Analyte event detection method and device based on machine learning and medium

Abstract

The invention discloses an analyte event detection method and device based on machine learning and a medium, and the detection method comprises the steps: obtaining a plurality of models, each model being a machine learning model obtained by training first data of a first time period; acquiring second data of a second time period corresponding to the first time period; processing the corresponding second data through a plurality of models to predict a plurality of first estimated values at a target moment after a plurality of second time periods; determining a plurality of second estimation values based on the plurality of first estimation values, determining corresponding analyte fluctuation information based on each first estimation value of at least a part of the plurality of first estimation values and corresponding second data, and determining a second estimation value corresponding to each first estimation value based on comparing the analyte fluctuation information with preset fluctuation; determining a combined estimate based on a combination of the plurality of second estimates; and acquiring an analyte event at the target time based on the combined estimate. According to the invention, the accuracy of detecting the analyte event can be improved.

Description

Technical Field

[0001] The present disclosure generally relates to the field of intelligent medical technology, and more particularly to a method, device and medium for detecting analyte events based on machine learning. Background Art

[0002] Diabetes has become a public health problem worldwide. If diabetes is not well controlled, it may cause a series of complications. At present, continuous glucose monitoring systems are widely used in blood sugar control. Through continuous glucose monitoring systems, users can obtain glucose concentrations and changes in glucose concentrations in real time, and intervene in time to reduce the risk of hyperglycemia or hypoglycemia, thereby reducing the occurrence of complications. Therefore, if the future glucose concentration can be known in advance, the occurrence of complications can be further reduced.

[0003] In recent years, researchers have used machine learning technology to analyze glucose data to predict future glucose concentrations. Specifically, in some machine learning-based studies, the user's glucose data over a period of time is often used to train the model, and the trained model is used to predict future glucose concentrations, and then glucose events (such as hyperglycemia or hypoglycemia) are obtained based on the predicted glucose concentration.

[0004] However, the above methods generally train a single model to characterize the changing pattern of glucose concentration and thus predict future glucose concentrations. However, the changes in the user's glucose concentration often do not follow a fixed pattern. A model that performs well in some changes may not accurately predict glucose concentrations in other changes, and may provide erroneous glucose events. Summary of the invention

[0005] The present disclosure is proposed in view of the above-mentioned situation, and its purpose is to provide a method, device and medium for detecting analyte events based on machine learning, which can improve the accuracy of detecting analyte events.

[0006] To this end, a first aspect of the present disclosure provides a method for detecting an analyte event based on machine learning, comprising: obtaining a plurality of models, each of the plurality of models being a machine learning model obtained by training with first data of a first time period, the first data being a time series provided by an analyte sensor continuously measuring an analyte level of a target object; obtaining second data of a second time period corresponding to the first time period, wherein at least a portion of the first time period is before the corresponding second time period; processing the corresponding second data by each of the plurality of models to predict a plurality of first estimated values ​​at a target moment after the plurality of second time periods; determining a plurality of second estimated values ​​based on the plurality of first estimated values, wherein analyte fluctuation information corresponding to each first estimated value is determined based on each first estimated value of at least a portion of the plurality of first estimated values ​​and the second data corresponding to each first estimated value, comparing the analyte fluctuation information with a preset fluctuation, and determining the second estimated value corresponding to each first estimated value based on comparing the analyte fluctuation information with the preset fluctuation; determining a combined estimated value based on a combination of the plurality of second estimated values; and obtaining the analyte event at the target moment based on the combined estimated value.

[0007] In the first aspect of the present disclosure, by combining the estimated values ​​of multiple models to determine the combined estimated value, it is possible to integrate the advantages of different models and compensate for the shortcomings of different models, thereby improving the accuracy of the combined estimated value, and thereby improving the accuracy of detecting analyte events. In addition, the first data is a time series, which enables the model to learn the characteristics of the change of analyte levels over time, thereby improving the accuracy of the predicted estimated value. In addition, the combined estimated value is determined in combination with the analyte fluctuation information, and the target estimated value is constrained at the application level of the model, which is conducive to reducing the prediction deviation, thereby reducing the requirements for the model. In addition, the combined estimated value is determined in combination with the analyte fluctuation information, and the difficulty of combining multiple estimated values ​​corresponding to multiple models can also be reduced.

[0008] In addition, in the detection method involved in the first aspect of the present disclosure, optionally, the parameters of at least a portion of the multiple models are parameters with the smallest time complexity. In this case, the training efficiency of the model and the prediction efficiency during application can be improved, which is beneficial for online training, which has high real-time requirements. In addition, it is also beneficial to improve the timeliness of predicting the level of the analyte.

[0009] In addition, in the detection method involved in the first aspect of the present disclosure, optionally, the multiple models include at least a first model with a settable order and which sorts the corresponding first data to find the regularity of the analyte level in the first data, and the order of the first model is set to the minimum value of the settable order. In this case, the first model sorts the first data to find the regularity, which can reduce the requirements for the distribution regularity of the first data, and is conducive to predicting the analyte level, which may be affected by some factors.

[0010] In addition, in the detection method involved in the first aspect of the present disclosure, optionally, the multiple models include at least a second model with a settable kernel function and a penalty item coefficient, the kernel function of the second model is set to the kernel function with the lowest time complexity among the settable kernel functions, and the penalty item coefficient of the second model is set to the minimum value of the settable penalty item coefficient.

[0011] In addition, in the detection method involved in the first aspect of the present disclosure, optionally, in the training, the input of the second model is each of a plurality of different feature vectors generated by dividing the corresponding first data, each of the feature vectors including a plurality of analyte levels. In this case, the plurality of different feature vectors can enable the model to learn the fluctuation characteristics of multiple groups of analyte levels, thereby improving the accuracy of the predicted estimated value.

[0012] In addition, in the detection method involved in the first aspect of the present disclosure, optionally, the multiple models include at least a third model based on a neural network and / or a fourth model based on autoregression, and the weight of the input layer and the bias of the hidden layer of the third model adopt fixed values. In this case, the weight of the input layer and the bias of the hidden layer of the third model are determined, which can make the third model have the advantages of fewer training parameters and faster learning speed. In addition, the fixed value can make the estimated value repeatable, which is conducive to comparing the performance of multiple trainings, and can also be conducive to analyzing the reasons for the prediction deviation when the model is applied.

[0013] In addition, in the detection method involved in the first aspect of the present disclosure, optionally, in determining the second estimated value, the first estimated value is adjusted based on the result of comparing the analyte fluctuation information with the preset fluctuation to obtain the second estimated value, the analyte fluctuation information is a rising rate and a falling rate, and in the adjustment, the rising rate and the falling rate are limited to be within the range represented by the preset fluctuation, or the analyte fluctuation information is a rising amplitude and a falling amplitude, and in the adjustment, the rising amplitude and the falling amplitude are limited to be within the range represented by the preset fluctuation. In this case, the constrained second estimated value can be simplified while taking into account the change of the analyte level over time.

[0014] In addition, in the detection method involved in the first aspect of the present disclosure, optionally, the combination is a weighted average, and the weight of the weighted average is a known value before the analyte sensor is used. In this case, when determining the target estimate value in combination with the analyte fluctuation information, the prediction bias of the model can be suppressed while simplifying the combination process. In addition, determining the target estimate value in combination with the analyte fluctuation information suppresses the prediction bias of the model and can also reduce the weight requirement for the weighted average.

