WO2023115751A1 - 血糖预测方法和装置、监测血糖水平的系统 - Google Patents

血糖预测方法和装置、监测血糖水平的系统 Download PDF

Info

Publication number
WO2023115751A1
WO2023115751A1 PCT/CN2022/085450 CN2022085450W WO2023115751A1 WO 2023115751 A1 WO2023115751 A1 WO 2023115751A1 CN 2022085450 W CN2022085450 W CN 2022085450W WO 2023115751 A1 WO2023115751 A1 WO 2023115751A1
Authority
WO
WIPO (PCT)
Prior art keywords
blood glucose
user
prediction
blood
data
Prior art date
Application number
PCT/CN2022/085450
Other languages
English (en)
French (fr)
Inventor
韩洋
蒋娟
Original Assignee
苏州百孝医疗科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 苏州百孝医疗科技有限公司 filed Critical 苏州百孝医疗科技有限公司
Publication of WO2023115751A1 publication Critical patent/WO2023115751A1/zh

Links

Images

Classifications

    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00Measuring for diagnostic purposes; Identification of persons
    • A61B5/145Measuring characteristics of blood in vivo, e.g. gas concentration, pH value; Measuring characteristics of body fluids or tissues, e.g. interstitial fluid, cerebral tissue
    • A61B5/14532Measuring characteristics of blood in vivo, e.g. gas concentration, pH value; Measuring characteristics of body fluids or tissues, e.g. interstitial fluid, cerebral tissue for measuring glucose, e.g. by tissue impedance measurement
    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00Measuring for diagnostic purposes; Identification of persons
    • A61B5/72Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235Details of waveform analysis
    • A61B5/7264Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • A61B5/7267Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00Measuring for diagnostic purposes; Identification of persons
    • A61B5/72Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7271Specific aspects of physiological measurement analysis
    • A61B5/7275Determining trends in physiological measurement data; Predicting development of a medical condition based on physiological measurements, e.g. determining a risk factor
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00Machine learning
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00Machine learning
    • G06N20/10Machine learning using kernel methods, e.g. support vector machines [SVM]
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • GPHYSICS
    • G16INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16HHEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/30ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for calculating health indices; for individual health risk assessment

