METHOD TO DETERMINE INDIVIDUALIZED INSULIN SENSITIVITY AND OPTIMAL INSULIN DOSE BY LINEAR REGRESSION, AND RELATED SYSTEMS

Abstract

This invention relates to a method and a device for predicting the glucose concentration of a subject and recommending therapeutic action. The responses of the user's glucose to administered doses of insulin, dietary carbohydrates, and other factors influencing glucose concentration are measured individually for a given user. Once these responses are learned as a function of time, the method and device can receive information about the factors that have been recently or will soon be administered and can recommend which other factors should also be administered.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application is a continuation of U.S. patent application Ser. No. 14/947,347, filed Nov. 20, 2015, pending, which application claims the benefit of Provisional Patent Application Ser. No. 62/083,191, filed Nov. 22, 2014, and titled "Method to Determine Individualized Insulin Sensitivity and Optimal Insulin Dose by Linear Regression," the disclosure of each of which is hereby incorporated herein in its entirety by this reference. This invention was not made with any government support and the government has no rights in this invention.

TECHNICAL FIELD

[0002] The present invention relates to a method for predicting the glucose concentration of a subject and recommending therapeutic action.

BACKGROUND

[0003] In the United States, approximately three million people have type 1 diabetes..sup.[1] To treat their condition, these patients depend either on multiple daily injections (MDI) of insulin or continuous subcutaneous insulin injection (CSII) by insulin pumps. Another 26 million people in the United States suffer from type 2 diabetes, many of whom become insulin dependent. In total, nearly six million Americans depend on insulin..sup.[2] For these patients, choosing the correct dose and type of insulin to take, and when to take it, remains a significant challenge. The goal of an insulin regimen is to maintain blood glucose concentration within a narrow range. Chronically high glucose levels, hyperglycemia, leads to severe health problems and premature death. Acute low glucose levels, hypoglycemia, can cause fainting, seizures, and death. Patients on MDI maintain a healthy glucose level by typically taking five to ten injections per day. These may include injection of a long-acting form of insulin before going to sleep and/or after waking, as well as short-acting insulin injections both before and after every meal and snack, with those following meals chosen to correct for prior insufficient doses. In order to make informed dosage decisions, most diabetics today monitor their blood glucose with small blood draws from a finger prick before each injection. An increasing number of diabetics also monitor their glucose using a continuous glucose monitor (CGM), which provides glucose readings with much higher frequency, commonly around once every five minutes. Although it has already been appreciated that this high frequency data can be used to alert users to problematic glucose levels as they arise, the wealth of data provided by these machines has, so far, been underutilized for the challenge of determining optimal insulin doses.

[0004] An "artificial pancreas" is a heavily researched future treatment device for diabetes that uses a closed loop between a CGM and an insulin pump. Given frequent, accurate readings from a CGM, the insulin pump should be able to determine, without human intervention, how much insulin to give. This will require the development of an algorithm that can precisely calculate the response of blood glucose to insulin.

[0005] Previous inventions in the field of glucose prediction have used the time-series of glucose over a narrow window in the recent past, possibly around 30 minutes, to predict glucose over a similar time frame in the future, without directly accounting for external factors like insulin, dietary carbohydrates, and exercise. The previous inventions also often focus on making universal predictions of insulin response that are not individualized to each patient and they predict a single, time-integrated response rather than response as a function of time.

[0006] U.S. Patent Publication No. 2011/0160555 to Reifman, published Jun. 30, 2011, for "Universal models for predicting glucose concentration in humans" "utilizes similarities in the short-term (30 minutes or less) dynamics of glucose regulation in different diabetic individuals to develop a single, universal autoregressive (AR) model for predicting future glucose levels across different patients." This patent does not account for external factors that cause changes in glucose or attempt to learn individualized models for different patients. Similarly, U.S. Pat. No. 9,076,107, issued Jul. 7, 2015, to Cameron et al. for "Neural network for glucose therapy recommendation" uses recent glucose trends to predict future glucose trends, with the model trained on data from multiple patients rather than individualized. U.S. Pat. No. 7,695,434, issued Apr. 13, 2010, to Malecha for "Medical device for predicting a user's future glycemic state" uses CGM data to predict future glycemic state using a Hidden Markov Model and not linear regression. This method does not use information other than the time series of glucose before the moment of prediction to predict future glucose.

