人工胰脏中血糖预测模型的关键问题研究

	人工胰脏中血糖预测模型的关键问题研究
其他题名	A Study on Key Problems of Blood Glucose Prediction Model in Artificial Pancreas for Type 1 Diabetes
	李鹏
学位类型	博士
导师	王弼陡
	2015-11
学位授予单位	中国科学院大学
学位专业	机械电子工程
关键词	人工胰脏一型糖尿病数据驱动时间序列模型葡萄糖-胰岛素生理模型模型降阶硬件实现
摘要	人工胰脏是传统胰岛素强化治疗的延伸，能协助一型糖尿病患者进行血糖管理，使其血糖维持在正常范围内。传统意义的人工胰脏包含三个部分：持续血糖监测装置（血糖传感器），以一定时间间隔测量血糖浓度；注射执行装置，定量注射胰岛素调节血糖浓度；闭环控制算法，根据血糖测量值计算出合适的胰岛素注射量和注射时间。由于技术条件的限制，现有的人工胰脏系统并不能完全满足一型糖尿病患者血糖管理需求，改善闭环算法是提升控制性能的关键。模型预测控制是人工胰脏最常用的闭环算法。血糖预测模型是预测控制的基础，控制性能与模型预测准确度息息相关。因此，为了提高人工胰脏闭环性能，本文集中研究血糖预测模型应用于人工胰脏闭环的关键问题。血糖预测模型功能是根据历史血糖测量值预测未来血糖水平的波动。根据建模原理不同，血糖预测模型可以归纳为数据驱动模型和葡萄糖-胰岛素生理模型两类。数据驱动时间序列模型最为常用，但其血糖预测精度受到血糖测量值不准确、辨识方法等众多因素的影响。本文从数据驱动时间序列模型入手，分析实际因素和辨识过程对时间序列血糖模型辨识的影响。仿真数据用于评估模型类别、个体差异和进食不确定性对辨识结果影响，临床数据用来评估参数估计方法对模型辨识的影响。葡萄糖-胰岛素生理模型是根据人体血糖代谢过程建立的一系列数学表达式。生理模型中多数参数都有实际的物理意义，所以可以量化代谢过程中的影响因素，如胰岛素敏感度等。因此，生理模型应用于人工胰脏闭环算法设计有其独特的优势。本文尝试将生理模型应用于闭环算法设计，但这个过程中实现存在以下难题。 1) 复杂的生理模型往往受限于众多参数和复杂的结构，并不能直接应用于人工胰脏闭环； 2) 在人工胰脏往小型化可穿戴化发展的过程中，硬件实现是生理模型应用的另一难题； 3) 昼夜节律，精神压力和运动等因素都会影响胰岛素敏感度，这是生理模型应用过程必须考量的因素。针对上述难题，本文以 Cobelli 葡萄糖-胰岛素生理模型为例，使用常用的模型降阶方法对 Cobelli 模型进行降阶化简。然后，提出完整的生理模型硬件实现流程，实现简化后 Cobelli 模型的 FPGA 硬件实现。最后，本文选取文献中量化的胰岛素敏感因子关键点，仿真实现了一型糖尿病患者 24 小时胰岛素敏感度变化曲线。一方面，本文将三种不同的模型降阶方法应用于 Cobelli 模型的降阶简化，获得一个适用于闭环算法设计的葡萄糖-胰岛素代谢模型，这三种方法分别是Pade 逼近，Routh 逼近以及系统辨识线性逼近。结果表明，模型降阶方法有效减少了模型参数，降低了模型复杂度，使 Cobelli 模型能够用于闭环算法的设计。另一方面，本文提出完整的硬件设计流程，以完成简化后 Cobelli 模型的 FPGA硬件实现。同时，仿真实现了 24 小时胰岛素敏感度变化的昼夜模式，用以评估胰岛素敏感度对生理模型的影响。总之，本文讨论了影响数据驱动模型应用于人工胰脏的潜在问题；提出了一种复杂生理模型简化和硬件实现方法，解决了生理模型应用于人工胰脏的难题，获得能用于人工胰脏中预测控制器的简化生理模型和硬件模型。这将在一定程度上提升人工胰脏系统的闭环控制性能。
其他摘要	An artificial pancreas system(APS) is an engineering approach to assisting patients with type 1 diabetes mellitus (T1DM) in maintianing glucose levels within a normal range. The APS consists of three components: a continuous glucose monitoring device to measure glucose concentrations at a regular interval: a control algotithm to compute appropriate rates of insulin infusions; and an insulin infusion device to deliver the computed insulin doses. Due to technical limitations, the existing APS can not satisfy the needs of blood glucose management for patients with T1DM. The key problem to improve the performance of an APS is the control algorithm. Model predictive control is one of the most commonly used control algorithms. The performance of a model predictive controller for an APS depends highly on the accuracy of a blood glucose prediction model. Therefore, this thesis will focus on the utilization of blood glucose prediction models in APS controllers and the key problems faced by these models. These models can predict future glucose trends based on measured glucose values. According the principle of modeling, the commonly used models for glucose prediction in APS can be divided into two major groups, namely data-driven models and physiological models. Although data-driven models are generally employed as the glucose prediction models in the APS, their prediction accuracy varies due to inaccuracies of glucose sensing and system identification methods. In this