CN114613509B - Artificial pancreas long-term adaptation individualized learning system based on Bayesian optimization - Google Patents

Artificial pancreas long-term adaptation individualized learning system based on Bayesian optimization Download PDF

Info

Publication number
CN114613509B
CN114613509B CN202210425970.0A CN202210425970A CN114613509B CN 114613509 B CN114613509 B CN 114613509B CN 202210425970 A CN202210425970 A CN 202210425970A CN 114613509 B CN114613509 B CN 114613509B
Authority
CN
China
Prior art keywords
blood sugar
value
blood glucose
hypoglycemia
parameter
Prior art date
Legal status (The legal status 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 status listed.)
Active
Application number
CN202210425970.0A
Other languages
Chinese (zh)
Other versions
CN114613509A (en
Inventor
史大威
蔡德恒
马牧远
王军政
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Institute of Technology BIT
Original Assignee
Beijing Institute of Technology BIT
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 Beijing Institute of Technology BIT filed Critical Beijing Institute of Technology BIT
Priority to CN202210425970.0A priority Critical patent/CN114613509B/en
Publication of CN114613509A publication Critical patent/CN114613509A/en
Application granted granted Critical
Publication of CN114613509B publication Critical patent/CN114613509B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/30—ICT 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
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/145—Measuring characteristics of blood in vivo, e.g. gas concentration, pH value; Measuring characteristics of body fluids or tissues, e.g. interstitial fluid, cerebral tissue
    • A61B5/14532—Measuring 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
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G06F18/24155—Bayesian classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

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

Abstract

The invention discloses an artificial pancreas long-term adaptation individualized learning system based on Bayesian optimization, which comprises an individual blood sugar condition evaluation module, an individualized AP parameter learning module and a Bayesian optimization module. The invention adopts the individualized learning system based on the Bayesian optimization for the artificial pancreas long-term adaptation to implement individualized AP self-adaptation, steadily adjusts AP parameters (BR), increases the normal blood sugar time proportion and obviously reduces the risk of hypoglycemia.

