東華大學圖書館 |

Prior Knowledge Driven Latent Variable Based Robust Model Predictive Control for Automated Insulin Delivery System /

Record Type:	Electronic resources : Monograph/item
Title/Author:	Prior Knowledge Driven Latent Variable Based Robust Model Predictive Control for Automated Insulin Delivery System // Xiaoyu Sun.
Author:	Sun, Xiaoyu,
Description:	1 electronic resource (169 pages)
Notes:	Source: Dissertations Abstracts International, Volume: 86-11, Section: B.
Contained By:	Dissertations Abstracts International86-11B.
Subject:	Biomedical engineering. -
Online resource:	https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31996804
ISBN:	9798315724506

Prior Knowledge Driven Latent Variable Based Robust Model Predictive Control for Automated Insulin Delivery System /
Sun, Xiaoyu,

Prior Knowledge Driven Latent Variable Based Robust Model Predictive Control for Automated Insulin Delivery System /Xiaoyu Sun. - 1 electronic resource (169 pages)

Source: Dissertations Abstracts International, Volume: 86-11, Section: B.

Automated insulin delivery (AID) systems enable the regulation of blood glucose concentration (BGC) levels of people living with type 1 diabetes mellitus (TIDM) with limited user interactions. The hybrid closed-loop insulin delivery systems rely on manually entered meal and exercise information by users, while the fully-automated AID systems work without any manually entered information. The AID systems compute the amount of insulin required by the body using a controller based on the information of BGC provided by continuous glucose monitoring (CGM) system, and deliver the insulin into the subcutaneous tissues with an insulin pump. Therefore, the controller is one of the key elements that influence the performance of BGC regulation. And the controllers based on model predictive control (MPC) algorithm have shown great potential in regulating the BGC level effectively in in-silico and clinical studies due to its ability for reducing the influence of time delay by using the predictive model. In the MPC-based AID systems, the optimal insulin is calculated by minimizing the difference between predicted glucose trajectory and its reference trajectory within the pre-defined constraints, along with additional objectives, such as minimizing the use of insulin.For the AID systems based on MPC strategies, the accuracy of the glucose concentration prediction model has a considerable effect on adjusting the BGC levels. As the model's prediction error becomes larger, the obtained insulin by solving the optimization becomes inaccurate, which leads to poor BGC regulation. Besides, CGM is the major signal resource for judging the BGC levels and computing the required insulin. However, sensor failures happen frequently for the CGM sensors, including step change, spike, missing signal, and so on. Without paying attention to the CGM sensor errors, the delivered insulin amount would be wrong, which may cause severe consequences, such as extremely low or hight BGC levels, ketoacidosis, and other complications.Motivated by the above considerations, this project aims to develop a robust MPC strategy for AID systems that is consistent with the related prior knowledge. Specifically, this project focuses on two aspects: (1) modelling the glucose metabolism more accurately and utilizing the prior knowledge to improve the modelling process and robustness of the model; (2) designing an MPC strategy that is robust to CGM sensor failures for fully-automated AID systems.Aim 1: (1) Based on the latent variables (LV), hybrid glucose concentration prediction models are proposed to forecast the future BGC levels. In the developed predictive models, the infused insulin is mapped to plasma insulin concentration which affects the transport, utilization, and consumption of plasma glucose. By projecting the variables related to glucose metabolism into the LV space, it enables the model to overcome the influence of sensor noise and establishes a more accurate relationship between glucose, infused insulin, and physical activity. (2) In order to address the issues caused by potential CGM sensor failures, a robust LV-based glucose concentration prediction model is developed. In the proposed model, the prior knowledge of the potential relationship between input variables and output variables are incorporated into the LV extraction process. The potential CGM sensor failures are simplified as missing data and a mean conditional replacement structure is integrated to enable the model forecast future glucose values when short term glucose measurements are missing. The proposed robust model provides more accurate glucose concentration predictions even though some of the CGM values are not available, guaranteeing the model-based controller can obtain the optimal insulin amount in spite of short-term missing CGM data. Aim 2: (1) To address the significant variations between subjects, a personalized adaptive MPC strategy based on the LV model is proposed. In the personalized adaptive MPC strategy, the predictive model is developed and recursively updated using personalized data, such that the inter-subject variability could be captured by the model-based controller. To reduce the influence of disturbances like meal and physical activities, and reduce the risk of hypoglycemia, the hyper-parameters in the MPC strategy are adaptively tuned according to the variations in CGM signal, such that the controller could be aggressive enough to cover the effect of meal and conservative enough to avoid the overdose of insulin. Meanwhile, the control target is modified during the physical activity periods such that the exercise-induced hypoglycemia can be avoided. (2) In order to provide effective regulation of the patient's BGC level when the CGM data is missing, the robust LV-based model is incorporated and a robust MPC strategy for the AID systems is proposed. In the robust MPC strategy, the missing CGM signals are replaced with their conditional means and therefore p-reserve most information of the corresponding variables in the LV space. The robust MPC strategy can further be promoted for other categories of CGM sensor failures by incorporating with a sensor failure detection algorithm. The performance of the LV-based models are tested with a public available dataset and simulation data generated with the FDA-accepted UVa/Padova simulator and the multivariable glucose-insulin-physical activity simulator (mGIPsim) developed by Dr. Cinar's group. The results indicate that the developed model can provide more accurate prediction for the future BGC levels and is robust to short-term missing CGM signals. The efficiency of the proposed MPC strategies for the AID systems are investigated with the UVa/Padova simulator as well as mGIPsim. The results demonstrate that the personalized adaptive MPC strategy can improve the regulation performance of the AID system and the robust MPC strategy enables the AID system control BGC levels of the patient even some of the CGM values are missing without compromise the performance of the system.