Definitions

  • the present application relates to the field of physiological signal processing, for example, to a blood sugar prediction method and device, and a system for monitoring blood sugar levels.
  • Diabetes is a metabolic disorder that affects a large portion of the world's population. Prompt and correct diagnosis and treatment of diabetes is essential for a patient to lead a relatively healthy lifestyle. In general, the treatment of diabetes needs to depend on the glucose concentration in the individual's blood at the current time and/or at a future time.
  • the blood glucose concentration continuous monitoring system can continuously provide sensor blood glucose measurement signals representing real-time blood glucose concentration at a certain frequency through implantable or non-implantable blood glucose sensitive sensors. properties, including photoelectric sensors, electrochemical sensors, optical absorption or optical penetration sensors, etc.
  • continuous blood glucose monitoring systems can provide patients with real-time blood glucose levels, for therapeutic purposes, there is an urgent need for predictive methods that can provide blood glucose concentration data, such as predicting blood glucose concentration data after half an hour or more .
  • Some data processing models are applied to the field of blood glucose concentration data prediction, such as Support Vector Regression (SVR) model, neural network (Neural Network, NN) model, autoregressive moving average (Autoregressive Moving Average, ARMA) model and Long Short Term Memory (LSTM) network, etc.
  • SVR Support Vector Regression
  • NN neural network
  • ARMA Autoregressive Moving Average
  • LSTM Long Short Term Memory
  • the method of using a single model for blood sugar prediction is to train the initialized deep learning model based on the relevant data of the patient, and perform rolling prediction of the patient's blood sugar. Due to the differences in the characteristics of different training models, a single prediction model trained with the same training data has different accuracy, response speed, and requirements for input information in different situations.
  • the ARMA model uses less data and resource occupancy, and is a fast and relatively mature model, but in actual use, when the selected second time period is longer, such as 30-60 minutes, the error is large . Therefore, this model is mainly used to quickly obtain the prediction of blood glucose concentration in the near future, such as 10 minutes in the future, or a shorter time.
  • the SVR model does not consider user data, such as diet and insulin intake, on blood glucose concentrations during training. Therefore, when there is a rapid fluctuation of blood sugar caused by the above user data, the prediction of the SVR model has a hysteresis.
  • Both the NN model and the LSTM model can take into account the influence of user data on blood glucose concentration during training and can combine user data (such as diet, insulin, etc.) to make predictions.
  • user data such as diet, insulin, etc.
  • human input may lead to data anomalies, etc. Therefore, when the training data deviates or the user data used in the prediction process deviates, the prediction results of the model will be affected to a certain extent.
  • the autoregressive model, support vector regression model, and long-term short-term memory network model can be used to predict the blood sugar concentration data of patients, and the blood sugar prediction results of multiple models can be obtained. Calculate the weight of each model, and then linearly combine the blood glucose prediction results of each model and the weight of each model to obtain the predicted value of the combined prediction model. After multi-model learning, the linear combination of blood glucose prediction results according to the fixed weight of each model still cannot achieve effective prediction accuracy and prediction efficiency, because the short-term changes in blood glucose concentration brought about by the user input data are not taken into account.
  • the present application provides a blood sugar prediction method and device, and a system for monitoring blood sugar levels.
  • This application provides a blood sugar prediction method, including:
  • the blood glucose trace data includes current blood glucose collection data and historical blood glucose collection data
  • the weighting factor set includes multiple weights classified based on multiple real-time scenarios Factor groups, each weighting factor group includes at least two weighting factors, and each weighting factor is determined based on the preset multi-model fusion target standard error and the prediction standard errors of the at least two blood glucose prediction models;
  • a blood sugar prediction result is obtained.
  • the present application also provides a blood sugar prediction device, including:
  • the blood glucose trace data acquisition module is configured to obtain the user's blood glucose trace data, wherein the blood glucose trace data includes current blood glucose collection data and historical blood glucose collection data;
  • a blood sugar pre-assessment module configured to input the user's blood sugar trajectory data into at least two blood sugar prediction models, and obtain at least two blood sugar pre-evaluation results respectively output by the at least two blood sugar prediction models;
  • the current scene determination module is configured to determine the current scene where the user is based on the rate of change of blood glucose concentration determined by the blood glucose trace data and/or the state of the current blood glucose collection data;
  • the weighting factor group selection module is configured to select the weighting factor group corresponding to the at least two blood glucose prediction models in the current scenario based on the current scenario, wherein the weighting factor set includes a set of weighting factors based on multiple Multiple weighting factor groups for classifying real-time scenarios, each weighting factor group includes at least two weighting factors, each weighting factor is based on the preset standard error of the multimodal fusion target and the prediction standard of the at least two blood sugar prediction models error determined;
  • the blood sugar prediction module is configured to obtain a blood sugar prediction result based on the at least two blood sugar pre-evaluation results and the weighting factor groups corresponding to the at least two blood sugar prediction models in the current scene.
  • the present application also provides a system for monitoring blood sugar levels, comprising:
  • a sensor configured to obtain a user's blood glucose measurement
  • a wireless transmitter configured to transmit said blood glucose measurement
  • a mobile computing device comprising:
  • a wireless receiver configured to receive said blood glucose measurement
  • a memory configured to store data comprising said blood glucose measurement
  • a processor configured to process data stored in the memory, and a software application program including instructions stored in the memory, the instructions implement the blood glucose prediction method as described above when executed.
  • the present application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and operable on the processor, and the processor implements the blood sugar prediction method when executing the program.
  • the present application also provides a non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the blood sugar prediction method is realized.
  • FIG. 1 is a schematic flow chart of a method for predicting blood sugar provided in an embodiment of the present application
  • Fig. 2 is a schematic diagram of the effect of the blood sugar pre-evaluation results independently predicted by at least two blood sugar prediction models provided by the embodiment of the present application;
  • Fig. 3 is a schematic diagram of the effect of a multi-model fusion blood sugar prediction result provided by the embodiment of the present application;
  • Fig. 4 is a schematic structural diagram of a blood sugar prediction device provided by an embodiment of the present application.
  • Fig. 5 is a schematic structural diagram of an electronic device provided by an embodiment of the present application.
  • a blood sugar prediction method of the present application is described below in conjunction with FIG. 1, the method includes:
  • Blood glucose trace data includes, but is not limited to, blood glucose concentration data with a time stamp, and data associated with blood glucose concentration.
  • Methods of obtaining blood glucose trace data include, but are not limited to, acquisition through a glucose concentration sensor, user input, network transfer collection, or other collection Way.
  • the user's current blood glucose concentration data with time stamp and the user's historical blood glucose concentration data with time stamp that are continuously output by the glucose concentration sensor in a certain period can be acquired.
  • the current blood glucose collection data is the blood glucose data collected at the current moment, which is real-time.
  • the historical blood glucose collection data is the historical blood glucose collection data, which can be the historical blood glucose collection data of the first time period from the current moment, and the first time period can be the past 30 Minutes to 2 hours, the historical blood glucose collection data can be represented as a period of historical data waveform with the time of the past first time period as the horizontal axis and the collected data as the vertical axis.
  • At least two blood sugar prediction models can be supervised models that are trained and tested through large data samples and real labels respectively.
  • the loss function of the model training iteration uses the mean absolute error (Mean Absolute Error, MAE), which is the L1 loss as Loss function; each blood glucose prediction model can predict a blood glucose pre-evaluation result based on the user's blood glucose trajectory data.
  • MAE mean Absolute Error
  • the at least two blood glucose pre-evaluation results respectively obtained according to the at least two blood glucose prediction models are the predicted values of blood glucose concentration after the second time period in the future predicted based on the current blood glucose collection data and the historical blood glucose collection data, and the second time period may be In the next 1 minute to 2 hours, the blood sugar pre-evaluation result can be represented as a future pre-evaluation data waveform with the time after the second time period as the horizontal axis and the predicted value of blood sugar concentration as the vertical axis.
  • the predicted value of the blood glucose concentration at the second moment in the future after the second time period can be predicted.
  • the data is moved backwards in real time, it is equivalent to dynamically adjusting the start time and end time of the first time period (equivalent to the current time) in real time.
  • the first time can be the end time or close to the end time, using at least two
  • the blood glucose prediction model can respectively obtain a set of predicted blood glucose concentration values at multiple second moments (synchronously and backwardly translated with the current moment) to form the future pre-evaluation data waveform.
  • the rate of change of blood glucose concentration is determined by a value close to the current moment in the current blood glucose collection data and historical blood glucose collection data in the blood glucose trajectory data.
  • the status of the current blood glucose collection data includes but not limited to: the current blood glucose collection data is abnormal (including abnormal blood glucose concentration data), the current blood glucose collection data is normal (including normal blood glucose concentration data and normal user input).
  • the priority of the state of the current blood glucose collection data is higher than the priority of the rate of change of blood glucose concentration.
  • Abnormal blood sugar concentration data is judged based on big data, data history and experience of previous users.
  • Abnormal blood glucose concentration data may be caused by sensor abnormality, network abnormality, data exceeding a certain range, or data missing.
  • weighting factor set corresponding to the at least two blood glucose prediction models in the current scene from the weighting factor set, wherein the weighting factor set includes multiple values classified based on multiple real-time scenarios weighting factor groups, each weighting factor group includes at least two weighting factors, and each weighting factor is determined based on the preset multi-model fusion target standard error and the prediction standard errors of the at least two blood sugar prediction models.
  • weighting factor groups corresponding to at least two blood glucose prediction models corresponding to the current scene may be determined.
  • the weighting factor group in each prediction process is dynamic and selected from the weighting factor set based on the current scene where the user is at the current moment, that is to say, the weighting factor group at different moments may be the same or different, is Determined based on the current scene.
  • the set of weighting factors is pre-stored and classified based on different real-time scenarios.
  • the weighting factor for each blood glucose prediction model is related to the prediction standard error of each blood glucose prediction model.
  • the current scene corresponds to the current moment
  • multiple real-time scenes refer to a collection of all possible current scenes corresponding to multiple current moments, and the current scenes corresponding to different moments may be the same or different.
  • Each moment is dynamically panned and advanced in real time based on the change of time, and the current scene at the current moment also changes with the panning of time.
  • the preset multi-mode fusion target standard error is a preset threshold, which can be set according to experience, characteristics of multiple models, and characteristics of multiple scenes.
  • At least two models corresponding to a real-time scene correspond to at least two
  • the weighting factor is jointly determined based on the preset multimodal fusion target standard error and the prediction standard errors of the at least two blood glucose prediction models.
  • the blood sugar prediction result includes the predicted value of blood sugar concentration corresponding to the current moment after the second time in the future, and also includes a plurality of sets of predicted blood sugar concentration values at the second time. is a section of future forecast data waveform on the horizontal axis and the predicted value of blood sugar concentration on the vertical axis.
  • the blood sugar forecast result comprehensively considers factors such as the prediction accuracy of multiple models in different scenarios, so it is closer to the user's real blood sugar level. The results are accurate and reliable.
  • Figure 2 is a schematic diagram of the effect of a blood sugar pre-evaluation result independently predicted by at least two blood sugar prediction models provided by the embodiment of the application of the present invention
  • Figure 3 is a schematic diagram of the effect of a multi-model fusion of the blood sugar prediction result provided by the embodiment of the application of the present invention
  • the abscissa represents time, which means the 1st to 500th 3 minutes
  • the ordinate represents the predicted future blood glucose concentration values at different times.
  • Fig. 2 shows the future blood glucose concentration values and real blood glucose concentration labels measured respectively by using the convolutional recurrent neural network (Convolutional Recurrent Neural Network, CRNN) model and the radial basis function (Radial Basis Function, RBF) model;
  • CRNN convolutional Recurrent Neural Network
  • RBF Radial basis Function
  • the first arrow 1 in Fig. 3 represents the diet data input by the user corresponding to this moment ( point)
  • the second arrow 2 represents the diet data (point) input by the user at another moment
  • the third arrow 3 represents insulin injection.
  • the prediction results in Figure 3 are better than the prediction results of a single model in Figure 2, no matter when the user is in a stable interval scene, or a slow-rising interval scene, the user input is correct, or a fast-rising interval scene.
  • the average root mean square error (Root Mean Square Error , RMSE)
  • This application pre-sets weighting factor sets based on different blood sugar prediction models and different real-time scenarios where the user is in, and dynamically selects the weighting factor set that best matches the current scene in the weighting factor set, and outputs the output of at least two blood sugar prediction models.
  • At least two blood glucose pre-evaluation results are assigned to at least two weighting factors of the weighting factor group, comprehensively considering the prediction accuracy, response speed, user's personalized prediction and other factors of different models in different scenarios, and strengthening the performance in the current scenario.
  • the advantages of the model weakening the shortcomings of the model with poor performance, to achieve a prediction result that best matches the current scene, and to get a blood sugar prediction result that is closer to the user's real blood sugar level.
  • the prediction result is accurate and reliable, and the response speed is fast. User experience.
  • the blood sugar track data includes current blood sugar collection data and historical blood sugar collection data, including:
  • the blood glucose trace data includes current blood glucose collection data and historical blood glucose collection data;
  • the network is a wired or wireless network, and associates the user's blood glucose measurement device Including but not limited to blood glucose meters with glucose sensors, blood collection measurement equipment, and other physiological data collection equipment that can collect blood glucose concentration data.
  • the blood glucose measuring device is a continuous blood glucose monitoring device capable of continuously collecting blood glucose concentration data in real time.
  • the current blood glucose collection data includes the first blood glucose measurement value at the current moment and its corresponding first time stamp
  • the historical blood glucose collection data includes a plurality of historical blood glucose measurement values continuously distributed at preset time intervals and their corresponding multiple historical timestamps.
  • the preset time interval is the interval at which the continuous blood glucose monitoring device generates blood glucose, such as 3 minutes.
  • the historical blood glucose collection data includes a plurality of historical blood glucose measurement values continuously distributed in a period of 3 minutes and a plurality of corresponding historical time stamps in the first time period from the current moment.
  • the blood sugar track data includes current blood sugar collection data and historical blood sugar collection data, including:
  • the blood glucose trace data includes current blood glucose collection data and historical blood glucose collection data; Data associated with blood glucose levels, such as food intake and its volume.
  • the current blood glucose collection data includes the first blood glucose associated data at the current moment input by the user and its corresponding first associated time stamp; the historical blood glucose collection data includes a plurality of historical blood glucose associated data at the historical moment input by the user and their corresponding time stamps.
  • a plurality of historical associated time stamps, the first blood glucose associated data and the historical blood glucose associated data respectively include one or more events associated with blood glucose concentration.
  • the first blood glucose associated data at the current moment and its corresponding first associated timestamp are used to determine the current scene, and multiple historical blood glucose associated data at historical moments and their corresponding multiple historical associated timestamps can be used to determine the scene corresponding to the historical moment , and can also be used to assist in calibrating the user's blood glucose concentration at historical moments.
  • the one or more events are associated with one or more of food consumption, beverage consumption, exercise, sleep, and administration of substances.
  • Administration of a substance includes administration of food, drink, drug, or insulin.
  • One or more events are each associated with a blood glucose concentration.
  • each blood sugar prediction model is trained in the following manner:
  • Each blood sugar prediction model is obtained by training based on blood sugar concentration sample data and pre-determined real blood sugar concentration tags;
  • the concentration label is the actual measured blood glucose concentration value corresponding to some specific moments in the future corresponding to the blood glucose concentration sample data.
  • Each blood sugar prediction model can use big data for training without user personalized data, and can also achieve better prediction accuracy.
  • the training model is periodically iteratively updated.
  • the 10,000 pieces of data can be used as the sample data of blood glucose concentration, and the corresponding future of the user's actual measurement
  • the blood glucose concentration value at a specific moment is used as the real blood glucose concentration label, and the user's data is used to train the model.
  • the living habits, location, diabetes type, age, etc. are all personalized based on the user's own settings, which can be more personalized model, the predicted result is closer to the real situation of the user.
  • the user's blood glucose trajectory data is input into at least two blood glucose prediction models, and at least two blood glucose pre-evaluation results respectively output by the at least two blood glucose prediction models are obtained, include:
  • Each blood glucose prediction model is obtained by performing training based on the blood glucose concentration sample data in the area where the user is located and the corresponding predetermined real blood glucose concentration label.
  • the division of the region can be made according to factors such as country, preset large region (such as Northeast China, North China, etc.), because users in each specific region may have similar living habits , environment, users in the same area have a certain regularity in the change of blood sugar concentration, so the blood sugar concentration sample data in the area where the user is located and the corresponding pre-determined real blood sugar concentration label in the area are used for training.