Description

Artificial pancreas long-term adaptation individualized learning system based on Bayesian optimization
Technical Field
The invention relates to the technical field of artificial pancreas control, in particular to an artificial pancreas long-term adaptation individualized learning system based on Bayesian optimization.
Background
In humans, blood glucose concentration is tightly controlled by insulin and glucagon secreted by the pancreas. Insulin is secreted by pancreatic beta cells for lowering blood glucose concentration; glucagon is secreted by pancreatic alpha cells, which increases blood glucose concentration. Type i diabetes is caused by autoimmune destruction of the pancreatic cells, loss of insulin secretion. Of these, type i diabetes accounts for 5-15% of about 3.66 billion diabetic patients worldwide, and its incidence increases at a rate of 3.9% per year. According to the investigation and research of the current situation of Chinese adult diabetes prevalence and control, the prevalence rate of diabetes of adults 18 years old and older in China is 11.6%, the prevalence rate of diabetes in the early stage is 50.1%, the treatment rate of diabetes patients in China is only 25.8%, and the effective treatment of diabetes becomes one of the most important and troublesome public health problems in China.
Early in the 20 th century, it was discovered that insulin changed type i diabetes from a fatal disease to a chronic disease requiring lifelong insulin replacement therapy. Wearing an Artificial Pancreas (AP) to supplement insulin in vitro becomes an important means for controlling blood sugar at present. The AP system is intended to automatically control and regulate the blood glucose in the normal range of patients by producing insulin pellets from type i diabetic patients. A traditional AP consists of three components: a glucose sensor, an infusion pump and a controller automatically adjust hormone delivery based on glucose measurements to adjust glucose concentration, effectively improving glycemic control.
Currently, AP systems rely primarily on the setting of AP parameters, such as the Basal Rate (BR) of insulin to adjust insulin dosage in real time. The regulation effect of the method on the blood sugar of the patient is basically determined by the accuracy of the initially set AP parameter (BR), the accuracy of the AP parameter (BR) is greatly dependent on the treatment judgment of a doctor and the experience of the patient, and the BR required by the patient to maintain normal blood sugar level is changed along with the physiological state for a long time. When the AP parameter setting is unreasonable or the BR variation range required by a patient is large, the influence of wrong AP parameters on the blood sugar control effect cannot be effectively corrected due to the limited real-time adjustment capability of the AP system. The blood sugar control effect is poor due to various reasons, the blood sugar control effect is not obvious when the AP is worn, the hyperglycemic events or the hypoglycemic events occur frequently, and the difference among patients is large.
In addition, during the process that the AP system is worn by a patient for a long time, abundant blood sugar monitoring data and insulin infusion data are generated, and data support can be provided for individualized medical treatment. However, it is still a challenge how to fully mine the blood sugar metabolism change rules contained in the data, systematically adjust the AP parameters (BR) under life interference, and simultaneously search for parameter constraints to ensure the AP parameters to be adjusted steadily without hypoglycemia, so as to realize long-term safe and good blood sugar closed-loop control. Therefore, the method for establishing the data-driven AP individualized adaptive model by utilizing the historical data of the patient to form the system to adjust the AP parameters (BR) is of great significance.
In the prior art, a patent CN201310312769.2 discloses an individualized insulin therapy pump and a basic infusion rate optimization method thereof, which establishes a relation model between a blood glucose value of a patient and a basic insulin infusion rate at any time according to blood glucose data fed back by the patient in real time, simulates a blood glucose change condition of the patient at any time when the basic infusion rate changes within a variable range of the basic infusion rate, and evaluates and selects an optimal basic infusion rate value. However, the patent adopts a simple linear relation when determining a relation model of the blood sugar value and the insulin basal infusion rate, neglects the complexity of the blood sugar metabolism rule, and is difficult to accurately simulate the blood sugar change condition under different basal infusion rates.
Patent cn201780048242.X discloses a method and system for determining a basal rate adjustment of insulin based on a risk associated with a glucose state of a diabetic patient. The method detects a glucose state of the person based on the glucose measurement signals received by the AP and determines a current risk metric associated with the detected glucose state, the current risk metric associated with the detected glucose state being determined from a weighted average of cumulative risk values for return paths generated from a glucose state distribution around the detected glucose state. Finally, an adjustment to a basal rate of the therapy delivery device is calculated from a current risk metric associated with the detected glucose state and a reference risk metric associated with a reference glucose level. The patent utilizes real-time blood sugar to adjust the insulin basal rate, belongs to the research of an artificial pancreas control method, does not pay attention to how to systematically adjust the insulin basal rate based on historical blood sugar monitoring and insulin infusion data of a patient, and realizes long-term safe and good blood sugar control.
Disclosure of Invention
The invention aims to provide an individualized learning system for long-term adaptation of an artificial pancreas based on Bayesian optimization, which implements individualized AP self-adaptation, steadily adjusts AP parameters (BR), increases the normal blood sugar time ratio and obviously reduces the risk of hypoglycemia.
In order to achieve the aim, the invention provides an artificial pancreas long-term adaptation individualized learning system based on Bayesian optimization, which comprises an individual blood sugar condition evaluation module, an individualized AP parameter (BR) learning module and a Bayesian optimization module.
Preferably, the individual blood glucose condition evaluation module is configured to construct a blood glucose condition evaluation function y (z) according to historical blood glucose data in a period:
Figure BDA0003608516320000031
wherein z is the collected historical blood glucose data, y is the evaluation index of the blood glucose state, T represents the collection period of the historical blood glucose data, u 1 (125-135mg/dL)、u 2 (70-80mg/dL)、u 3 (70 mg/dL) represents the average value of the range of normal blood glucose, the lower limit of normal blood glucose, and the standard value of hypoglycemia, respectively. 0-t 1 Indicating that the blood glucose level is greater than the lower normal blood glucose limit u during the period T 2 Total time of (c); t is t 1 -t 2 Represents that the blood glucose level is less than the lower normal blood glucose limit u during the period T 2 But greater than a standard value u for hypoglycemia 3 Total time of (d); t is t 2 -T represents a blood glucose level less than a standard value u for hypoglycemia during the period T 3 The total time of (c).
The first term is a hypoglycemic penalty term: a is a hypoglycemia penalty term coefficient which represents that when hypoglycemia occurs, the target function value is obviously reduced compared with the target function value when hypoglycemia does not occur; the second term is an abnormal blood sugar penalty term: b is an abnormal blood sugar penalty term coefficient which represents that when abnormal blood sugar occurs but hypoglycemia does not occur, the objective function value is obviously reduced compared with the objective function values in the range of normal blood sugar; the third term represents the effect of the range of normal blood glucose on the objective function value, and C is the coefficient of this term. In general, the approximate relationship of A, B, and C may be considered to be A > B > C.
Preferably, 0 to t 1 The total time is that the blood sugar value is greater than the lower limit value u of normal blood sugar 2 May be discontinuous between the various time periods. t is t 1 The times for the respective time periods are simply added. t is t 1 -t 2 Total time is bloodThe sugar value is greater than the lower normal blood sugar value u 2 May be discontinuous between time periods. t is t 2- t 1 The times for the respective time periods are simply added. t is t 2 -Ttotal time is blood glucose value greater than lower normoglycemic limit u 2 May be discontinuous between time periods. T minus T 2 The times for the respective time segments are simply added.
Preferably, the individualized AP parameter learning module learns the rule between the historical blood glucose condition evaluation y and the AP parameter (BR) of the diabetic patient by adopting a Gaussian process to obtain a blood glucose condition y prediction model under different AP parameters (BR).
Preferably, the bayesian optimization module is configured to safely select an optimal AP parameter (BR) according to the blood glucose prediction model to obtain an AP parameter (BR) of a next period.
Preferably, the individual blood sugar condition evaluation module comprises a parameter determination unit, a blood sugar data acquisition unit and a function calculation unit.
The parameter determining unit is used for obtaining the average value u of the time interval T, the period T and the normal blood sugar range according to the blood sugar data in the AP system 1 (125-135 mg/dL) and lower limit of normal blood sugar u 2 (70-80 mg/dL) and a standard value u for hypoglycemia 3 (70 mg/dL) respectively determining blood sugar condition evaluation function coefficients A, B and C, and reflecting the hypoglycemic effect (generating a reduction effect on the blood sugar condition evaluation function value) and the obvious increase effect of the proportion of normal blood sugar on the function value by limiting the function coefficients.
Preferably, the coefficients a, B, C are determined as follows:
s11, classifying the blood sugar condition evaluation function curves when hypoglycemia does not occur;
(1) Function curve initial time t 0 The blood sugar is in the normal range and the final time t T Still in the normal range.
(2) Function curve initial time t 0 The blood sugar is in the normal range, and the final time t T Between the lower normal glycemic limit and the hypoglycemic norm value.
(3) Letter boxNumber curve initial time t 0 The blood sugar is between the lower normal blood sugar limit and the hypoglycemic standard value, and the final time t T Still between the lower normal glycemic limit and the hypoglycemic norm value.
S12, under the ideal extreme condition, respectively constructing 3 blood sugar condition evaluation function expressions according to 3 classifications of the blood sugar condition evaluation function curve in the step S1, and determining the minimum value y of the blood sugar condition evaluation function when hypoglycemia does not occur under the 3 classification conditions (1) 、y (2) 、y (3) 。
