A continuous blood glucose dangerous state early warning method and device based on a deep learning model

Abstract

The application provides a continuous blood glucose dangerous state early warning method and device based on a deep learning model, and the method comprises the following steps: constructing a blood glucose prediction model based on an LSTM neural network; performing data preprocessing on blood glucose data sampled by a dynamic blood glucose monitoring device; using a projection neural network to predict blood glucose data of a next moment of a patient according to the processed blood glucose data, predicting future blood glucose values of the patient and whether the patient is in a dangerous state; performing dangerous state early warning, and providing an individualized treatment scheme for each patient according to the predicted dangerous state of the patient. The application predicts short-time blood glucose conditions of a patient by using continuous blood glucose monitoring, and predicts whether the blood glucose level of the patient in a short time enters a dangerous range.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of blood glucose prediction, in particular to a continuous blood glucose dangerous state early warning method and device based on a deep learning model. BACKGROUND

[0002] The previous method only stays in the prediction of blood glucose value, and the specific method includes regression analysis, deep learning prediction model and other methods, and the future blood glucose value is predicted according to the past blood glucose value. However, it is not enough to only provide the predicted blood glucose value: first, different patients may have different sensitivities to dangerous blood glucose due to different constitutions and states, that is, different patients should have different dangerous blood glucose thresholds. For example, some patients may secrete hormones when the blood glucose is reduced to 3.0 mmol / L, so as to rapidly increase the blood glucose to the normal range through their own endocrine, and this short-time, occasional and self-adjusted normal blood glucose level is not dangerous, but the traditional method will misreport that the patient is in a dangerous state when predicting that the future blood glucose is 3.0 mmol / L. Secondly, it is not very intuitive for doctors to understand only the blood glucose value. A better display method should not only enable doctors to quickly obtain the prediction of the future blood glucose value by the model, but also enable doctors to quickly infer the future blood glucose trend of the patient from the display content, and only giving the predicted blood glucose value cannot achieve this purpose. SUMMARY

[0003] In view of the defects in the prior art, the purpose of the present application is to provide a continuous blood glucose dangerous state early warning method and device based on a deep learning model, which predicts the short-time blood glucose condition of the patient by using continuous blood glucose monitoring, and predicts whether the blood glucose level of the patient in a short time enters a dangerous range.

[0004] To solve the above problems, the technical scheme of the present application is as follows:

[0005] A continuous blood glucose dangerous state early warning method based on a deep learning model, comprising the following steps:

[0006] Constructing a blood glucose prediction model based on an LSTM neural network;

[0007] Performing data preprocessing on the blood glucose data sampled by a dynamic blood glucose monitoring device;

[0008] According to the processed blood glucose data, using a projection neural network to predict the blood glucose data of the patient at the next time, and predicting the future blood glucose value of the patient and whether the patient is in a dangerous state;

[0009] Performing dangerous state early warning, and providing an individualized treatment plan for each patient according to the predicted dangerous state of the patient.

[0010] Preferably, the step of constructing the blood glucose prediction model based on the LSTM neural network specifically comprises: the LSTM neural network encodes the basic information, the insulin type and the drug type through an encoding layer to obtain corresponding non-temporal feature layer vector expressions, then splices the time series dose information of the blood glucose, the insulin and the drug in the time dimension to obtain a feature expression representing the medication and personal condition of the patient at each time point, the time series information passes through an LSTM layer to obtain a time series feature layer vector expression, and then the time series feature layer vector and the non-temporal feature layer vector expression are spliced to obtain an overall feature layer vector expression of the patient, and the blood glucose prediction value is output after a plurality of non-linear layers are mapped.

[0011] Preferably, the step of performing data preprocessing on the blood glucose data sampled by the dynamic blood glucose monitoring device specifically comprises:

[0012] filtering the blood glucose time series information using a Savgol filter to define a danger degree;

[0013] After the danger degree is obtained, whether the blood glucose is in an ascending phase or a descending phase in a period of time is judged according to the maximum value and the minimum value;

[0014] filtering adjacent local minimum value point pairs and maximum value point pairs;

[0015] deleting the maximum value points and the minimum value points in the descending and ascending trend lines.

[0016] Preferably, the step of filtering the blood glucose time series information using a Savgol filter to define a danger degree specifically comprises: the blood glucose time series information is filtered in advance using a Savgol filter to define the danger degree as:

[0017]

[0018] wherein i is a time point, S i is the blood glucose value at the i th time point, 1 is a low blood glucose threshold value, L is a high blood glucose threshold value, and C is a score coefficient for low blood glucose.

