Department of Information Engineering, University of Padova, 35137 Padova, Italy.
IEEE Trans Biomed Eng. 2012 Jun;59(6):1550-60. doi: 10.1109/TBME.2012.2188893. Epub 2012 Feb 24.
Diabetes mellitus is one of the most common chronic diseases, and a clinically important task in its management is the prevention of hypo/hyperglycemic events. This can be achieved by exploiting continuous glucose monitoring (CGM) devices and suitable short-term prediction algorithms able to infer future glycemia in real time. In the literature, several methods for short-time glucose prediction have been proposed, most of which do not exploit information on meals, and use past CGM readings only. In this paper, we propose an algorithm for short-time glucose prediction using past CGM sensor readings and information on carbohydrate intake. The predictor combines a neural network (NN) model and a first-order polynomial extrapolation algorithm, used in parallel to describe, respectively, the nonlinear and the linear components of glucose dynamics. Information on the glucose rate of appearance after a meal is described by a previously published physiological model. The method is assessed on 20 simulated datasets and on 9 real Abbott FreeStyle Navigator datasets, and its performance is successfully compared with that of a recently proposed NN glucose predictor. Results suggest that exploiting meal information improves the accuracy of short-time glucose prediction.
糖尿病是最常见的慢性疾病之一，其管理中的一个临床重要任务是预防低血糖/高血糖事件。这可以通过利用连续血糖监测 (CGM) 设备和合适的短期预测算法来实现，这些算法能够实时推断未来的血糖水平。在文献中，已经提出了几种短期血糖预测方法，其中大多数都没有利用关于膳食的信息，并且仅使用过去的 CGM 读数。在本文中，我们提出了一种使用过去的 CGM 传感器读数和碳水化合物摄入信息进行短期血糖预测的算法。预测器结合了神经网络 (NN) 模型和一阶多项式外推算法，分别用于描述血糖动力学的非线性和线性成分。通过之前发表的生理模型来描述餐后血糖出现率的信息。该方法在 20 个模拟数据集和 9 个真实的 Abbott FreeStyle Navigator 数据集上进行了评估，并成功地将其性能与最近提出的 NN 血糖预测器进行了比较。结果表明，利用膳食信息可以提高短期血糖预测的准确性。
