基于GRU网络的血糖预测方法研究

BLOOD GLUCOSE PREDICTION METHOD BASED ON GRU

针对连续血糖监测数据(Continues Glucose Monitoring,CGM)存在强烈的时变性、复杂非线性和非平稳性等问题,提出一种基于门控循环网络(Gated Recurrent Unit,GRU)的血糖预测模型。对原始时间序列数据进行平稳化处理,利用自相关系数确立模型输入序列长度,进而将原始数据序列转化为监督型学习样本。在此基础上构建GRU血糖预测模型,并与基本RNN网络、长短记忆网络、支持向量回归进行对比。结果表明,该方法具有较高预测精度,其预测步长为20步的均方根误差和平均绝对百分误差分别为0.761 2 mmol/L和7.342 7%,可为血糖闭环控制系统提供支持。

In view of the problems of continuous glucose monitoring(CGM),such as strong time-varying,complex nonlinear and non-stationary,this paper proposes a blood glucose prediction model based on GRU.The original time series data were smoothed for stationarity.The autocorrelation coefficient was used to determine the length of input sequence,and then the original data series was transformed into supervised learning samples.On this basis,the GRU blood glucose prediction model was constructed,and compared with the basic RNN,LSTM and SVM.The results show that our method has high prediction accuracy.The RMSE and MAPE of the prediction step size of 20 steps are 0.7612 mmol/L and 7.3427%respectively,which can provide support for the closed-loop control system of blood glucose.