Blood glucose data is equal to the normal blood glucose lower limit:
Figure BDA0003608516320000051
initial time t 0 Until the blood sugar data in t is equal to the lower limit of normal blood sugar value, t to the final time t T The internal blood glucose data is equal to the hypoglycemic normative value:
Figure BDA0003608516320000052
the blood glucose data are all at the hypoglycemic standard value:
Figure BDA0003608516320000053
s13, comparing the minimum value y of the evaluation function of the blood sugar condition without hypoglycemia under the three classification conditions in the step S12 (1) 、y (2) 、y (3) Selecting y (1) 、y (2) 、y (3) The minimum value of the medium is used as the minimum value y of the evaluation function of the blood sugar condition when hypoglycemia does not occur 1min ；
S14, determining the maximum value y of the blood sugar condition evaluation function when hypoglycemia occurs by approximating a slope in ideal condition 1max . The simulated ideal blood glucose curve is:
Figure BDA0003608516320000054
where Δ t is the blood glucose measurement time interval and a is the confidence, i.e., if and only if the actual blood glucose is less than (1- α) u 3 When it is, hypoglycemia is considered to occur. Maximum value y of evaluation function of blood sugar status in occurrence of hypoglycemia 1max Comprises the following steps:
Figure BDA0003608516320000061
s15, making the minimum value y of the blood sugar state evaluation function obtained in the step S13 1min Is larger than the maximum value y of the blood sugar condition evaluation function obtained in step S14 1max Obtaining a specific limiting relation among blood sugar condition evaluation function coefficients A, B and C;
s16, when the blood sugar is in the normal range, taking ideal extreme conditions (the blood sugar value is in the upper limit or the lower limit of normal blood sugar) to obtain the minimum value y of the blood sugar condition evaluation function 2min ：
Figure BDA0003608516320000062
S17, calculating the maximum value y of the blood sugar condition evaluation function when the blood sugar is positioned between the lower normal blood sugar limit and the hypoglycemic limit 2max . Maximum value y of evaluation function of blood sugar status 2max Still determined by the ideal case approximating a diagonal line. And (3) simulating an ideal blood sugar curve:
Figure BDA0003608516320000063
where Δ t is the blood glucose measurement time interval and γ is the confidence, i.e., if and only if the actual blood glucose is less than (1- γ) u 3 When, hypoglycemia is considered to occur. Maximum value y of blood glucose evaluation function 2max Comprises the following steps:
Figure BDA0003608516320000064
s18, making the minimum value y of the blood sugar state evaluation function obtained in the step S16 2min Is larger than the maximum value y of the blood sugar condition evaluation function obtained in step S17 2max Obtaining another specific limiting relation between the blood sugar condition evaluation function coefficients B and C;
s19, through specific calculation in the step S15 and the step S18, the limit relation B (u) of the blood sugar condition evaluation function coefficients A, B and C is obtained 2 -u 3 )T＜4.5Aαu 3 Δt，C(u 1 -u 2 )T＜4.5Bγu 2 At is measured. Setting specific coefficient values meeting the limit relation according to experimental needs, wherein the coefficient values are not changed after being determined;
preferably, the hypoglycemic evidence is: the blood sugar data of three continuous samplings are all less than the standard value u of hypoglycemia 3 ；
The basis that the blood glucose vector is in an abnormal state is as follows: the blood sugar concentration of the continuous three times of sampling is lower than the normal range;
since the blood glucose data curves are diversified in the actual case, the determination of each simulation curve is performed in the ideal extreme case.
Preferably, the blood sugar data measured by the blood sugar sensor has certain error, and the confidence coefficient alpha is set by the invention to be that: when the actual blood sugar is positioned in [ (1-alpha) u 3 ，u 3 ]In the interval range, z and u 3 Approximately equal, and not considered hypoglycemic, when (T-2 Δ T) the blood glucose level is (1- α) u 3 And (T-3. DELTA.t) a blood glucose value of u 3 Then, the maximum value y of the evaluation function of the blood sugar condition when hypoglycemia occurs can be calculated and determined 1max . The confidence α represents: the probability of hypoglycemia occurring in real life is alpha.
Preferably, the confidence level γ is set by the present invention to be: when the actual blood sugar is located at [ (1-gamma) u 2 ，u 2 ]In the interval, z and u 2 Approximately equal, blood glucose is not considered abnormal. When the blood sugar value is (1-gamma) u when the (T-2 delta T) 2 (T-3. DELTA.t) blood sugar value u 2 While y can be determined 2max . With the confidence γ, the probability that the blood glucose is in an abnormal state in an actual situation is γ.
Preferably, the blood glucose data acquisition unit is used for acquiring historical blood glucose data and a BR set value of the diabetic patient in a certain period T when the diabetic patient wears the AP. Wherein, blood sugar data of the patient is obtained at intervals of t in the blood sugar sensing period in the AP. The period T is much larger than the time interval T.
Preferably, the function calculation unit performs linear interpolation on the measured blood glucose data points by the AP in a certain period according to the blood glucose condition evaluation function to obtain a blood glucose data piecewise function in a certain period T, and calculates a final blood glucose condition evaluation function value by piecewise weighted integration.
Preferably, the individualized AP parameter (BR) learning module is used for collecting the AP parameter (BR) value sets X obtained in the previous n periods 1 Evaluation function value set Y of blood sugar condition 1 Respectively as training input and training output of a blood sugar condition prediction model based on the Gaussian process to obtain a blood sugar condition y prediction model M1 under different AP parameters (BR), wherein the blood sugar condition y under different AP parameters (BR) has a mean value mu and a variance sigma 2 A gaussian distribution of (a).
When the number of data points is too small, the information collected during the gaussian prediction is insufficient, which may result in the accuracy of the prediction being reduced. N =4 is generally selected, and after 4 sets of data are collected, individualized AP parameter (BR) learning is performed.
Preferably, the bayesian optimization module selects an optimal AP parameter (BR) without hypoglycemia in the prediction model M1 by using a safety selection algorithm, and performs AP adaptation using the optimal AP parameter (BR) in a next period. The safety selection algorithm comprises the following steps:
s21, determining a prior safety set S according to experience of doctors and patients 0 And sample set X, S 0 Including the range of insulin basal rates considered by physicians and patients to be free of hypoglycemia and hyperglycemia. Historical blood glucose data was collected for the different insulin basal rates for the previous n cycles. Sample set X includes the basal rate of insulin for the first n cycles (n.gtoreq.4). Determining a verification set D, wherein the range is that the insulin basic rate in the current period fluctuates up and down by 50%, and an independent variable x represents the insulin basic rate;
s22, forecasting model M by utilizing Gaussian process 1 Definition of the Liphoz constant L of the Microcovariance nucleus k And error limit delta L ,
Figure BDA0003608516320000091
δ L E (0, 1). For a zero-mean gaussian process defined by a covariance kernel k (·, ·), the continuous partial derivative reaches a partial derivative kernel of fourth order:
Figure BDA0003608516320000092
and also,
Figure BDA0003608516320000093
represents the partial derivative kernel over the sample set X with the maximum distance r = max (X-X ') X, X' ∈ X
Figure BDA0003608516320000094
The liphoz constant of (a);
s23, using L in step S22 k ，
Figure BDA0003608516320000095
And r calculating a prediction model M 1 Lipschitz constant of unknown function of (1):
Figure BDA0003608516320000096
s24, predicting the model M through the Gaussian process 1 Obtaining different AP parameters (BR) on the verification set D, wherein the obedience mean value is mu and the variance is sigma 2 (ii) a Gaussian distribution of glycemic status y and a Leptochis constant L f Calculating the confidence upper limit u (x) = mu (x) + beta for each point x on the verification set D by taking deltax as the AP parameter (BR) interval 0.5 σ (x), lower confidence limit l (x) = μ (x) - β 0.5 Sigma (x), the parameter beta is a constant, and the confidence interval omega (x) is the difference between the upper confidence limit and the lower confidence limit;
the compactness of the confidence interval is controlled by the parameter β in step S24, which is generally set by the user according to the experimental situation.
S25, using the confidence limits 1 (x) of each point in step S24 and the safety set S in step S21 0 Selecting the lower confidence limit 1 (x) with the minimum value in the safety set as a safety limit h, h: = min l (x), x ∈ S 0 ；
S26, utilizing the Lipschitz constant L in the step S23 f Step S24 is to trust the upper limit set u (x), the lower limit set l (x), the confidence interval set omega (x), and step S25 is to secure the set S 0 Updating, wherein the updated security set is called a prior security set S t ；
Figure BDA0003608516320000101
S27, utilizing the Lipschitz constant L in the step S23 f Confidence upper bound u (x), confidence lower bound l (x) in step S24, security bound h in step S25 and a priori secure set S in step S26 t Computing a set of potential maximizers M t Extended Security set G t ；
M t ＝{x∈S t |u(x)≥max l(x′)，x′∈S t }
g(x)：＝|{x′∈D\S t |u(x)-L f d(x，x′)≥h}|
G t ＝{x∈S t |g(x)＞0}
S28, potential maximizer set M in step S27 t And extended Security set G t And in the merging set, selecting the maximum confidence interval value max omega (x), x belongs to M t ∪G t And selecting the AP parameter (BR) corresponding to the value as the optimal AP parameter (BR).
S29, calculating posterior safety set S 0 . Selecting the optimal AP parameter (BR) in the step S28 as the insulin basal rate in the (n + 1) th cycle, collecting blood sugar data, adding the (n + 1) th data to update the sample set X, and selecting the left and right limits in the sample set as a posterior safety set S when hypoglycemia or hyperglycemia does not occur at the time 0 An upper boundary and a lower boundary. If hypoglycemia or hyperglycemia occurs, the point is removed and the adjacent known function point is selected as the upper or lower boundary.
S30, posterior safety set S in step S29 0 As a safety set S at the n +1 th cycle 0 . Step S21 is repeated.
