Blood glucose real-time filtering model training method and related device and storage medium

Abstract

The application relates to the technical field of blood glucose detection, and in particular to a blood glucose real-time filtering model training method and related equipment and a storage medium. The blood glucose real-time filtering model training method comprises the following steps: performing normalization processing on n original blood glucose data respectively to obtain n normalized data, the original blood glucose data being collected by a CGM device implanted in a living body at least in part, each original blood glucose data corresponding to a collection time; performing offline smoothing processing on the n original blood glucose data to obtain n smoothed blood glucose data, performing normalization processing on the n smoothed blood glucose data respectively to obtain n label data; inputting the n normalized data into a blood glucose real-time filtering model to obtain n result data output by the blood glucose real-time filtering model; and updating parameters of the blood glucose real-time filtering model according to differences between the n result data and the n label data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of blood glucose detection, and in particular to a training method of a blood glucose real-time filtering model and related equipment and a storage medium. BACKGROUND

[0002] Continuous glucose monitoring (CGM) is a technology that monitors the changes in blood glucose levels of diabetic patients by detecting the glucose levels in the interstitial fluid in real time. Traditional blood glucose monitoring methods, such as finger blood sampling, can only provide discrete blood glucose data and cannot accurately reflect the changes in blood glucose. The CGM technology can provide continuous blood glucose data by monitoring the glucose concentration in real time through a sensor implanted in the subcutaneous tissue, and can provide more accurate blood glucose level information for diabetic patients.

[0003] However, the raw blood glucose data obtained by the CGM device often has noise interference and instability, which brings certain difficulties to the accurate judgment of the blood glucose level. For example, the user wearing the CGM device has made a large amplitude of body movement at the blood glucose collection moment, which causes the position of the detection electrode of the CGM device in the user's body to have a short-term abnormal change, and thus causes the raw blood glucose data measured by the detection electrode to have a large deviation from the actual situation. SUMMARY

[0004] The purpose of the present application is to provide a training method of a blood glucose real-time filtering model and related equipment and a storage medium, so that the CGM device can obtain smooth and accurate blood glucose data.

[0005] In a first aspect, the present application provides a training method of a blood glucose real-time filtering model, comprising:

[0006] n original blood glucose data are respectively normalized to obtain n normalized data, wherein the original blood glucose data are collected by a CGM device implanted at least partially in a living body, each original blood glucose data corresponds to a collection moment, and n is a positive integer;

[0007] The n original blood glucose data are offline smoothed to obtain n smoothed blood glucose data, and the n smoothed blood glucose data are respectively normalized to obtain n label data;

[0008] The n normalized data are input into a blood glucose real-time filtering model to obtain n result data output by the blood glucose real-time filtering model;

[0009] According to the difference between the n result data and the n label data, the parameters of the blood glucose real-time filtering model are updated.

[0010] In some possible implementations, the normalization process for each of the n raw blood glucose data includes:

[0011] The n raw blood glucose data are normalized using the following formula:

[0012]

[0013] in, For the original blood glucose data, x i For the normalized data, x max and x min It is a preset constant, and x max Not less than the maximum value among the n raw blood glucose data, x min The value is not greater than the minimum value among the n original blood glucose data, i = 1, 2, 3...n.

[0014] In some possible implementations, the normalization process for the n smoothed blood glucose data includes:

[0015] The n smoothed blood glucose data are normalized using the following formula:

[0016]

[0017] in, For the smoothed blood glucose data, y i For the label data, y max and y min It is a preset constant, and y max Not less than the maximum value among the n raw blood glucose data, y min The value is not greater than the minimum value among the n original blood glucose data, i = 1, 2, 3...n.

[0018] In some possible implementations, y max =x max y min =x min .

[0019] In some possible implementations, the offline smoothing process of the n raw blood glucose data to obtain n smoothed blood glucose data includes:

[0020] The n raw blood glucose data are processed offline to obtain n preliminary smoothed data.

[0021] If any of the n preliminary smoothed data is greater than y max The first preliminary smoothed data is then assigned the value y. max If there is a value less than y among the n preliminary smoothing data...min a second preliminary smoothing data, then the second preliminary smoothing data is assigned as y min , thereby obtaining the n smoothing blood glucose data.

