Glucose predictor based on regularization networks with adaptively chosen kernels and regularization parameters
The invention relates to a method and a device for predicting a glycaemic profile of a subject. A multistage algorithm is employed comprising a prediction setting stage specifying a functional space for the prediction and a prediction execution stage specifying a predicted future glycaemic state of the subject in the functional space as a continuous function of time.
1. A glucose prediction device comprising:
an input device structured to receive information indicative of a physiologic condition of a subject;
a processing device comprising an adaptive regularization network that is structured to predict a future glucose profile;
an output device structured to convey the future glucose profile; and
an alarm;
wherein the future glucose profile is the glycaemic state of the subject as a continuous function of time;
wherein the adaptive regularization network is adapted to perform a multistage prediction process and comprises a supervising learning machine and a supervised learning machine that allows the future glucose profile to be predicted with irregularly sampled data in the information received by the input device;
wherein (a) the supervising learning machine is adapted to (i) compress the information received by the input device and (ii) run the compressed information through a pre-constructed machine to produce kernel parameters and a regularization parameter; and (b) the supervised learning machine is adapted to calculate the future glucose profile as a function of the information received by the input device and the kernel parameters and the regularization parameter produced by the supervising learning machine; and
wherein the alarm is structured to alert the subject if the future glucose profile includes an impending hypo- or hyperglycaemic event.
2. The device according to claim 1 , wherein the pre-constructed machine is produced by a linear fitting, and wherein the kernel parameters and the regularization parameter is based on the linear fitting.
3. The device according to claim 1 , wherein the multistage prediction process comprises a reproducing kernel Hilbert space.
4. The device according to claim 3 , wherein the reproducing kernel Hilbert space is selected based on a data pool of training data segments, wherein each training data segment is indicative of the physiologic condition of the subject at a point in time.
5. The device according to claim 4 , wherein the data pool is obtained from a predetermined data pool.
6. The device according to claim 4 , wherein the data pool is obtained from a continuously updated data pool.
7. The device according to claim 1 , wherein the information indicative of the physiologic condition of the subject is at least one measurement of a body characteristic.
8. The device according to claim 7 , wherein the body characteristic is blood or tissue glucose.
9. The device according to claim 7 , wherein the input device is further adapted to receive information related to a therapeutic treatment.
10. The device according to claim 9 , wherein the information related to the therapeutic treatment is past delivery of a glucose regulating agent.
11. The device according to claim 1 , wherein the input device is further adapted to receive information related to a meal consumed or to be consumed by the subject.
12. A computer-implemented method for predicting a future glucose profile of a subject that is the glycaemic state of the subject as a continuous function of time, the method comprising:
providing a glucose prediction device comprising a processing device;
receiving, by the processing device, information indicative of a physiologic condition of the subject;
specifying, by the processing device, a functional space for the predicted glucose profile by compressing the received information and running the compressed information through a pre-constructed machine to produce kernel parameters and a regularization parameter;
calculating, by the processing device, predicted glucose values as a continuous function of time based on the received information, the kernel parameters, and the initial regularization parameter to thereby produce the predicted future glucose profile of the subject; and
alerting the user, by the glucose prediction device, if the future glucose profile of the subject includes an impending hypo- or hyperglycaemic event.
13. The method according to claim 12 , further comprising producing the kernel parameters by a regularized learning algorithm in a reproducing kernel Hilbert space defined by at least approximately minimizing an error function.
14. The computer-implemented method according to claim 12 , further comprising producing the pre-constructed machine by
receiving a data pool of training data segments, wherein each training data segment is indicative of a physiologic condition of the subject at respective points in time;
compressing each training data segment to obtain a corresponding set of compressed training data segments;
determining a set of training parameters parameterising the functional space, wherein each training parameter of the training set minimizes a prediction error function indicative of a deviation between a physiologic condition and a predicted physiologic condition, wherein the predicted physiologic condition is predicted based on the functional space parameterised by said training parameter; and
constructing, from the set of training parameters, a non-linear mapping between compressed input data segments and a set of kernel parameters parameterising the respective kernels.
15. The method according to claim 14 , wherein constructing the non-linear mapping comprises determining a mechanism for determining a Tikhonov regularization parameter.
16. The method according to claim 14 , wherein the prediction error function penalises a delay or failure in the prediction of physiological events classified as dangerous more heavily than a delay or failure in the prediction of physiological events classified as normal.
17. The method according to claim 14 , wherein the data pool is obtained from a predetermined data pool.
18. The method according to claim 14 , wherein the data pool is obtained from a continuously updated data pool.
19. The method according to claim 12 , wherein the information indicative of the physiologic condition of the subject is at least one measurement of a body characteristic.
