第三部分描述了用于本研究中模型开发和评估的公开可用的俄亥俄数据集/UVA、PD/directnet1/directnet2 数据集

第三部分描述了用于本研究中模型开发和评估的公开可用的俄亥俄数据集/UVA、PD/directnet1/directnet2 数据集

第四节介绍了 BGL 预测任务的开发方法。
第五节进行了实验验证和讨论。
第六节得出结论。

To develop the forecasting algorithms, the publicly available Ohio T1DM dataset [27], [28] was used, containing eight weeks’ worth of data for 12 people with T1DM. The data of six patients (PID numbers 559, 563, 570, 575, 588, and 591) were released in 2018 for the first BGL prediction challenge [28], followed by releasing data for an additional six patients (PID numbers 540, 544, 552, 567, 584, and 596) regarding the second BGL prediction challenge in 2020 [27].

Data contributors included five females and seven males who were in the age range of 20-80 years old at the time of data collection. Contributors were on insulin pump therapy. There were two separate XML files for each participant for training and testing sets. The last 10 days’ worth of data for each contributor was allocated as the testing set, and the rest belonged to the training set.

The original Ohio dataset included CGM data collected every five minutes using the Medtronic Enlite CGM sensor, alongside other types of data collected from a physical activity band, physiological sensor, and self-reported life-events. In the work presented in this paper we have used CGM data only and this is in line with other work reported in the literature [21], [29]–[33]. The problem in our case is defined as a univariate time series forecasting problem that allows objective evaluation of the proposed methodology whilst still alleviates the complexity and variability associated with using models having varying capabilities for handling multivariate variables.

Table I summarises the information in terms of gender, age,and the number of data points in the Ohio dataset. More information about the dataset can be found in [27], [28].