[0022] 6. The method of claim 5, wherein x max is an upper limit value of a range of the CGM device, x min is a lower limit value of the range of the CGM device.

[0023] 7. The method of claim 1, wherein the updating the parameters of the blood glucose real-time filtering model according to the difference between the n result data and the n label data comprises:

[0024] constructing a loss function wherein, y i is the label data, F θ (y i |x i ) is the result data.

[0025] updating the parameters of the blood glucose real-time filtering model according to the loss function Loss.

[0026] In some possible implementation manners, the offline smoothing processing comprises at least one of the following:

[0027] Kalman filtering processing, Savitzky-Golay filtering processing, and moving average filtering processing.

[0028] In some possible implementation manners, the blood glucose real-time filtering model is a time-domain convolution network model.

[0029] In some possible implementation manners, in the n collection time instants corresponding to the n original blood glucose data, a time interval between each two adjacent collection time instants is equal to one collection period of the CGM device.

[0030] In some possible implementation manners, the blood glucose real-time filtering model is installed in the CGM device, and the method is performed by the CGM device.

[0031] In some possible implementation manners, the CGM device comprises:

[0032] a bioelectric sensor having a detection electrode implanted under the skin of a living body;

[0033] a transmitter electrically connected to the bioelectric sensor;

[0034] a receiver wirelessly connected to the transmitter;

[0035] The blood glucose real-time filtering model is installed in the transmitter or the receiver.

[0036] In a second aspect, the present application provides a model training device, comprising:

[0037] A first processing module is configured to perform normalization processing on n original blood glucose data respectively to obtain n normalized data, wherein the original blood glucose data are collected by a CGM device implanted at least partially in a living body, each of the original blood glucose data corresponds to a collection time, and n is a positive integer.

[0038] A second processing module is configured to perform offline smoothing processing on the n original blood glucose data to obtain n smoothed blood glucose data, and perform normalization processing on the n smoothed blood glucose data to obtain n label data.

[0039] An input module is configured to input the n normalized data into a blood glucose real-time filtering model to obtain n result data output by the blood glucose real-time filtering model.

[0040] An update module is configured to update parameters of the blood glucose real-time filtering model according to a difference between the n result data and the n label data.

[0041] In a third aspect, the present application provides a model training device, comprising:

[0042] a memory,

[0043] a processor,

[0044] program instructions stored in the memory and executable by the processor;

[0045] When the program instructions are executed by the processor, the model training device performs the method of the first aspect.

[0046] In a fourth aspect, the present application provides a computer-readable storage medium storing program instructions, wherein when the program instructions are run on a computer device, the computer device performs the method of the first aspect.

[0047] In a fifth aspect, the present application provides a CGM device, comprising:

[0048] a bioelectric sensor having a detection electrode implanted under the skin of a living body;

[0049] a transmitter electrically connected to the bioelectric sensor;

[0050] a receiver wirelessly connected to the transmitter;

[0051] The transmitter or the receiver is installed with a blood glucose real-time filtering model, which is trained by the method of the first aspect.

[0052] In a sixth aspect, the present application provides a CGM device, comprising:

[0053] A bioelectric sensor having a detection electrode implanted under the skin of a living body;

[0054] A transmitter electrically connected to the bioelectric sensor;

[0055] A receiver wirelessly connected to the transmitter;

[0056] The receiver or the receiver includes a memory, a processor, program instructions stored in the memory and executable by the processor, and when the program instructions are executed by the processor, the receiver or the receiver executes the method of the first aspect.

[0057] According to the training method of the blood glucose real-time filtering model provided by the present application, the n original blood glucose data are normalized to obtain n normalized data, wherein the original blood glucose data are collected by at least partially implanting a CGM device in a living body, each original blood glucose data corresponds to a collection time, and n is a positive integer; the n original blood glucose data are offline smoothed to obtain n smoothed blood glucose data, and the n smoothed blood glucose data are normalized to obtain n label data; the n normalized data are input into the blood glucose real-time filtering model to obtain n result data output by the blood glucose real-time filtering model; and the parameters of the blood glucose real-time filtering model are updated according to the difference between the n result data and the n label data. In this way, the trained blood glucose real-time filtering model can output smoothed and accurate blood glucose data that are more consistent with the true situation. BRIEF DESCRIPTION OF DRAWINGS

[0058] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings of the embodiments will be briefly introduced below. Obviously, the drawings described below only relate to some embodiments of the present application, but not limit the present application.

