Method and system for closed-loop control of an artificial pancreas

• US 9,517,306 B2
• Filed: 03/15/2013
• Issued: 12/13/2016
• Est. Priority Date: 03/15/2013
• Status: Active Grant

• Current Assignee: JDRF International
• Original Assignee: Animas Corporation (Johnson & Johnson)
• Inventors: Morales, Carlos
• Primary Examiner(s): Skibinsky, Anna
• Application Number:
  US13/840,429
• Publication Number:
  US 20140276555A1
• Time in Patent Office: 1,369 Days
• Field of Search: None
• US Class Current: 1/1
• CPC Class Codes: A61M 2205/3569 sublocal, e.g. between console and disposable
• CPC Class Codes: A61M 5/1782 Devices aiding filling of syringes in situ combination of a vial and a syringe for transferring or mixing their contents A61J1/2096, filling of medical containers in general B65B3/003
• CPC Class Codes: G16H 20/17 delivered via infusion or injection
• CPC Class Codes: G16H 40/60 for the operation of medical equipment or devices

Abstract

A local extremum of a function ƒ is determined by computing a Jacobian of ƒ at a test point x by adding an imaginary part to x and a Hessian of ƒ by adding two imaginary parts to a multicomplex copy of x and extracting a third imaginary part. Solving a system of equations defined by the Jacobian and Hessian yields a delta; the process is repeated until convergence. This method is used in each of a series of time intervals to compute an insulin-delivery amount for an insulin pump. ƒ is a model-predictive-control cost function; x is a set of successive candidate insulin delivery amounts beginning from a selected time interval. A system includes a glucose monitor and a controller using glucose measurement data therefrom to determine an insulin delivery amount for a time interval by minimizing ƒ; an insulin pump provides insulin corresponding to the delivery amount.

Claims

1. Apparatus for the delivery of insulin, the apparatus comprising:

a) a glucose monitor adapted to measure respective glucose levels of a subject at discrete time intervals and provide respective glucose measurement data indicating each measured glucose level;

b) an insulin infusion pump configured to deliver insulin in response to a delivery control signal;

c) a memory configured to store a plurality of basal insulin delivery amounts at respective ones of the discrete time intervals; and

d) a model predictive controller adapted to, for each of a plurality of the discrete time intervals;

i) receive the glucose measurement data for that time interval from the glucose monitor;

ii) determine an insulin delivery amount for that time interval using model predictive control based on a selected target glucose concentration range, the received glucose measurement data, the stored basal insulin delivery profile amounts for that time interval and n−

1 successive time intervals; and

iii) provide to the insulin infusion pump a delivery control signal corresponding to the determined insulin delivery amount, so that a corresponding amount of insulin is delivered by the infusion pump;

e) in which the model predictive controller is adapted to determine the insulin delivery amount by mathematical minimization of a cost function that computes a cost metric correlated with physiological-fluid glucose-level excursions from a selected target glucose range for a particular set of n successive candidate insulin delivery amounts beginning from the selected one of the time intervals, and the model predictive controller, in order to carry out the mathematical minimization, is adapted to;

i) select n candidate insulin delivery values for a real-type n-dimensional test point;

ii) for each dimension i of the n dimensions;

A) set a complex-type n-dimensional computation point c equal to the test point;

B) set an imaginary part of element i of c equal to a nonzero increment;

C) compute a complex-type value of the cost function at c; and

D) divide the imaginary part of the complex-type computed function value by the increment to form element i of an approximate Jacobian of the function at the test point;

iii) for each pair of dimensions (i,j), each i and j one of the n dimensions;

A) set a multicomplex-type n-dimensional computation point b equal to the test point;

B) set a first imaginary part of element i of b equal to a nonzero first increment;

C) set a second imaginary part of element j of b equal to a nonzero second increment;

D) compute a multicomplex-type value of the cost function at b; and

E) divide the third imaginary part of the computed multicomplex-type function value by the product of the first and second increments to form element (i,j) of an approximate Hessian of the function at the test point;

iv) solve a system of equations defined by the approximate Hessian and the approximate Jacobian to find a delta;

v) modify the test point according to the delta to form a next point;

vi) determine whether the next point satisfies selected convergence criteria;

vii) if not, assign the value of the next point to the test point and repeat the computation of the approximate Jacobian and Hessian of the function, solution of the system of equations, modification of the test point, and determination of whether the next point satisfies the selected convergence criteria; and

viii) if so, select the first element of the next point as the candidate insulin delivery amount for the selected one of the time intervals;

f) and further in which the model predictive controller is adapted to, in order to compute the cost function for each input set of n successive candidate insulin delivery amounts;

i) predict an excursion of the glucose level from a selected target glucose range using at least some of the glucose measurements, a glucose-insulin dynamic model of the subject, and an input set of n candidate insulin delivery amounts;

ii) compute a deviation of the candidate insulin delivery amounts from respective selected basal delivery amounts; and

iii) form a weighted sum of respective values representing the predicted excursion and the computed deviation using respective selected weights, such that the output of the cost function includes the weighted sum; and

g) deliver, with the insulin infusion pump, the determined insulin delivery amount of insulin to the subject, as provided by the delivery control signal from the controller.