[0015] In addition, in the detection method involved in the first aspect of the present disclosure, optionally, the analyte level is glucose concentration, and the analyte event is a hyperglycemic event and / or a hypoglycemic event. Thus, the accuracy of detecting hyperglycemic events and / or hypoglycemic events can be improved.

[0016] In addition, in the detection method involved in the first aspect of the present disclosure, optionally, the preset fluctuation is determined by a baseline of a predetermined object related to the target object.

[0017] A second aspect of the present disclosure provides an electronic device, the electronic device comprising a processor and a memory, the processor executing a program stored in the memory to implement the detection method involved in the first aspect of the present disclosure.

[0018] A third aspect of the present disclosure provides a computer-readable storage medium, wherein the computer-readable storage medium stores at least one instruction, and when the at least one instruction is executed by a processor, the detection method involved in the first aspect of the present disclosure is implemented.

[0019] According to the present disclosure, a method, device and medium for detecting analyte events based on machine learning are provided, which can improve the accuracy of detecting analyte events. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The present disclosure will now be explained in further detail, by way of example only, with reference to the accompanying drawings.

[0021] Figure 1 is a schematic diagram showing a detection environment of a glucose event involved in an example of the present disclosure.

[0022] Figure 2 is an exemplary flow chart showing a detection method involved in the examples of the present disclosure.

[0023] Figure 3 Detailed description is given of an exemplary flowchart of determining a second estimated value based on a first estimated value and first fluctuation information according to an example of the present disclosure.

[0024] Figure 4 is an exemplary flow chart showing the determination of a combined estimate based on multiple models involved in the examples of the present disclosure.

[0025] Figure 5 is an exemplary schematic diagram showing a third model involved in the example of the present disclosure.

[0026] Fig. 6A is a first comparative schematic diagram showing three glucose concentrations involved in the examples of the present disclosure; Figure 6B This is a diagram showing the examples of the present disclosure. Fig. 6A A partial enlarged view of the middle area A1; Fig. 7A is a second comparative schematic diagram showing three glucose concentrations involved in the examples of the present disclosure; Figure 7B This is a diagram showing the examples of the present disclosure. Fig. 7A A partial enlarged view of the middle area A2; Fig. 8A is a third comparative schematic diagram showing three glucose concentrations involved in the examples of the present disclosure; Figure 8B This is a diagram showing the examples of the present disclosure. Fig. 8A A partial enlarged view of the middle area A3.

[0027] Fig. 9 is an exemplary flowchart showing an online training model involved in the examples of the present disclosure. DETAILED DESCRIPTION

[0028] Hereinafter, the preferred embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. In the following description, the same symbols are assigned to the same components, and repeated descriptions are omitted. In addition, the accompanying drawings are only schematic diagrams, and the ratio of the dimensions of the components or the shapes of the components may be different from the actual ones.

[0029] It should be noted that the terms "including" and "having" and any variations thereof in the present disclosure, such as a process, method, system, product or device that includes or has a series of steps or units, are not necessarily limited to those steps or units clearly listed, but may include or have other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0030] The disclosed example relates to a method for detecting an analyte event based on machine learning, and an analyte event can be detected based on a predicted analyte level. Specifically, the analyte level of a target object for a period of time can be used to predict the analyte level of another period of time, and then the analyte event of another period of time can be detected. In some examples, the above-mentioned period of time can be a time period before and at the current moment, and another period of time can be a time period after the current moment. In this case, it is possible to detect analyte events in future time periods in order to prevent in advance, thereby helping to reduce the occurrence of unfavorable analyte events.

[0031] In addition, the detection method of analyte events based on machine learning involved in the examples of the present disclosure may also be referred to as a detection method, an early warning method, or an identification method, etc. The detection method involved in the examples of the present disclosure can be applied to the scenario of controlling the analyte.

[0032] The detection method involved in the example of the present disclosure, when applied to analyte event detection of glucose concentration (that is, the analyte is glucose), predicts the glucose concentration in future time periods based on the model and glucose data provided by the target object, so as to issue alarm information (such as high and low blood sugar alarms) when necessary, which can help the target objects better manage their glucose levels, prevent symptoms of hypoglycemia and hyperglycemia, and thus improve the quality of life.

[0033] In some examples, the analyte level involved in the examples of the present disclosure can be obtained by an analyte sensor. Specifically, the analyte level can be measured by an analyte sensor worn by the target object.

[0034] The analyte sensor involved in the examples of the present disclosure can continuously measure the analyte level of the target object. In this case, compared with the single-point analyte level (such as finger blood), the continuous multiple analyte levels are conducive to obtaining the change trend of the analyte. In some examples, the analyte sensor can periodically measure and record a signal related to the analyte level (hereinafter referred to as the analyte signal), and the analyte level can be obtained based on the analyte signal.

[0035] For example, the analyte sensor can measure and record the analyte signal every 1 minute, 2 minutes, 3 minutes, 5 minutes, or 10 minutes. In some examples, the analyte sensor can obtain the analyte signal of the target object, the analyte signal can match the actual analyte level of the body fluid of the target object, and the analyte level obtained by processing the analyte signal can be used as the analyte level of the body fluid of the target object. In some examples, the analyte signal can be a signal related to an electrochemical reaction. Preferably, the analyte signal can be a current signal.

[0036] The target object involved in the examples of the present disclosure may be an object with life characteristics. In some examples, the target object may be a patient suffering from diabetes, hypopituitarism, thyroid disease, or obesity.

[0037] In some examples, when the analyte is glucose, the analyte sensor may be a glucose sensor, the analyte level may be a glucose concentration, and the analyte event may be a glucose event. However, the analyte of the example disclosed herein may not be limited to glucose. For example, by changing the analyte sensor, other body fluid component data other than glucose concentration may also be obtained. The body fluid components here may be, for example, acetylcholine, amylase, bilirubin, cholesterol, chorionic gonadotropin, creatine kinase, creatine, creatinine, DNA, fructosamine, glucose, glutamine, growth hormone, hormone, ketone body, lactate, oxygen, peroxide, prostate specific antigen, prothrombin, RNA, thyroid stimulating hormone, troponin, etc.

[0038] For the convenience of description, some examples are described below using the analyte level as glucose concentration as an example. It should be noted that this does not limit the present disclosure. Unless there is a contradiction, the relevant description is also applicable to other analyte levels. In addition, the analyte level measured by the analyte sensor is referred to as a measured value, and the analyte level predicted by the model is referred to as an estimated value.

[0039] Examples of the present disclosure will be described in detail below with reference to the accompanying drawings. Figure 1 is a schematic diagram showing a detection environment of a glucose event involved in an example of the present disclosure.

[0040] In some examples, when the analyte level is glucose concentration, the detection method disclosed herein can be applied to Figure 1 In the test environment shown. Figure 1 In some examples, the detection environment may include an on-body device 900, a receiving device 800, and a detection device 100. The analyte sensor may be at least partially disposed on the on-body device 900, and the on-body device 900 may be disposed on the skin surface of the target object. The on-body device 900 may be connected to the receiving device 800 via a communication link. In some examples, the receiving device 800 may be integrated with the detection device 100 or connected to the detection device 100 via a communication link.