[0019] Preferably, the step of obtaining the danger degree and judging whether the blood glucose is in an ascending phase or a descending phase in a period of time according to the maximum value and the minimum value specifically comprises: from a local minimum value to a local maximum value is an ascending phase, and from a local maximum value to a local minimum value is a blood glucose descending phase, and the local minimum value of the time series blood glucose S i satisfies:

[0020] s i = min(s j ), j∈[i-h, i=h+1,..., i+h], and the local maximum value satisfies:

[0021] s i= max(s j ), j e [i-h, i = h+1,... i+h].

[0022] Preferably, the step of predicting the blood glucose data of the patient at the next time point, predicting the future blood glucose value of the patient and whether the patient is in a dangerous state according to the processed blood glucose data using a projection neural network specifically comprises: calculating how many steps the current patient is away from the next dangerous state according to the low blood glucose threshold 1, the high blood glucose threshold L and the local extreme point, the local minimum value below 1 and the local maximum value above L are dangerous states, defining the blood glucose sequence S i from l to the local minimum value and from L to the local maximum value as "entering a dangerous state", from the local minimum value to l and from the local maximum value to L as "leaving a dangerous state", calculating how many steps away from the dangerous state and how many steps away from the dangerous state, denoted as t i .

[0023] Preferably, the step of predicting the blood glucose data of the patient at the next time point, predicting the future blood glucose value of the patient and whether the patient is in a dangerous state according to the processed blood glucose data using a projection neural network further comprises: using a projection neural network to predict the data at the next time point, the input of the projection neural network is the intermediate layer feature of the LSTM blood glucose prediction network, and the output is the data S T+1 at the next time point and t T+1 .

[0024] Further, the present application also provides a continuous blood glucose dangerous state early warning device based on a deep learning model, the device comprising a processor and a memory for storing executable instructions of the processor, the processor being configured to execute the executable instructions to perform the continuous blood glucose dangerous state early warning method based on a deep learning model as described above.

[0025] Compared with the prior art, the present application predicts the short-term blood glucose condition of the patient using continuous blood glucose monitoring, and predicts whether the blood glucose level of the patient in a short time enters a dangerous range, not only can accurately predict the future blood glucose value of the patient, but also can early warn the high and low blood glucose condition that may appear in the future, predict the dangerous state of the patient, so that the doctor can provide a more personalized treatment plan for each patient. BRIEF DESCRIPTION OF DRAWINGS

[0026] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments, made with reference to the accompanying drawings:

[0027] Figure 1 The flow chart of the continuous blood glucose dangerous state early warning method based on a deep learning model provided by the embodiments of the present application;

[0028] Figure 2 An LSTM neural network structure schematic diagram provided for an embodiment of the present application is shown in FIG. 1.

[0029] Figure 3 A blood glucose data preprocessing flowchart provided for an embodiment of the present application is shown in FIG. 2.

[0030] Figure 4 A patient state curve diagram provided for an embodiment of the present application is shown in FIG. 3. DETAILED DESCRIPTION

[0031] The present application will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the present application, and make the technical solutions and beneficial effects of the present application apparent.

[0032] Specifically, Figure 1 A continuous blood glucose risk state early warning method flowchart based on a deep learning model provided for an embodiment of the present application is shown in FIG. 4. Figure 1 As shown, the method comprises the following steps:

[0033] S1: Constructing a blood glucose prediction model based on an LSTM neural network;

[0034] Specifically, an LSTM neural network is used to train a blood glucose prediction model, and the input data is from 218 patients wearing continuous glucose monitoring devices (CGM) collected by Zhongshan Hospital from 2019 to 2021, and the collected information includes basic information (height, weight, blood pressure, etc.), laboratory test information (such as blood routine, urine routine, blood biochemistry, etc.), blood glucose time series data (sampling interval is 15 minutes), insulin use record, hypoglycemic, hormone drug use record. The LSTM neural network first encodes the basic information, insulin type and drug type (non-time series) through the encoding layer to obtain the corresponding non-time series feature layer vector expression, and then splices the time series dose information of blood glucose, insulin and drugs in the time dimension to obtain the feature expression representing the medication and personal condition of the patient at each time. Subsequently, the time series information is obtained through the LSTM layer to obtain the time series feature layer vector expression, and then the time series feature layer vector and the non-time series feature layer vector expression are spliced to obtain the overall feature layer vector expression of the patient. After mapping through multiple nonlinear layers, the blood glucose prediction value is output. The LSTM neural network structure is shown in FIG. 1. Figure 2 The trainer adopts Adam and trains for 15 rounds.