[0007] U.S. Pat. No. 7,404,796, issued Jul. 29, 2008, to Ginsberg for "System for determining insulin dose using carbohydrate to insulin ratio and insulin sensitivity factor" is a method for finding individualized carbohydrate-to-insulin ratios (CIRs) and insulin sensitivity factors (ISFs). At its greatest level of detail, that method gives the integrated effect of a unit of insulin or a carbohydrate on a user's glucose concentration, and not the effect as a function of time. That method also does not use linear regression.

[0008] U.S. Patent Publication No. 2014/0073892, published Mar. 13, 2014, to Jette Randloev et al., for "Glucose predictor based on regularization networks with adaptively chosen kernels and regularization parameters" describes a method for predicting glucose based on data sets that are sparsely sampled in time, which does not make use of the glucose time series data made available by CGMs and the insulin time series data made available by insulin pumps. That method uses regularization networks with adaptively chosen kernels and regularization parameters and not linear regression.

BRIEF SUMMARY

[0009] This patent is for a linear regression-based method to learn a user's blood glucose response to short-acting insulin, long-acting insulin, dietary carbohydrates, and lifestyle factors such as exercise. This method learns a unique response for each user. The calculated response curve is a function of time, not merely the total time-integrated effect of 1 unit of a given factor. After obtaining these response functions, in a second step, this method determines optimal insulin dosages in response to the user's relevant activities. This separate step allows the method to assign an asymmetrical cost function for hyperglycemia and hypoglycemia. This asymmetry is necessary because a positive fluctuation that results in only mild hyperglycemia and no immediate problems could, if only the sign of the fluctuation were reversed, cause severe hypoglycemia and death: Including this information in the same step as learning the user's response to insulin and carbohydrates would distort the user's true response. From the time series of a given user's glucose, insulin, exercise, and carbohydrate intake, the method learns the optimal dose and time of long-acting insulin, the optimal ratio of short-acting insulin to dietary carbohydrates, the optimal correction dose of short-acting insulin for hyperglycemia, and the typical effect of exercise on the user's blood glucose.

[0010] The two steps, learning the user's response curves and suggesting the optimal insulin dose, can be decoupled. Once the time dependent response curves are calculated, they can be integrated to determine the cumulative effect of each kind of insulin and dietary carbohydrates. The insulin:carb ratio can then be calculated as the total effect of 1 unit of insulin divided by the total effect of 1 carbohydrate. Alternatively, clinically averaged response functions from many patients, in combination with the user's weight, can be used as an approximate input for the second step, determining the optimal insulin dose.

BRIEF DESCRIPTION OF THE DRAWING

[0011] FIG. 1 is a flow diagram depicting the input and output to the two stages of the process of determining individualized insulin sensitivity and optimal insulin dose.

DETAILED DESCRIPTION

[0012] Step 1 of the method linearly regresses changes in glucose over possibly overlapping time intervals onto the following equation:

.DELTA. g i , j = ( EGP 0 - P * g i ) * .DELTA. t i , j + RGC + j = l N short .theta. short , j S j + j = l N long .theta. long , j L j + j = l N carb .theta. carb , j C j ##EQU00001##

where,

[0013] g.sub.i is the i.sup.th glucose measure in time,

[0014] .DELTA.g.sub.i,j=g.sub.i+j-g.sub.i,

[0015] .DELTA.t.sub.i,j=t.sub.i+j-t.sub.i

[0016] EGP.sub.0 is endogenous glucose production extrapolated in 0 glucose, a fit parameter; p is suppression of endogenous glucose production, a fitting parameter;

[0017] RGC=A(g.sub.i-g.sub.RGC) if g>g.sub.RGC and RGC=0 otherwise, represents renal glucose clearance where A and g.sub.RGC--the renal glucose clearance threshold--are fit parameters;