[0041] Continue to refer Figure 1 In some examples, the glucose signal measured by the analyte sensor can be received by the receiving device 800, which can process the received glucose signal to obtain the glucose concentration and send it to the detection device 100, which can receive the glucose concentration and detect the glucose event based on the predicted glucose concentration. That is, the detection device 100 can be used to implement the detection method involved in the examples of the present disclosure.

[0042] In some examples, the glucose signal measured by the analyte sensor may be transmitted to the receiving device 800 via a communication component (eg, a Bluetooth module) of the on-body device 900 .

[0043] In addition, as described above, in some examples, the receiving device 800 may be integrated with the detection device 100. Specifically, the application of the receiving device 800 may be integrated with the application of the detection device 100. Thus, compared with the solution where the detection device 100 is far away from the receiving device 800, the analyte event can be obtained in a relatively real-time manner, which is conducive to improving the timeliness of the alarm information provided to the target object. For example, if the receiving device 800 is a mobile APP (Application) end, the application of the detection device 100 can be deployed on the mobile APP end.

[0044] In other examples, the application of the detection device 100 may also be deployed in a server, and the receiving device 800 may obtain the alarm information provided by the glucose event by requesting the server interface. Thus, the convenience of updating the detection device 100 can be improved.

[0045] Figure 2 is an exemplary flow chart showing a detection method involved in the examples of the present disclosure.

[0046] As described above, the detection method involved in the examples of the present disclosure can detect analyte events based on predicted analyte levels (also referred to as predicted estimated values). In addition, the component for predicting analyte levels can be a model.

[0047] To this end, the present disclosure also provides an exemplary process of a detection method based on machine learning. Figure 2 , the exemplary process may include: obtaining a target estimate of a target object at a target time using a model (step S101), and obtaining an analyte event at the target time based on the target estimate at the target time (step S102).

[0048] refer to Figure 2 In some examples, in step S101 , the target estimate at the target time may be determined by the output of the model.

[0049] In some examples, the number of models may be one or more. Preferably, the number of models may be multiple. In this case, when the estimated values ​​of multiple models are combined to obtain the target estimated value, the advantages of different models can be combined, and the disadvantages of different models can be compensated for each other, thereby further improving the accuracy of the target estimated value. In some examples, the types of multiple models may be different (described later). In this case, the advantages of different types of models can be combined, and the disadvantages of different types of models can be compensated for each other, thereby further improving the accuracy of the target estimated value.

[0050] In some examples, the target estimate may be determined by any one of the models. Specifically, the target estimate may be an estimate associated with the output of any one of the models (e.g., a first estimate or a second estimate described later). In some examples, the target estimate may be determined by combining a plurality of models. Specifically, the target estimate may be determined based on a combination of a plurality of estimates, each estimate being associated with the output of a model. In addition, the target estimate determined by combining a plurality of models may also be referred to as a combined estimate.

[0051] In some examples, the target estimate can be determined based on a first estimate associated with the output of the model and analyte fluctuation information (hereinafter referred to as the first fluctuation information) associated with the first estimate. In this case, the target estimate is determined in combination with the analyte fluctuation information, and the target estimate is constrained at the application level of the model, which is conducive to reducing the prediction bias, and thus can reduce the requirements for the model (such as the number of models or model types). In addition, when multiple models are used, the difficulty of combining multiple estimates corresponding to multiple models can also be reduced (for example, when a fixed weight is used for combination, a more accurate target estimate can also be obtained).

[0052] In some examples, the first estimated value related to the output of the model may be the output of the model. In other words, the first estimated value may be directly output by the model.

[0053] In addition, the analyte fluctuation information may be fluctuation information of multiple analyte levels within a period of time. In this case, the target estimate can be constrained by taking into account the change of the analyte level over time, which is more in line with the analyte fluctuation characteristics (for example, the analyte level generally does not suddenly increase or decrease in a short period of time) compared to constraining the target estimate based on a single analyte level, and can reduce the impact of random deviations of the estimate.

[0054] In addition, the first fluctuation information may be fluctuation information of a plurality of analyte levels including the first estimated values. In some examples, the first fluctuation information may be determined by each first estimated value and target data corresponding to each first estimated value. The target data corresponding to each first estimated value may be data for predicting each first estimated value, such as second data corresponding to each first estimated value (described later).

[0055] In some examples, the fluctuation information may be at least one of waveform characteristics, rate of change, amplitude of change, and variability. Preferably, the fluctuation information may be rate of change or amplitude of change. In this case, the constrained second estimate can be simplified while considering the change of the analyte level over time. In addition, the selection of the fluctuation information and the setting of the corresponding parameters (e.g., preset fluctuation and adjustment step) may be set according to the performance of the model.

[0056] Figure 3 Detailed description is given of an exemplary flowchart of determining a second estimated value based on a first estimated value and first fluctuation information according to an example of the present disclosure.

[0057] As described above, in some examples, the target estimated value may be determined based on the first estimated value and the first fluctuation information. In some examples, the second estimated value may be determined based on the first estimated value and the first fluctuation information, and the target estimated value at the target time may be determined based on the second estimated value.

[0058] In some examples, in determining the second estimated value, the first estimated value may be adjusted based on the first fluctuation information, and the adjusted first estimated value may be used as the second estimated value. In addition, the adjustment may be a process of making the adjusted first estimated value meet the predetermined requirements. It should be noted that if the first estimated value already meets the predetermined requirements, the first estimated value may also be maintained.

[0059] In some examples, the predetermined requirement may be a preset fluctuation. Specifically, the second estimated value may be determined based on comparing the first fluctuation information with the preset fluctuation. In some examples, in determining the second estimated value, the first estimated value may be adjusted based on the result of comparing the analyte fluctuation information with the preset fluctuation to obtain the second estimated value.

[0060] In some examples, during the adjustment, the first fluctuation information may be limited to be within a range represented by a preset fluctuation. To this end, the present disclosure also provides an exemplary process for determining a second estimated value based on a first estimated value and the first fluctuation information. It should be noted that this does not limit the present disclosure. Figure 3 , the exemplary process includes the following steps:

[0061] Step S201: Obtaining a preset fluctuation.

[0062] Step S202: Determine first fluctuation information corresponding to the first estimated value based on the first estimated value and the target data corresponding to the first estimated value. Specifically, a set of data that changes over time can be formed based on the first estimated value and at least a portion of the analyte level in the target data, and fluctuation information of the set of data can be obtained as the first fluctuation information.

[0063] Step S203: Compare the first fluctuation information with the preset fluctuation. In some examples, in step S203, the result of comparing the first fluctuation information with the preset fluctuation can be obtained. In addition, the result can be used to adjust the first estimated value to obtain the second estimated value. In some examples, if the result is that the first fluctuation information is within the range represented by the preset fluctuation (that is, when the result is the first result), step S204 can be executed, otherwise (that is, when the result is the second result), step S205 can be executed.

[0064] Step S204: adjusting the first estimated value based on the preset fluctuation, and using the adjusted first estimated value as the second estimated value. In some examples, during the adjustment, the first estimated value may be modified based on the preset fluctuation until the fluctuation information corresponding to the modified first estimated value is within the range represented by the preset fluctuation.