  • a model with regionalization is obtained, and the predicted result is closer to the real situation of users in a specific area.
  • the user's blood glucose trajectory data is input into at least two blood glucose prediction models, and at least two blood glucose pre-evaluation results respectively output by the at least two blood glucose prediction models are obtained, include:
  • each blood glucose prediction model is trained in the following manner:
  • Each blood glucose prediction model is obtained by performing training based on the blood glucose concentration sample data of the type of diabetes to which the user belongs and the corresponding predetermined real blood glucose concentration label.
  • Also before the current scene where the user is located is determined based on the blood glucose concentration change rate determined from the blood glucose trajectory data and/or the state of the current blood glucose collection data ,Also includes:
  • the second value includes a second blood glucose measurement value and its corresponding second time stamp, the second time stamp is associated with the first time stamp.
  • the calculation method of the rate of change of blood glucose concentration is:
  • the third time period can be 1 minute to 30 minutes. For example, you can select the data of 3 minutes before the current time. If there is missing or abnormal data, you can select other data within the third time period from the current time. data.
  • the determination of the current scene where the user is located is based on the rate of change of blood sugar concentration determined from the blood sugar trajectory data and/or the state of the current blood sugar collection data, including :
  • the first preset threshold is set to 0.05mmol/L/min; when the rate of change of blood glucose concentration is greater than the first If the preset threshold value is not greater than the second preset threshold value, it is determined that the user is in the slow-speed up-and-down interval scene; for example, the second preset threshold value is set to 0.1mmol/L/min; when the rate of change of the blood glucose concentration is greater than the second If the preset threshold is not greater than the third preset threshold, it is determined that the user is in a scene of a medium-speed up-and-down interval; for example, the third preset threshold is set to 0.15mmol/L/min; when the rate of change of the blood glucose concentration is greater than the third A preset threshold is used to determine that the user is in a scene of a rapid ups and downs interval.
  • the settings of the first preset threshold, the second preset threshold, and the third preset threshold are selected based on a large amount of experimental data, including but not limited to the above numerical ranges.
  • the determination of the current scene where the user is located is based on the rate of change of blood sugar concentration determined from the blood sugar trajectory data and/or the state of the current blood sugar collection data, including :
  • the user's blood glucose trace data obtained from the blood glucose measurement device associated with the user is abnormal, it is determined that the user is in an abnormal blood glucose measurement scene; the user's blood glucose trace data obtained from the blood glucose measurement device associated with the user may be abnormal It is the abnormality of blood glucose trace data caused by sensor abnormality, abnormal network of data transmission, or other circumstances.
  • the abnormality is judged by setting thresholds based on user history and some experience. There are many ways to judge whether blood glucose trace data is abnormal. When the blood glucose trace data is abnormal, the blood glucose value at the current moment cannot be used for future prediction.
  • the abnormality of the first blood glucose-related data at the current moment may be that the user has entered an event that cannot be achieved. For example, if the diet input is 5 kg, it can be determined as an abnormal user input, and if the diet input is 200 grams, it can be confirmed that the user input is normal. .
  • the abnormality or normality is judged by setting thresholds based on user history and some experience. When the user input is abnormal, the data input by the user cannot be used for future prediction, and when the user input is normal, the data input by the user can be used for future prediction.
  • the scene with abnormal blood glucose measurement value is the first priority
  • the scene with abnormal user input and the scene with normal user input are the second priority
  • Interval scenarios, medium-speed ascending and descending interval scenarios, and fast ascending and descending interval scenarios have a third priority, the first priority is greater than the second priority, and the second priority is greater than the third priority.
  • the first priority is the factor with the highest priority when determining the current scene where the user is in. If there is no first priority, then consider the factors of the second priority. If there is neither the first priority nor the first priority If there is a second priority, then consider the factors of the third priority. For example, in the scene of abnormal blood glucose measurement value, since the blood glucose concentration obtained at the current moment is inaccurate, it is impossible to predict the corresponding future blood glucose concentration value, and the prediction result is not output at this time. In a variety of different real-time scenarios, the weighting factor groups of multiple blood sugar prediction models are shown in Table 2 below.
  • each weighting factor group includes at least two weighting factors, and each weighting factor is based on the preset multi-model fusion target standard error and the prediction of the at least two blood sugar prediction models Standard errors are determined, including:
  • m represents the total amount of blood glucose concentration sample data of each blood glucose prediction model
  • i represents a variable from 1 to m
  • i and j are positive integers greater than or equal to 1
  • y i represents the ith of each blood glucose prediction model
  • Blood glucose pre-assessment results Represents the i-th real blood glucose concentration label of each blood glucose prediction model; select a set of a j ,...,b j that makes RMSE final smaller than the preset multi-modal fusion target standard error or makes RMSE final reach the minimum value as at least two
  • the weighting factors of the blood glucose prediction models in the jth real-time scene, the above ... means other models except a and b, and may only include the two models a and b.
  • the calculation formula of the prediction standard error RMSE final of the at least two blood sugar prediction models is:
  • RMSE final a j 2 *RMSE ARMA +b j 2 *RMSE SVR +c j 2 *RMSE LSTM +d j 2 *RMSE NN
  • the formula for calculating the prediction standard error RMSE of each blood glucose prediction model in a real-time scenario is:
  • the m represents the total amount of blood glucose concentration sample data
  • i represents a variable from 1 to m
  • i and j are both positive integers greater than or equal to 1
  • y i represents the i-th blood glucose pre-evaluation result, Indicates the i-th real blood glucose concentration label
  • the prediction standard error RMSE of each blood glucose prediction model is different in different scenarios, for example, the RMSE in the above formula can be the prediction standard error of each model in the current j scenario
  • y i is the estimated result of blood sugar in scene j
  • the preset target standard error of multimodal fusion is selected based on historical experience, for example, 0.5mmol/L.
  • a group of a j , b j , c j , d j that minimizes the RMSE final is preferentially selected as the weighting of the ARMA model, SVR model, LSTM model, and NN model in the jth real-time scene factor.
  • a minimum value in RMSE final When only two models are fused, then there is a minimum value in RMSE final .
  • the above four models all have the ability to use historical blood glucose data to predict future blood glucose values.
  • the content that must be input in the process of predicting blood glucose concentration is historical blood glucose concentration data and its time stamp. Due to the differences in the characteristics of different training models, the accuracy, response speed, and demand for input information of a single prediction model trained using the same training data are different in different situations. Based on the same personalized data, different single predictions The model will obtain different predicted blood sugar concentration data, so multiple blood sugar prediction models are considered comprehensively, and different factors are assigned to each model based on different scenarios, which can strengthen the advantages of some models that perform well in the current scenario and weaken some performances. The disadvantage of a bad model is to achieve a prediction result that best matches the current scene.
  • each weighting factor group includes at least two weighting factors, and each weighting factor is based on the preset multi-model fusion target standard error and the prediction of the at least two blood sugar prediction models Standard errors are determined, including:
  • Each weighting factor group includes at least two weighting factors, and each weighting factor is determined based on a preset multimodal fusion target standard error and prediction standard errors of the at least two blood glucose prediction models and is periodically iteratively updated.
  • the corresponding weighting factors are also iteratively updated synchronously, and each weighting factor is associated with the prediction standard errors of multiple blood glucose prediction models.
  • the autoregressive model, the support vector regression model and the long-term short-term memory network model can be used for prediction, and the blood glucose prediction results of each model can be obtained, and then the weights of the models can be calculated separately, and then the blood glucose prediction results of multiple models can be calculated. and multiple model weights are linearly combined to obtain the predicted value of the combined forecasting model.
  • the models with good prediction effects get greater weights. Every time multiple blood glucose prediction models are iteratively updated, the prediction standard error of each blood glucose prediction model is also updated accordingly, so the weighting factor corresponding to each blood glucose prediction model is also updated iteratively.
  • the blood sugar prediction result is obtained based on the at least two blood sugar pre-evaluation results and their corresponding weighting factor groups in the current scene, including:
  • the at least two blood glucose pre-evaluation results are respectively multiplied by corresponding weighting factors in the corresponding weighting factor group in the current scene, and then summed to obtain a blood glucose prediction result after a preset time period corresponding to the current moment .
  • the preset time period is the same as the above-mentioned second time period.
  • Glu represents the final blood glucose concentration prediction result
  • a j , b j , c j ... d j represent the weighting factor group corresponding to the jth real-time scene
  • n represents the total number of scenes, and there can be multiple scenes.
  • Glu A represents the pre-evaluation result of blood glucose using the a model
  • Glu B represents the pre-evaluation result of the blood glucose using the b model
  • the weighting factors of multiple models in multiple scenarios are shown in Table 3 below:
  • the A-M model represents a model that can predict data based on certain data, including the SVR model, NN model, ARMA model, and LSTM model.
  • the blood glucose prediction method described in the present application after obtaining the blood glucose prediction result based on the at least two blood glucose pre-evaluation results and their corresponding weighting factor groups in the current scenario, further includes:
  • the visualization of the blood glucose prediction result is realized by using at least one display module; the display module may be configured to display the blood glucose concentration prediction value at the second moment after the second time period starting from the current moment.
  • At least one alarm module is used to realize the alarm prompt of the blood sugar prediction result based on the preset blood sugar threshold.
  • the alarm module can be integrated with the display module in the same display device, or can be set independently.
  • the preset blood sugar threshold set in combination with user settings and historical experience
  • the patient's blood sugar value is mainly affected by diet (carbohydrate) and insulin.
  • the patient records the start time, intake, and insulin injection time and injection volume of the diet through the receiving device, such as a mobile application, and displays it on the display interface of the device. Display blood glucose history curve and diet and insulin.
  • the hybrid model includes: SVR model, NN model and LSTM model.
  • the input of the SVR model prediction process is only the blood glucose data and corresponding timestamps of the past 2 hours, while the inputs of the other two models can include blood glucose, diet and insulin and their corresponding timestamps.
  • the three models are calculated separately to finally obtain the predicted blood glucose value of the patient for a period of time in the future.
  • the length of time can be set by the user according to the needs, and the minimum set time interval is the interval for the continuous blood glucose monitoring equipment to generate blood glucose, such as 3 minutes.
  • the optimal length of prediction is recommended to be 30 minutes.
  • the SVR model does not consider the influence of diet and insulin, when the user records the input of diet or insulin, the blood sugar of the human body begins to rise or fall correspondingly.
  • the results of training can be obtained within a certain period of time, and the SVR model fails to timely It is predicted that blood sugar will rise or fall, therefore, at this time, the weighting factor a of SVR blood sugar in the following formula is small or 0.
  • the diet and insulin data are values manually entered by the user. According to the user's historical habits, when the diet or insulin values deviate too much from common sense or user habits, it is considered that the user has entered inaccurate information. In order to prevent inaccurate information from affecting the final prediction results, the weighting factors (b and c) of the prediction results of the NN model and LSTM model are reduced or set to 0.
  • the insulin infusion may be transmitted by the insulin infusion device to the receiving device over the first network.
  • dietary data can be obtained by software that automatically recognizes calories in pictures of food.
  • the SVR model, NN model and LSTM model are trained using Chinese data during the training process, and the training data includes blood sugar, diet and insulin information.
  • the training data includes blood sugar, diet and insulin information.
  • multiple basic models are obtained through training. Determine the weighting factors of multiple models at different times and under different circumstances by comparing the differences between multiple models and actual values at different times. Differences were assessed by RMSE.
  • Glu Glu SVR *a 1 +Glu NN *b 1 +Glu LSTM *c 1
  • the weighting factors of the two models are set when the user data is generated. Zero, the weighting factors of the other two models are determined according to the above RMSE.
  • RMSE final c 2 *RMSE LSTM +d 2 *RMSE NN
  • RMSE final c 2 *RMSE LSTM +(1-c) 2 *RMSE NN
  • the c value when RMSE final is the smallest can be obtained, and the value of c is determined by RMSE LSTM and RMSE NN .
  • c represents the weight factor of the LSTM model in this scenario
  • d represents the weight factor of the NN model in this scenario.
  • the weighting factor recovery time of the ARMA model and the SVR model is determined according to the time when the impact of user events on blood sugar decreases. For example, after 2 hours after a meal, the impact of dietary events on blood sugar values decreases, and the selection of weighting factors is based on the blood sugar in the first case. Volatility interval selection.
  • the weighting factors of the LSTM model and the NN model that will use user data in the prediction process can be set to zero, and the recovery time depends on blood sugar changes The time at which the rate or user event's effect on blood glucose decreases, such as 2 hours after a meal, is determined. Whether the user data is wrong is judged by the electronic device according to the data history and experience of previous users.
  • the blood sugar prediction device includes:
  • the blood glucose trace data acquisition module 10 is configured to obtain the user's blood glucose trace data, and the blood glucose trace data includes current blood glucose collection data and historical blood glucose collection data;
  • the blood glucose pre-assessment module 20 is configured to input the user's blood glucose trace data respectively At least two blood glucose prediction models, obtaining at least two blood glucose pre-evaluation results respectively output by the at least two blood glucose prediction models;
  • the current scene determination module 30 is configured to be based on the blood glucose concentration change rate determined by the blood glucose trajectory data and/or Or the state of the current blood glucose collection data, determine the current scene where the user is in;
  • the weighting factor group selection module 40 is set to select the at least two blood glucose prediction models in the weighting factor set based on the current scene The weighting factor group corresponding to the current scene, wherein the weighting factor set includes a plurality of weighting factor groups classified based on multiple real-time scenes, each weighting factor group includes at least two weighting factors, and each weighting factor is based on determined by the preset multimodal
  • the blood sugar prediction device corresponds one-to-one to the blood sugar prediction method in the above embodiments, any embodiment of the above blood sugar prediction method is also applicable to the blood sugar prediction device, and details will not be repeated here.
  • the present application also provides a system for monitoring blood sugar levels, comprising:
  • a sensor configured to obtain a blood glucose measurement of a user; a wireless transmitter configured to transmit the blood glucose measurement; and a mobile computing device comprising: a wireless receiver configured to receive the blood glucose measurement; a memory configured to store data comprising said blood glucose measurement; a processor configured to process said data, and a software application comprising instructions stored in said memory which when executed by said processor to perform The blood sugar prediction methods of the above-mentioned multiple embodiments.
  • FIG. 5 is a schematic structural diagram of an electronic device provided by an embodiment of the present application.
  • the electronic device may include: a processor (processor) 510, a communication interface (Communications Interface) 520, a memory (memory) 530, and a communication bus 540, wherein, The processor 510 , the communication interface 520 , and the memory 530 communicate with each other through the communication bus 540 .
  • the processor 510 can invoke logic instructions in the memory 530 to execute the blood glucose prediction method of the above-mentioned embodiment.
  • the above logic instructions in the memory 530 may be implemented in the form of software function units and be stored in a computer-readable storage medium when sold or used as an independent product.
  • the technical solution of the present application can be embodied in the form of a software product in essence.
  • the computer software product is stored in a storage medium and includes a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) ) Execute all or part of the steps of the method described in the embodiments of the present application.
  • the aforementioned storage media include: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), magnetic disk or optical disc and other media that can store program codes. .
  • the present application also provides a computer program product
  • the computer program product includes a computer program stored on a non-transitory computer-readable storage medium
  • the computer program includes program instructions, and when the program instructions are executed by a computer During execution, the computer can execute the blood sugar prediction method provided by the above-mentioned embodiments.
  • the present application also provides a non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the method for predicting blood sugar provided by the above-mentioned embodiments is implemented.
  • the device embodiments described above are only illustrative, and the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in One place, or it can be distributed to multiple network elements. Part or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.
  • the embodiments can be implemented by means of software plus a necessary general-purpose hardware platform, or by means of hardware.
  • the above-mentioned technical solutions can be embodied in the form of software products in essence, and the computer software products can be stored in computer-readable storage media, such as ROM/RAM, magnetic disks, optical disks, etc., and include multiple instructions to make a computer device (It may be a personal computer, a server, or a network device, etc.) executes the methods described in the embodiments or some parts of the embodiments.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Theoretical Computer Science (AREA)
  • Medical Informatics (AREA)
  • Biomedical Technology (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Biophysics (AREA)
  • Public Health (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Molecular Biology (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Pathology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Animal Behavior & Ethology (AREA)
  • Veterinary Medicine (AREA)
  • Surgery (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Signal Processing (AREA)
  • Psychiatry (AREA)
  • Physiology (AREA)
  • Computational Linguistics (AREA)
  • Primary Health Care (AREA)
  • Epidemiology (AREA)
  • Databases & Information Systems (AREA)
  • Fuzzy Systems (AREA)
  • Emergency Medicine (AREA)
  • Optics & Photonics (AREA)
  • Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
  • Investigating Or Analysing Biological Materials (AREA)

