Comparing filters and a non-mechanistic forecasting machine.

The different filter configurations include all filters off, state-only filters and dual (state and parameter) filters enabled. The non-mechanistic forecast is provided by the Gaussian Process Model Regression (GPMR). In all but one case, the dual UKF based on the ultradian model provides the most accurate forecast—although model averaging in Table 4 provides an even more accurate forecast. While the GPMR does have a potentially useful MSE, it is not as useful for tracking the mean glucose state because it is trained only on glucose measurements surrounding meal times.

不同的滤波配置包含三种形式：所有滤波器关闭、仅启用状态滤波器，以及启用双（状态与参数）滤波器。非机理预测任务由高斯过程模型回归（Gaussian Process Model Regression, GPMR）完成。除某一特定场景外，基于超昼夜模型的双无迹卡尔曼滤波（Unscented Kalman Filter, UKF）可获得精度最高的预测结果——尽管表4中的模型平均方案可实现更高的预测精度。尽管高斯过程模型回归（GPMR）的均方误差（Mean Squared Error, MSE）具备一定应用价值，但由于其仅基于进餐时段周边的血糖测量数据进行训练，因此在追踪平均血糖状态方面的实用性相对不足。