[0018] 0.sub.short,j are responses to short-acting insulin doses S.sub.j administered j time periods before t.sub.i, where each time period could be 30 to 60 minutes, are fit parameters;

[0019] 0.sub.long,j are responses to long-acting insulin doses L.sub.j administered j time periods before t.sub.i, where each time period could be 30 to 60 minutes, are fit parameters; and

[0020] 0.sub.carb,j are responses to dietary carbohydrates C.sub.j administered j time periods before t.sub.i, where each time period could be 15 to 60 minutes, are fit parameters.

[0021] As another example linear-fit form, step 1 of the method could fit glucose changes to the following equation:

.DELTA. g i , j = ( EGP 0 - p * g i ) * .DELTA. t i , j + RGC + j = l N short k = 0 d short .theta. short , k ( t i - t short , j ) k S j + j = l N long k = 0 d long .theta. long , k ( t i - t long , j ) k L j + j = l N carb k = 0 d carb .theta. carb , k ( t i - t carb , j ) k C j ##EQU00002##

where,

[0022] S.sub.j are doses of short-acting insulin taken within some timeframe of t.sub.i, possibly 6 hours;

[0023] L.sub.j are doses of long-acting insulin taken within some timeframe of t.sub.i, possibly 24 hours;

[0024] C.sub.j are dietary carbohydrates eaten within some timeframe of t.sub.i, possibly 3 hours; and d.sub.short, d.sub.long, d.sub.carb are the degrees of polynomial response functions.

[0025] For any linear-fit form, the length of the time periods for the response curves in step 1 is chosen such that the rate of change of glucose to that factor has an insignificant fluctuation over the time period. The number of time periods to look back from each g.sub.i--which could be different for each factor--is chosen such that the number of periods multiplied by the length of each period is longer than the time for the insulin, carbohydrate, or exercise to clear the body. Endogeneous glucose production could also be set as a separate factor for different times of day, possibly blocks of one to six hours.

[0026] In step 2, the method could use the response functions learned in step 1 to suggest appropriate insulin doses. The output could be carbohydrate ratios, correction doses, and long-acting insulin doses for a user on MDI, or basal and bolus rates for CSII, to be programmed manually in an existing open-loop pump system or to be set automatically in a future, closed-loop artificial pancreas. The method takes the actual time series of food and exercise and combines them with the times, but not the sizes, of actual insulin doses, as well as the response functions to these factors learned in step 1. The method then calculates optimal insulin doses for each relevant event, where the ideal dose is the one that minimizes a cost function that penalizes hyperglycemia and hypoglycemia. The cost function for excursions outside the optimal glucose range could be asymmetrical in order to avoid hypoglycemia more strongly than hyperglycemia. By then averaging over dose size per carbohydrate at each meal, the method can be used to suggest a best insulin:carb ratio. Similarly, by averaging over the ideal correction dose for a given glucose level at each episode of hyperglycemia, the method can suggest the proper correction dose for different levels of hyperglycemia. Step 2 will also assume that a fixed amount of long-acting insulin is taken each day, and it can suggest its ideal dose and time of application.

REFERENCES

[0027] [1] JDRF, Statistics: JDRF and Diabetes, (2014)

[0028] [2] Centers for Disease Control and Prevention, Number (in Millions) of Adults with Diabetes by Diabetes Medication Status, United States, 1997-2011, (2013)

Claims

1-7. (canceled)

8. A diabetes management system for recommending a size of a user insulin dose, the system comprising: a glucose monitor configured to provide user glucose readings; an insulin delivery mechanism; a memory module, the memory module configured to receive and store a record of insulin doses previously administered to a user as a function of time, the memory module configured to receive and store a confidence interval of a response of a user's glucose concentration to the insulin doses; and a processor module configured to use a mathematical model to calculate an optimal dose of insulin to be taken by the user, the mathematical model comprising: projecting a time series of future glucose concentrations over a predetermined amount of time; and minimizing an error function that penalizes deviations away from a predetermined ideal glucose level, wherein the error function penalizes deviations of the glucose concentration below the predetermined ideal glucose level more than deviations of the glucose concentration above the predetermined ideal glucose level; and a display module to display the optimal dose of insulin.