Abstract

一种血糖预测方法和装置、监测血糖水平的系统。血糖预测方法包括:获取用户的血糖轨迹数据;将用户的血糖轨迹数据输入至少两个血糖预测模型,得到至少两个血糖预测模型分别输出的至少两个血糖预评估结果;基于由血糖轨迹数据确定的血糖浓度变化率和/或当前血糖采集数据的状态,确定用户所处的当前场景;基于当前场景,在加权因子集中选取至少两个血糖预测模型在当前场景下对应的加权因子组,其中,加权因子集包含基于多个实时场景进行分类的多个加权因子组;基于至少两个血糖预评估结果及至少两个血糖预测模型在当前场景下对应的加权因子组,得到血糖预测结果。

Description

血糖预测方法和装置、监测血糖水平的系统
本申请要求在2021年12月21日提交中国专利局、申请号为202111565948.8的中国专利申请的优先权,该申请的全部内容通过引用结合在本申请中。
技术领域
本申请涉及生理信号处理领域,例如涉及一种血糖预测方法和装置、监测血糖水平的系统。
背景技术
糖尿病是一种代谢紊乱性疾病,影响着世界人口中的很大一部分。及时正确地诊断和治疗糖尿病对于病患相对健康的生活方式至关重要。通常,糖尿病的治疗需要依赖于当前时刻和/或将来时刻中个体血液中的葡萄糖浓度。
随着生物传感器的技术发展,血糖浓度连续性监控系统(按一定频率连续提供实时血糖浓度数据)的出现,给病患提供了更好地了解其血糖浓度变化水平的途径,该系统提供了控制血糖浓度的数据基础,这对于糖尿病的管理而言是非常有用的。血糖浓度连续性监控系统可通过植入式或非植入式血糖敏感型传感器按一定频率连续提供表示实时血糖浓度的传感器血糖测量信号,这种传感器可以测量血液、人体组织或人体一部位的多种属性,包括光电传感器、电化学传感器、光学吸收或光学穿透传感器等。
尽管血糖浓度连续性监控系统可向病患提供实时血糖浓度水平,出于治疗的目的,还亟需要能够提供血糖浓度数据的预测方法,比如预测半小时后或更长时间段之后的血糖浓度数据。
一些数据处理的模型被运用至血糖浓度数据预测领域,如支持向量回归(Support Vector Regression,SVR)模型,神经网络(Neural Network,NN)模型,自回归滑动平均(Autoregressive Moving Average,ARMA)模型和长短期记忆(Long Short Term Memory,LSTM)网络等,可以采用单一模型的方式进行血糖预测,也可采用多重模型的混合模型进行血糖预测。采用单一模型进行血糖预测的方法,是基于病患的相关数据,对初始化的深度学习模型进行训练,对病患的血糖进行滚动预测。由于不同训练模型的特性差异,使用相同的训练数据训练得到的单一预测模型在不同的情况下准确性、响应速度、对输入信息的需求各不相同,因此,基于相同的个性化数据,不同的单一预测模型将会获得不同的预测血糖浓度数据。如ARMA模型使用的数据和资源占用率较少,是一种快速且相对成熟的模型,但是在实际的使用中,当选择的第二时间段较长, 如30-60分钟时,误差较大。因此该模型被主要用于快速得到临近的未来时间的血糖浓度预测中,如未来10分钟,或更短的时间。SVR模型在训练过程中没有考虑用户数据,如饮食和胰岛素的摄入等对血糖浓度的影响。因此当出现由上述用户数据引起的血糖快速波动时,SVR模型的预测有滞后性。NN模型和LSTM模型均可以在训练中考虑到用户数据对血糖浓度的影响并可以结合用户数据(如饮食、胰岛素等)进行预测。但是由于部分用户数据为人为输入的数据,人为输入可能导致数据异常等情况,因此当训练的数据出现偏差或预测过程中使用的用户数据出现偏差时,模型的预测结果则会受一定的影响。
在采用混合模型进行血糖预测的方法中,可以针对病患的血糖浓度数据,分别应用自回归模型、支持向量回归模型和长短期记忆网络模型进行预测,得到多个模型的血糖预测结果,再分别计算每个模型的权重,进而对每个模型的血糖预测结果及每个模型的权重进行线性组合,得到组合预测模型预测值。多模型学习后按每个模型的固定权重对血糖预测结果进行线性组合,仍无法达到有效的预测精度和预测效率,由于未考虑到用户输入数据所带来的短时间内血糖浓度变化规律对预测结果的影响,譬如用户目前的血糖浓度水平、胰岛素使用量、摄入或消耗的碳水化合物、运动量、用药等或出现错误数据情况,未考虑到用户的个性化预测、尤其是无法确保不同场景下的血糖预测精度,导致用户体验感差。
发明内容
本申请提供一种血糖预测方法和装置、监测血糖水平的系统。
本申请提供一种血糖预测方法,包括:
获取用户的血糖轨迹数据,其中,所述血糖轨迹数据包括当前血糖采集数据和历史血糖采集数据;
将所述用户的血糖轨迹数据输入至少两个血糖预测模型,得到所述至少两个血糖预测模型分别输出的至少两个血糖预评估结果;
基于由所述血糖轨迹数据确定的血糖浓度变化率和/或所述当前血糖采集数据的状态,确定所述用户所处的当前场景;
基于所述当前场景,在加权因子集中选取所述至少两个血糖预测模型在所述当前场景下对应的加权因子组,其中,所述加权因子集包含基于多个实时场景进行分类的多个加权因子组,每个加权因子组包含至少两个加权因子,每个加权因子是基于预设的多模融合目标标准误差以及所述至少两个血糖预测模型的预测标准误差确定的;
基于所述至少两个血糖预评估结果及所述至少两个血糖预测模型在所述当前场景下对应的加权因子组,得到血糖预测结果。
本申请还提供了一种血糖预测装置,包括:
血糖轨迹数据获取模块,设置为获取用户的血糖轨迹数据,其中,所述血糖轨迹数据包括当前血糖采集数据和历史血糖采集数据;
血糖预评估模块,设置为将所述用户的血糖轨迹数据输入至少两个血糖预测模型,得到所述至少两个血糖预测模型分别输出的至少两个血糖预评估结果;
当前场景确定模块,设置为基于由所述血糖轨迹数据确定的血糖浓度变化率和/或所述当前血糖采集数据的状态,确定所述用户所处的当前场景;
加权因子组选取模块,设置为基于所述当前场景,在加权因子集中选取所述至少两个血糖预测模型在所述当前场景下对应的加权因子组,其中,所述加权因子集包含基于多个实时场景进行分类的多个加权因子组,每个加权因子组包含至少两个加权因子,每个加权因子是基于预设的多模融合目标标准误差以及所述至少两个血糖预测模型的预测标准误差确定的;
血糖预测模块,设置为基于所述至少两个血糖预评估结果及所述至少两个血糖预测模型在所述当前场景下对应的加权因子组,得到血糖预测结果。
本申请还提供了一种监测血糖水平的系统,包括:
传感器,设置为获取用户的血糖测量值;
无线发射器,设置为发射所述血糖测量值;
以及
移动计算装置,其包括:
无线接收器,设置为接收所述血糖测量值;
存储器,设置为存储包含所述血糖测量值的数据;
处理器,设置为处理所述存储器存储的数据,以及软件应用程序,其包含存储于所述存储器中的指令,所述指令执行时实现如上述所述的血糖预测方法。
本申请还提供一种电子设备,包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序,所述处理器执行所述程序时实现所述的血糖预测方法。
本申请还提供一种非暂态计算机可读存储介质,其上存储有计算机程序,该计算机程序被处理器执行时实现所述的血糖预测方法。
附图说明
图1是本申请实施例提供的一种血糖预测方法的流程示意图;
图2是本申请实施例提供的一种至少两个血糖预测模型分别单独预测的血糖预评估结果效果示意图;
图3是本申请实施例提供的一种多模型融合的血糖预测结果效果示意图;
图4是本申请实施例提供的一种血糖预测装置的结构示意图;
图5是本申请实施例提供的一种电子设备的结构示意图。
具体实施方式
下面将结合本申请中的附图,对本申请中的技术方案进行描述,所描述的实施例是本申请一部分实施例。
下面结合图1描述本申请的一种血糖预测方法,该方法包括:
S1、获取用户的血糖轨迹数据,所述血糖轨迹数据包括当前血糖采集数据和历史血糖采集数据。
血糖轨迹数据包括但不限于带有时间戳的血糖浓度数据、与血糖浓度相关联的数据,血糖轨迹数据的获取方式包括但不限于通过葡萄糖浓度传感器获取、用户输入、由网络传递采集或其他采集方式。例如可以获取葡萄糖浓度的传感器按一定周期连续输出的带有时间戳的用户当前血糖浓度数据和带有时间戳的用户历史血糖浓度数据。当前血糖采集数据是当前时刻采集的血糖数据,具有实时性,历史血糖采集数据是历史采集的血糖数据,可以是距离当前时刻第一时间段的历史血糖采集数据,第一时间段可以为过去30分钟至2小时,历史血糖采集数据可以体现为以过去第一时间段的时间为横轴、以采集数据为纵轴的一段历史数据波形。
S2、将所述用户的血糖轨迹数据输入至少两个血糖预测模型,得到所述至少两个血糖预测模型分别输出的至少两个血糖预评估结果。
至少两个血糖预测模型均可以是有监督的、分别经过大数据样本和真实标签进行训练并测试的模型,模型训练迭代的损失函数使用平均绝对误差(Mean Absolute Error,MAE),即将L1损失作为损失函数;每个血糖预测模型可以基于用户的血糖轨迹数据,预测一个血糖预评估结果。根据至少两个血糖预测模型分别得到的至少两个血糖预评估结果均是基于当前血糖采集数据和历史血糖采集数据所预测的未来第二时间段之后的血糖浓度预测值,第二时间段可以为未来1分钟至2小时,血糖预评估结果可以体现为以第二时间段之后的时间为 横轴、以血糖浓度预测值为纵轴的一段未来预评估数据波形。对应于第一时间段的第一时刻的采集数据,可以预测出第二时间段之后、在未来第二时刻的血糖浓度预测值。由于数据是实时向后平移推进的,相当于实时动态调整第一时间段的起点时刻和终点时刻(相当于当前时刻),例如,第一时刻可以是终点时刻或接近终点时刻,采用至少两个血糖预测模型分别可以获得在多个第二时刻(与当前时刻同步向后平移推进)的血糖浓度预测值集合,形成所述未来预评估数据波形。
S3、基于由所述血糖轨迹数据确定的血糖浓度变化率和/或所述当前血糖采集数据的状态,确定所述用户所处的当前场景。
血糖浓度变化率是由所述血糖轨迹数据中的当前血糖采集数据和历史血糖采集数据中的一个接近当前时刻的值来确定的,当前血糖采集数据的状态包括但不限于:当前血糖采集数据异常(包含血糖浓度数据异常)、当前血糖采集数据正常(包含血糖浓度数据正常和用户输入正常)。在确定所述用户所处的当前场景时,当前血糖采集数据的状态的优先级高于血糖浓度变化率的优先级,在当前血糖采集数据正常时,若有用户输入,则优先考虑基于用户输入确定所述用户所处的当前场景,若无用户输入,则基于血糖浓度变化率来确定所述用户所处的当前场景。
用户输入是否异常是根据以往的用户的数据历史和经验进行判断。血糖浓度数据异常是根据大数据、以往的用户的数据历史和经验进行判断。血糖浓度数据异常可能是由于传感器异常、网络异常导致的数据超出一定范围、或者数据缺失等原因造成的。
S4、基于所述当前场景,在加权因子集中选取所述至少两个血糖预测模型在所述当前场景下对应的加权因子组,其中,所述加权因子集包含基于多个实时场景进行分类的多个加权因子组,每个加权因子组包含至少两个加权因子,每个加权因子是基于预设的多模融合目标标准误差以及所述至少两个血糖预测模型的预测标准误差确定的。
基于S3所确定的所述用户所处的当前场景,可以确定出当前场景对应的至少两个血糖预测模型对应的加权因子组。每一次预测过程中的加权因子组是动态的、基于当前时刻用户所处的当前场景在加权因子集中选取的,也就是说,不同时刻的加权因子组可能是相同,也可能是不同的,是基于当前场景确定的。加权因子集是预存的、且预先基于不同的实时场景进行分类的。每个血糖预测模型的加权因子与每个血糖预测模型的预测标准误差相关。当前场景是对应于当前时刻的,多个实时场景是指多个当前时刻对应的所有可能出现的当前场景的集合,不同时刻对应的当前场景可能相同,也可能不同。每个时刻是基于时 间的变化而实时动态平移推进的,当前时刻的当前场景也是随着时间的平移推进而变化的。预设的多模融合目标标准误差是预先设定好的阈值,可以根据经验、多个模型的特点、多个场景的特点等设置,一个实时场景所对应的至少两个模型对应的至少两个加权因子是基于预设的多模融合目标标准误差以及所述至少两个血糖预测模型的预测标准误差共同确定的。
S5、基于所述至少两个血糖预评估结果及所述至少两个血糖预测模型在所述当前场景下对应的加权因子组,得到血糖预测结果。
基于在所述当前场景下对应的加权因子组,赋予至少两个血糖预评估结果以至少两个加权因子,并对其进行线性组合,可得到血糖预测结果。血糖预测结果包括对应于当前时刻在未来第二时刻之后的血糖浓度预测值,也包含了多个第二时刻的血糖浓度预测值集合,血糖预测结果可以体现为未来第二时间段之后的以时间为横轴、以血糖浓度预测值为纵轴的一段未来预测数据波形,该血糖预测结果综合考虑了多个模型在不同场景下的预测精度等因素,因此更加接近用户的真实血糖水平,该预测结果精准可靠。