20. The method according to claim 19 , wherein the body characteristic is blood or tissue glucose.
21. The method according to claim 12 , further comprising receiving, by the processing device, information related to a therapeutic treatment.
22. The method according to claim 21 , wherein the information related to the therapeutic treatment is past delivery of a glucose regulating agent.
23. The method according to claim 12 , further comprising receiving, by the processing device, information related to a meal consumed or to be consumed by the subject.
24. A glucose prediction device comprising:
an input device structured to receive information indicative of a physiologic condition of a subject, wherein the information comprises blood or tissue glucose measurements;
a processing device comprising an adaptive regularization network that is structured to predict a future glucose profile;
an output device structured to convey the future glucose profile; and
an alarm;
wherein the future glucose profile is the glycaemic state of the subject as a continuous function of time;
wherein the adaptive regularization network is adapted to perform a multistage prediction process and comprises a supervising learning machine and a supervised learning machine that allows the future glucose profile to be predicted with irregularly sampled data in the blood or tissue glucose measurements received by the input device;
wherein (a) the supervising learning machine is adapted to (i) compress the information received by the input device and (ii) run the compressed information through a pre-constructed machine to produce kernel parameters and a regularization parameter; and (b) the supervised learning machine is adapted to calculate the future glucose profile as a function of the information received by the input device and the kernel parameters and the regularization parameter produced by the supervising learning machine; and
wherein the alarm is structured to alert the subject if the future glucose profile includes an impending hypo- or hyperglycaemic event.
25. The device according to claim 24 , wherein the pre-constructed machine is produced by a linear fitting, and wherein the kernel parameters and the regularization parameter is based on the linear fitting.
26. The device according to claim 24 , wherein the multistage prediction process comprises a reproducing kernel Hilbert space.
27. The device according to claim 26 , wherein the reproducing kernel Hilbert space is selected based on a data pool of training data segments, wherein each training data segment is indicative of the physiologic condition of the subject at a point in time.
28. The device according to claim 27 , wherein the data pool is obtained from a predetermined data pool.
29. The device according to claim 27 , wherein the data pool is obtained from a continuously updated data pool.
30. The device according to claim 24 , wherein the input device is further adapted to receive information related to a therapeutic treatment.
31. The device according to claim 30 , wherein the information related to the therapeutic treatment is past delivery of a glucose regulating agent.
32. The device according to claim 24 , wherein the input device is further adapted to receive information related to a meal consumed or to be consumed by the subject.
33. A computer-implemented method for predicting a future glucose profile of a subject that is the glycaemic state of the subject as a continuous function of time, the method comprising:
providing a glucose prediction device comprising a processing device;
receiving, by the processing device, information indicative of a physiologic condition of the subject, wherein the information comprises blood or tissue glucose measurements;
specifying, by the processing device, a functional space for the predicted glucose profile by compressing the received information and running the compressed information through a pre-constructed machine to produce kernel parameters and a regularization parameter, wherein the kernel parameters are produced by a regularized learning algorithm in a reproducing kernel Hilbert space defined by at least approximately minimizing an error function;
calculating, by the processing device, predicted glucose values as a continuous function of time based on the received information, the kernel parameters, and the initial regularization parameter to thereby produce the predicted future glucose profile of the subject; and
alerting the user, by the glucose prediction device, if the future glucose profile of the subject includes an impending hypo- or hyperglycaemic event.
34. The method according to claim 33 , further comprising producing the pre-constructed machine by
receiving a data pool of training data segments, wherein each training data segment is indicative of a physiologic condition of the subject at respective points in time;
compressing each training data segment to obtain a corresponding set of compressed training data segments;
determining a set of training parameters parameterising the functional space, wherein each training parameter of the training set minimizes a prediction error function indicative of a deviation between a physiologic condition and a predicted physiologic condition, wherein the predicted physiologic condition is predicted based on the functional space parameterised by said training parameter; and
constructing, from the set of training parameters, a non-linear mapping between compressed input data segments and a set of kernel parameters parameterising the respective kernels.
35. The method according to claim 34 , wherein constructing the non-linear mapping comprises determining a mechanism for determining a Tikhonov regularization parameter.
36. The method according to claim 34 , wherein the prediction error function penalises a delay or failure in the prediction of physiological events classified as dangerous more heavily than a delay or failure in the prediction of physiological events classified as normal.
37. The method according to claim 34 , wherein the data pool is obtained from a predetermined data pool.
38. The method according to claim 34 , wherein the data pool is obtained from a continuously updated data pool.
39. The method according to claim 33 , further comprising receiving, by the processing device, information related to a therapeutic treatment.
40. The method according to claim 39 , wherein the information related to the therapeutic treatment is past delivery of a glucose regulating agent.
41. The method according to claim 33 , further comprising receiving, by the processing device, information related to a meal consumed or to be consumed by the subject.