[0035] S2: Data preprocessing of blood glucose data sampled by dynamic blood glucose monitoring equipment;

[0036] Specifically, the data preprocessing process of the blood glucose data sampled by the dynamic blood glucose monitoring equipment is as shown in FIG. 2.Figure 3 as shown, comprising the following steps:

[0037] S21: using Savgol filter to filter the blood glucose time series information, defining the danger degree;

[0038] Because the blood glucose data is sampled by dynamic blood glucose monitoring equipment, and there is fluctuation in a very short time, which is not conducive to model prediction, Savgol filter is used to filter the blood glucose time series information in advance. The principle of Savgol filter is to perform polynomial fitting on the local signal, and the parameters include window length and polynomial order. From the results, it is found that the effect of using a 5th order polynomial with a window length of 25 units to filter the blood glucose data is the best. Then, the danger degree is defined as:

[0039]

[0040] where i is the time point, S i is the blood glucose value at time i, l is the low blood glucose threshold, L is the high blood glucose threshold, and C is the score coefficient for low blood glucose. The reason for designing the low blood glucose score coefficient is that low blood glucose usually has a greater degree of harm to the patient, which may cause shock and other conditions, so a larger coefficient is set to expand its influence.

[0041] S22: After obtaining the danger degree, determine whether the blood glucose is in the rising stage or the falling stage in a period of time according to the maximum and minimum values;

[0042] After obtaining the danger degree, determine whether the blood glucose is in the rising stage or the falling stage in a period of time according to the maximum and minimum values, i.e. from local minimum to local maximum is the rising stage, and from local maximum to local minimum is the falling stage of blood glucose. Define the local minimum of time series blood glucose S i s i = min(s j ), j ∈ [i-h, i = h+1,... i+h], and the local maximum satisfies: s i = max(s j ), j ∈ [i-h, i = h+1,... i+h], and h = 8 can be taken in practice, i.e. the local minimum and maximum blood glucose points within two hours.

[0043] S23: filter adjacent local minimum point pairs and maximum point pairs;

[0044] However, due to the complex and changeable actual patient blood glucose, there are usually two adjacent local minimum values or two adjacent local maximum values, which should actually be ignored, because two adjacent local minimum values or maximum values mean that the patient's blood glucose has been in a low (high) state for a period of time, and the larger and smaller extreme points can be removed.

[0045] Another situation that should be ignored is that max(s i-h , s i-h+1 ,... s i+h )-min(s i-h , s i-h+1， ... s i+h )<∈, that is, the patient's blood glucose fluctuates little in this time window. For this situation, the local extreme values in the i-h, i=h+1,... i+h time are removed.

[0046]

[0047] S24: Remove the maximum and minimum points in the falling and rising trend lines.

[0048] After the above processing, the blood glucose curve has been smoothed, not only reflecting the rising and falling trend of blood glucose, but also marking the local maximum and minimum values.

[0049] S3: According to the processed blood glucose data, a projection neural network is used to predict the patient's blood glucose data at the next time, to predict the patient's future blood glucose value and whether the patient is in a dangerous state.

[0050] Specifically, according to the blood glucose sequence, the number of steps remaining to reach the dangerous state and the number of steps that have passed to leave the dangerous state are calculated, that is, according to the low blood glucose threshold 1, the high blood glucose threshold L and the above local extreme points, the number of steps remaining to reach the next dangerous state for the current patient is calculated. The dangerous state is defined at the local extreme point, however, not all extreme points are dangerous states, specifically, the local minimum value below 1 and the local maximum value above L are dangerous states. Define the blood glucose sequence S i from 1 to the local minimum value and from L to the local maximum value as "entering a dangerous state", and from the local minimum value to 1 and from the local maximum value to L as "leaving a dangerous state". According to this, the number of steps remaining to reach the dangerous state and the number of steps that have passed to leave the dangerous state are calculated, denoted as t i .

[0051] At this point, S i and t i are obtained, and a projection neural network is used to predict the data at the next time, the input of the projection neural network uses the intermediate layer features of the LSTM blood glucose prediction network in the above step S1, and the output is the data S​T+1 and t T+1 The network consists of two fully connected layers, with Leaky ReLU used as the non-linear activation function in between. For each patient, we trained the Adam optimizer for 1000 epochs, minimizing the L2 norm of the predicted and actual values. A two-dimensional patient state curve was then plotted, as shown below. Figure 4 As shown, Figure 4 The black dots represent patient history, and the arrows between two dots indicate the time sequence. The horizontal and vertical axes represent t, respectively. i and S i The background color in the image represents the level of danger, with lighter colors indicating danger and darker colors indicating safety. This patient's dangerous condition is characterized by hyperglycemia, but not hypoglycemia. From... Figure 4 As can be seen, the patient exhibits distinct paths into and out of a dangerous state (three similar paths in the upper right corner) and out of that state (four similar paths in the upper left corner). This graph depicts the trend of blood glucose changes during the patient's hospitalization and the changes during state transitions. Based on this graph, we can not only estimate the patient's future blood glucose levels but also estimate whether the patient is in a dangerous state and approximately how long until the next dangerous state. The triangle in the graph shows the patient's current state as (8.4, 5), where the first value represents blood glucose and the second represents 5 time units until the dangerous state. The asterisks indicate the projected model's prediction of the patient's state 15 minutes later. It can be seen that our network prediction (cross) highly matches the actual result (asterisk), indicating that our network successfully predicted the patient's blood glucose S 15 minutes later. T+1 and the number of dangerous state steps t T+1 .