[0059] Figure 1 is a training method flowchart of a blood glucose real-time filtering model provided by an embodiment of the present application.

[0060] Figure 2 is a training method flowchart of a blood glucose real-time filtering model provided by an embodiment of the present application. Figure 1 is a flowchart of step S102 in the method.

[0061] Figure 3 is a comparison curve of blood glucose data provided by an embodiment of the present application, wherein the origin data is original blood glucose data, and the fiter data is blood glucose result data obtained through the output of the trained blood glucose real-time filtering model and the inverse normalization processing. Figure 1 is a flowchart of step S104 in the method.

[0062] Figure 4 is a comparison curve of blood glucose data provided by an embodiment of the present application, wherein the origin data is original blood glucose data, and the fiter data is blood glucose result data obtained through the output of the trained blood glucose real-time filtering model and the inverse normalization processing. DETAILED DESCRIPTION

[0063] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions of the embodiments of the present application will be described clearly and completely below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, rather than all the embodiments of the present application. Based on the described embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without any creative effort belong to the scope of protection of the present application. It can be understood that some technical means of the various embodiments described herein can be mutually replaced or combined without conflict.

[0064] In the description of the present application, if there are terms such as “first”, “second”, etc., they are only used to distinguish the described objects, and do not have any order or technical meaning. Therefore, the objects defined with “first”, “second”, etc. can be explicitly or implicitly included one or more of the objects, and for example, the term “first element” itself does not mean the existence of “second element”, and the term “second element” itself does not mean the existence of “first element”. In addition, “one” or “an” and the like do not represent a quantity limitation, but represent the existence of at least one, and “multiple” represents no less than two.

[0065] In the description of the present application, if there are terms such as “based on”, “according to”, they are used to describe one or more factors that affect the determination. The terms do not exclude additional factors that affect the determination. That is, the determination can be based only on these factors or at least partially based on these factors. For example, the phrase “determining B based on A”, in this case, A is a factor that affects the determination of B, this phrase does not exclude the possibility that the determination of B can also be based on C.

[0066] In the description of the present application, reference to “one embodiment” or “some embodiments” and the like means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the present application. Therefore, the appearance of the phrases “in one embodiment”, “in some embodiments”, “in other some embodiments”, “in yet some embodiments” and the like in different places in the specification does not necessarily all refer to the same embodiment, but means “one or more but not all embodiments”, unless otherwise specifically emphasized.

[0067] The following describes a training method of a blood glucose real-time filtering model provided in embodiments of the present application, which comprises: Figure 1 The training method of the blood glucose real-time filtering model provided in embodiments of the present application comprises:

[0068] S101, n original blood glucose data are normalized respectively to obtain n normalized data, wherein the original blood glucose data are collected by a CGM device implanted at least partially in a living body, each original blood glucose data corresponds to a collection time, and n is a positive integer.

[0069] The n original blood glucose data are all collected by a CGM device implanted at least partially in a living body (usually a human body) at n different time points. For example, the CGM device can continuously collect n original blood glucose data with time stamps from the living body at a certain period, so that the interval between every adjacent two collection times in the n collection times is equal to the collection period of the CGM device, for example, 1 minute.

[0070] In some embodiments, the CGM device comprises a bioelectric sensor, a transmitter and a receiver. The transmitter and the bioelectric sensor are held by the same shell and electrically connected, and the shell can be adhered to the skin surface of a user by an adhesive, wherein the transmitter can be completely accommodated in the shell, the detection electrode of the bioelectric sensor extends from below the shell and is implanted into the user's body, i.e., under the user's skin, to obtain the original blood glucose data of the user based on the electrochemical reaction of the detection electrode and the glucose in the tissue fluid in the user's body, and the rest of the bioelectric sensor is accommodated in the shell. The transmitter and the receiver also have wireless communication function, and wireless communication connection can be established between them. In practice, the transmitter can send the original blood glucose data obtained by the bioelectric sensor to the receiver in a wireless communication manner, and the receiver can process the original blood glucose data received thereby (for example, by the blood glucose real-time filtering model described below) and then display it to the user. The receiver can be a mobile phone or a smart wearable device, such as a smart watch.