Therefore, the invention adopts the individualized learning system for the artificial pancreas long-term adaptation based on the Bayesian optimization, fully excavates the historical data information of the diabetic and the AP by using an artificial intelligence method, establishes a blood glucose condition prediction model under different AP parameters (BR), and safely selects the AP parameters (BR) by using the Bayesian optimization and safety selection ideas. Finally, individualized AP self-adaptation is realized, a data-driven AP self-adaptation model is established, AP parameters (BR) are steadily adjusted for a long time, the normal blood sugar time proportion is increased, the risk of hypoglycemia is obviously reduced, and the performance of a long-term self-adaptation AP system is automatically improved under the condition of minimum manual intervention.
Generally, the invention designs an AP long-term adaptive individualized safe learning system based on the Gaussian process by utilizing the historical blood glucose data of a patient in a certain period and AP system parameters (BR) in different periods, the system fully excavates the internal relation between the information of the historical blood glucose data of the patient and the AP system parameters (BR), and designs an individual blood glucose condition evaluation function, thereby being convenient for accurately and effectively evaluating the blood glucose condition level of the patient in a certain period; meanwhile, learning the rule between the historical blood sugar condition of the diabetic and the AP parameter (BR) by using a Gaussian process to obtain blood sugar condition prediction models under different AP parameters (BR); and determining the optimal AP parameter (BR) under the condition of no hypoglycemia by using a safety selection algorithm based on a blood sugar condition prediction model by using Bayesian optimization solution, so as to update the AP parameter (BR) within a safety range to adapt to the AP for a long time, increase the time proportion of normal blood sugar and obviously reduce hypoglycemia. Finally, the invention considers the difference of individual blood sugar data, establishes an individual safe learning system for long-term adaptation of AP by utilizing the rule between the historical data of each patient and the AP parameter (BR), and updates the blood sugar condition prediction model in a certain period so as to better improve the AP performance.
The technical solution of the present invention is further described in detail by the accompanying drawings and embodiments.
Drawings
FIG. 1 is a diagram of an adaptive AP framework provided by the present invention;
FIG. 2 is a block diagram of an individualized learning system for long term adaptation of APs provided by the present invention;
FIG. 3 is a flow chart of the implementation steps provided by the present invention;
FIG. 4 is a graph of blood glucose data for a patient having an inappropriate AP parameter (BR) provided by the present invention;
FIG. 5 is a graph of optimized AP parameters (BR) for a patient according to the present invention.
Detailed Description
The technical scheme of the invention is further explained by the attached drawings and the embodiment.
Unless defined otherwise, technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention belongs.
It will be evident to those skilled in the art that the invention is not limited to the details of the foregoing illustrative embodiments, and that the present invention may be embodied in other specific forms without departing from the spirit or essential attributes thereof. The present embodiments are therefore to be considered in all respects as illustrative and not restrictive, the scope of the invention being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein, and any reference signs in the claims are not intended to be construed as limiting the claim concerned.
Furthermore, it should be understood that although the present description refers to embodiments, not every embodiment may contain only a single embodiment, and such description is for clarity only, and those skilled in the art should integrate the description, and the embodiments may be combined as appropriate to form other embodiments understood by those skilled in the art. These other embodiments are also covered by the scope of the present invention.
It should be understood that the above-mentioned embodiments are only for explaining the present invention, and the protection scope of the present invention is not limited thereto, and any person skilled in the art should be able to cover the technical scope of the present invention and the equivalent replacement or change of the technical solution and the inventive concept thereof in the technical scope of the present invention.
The use of the word "comprising" or "comprises" and the like in the present invention means that the element preceding the word covers the element listed after the word and does not exclude the possibility of also covering other elements. The terms "inner", "outer", "upper", "lower", and the like indicate orientations or positional relationships based on orientations or positional relationships shown in the drawings, merely for convenience in describing the present invention and simplifying the description, but do not indicate or imply that the referred devices or elements must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as limiting the present invention, and when the absolute position of the described object is changed, the relative positional relationships may be changed accordingly. In the present invention, unless otherwise explicitly stated or limited, the terms "attached" and the like are to be understood broadly, e.g., as being fixedly attached, detachably attached, or integral; either directly or indirectly through intervening media, either internally or in any other relationship. The specific meanings of the above terms in the present invention can be understood by those skilled in the art according to specific situations. The term "about" as used herein has the meaning well known to those skilled in the art, and preferably means that the term modifies a value within the range of ± 50%, ± 40%, ± 30%, ± 20%, ± 10%, ± 5% or ± 1% thereof.
All terms (including technical or scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs unless specifically defined otherwise. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.
Techniques, methods, and apparatus known to those of ordinary skill in the relevant art may not be discussed in detail but are intended to be part of the specification where appropriate.
The disclosures of the prior art documents cited in the present description are incorporated by reference in their entirety and are therefore part of the present disclosure.
As shown in fig. 1, the adaptive AP includes an insulin real-time regulation system and an AP long-term adaptive individualized learning system. The real-time insulin regulating system includes insulin pump, blood sugar sensor and controller. After inputting the corresponding BR and other AP parameters, the controller automatically adjusts the injected insulin based on the measured blood glucose concentration. And the individualized learning system adapted to the AP for a long time updates BR in the insulin real-time regulation system according to the blood sugar value measured by the blood sugar sensor in a period.
As shown in fig. 2, an individualized learning system for long-term AP adaptation based on a gaussian process includes an individual blood glucose condition evaluation module, an individualized AP parameter (BR) learning module, and a bayesian optimization module; meanwhile, each parameter in the invention is determined based on the simulation of UVA/Padova T1DM blood sugar metabolism simulator certified by FDA, and ADRC is adopted in the control algorithm of the controller in the embodiment of the real-time insulin regulation system.
The individual blood sugar condition evaluation module is used for constructing a blood sugar condition evaluation function y (z) according to historical blood sugar data in a period:
Figure BDA0003608516320000141
wherein z is collected historical blood glucose data, y is a blood glucose state evaluation index, T represents a historical blood glucose data collection period, u 1 (125-135mg/dL)、u 2 (70-80mg/dL)、u 3 (70 mg/dL) represents the average value of the range of normal blood glucose, the lower limit of normal blood glucose, and the standard value of hypoglycemia, respectively. 0-t 1 Representing that the blood glucose level is greater than the lower normal blood glucose limit value u during the period T 2 Total time of (d); t is t 1 -t 2 Represents that the blood glucose level is less than the lower normal blood glucose limit u during the period T 2 But greater than the standard value u for hypoglycemia 3 Total time of (c); t is t 2 -T represents the criterion that the blood glucose level is less than hypoglycemia during period TValue u 3 The total time of (c).
The first term is a hypoglycemic penalty term: a is a hypoglycemia penalty term coefficient which represents that when hypoglycemia occurs, the target function value is obviously reduced compared with the target function value when hypoglycemia does not occur; the second term is an abnormal blood sugar penalty term: b is an abnormal blood sugar penalty term coefficient which represents that when abnormal blood sugar occurs but hypoglycemia does not occur, the objective function value is obviously reduced compared with the objective function value within the range of normal blood sugar; the third term represents the effect of the range of normal blood glucose on the objective function value, and C is the coefficient of this term. In general, the approximate relationship of A, B, and C can be considered to be A > B > C.
The individualized AP parameter learning module learns the rule between the historical blood sugar condition y of the diabetic and the AP parameter (BR) by adopting a Gaussian process to obtain a blood sugar condition y prediction model under different AP parameters (BR).
And the Bayesian optimization module is used for safely selecting the optimal AP parameter (BR) according to the blood sugar condition prediction model to obtain the next period AP parameter (BR).
That is, as shown in fig. 3, the system designed by the present invention includes the following steps:
(1) Historical blood glucose data and AP parameter (BR) historical data of a diabetic patient after wearing the AP are collected. And designing an individual blood sugar condition evaluation function, accurately and effectively evaluating the blood sugar condition level of a patient in a certain period, and obtaining a patient blood sugar condition function value in the certain period.
(2) Selecting a proper mean value and a proper kernel function by using the blood glucose condition function values of the patients in multiple cycles in the step (1) and AP parameter (BR) historical data, designing a corresponding training method, training a Gaussian process by using the data, and establishing an individualized blood glucose condition prediction model under the AP parameter (BR) under life interference according to different environments. In addition, the system can utilize the new history data at regular intervals to relearn the blood sugar condition rule and the AP parameter (BR) and update the blood sugar condition prediction model.
(3) And (3) solving the blood sugar condition prediction model determined in the step (2) by using Bayesian optimization. Due to the nature of gaussian process prediction, the glycemic condition predictor function has an expectation and variance. The AP parameters (BR) can be updated within a safe range by using a safe selection algorithm to determine the optimal AP parameters (BR) without hypoglycemia. The normal blood sugar time ratio is increased, and the hypoglycemia is obviously reduced.
