Machine learning-based system for estimating glucose values

Abstract

An optimized population model that estimates blood glucose values for a population of users is generated by mapping received data for the population of users over a time period to a sequence of estimated blood glucose values for the population of users over the time period. Discrete blood glucose measurement data for each user, user activity data for each user, and other contextual data for each user can be processed via a supervised machine learning model to learn a transfer function for a population model that estimates blood glucose values for the population of users. One or more parameters of the learning model can be adjusted to generate the optimized population model.

1 . A method, comprising:
 receiving data for each user within a population of users from a number of different input channels over a time period, the received data for each user comprising: discrete blood glucose measurement data for that user; user activity data for that user; and other contextual data for that user;   processing the received data for the population of users to generate an input data set;   processing the input data set, via a supervised machine learning model, to learn a transfer function for a population model that estimates blood glucose values for the population of users by mapping the received data for the population of users over the time period to a sequence of estimated blood glucose values for the population of users over the time period, wherein the supervised machine learning model has parameters that are estimated to generate the transfer function for the population model; and   performing a parameter optimization process by adjusting one or more parameters of the supervised machine learning model to generate an optimized population model that estimates blood glucose values for the population of users by mapping the received data for the population of users over the time period to the sequence of estimated blood glucose values for the population of users over the time period.   
     
     
         2 . The method of  claim 1 , wherein performing the parameter optimization process comprises:
 iteratively adjusting one or more parameters of the supervised machine learning model;   continuously evaluating an objective function that measures a level of mathematical agreement between the estimated blood glucose values output by the machine learning model and actual measured blood glucose levels; and   generating the optimized population model for the population of users when the level of mathematical agreement between the estimated blood glucose values output by the machine learning model and the actual measured blood glucose levels reaches a desired threshold,   wherein the optimized population model for the population of users has the parameters of the transfer function set to values that were being evaluated when the mathematical agreement reached the desired threshold.   
     
     
         3 . The method of  claim 2 , wherein the optimized population model is an optimized window population model, the method further comprising:
 applying a window filter to the received data, prior to processing the received data for the population of users, to split the received data into a series of different time windows each having a period that is less than the time period that the received data was received over, wherein each time window includes a discrete time segment of the received data; and   wherein processing the received data for the population of users to generate the input data set, comprises:   processing each discrete time segment of the received data to generate a corresponding input data set based on the received data received over a particular time window; and   wherein processing the input data set, comprises:   processing each corresponding input data set, via the supervised machine learning model, to learn a corresponding transfer function for a corresponding population model that estimates blood glucose values for the population of users by mapping the received data for the population of users over the particular time window to the sequence of estimated blood glucose values for the population of users over the particular time window, wherein the supervised machine learning model has parameters that are estimated to generate each corresponding transfer function for the population model.   
     
     
         4 . The method of  claim 3 , further comprising:
 selecting the corresponding population model having a lowest error as the optimized population model.   
     
     
         5 . The method of  claim 3 , wherein the optimized window population model is an optimized extended window population model, the method further comprising:
 sequentially joining each of the sequences of estimated blood glucose values for the population of users for each particular time window to generate a joined set of estimated blood glucose values over a number of the particular time windows to generate the optimized extended window population model that estimates the blood glucose values for the population of users.   
     
     
         6 . The method of  claim 1 , wherein the discrete blood glucose measurement data for each user is measured using a sensor arrangement that provides discrete blood glucose measurements for that user, wherein the user activity data for each user is provided from a source of user activity data that correlates to activity of that user, and wherein the other contextual data for each user is provided from at least one other source associated with that user and provides other information about that user. 
     
     
         7 . The method of  claim 1 , wherein each input channel comprises a different data type being measured for that user over the time period. 
     
     
         8 . A system, comprising:
 one or more hardware-based processors configured by machine-readable instructions to:
 receive data for each user within a population of users from a number of different input channels over a time period, the received data for each user comprising: discrete blood glucose measurement data for that user; user activity data for that user; and other contextual data for that user; 
 process the received data for the population of users to generate an input data set; 
 process the input data set, via a supervised machine learning model, to learn a transfer function for a population model that estimates blood glucose values for the population of users by mapping the received data for the population of users over the time period to a sequence of estimated blood glucose values for the population of users over the time period, wherein the supervised machine learning model has parameters that are estimated to generate the transfer function for the population model; and 
 perform a parameter optimization process by adjusting one or more parameters of the supervised machine learning model to generate an optimized population model that estimates blood glucose values for the population of users by mapping the received data for the population of users over the time period to the sequence of estimated blood glucose values for the population of users over the time period. 
   
     
     
         9 . The system according to  claim 8 , wherein the one or more hardware-based processors are further configured by machine-readable instructions to:
 iteratively adjust one or more parameters of the supervised machine learning model;   continuously evaluate an objective function that measures a level of mathematical agreement between the estimated blood glucose values output by the machine learning model and actual measured blood glucose levels; and   generate the optimized population model for the population of users when the level of mathematical agreement between the estimated blood glucose values output by the machine learning model and the actual measured blood glucose levels reaches a desired threshold, wherein the optimized population model for the population of users has the parameters of the transfer function set to values that were being evaluated when the mathematical agreement reached the desired threshold.   
     