[0071] In order to reduce the influence of large values in the n original blood glucose data and process multiplicative noise, the n original blood glucose data are normalized in step S101 to be scaled to the range of [-1, 1].

[0072] Specifically, the n original blood glucose data can be normalized respectively by the following formula (logarithmic transformation):

[0073]

[0074] wherein, is the original blood glucose data, x ifor normalization, x max and x min are preset constants, and x max is not less than the maximum value in the n original blood glucose data, x min is not greater than the minimum value in the n original blood glucose data, i = 1, 2, 3…n

[0075] It can be understood that, in the case that x max is not less than the maximum value in the n original blood glucose data, and x min is not greater than the minimum value in the n original blood glucose data, each of the normalized data obtained can be ensured to be within the range of [-1, 1].

[0076] In some embodiments, x max may be an upper limit value of the range of the CGM device, x min may be a lower limit value of the range of the CGM device, such as x min = 2.2 mol / L, x max = 22.2 mol / L. In general, the range of the CGM device is the reading range of the CGM device.

[0077] S102, offline smoothing processing is performed on the n original blood glucose data to obtain n smoothed blood glucose data, and the n smoothed blood glucose data are respectively normalized to obtain n label data.

[0078] In this step S102, the offline smoothing processing can include at least one of Kalman filtering, Savitzky-Golay filtering processing, and moving average filtering processing. Since Kalman filtering, Savitzky-Golay filtering, and moving average filtering are well-known techniques in the field, they will not be described herein.

[0079] In this step S102, the n smoothed blood glucose data can be respectively normalized by the following formula (logarithmic transformation):

[0080]

[0081] wherein, is the smoothed blood glucose data, y i is the label data, y max and y min are preset constants, and y max is not less than the maximum value in the n original blood glucose data, y min is not greater than the minimum value in the n original blood glucose data, i = 1, 2, 3…n.

[0082] In general, the aforementioned y max , y min , x maxand x min are set as follows: y max = x max , y min = x min . For example, x max and y max are both set as the upper limit value of the range of the CGM device, and x min and y min are both set as the lower limit value of the range of the CGM device.

[0083] In some embodiments, the offline smoothing of the n original blood glucose data to obtain n smoothed blood glucose data in S102 specifically comprises:

[0084] S102a, offline smoothing of the n original blood glucose data to obtain n preliminary smoothed data;

[0085] S102b, if there is a preliminary smoothed data greater than y max , i.e. a first preliminary smoothed data, among the n preliminary smoothed data, the first preliminary smoothed data is assigned a value of y max , and if there is a preliminary smoothed data less than y min , i.e. a second preliminary smoothed data, among the n preliminary smoothed data, the second preliminary smoothed data is assigned a value of y min , thereby obtaining the aforementioned n smoothed blood glucose data.

[0086] The values of y max and y min are preset and generally equal to x max and x min respectively, so that there can be data greater than y max or / and data less than y min among the n preliminary smoothed data obtained by directly offline smoothing of the n original blood glucose data in S102a. In this case, if these preliminary smoothed data are directly processed according to the above formula, the purpose of normalization cannot be achieved. In this regard, in S102b, those preliminary smoothed data greater than y max are assigned a value of y max as the corresponding final smoothed blood glucose data, and those preliminary smoothed data less than y min are assigned a value of y min as the corresponding final smoothed blood glucose data, thereby avoiding the aforementioned problem. In addition, the preliminary smoothed data assigned a value of y max or y min generally differs very little from y max or y min , so that the training accuracy of the model will not be affected.

[0087] S103, inputting the n normalized data into the trained blood glucose real-time filtering model to obtain n result data output by the blood glucose real-time filtering model.

