人工胰脏中数据驱动个体血糖代谢模型的辨识

Data-driven model identification for individual glucose metabolism in an artificial pancreas system

数据驱动时间序列模型是人工胰脏系统中最常用的一类血糖预测模型,但其血糖预测精度受到进食不确定性和胰岛素敏感性波动等实际因素的影响。本文从真实血糖测量数据入手,提出基于卡尔曼滤波参数估计的带输入误差滑动平均模型的辨识方法,将辨识结果与最小二乘法辨识结果进行对比。结果表明,本文提出的辨识方法具有辨识精度高(FIT:90.05±3.12%v.s.54.41±9.56%)、能有效抵消实际因素的影响、对不同特征的个体能获得稳定的辨识结果等优势。

Although data-driven time sequence models are generally employed as the glucose prediction models in APS,their glucose prediction accuracy is influenced by insulin sensing fluctuation,diet uncertainty and other factors. In this paper,starting from real clinical glucose measurement data,a new identification method called data-driven model identification method is proposed,which is based on the autoregressive moving average model with exogenous error inputs and its parameters are estimated with Kalman filter parameter estimation（KF-ARMAX）. The glucose prediction identification results for the proposed method were compared with those for the LS-ARMAX（ARMAX identification based on least square parameter estimation） method; the proposed method has the following advantages of higher identification accuracy（ FIT：90. 05 ± 3. 12% vs. 54. 41 ± 9. 56%）,being able to counteract the effects of various practical factors,being able to obtain relatively stable identification results for the individuals with different characteristics.