9. The system of claim 8, wherein the mathematical model is a linear regression algorithm where a dependent variable is change in glucose over possibly overlapping time intervals and independent variables are doses of insulin, dietary carbohydrates, and other factors that may affect glucose summed over windows of time prior to the time window during which the glucose change is being regressed.

10. The system of claim 9, wherein the other factors that may affect glucose include a duration of exercise engaged in by a user.

11. The system of claim 8, wherein the mathematical model is a linear regression algorithm where a dependent variable is change in glucose over possibly overlapping time intervals and independent variables are doses of insulin, dietary carbohydrates, and other factors that may affect glucose multiplied by a polynomial dependent on the time delay between when the dose was administered and when the glucose response is being measured.

12. The system of claim 11, wherein the other factors that may affect glucose include duration of exercise engaged in by the user.

13. The system of claim 8, wherein the insulin delivery mechanism comprising a delivery mechanism for delivering short-acting insulin.

14. The system of claim 13, further comprising a second insulin delivery mechanism comprising a delivery mechanism for delivering long-acting insulin.

15. The system of claim 14, wherein the system is configured to suggest an ideal dose of the long-acting insulin and an ideal time to dose the long-acting insulin.

16. The system of claim 13, wherein the system suggests an ideal correction dose of the short-acting insulin for different levels of hyperglycemia.

17. A diabetes management system for recommending the size of a user insulin dose, the system comprising: a glucose monitor configured to provide a user glucose readings; an insulin delivery mechanism; a memory module, the memory module configured to receive and store a record of insulin doses previously administered to the user as a function of time; and a processor module configured to: (i) derive rates of change in a user's glucose concentration as a function of time from a plurality of glucose concentrations stored in the memory module, (ii) derive a glucose response as a function of time that is a fit to a plurality of glucose concentrations and insulin doses, the fit being based on a mathematical model, (iii) integrate the glucose response as a function of time to obtain a user's total glucose response to insulin doses, and (iv) calculate an optimal dose of insulin to be taken by the user by projecting a time series of future glucose concentrations over a predetermined amount of time and minimizing an error function that penalizes deviations away from a predetermined ideal glucose level; and a display module to display the optimal dose of insulin.

18. The system of claim 17, wherein the mathematical model is a linear regression algorithm where a dependent variable is change in glucose over possibly overlapping time intervals and independent variables are doses of insulin, dietary carbohydrates, and other factors that may affect glucose summed over windows of time prior to the time window during which the glucose change is being regressed.

19. The system of claim 18, wherein the other factors that may affect glucose include a duration of exercise engaged in by the user.

20. The system of claim 17, wherein the mathematical model is a linear regression algorithm where a dependent variable is change in glucose over possibly overlapping time intervals and independent variables are doses of insulin, dietary carbohydrates, and other factors that may affect glucose multiplied by a polynomial dependent on the time delay between when the dose was administered and when the glucose response is being measured.

21. The system of claim 20, wherein the other factors that may affect glucose include duration of exercise engaged in by the user.

22. The system of claim 17, wherein the error function penalizes deviations of the glucose concentration below the predetermined ideal glucose level more than deviations of the glucose concentration above the predetermined ideal glucose level, whereby a glucose concentration below the ideal level is made less likely.

23. The system of claim 17, wherein the insulin delivery mechanism is a delivery mechanism for delivering short-acting insulin.

24. The system of claim 23, further comprising a second insulin delivery mechanism for delivering long-acting insulin.

25. The system of claim 24, wherein the system is configured to suggest an ideal dose of the long-acting insulin and an ideal time to dose the long-acting insulin.

26. The system of claim 23, wherein the system suggests an ideal correction dose of the short-acting insulin for different levels of hyperglycemia.