[0088] The blood glucose real-time filtering model can be a Temporal Convolutional Network (TCN) model, which mainly includes three parts of causal convolution, dilated convolution and residual block, wherein,

[0089] The causal convolution refers to the data at time t in the i-th layer, which only depends on the influence of the values at time t and before time t in the (i-1)-th layer, that is, if the input is X=(x1, x2, …, x n )∈R n×1 (one-dimensional time series data), then the causal convolution at x n is

[0090] The dilated convolution is different from the traditional convolution, and the dilated convolution allows the input to exist interval sampling during convolution. The sampling rate layer d controls, for example, d=1 represents that each point is sampled during input, and d=2 represents that every 2 points are sampled as input during input. If the input is X=(x1, x2, …, x n )∈R n×1 , then the dilated convolution of d at x n is

[0091] The residual block adds a constant mapping of the identity mapping of the cross-layer connection, and learns the transformation function H(x)=x.

[0092] The initial parameters of the blood glucose real-time filtering model can be pre-set by human or be system default values. After the n normalized data obtained through step S101 are input into the blood glucose real-time filtering model, the blood glucose real-time filtering model outputs n result data (here, it should be understood that the n result data are not the prediction of the blood glucose level of the living body at future time, and the time points corresponding to the n result data are consistent with the collection time of the n original blood glucose data, which are the speculation of the current or / and historical true blood glucose level). However, since the initial parameters of the blood glucose real-time filtering model are pre-set according to experience or other considerations, the initial parameters are likely to be not optimal, and thus the n result data can be greatly different from the true blood glucose condition.

[0093] S104, updating the parameters of the blood glucose real-time filtering model according to the difference between the n result data and the n label data.

[0094] In some embodiments, the blood glucose real-time filtering model can be installed in the GCM device, such as in the receiver of the GCM device. In this way, after the blood glucose real-time filtering model in the receiver is trained to update to appropriate parameters, when the transmitter sends the raw blood glucose data acquired by the bioelectric sensor to the receiver in a wireless communication manner, the receiver can input the raw blood glucose data received by it into the blood glucose real-time filtering model to obtain high-quality result data close to the real blood glucose level output by the blood glucose real-time filtering model (and after inverse normalization processing) and display the high-quality result data on the display screen of the receiver.

[0095] As mentioned above, because the parameters of the blood glucose real-time filtering model are not optimal before training, the result data obtained in step S102 can be quite different from the real blood glucose situation. Although high-quality blood glucose data closer to the real level of the user can be measured by other blood glucose measurement methods (such as real high-level chemical experiments), and the aforementioned blood glucose real-time filtering model can be trained according to the difference between the high-quality blood glucose data and the result data obtained in S102, the determination cost of real high-level chemical experiments is high, and this determination method can generally only be implemented by the manufacturer before the CGM device is sold, and cannot be implemented by the user who has purchased the CGM device, so the basis data (i.e. raw blood glucose data) for model training and the high-quality blood glucose data measured by high-level chemical experiments are not measured by the user himself / herself who has purchased the CGM device, and different users have individual differences, so that the blood glucose real-time filtering model finally trained can not be suitable for the user, that is, the parameters of the blood glucose real-time filtering model installed in the CGM device and already trained by the manufacturer can not be suitable for a specific user who has purchased the CGM device. In addition, the determination time interval of the aforementioned n result data is usually very short, for example, 1 minute, however, it is not practical to extract the blood of the organism every 1 minute to perform real high-level chemical experiments, so the number of label data obtained by real high-level chemical experiments cannot completely correspond to the number of raw blood glucose data collected by the CGM device, not to mention one-to-one correspondence at each collection time.

[0096] Advantageously, in step S104 of the embodiments of the present application, the label data used is not obtained by real high-level chemical experiments, but is obtained by offline smoothing and normalization processing of the original blood glucose data in S101, thereby avoiding the aforementioned problems and ensuring that the label data corresponds completely in quantity and collection time to the original blood glucose data. In implementation, after the detection electrode of the CGM device is implanted under the skin of the user and is ready, the user can trigger the "calibration mode" (or "training mode") of the CGM device to train the blood glucose real-time filtering model installed in the CGM device as described in S101 to S103 within a set time period (for example, within two hours or two days after the current time). Since the basic data (i.e., the original blood glucose data) for training the blood glucose real-time filtering model and the label data are both obtained based on measurements on the user himself, the parameters of the trained blood glucose real-time filtering model can better adapt to the individual differences of the user. Moreover, since the basic data for training the blood glucose real-time filtering model and the label data correspond completely in sample quantity and corresponding collection time, the training accuracy of the model is improved.