English

ISBN: 9798315724506Subjects--Topical Terms:

535387
Biomedical engineering.
Subjects--Index Terms:

Artificial pancreas system

Prior Knowledge Driven Latent Variable Based Robust Model Predictive Control for Automated Insulin Delivery System /
LDR:07619nmm a22004453i 4500 001 2434874
005 20251001062545.5
006 m o d
007 cr|nu||||||||
008 261011s2025 miu||||||m |||||||eng d
020 $a 9798315724506
035 $a (MiAaPQD)AAI31996804
035 $a AAI31996804
040 $a MiAaPQD $b eng $c MiAaPQD $e rda
100 1 $a Sun, Xiaoyu, $e author. $3 3825006
245 1 0 $a Prior Knowledge Driven Latent Variable Based Robust Model Predictive Control for Automated Insulin Delivery System / $c Xiaoyu Sun.
264 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2025
300 $a 1 electronic resource (169 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 86-11, Section: B.
500 $a Advisors: Cinar, Ali Committee members: Kawaji, Keigo; Bhushan, Abhinav; Rashid, Mudassir.
502 $b Ph.D. $c Illinois Institute of Technology $d 2025.
520 $a Automated insulin delivery (AID) systems enable the regulation of blood glucose concentration (BGC) levels of people living with type 1 diabetes mellitus (TIDM) with limited user interactions. The hybrid closed-loop insulin delivery systems rely on manually entered meal and exercise information by users, while the fully-automated AID systems work without any manually entered information. The AID systems compute the amount of insulin required by the body using a controller based on the information of BGC provided by continuous glucose monitoring (CGM) system, and deliver the insulin into the subcutaneous tissues with an insulin pump. Therefore, the controller is one of the key elements that influence the performance of BGC regulation. And the controllers based on model predictive control (MPC) algorithm have shown great potential in regulating the BGC level effectively in in-silico and clinical studies due to its ability for reducing the influence of time delay by using the predictive model. In the MPC-based AID systems, the optimal insulin is calculated by minimizing the difference between predicted glucose trajectory and its reference trajectory within the pre-defined constraints, along with additional objectives, such as minimizing the use of insulin.For the AID systems based on MPC strategies, the accuracy of the glucose concentration prediction model has a considerable effect on adjusting the BGC levels. As the model's prediction error becomes larger, the obtained insulin by solving the optimization becomes inaccurate, which leads to poor BGC regulation. Besides, CGM is the major signal resource for judging the BGC levels and computing the required insulin. However, sensor failures happen frequently for the CGM sensors, including step change, spike, missing signal, and so on. Without paying attention to the CGM sensor errors, the delivered insulin amount would be wrong, which may cause severe consequences, such as extremely low or hight BGC levels, ketoacidosis, and other complications.Motivated by the above considerations, this project aims to develop a robust MPC strategy for AID systems that is consistent with the related prior knowledge. Specifically, this project focuses on two aspects: (1) modelling the glucose metabolism more accurately and utilizing the prior knowledge to improve the modelling process and robustness of the model; (2) designing an MPC strategy that is robust to CGM sensor failures for fully-automated AID systems.Aim 1: (1) Based on the latent variables (LV), hybrid glucose concentration prediction models are proposed to forecast the future BGC levels. In the developed predictive models, the infused insulin is mapped to plasma insulin concentration which affects the transport, utilization, and consumption of plasma glucose. By projecting the variables related to glucose metabolism into the LV space, it enables the model