blood glucose prediction algorithms, and use the feature fusion matrix as the input of these blood glucose prediction algorithms to obtain blood glucose prediction results.

2. The method according to claim 1, characterized in that The signal-to-noise ratios of the ECG signal and the PPG signal are both greater than a preset threshold.

3. The method according to claim 1, characterized in that When the distance between adjacent R points and the distance between adjacent P points do not meet the requirements, the ECG signal and the PPG signal are reacquired.

4. The method according to any one of claims 1 to 3, characterized in that: The extracting time series statistical features from the ECG block and the PPG block comprises: Performing discrete wavelet transform on the data block, performing signal decomposition at different levels, and obtaining signal detail coefficients corresponding to different levels; the data block includes the ECG block and the PPG block; According to the wavelet signal reconstruction principle, signals with different detail coefficients are reconstructed respectively to obtain a first reconstructed signal; The reconstructed signal is fused with the data block to obtain a data block matrix; the data block matrix includes an ECG block matrix and a PPG block matrix; The time series statistical characteristics of each element in the ECG block matrix and the PPG block matrix are calculated respectively, and the time series characteristics include at least the kurtosis, skewness, signal mobility, signal complexity, fractal dimension, power spectrum entropy, Kolmogorov entropy, and Shannon entropy of each element.

5. The method according to any one of claims 1 to 3, characterized in that: The step of determining a time series statistical feature matrix according to the time series statistical feature comprises: Based on the feature selection method, the time series statistical features are screened to obtain effective features; wherein the effective features satisfy: the time series statistical features rank in the top 50% of the feature importance among all the features and the time series statistical features are the intersection of three feature selection algorithms, and the three feature selection algorithms include: a univariate feature selection algorithm, a recursive feature elimination algorithm, and an L1-based feature selection algorithm; The time series statistical feature matrix is ​​generated according to the effective features.

6. The method according to any one of claims 1 to 3, characterized in that: Extracting spatial morphological features from the ECG segment and the PPG segment includes: Using the ECG segment as the input of the first residual network model and the PPG segment as the input of the second residual network model to extract spatial morphological features; Among them, the first residual network model and the second residual network model both include two weight layers, and the input is directly connected to the output layer using a jumper.

7. The method according to any one of claims 1 to 3, characterized in that: Determining a spatial morphological feature matrix according to the spatial morphological features includes: The spatial morphological features are used as input of a deep neural network, multiple hidden layers are used to perform feature compression and dimension reduction on the spatial morphological features, and the spatial morphological feature matrix is ​​generated in the output layer of the deep neural network.

8. The method according to any one of claims 1 to 3, characterized in that: The method of using the time series statistical feature matrix, the spatial morphological feature matrix and the feature fusion matrix as inputs of a plurality of blood glucose prediction algorithms to obtain blood glucose prediction results includes: Determine several blood sugar prediction algorithms, use the time series statistical feature matrix as the input of these blood sugar prediction algorithms, use the spatial morphological feature matrix as the input of these blood sugar prediction algorithms, and use the feature fusion matrix as the input of these blood sugar prediction algorithms, to obtain several different blood sugar prediction models; Acquire a number of historical blood sugar prediction values ​​and current blood sugar prediction values ​​of the several different blood sugar prediction models; Using the several historical blood sugar prediction values ​​and the current blood sugar prediction value as inputs of a fuzzy integral algorithm to obtain corresponding blood sugar prediction values; Calculating the prediction error of each of the blood glucose prediction models according to the blood glucose prediction value; Determining the weight of each of the blood glucose prediction models according to the size of the prediction error; According to the weight of each blood sugar prediction model, the output value of each blood sugar prediction model is fused to obtain the blood sugar prediction result.

Images