[0065] In some examples, during the adjustment, the adjustment step size may be increased to the previous value of the first estimate to obtain the second estimate. Thus, the risk of the second estimate deviating too much from the normal analyte variation can be reduced. In addition, the previous value may be the analyte level at the previous moment of the target moment, and the analyte level at the previous moment may be a measured value or an estimated value according to the actual situation.

[0066] In addition, the adjustment step size may be a positive number or a negative number depending on the direction of the adjustment. In some examples, the adjustment step size may be determined based on the fluctuation information of the surrounding area at the reference target time and the fluctuation law of the analyte. Thus, the accuracy of the second estimated value can be improved.

[0067] In some examples, the fluctuation law of the analyte may include that if the analyte level exceeds a predetermined value indicating that the analyte level is too high, the fluctuation of the analyte becomes more and more gentle. In some examples, the fluctuation law of the analyte may also include that the fluctuation of the analyte does not jump in a short period of time. It should be noted that more fluctuation laws of analytes may be added according to the situation to determine the adjustment step size.

[0068] Step S205: The first estimated value is used as the second estimated value. That is, the first estimated value is kept unchanged. As described above, preferably, the fluctuation information can be a rate of change or a magnitude of change. In addition, in some examples, during the adjustment, the first fluctuation information can be limited to be within the range represented by a preset fluctuation. For this purpose, the present disclosure also provides an example of limiting the first fluctuation information when the first fluctuation information is a rate of change or a magnitude of change.

[0069] In some examples, the first fluctuation information may be a change rate, and the change rate may be a rising rate and a falling rate. Specifically, in adjusting the first estimated value, the rising rate and the falling rate may be limited to be within a range represented by a preset fluctuation.

[0070] In some examples, the first fluctuation information may be a change range, and the change range may be an increase range or a decrease range. Specifically, in adjusting the first estimated value, the increase range or the decrease range may be limited to be within a range represented by a preset fluctuation.

[0071] In addition, the preset fluctuation can be determined by a baseline of a predetermined subject related to the target subject. The baseline can be a fluctuation condition that the analyte level of the predetermined subject should be in.

[0072] In some examples, the predetermined object may be the target object, the group to which the target object belongs, and at least one type of object among all target objects. Preferably, the predetermined object may be the target object itself. That is, the baseline may be the target object itself. In this case, the second estimated value may be more closely aligned with the target object itself, which is conducive to improving the accuracy of the second estimated value.

[0073] In some examples, the baseline can be a predetermined fixed value, such as a baseline common to all target subjects. In some examples, the baseline can be obtained by analyzing the analyte level, such as a baseline of a population to which the target subject belongs or a baseline of the target subject.

[0074] In some examples, for the target object's baseline, the target object's historical analyte fluctuations can be analyzed to determine the fluctuations (eg, waveform characteristics or rate of change) that the target object's analyte level should be in and use it as the target object's baseline.

[0075] In some examples, the baseline of the target object can be dynamically updated. For example, the baseline of the target object can be dynamically updated according to changes in the analyte level of the target object. In this case, the baseline can adapt to changes in the physiological condition of the target object, which is conducive to improving the accuracy of the second estimate.

[0076] In some examples, the baseline for the group to which the target object belongs can be determined by counting the analyte levels of the corresponding group. In some examples, the baseline common to all target objects can be that the analyte level should be within the maximum range of the analyte level.

[0077] As described above, the second estimated value may be determined based on the first estimated value and the first fluctuation information. In other examples, the second estimated value may also be determined in combination with the range in which the single first estimated value should be and the first fluctuation information. In other examples, the second estimated value may also be determined based only on the range in which the single first estimated value should be.

[0078] Figure 4 is an exemplary flow chart showing the determination of a combined estimate based on multiple models involved in the examples of the present disclosure.

[0079] As mentioned above, the number of models can be multiple. For this purpose, the example disclosed herein also provides an exemplary process for determining a combined estimated value based on multiple models (that is, the above-mentioned target estimated value is a combined estimated value), see Figure 4 , the exemplary process includes the following steps:

[0080] Step S301: Acquire multiple models corresponding to multiple first time periods. In addition, each of the multiple models may be a machine learning model. Each model may be obtained by training the first data of the first time period of the target object. Specifically, the first data of a first time period may be used to train a model, and multiple models may correspond to multiple first time periods.

[0081] In addition, multiple first time periods may not be exactly the same. For example, multiple first time periods may be selected from 5 minutes, half an hour, and 1 hour. In some examples, the starting endpoints of multiple first time periods may be at least partially different, and the ending endpoints of multiple first time periods may be the same, which is specifically related to the corresponding model. In some examples, the corresponding first time period can be adjusted according to the performance of the model. Taking the performance as the accuracy rate as an example, multiple experiments corresponding to different time periods can be implemented, and the accuracy of the model in the corresponding time period can be analyzed according to the results of the multiple experiments, and then it is determined how long the time period is to be selected (that is, the first time period corresponding to the model is determined).

[0082] As described above, the analyte level can be obtained by the analyte sensor, and therefore, the first data can be analyte data provided by the analyte sensor continuously measuring the analyte level of the target object. In some examples, the first data can be a time series. In this case, the model can learn the characteristics of the analyte level changing over time, thereby improving the accuracy of the predicted estimate.

[0083] In some examples, the time series can be a sequence formed by arranging the values ​​of the analyte level at different times in chronological order. In this case, the dimension of the first data is the smallest, which is conducive to improving the training efficiency and the prediction efficiency when the model is applied. In some examples, the first data may also include other indicators that change with time in addition to the analyte level. In some examples, the first data may also include other indicators that do not change with time within a certain period of time (such as age).

[0084] In some examples, before step S301, initial data of a predetermined time period may be obtained, and multiple first data corresponding to multiple first time periods may be retrieved from the initial data. For example, initial data of the current hour may be input, and multiple first data of a time period corresponding to the model may be selected therefrom.

[0085] In addition, the predetermined time period may be determined by the length of the plurality of first time periods. In some examples, the length of the predetermined time period may be equal to the maximum length of the plurality of first time periods. In this case, when the predetermined time period is the shortest, the initial data may include a plurality of first data of the plurality of first time periods, so that the analyte data measured by the analyte sensor may meet the requirements of the training data of the plurality of models.

[0086] In some examples, when the analyte sensor is initially worn, the training of multiple models may be started after waiting for a predetermined period of time. For example, if the longest length of the multiple first time periods is 1 hour, the training may be started after waiting for 1 hour.

[0087] In some examples, the predetermined time period may be a time period closest to and including the current moment. In this case, the latest first data can be obtained, and when predicting the estimated value of the target moment after the current moment, the influence of the uncertainty factors before the target moment on the estimated value of the target moment can be reduced, thereby improving the accuracy of the estimated value.

[0088] In other examples, the stored analyte levels may be directly retrieved to obtain a plurality of first data without first obtaining the initial data.

[0089] Step S302: Acquire second data of a second time period corresponding to the first time period. Specifically, one second time period may correspond to one first time period, and multiple second time periods may correspond to multiple first time periods.

[0090] In some examples, the second time period may correspond to the first time period through a model training process. For example, when training a corresponding model, if the time period corresponding to each input model data is 30 minutes, then when the model is applied, the analyte level for 30 minutes may be directly obtained as the input of the model or the analyte level for 30 minutes may be processed (e.g., formed into a feature vector) corresponding to the model training process as the input of the model.