[0097] In some embodiments, referring to Figure 3 , step S104 specifically includes:

[0098] S104a, based on the difference between the n result data and the n label data, a loss function Loss is constructed;

[0099] S104b, according to the loss function, the parameters of the blood glucose real-time filtering model are updated;

[0100] wherein the loss function y i is the label data, F θ (y i |x i ) is the result data.

[0101] The method of steps S104a and S104b for updating the parameters of the blood glucose real-time filtering model is based on the consideration that L mse is the result data F θ (y i |x i ) and the label data y i is the mean square error loss function (for the sake of distinction and description, the aforementioned Loss can be referred to as the comprehensive loss function), and the component of the mean square error loss function is added to the comprehensive loss function, and the purpose is to make the result data F θ (y i |x i ) output by the trained blood glucose real-time filtering model as close as possible to the label data yi However, the result data F output by the real-time blood glucose filtering model after training... θ (y i |x i (Approximate label data y) i This is insufficient because the resulting data, when arranged chronologically, may not be smooth enough and may not match the actual situation (the actual blood glucose values ​​should be smooth over time). Therefore, to further achieve data smoothing, this application's embodiments add a second-order difference energy L loss function to the overall loss function Loss. ESOD The weight of the data means that the output data of the trained real-time blood glucose filtering model balances accuracy and smoothness, and is more consistent with the actual blood glucose situation of the organism.

[0102] It should be understood that in the various embodiments of this application, the execution order of each step should be determined by its function and internal logic. The order of each step number does not imply the order of execution and does not constitute any limitation on the implementation process of the embodiment. For example, the aforementioned steps S101 and S102 can be performed simultaneously or in any order. This application does not limit this, but step S104 needs to be executed after steps S101 and S102.

[0103] The inventors trained the model using the aforementioned method based on raw blood glucose data from 20 real patients. Specifically, 6720 raw blood glucose data points were continuously collected from each patient over 14 days using a CGM device, resulting in 6720 × 20 raw blood glucose data points. These 6720 × 20 raw blood glucose data points were then divided into a training set and a test set in a 7:3 ratio. Specifically, 6720 × 14 raw blood glucose data points were used for training the real-time blood glucose filtering model (14 training sets), while the remaining 6720 × 6 raw blood glucose data points were used to test the effectiveness of the trained real-time blood glucose filtering model (6 test sets). The training parameters of the real-time blood glucose filtering model are shown in Table 1. The results show that the final loss on the training set was 0.0023, and the loss on the test set was 0.0021.

[0104] Table 1: Model Training Parameters

[0105] input step output step epoch batch-size optimizer initial learning rate 15 15 30 32 Adam 0.0001

[0106] Compare MARD (Mean Absolute Error) and ESOD (Second Difference Energy) on six test sets, and with the aforementioned L ESODThe calculation method of the Origin-MARD is similar to that of the Filter-MARD, and is not described herein again. In Table 2, the Origin-MARD is the mean absolute error of the initial blood glucose data and the fingertip blood glucose data (BG), and the Filter-MARD is the mean absolute error of the blood glucose data output by the trained blood glucose real-time filtering model and the fingertip blood glucose data. The Origin-ESOD is the second-order difference energy of the initial blood glucose data, and the greater the ESOD, the less smooth the time series is; on the contrary, the smoother the time series is. The Filter-ESOD is the second-order difference energy of the blood glucose data output by the trained blood glucose real-time filtering model. In the foregoing, the fingertip blood glucose data is obtained by collecting the fingertip blood of the patient and then detected by a chemical experiment. Specifically, in the comparative experiment, the patient is required to collect fingertip blood 5 times a day to obtain 5 fingertip blood glucose data.

[0107] By comparing the Origin-MARD and the Filter-MARD, it can be seen that the average MARD of the smoothed data is 13.17, which is very close to the original 13.10; by comparing the Origin-ESOD and the Filter-ESOD, it can be seen that the ESOD of the smoothed data is reduced by 404. In general, the method proposed in the present application can smooth the original data while ensuring the accuracy of the blood glucose data.