thesis, we evaluate the effect of practical issues and identification methods on the time-series glucose prediction models. The simulation studies investigate the difference results among model types, and the clinical datasets evaluate the effect of parameters estimation methods on model identification. The glucose-insulin physiological models consist of a series of mathematical functions which describe blood glucose metabolism in human body. Most of parameters in physiological models have actual physical meanings, so the influencing factors of physiological process, such as insulin sensitivity, etc., can be quantified in these models. Therefore, these physiological models have unique advantages to be applied to the design and deployment of an APS controller. In this thesis, we will attempt to overcome the obstacles of implementing these physiological models in the APS controller. 1. The utilization of the sophisticated physiological models are confined by their numerous parameters and complex structure, 2. The hardware implementation is another problem for the physiological models to be applied in the miniaturized and wearable APS. 3. The glucose-insulin kinetics are affected by many factors including insulin sensitivity, which is the must-be-considered factor. Cobelli's glucose-insulin interactions accepted by Food Drug Administration as a substitute to animal trials. In this thesis, Cobelli's model is simplified by Pade approximant method and implemented on a field programmable gate array(FPGA) based platform as a hardware glucose predicton model. In addition, 24-hour profiles of the insulin sensitivity variation are simulated based on a diurnal pattern of insulin sensitivity. On the one hand, three model order reduction methods, namely Pade approximant, Routh approximant and system identification approximant, are used to obtain a simplified model that are suitable for the design of the APS controller. The results show that the proposed simplified model can describe the insulin-glucose metabolism process rather accurately as well as can be easily implemented. On the other hand, an entire design flow of hardware implimentation for simplified model is proposed, and Cobelli's model is implemented in an FPGA successfully. In addition, a diurnal pattern of insulin sensitivity variation for a whole day is also implemented using lookup tables in order to evaluate the effect of insulin sensitivity. In a word, this thesis discusses the potential problems of data-driven model applied to APS, and presents a method of model simplification and its hardware implementation for a complex physiological model. The successful hardware implementation of Cobelli's model will promote a wider adoption of this model that can subsititute animal trials, provide fast and reliable glucose and insulin estimation, and ultimately assist the further development of an artificial pancreas system.
语种	中文
文献类型	学位论文
条目标识符	http://ir.ciomp.ac.cn/handle/181722/49224
专题	中科院长春光机所知识产出
推荐引用方式 GB/T 7714	李鹏. 人工胰脏中血糖预测模型的关键问题研究[D]. 中国科学院大学,2015.