It should be noted that the hypoglycemia criterion is: the blood sugar data of three continuous samplings are all smaller than the standard value u of hypoglycemia 3 。
As shown in FIG. 4, the individualized AP parameter (BR) learning module sets X the AP parameter (BR) values obtained in the first n periods 1 And blood sugar condition evaluation function value set y 1 Respectively as the training input and the training output of the blood sugar condition prediction model based on the Gaussian process to obtain a blood sugar condition y prediction model M1 under different AP parameters (BR), wherein y obeys that the mean value is mu and the variance is sigma 2 A gaussian distribution of (a).
And the Bayesian optimization module selects the optimal AP parameter (BR) under the condition of no hypoglycemia in the prediction model M1 by using a safety selection algorithm, and uses the optimal AP parameter (BR) in the next period to perform AP self-adaptation.
In addition, as time goes on, in a certain period, the patient generates new blood sugar historical data and AP parameters (BR), and the Gaussian process can be retrained by using the new data so as to realize AP self-adaptation and improve AP performance.
The individual blood sugar condition evaluation module comprises a parameter determination unit, a blood sugar data acquisition unit and a function calculation unit;
the parameter determining unit is used for obtaining the average value u of the time interval T, the period T and the normal blood sugar range according to the blood sugar data in the AP system 1 (125-135 mg/dL) lower limit of Normal blood sugar u 2 (70-80 mg/dL) and a standard value u for hypoglycemia 3 (70 mg/dL) respectively determining blood sugar condition evaluation function coefficients A, B and C, and reflecting the hypoglycemic effect (generating a reduction effect on a blood sugar condition evaluation function value) and the obvious increase effect of the proportion of the normal blood sugar on the function value through limiting the function coefficients.
The determination method of the coefficients A, B and C comprises the following steps:
s11, classifying the blood sugar condition evaluation function curves when hypoglycemia does not occur;
(1) Function curve initiationAt time t 0 The blood sugar is in the normal range, and the final time t T Still in the normal range.
(2) Function curve initial time t 0 The blood sugar is in the normal range and the final time t T Between the lower normal glycemic limit and the hypoglycemic norm.
(3) Function curve initial time t 0 The blood sugar is between the lower normal blood sugar limit and the hypoglycemic standard value, and the final time t T Still between the lower normal glycemic limit and the hypoglycemic norm value.
S12, under the ideal extreme condition, respectively constructing 3 blood sugar condition evaluation function expressions according to 3 classifications of the blood sugar condition evaluation function curve in the step S11, and determining the minimum value y of the blood sugar condition evaluation function when hypoglycemia does not occur under the 3 classification conditions (1) 、y (2) 、y (3) ；
(1) Blood glucose data is equal to the normal blood glucose lower limit:
Figure BDA0003608516320000171
(2) Initial time t 0 Until the blood sugar data in t is equal to the lower limit of normal blood sugar value, t to the final time t T Internal blood glucose data equal to the hypoglycemic norm:
Figure BDA0003608516320000172
(3) The blood glucose data are all at the hypoglycemic standard value:
Figure BDA0003608516320000173
s13, comparing the minimum value y of the evaluation function of the blood sugar condition without hypoglycemia under the three classification conditions in the step S12 (1) 、y (2) 、y (3) Selecting y (1) 、y (2) 、y (3) The medium-minimum value is taken as the blood sugar condition when hypoglycemia does not occurMinimum value y of merit function 1min ；
S14, determining the maximum value y of the blood sugar condition evaluation function when hypoglycemia occurs by approximating a slope in ideal condition 1max . The simulated ideal blood glucose curve is:
Figure BDA0003608516320000174
where Δ t is the blood glucose measurement time interval and a is the confidence level, i.e., if and only if the actual blood glucose is less than (1- α) u 3 When it is, hypoglycemia is considered to occur. Maximum value y of evaluation function of blood sugar status in occurrence of hypoglycemia 1max Comprises the following steps:
Figure BDA0003608516320000181
s15, making the minimum value y of the blood sugar state evaluation function obtained in the step S13 1min Is larger than the maximum value y of the blood sugar condition evaluation function obtained in step S14 1max Obtaining a specific limit relation among blood sugar condition evaluation function coefficients A, B and C;
s16, when the blood sugar is in the normal range, taking ideal extreme conditions (the blood sugar value is in the upper limit or the lower limit of normal blood sugar) to obtain the minimum value y of the blood sugar condition evaluation function 2min ：
Figure BDA0003608516320000182
S17, calculating the maximum value y of the blood sugar condition evaluation function when the blood sugar is positioned between the lower normal blood sugar limit and the hypoglycemic limit 2max . Maximum value y of evaluation function of blood sugar status 2max Still determined by the ideal case approximating a diagonal line. And (3) simulating an ideal blood sugar curve:
Figure BDA0003608516320000183
where Δ t is the blood glucose measurement time interval and γ is the confidence level, i.e., if and only if the actual blood glucose is less than (1- γ) u 3 When, hypoglycemia is considered to occur. Maximum value y of evaluation function of blood sugar status 2max Comprises the following steps:
Figure BDA0003608516320000184
s18, making the minimum value y of the blood sugar state evaluation function obtained in the step S16 2min Is larger than the maximum value y of the blood glucose evaluation function obtained in step S17 2max Obtaining another specific limiting relation between the blood sugar condition evaluation function coefficients B and C;
s19, through specific calculation in the step S15 and the step S18, the limit relation B (u) of the blood sugar condition evaluation function coefficients A, B and C is obtained 2 -u 3 )T＜4.5Aαu 3 Δt，C(u 1 -u 2 )T＜4.5Bγu 2 At. Setting specific coefficient values meeting the limiting relation according to experimental needs, wherein the coefficient values are not changed after being determined;
it should be noted that, because the curves are diversified in practical situations, it is not easy to construct a proper ideal oblique line for representing the minimum value of y under the condition of no occurrence of hypoglycemia, so that the minimum value y of the evaluation function of the blood glucose condition under the condition of no occurrence of hypoglycemia is determined under the ideal extreme condition 1min 。
Further, the blood sugar data measured by the blood sugar sensor has certain error, and the invention sets the confidence coefficient alpha to be considered when the actual blood sugar is positioned in [ (1-alpha) u 3 ，u 3 ]In the interval, z and u 3 Approximately equal, and not considered hypoglycemic, blood glucose values are (1-. Alpha.) u when (T-2. DELTA.t) 3 (T-3. DELTA.t) blood glucose value of u 3 The maximum value y of the evaluation function of the blood sugar condition when hypoglycemia occurs can be calculated and determined 1max . The confidence α represents: the probability of hypoglycemia occurring in real world is α.
Further, the invention sets the confidence level gamma to be considered when the actual blood sugar is positioned in [ (1-gamma) u 2 ，u 2 ]In the interval, z and u 2 Approximately equal, blood glucose is not considered abnormal. When the blood sugar value is (1-gamma) u when the (T-2 delta T) 2 (T-3. DELTA.t) blood sugar value u 2 Then y can be determined 2max . With the confidence γ, the probability that the blood glucose is in an abnormal state in an actual situation is γ.
For example: and calculating the limits among the coefficients A, B and C of the blood glucose condition evaluation function by using the blood glucose historical data acquired in the period T and the time interval delta T, and selecting proper coefficients A, B and C to construct the blood glucose condition evaluation function.
The obvious reduction effect of hypoglycemia on the blood sugar condition evaluation function is ensured by the limit relation among the blood sugar condition evaluation function coefficients A, B and C. The function coefficients B and C are evaluated according to the blood sugar condition, and the function of the normal blood sugar range on the obvious rise of the objective function value is ensured. And after the blood sugar condition evaluation function coefficients A, B and C obtain the limiting relation, the corresponding numerical values of A, B and C are set by self and are not changed.
The blood sugar data acquisition unit is used for acquiring blood sugar historical data and BR set values of a diabetic patient in a certain period T under the condition of wearing the AP. Wherein, blood sugar data of the patient is obtained at intervals of delta t in the blood sugar sensing period in the AP. The period T is much greater than the time interval Δ T;
that is, the invention collects the blood sugar history data with the period T and the AP parameter (BR) of the diabetic patient wearing the AP; the blood glucose sensor samples the blood glucose of a patient at intervals of delta t, and the sampled blood glucose data comprises pre-meal blood glucose data and post-meal blood glucose data of the patient. The period T should be much larger than the sampling time at. To reflect the interference of meals with the patient's blood glucose levels, the present invention assumes that the patient consumes approximately constant amounts of carbohydrates at each meal, with the patient consuming 50g of carbohydrates at breakfast and 75g of carbohydrates at lunch and dinner, respectively. These data were collected using a UVA/Padova T1DM simulator simulation.
For example, daily blood glucose data of a patient is collected at a period of 24 hours, and the blood glucose sensor samples blood glucose of the patient at intervals of 5 min. A total of 288 blood glucose values were collected and the AP parameters (BR) at this period were recorded and recorded as 1 sample.
And the function calculation unit performs linear interpolation on the measured blood glucose data points by the AP in a certain period according to the blood glucose condition evaluation function to obtain a blood glucose data piecewise function in a certain period T, and calculates to obtain a final blood glucose condition evaluation function value through piecewise weighted integration.
Further, after selecting an appropriate blood glucose condition evaluation function coefficient, 288 blood glucose history data are linearly interpolated to obtain a blood glucose data function, and the blood glucose data function is substituted into the blood glucose condition evaluation function to perform a piecewise weighted integral to calculate a blood glucose condition evaluation function value. This value and the AP parameter (BR) form a blood glucose condition data pair processed by a gaussian process.