[0108] Table 2: Comparison of MARD and ESOD on the test set

[0109] serial number Origin-MARD Origin-ESOD Filter-MARD Filter-ESOD 1 12.30 956.99 12.29 445.13 2 11.24 858.05 11.41 339.19 3 16.74 1018.22 16.73 400.82 4 14.49 495.82 14.46 230.24 5 12.54 491.51 12.74 189.75 6 11.29 406.89 11.38 196.35 average 13.10 704.58 13.17 300.25

[0110] The embodiment of the present application also provides a model training device, comprising:

[0111] A first processing module is configured to perform normalization processing on n original blood glucose data respectively to obtain n normalized data, wherein the original blood glucose data is collected by a CGM device implanted at least partially in a living body, each of the original blood glucose data corresponds to a collection time point, and n is a positive integer.

[0112] A second processing module is configured to perform offline smoothing processing on the n original blood glucose data to obtain n smoothed blood glucose data, and perform normalization processing on the n smoothed blood glucose data respectively to obtain n label data.

[0113] An input module is configured to input the n normalized data into a blood glucose real-time filtering model to obtain n result data output by the blood glucose real-time filtering model.

[0114] An updating module is configured to update parameters of the blood glucose real-time filtering model according to a difference between the n result data and the n label data.

[0115] The embodiment of the present application further provides a model training device, comprising a memory, a processor, program instructions stored in the memory and executable by the processor; when the program instructions are executed by the processor, the model training device executes the method described above. The model training device can be a CGM device (for example, a transmitter or a receiver of the CGM device), can also be a PC, and can also be a server in communication connection with the CGM device.

[0116] The embodiment of the present application further provides a computer readable storage medium, which stores program instructions, when the program instructions are run on a computer device, the computer device executes the method described above. The computer device can be a CGM device, can also be a PC, and can also be a server in communication connection with the CGM device.

[0117] The embodiment of the present application further provides a CGM device, comprising: a bioelectric sensor, having a detection electrode implanted under the skin of a living body; a transmitter, electrically connected with the bioelectric sensor; a receiver, in wireless communication connection with the transmitter; wherein the receiver is installed with a blood glucose real-time filtering model, and the blood glucose real-time filtering model is obtained by the method described above.

[0118] The embodiment of the present application further provides a CGM device, comprising: a bioelectric sensor, having a detection electrode implanted under the skin of a living body; a transmitter, electrically connected with the bioelectric sensor; a receiver, in wireless communication connection with the transmitter; wherein the transmitter or the receiver comprises a memory, a processor, program instructions stored in the memory and executable by the processor, when the program instructions are executed by the processor, the transmitter or the receiver executes the method described above. In this embodiment, the blood glucose real-time filtering model obtained by the method described above can be reserved in the transmitter or the receiver to play a role in the subsequent blood glucose monitoring.

Claims

1. A method for training a real-time blood glucose filtering model, characterized in that, The method comprises the following steps: normalizing n original blood glucose data respectively to obtain n normalized data, wherein the original blood glucose data are collected by a CGM device implanted at least partially in a living body, each of the original blood glucose data corresponds to a collection time, and n is a positive integer; performing offline smoothing processing on the n original blood glucose data to obtain n smoothed blood glucose data, and normalizing the n smoothed blood glucose data respectively to obtain n label data; inputting the n normalized data into a blood glucose real-time filtering model to obtain n result data output by the blood glucose real-time filtering model; updating parameters of the blood glucose real-time filtering model according to a difference between the n result data and the n label data.

2. The method of claim 1, wherein, The step of normalizing the n original blood glucose data respectively comprises the following steps: The step of normalizing the n original blood glucose data respectively comprises the following steps: wherein, is the original blood glucose data, x i is the normalized data, x max and x min is a preset constant, and x max is not less than the maximum value in the n original blood glucose data, x min is not greater than the minimum value in the n original blood glucose data, i = 1, 2, 3...n.

3. The method of claim 2, wherein, The step of normalizing the n smoothed blood glucose data respectively comprises the following steps: The step of normalizing the n smoothed blood glucose data respectively comprises the following steps: wherein, is the smoothed blood glucose data, y i is the label data, y max and y min is a preset constant, and y max is not less than the maximum value in the n original blood glucose data, y min is not greater than the minimum value in the n original blood glucose data, i = 1, 2, 3...n.