     
         10 . The system according to  claim 9 , wherein the optimized population model is an optimized window population model, and wherein the one or more hardware-based processors are further configured by machine-readable instructions to:
 apply a window filter to the received data, prior to processing the received data for the population of users, to split the received data into a series of different time windows each having a period that is less than the time period that the received data was received over, wherein each time window includes a discrete time segment of the received data; and   process each discrete time segment of the received data to generate a corresponding input data set based on the received data received over a particular time window; and   process each corresponding input data set, via the supervised machine learning model, to learn a corresponding transfer function for a corresponding population model that estimates blood glucose values for the population of users by mapping the received data for the population of users over the particular time window to the sequence of estimated blood glucose values for the population of users over the particular time window, wherein the supervised machine learning model has parameters that are estimated to generate each corresponding transfer function for the population model.   
     
     
         11 . The system according to  claim 10 , wherein the one or more hardware-based processors are further configured by machine-readable instructions to
 select the corresponding population model having a lowest error as the optimized population model.   
     
     
         12 . The system according to  claim 10 , wherein the optimized window population model is an optimized extended window population model, and wherein the one or more hardware-based processors are further configured by machine-readable instructions to:
 sequentially join each of the sequences of estimated blood glucose values for the population of users for each particular time window to generate a joined set of estimated blood glucose values over a number of the particular time windows to generate the optimized extended window population model that estimates the blood glucose values for the population of users.   
     
     
         13 . The system according to  claim 14 , the discrete blood glucose measurement data for each user is measured using a sensor arrangement that provides discrete blood glucose measurements for that user, wherein the user activity data for each user is provided from a source of user activity data that correlates to activity of that user, and wherein the other contextual data for each user is provided from at least one other source associated with that user and provides other information about that user. 
     
     
         14 . At least one non-transient computer-readable medium having instructions stored thereon that are configurable to cause at least one processor to perform a method, the method comprising:
 receiving data for each user within a population of users from a number of different input channels over a time period, the received data for each user comprising: discrete blood glucose measurement data for that user; user activity data for that user; and other contextual data for that user;   processing the received data for the population of users to generate an input data set;   processing the input data set, via a supervised machine learning model, to learn a transfer function for a population model that estimates blood glucose values for the population of users by mapping the received data for the population of users over the time period to a sequence of estimated blood glucose values for the population of users over the time period, wherein the supervised machine learning model has parameters that are estimated to generate the transfer function for the population model; and   performing a parameter optimization process by adjusting one or more parameters of the supervised machine learning model to generate an optimized population model that estimates blood glucose values for the population of users by mapping the received data for the population of users over the time period to the sequence of estimated blood glucose values for the population of users over the time period.   
     
     
         15 . The computer-readable medium according to  claim 14 , the method further comprising:
 iteratively adjusting one or more parameters of the supervised machine learning model;   continuously evaluating an objective function that measures a level of mathematical agreement between the estimated blood glucose values output by the machine learning model and actual measured blood glucose levels; and   generating the optimized population model for the population of users when the level of mathematical agreement between the estimated blood glucose values output by the machine learning model and the actual measured blood glucose levels reaches a desired threshold,   wherein the optimized population model for the population of users has the parameters of the transfer function set to values that were being evaluated when the mathematical agreement reached the desired threshold.   
     
     
         16 . The computer-readable medium according to  claim 15 , wherein the optimized population model is an optimized window population model, the method further comprising:
 applying a window filter to the received data, prior to processing the received data for the population of users, to split the received data into a series of different time windows each having a period that is less than the time period that the received data was received over, wherein each time window includes a discrete time segment of the received data; and   wherein processing the received data for the population of users to generate the input data set, comprises:   processing each discrete time segment of the received data to generate a corresponding input data set based on the received data received over a particular time window; and   wherein processing the input data set, comprises:   processing each corresponding input data set, via the supervised machine learning model, to learn a corresponding transfer function for a corresponding population model that estimates blood glucose values for the population of users by mapping the received data for the population of users over the particular time window to the sequence of estimated blood glucose values for the population of users over the particular time window, wherein the supervised machine learning model has parameters that are estimated to generate each corresponding transfer function for the population model.   
     
     
         17 . The computer-readable medium according to  claim 16 , the method further comprising:
 selecting the corresponding population model having a lowest error as the optimized population model.   
     
     
         18 . The computer-readable medium according to  claim 16 , wherein the optimized window population model is an optimized extended window population model, the method further comprising:
 sequentially joining each of the sequences of estimated blood glucose values for the population of users for each particular time window to generate a joined set of estimated blood glucose values over a number of the particular time windows to generate the optimized extended window population model that estimates the blood glucose values for the population of users.   
     
     
         19 . The computer-readable medium according to  claim 14 , the discrete blood glucose measurement data for each user is measured using a sensor arrangement that provides discrete blood glucose measurements for that user, wherein the user activity data for each user is provided from a source of user activity data that correlates to activity of that user, and wherein the other contextual data for each user is provided from at least one other source associated with that user and provides other information about that user. 
     
     
         20 . The computer-readable medium according to  claim 14 , wherein each input channel comprises a different data type being measured for that user over the time period.