[0091] In addition, at least a portion of the first time period is before the corresponding second time period. Thus, the model trained by the historical analyte data can be combined to predict the subsequent estimated value.

[0092] In addition, the second data may be data corresponding to the first data when the model is applied, and the relevant content will not be repeated. It should be noted that the analyte level in the second data may include only measured values, both measured values ​​and estimated values, or only estimated values, which is specifically related to the second time period corresponding to the estimated value of the predicted target moment. In some examples, if there is a measured value at the corresponding moment of the second time period and the measured value is a normal value, the measured value can be used as the analyte level corresponding to the moment, otherwise the model is used to obtain the estimated value of the moment and used as the analyte level corresponding to the moment. Thus, the risk of random deviation of the measured value can be reduced.

[0093] Step S303: Process the corresponding second data through each model of the multiple models to predict multiple first estimated values ​​at the target moment after multiple second time periods. Specifically, for a single first time period, a model can be obtained by training the first data of the first time period, and the model can be used to process the second data of the second time period corresponding to the first time period to predict the first estimated value of the target moment after the second time period. For example, multiple second data can be input into multiple models respectively to obtain multiple first estimated values ​​within a certain period of time in the future.

[0094] In some examples, if there is a moment without a measurement value in the second time period corresponding to the target moment, a first estimated value of at least one moment before the target moment can be predicted first, and then a first estimated value of the target moment can be obtained based on the first estimated value of at least one moment.

[0095] In some examples, when obtaining a first estimate at a moment, the first estimate can be used as part of the input for predicting a first estimate at the next moment until the first estimate at the target moment is obtained. Thus, multiple first estimates within a period of time after the second time period (hereinafter referred to as the target time range) can be obtained. For example, the target time range can be 1 minute, 5 minutes, 10 minutes, 20 minutes, 30 minutes, 40 minutes, 50 minutes, or 1 hour, etc.

[0096] Specifically, the second data of the second time period is substituted into the model to obtain the first estimated value at the next moment of the second time period, and then the first estimated value at the next moment is substituted into the model to obtain the first estimated value at the next moment, and finally the first estimated values ​​at multiple moments after the second time period, such as 1 minute, 2 minutes, 3 minutes, etc. are obtained. That is, the estimated value at the next moment is predicted and then substituted into the model to obtain the estimated value at the next moment.

[0097] For example, for the next 10 minutes, 10 records need to be predicted, and for the next 30 minutes, 30 records need to be predicted. It should be noted that the above examples do not limit the intervals between moments of the present disclosure, and the intervals between moments are related to the intervals between analyte levels in the first data used for training.

[0098] In some examples, the target time range is adjustable. For example, the target object can adjust the target time range according to the control requirements of the analyte. That is, the target object can set how long in the future the analyte level needs to be predicted. In some examples, the target time range can have a default value, for example, for glucose concentration, the target time range can default to predicting the glucose concentration in the next 30 minutes.

[0099] Step S304: Determine multiple second estimated values ​​based on multiple first estimated values. In some examples, in step S304, a second estimated value corresponding to each first estimated value of at least a portion of the multiple first estimated values ​​may be determined based on each first estimated value of at least a portion of the multiple first estimated values ​​and the corresponding first fluctuation information. That is, at least a portion of the multiple first estimated values ​​needs to be adjusted in combination with the corresponding first fluctuation information.

[0100] In some examples, the performance of the model can be analyzed to determine whether it is necessary to adjust in combination with the corresponding first fluctuation information. In some examples, the performance of the model can be analyzed before the analyte sensor is used to determine whether to start the adjustment process. Thus, the consumption of hardware resources can be reduced. In some examples, for models whose performance stability does not fall within a predetermined stable range (i.e., the performance is not stable enough), the adjustment process can be started when the model is applied.

[0101] For relevant content about determining the second estimated value corresponding to each first estimated value based on each first estimated value and the corresponding first fluctuation information, please refer to the above description of determining the second estimated value based on the first estimated value and the first fluctuation information, wherein the target data is the second data.

[0102] In some examples, in step S304, each first estimated value except at least a portion of the plurality of first estimated values ​​may be directly used as the second estimated value.

[0103] Step S305: Determine a combined estimated value based on a combination of multiple second estimated values. In some examples, the combination can be a weighted average. For example, in the case of four models, four second estimated values ​​can be obtained, and weights are assigned to the four second estimated values ​​respectively and then summed to obtain a combined estimated value. In addition, the weight of the weighted average can be dynamic or fixed. Preferably, when determining the target estimated value in combination with the analyte fluctuation information, the weight of the weighted average can be a known value before the analyte sensor is used. Thereby, the prediction bias of the model can be suppressed and the combination process can be simplified. In addition, determining the target estimated value in combination with the analyte fluctuation information suppresses the prediction bias of the model and can also reduce the weight requirement for the weighted average.

[0104] Return to reference Figure 2 In some examples, in step S102, the analyte event at the target moment may be obtained by comparing the target estimate value at the target moment with a preset analyte level.

[0105] In addition, the analyte event can be an event indicating that the analyte level is too high or too low. In some examples, the analyte event can be a high analyte event and / or a low analyte event. For example, if the analyte is glucose, the analyte event can be a hyperglycemic event and / or a hypoglycemic event. Thus, the accuracy of detecting hyperglycemic events and / or hypoglycemic events can be improved.

[0106] In addition, the preset analyte level can be related to the conditions for generating the analyte event. In some examples, the preset analyte level can be related to the group to which the target object belongs. For example, for glucose concentration, a normal person's fasting glucose concentration below 2.8mmol / L (millimoles per liter) can be considered to have a hypoglycemic event, and a diabetic patient's glucose concentration not greater than 3.9mmol / L can be considered to have a hypoglycemic event. In other examples, the preset analyte level can be adjusted by the target object. In some examples, the preset analyte level can include a plurality of sub-preset analyte levels corresponding to different degrees of urgency. Thus, the analyte events can be graded according to the degree of urgency.

[0107] In some examples, in step S102 , a notification (eg, alarm information) associated with the analyte event may also be generated to prompt the target subject.

[0108] In some other examples, step S102 is not necessary, that is, the target estimate value may not be used to obtain the analyte event at the target time, and the target estimate value may be used for any purpose related to analyte analysis.

[0109] As described above, in some examples, the number of models may be one or more. In some examples, the parameters of at least a portion of the multiple models are parameters with the smallest time complexity. In this case, the training efficiency of the model and the prediction efficiency during application can be improved, which is beneficial for online training, a scenario with high real-time requirements. In addition, it is also beneficial to improve the timeliness of predicting the level of the analyte.

[0110] Preferably, when determining the target estimated value in combination with the analyte fluctuation information, the parameters of at least a portion of the multiple models may be parameters with the minimum time complexity. In this case, the training efficiency of the model and the prediction efficiency during application can be improved while suppressing the prediction deviation of the model. In addition, determining the target estimated value in combination with the analyte fluctuation information, suppressing the prediction deviation of the model, and reducing the time complexity requirement for the model can also be achieved.

[0111] In addition, the examples of the present disclosure also provide some examples of models. In some examples, the multiple models may include at least one of the first model, the second model, the third model 110 and the fourth model. In addition, as described above, the types of the multiple models may be different.