图2是本发明申请实施例提供的一种至少两个血糖预测模型分别单独预测的血糖预评估结果效果示意图,图3是本发明申请实施例提供的一种多模型融合的血糖预测结果效果示意图。在图2-3中,横坐标表示时间,含义是第1-500个3分钟,纵坐标表示不同时刻所预测的未来血糖浓度值。图2示出了使用卷积递归神经网络(Convolutional Recurrent Neural Network,CRNN)模型、径向基函数(Radial Basis Function,RBF)模型分别测得的未来血糖浓度值以及真实血糖浓度标签;图3示出了基于CRNN、RBF两个模型考虑当前场景进行权重因子融合后测得的未来血糖浓度值以及真实血糖浓度标签,图3中的第一个箭头1表示该时刻对应的用户输入的饮食数据(点),第二箭头2表示另一时刻用户输入的饮食数据(点),第三个箭头3表示胰岛素注入,在用户输入正确的情况下,是基于用户输入的数据确定当前场景和权重因子。可以看出,图3中预测的结果无论是当用户处于平稳区间场景,还是缓速升降区间场景、用户输入正确场景或快速升降区间场景,相较于图2中单个模型的预测结果都是更为接近真实血糖浓度标签,得到的结果更精准,CRNN模型、RBF模型、以及两个模型融合得到的混合模型分别在图2-3所示的所有时刻的平均均方根差(Root Mean Square Error,RMSE),也就是预测标准误差如下表1所示。
表1
模型种类 RMSE
LSTM(CRNN) 11.1
RBF(RBF) 11.8
混合模型 9.1
本申请通过基于不同的血糖预测模型和用户所处的不同实时场景预先设置加权因子集,在该加权因子集中动态选取与当前场景匹配最优的加权因子组,将至少两个血糖预测模型输出的至少两个血糖预评估结果分别赋予加权因子组的至少两个加权因子,综合考虑了不同模型在不同场景下的预测精度、响应速度、用户的个性化预测等因素,在当前场景下加强表现好的模型的优点,弱化表现不好的模型的缺点,达到一个与当前场景最匹配的预测结果,能够得到更加接近用户真实血糖水平的血糖预测结果,该预测结果精准可靠,响应速度快,提高了用户体验感。
根据本申请所述的血糖预测方法,其中,所述获取用户的血糖轨迹数据,所述血糖轨迹数据包括当前血糖采集数据和历史血糖采集数据,包括:
通过网络从关联所述用户的血糖测量设备处获取用户的血糖轨迹数据,所述血糖轨迹数据包括当前血糖采集数据和历史血糖采集数据;网络为有线或无线网络,关联所述用户的血糖测量设备包括但不限于含葡萄糖传感器的血糖仪、采血测量设备、其他可以采集血糖浓度数据的生理数据采集设备。例如,所述血糖测量设备是能够连续实时采集血糖浓度数据的连续血糖监测设备。
所述当前血糖采集数据包括当前时刻的第一血糖测量值及其对应的第一时间戳,所述历史血糖采集数据包括按预设时间间隔连续分布的多个历史血糖测量值及其对应的多个历史时间戳。
预设时间间隔为连续血糖监测设备产生血糖的间隔,如3分钟。历史血糖采集数据包括距离当前时刻第一时间段的、以3分钟为周期连续分布的多个历史血糖测量值及其对应的多个历史时间戳。
根据本申请所述的血糖预测方法,其中,所述获取用户的血糖轨迹数据,所述血糖轨迹数据包括当前血糖采集数据和历史血糖采集数据,包括:
获取用户输入的血糖轨迹数据,所述血糖轨迹数据包括当前血糖采集数据和历史血糖采集数据;用户输入的血糖轨迹数据包括用户手动输入、或利用手机应用程序(Application,APP)获取的图片识别出的与血糖浓度相关联的数据,例如食物摄入及其摄入量等。
所述当前血糖采集数据包括用户输入的当前时刻的第一血糖关联数据及其对应的第一关联时间戳;所述历史血糖采集数据包括用户输入的历史时刻的多个历史血糖关联数据及其对应的多个历史关联时间戳,所述第一血糖关联数据、 历史血糖关联数据分别包括与血糖浓度相关联的一个或多个事件。
当前时刻的第一血糖关联数据及其对应的第一关联时间戳用于确定当前场景,历史时刻的多个历史血糖关联数据及其对应的多个历史关联时间戳可用于确定历史时刻对应的场景,也可以用于辅助校准用户在历史时刻的血糖浓度。
根据本申请所述的血糖预测方法,其中,所述一个或多个事件与食物消耗、饮料消耗、锻炼、睡眠以及物质的施予中的一个或多者相关联。物质的施予包含食物、饮料、药物、或胰岛素的施予。一个或多个事件均与血糖浓度相关联。
根据本申请所述的血糖预测方法,其中,每个血糖预测模型按照以下方式进行训练:
基于血糖浓度样本数据以及预先确定的真实血糖浓度标签进行训练,得到每个血糖预测模型;血糖浓度样本数据可以包含一些用户的大数据中的一些历史血糖浓度数据和历史用户输入的数据,真实血糖浓度标签是实际测得的与血糖浓度样本数据对应的未来一些特定时刻的真实血糖浓度值。每个血糖预测模型在没有用户个性化数据的前提下,可以利用大数据进行训练,也可以达到较好的预测精度。
或周期性地基于所述用户迭代更新的血糖浓度样本数据以及对应的真实血糖浓度标签进行训练,得到每个血糖预测模型。
当一个用户的数据量每累计到一定程度(如10000条数据)后就周期性的迭代更新训练模型,此时就可以将该10000条数据作为血糖浓度样本数据,将该用户实测的对应的未来一些特定时刻的血糖浓度值作为真实血糖浓度标签,以用户的数据来训练模型,生活习惯、所处地域、糖尿病类型、年龄等都是基于用户自身进行个性化设定的,可以得到更加个性化的模型,预测的结果更接近用户的真实情况。
根据本申请所述的血糖预测方法,其中,所述将所述用户的血糖轨迹数据输入至少两个血糖预测模型,得到所述至少两个血糖预测模型分别输出的至少两个血糖预评估结果,包括:
将所述用户的血糖轨迹数据输入至少两个血糖预测模型,得到所述至少两个血糖预测模型分别输出的至少两个血糖预评估结果;其中,所述用户的血糖轨迹数据还包括所述用户的所在区域;相应的,每个血糖预测模型按照以下方式进行训练:
基于所述用户所在区域的血糖浓度样本数据以及对应的预先确定的真实血糖浓度标签进行训练,得到每个血糖预测模型。
在预测时,考虑到用户的所在区域,区域的划分可以是按照国别、预设的 大区域(例如东北、华北等)等因素进行区别,因为每个特定区域的用户可能存在相似的生活习性、环境,同一个区域的用户对于血糖浓度的变化具有一定的规律性,因此训练时采用所述用户所在区域的血糖浓度样本数据以及对应的预先确定的该区域的真实血糖浓度标签进行训练,可以得到具有地域化的模型,预测的结果更接近在特定区域中用户的真实情况。
根据本申请所述的血糖预测方法,其中,所述将所述用户的血糖轨迹数据输入至少两个血糖预测模型,得到所述至少两个血糖预测模型分别输出的至少两个血糖预评估结果,包括:
将所述用户的血糖轨迹数据输入至少两个血糖预测模型,得到所述至少两个血糖预测模型分别输出的至少两个血糖预评估结果;其中,所述用户的血糖轨迹数据还包括所述用户所属的糖尿病类型;相应的,每个血糖预测模型按照以下方式进行训练:
基于所述用户所属糖尿病类型的血糖浓度样本数据以及对应的预先确定的真实血糖浓度标签进行训练,得到每个血糖预测模型。
在预测时,考虑到用户的糖尿病类型,糖尿病类型包括1型糖尿病、2型糖尿病、妊娠糖尿病,因为每个糖尿病类型的用户可能存在相似的血糖浓度规律、以及相似的食物、药物敏感性等,每个糖尿病类型的用户对于血糖浓度的变化具有一定的规律性,因此训练时采用所述用户所属糖尿病类型的血糖浓度样本数据以及患有该糖尿病类型的用户对应的预先确定的真实血糖浓度标签进行训练,可以得到具有按照用户所属糖尿病类型进行区分的模型,预测的结果更接近患有该糖尿病类型的用户的真实情况。
根据本申请所述的血糖预测方法,其中,在所述基于由所述血糖轨迹数据确定的血糖浓度变化率和/或所述当前血糖采集数据的状态,确定所述用户所处的当前场景之前,还包括:
基于所述第一血糖测量值及其对应的第一时间戳、以及在所述多个历史血糖测量值及其对应的多个历史时间戳中选取的第二值,确定所述血糖浓度变化率;所述第二值包括第二血糖测量值及其对应的第二时间戳,所述第二时间戳与所述第一时间戳相关联。
血糖浓度变化率的计算方式为:|第一血糖测量值-第二血糖测量值|/(第二时间戳-第一时间戳),第二值的选择可在距离当前时刻的第三时间段内进行选择,第三时间段可以是1分钟至30分钟,例如,可选择当前时刻之前的3分钟时的数据,若有数据缺失或异常,可选择其他距离当前时刻的第三时间段内的数据。
根据本申请所述的血糖预测方法,其中,所述基于由所述血糖轨迹数据确 定的血糖浓度变化率和/或所述当前血糖采集数据的状态,确定所述用户所处的当前场景,包括:
当所述血糖浓度变化率不大于第一预设阈值,确定所述用户处于平稳区间场景;例如,第一预设阈值设置为0.05mmol/L/min;当所述血糖浓度变化率大于第一预设阈值、且不大于第二预设阈值,确定所述用户处于缓速升降区间场景;例如,第二预设阈值设置为0.1mmol/L/min;当所述血糖浓度变化率大于第二预设阈值、且不大于第三预设阈值,确定所述用户处于中速升降区间场景;例如,第三预设阈值设置为0.15mmol/L/min;当所述血糖浓度变化率大于第三预设阈值,确定所述用户处于快速升降区间场景。
所述第一预设阈值、第二预设阈值、第三预设阈值的设置是基于大量试验数据进行选取的,包括但不限于以上数值区间。
根据本申请所述的血糖预测方法,其中,所述基于由所述血糖轨迹数据确定的血糖浓度变化率和/或所述当前血糖采集数据的状态,确定所述用户所处的当前场景,包括:
当从关联所述用户的血糖测量设备处获取的用户的血糖轨迹数据异常,确定所述用户处于血糖测量值异常场景;从关联所述用户的血糖测量设备处获取的用户的血糖轨迹数据异常可能是传感器异常、传输数据的网络异常、或其他情况导致的血糖轨迹数据异常,该异常是基于用户历史情况以及一些经验进行设置阈值来判定的,血糖轨迹数据是否异常的判定方式有多种,当血糖轨迹数据异常时,当前时刻的血糖值已经不能用于未来的预测。
当用户输入的当前时刻的第一血糖关联数据异常,确定所述用户处于用户输入异常场景;当用户输入的当前时刻的第一血糖关联数据正常,确定所述用户处于用户输入正常场景;用户输入的当前时刻的第一血糖关联数据异常可能是用户在输入时输入了一个不可能达到的事件,例如可以将饮食输入为5千克确定为用户输入异常,饮食输入为200克则确认为用户输入正常。该异常或正常同理是基于用户历史情况以及一些经验进行设置阈值来判定的。当用户输入异常时,该用户输入的数据则不能用于未来的预测,当用户输入正常时,该用户输入的数据可以用于未来的预测。
在确定所述用户处于的当前场景时,所述血糖测量值异常场景为第一优先级;所述用户输入异常场景和用户输入正常场景为第二优先级,所述平稳区间场景、缓速升降区间场景、中速升降区间场景以及快速升降区间场景为第三优先级,所述第一优先级大于第二优先级,所述第二优先级大于第三优先级。
第一优先级作为确定所述用户所处的当前场景时最优先考虑的因素,如果 不存在第一优先级的情况,则考虑第二优先级的因素,若既不存在第一优先级也不存在第二优先级的情况,则考虑第三优先级的因素。例如,血糖测量值异常场景下,由于当前时刻获取的血糖浓度值不准,也就无法预测对应的未来时刻的血糖浓度值,此时也就不输出预测结果。在多种不同的实时场景下,多个血糖预测模型的加权因子组如下表2所示。
表2
Figure PCTCN2022085450-appb-000001
根据本申请所述的血糖预测方法,其中,每个加权因子组包含至少两个加权因子,每个加权因子是基于预设的多模融合目标标准误差以及所述至少两个血糖预测模型的预测标准误差确定的,包括:
设定至少两个血糖预测模型在第j个实时场景下的多个加权因子分别为a j,…,b j,且满足a j+…+b j=1;所述至少两个血糖预测模型的预测标准误差RMSE final的计算公式为:
Figure PCTCN2022085450-appb-000002
Figure PCTCN2022085450-appb-000003
表示至少两个血糖预测模型中的a模型在j场景下的预测标准误差,
Figure PCTCN2022085450-appb-000004
表示至少两个血糖预测模型中的b模型在j场景下的预测标准误差,每个 血糖预测模型的预测标准误差RMSE的计算公式为:
Figure PCTCN2022085450-appb-000005
m表示每个血糖预测模型的血糖浓度样本数据的总量,i表示从1至m的变量,i、j均为大于或等于1的正整数,y i表示每个血糖预测模型的第i个血糖预评估结果,
Figure PCTCN2022085450-appb-000006
表示每个血糖预测模型的第i个真实血糖浓度标签;选取使RMSE final小于预设的多模融合目标标准误差或使RMSE final达到最小值的一组a j,…,b j作为至少两个血糖预测模型分别在第j个实时场景下的加权因子,以上…表示除a、b以外的其他模型,也可能只包含a、b这两种模型。
可以设定至少两个血糖预测模型(包括ARMA模型、SVR模型、LSTM模型和NN模型),在第j个实时场景下的多个加权因子分别为a j,b j,c j,d j,且满足a j+b j+c j+d j=1;所述至少两个血糖预测模型的预测标准误差RMSE final的计算公式为:
RMSE final=a j 2*RMSE ARMA+b j 2*RMSE SVR+c j 2*RMSE LSTM+d j 2*RMSE NN
在一实时场景下的每个血糖预测模型的预测标准误差RMSE的计算公式为:
Figure PCTCN2022085450-appb-000007
所述m表示血糖浓度样本数据的总量,i表示从1至m的变量,i、j均为大于或等于1的正整数,y i表示第i个血糖预评估结果,
