基于长短时记忆的1型糖尿病患者运动模式下的血糖预测  

Research on long short-term memory-based glucose prediction in patients with type 1 diabetes mellitus under exercise mode

作　　者：刘昱鑫 郁磊 常利利 李嘉旭 张晴 LIU Yuxin;YU Lei;CHANG Lili;LI Jiaxu;ZHANG Qing(Complex System Research Center,Shanxi University,Taiyuan 030006,China;Shanxi Key Laboratory of Mathematical Techniques and Big Data Analysis on Disease Control and Prevention,Shanxi University;Department of Mathematics,University of Louisville;Commonwealth Scientific and Industrial Research Organisation)

机构地区：[1]山西大学复杂系统研究所,太原030006 [2]山西大学疾病防控的数学技术与大数据分析山西省重点实验室 [3]路易斯维尔大学数学系 [4]澳大利亚联邦科学与工业研究院

出　　处：《山西医科大学学报》2022年第2期215-220,共6页Journal of Shanxi Medical University

基　　金：国家自然科学基金资助项目(61803242);国家留学基金管理委员会资助项目(留金美[2020]1417号);山西省重点研发计划资助项目(201703D421012,201803D31032);中央引导地方科技发展专项资金资助项目(YDZX20191400002563)。

摘　　要：目的基于长短时记忆(long short-term memory,LSTM)网络,构建1型糖尿病(type 1 diabetes mellitus,T1DM)患者运动模式下的血糖预测模型,并研究该模型预测精度。方法使用UVa/Padova 1型糖尿病血糖代谢仿真器(type 1 diabetes metabolic simulator,T1DMS)模拟了30名受试者在不同的运动时间及运动时长下的血糖数据。利用LSTM神经网络构建糖尿病患者运动模式下的血糖预测模型。将预测结果与支持向量回归(support vector regression,SVR)、循环神经网络(recurrent neural network,RNN)进行对比。研究不同输入特征对模型预测性能的影响,并分析仅使用历史血糖值作为模型输入时高、低血糖预测的准确率,通过克拉克误差网络分析(error grid analysis,EGA)对血糖预测结果进行评估。结果在运动模式下,相比于SVR与RNN,基于LSTM的血糖预测模型在不同预测时长下均具有更小的预测误差。当模型输入序列时长达25 min时,高、低血糖事件在20 min预测时长下的预测准确率分别可达94.12%和90.59%,EGA的A、B区域占比可达100%。结论在运动模式下,基于LSTM的血糖预测模型表现突出。基于LSTM的血糖预测模型在仅使用历史血糖值作为模型输入的条件下,也可以实现血糖的短期预测,且能够较准确预测到高、低血糖事件的发生。Objective To build a glucose prediction model based on long short-term memory(LSTM)network for type 1 diabetes mellitus(T1DM)patients under exercise mode and study its performance.Methods The blood glucose data from 30 subjects with diffe-rent exercise times and durations were generated using UVa/Padova type 1 diabetes metabolic simulator(T1DMS).The LSTM neural network that was capable of forecasting glucose levels under exercise mode for diabetes was established.Prediction results of the model were compared with support vector regression(SVR)and recurrent neural network(RNN).The model performance was studied under different input features.And the accuracy for predicting hyperglycemia and hypoglycemia was evaluated when only glucose values served as model input.The glucose prediction results were assessed by Clark error grid analysis(EGA).Results In exercise mode,the LSTM-based glucose prediction model had smaller error at all prediction horizons than SVR and RNN.When the sequence served as model input was 25 min,the accuracy for predicting hyperglycemia and hypoglycemia reached 94.12%and 90.59%at 20 min prediction horizon,respectively,and the regions A and B of EGA accounted for 100%.Conclusion In exercise mode,the LSTM-based glucose prediction model has outstanding performance.Short-term glucose prediction can be achieved with only historical glucose values as model input,and the occurrence of hyperglycemia and hypoglycemia could be predicted more accurately.

关 键 词：血糖预测 长短时记忆 运动模式 Ⅰ型糖尿病 克拉克误差网络分析 

分 类 号：R587.1[医药卫生—内分泌]