[0112] For the convenience of description, unless otherwise specified, when describing the first model, the second model, the third model 110 and the fourth model below, the first data may be data of one analyte level (i.e., one analyte level) per minute, which does not represent a limitation of the present disclosure.

[0113] In some examples, the first model may be a model that sorts the first data to find the regularity of the analyte level in the first data. In this case, the first model sorts the first data to find the regularity, which can reduce the requirements for the distribution regularity of the first data, and is conducive to predicting the analyte level, which may be affected by some factors (such as the living habits, behavior or environment of the target object).

[0114] In some examples, the first model may be set to an order. In some examples, when the parameter with the smallest time complexity is used, the order of the first model may be set to the minimum value of the settable order of the first model. In addition, when the interval between the analyte levels in the first data is less than a preset interval, the accuracy of the predicted estimated value can be further improved, wherein the preset interval may be less than 5 minutes. Preferably, the preset interval may be 1 minute.

[0115] In some examples, the first model may be a model with the smallest order in the gray prediction model. Specifically, the first model may be a GM (1, 1) model, that is, the differential equation is of 1st order and there is 1 variable in the differential equation. In some examples, when the first model is a GM (1, 1) model, the first time period may be half an hour.

[0116] In some examples, the first data may be first subjected to a first-order accumulation and then used as the input of the GM (1, 1) model, and the output of the GM (1, 1) model may be an estimated value. In addition, the first data used for the first-order accumulation of the GM (1, 1) model may be one-dimensional.

[0117] Specifically, the data sequence in the first data can be logarithmically transformed and first-order accumulated to generate a new data sequence, the new data sequence can be used to construct a first-order whitened differential equation, the least squares method can be used to calculate the development coefficient and gray usage in the whitened differential equation to determine the formula representing the estimated value corresponding to the new data sequence, and the formula representing the estimated value corresponding to the new data sequence can be inversely logarithmically transformed to obtain the prediction formula.

[0118] During training, the prediction formula may be used to obtain multiple estimated values ​​corresponding to the first data. During application, the prediction formula may be used to obtain multiple estimated values ​​corresponding to the second data.

[0119] In addition, for the GM(1,1) model, a metabolic model can be used to obtain the corresponding estimated value. Specifically, the new analyte levels closest to the target time can be continuously added, the historical analyte levels farthest from the target time can be removed, and the GM(1,1) model can be retrained to predict the estimated value at the target time.

[0120] As described above, at least a portion of the model can determine the first estimated value in combination with the analyte fluctuation information. The inventors have found through research that the first model can reduce the risk of excessive prediction deviation by constraining the first estimated value through the rate of change. In some examples, for the first model, the first fluctuation information can be a rising rate and a falling rate, and the preset fluctuation can include a preset falling rate and a preset rising rate. The first estimated value is constrained by the preset fluctuation to limit the first fluctuation information to be within the range represented by the preset fluctuation.

[0121] Taking the GM (1, 1) model as an example, the preset decreasing rate can be 0.5, the preset increasing rate can be 0.5, and the adjustment step can be ±0.05. When the decreasing rate is greater than 0.5, the first estimated value can be adjusted to the previous value plus -0.05. When the increasing rate is greater than 0.5, the first estimated value can be adjusted to the previous value plus 0.05.

[0122] In some examples, the second model may be a model that maps the first data into a high-dimensional space so that the first data is more easily separated in the high-dimensional space. In some examples, there may be multiple parameters in the second model that need to be set. In some examples, the multiple parameters of the second model may include a kernel function and a penalty term coefficient. That is, the second model may set a kernel function and a penalty term coefficient.

[0123] In some examples, when the parameters with the lowest time complexity are used, the kernel function of the second model is set to the kernel function with the lowest time complexity among the settable kernel functions, and the penalty item coefficient of the second model is set to the minimum value of the settable penalty item coefficient. In some examples, the kernel function with the lowest time complexity may be a linear kernel function, and the minimum value of the settable penalty item coefficient may be 1.

[0124] In some examples, the training data of the second model can be a plurality of different feature vectors generated by partitioning the first data. Each feature vector has a plurality of analyte levels. In this case, the plurality of different feature vectors can enable the model to learn the fluctuation characteristics of multiple groups of analyte levels, thereby improving the accuracy of the predicted estimated value. In addition, each feature vector can be used as an input to the second model, and the output of the second model can be an estimated value. In addition, the first data used by the second model to partition the feature vectors can be one-dimensional.

[0125] In some examples, the first data of the first time period can be divided into a preset number of feature vectors, each feature vector including the preset number of analyte levels. In some examples, the preset number can be half of the amount of data in the first data. In this case, the model can learn multiple fluctuation features covering half of the data, which is more in line with the fluctuation characteristics of the analyte, and there is overlap between the feature vectors, thereby increasing the correlation between the fluctuation features and further fitting the fluctuation characteristics of the analyte.

[0126] In some examples, the second model may be support vector regression (SVR). In addition, when using parameters with the minimum time complexity, the kernel function of the support vector regression may be a linear kernel function, and the penalty term coefficient may be 1.

[0127] In some examples, for the second model being support vector regression, the first time period may be 1 hour (i.e., 60 minutes). For example, the second data having 60 analyte levels in the first hour may be selected and divided into 30 feature vectors, each feature vector having 30 values, wherein the first feature vector may include the 1st to 30th analyte levels in the second data, the second feature vector may include the 2nd to 31st analyte levels in the second data, and so on, until the last feature vector is obtained.

[0128] As described above, for at least a portion of the models, the first estimated value can be determined in combination with the analyte fluctuation information. The inventors have found through research that the second model can reduce the risk of excessive prediction deviation by constraining the first estimated value through the amplitude of change. In some examples, for the second model, the first fluctuation information can be the amplitude of change, and the preset fluctuation can include a preset decrease amplitude and a preset increase amplitude. The first estimated value is constrained by the preset fluctuation to limit the first fluctuation information to be within the range represented by the preset fluctuation.

[0129] Taking support vector regression as an example, the preset decrease can be 2, the preset increase can be 0.2, and the adjustment step can be 0.1 and -1. When the increase relative to the previous value is greater than 0.2, it means that the growth rate is too fast, and the first estimated value is adjusted to the previous value plus 0.1. If the decrease relative to the previous value is greater than 2, it means that the decrease is too fast, and the first estimated value is adjusted to the previous value plus -1.

[0130] Figure 5 is an exemplary schematic diagram showing a third model 110 involved in the example of the present disclosure.

[0131] In some examples, the third model 110 may be a model based on a neural network. In some examples, the third model 110 may be a model built based on a feedforward neural network. In addition, the input of the third model 110 may be an index of the analyte level in the first data, and the output of the third model 110 may be an estimated value.

[0132] refer to Figure 5 In some examples, the third model 110 may include an input layer 111 , a hidden layer 112 , and an output layer 113 .

[0133] In addition, the input layer 111 may include a plurality of input nodes. The inputs of the plurality of input nodes may be indexes of analyte levels. Figure 5 , the multiple input nodes may include node I1, ..., node Ii, ..., node ID, wherein node Ii represents the i-th node among the multiple input nodes, node ID represents the D-th node among the multiple input nodes, and D represents the data amount of the first data.