The individualized AP parameter (BR) learning module collects the AP parameter (BR) value sets X obtained in the previous n periods 1 With blood glucose condition evaluation function value set y 1 Respectively as training input and training output of a blood sugar condition prediction model based on a Gaussian process to obtain a blood sugar condition y prediction model M1 under different AP parameters (BR), wherein y obeys a mean value mu and a variance sigma 2 Gaussian distribution of (a).
For example, using a known set of AP parameter (BR) values X obtained for the first n cycles 1 With blood glucose condition evaluation function value set y 1 Namely, N blood sugar condition data pairs are obtained, a blood sugar condition y prediction model M1 under different AP parameters (BR) is obtained, and a covariance function in a Gaussian process is selected as a Gaussian square exponential kernel function:
Figure BDA0003608516320000211
in the formula, x is an AP parameter (BR) in different periods, and λ and h are hyper-parameters to be obtained by training and are respectively an input scale parameter and an output scale parameter. For a series of data pairs (x) i ，y i )，i∈[1，n]There is a covariance matrix:
Figure BDA0003608516320000212
through the Gaussian process training, a blood glucose condition y prediction model M1 under different AP parameters (BR) can be obtained through training.
And the Bayesian optimization module selects the optimal AP parameter (BR) under the condition of no hypoglycemia in the prediction model M1 by using a safety selection algorithm, and uses the optimal AP parameter (BR) in the next period to perform AP self-adaptation. The safety selection algorithm comprises the following steps:
s21, according to experience of doctors and patientsDetermining a priori security set S 0 And sample set X, S 0 Including the range of basal rates of insulin that physicians and patients believe will not experience hypoglycemia and hyperglycemia. And collecting blood glucose historical data under different insulin basal rates in the previous n periods. Sample set X includes the basal rate of insulin for the first n cycles (n.gtoreq.4). Determining a verification set D, wherein the range is that the insulin basic rate in the current period fluctuates up and down by 50%, and an independent variable x represents the insulin basic rate;
s22, utilizing a Gaussian process prediction model M 1 Definition of the Lipschitz constant L of Microcovariance nucleus k And error limit delta L ,
Figure BDA0003608516320000221
δ L E (0, 1). For a zero-mean gaussian process defined by a covariance kernel k (·, ·), the continuous partial derivative reaches a partial derivative kernel of fourth order:
Figure BDA0003608516320000222
and also,
Figure BDA0003608516320000223
represents the partial derivative kernel over a sample set X with a maximum distance r = max (X-X ') X, X' e.X
Figure BDA0003608516320000224
The lipschitz constant of (a);
s23, utilizing L in the step S22 k ，
Figure BDA0003608516320000225
And r, calculating a prediction model M 1 Liphoz constant of unknown function (iii):
Figure BDA0003608516320000226
s24, predicting the model M through the Gaussian process 1 Obtaining different AP parameters (BR) on the verification set D, wherein the obedience mean value is mu and the variance is sigma 2 Of a Gaussian distributionGlycemic status y and Lipschitz constant L f Calculating the confidence upper limit u (x) = mu (x) + beta for each point x on the verification set D by taking deltax as the AP parameter (BR) interval 0.5 σ (x), lower confidence limit 1 (x) = μ (x) - β 0.5 σ (x), wherein the parameter beta is a constant, and the confidence interval ω (x) is the difference between the upper confidence limit and the lower confidence limit;
the compactness of the confidence interval is controlled by the parameter β in step S24, which is generally set by the user according to the experimental situation.
S25, using the confidence limits 1 (x) of each point in step S24 and the safety set S in step S21 0 Selecting the lower confidence limit 1 (x) with the minimum value in the safety set as a safety limit h, h: = mini (x), x ∈ S 0 ；
S26, utilizing the Lipschitz constant L in the step S23 f In step S24, the upper limit set u (X), the lower limit set l (X), the confidence interval set ω (X), and the safety limit h in step S25 are signaled for the safety set S 0 Updating, wherein the updated security set is called a prior security set S t ；
Figure BDA0003608516320000231
S27, utilizing the Lipschitz constant L in the step S23 f Confidence upper bound u (x), confidence lower bound l (x) in step S24, security bound h in step S25 and a priori secure set S in step S26 t Computing a set of potential maximizers M t Extended Security set G t ；
M t ＝{x∈S t |u(x)≥maxl(x)，x∈S t }
g(x)：＝|{x∈D\S t |u(x)-Ld(x，x)≥h}|
G t ＝{x∈S t |g(x)＞0}
S28, potential maximizer set M in step S27 t And extended Security set G t And concentrating, and selecting the maximum confidence interval value max omega (x) x epsilon M t ∪G t And selecting the AP parameter (BR) corresponding to the value as the optimal AP parameter (BR).
S29, calculating posterior safety set S 0 . Selecting the optimal AP parameter (BR) in the step S28 as the insulin basal rate in the (n + 1) th cycle, collecting blood sugar data, adding the (n + 1) th data to update the sample set X, and selecting the left and right limits in the sample set as the posterior safety set S when hypoglycemia or hyperglycemia does not occur at this time 0 An upper boundary and a lower boundary. If hypoglycemia or hyperglycemia occurs, the point is removed and the adjacent known function point is selected as the upper or lower boundary.
S30, posterior safety set S in step S29 0 As a safety set S at the n +1 th cycle 0 . Step S21 is repeated.
Therefore, in order to obtain the optimal AP parameter (BR) under the condition of no hypoglycemia for the blood glucose state y prediction model M1 under different AP parameters (BR), the invention firstly sets the known prior safety set S 0 Sample set X.
It should be noted that: initial secure set S 0 Using a secure set S for a set of secure AP parameters (BR) known a priori by the patient 0 AP parameters (BR) in (1), AP performance is not the same, glycemic effect is guaranteed only by patient and doctor experience, so patient uses initial safety set S 0 Medium BR, hypoglycemia or hyperglycemia may still occur. After the safety set is updated, the hypoglycemic and hyperglycemic points are removed, and the posterior safety set S is selected 0 Medium BR does not experience hypoglycemia.
Further, the glycemic status evaluation function heavily considers the effect of hypoglycemia, and the safety margin h: = min l t (x)，x∈S 0 And the blood sugar effect values in the prior safety set are all larger than h under the existing known information, so that the blood sugar effect is continuously controlled.
It should be noted that, from the condition of Lepruschitz, under the existing information, the prior safety decision set S t The values of the middle elements are all larger than a safety limit h. On the premise of ensuring safety, the probability of selecting elements in a potential maximizer set Mt increases the evaluation function value of the blood sugar condition. And on the premise of ensuring safety, selecting elements in the expanded safety set Gt to have a probability expanded safety set.
And finally, selecting the optimal AP parameter (BR) standard under the condition of no hypoglycemia by the safety selection algorithm through a prediction model as follows: x is a radical of a fluorine atom t ∈arg maxω t (x)x∈M t ，x∈G t 。
Finally, the invention utilizes a UVA/Padova T1DM blood glucose metabolism simulator to perform simulation verification on the system performance, an AP controller implements an exemplary control algorithm and adopts ADRC, aiming at a certain patient, AP parameters (BR) are safely selected according to the steps from one step to three, and finally individualized AP self-adaptation is realized, blood glucose data of the patient with improper AP parameters (BR) is shown in figure 4, blood glucose data of the patient after 8 safe selections of the AP parameters (BR) is shown in figure 5, and insulin basal rate, blood glucose mean value and blood glucose normal occupation ratio in 8 updating processes are shown in table 1. As can be seen from fig. 4, fig. 5 and table 1, the present invention robustly adjusts the AP parameters (BR), increases the euglycemic time ratio, significantly reduces the risk of hypoglycemia, and automatically improves the performance of the long-term adaptive AP system with minimal manual intervention.
TABLE 1 AP parameter (BR) update data sheet 8 times for a patient
Basal rate Secure collections Prediction horizon Predicted point Mean blood glucose Normal ratio of occupation
0.0300 0.0450-0.1200 0.0315-0.0585 0.0580 178.2743 38.54％
0.0580 0.0450-0.1200 0.0406-0.0754 0.0754 150.3750 87.85％
0.0754 0.0450-0.1200 0.0528-0.0980 0.0980 137.2951 94.79％
0.0980 0.0450-0.1200 0.0686-0.1274 0.0949 121.3125 100.00％
0.0949 0.0450-0.1200 0.0664-0.1234 0.0681 124.6979 100.00％
0.0681 0.0450-0.1200 0.0476-0.0885 0.0885 139.7743 93.06％
0.0885 0.0450-0.1200 0.0620-0.1151 0.0967 129.1042 97.22％
0.0967 0.0450-0.1200 0.0677-0.1257 0.0703 123.6250 100.00％
0.0703 0.0450-0.1200 0.0492-0.0914 135.0868 98.96％
Therefore, the invention adopts the individualized learning system for the long-term adaptation of the artificial pancreas based on the Bayesian optimization, which comprises an individual blood sugar condition evaluation module, an individualized AP parameter (BR) learning module and a Bayesian optimization module. The individual blood glucose condition evaluation module collects historical blood glucose data and AP parameter (BR) historical data of the diabetic patient after wearing the AP. An individual blood sugar condition evaluation function is designed, the blood sugar condition level of a patient in a certain period is accurately and effectively evaluated, and a blood sugar condition function value of the patient in the certain period is obtained. The individualized AP parameter (BR) learning module establishes an individualized blood glucose condition prediction model under the AP parameter (BR) under life interference according to different environments by utilizing data consisting of the blood glucose condition function data of the patient and the AP parameter (BR) to a training Gaussian process, and relearns the blood glucose condition rule and the AP parameter (BR) at regular intervals and updates the blood glucose condition prediction model. And the Bayesian optimization module determines the optimal AP parameter (BR) under the condition of no hypoglycemia by using a safety selection algorithm for the blood sugar condition prediction model, and updates the AP parameter (BR) in a safety range. The normal blood sugar time ratio is increased, and the hypoglycemia is obviously reduced.
Finally, it should be noted that: the above embodiments are only for illustrating the technical solutions of the present invention and not for limiting the same, and although the present invention is described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that: modifications and equivalents may be made to the disclosed embodiments without departing from the spirit and scope of the present invention.