4. The method of claim 3, wherein, y max = x max , y min = x min .

5. The method according to claim 3 or 4, characterized in that, The step of performing offline smoothing processing on the n original blood glucose data to obtain n smoothed blood glucose data comprises the following steps: performing offline smoothing processing on the n original blood glucose data to obtain n preliminary smoothed data; If there is a first preliminary smoothed data greater than y max in the n preliminary smoothed data, the first preliminary smoothed data is assigned as y max If there is a second preliminary smoothed data less than y min in the n preliminary smoothed data, the second preliminary smoothed data is assigned as y min , thereby obtaining the n smoothed blood glucose data.

6. The method of claim 5, wherein, x max is an upper limit value for the range of the CGM device, x min is a lower limit value for the range of the CGM device.

7. The method of claim 1, wherein, The step of updating the parameters of the blood glucose real-time filtering model according to the difference between the n result data and the n label data comprises the following steps: Constructing loss function wherein, y i is the label data, F θ (y i | x i ) is the result data; The step of updating the parameters of the blood glucose real-time filtering model according to the loss function Loss comprises the following steps:

8. The method of claim 1, wherein, The offline smoothing processing comprises at least one of the following: Kalman filtering processing, Savitzky-Golay filtering processing, and moving average filtering processing.

9. The method of claim 1, wherein, The blood glucose real-time filtering model is a time-domain convolution network model.

10. The method of claim 1, wherein, In the n collection times corresponding to the n original blood glucose data, a time interval between every two adjacent collection times is equal to a collection period of the CGM device.

11. The method according to any one of claims 1 to 10, characterized in that, The blood glucose real-time filtering model is installed in the CGM device, and the method is performed by the CGM device.

12. The method of claim 10, wherein, The CGM device comprises: a bioelectric sensor having a detection electrode implanted under the skin of a living body; a transmitter electrically connected to the bioelectric sensor; a receiver wirelessly connected to the transmitter; The blood glucose real-time filtering model is installed in the transmitter or the receiver.

13. A model training apparatus, comprising: The method comprises the following steps: a first processing module configured to normalize n original blood glucose data respectively to obtain n normalized data, wherein the original blood glucose data are collected by a CGM device implanted at least partially in a living body, each of the original blood glucose data corresponds to a collection time, and n is a positive integer; a second processing module configured to perform offline smoothing processing on the n original blood glucose data to obtain n smoothed blood glucose data, and normalize the n smoothed blood glucose data respectively to obtain n label data; an input module configured to input the n normalized data into a blood glucose real-time filtering model to obtain n result data output by the blood glucose real-time filtering model; an updating module configured to update parameters of the blood glucose real-time filtering model according to a difference between the n result data and the n label data. An updating module is configured to update parameters of the blood glucose real-time filtering model according to differences between the n result data and n label data, wherein the n label data is obtained by performing offline smoothing processing on the n normalized data.

14. A model training apparatus, comprising: Comprise: a memory, a processor, program instructions stored in the memory and executable by the processor; When the program instructions are executed by the processor, the model training apparatus executes the method of any one of claims 1 to 11.

15. A computer readable storage medium storing program instructions, wherein the program instructions comprise instructions for causing a computer to perform the method of any one of claims 1-14. When the program instructions are executed on the computer device, the computer device executes the method of any one of claims 1 to 12.

16. A CGM device comprising: a bioelectric sensor having a detection electrode implanted under the skin of a living body; a transmitter electrically connected to the bioelectric sensor; a receiver wirelessly connected in communication with the transmitter; characterized in that the transmitter or the receiver is installed with a blood glucose real-time filtering model, and the blood glucose real-time filtering model is trained by the method of any one of claims 1 to 12.

17. A CGM device comprising: a bioelectric sensor having a detection electrode implanted under the skin of a living body; a transmitter electrically connected to the bioelectric sensor; a receiver wirelessly connected in communication with the transmitter; characterized in that the receiver or the receiver comprises a memory, a processor, program instructions stored in the memory and executable by the processor, and when the program instructions are executed by the processor, the receiver or the receiver executes the method of any one of claims 1 to 12.

Images