[0134] In addition, the hidden layer 112 may include a plurality of hidden nodes. Figure 5 , the multiple hidden nodes may include nodes H1, ..., nodes Hj, ..., nodes HL, where node Hj represents the jth node among the multiple hidden nodes, node HL represents the Lth node among the multiple hidden nodes, and L represents the number of nodes in the hidden layer 112. In some examples, the number of nodes in the hidden layer 112 may be greater than the number of nodes in the input layer 111.

[0135] In addition, the output layer 113 may include multiple output nodes. The outputs of the multiple output nodes may be estimated values. Figure 5 , the multiple output nodes may include nodes O1, ..., nodes Ok, ..., nodes OM, where node Ok represents the kth node among the multiple output nodes, node OM represents the Mth node among the multiple output nodes, and M represents the number of output nodes. In some examples, the number of nodes in the output layer 113 may be greater than the number of nodes in the hidden layer 112.

[0136] In some examples, the third model 110 may select the weights of the input layer 111 and the bias of the hidden layer 112, and the learning process calculates the weights of the output layer 113. Thus, the third model 110 has the advantages of fewer training parameters and faster learning speed. As an example, Figure 5 The straight line L1 represents the unknown quantity between the node I1 and the node H1. Specifically, the unknown quantity may be the weight of the node I1 and the bias of the node H1. Figure 5 The middle straight line L2 represents the unknown quantity between the node H1 and the node O1 , and the unknown quantity may be the weight of the node O1 .

[0137] In some examples, the weight of the input layer 111 and the bias of the hidden layer 112 may be random or fixed. That is, the weight of the output layer 113 is an unknown quantity. Preferably, the weight of the input layer 111 and the bias of the hidden layer 112 may adopt a fixed value. In this case, the estimated value can be made repeatable, that is, the third model 110 trained with the same first data can obtain the same estimated value, which is conducive to comparing the performance of multiple trainings and can also be conducive to analyzing the cause of prediction deviation when the model is applied.

[0138] In some examples, the weight of the input layer 111 and the bias of the hidden layer 112 can be an arithmetic progression. In this case, the weight of the input layer 111 and the bias of the hidden layer 112 are also fixed while the data flowing between the nodes can be evenly distributed. That is, the data can be evenly distributed and the weight of the input layer 111 and the bias of the hidden layer 112 can be fixed. In addition, by using an arithmetic progression, the risk of excessive prediction deviation can be reduced, and the stability of the prediction of the third model 110 can be improved.

[0139] In some examples, the third model 110 may be an extreme learning machine (ELM). In some examples, for the extreme learning machine, the first time period may be half an hour. To this end, the present disclosure also provides some parameter settings using the extreme learning machine as an example, which should be noted that does not represent a limitation of the present disclosure. Specifically, for the extreme learning machine, half an hour may correspond to 30 analyte levels, the number of nodes in the hidden layer 112 may be 40, and the weight of the input layer 111 may be represented by an arithmetic progression of {-1, -0.95, -0.9, -0.85, ..., 0.95}, and the bias of the hidden layer 112 may be represented by an arithmetic progression of {-0.1, -0.093, ..., 0.093}.

[0140] As described above, the weight of the output layer 113 is an unknown quantity. In some examples, in order to prevent the occurrence of a singular matrix, the formula of the weight of the output layer 113 may be improved. Specifically, the weight of the output layer 113 may satisfy the formula:

[0141] β=(H T H+λE) -1 H T T,

[0142] Where β represents the weight of the output layer 113, λE is used to prevent the appearance of a singular matrix, E represents the identity matrix, H represents the output matrix of the hidden layer 112, and H T represents the transpose of H, and T represents the target matrix of the first data (ie, the expected output). In some examples, λ may be 0.01.

[0143] As described above, for at least a portion of the models, the first estimate may be determined in combination with the analyte fluctuation information. The inventors have found through analysis that when using an arithmetic progression, the stability of the third model 110 is within a predetermined stability range, and the first estimate may not be constrained by the analyte fluctuation information.

[0144] In some examples, the fourth model may be a fourth model based on autoregression. In some examples, the fourth model may be an AR model (Autoregressive model). When processing training data for an AR model, the first data may be subjected to standardized preprocessing and used as input to the fourth model. In this case, the input of the fourth model may present an assimilation trend, thereby improving the accuracy of the estimated value. In some examples, the standardized preprocessing may satisfy the formula:

[0145]

[0146] Wherein, X is the first data before normalization, and Y is the data after normalization.

[0147] In some examples, the order of the AR model may be 1. That is, the simplest first-order AR model is used. In some examples, for the AR model, the first time period may be 5 minutes.

[0148] In some examples, the AR model can satisfy the formula: t =b1x t-1 +b0,

[0149] Among them, x t-1 is the analyte level at time t-1, x t is the analyte level at the next moment after the t-1th moment, b1 is the regression coefficient, and b0 is white noise. b1 is obtained through training.

[0150] As described above, for at least a portion of the models, the first estimated value can be determined in combination with the analyte fluctuation information. After research, the inventors found that the fourth model constrains the first estimated value by the rate of change, which can reduce the risk of excessive prediction deviation. In some examples, for the fourth model, the first fluctuation information can be a rising rate and a falling rate, and the preset fluctuation can include a preset falling rate and a preset rising rate. The first estimated value is constrained by the preset fluctuation to limit the first fluctuation information to be within the range represented by the preset fluctuation.

[0151] Taking the AR model as an example, the preset decline rate can be 0.1, the preset rise rate can be 0.1, and the adjustment step size can be 9 / 10 of the previous differential value. When the decline rate is greater than 0.1, the first estimated value is adjusted to the previous value plus 9 / 10 of the previous differential value. If the rise rate is greater than 0.1, the first estimated value is adjusted to the previous value plus 9 / 10 of the previous differential value. In addition, if the first estimated value is greater than a predetermined value indicating that the analyte level is too high (for example, the first estimated value is greater than 10), the first estimated value can be adjusted to the previous value plus the previous differential value × 5 / first estimated value.

[0152] In addition, the previous differential value may be the difference between the values ​​at two moments before the target moment. For example, if the first estimated value at the 9th minute is predicted, the previous differential value may be the analyte level at the 8th minute minus the analyte level at the 7th minute.

[0153] As described above, in some examples, the plurality of models may include at least one of the first model, the second model, the third model 110 and the fourth model. To this end, the present disclosure also provides examples of the plurality of models.

[0154] In some examples, the multiple models may include at least a first model. In some examples, the multiple models may include at least a first model and a second model. In some examples, the multiple models may include at least a first model, a second model, and a third model 110. In some examples, the multiple models may include at least a first model, a second model, and a fourth model. In some examples, the multiple models may include at least a first model, a second model, a third model 110, and a fourth model.

[0155] Preferably, the multiple models may include a first model, a second model, a third model 110 and a fourth model. The inventors have found through research that these models cooperate with each other to further improve the accuracy of the combined estimation value.

[0156] As described above, the weight of the weighted average can be a fixed value. To this end, the example of the present disclosure also provides an example of weighted average of the first model, the second model, the third model 110 and the fourth model, wherein the first model and the fourth model have the highest weights, and the second model and the third model 110 have the lowest weights. In some examples, the combined estimate may satisfy the formula:

[0157] finalResult=M1×0.3+M4×0.3+M2×0.2+M3×0.2.

