Model characteristics including default variable importance of random forest models used to determine individual feature contributions.

The feature’s importance is calculated by removing the feature of interest from the combined model, retrain, evaluate the model by correlating measured and predicted postprandial glucose responses and subsequently subtract the correlation R from the one obtained from the combined model including all of the features. Hyperparameter names in the models correspond to parameter names from the h2o package in R. Training and test evaluations as the mean squared error, root mean squared error, mean absolute error, root mean squared logarithmic error, mean of the residuals, correlation between predicted and measured postprandial plasma glucose concentrations including confidence intervals and P values, respectively, are presented.
(XLSX)

本研究中特征重要性的计算方法如下：从组合模型中移除目标特征，重新训练模型，通过关联实测与预测的餐后血糖响应对模型进行评估，随后将该评估得到的相关系数R值与包含全部特征的组合模型所得到的R值相减。模型中的超参数名称与R语言h2o包中的参数名称保持一致。训练与测试集的评估指标包括均方误差、均方根误差、平均绝对误差、均方根对数误差、残差均值，以及预测值与实测餐后血浆葡萄糖浓度间的相关性，各项指标均附带置信区间与P值。
(XLSX)