Claims (9)

1. An artificial pancreas long-term adaptation individualized learning system based on Bayesian optimization is characterized in that: the system comprises an individual blood sugar condition evaluation module, an individual AP parameter learning module and a Bayesian optimization module;
the individual blood sugar condition evaluation module constructs a blood sugar condition evaluation function y (z) according to historical blood sugar data in a period:
Figure FDA0003806518600000011
wherein z is the collected historical blood glucose data, y is the blood glucose state evaluation, T represents the historical blood glucose data collection period, u 1 、u 2 、u 3 Respectively representing the average value of the range of normal blood sugar, the lower limit of normal blood sugar and the standard value of hypoglycemia;
0-t 1 indicating that the blood glucose level is greater than the lower normal blood glucose limit u during the period T 2 Total time of (d); t is t 1 -t 2 Represents that the blood glucose level is less than the lower normal blood glucose limit u during the period T 2 But greater than the standard value u for hypoglycemia 3 Total time of (d); t is t 2 -T represents a blood glucose value less than a standard value u for hypoglycaemia during the period T 3 Total time of (d);
the first term is a hypoglycemic penalty term: a is a hypoglycemia penalty term coefficient which represents that when hypoglycemia occurs, the target function value is obviously reduced compared with the target function value when hypoglycemia does not occur; the second term is an abnormal blood sugar penalty term: b is an abnormal blood sugar penalty term coefficient which represents that when abnormal blood sugar occurs but hypoglycemia does not occur, the objective function value is obviously reduced compared with the objective function values in the range of normal blood sugar; the third term represents the effect of the range of normal blood glucose on the objective function value, and C is the coefficient of this term.
2. The system of claim 1, wherein the system is characterized in that: the individualized AP parameter learning module is used for establishing a rule between the historical blood sugar state evaluation y of the diabetic and the AP parameter by utilizing a Gaussian process to obtain a prediction model of the blood sugar state evaluation y under different AP parameters;
the individualized AP parameter learning module is used for acquiring AP parameter value sets X in the first n periods 1 Evaluation function value set Y of blood sugar condition 1 Respectively as training input and training output of a blood sugar condition prediction model based on a Gaussian process to obtain a blood sugar condition evaluation y prediction model M1 under different AP parameters, wherein the blood sugar condition evaluation y obeys a mean value of mu and a variance of sigma under different AP parameters 2 A gaussian distribution of (a).
3. The system of claim 1, wherein the system is characterized in that: and the Bayesian optimization module is used for safely selecting the optimal AP parameter according to the blood sugar condition prediction model to obtain the AP parameter of the next period.
4. The system of claim 1, wherein the system is characterized in that: the individual blood sugar condition evaluation module comprises a parameter determination unit, a blood sugar data acquisition unit and a function calculation unit;
the parameter determining unit obtains the average value u of the time interval T, the period T and the normal blood sugar range according to the blood sugar data in the AP system 1 Lower limit of normal blood glucose u 2 And a standard value u for hypoglycemia 3 And determining blood sugar condition evaluation function coefficients A, B and C respectively, and reflecting the obvious rising effect of the proportion of hypoglycemia influence and normal blood sugar on function values by limiting the function coefficients.
5. The system for individualized learning of long-term adaptation of artificial pancreas based on Bayesian optimization as recited in claim 4, wherein the coefficients A, B and C are determined by the following steps:
s11, classifying the blood sugar condition evaluation function curves when hypoglycemia does not occur;
(1) Function curve initial time t 0 The blood sugar is in the normal range and the final time t T Is still in the normal range;
(2) Function curve initial time t 0 The blood sugar is in the normal range, and the final time t T Is between the lower normal glycemic limit and the hypoglycemic norm value;
(3) Function curve initial time t 0 The blood sugar is between the lower normal blood sugar limit and the hypoglycemic standard value, and the final time t T Is still between the lower normal glycemic limit and the hypoglycemic standard value;
s12, under the ideal extreme condition, respectively constructing 3 blood sugar condition evaluation function expressions according to 3 classifications of the blood sugar condition evaluation function curve in the step S1, and determining the minimum value y of the blood sugar condition evaluation function when hypoglycemia does not occur under the 3 classification conditions (1) 、y (2) 、y (3) ；
Blood glucose data is equal to the normal blood glucose lower limit:
Figure FDA0003806518600000031
initial time t 0 Until the blood sugar data in t is equal to the lower limit of normal blood sugar value, t to the final time t T The internal blood glucose data is equal to the hypoglycemic normative value:
Figure FDA0003806518600000032
the blood glucose data are all at a hypoglycemic standard value:
Figure FDA0003806518600000033
s13, comparing the minimum value y of the evaluation function of the blood sugar condition without hypoglycemia under the three classification conditions in the step S12 (1) 、y (2) 、y (3) Selecting y (1) 、y (2) 、y (3) Minimum value as minimum value y of evaluation function of blood sugar status when hypoglycemia does not occur 1min ；
S14, determining the maximum value y of the blood sugar condition evaluation function when hypoglycemia occurs by approximating a slope in ideal condition 1max The simulated ideal blood glucose curve is:
Figure FDA0003806518600000034
where Δ t is the blood glucose measurement time interval and α is the confidence, i.e., if and only if the actual blood glucose is less than (1- α) u 3 When hypoglycemia occurs, the maximum value y of the evaluation function of the blood sugar condition is considered 1max Comprises the following steps:
Figure FDA0003806518600000035
s15, minimizing the blood sugar evaluation function obtained in step S13Value y 1min Is larger than the maximum value y of the blood sugar condition evaluation function obtained in step S14 1max Obtaining a specific limiting relation among blood sugar condition evaluation function coefficients A, B and C;
s16, when the blood sugar is in the normal range, taking the ideal extreme condition, namely that the blood sugar is in the upper limit or the lower limit of normal blood sugar, and obtaining the minimum value y of the blood sugar condition evaluation function 2min ：
Figure FDA0003806518600000041
S17, calculating the maximum value y of the blood sugar condition evaluation function when the blood sugar is positioned between the lower normal blood sugar limit and the hypoglycemic limit 2max (ii) a Maximum value y of blood glucose evaluation function 2max Still approximate a slash through the ideal case to confirm; and (3) simulating an ideal blood sugar curve:
Figure FDA0003806518600000042
where Δ t is the blood glucose measurement time interval and γ is the confidence level, i.e., if and only if the actual blood glucose is less than (1- γ) u 3 When, hypoglycemia is considered to have occurred;
maximum value y of blood glucose evaluation function 2max Comprises the following steps:
Figure FDA0003806518600000043
s18, making the minimum value y of the blood sugar state evaluation function obtained in the step S16 2min Is larger than the maximum value y of the blood glucose evaluation function obtained in step S17 2max Obtaining another specific limiting relation between the blood sugar condition evaluation function coefficients B and C;
s19, obtaining the limiting relation B (u) of the blood sugar condition evaluation function coefficients A, B and C through specific calculation in the steps S15 and S18 2 -u 3 )T＜4.5Aαu 3 Δt，C(u 1 -u 2 )T＜4.5Bγu 2 Δ t; and setting specific coefficient values meeting the limiting relation according to experimental needs, wherein the coefficient values are not changed after being determined.
6. The system of claim 5, wherein the system comprises: the standard of hypoglycemia is that blood sugar data sampled for three times continuously are all smaller than a standard value u of hypoglycemia 3 ；
The blood glucose vector is in an abnormal state standard that the blood glucose concentration is lower than a normal range after three times of continuous sampling.
7. The system of claim 4, wherein the system comprises: the blood sugar data acquisition unit is used for acquiring historical blood sugar data and a set value of an insulin basic rate BR of a diabetic patient in a certain period T when the diabetic patient wears the AP, wherein the blood sugar data of the diabetic patient is acquired at intervals of time T in a blood sugar sensing period in the AP.
8. The system of claim 4, wherein the system comprises: and the function calculation unit performs linear interpolation on the measured blood glucose data points by the AP in a certain period according to the blood glucose condition evaluation function to obtain a blood glucose data piecewise function in a certain period T, and calculates to obtain a final blood glucose condition evaluation function value through piecewise weighted integration.
9. The system of claim 1, wherein the system is characterized in that: the Bayesian optimization module selects the optimal AP parameter under the condition of no hypoglycemia in the prediction model M1 by using a safety selection algorithm, and uses the optimal AP parameter in the next period to perform AP self-adaptation;
the safety selection algorithm comprises the following steps:
s21, determining a prior safety set S according to experience of doctors and patients 0 And sample set X, S 0 A range of basal rates of insulin including those considered by physicians and patients to be non-hypoglycemic and non-hyperglycemic; collecting historical blood glucose data under different insulin basal rates in the previous n periods,the sample set X comprises the insulin basic rate under the previous n cycles, wherein n is more than or equal to 4, the verification set D is determined, the range is that the insulin basic rate in the current cycle fluctuates by 50 percent from top to bottom, and the independent variable X represents the insulin basic rate;
s22, forecasting model M by utilizing Gaussian process 1 Definition of the Lipschitz constant L of Microcovariance nucleus k And error limit delta L ，
Figure FDA0003806518600000061
δ L E (0, 1); for a zero-mean gaussian process defined by a covariance kernel k (·, ·), its continuous partial derivatives reach a partial derivative kernel of order four:
Figure FDA0003806518600000062
and the number of the first and second electrodes,
Figure FDA0003806518600000063
represents the partial derivative kernel over a sample set X with a maximum distance r = max (X-X ') X, X' e.X
Figure FDA0003806518600000064
The liphoz constant of (a);
s23, using L in step S22 k ，
Figure FDA0003806518600000065
And r calculating a prediction model M 1 Liphoz constant of unknown function (iii):
Figure FDA0003806518600000066
s24, predicting the model M through the Gaussian process 1 Obtaining different AP parameters on the verification set D, wherein the obedience mean value is mu and the variance is sigma 2 (ii) the evaluation of the glycemic State of Gaussian distribution of (y) and the Leptozetz constant L f Calculating the confidence upper limit u (x) = mu (x) + beta of each point x on the verification set D by taking deltax as an AP parameter interval 0.5 σ (x), lower confidence limit l (x) = μ (x) - β 0.5 σ (x), wherein the parameter beta is a constant, and the confidence interval ω (x) is the difference between the upper confidence limit and the lower confidence limit;
s25, using the confidence lower limit l (x) of each point in the step S24 and the safety set S in the step S21 0 Selecting the lower confidence limit l (x) with the minimum value in the safety set as the safety limit h, h = minl (x), x belongs to S 0 ；
S26, utilizing the Lipschitz constant L in the step S23 f In step S24, the upper limit set u (x), the lower limit set l (x), the confidence interval set ω (x), and the safety limit h in step S25 are signaled for the safety set S 0 Updating is carried out, and the updated security set is called a prior security set S t ；
Figure FDA0003806518600000071
S27, utilizing the Lipschitz constant L in the step S23 f Confidence upper bound u (x), confidence lower bound l (x) in step S24, security bound h in step S25 and a priori secure set S in step S26 t Computing a set of potential maximizer M t Expanding the secure set G t ；
M t ＝{x∈S t |u(x)≥maxl(x′)，x′∈S t }
g(x)＝|{x′∈D\S t |u(x)-L f d(x，x′)≥h}|
G t ＝{x∈S t |g(x)＞0}
S28, potential maximizer set M in step S27 t And extended Security set G t And in the union set, selecting the maximum confidence interval value max omega (x), wherein x belongs to M t UG t Selecting an AP parameter corresponding to max omega (x) as an optimal AP parameter;
s29, calculating posterior safety set S 0 (ii) a Selecting the optimal AP parameter in the step S28 as the insulin basal rate in the (n + 1) th cycle, collecting blood sugar data, adding the (n + 1) th data to update the sample set X, and selecting the sample when the hypoglycemia or hyperglycemia does not occur at the timeCentralizing left and right bounds as posterior safety set S 0 An upper boundary and a lower boundary; if hypoglycemia or hyperglycemia occurs, removing the AP parameter, and selecting an adjacent known function point as an upper boundary or a lower boundary;
s30, posterior safety set S in step S29 0 As a safety set S at the n +1 th cycle 0 (ii) a Step S21 is repeated.
CN202210425970.0A 2022-04-21 2022-04-21 Artificial pancreas long-term adaptation individualized learning system based on Bayesian optimization Active CN114613509B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210425970.0A CN114613509B (en) 2022-04-21 2022-04-21 Artificial pancreas long-term adaptation individualized learning system based on Bayesian optimization