[0158] Among them, finalResult represents the combined estimation value, M1 represents the first model, M2 represents the second model, M3 represents the third model 110, and M4 represents the fourth model.

[0159] Fig. 6A is a first comparative schematic diagram showing three glucose concentrations involved in the examples of the present disclosure, Figure 6B This is a diagram showing the examples of the present disclosure. Fig. 6A A partial enlarged view of the middle area A1. Fig. 7A is a second comparative schematic diagram showing three glucose concentrations involved in the examples of the present disclosure, Figure 7B This is a diagram showing the examples of the present disclosure. Fig. 7A A partial enlarged view of the middle area A2. Fig. 8A is a third comparative schematic diagram showing three glucose concentrations involved in the examples of the present disclosure, Figure 8B This is a diagram showing the examples of the present disclosure. Fig. 8A A partial enlarged view of the middle area A3.

[0160] As described above, the inventors have found through research that these models cooperate with each other to further improve the accuracy of the combined estimated value. In some examples, the curves of the combined estimated values ​​corresponding to the four models are almost close to the analyte curve. Taking glucose as an example, the first model selects the GM model, the second model selects the SVR model, the third model 110 selects the AR model, and the fourth model selects the EML model. The initial data input is the historical data of the current hour, and the historical data is one minute per transaction. The combined estimated value within 30 minutes is predicted. An example of the main settings of the model is shown in Table 1 below.

[0161] Table 1 Example of the main settings of the model

[0162] refer to Fig. 6A , Figure 6B , Fig. 7A , Figure 7B , Fig. 8A and Figure 8B Among them, the three types of glucose concentrations are measured values, finger blood values ​​and combined estimated values. The circle represents the measured value, the multiplication sign represents the finger blood value, and the inverted triangle represents the combined estimated value. It can be seen from the three comparative schematic diagrams that the distribution of the combined estimated value is close to the distribution of the measured value.

[0163] Fig. 9is an exemplary flow chart showing the online training model involved in the examples of the present disclosure. In some examples, as the data collected by the analyte sensor is continuously updated, at least one model may also be continuously trained and updated online. In some examples, at least one model involved in the examples of the present disclosure may be updated in response to receiving analyte data that satisfies the training conditions. In some examples, the training condition may be receiving the analyte level newly measured by the analyte sensor. That is, in response to receiving the analyte level newly measured by the analyte sensor, the model may be trained and updated in combination with the newly measured analyte level, and the updated model may be used to predict the first estimate.

[0164] In addition, online training can be to update the model during the application of the model (also referred to as in use). To this end, the example disclosed in this disclosure also provides an exemplary process of online training, refer to Fig. 9 , the exemplary process includes:

[0165] Step S401: Acquire multiple measurement values ​​before the current moment as initial data.

[0166] Step S402: According to the first time period of each model, corresponding measurement values ​​are retrieved from the initial data to form the first data of each model. In some examples, the measurement values ​​of the latest time period are retrieved.

[0167] Step S403: Train each model using the first data of each model.

[0168] Step S404: Use the model to obtain a target estimated value of the target object at a target time. The target time may be after the current time. For details, see the relevant description in step S101.

[0169] Step S405: Determine whether analyte data satisfying the training condition is received.

[0170] If analyte data satisfying the training condition is received, step S406 is executed, otherwise, step S405 is continued to be executed.

[0171] Step S406: Update the initial data using the received analyte data. After executing step S406, step S402 may be executed.

[0172] The example of the present disclosure also discloses an electronic device, including a processor and a memory. The processor executes a program stored in the memory to implement one or more steps in the above detection method.

[0173] The examples of the present disclosure also disclose a computer-readable storage medium, which can store at least one instruction, and when the at least one instruction is executed by a processor, one or more steps in the above-mentioned detection method are implemented.

[0174] Although the present disclosure is specifically described above in conjunction with the accompanying drawings and examples, it is to be understood that the above description does not limit the present disclosure in any form. Those skilled in the art may modify and change the present disclosure as needed without departing from the essential spirit and scope of the present disclosure, and these modifications and changes all fall within the scope of the present disclosure.

Claims

1. A method for detecting analyte events based on machine learning, characterized in that: include: Acquire a plurality of models, each of the plurality of models being a machine learning model trained by first data in a first time period, the first data being a time series provided by an analyte sensor continuously measuring an analyte level of a target object; Acquire second data of a second time period corresponding to the first time period, wherein at least a portion of the first time period is before the corresponding second time period; process the corresponding second data through each of the multiple models to predict multiple first estimated values ​​at target moments after the multiple second time periods; determine multiple second estimated values ​​based on the multiple first estimated values, wherein, based on each first estimated value of at least a portion of the multiple first estimated values ​​and the second data corresponding to each first estimated value, determine analyte fluctuation information corresponding to each first estimated value, compare the analyte fluctuation information with a preset fluctuation, and determine the second estimated value corresponding to each first estimated value based on comparing the analyte fluctuation information with the preset fluctuation; determine a combined estimated value based on a combination of the multiple second estimated values; and acquire the analyte event at the target moment based on the combined estimated value.

2. The detection method according to claim 1, characterized in that: Parameters of at least a portion of the multiple models are parameters with the smallest time complexity.

3. The detection method according to claim 1, characterized in that: The multiple models include at least a first model with a configurable order and which is obtained by sorting the corresponding first data to find the regularity of the analyte level in the first data, and the order of the first model is set to the minimum value of the configurable order.

4. The detection method according to claim 3, characterized in that: The multiple models include at least a second model with a settable kernel function and a penalty coefficient, the kernel function of the second model is set to the kernel function with the lowest time complexity among the settable kernel functions, and the penalty coefficient of the second model is set to the minimum value of the settable penalty coefficient.

5. The detection method according to claim 4, characterized in that: In the training, the input of the second model is each of a plurality of different feature vectors generated by partitioning the corresponding first data, each of the feature vectors including a plurality of the analyte levels.

6. The detection method according to claim 4 or 5, characterized in that: The multiple models include at least a third model based on a neural network and / or a fourth model based on autoregression, and the weights of the input layer and the bias of the hidden layer of the third model adopt fixed values.

7. The detection method according to claim 1, characterized in that: In determining the second estimate, adjusting the first estimate based on a result of comparing the analyte fluctuation information with the preset fluctuation to obtain the second estimate, The analyte fluctuation information is a rising rate and a falling rate, and in the adjustment, the rising rate and the falling rate are limited to be within the range of the preset fluctuation representation, or The analyte fluctuation information is an increase amplitude and a decrease amplitude. In the adjustment, the increase amplitude and the decrease amplitude are limited to be within the range represented by the preset fluctuation.

8. The detection method according to claim 1, characterized in that: The combination is a weighted average, and the weight of the weighted average is a known value before the analyte sensor is used.

9. The detection method according to claim 1, characterized in that: The analyte level is glucose concentration, and the analyte event is a hyperglycemic event and / or a hypoglycemic event.

10. The detection method according to claim 1, characterized in that: The preset fluctuation is determined by a baseline of a predetermined object related to the target object.

11. An electronic device, characterized in that: The electronic device comprises a processor and a memory, and the processor executes a program stored in the memory to implement the detection method according to any one of claims 1 to 10.

12. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one instruction, and when the at least one instruction is executed by a processor, the detection method according to any one of claims 1 to 10 is implemented.

Images