2. The apparatus according to claim 1, wherein the model predictive controller is further adapted to:

a) predict an excursion of a glucose level of the subject from the selected target glucose range using a safety model and at least some of the glucose measurement data for a plurality of the time intervals; and

b) reduce the determined insulin delivery amount according to the predicted excursion.

3. The apparatus according to claim 1, wherein the glucose monitor includes a plurality of glucose sensors.

4. The apparatus according to claim 1, wherein the multicomplex type is a bicomplex type and the third imaginary part is a combined imaginary part.

5. The apparatus according to claim 1, wherein the multicomplex type is a quaternion type, and the first, second, and third imaginary parts are i_Q, j_Q, and k_Q parts, respectively.

6. A method to control an infusion pump responsive to a controller that receives data from a glucose sensor, the method comprising:

a) measuring respective glucose levels of a physiological fluid from a subject for each time interval of a series of discrete time intervals using the glucose sensor; and

b) using a controller, automatically calculating an insulin delivery amount for a selected one of the time intervals by mathematically minimizing a cost function that computes a cost metric correlated with physiological-fluid glucose-level excursions from a selected target glucose range for a particular set of n successive candidate insulin delivery amounts beginning from the selected one of the time intervals, the minimizing including;

i) selecting n candidate insulin delivery values for a real-type n-dimensional test point;

ii) computing an approximate Jacobian of the cost function at the test point by evaluating the cost function, in which the model predictive controller computes the cost function for each input set of n successive candidate insulin delivery amounts by;

A) predicting an excursion of the glucose level from a selected target glucose range using at least some of the glucose measurements, a glucose-insulin dynamic model of the subject, and the candidate insulin delivery amounts;

B) computing a deviation of the candidate insulin delivery amounts from respective selected basal delivery amounts; and

C) forming a weighted sum of respective values representing the predicted excursion and the computed deviation using respective selected weights such that the output of the cost function includes the weighted sum;

iii) computing an approximate Hessian of the cost function at the test point by evaluating the cost function;

iv) solving a system of equations defined by the approximate Hessian and the approximate Jacobian to find a delta;

v) modifying the test point according to the delta to form a next point;

vi) determining whether the next point satisfies selected convergence criteria;

vii) if not, assigning the value of the next point to the test point and repeating the computing of the approximate Jacobian and Hessian, solving the system of equation, modifying the test point and the determining step; and

viii) if so, selecting a first element of the next point as the candidate insulin delivery amount for the selected one of the time intervals;

(c) commanding, with a control signal from the controller, the infusion pump to deliver the calculated insulin delivery amount to the subject; and

d) delivering, with the infusion pump, the calculated insulin delivery amount of insulin to the subject.

7. The method according to claim 6, in which:

a) the computing-approximate Jacobian step includes, for each dimension i of the n dimensions;

i) setting a complex-type n-dimensional computation point c equal to the test point;

ii) setting an imaginary part of element i of c equal to a nonzero increment;

iii) computing a complex-type value of the cost function at c; and

iv) dividing the imaginary part of the complex-type computed function value by the increment to form element i of the Jacobian; and

b) the computing-approximate-Hessian step includes, for each pair of dimensions (i,j), each i and j one of the n dimensions;

i) setting a multicomplex-type n-dimensional computation point b equal to the test point;

ii) setting a first imaginary part of element i of b equal to a nonzero first increment;

iii) setting a second imaginary part of element j of b equal to a nonzero second increment;

iv) computing a multicomplex-type value of the cost function at b; and

v) dividing the third imaginary part of the computed multicomplex-type function value by the product of the first and second increments to form element (i,j) of the Hessian.

8. The method according to claim 7, wherein the multicomplex type is a bicomplex type and the third imaginary part is a combined imaginary part.

9. The method according to claim 7, wherein the multicomplex type is a quaternion type, and the first, second, and third imaginary parts are i_Q, j_Q, and k_Q parts, respectively.

10. The method according to claim 6, wherein the determining-approved-amount step includes providing the candidate insulin delivery amount as the approved insulin delivery amount.

11. The method according to claim 6, wherein the determining-approved-amount step includes reducing the candidate delivery amount according to a safety model to provide the approved insulin delivery amount.

12. The method according to claim 6, wherein the solving step includes executing at least two iterations of a conjugate-gradient algorithm.

13. The method according to claim 6, further including selecting a value for a trust region, in which the modifying step includes computing the next point as the sum of the test point and the delta and clipping the sum within the trust region, and the determining step includes:

a) determining whether the value of the function at the next point is lower than the value of the function at the test point;

b) if not, contracting the trust region and determining that the next point does not satisfy the selected convergence criteria; and

c) if so, determining whether the next point satisfies the selected convergence criteria and, if not, expanding the trust region.