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210425970.0A CN114613509B (en) 2022-04-21 2022-04-21 Artificial pancreas long-term adaptation individualized learning system based on Bayesian optimization

Publications (2)

Publication Number Publication Date
CN114613509A CN114613509A (en) 2022-06-10
CN114613509B true CN114613509B (en) 2022-11-08

Family

ID=81870100

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210425970.0A Active CN114613509B (en) 2022-04-21 2022-04-21 Artificial pancreas long-term adaptation individualized learning system based on Bayesian optimization

Country Status (1)

Country Link
CN (1) CN114613509B (en)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109999270A (en) * 2019-03-22 2019-07-12 北京理工大学 A kind of adaptive automatic disturbance rejection controller of artificial pancreas based on change of blood sugar trend
CN110111898A (en) * 2019-04-26 2019-08-09 北京理工大学 Multifunction manual pancreas control module and method based on mobile terminal
CN112133439A (en) * 2020-09-07 2020-12-25 北京理工大学 Pre-meal insulin dose individualized decision making system based on Gaussian process
CN112927802A (en) * 2021-01-28 2021-06-08 北京理工大学 Expert experience assisted preprandial insulin dosage learning optimization decision system

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210259591A1 (en) * 2020-02-20 2021-08-26 Dexcom, Inc. Machine learning in an artificial pancreas

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109999270A (en) * 2019-03-22 2019-07-12 北京理工大学 A kind of adaptive automatic disturbance rejection controller of artificial pancreas based on change of blood sugar trend
CN110111898A (en) * 2019-04-26 2019-08-09 北京理工大学 Multifunction manual pancreas control module and method based on mobile terminal
CN112133439A (en) * 2020-09-07 2020-12-25 北京理工大学 Pre-meal insulin dose individualized decision making system based on Gaussian process
CN112927802A (en) * 2021-01-28 2021-06-08 北京理工大学 Expert experience assisted preprandial insulin dosage learning optimization decision system

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
"Multivariate learning framework for long-term adaptation in the artificial pancreas";Dawei Shi et al;《Bioengineering & Translational Medicine》;20190131;第4卷(第1期);第61-74页 *
Dawei Shi et al."Multivariate learning framework for long-term adaptation in the artificial pancreas".《Bioengineering & Translational Medicine》.2019,第4卷(第1期), *
基于胰岛素基础率估计的人工胰腺系统自抗扰控制;史大威 等;《自动化学报》;20210531;第47卷(第5期);全文 *

Also Published As

Publication number Publication date
CN114613509A (en) 2022-06-10

Similar Documents

Publication Publication Date Title
US8690820B2 (en) Automatic insulin pumps using recursive multivariable models and adaptive control algorithms
CN105050539B (en) For the daily periodic object intervals control in the Model Predictive Control problem of the artificial pancreas of type 1 diabetes application
Harvey et al. Quest for the artificial pancreas: combining technology with treatment
EP2400882B1 (en) Method and system for automatic monitoring of diabetes related treatments
Khodaei et al. Physiological closed-loop control (PCLC) systems: Review of a modern frontier in automation
CN112133439B (en) Pre-meal insulin dose individualized decision making system based on Gaussian process
JP7019127B2 (en) Insulin assessment based on reinforcement learning
CN112927802B (en) Expert experience assisted preprandial insulin dosage learning optimization decision system
Boiroux et al. Overnight control of blood glucose in people with type 1 diabetes
CN110289094A (en) Decision-making technique is precisely administered in a kind of insulin based on Expert Rules
Zarkogianni et al. Personalized glucose-insulin metabolism model based on self-organizing maps for patients with Type 1 Diabetes Mellitus
CN114903648B (en) Double-hormone artificial pancreas controller based on ESO and model predictive control
CN116139367A (en) Self-adaptive insulin infusion method and system
CN114613509B (en) Artificial pancreas long-term adaptation individualized learning system based on Bayesian optimization
Boiroux et al. A nonlinear model predictive control strategy for glucose control in people with type 1 diabetes
Batora et al. Bihormonal control of blood glucose in people with type 1 diabetes
Matabuena et al. Multilevel functional distributional models with application to continuous glucose monitoring in diabetes clinical trials
Hettiarachchi et al. G2P2C—A modular reinforcement learning algorithm for glucose control by glucose prediction and planning in Type 1 Diabetes
WO2023092908A1 (en) Artificial pancreas control method, readable storage medium, and blood glucose management system
Taghvafard et al. Design of intermittent control for cortisol secretion under time-varying demand and holding cost constraints
Ata et al. Model based control of artificial pancreas under meal disturbances
Sun et al. Impact of errors in carbohydrate estimation on control of blood glucose in type 1 diabetes
Rashid et al. Predictive control with variable delays in plasma insulin action for artificial pancreas
Sevil et al. Automated insulin delivery systems for people with type 1 diabetes
CN114464291B (en) MDI dosage suggestion system based on Bayesian optimization

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant