基于自适应噪声完整聚合经验模态分解-极限学习机的短期血糖预测

摘要

糖尿病患者的血糖浓度时间序列具有时变、非线性和非平稳的特点,为提高血糖预测精度,提出一种自适应噪声的完整聚合经验模态分解(CEEMDAN)与极限学习机(ELM)相结合的短期血糖预测模型.首先,利用CEEMDAN方法将患者的血糖浓度时间序列进行分解,得到不同频段的血糖分量IMF(本征模态函数)和残余分量,以降低血糖时间序列的非平稳性;然后对各血糖分量1MF和残余分量分别构建极限学习机,并将各极限学习机的预测结果融合,获得患者未来血糖浓度的预测值,提高预测精度;在此基础上,进行低血糖预警.利用从河南省人民医院内分泌科采集的56例患者的数据进行模型检验,结果表明:与ELM模型和EMD-ELM模型相比,CEEMDAN-ELM短期血糖预测模型提前45 min的预测仍可达到较高预测水平(RMSE=0.205 1,MAPE=2.1164％);低血糖预警虚警率和漏警率分别为0.97％和7.55％.血糖预测时间的延长,可以为医生和患者提供充足时间进行血糖浓度控制,提高糖尿病治疗的效果.%A short-term prediction model of blood glucose based on the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and extreme learning machine (ELM) was proposed to improve the forecasting accuracy of blood glucose,which is timing-varying,nonlinear and non-stationary in diabetics.Firstly,by means of CEEMDAN,time sequence of blood glucose was decomposed into several intrinsic mode function (IMF) with different frequencies and a residual to reduce the non-stationary.Next,ELMs were built for each IMF and residual component to improve the forecasting accuracy,and the all forecasts of ELMs were fused to produce the prediction of blood glucose.Finally,an early alarm algorithm of hypoglycemia was proposed based on the short-term prediction model.The model was verified by 56 cases of diabetic in Department of Endocrinology of Henan Province People's Hospital.The experimental results showed that comparing to ELM model and EMD-ELM model,the proposed prediction model of blood glucose based on CEEMDAN-ELM could achieve 45 min prediction in advance,the RMSE was 0.2051 and the MAPE was 2.1164％.The false alarm rate and missing alarm rate of early alarm algorithm of hypoglycemia were 0.97％ and 7.55％ respectively.The 45 min prediction ahead provided sufficient time for doctors and diabetics to control the blood glucose concentration,especially for diabetics with hypoglycemia.