Figure PCTCN2022085450-appb-000008
表示第i个真实血糖浓度标签;每个血糖预测模型的预测标准误差RMSE在不同的场景下是不同的,例如,上式中的RMSE可以是在当前j场景下每个模型的预测标准误差,y i是在j场景下的血糖预估结果,
Figure PCTCN2022085450-appb-000009
是在j场景下的第i个真实血糖浓度标签。
选取使RMSE final小于预设的多模融合目标标准误差的一组a j,b j,c j,d j作为ARMA模型、SVR模型、LSTM模型、NN模型分别在第j个实时场景下的加权因子。预设的多模融合目标标准误差是基于历史经验选定的,例如为0.5mmol/L。当RMSE final存在最小值时,优先选取使RMSE final最小的一组a j,b j,c j,d j作为ARMA模型、SVR模型、LSTM模型、NN模型分别在第j个实时场景下的加权因子。当只有两个模型融合时,此时即RMSE final存在最小值。当RMSE final不存在最小值时,则选取使RMSE final小于预设的多模融合目标标准误差的一组a j,b j,c j,d j作为ARMA模型、SVR模型、LSTM模型、NN模型分别在第j个实时场景下的加权因子。
以上四种模型均具有利用历史血糖数据预测未来血糖值的能力,在进行血 糖浓度的预测过程中必须输入的内容为历史的血糖浓度数据和其时间戳。由于不同训练模型的特性差异,使用相同的训练数据训练得到的单一预测模型在不同的情况下准确性、响应速度、对输入信息的需求各不相同,基于相同的个性化数据,不同的单一预测模型将会获得不同的预测血糖浓度数据,因此综合考虑多个血糖预测模型,基于不同的场景对每个模型赋予不同的因子,能够加强在当前场景下一些表现好的模型的优点,弱化一些表现不好模型的缺点,达到一个与当前场景最匹配的预测结果。
根据本申请所述的血糖预测方法,其中,每个加权因子组包含至少两个加权因子,每个加权因子是基于预设的多模融合目标标准误差以及所述至少两个血糖预测模型的预测标准误差确定的,包括:
每个加权因子组包含至少两个加权因子,每个加权因子是基于预设的多模融合目标标准误差以及所述至少两个血糖预测模型的预测标准误差确定并周期性地迭代更新的。当血糖预测模型迭代更新时,对应的加权因子也同步迭代更新,每个加权因子是与多个血糖预测模型的预测标准误差相关联的。
本申请实施例可以应用自回归模型、支持向量回归模型和长短期记忆网络模型进行预测,得到每个模型的血糖预测结果,再分别计算所述模型的权重,进而对多个模型的血糖预测结果及多个模型权重进行线性组合,得到组合预测模型预测值。同时还通过不断更新多个模型的权重,使得预测效果好的模型获得更大的权重。每一次多个血糖预测模型迭代更新时,每个血糖预测模型的预测标准误差也相应更新,因此每个血糖预测模型对应的加权因子也会迭代更新。
根据本申请所述的血糖预测方法,其中,所述基于所述至少两个血糖预评估结果及其在所述当前场景下对应的加权因子组,得到血糖预测结果,包括:
将所述至少两个血糖预评估结果分别与所述当前场景下对应的加权因子组中对应的加权因子相乘后求和,得到与所述当前时刻对应的预设时间段之后的血糖预测结果。该预设时间段同上述第二时间段。
血糖预测结果的计算如下式所示:
Glu=Glu A*a j+Glu B*b j+Glu C*c j+…+Glu M*d j,j=1,2,3…n
Glu代表最终血糖浓度预测结果,a j,b j,c j…d j代表第j个实时场景对应的加权因子组,n表示场景总量,可以有多个场景,在本申请实施例中的第1至7个场景中,Glu A代表采用a模型血糖预评估结果,Glu B代表采用b模型血糖预评估结果,其余含义类似。多个场景下多个模型的加权因子如下表3所示:
表3
  场景1 场景2 场景n
A模型 a 1 a 2 a n
B模型 b 1 b 2 b n
C模型 c 1 c 2 c n
M模型 m 1 m 2 m n
A-M模型代表包含SVR模型,NN模型,ARMA模型和LSTM模型等在内的可以基于一定的数据进行数据预测的模型。
根据本申请所述的血糖预测方法,其中,所述基于所述至少两个血糖预评估结果及其在所述当前场景下对应的加权因子组,得到血糖预测结果之后,还包括:
利用至少一显示模块实现所述血糖预测结果的可视化;显示模块可以配置为显示以当前时刻为起点、第二时间段之后在第二时刻的血糖浓度预测值。
和/或,利用至少一告警模块基于预设血糖阈值实现所述血糖预测结果的告警提示。
告警模块可以与显示模块集成在同一个显示设备中,也可以独立设置,当血糖预测结果超出预设血糖阈值(结合用户设置和历史经验等因素设定),则对血糖浓度进行告警提示。
为了说明本申请的血糖预测方法,结合不同的当前场景,提供以下具体实施例。
患者的血糖数值主要受饮食(碳水化合物)和胰岛素的影响,患者通过接收设备,如移动应用程序记录摄入饮食的开始时间、摄入量以及胰岛素的注射时间和注射量,在设备的显示界面显示血糖历史曲线和饮食和胰岛素。
在第一个具体实施例中,混合模型包括:SVR模型,NN模型和LSTM模型。其中,SVR模型预测过程中的输入仅有过去2小时的血糖数据和对应的时间戳,而另外两种模型的输入可以包含血糖、饮食和胰岛素和其对应的时间戳。三种模型分别进行运算最终得到该患者未来一段时间的血糖预测值,时间的长度可以由用户根据需求进行设置,设置的时间间隔最小值为连续血糖监测设备产生血糖的间隔,如3分钟。一般情况下由于人体代谢复杂,且用户的未来行为很难预测,预测的最佳长度推荐为30分钟。
由于SVR模型未考虑饮食和胰岛素的影响,因此当出现用户记录饮食或胰 岛素的输入后,人体的血糖开始出现对应的上升或下降,通过训练的结果可得在一定时间内,SVR模型未能及时预测血糖上升或下降,因此,此时下式中的SVR血糖的加权因子a较小或者为0。
Glu=Glu SVR*a+Glu NN*b+Glu LSTM*c
在一种情况下,饮食和胰岛素的数据为用户手动输入的数值,根据用户的历史习惯,当饮食或胰岛素的数值过分偏离常识或用户习惯时,认为用户输入了不准确的信息。为了防止不准确的信息影响最终的预测结果,将NN模型和LSTM模型的预测结果的加权因子(b和c)减小或设置为0。
在一种情况下,胰岛素输入量可以由胰岛素输注设备通过第一网络传输至接收设备。
在一种情况下,饮食数据可以由自动识别食物图片中的热量的软件获得。
在第二个具体实施例中,SVR模型,NN模型和LSTM模型在训练过程中使用中国人的数据进行训练,训练的数据中包含血糖,饮食和胰岛素信息。以30分钟为目标,通过训练得到多个基本的模型。通过对比多个模型和实际值在不同时间的差异确定多个模型在不同时间和不同情况下的加权因子。差异通过RMSE进行评估。
如,在血糖平稳(每分钟血糖变化率小于0.016mmol/L)的情况下,以30分钟为目标时间,三种模型均有一定的差异,如三种模型和真实值的平均差异均在1mmol左右。但三种模型的预测值与真实值的大小关系不同,因此,通过尝试不同的加权因子得到一组加权因子:
Glu=Glu SVR*a 1+Glu NN*b 1+Glu LSTM*c 1
使得预测结果Glu与30分钟后的真实值相差最小。
在血糖快速波动,例如是由于饮食和胰岛素的摄入导致的快速波动的过程中,SVR模型的表现较差,则得到另外的加权因子组,最终得到加权因子集,该集合和三种模型一起被部署到接收设备中和云端服务器。
由于饮食结构和身体代谢的不同,利用其他地区的人群进行训练得到不同的模型和对应的加权因子集。在用户使用的过程中,选择自己所在的国家和地区来调用不同的模型和加权因子集进行预测。
在第三个具体实施例中,当出现用户记录的饮食、胰岛素等事件的时候,由于ARMA模型或SVR模型无法响应用户输入的事件,因此在用户数据产生的时候此两种模型的加权因子置零,另外两种模型的加权因子的确定,根据上述RMSE进行确定。
RMSE final=c 2*RMSE LSTM+d 2*RMSE NN
此时,设定c+d=1
RMSE final=c 2*RMSE LSTM+(1-c) 2*RMSE NN
根据上式可以获得RMSE final最小的时候的c值,且c的取值由RMSE LSTM和RMSE NN确定。上式中,c表示LSTM模型在该场景下的权重因子,d表示NN模型在该场景下的权重因子。
ARMA模型和SVR模型的加权因子恢复时间根据用户事件对血糖的影响减小的时间确定,如餐后2小时后饮食事件对血糖值的影响减小,加权因子的选择根据第一种情况的血糖波动区间选择。
当系统检测到错误的用户数据时,为了避免错误的数据传递到系统中,该场景下,可以将预测过程中会使用用户数据的LSTM模型和NN模型的加权因子置零,恢复时间根据血糖变化率或用户事件对血糖的影响减小的时间确定,如餐后2小时。用户数据是否错误由电子设备根据以往的用户的数据历史和经验进行判断。
当与传感器耦合的电子设备监测到传感器异常时,传感器得到的血糖结果已经不能用于未来的预测时,所有加权因子全部为零,即停止产生血糖预测,暂停血糖预测结果输出。待传感器恢复正常工作,且正常工作产生的数据量可以满足多个模型需求时再根据以上情况提供血糖预测。
参见图4,下面对本申请提供的血糖预测装置进行描述,下文描述的血糖预测装置与上文描述的血糖预测方法可相互对应参照,所述血糖预测装置包括:
血糖轨迹数据获取模块10,设置为获取用户的血糖轨迹数据,所述血糖轨迹数据包括当前血糖采集数据和历史血糖采集数据;血糖预评估模块20,设置为将所述用户的血糖轨迹数据分别输入至少两个血糖预测模型,得到所述至少两个血糖预测模型分别输出的至少两个血糖预评估结果;当前场景确定模块30,设置为基于由所述血糖轨迹数据确定的血糖浓度变化率和/或所述当前血糖采集数据的状态,确定所述用户所处的当前场景;加权因子组选取模块40,设置为基于所述当前场景,在加权因子集中选取所述至少两个血糖预测模型在所述当前场景下对应的加权因子组,其中,所述加权因子集包含基于多个实时场景进行分类的多个加权因子组,每个加权因子组包含至少两个加权因子,每个加权因子是基于预设的多模融合目标标准误差以及所述至少两个血糖预测模型的预测标准误差确定的;血糖预测模块50,设置为基于所述至少两个血糖预评估结果及所述至少两个血糖预测模型在所述当前场景下对应的加权因子组,得到血糖预测结果。
由于血糖预测装置与上述实施例中的血糖预测方法是一一对应的,上述血糖预测方法中的任一实施例同样适用于该血糖预测装置,此处不再一一赘述。
本申请还提供了一种监测血糖水平的系统,包括:
传感器,设置为获取用户的血糖测量值;无线发射器,设置为发射所述血糖测量值;以及移动计算装置,其包括:无线接收器,设置为接收所述血糖测量值;存储器,设置为存储包含所述血糖测量值的数据;处理器,设置为处理所述数据,以及软件应用程序,软件应用程序包含存储于所述存储器中的指令,所述指令当由所述处理器执行时以执行上述多个实施例的血糖预测方法。
图5是本申请实施例提供的一种电子设备的结构示意图,该电子设备可以包括:处理器(processor)510、通信接口(Communications Interface)520、存储器(memory)530和通信总线540,其中,处理器510,通信接口520,存储器530通过通信总线540完成相互间的通信。处理器510可以调用存储器530中的逻辑指令,以执行上述实施例血糖预测方法。
此外,上述的存储器530中的逻辑指令可以通过软件功能单元的形式实现并作为独立的产品销售或使用时,可以存储在一个计算机可读取存储介质中。本申请的技术方案本质上可以以软件产品的形式体现出来,该计算机软件产品存储在一个存储介质中,包括多个指令用以使得一台计算机设备(可以是个人计算机,服务器,或者网络设备等)执行本申请实施例所述方法的全部或部分步骤。而前述的存储介质包括:U盘、移动硬盘、只读存储器(Read-Only Memory,ROM)、随机存取存储器(Random Access Memory,RAM)、磁碟或者光盘等多种可以存储程序代码的介质。
另一方面,本申请还提供一种计算机程序产品,所述计算机程序产品包括存储在非暂态计算机可读存储介质上的计算机程序,所述计算机程序包括程序指令,当所述程序指令被计算机执行时,计算机能够执行上述实施例所提供的血糖预测方法。
又一方面,本申请还提供一种非暂态计算机可读存储介质,其上存储有计算机程序,该计算机程序被处理器执行时实现以执行上述实施例所提供的血糖预测方法。
以上所描述的装置实施例仅仅是示意性的,其中所述作为分离部件说明的单元可以是或者也可以不是物理上分开的,作为单元显示的部件可以是或者也可以不是物理单元,即可以位于一个地方,或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部模块来实现本实施例方案的目的。
通过以上的实施方式的描述,实施方式可借助软件加必需的通用硬件平台 的方式来实现,也可以通过硬件。上述技术方案本质上可以以软件产品的形式体现出来,该计算机软件产品可以存储在计算机可读存储介质中,如ROM/RAM、磁碟、光盘等,包括多个指令用以使得一台计算机设备(可以是个人计算机,服务器,或者网络设备等)执行实施例或者实施例的一些部分所述的方法。