[0052] S4: Provide early warning of dangerous conditions and offer personalized treatment plans for each patient based on the predicted patient risk status.

[0053] This invention not only accurately predicts future blood glucose levels but also predicts the patient's current state, i.e., their risk level. This allows doctors to provide more personalized treatment plans for each patient. For example, when a patient appears to be moving towards a lighter-colored danger zone, the final risk level can be estimated based on previous similar blood glucose spikes. If the patient's blood glucose was well controlled in previous spikes, the previous treatment plan can be continued; if the doctor is dissatisfied with the previous plan and believes the blood glucose is too high, the insulin dose can be increased, thereby forcing the patient's current state to move towards a darker, safer direction.

[0054] The application further discloses a continuous blood glucose dangerous state early warning device based on a deep learning model, which comprises a processor and a memory for storing executable instructions of the processor, and the processor is configured to execute the executable instructions to perform the continuous blood glucose dangerous state early warning method based on the deep learning model.

[0055] In summary, the application predicts the short-time blood glucose condition of a patient by using continuous blood glucose monitoring, and predicts whether the blood glucose level of the patient in a short time enters a dangerous range, which can not only accurately predict the future blood glucose value of the patient, but also can early warn the high and low blood glucose condition that may occur in a future period of time, predict the dangerous state of the patient, so that the doctor can provide a more personalized treatment plan for each patient.

[0056] The specific embodiments of the application are described above. It should be understood that the application is not limited to the specific embodiments described above, and various changes or modifications can be made by those skilled in the art within the scope of the claims, which does not affect the essential content of the application. The embodiments of the present application and the features in the embodiments can be arbitrarily combined without conflict.

Claims

1.A device for early warning of continuous blood glucose dangerous state based on a deep learning model, characterized in that, The device performs the following steps: Constructing a blood glucose prediction model based on an LSTM neural network; Data preprocessing is performed on blood glucose data sampled by a dynamic blood glucose monitoring device, specifically including: Filtering blood glucose time series information using a Savgol filter to define a dangerous degree, specifically including: where i is a time point, S i is a blood glucose value at time i, I is a low blood glucose threshold value, L is a high blood glucose threshold value, and C is a score coefficient for low blood glucose. After the degree of danger is obtained, it is judged according to the maximum and minimum whether the blood glucose is in an ascending stage or a descending stage in a period of time, specifically including: from a local minimum to a local maximum is an ascending stage, and from a local maximum to a local minimum is a blood glucose descending stage, and the local minimum of the time-series blood glucose S i satisfies: s i = min(s j ), j e [i - h, i = h + 1,... i + h], local maximum satisfies: s i = max(s j ), j e [-h, i = h + 1,... i + h] Filtering adjacent local minimum point pairs and maximum point pairs; Deleting maximum points and minimum points in the downward and upward trend lines; According to the processed blood glucose data, a projection neural network is used to predict the blood glucose data of the patient at the next time, predict the future blood glucose value of the patient and whether the patient is in a dangerous state, specifically comprising: according to a low blood glucose threshold value l, a high blood glucose threshold value L and a local extreme point, calculating how many steps the current patient is away from the next dangerous state, the local minimum value below l and the local maximum value above L are dangerous states, defining a blood glucose sequence S i from l to the local minimum value and from L to the local maximum value as "entering a dangerous state", from the local minimum value to l and from the local maximum value to L as "leaving a dangerous state", calculating how many steps are left to reach the dangerous state and how many steps have been taken to leave the dangerous state, denoted as t i ; using a projection neural network to predict the next time data, the input of the projection neural network is the intermediate layer feature of the LSTM blood glucose prediction network, and the output is the next time data S T+1 and t T+1 ; Performing a dangerous state early warning and providing personalized treatment plans for each patient according to the predicted patient dangerous state. 2.The deep learning model-based continuous blood glucose critical state early warning device according to claim 1, wherein, The step of constructing a blood glucose prediction model based on an LSTM neural network specifically includes: The LSTM neural network encodes basic information, insulin types, and drug types through an encoding layer to obtain corresponding non-temporal feature layer vector expressions, then splices blood glucose, insulin, and drug time series dose information in the time dimension to obtain feature expressions representing the medication and personal conditions of the patient at each time, the time series information is obtained through an LSTM layer to obtain a time series feature layer vector expression, then the time series feature layer vector and the non-temporal feature layer vector expression are spliced to obtain an overall feature layer vector expression of the patient, and after being mapped through multiple nonlinear layers, a blood glucose prediction value is output.

Images