to overcome the influence of sensor noise and establishes a more accurate relationship between glucose, infused insulin, and physical activity. (2) In order to address the issues caused by potential CGM sensor failures, a robust LV-based glucose concentration prediction model is developed. In the proposed model, the prior knowledge of the potential relationship between input variables and output variables are incorporated into the LV extraction process. The potential CGM sensor failures are simplified as missing data and a mean conditional replacement structure is integrated to enable the model forecast future glucose values when short term glucose measurements are missing. The proposed robust model provides more accurate glucose concentration predictions even though some of the CGM values are not available, guaranteeing the model-based controller can obtain the optimal insulin amount in spite of short-term missing CGM data. Aim 2: (1) To address the significant variations between subjects, a personalized adaptive MPC strategy based on the LV model is proposed. In the personalized adaptive MPC strategy, the predictive model is developed and recursively updated using personalized data, such that the inter-subject variability could be captured by the model-based controller. To reduce the influence of disturbances like meal and physical activities, and reduce the risk of hypoglycemia, the hyper-parameters in the MPC strategy are adaptively tuned according to the variations in CGM signal, such that the controller could be aggressive enough to cover the effect of meal and conservative enough to avoid the overdose of insulin. Meanwhile, the control target is modified during the physical activity periods such that the exercise-induced hypoglycemia can be avoided. (2) In order to provide effective regulation of the patient's BGC level when the CGM data is missing, the robust LV-based model is incorporated and a robust MPC strategy for the AID systems is proposed. In the robust MPC strategy, the missing CGM signals are replaced with their conditional means and therefore p-reserve most information of the corresponding variables in the LV space. The robust MPC strategy can further be promoted for other categories of CGM sensor failures by incorporating with a sensor failure detection algorithm. The performance of the LV-based models are tested with a public available dataset and simulation data generated with the FDA-accepted UVa/Padova simulator and the multivariable glucose-insulin-physical activity simulator (mGIPsim) developed by Dr. Cinar's group. The results indicate that the developed model can provide more accurate prediction for the future BGC levels and is robust to short-term missing CGM signals. The efficiency of the proposed MPC strategies for the AID systems are investigated with the UVa/Padova simulator as well as mGIPsim. The results demonstrate that the personalized adaptive MPC strategy can improve the regulation performance of the AID system and the robust MPC strategy enables the AID system control BGC levels of the patient even some of the CGM values are missing without compromise the performance of the system.
546 $a English
590 $a School code: 0091
650 4 $a Biomedical engineering. $3 535387
650 4 $a Pharmacology. $3 634543
650 4 $a Health sciences. $3 3168359
650 4 $a Pharmaceutical sciences. $3 3173021
653 $a Artificial pancreas system
653 $a Automated insulin delivery system
653 $a Model predictive control
653 $a Robust control
653 $a Type 1 diabetes mellitus
690 $a 0541
690 $a 0566
690 $a 0419
690 $a 0572
710 2 $a Illinois Institute of Technology. $b Biomedical Engineering. $e degree granting institution. $3 3825007
720 1 $a Cinar, Ali $e degree supervisor.
773 0 $t Dissertations Abstracts International $g 86-11B.
790 $a 0091
791 $a Ph.D.
792 $a 2025
856 4 0 $u https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31996804