Claims (18)

  1. 一种血糖预测方法,包括:
    获取用户的血糖轨迹数据,其中,所述血糖轨迹数据包括当前血糖采集数据和历史血糖采集数据;
    将所述用户的血糖轨迹数据输入至少两个血糖预测模型,得到所述至少两个血糖预测模型分别输出的至少两个血糖预评估结果;
    基于由所述血糖轨迹数据确定的血糖浓度变化率和所述当前血糖采集数据的状态中的至少之一,确定所述用户所处的当前场景;
    基于所述当前场景,在加权因子集中选取所述至少两个血糖预测模型在所述当前场景下对应的加权因子组,其中,所述加权因子集包含基于多个实时场景进行分类的多个加权因子组,每个加权因子组包含至少两个加权因子,每个加权因子是基于预设的多模融合目标标准误差以及所述至少两个血糖预测模型的预测标准误差确定的;
    基于所述至少两个血糖预评估结果及所述至少两个血糖预测模型在所述当前场景下对应的加权因子组,得到血糖预测结果。
  2. 根据权利要求1所述的血糖预测方法,其中,所述获取用户的血糖轨迹数据,其中,所述血糖轨迹数据包括当前血糖采集数据和历史血糖采集数据,包括:
    通过网络从关联所述用户的血糖测量设备处获取所述用户的血糖轨迹数据;
    其中,所述当前血糖采集数据包括当前时刻的第一血糖测量值及所述第一血糖测量值对应的第一时间戳,所述历史血糖采集数据包括按预设时间间隔连续分布的多个历史血糖测量值及所述多个历史血糖测量值分别对应的多个历史时间戳。
  3. 根据权利要求2所述的血糖预测方法,其中,所述获取用户的血糖轨迹数据,其中,所述血糖轨迹数据包括当前血糖采集数据和历史血糖采集数据,包括:
    获取用户输入的血糖轨迹数据;
    所述当前血糖采集数据包括用户输入的当前时刻的第一血糖关联数据及所述第一血糖关联数据对应的第一关联时间戳;所述历史血糖采集数据包括用户输入的历史时刻的多个历史血糖关联数据及所述多个历史血糖关联数据分别对应的多个历史关联时间戳,所述第一血糖关联数据、所述历史血糖关联数据分别包括与血糖浓度相关联的至少一个事件。
  4. 根据权利要求3所述的血糖预测方法,其中,所述至少一个事件与食物消 耗、饮料消耗、锻炼、睡眠以及物质的施予中的至少一者相关联。
  5. 根据权利要求1所述的血糖预测方法,其中,每个血糖预测模型按照以下方式进行训练:
    基于血糖浓度样本数据以及预先确定的真实血糖浓度标签进行训练,得到所述每个血糖预测模型;或者,
    周期性地基于用户迭代更新的血糖浓度样本数据以及所述用户迭代更新的血糖浓度样本数据对应的真实血糖浓度标签进行训练,得到所述每个血糖预测模型。
  6. 根据权利要求1所述的血糖预测方法,其中,所述用户的血糖轨迹数据还包括所述用户的所在区域;
    每个血糖预测模型按照以下方式进行训练:
    基于所述用户所在区域的血糖浓度样本数据以及所述血糖浓度样本数据对应的预先确定的真实血糖浓度标签进行训练,得到所述每个血糖预测模型。
  7. 根据权利要求1所述的血糖预测方法,其中,所述用户的血糖轨迹数据还包括所述用户所属的糖尿病类型;
    每个血糖预测模型按照以下方式进行训练:
    基于所述用户所属糖尿病类型的血糖浓度样本数据以及所述血糖浓度样本数据对应的预先确定的真实血糖浓度标签进行训练,得到所述每个血糖预测模型。
  8. 根据权利要求3所述的血糖预测方法,在所述基于由所述血糖轨迹数据确定的血糖浓度变化率和所述当前血糖采集数据的状态中的至少之一,确定所述用户所处的当前场景之前,还包括:
    基于所述第一血糖测量值及所述第一血糖测量值对应的第一时间戳、在所述多个历史血糖测量值中选取的第二血糖测量值、以及在所述多个历史血糖测量值分别对应的多个历史时间戳中选取的所述第二血糖测量值对应的第二时间戳,确定所述血糖浓度变化率;其中,所述第二时间戳与所述第一时间戳相关联。
  9. 根据权利要求8所述的血糖预测方法,其中,所述基于由所述血糖轨迹数据确定的血糖浓度变化率和所述当前血糖采集数据的状态中的至少之一,确定所述用户所处的当前场景,包括:
    在所述血糖浓度变化率不大于第一预设阈值的情况下,确定所述用户处于平稳区间场景;
    在所述血糖浓度变化率大于第一预设阈值、且不大于第二预设阈值的情况下,确定所述用户处于缓速升降区间场景;
    在所述血糖浓度变化率大于第二预设阈值、且不大于第三预设阈值的情况下,确定所述用户处于中速升降区间场景;
    在所述血糖浓度变化率大于第三预设阈值的情况下,确定所述用户处于快速升降区间场景;
    其中,所述第三预设阈值大于所述第二预设阈值,所述第二预设阈值大于所述第一预设阈值。
  10. 根据权利要求9所述的血糖预测方法,其中,所述基于由所述血糖轨迹数据确定的血糖浓度变化率和所述当前血糖采集数据的状态中的至少之一,确定所述用户所处的当前场景,包括:
    在从关联所述用户的血糖测量设备处获取的用户的血糖轨迹数据异常的情况下,确定所述用户处于血糖测量值异常场景;
    在所述用户输入的当前时刻的第一血糖关联数据异常的情况下,确定所述用户处于用户输入异常场景;
    在所述用户输入的当前时刻的第一血糖关联数据正常的情况下,确定所述用户处于用户输入正常场景;
    其中,在确定所述用户所处的当前场景的情况下,所述血糖测量值异常场景为第一优先级,所述用户输入异常场景和所述用户输入正常场景为第二优先级,所述平稳区间场景、所述缓速升降区间场景、所述中速升降区间场景以及所述快速升降区间场景为第三优先级,所述第一优先级大于所述第二优先级,所述第二优先级大于所述第三优先级。
  11. 根据权利要求1所述的血糖预测方法,其中,每个加权因子组包含至少两个加权因子,每个加权因子是基于预设的多模融合目标标准误差以及所述至少两个血糖预测模型的预测标准误差确定的,包括:
    设定所述至少两个血糖预测模型在第j个实时场景下的至少两个加权因子分别为a j,…,b j,且满足a j+…+b j=1,j为大于或等于1的正整数;
    所述至少两个血糖预测模型的预测标准误差RMSE final的计算公式为:
    Figure PCTCN2022085450-appb-100001
    Figure PCTCN2022085450-appb-100002
    表示至少两个血糖预测模型中的a模型在j场景下的预测标准误差,
    Figure PCTCN2022085450-appb-100003
    表示至少两个血糖预测模型中的b模型在j场景下的预测标准误差,每个血糖预测模型的预测标准误差RMSE的计算公式为:
    Figure PCTCN2022085450-appb-100004
    其中,m表示所述每个血糖预测模型的血糖浓度样本数据的总量,i表示从1至m的变量,i为大于或等于1的正整数,y i表示所述每个血糖预测模型的第i个血糖预评估结果,
    Figure PCTCN2022085450-appb-100005
    表示所述每个血糖预测模型的第i个真实血糖浓度标签;
    选取使RMSE final小于所述预设的多模融合目标标准误差或使RMSE final达到最小值的一组a j,…,b j作为所述至少两个血糖预测模型分别在所述第j个实时场景下的多个加权因子。
  12. 根据权利要求1所述的血糖预测方法,其中,
    每个加权因子是基于所述预设的多模融合目标标准误差以及所述至少两个血糖预测模型的预测标准误差确定并周期性地迭代更新的。
  13. 根据权利要求1所述的血糖预测方法,其中,所述基于所述至少两个血糖预评估结果及所述至少两个血糖预测模型在所述当前场景下对应的加权因子组,得到血糖预测结果,包括:
    将所述至少两个血糖预评估结果分别与所述加权因子组中对应的加权因子相乘后求和,得到与当前时刻对应的预设时间段之后的血糖预测结果。
  14. 根据权利要求1所述的血糖预测方法,在所述基于所述至少两个血糖预评估结果及所述至少两个血糖预测模型在所述当前场景下对应的加权因子组,得到血糖预测结果之后,还包括以下至少之一:
    利用至少一显示模块实现所述血糖预测结果的可视化;
    利用至少一告警模块基于预设血糖阈值实现所述血糖预测结果的告警提示。
  15. 一种血糖预测装置,包括:
    血糖轨迹数据获取模块,设置为获取用户的血糖轨迹数据,其中,所述血糖轨迹数据包括当前血糖采集数据和历史血糖采集数据;
    血糖预评估模块,设置为将所述用户的血糖轨迹数据输入至少两个血糖预测模型,得到所述至少两个血糖预测模型分别输出的至少两个血糖预评估结果;
    当前场景确定模块,设置为基于由所述血糖轨迹数据确定的血糖浓度变化率和所述当前血糖采集数据的状态中的至少之一,确定所述用户所处的当前场景;
    加权因子组选取模块,设置为基于所述当前场景,在加权因子集中选取所述至少两个血糖预测模型在所述当前场景下对应的加权因子组,其中,所述加权因子集包含基于多个实时场景进行分类的多个加权因子组,每个加权因子组 包含至少两个加权因子,每个加权因子是基于预设的多模融合目标标准误差以及所述至少两个血糖预测模型的预测标准误差确定的;
    血糖预测模块,设置为基于所述至少两个血糖预评估结果及所述至少两个血糖预测模型在所述当前场景下对应的加权因子组,得到血糖预测结果。
  16. 一种监测血糖水平的系统,包括:
    传感器,设置为获取用户的血糖测量值;
    无线发射器,设置为发射所述血糖测量值;
    以及
    移动计算装置,包括:
    无线接收器,设置为接收所述血糖测量值;
    存储器,设置为存储包含所述血糖测量值的数据;
    处理器,设置为处理所述存储器存储的数据,以及软件应用程序,所述软件应用程序包含存储于所述存储器中的指令,所述指令执行时实现如权利要求1至14任一项所述的血糖预测方法。
  17. 一种电子设备,包括存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序,其中,所述处理器执行所述程序时实现如权利要求1至14任一项所述的血糖预测方法。
  18. 一种非暂态计算机可读存储介质,存储有计算机程序,其中,所述计算机程序被处理器执行时实现如权利要求1至14任一项所述的血糖预测方法。
PCT/CN2022/085450 2021-12-21 2022-04-07 血糖预测方法和装置、监测血糖水平的系统 WO2023115751A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111565948.8 2021-12-21
CN202111565948.8A CN113951879B (zh) 2021-12-21 2021-12-21 血糖预测方法和装置、监测血糖水平的系统

Publications (1)

Publication Number Publication Date
WO2023115751A1 true WO2023115751A1 (zh) 2023-06-29

Family

ID=79473312

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/085450 WO2023115751A1 (zh) 2021-12-21 2022-04-07 血糖预测方法和装置、监测血糖水平的系统

Country Status (2)

Country Link
CN (1) CN113951879B (zh)
WO (1) WO2023115751A1 (zh)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113951879B (zh) * 2021-12-21 2022-04-05 苏州百孝医疗科技有限公司 血糖预测方法和装置、监测血糖水平的系统
CN114145738B (zh) * 2022-02-10 2022-06-24 苏州百孝医疗科技有限公司 分析物浓度数据生成方法和装置、监测分析物水平的系统
CN115862795A (zh) * 2022-09-22 2023-03-28 深圳可孚生物科技有限公司 一种基于预训练和数据分解两阶段的血糖预测方法
CN118078275A (zh) * 2024-04-29 2024-05-28 深圳市爱都科技有限公司 无创血糖检测的方法、装置、电子设备及介质

Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20090312621A1 (en) * 2006-08-08 2009-12-17 Koninklijke Philips Electronics N.V. Method and device for monitoring a physiological parameter
US20110040487A1 (en) * 2007-10-12 2011-02-17 Roman Hovorka Substance monitoring and control in human or animal bodies
CN104769595A (zh) * 2012-08-30 2015-07-08 美敦力迷你迈德公司 用于闭环胰岛素输注系统的防护技术
CN105190316A (zh) * 2013-02-21 2015-12-23 弗吉尼亚大学专利基金会 跟踪糖尿病中的平均血糖的变化
CN106415556A (zh) * 2014-04-10 2017-02-15 德克斯康公司 血糖紧迫性评估和警告界面
CN108292525A (zh) * 2015-08-21 2018-07-17 美敦力迷你迈德公司 用于糖尿病管理和控制的数据分析和认知递送
CN110650683A (zh) * 2017-03-24 2020-01-03 美敦力泌力美公司 患者专用的葡萄糖预测系统和方法
US20200383645A1 (en) * 2013-11-28 2020-12-10 Roche Diabetes Care, Inc. Method and device for analyzing continuously monitored physiological measurement values of a user
US20210077719A1 (en) * 2019-09-13 2021-03-18 Insulet Corporation Blood glucose rate of change modulation of meal and correction insulin bolus quantity
CN113951879A (zh) * 2021-12-21 2022-01-21 苏州百孝医疗科技有限公司 血糖预测方法和装置、监测血糖水平的系统

Family Cites Families (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
GB2443434A (en) * 2006-11-02 2008-05-07 Richard Butler Method for predicting nocturnal hypoglycaemia
JP2013501558A (ja) * 2009-08-10 2013-01-17 ディアベテス トールス スウェーデン アーべー 1組のデータ値を処理する装置および方法
US9119529B2 (en) * 2012-10-30 2015-09-01 Dexcom, Inc. Systems and methods for dynamically and intelligently monitoring a host's glycemic condition after an alert is triggered
US20180353112A1 (en) * 2017-06-09 2018-12-13 President And Fellows Of Harvard College Prevention of post-bariatric hypoglycemia using a novel glucose prediction algorithm and mini-dose stable glucagon
CN108766578A (zh) * 2018-05-16 2018-11-06 清华大学深圳研究生院 一种血糖预测方法及装置
CN113614850A (zh) * 2018-09-07 2021-11-05 明智数据系统公司(d/b/a一滴公司) 预测血糖浓度
CN110197724A (zh) * 2019-03-12 2019-09-03 平安科技(深圳)有限公司 预测糖尿病患病阶段的方法、装置及计算机设备
CN110085318A (zh) * 2019-03-12 2019-08-02 平安科技(深圳)有限公司 预测未来血糖值的方法、装置及计算机设备
CN109998560B (zh) * 2019-04-30 2023-12-22 苏州百孝医疗科技有限公司 分离供电动态血糖监测发射器、系统及信号采样方法
WO2020243576A1 (en) * 2019-05-31 2020-12-03 Ydo Wexler Systems for biomonitoring and blood glucose forecasting, and associated methods
CN111329491A (zh) * 2020-02-27 2020-06-26 京东方科技集团股份有限公司 一种血糖预测方法、装置、电子设备和存储介质
CN111631729B (zh) * 2020-05-14 2023-06-06 中国科学院深圳先进技术研究院 一种基于多模融合的低血糖预测方法和系统
CN111631730B (zh) * 2020-05-15 2023-06-06 中国科学院深圳先进技术研究院 基于传感数据和生理信息融合的低血糖预警方法和系统
CN111631704B (zh) * 2020-05-29 2021-06-25 中国科学院深圳先进技术研究院 基于心电与脑电信息结合的糖尿病前期检测系统和方法
CN112402731B (zh) * 2020-10-10 2023-05-26 广东食品药品职业学院 一种预防低血糖现象的闭环胰岛素输注系统
CN113380411B (zh) * 2021-07-19 2024-03-01 苏州百孝医疗科技有限公司 一种提高动物体分析物浓度连续监测过程中浓度变化实时趋势准确率的方法

Patent Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20090312621A1 (en) * 2006-08-08 2009-12-17 Koninklijke Philips Electronics N.V. Method and device for monitoring a physiological parameter
US20110040487A1 (en) * 2007-10-12 2011-02-17 Roman Hovorka Substance monitoring and control in human or animal bodies
CN104769595A (zh) * 2012-08-30 2015-07-08 美敦力迷你迈德公司 用于闭环胰岛素输注系统的防护技术
CN105190316A (zh) * 2013-02-21 2015-12-23 弗吉尼亚大学专利基金会 跟踪糖尿病中的平均血糖的变化
US20200383645A1 (en) * 2013-11-28 2020-12-10 Roche Diabetes Care, Inc. Method and device for analyzing continuously monitored physiological measurement values of a user
CN106415556A (zh) * 2014-04-10 2017-02-15 德克斯康公司 血糖紧迫性评估和警告界面
CN108292525A (zh) * 2015-08-21 2018-07-17 美敦力迷你迈德公司 用于糖尿病管理和控制的数据分析和认知递送
CN110650683A (zh) * 2017-03-24 2020-01-03 美敦力泌力美公司 患者专用的葡萄糖预测系统和方法
US20210077719A1 (en) * 2019-09-13 2021-03-18 Insulet Corporation Blood glucose rate of change modulation of meal and correction insulin bolus quantity
CN113951879A (zh) * 2021-12-21 2022-01-21 苏州百孝医疗科技有限公司 血糖预测方法和装置、监测血糖水平的系统

Also Published As

Publication number Publication date
CN113951879A (zh) 2022-01-21
CN113951879B (zh) 2022-04-05

Similar Documents

Publication Publication Date Title
WO2023115751A1 (zh) 血糖预测方法和装置、监测血糖水平的系统
McLellan et al. Wear compliance, sedentary behaviour and activity in free-living children from hip-and wrist-mounted ActiGraph GT3X+ accelerometers
AU2021245245B2 (en) System, device and method of dynamic glucose profile response to physiological parameters
US8182424B2 (en) Diary-free calorimeter
WO2018231302A1 (en) Techniques for providing customized exercise-related recommendations
CN107423549A (zh) 健康跟踪设备
KR20210110284A (ko) 혈당 농도 예보
Zhu Automatic detection of anomalies in blood glucose using a machine learning approach
RU2712395C1 (ru) Способ выдачи рекомендаций по поддержанию здорового образа жизни на основе параметров ежедневной деятельности пользователя, автоматически отслеживаемых в реальном времени, и соответствующая система (варианты)
US20210241916A1 (en) Forecasting and explaining user health metrics
CN111588384B (zh) 获得血糖检测结果的方法、装置及设备
CN107658001A (zh) 一种家庭用油健康管理方法和系统
CN113380411B (zh) 一种提高动物体分析物浓度连续监测过程中浓度变化实时趋势准确率的方法
KR20190063954A (ko) 인공지능 클라우드 영양소 변화 예측 방법과 인공지능 클라우드 영양소 대사량 측정 방법 및 그를 사용하는 질병관리 방법
CN114023418A (zh) 胰岛素推荐方法和装置、监控血糖水平的系统
US20220005580A1 (en) Method for providing recommendations for maintaining a healthy lifestyle basing on daily activity parameters of user, automatically tracked in real time, and corresponding system
Chew et al. The potential of artificial intelligence in enhancing adult weight loss: a scoping review
WO2021121226A1 (zh) 一种心电信号的预测方法、装置、终端以及存储介质
CN106446543A (zh) 健康数据处理方法、装置及服务器集群
Coyne et al. Heart rate variability and direct current measurement characteristics in professional mixed martial arts athletes
TWI510948B (zh) Health activity plan establishment system, establishment method and computer program product
Kalatzis et al. Interactive dimensionality reduction for improving patient adherence in remote health monitoring
WO2023151213A1 (zh) 分析物浓度数据生成方法和装置、监测分析物浓度的系统
CN115005787A (zh) 一种基于智能手表的卡路里管理系统
CN113569140A (zh) 信息推荐方法、装置、电子设备及计算机可读存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22909089

Country of ref document: EP

Kind code of ref document: A1