目的研究比较长短时记忆（LSTM）神经网络的单维输入、多维输入模型以及反向传播（BP）神经网络在动态血糖领域的预测效果。
方法研究获取2021年6月至2022年1月天津医科大学朱宪彝纪念医院18例2型糖尿病患者血糖值及其中1例患者的运动步数、饮食摄入热量数据，建立以血糖为数据的单维输入模型和以血糖、步数、摄入热量为数据的多维输入模型两种基于深度学习中LSTM神经网络的预测模型，设置预测未来血糖的时间梯度分别为6、12、24 h，再计算预测值与实际值的均方根误差（RMSE）来评估模型间的差异，并同时基于BP神经网络建立单维预测模型，并与LSTM神经网络结果进行对比。
结果6、12、24 h预测时，基于LSTM神经网络的单维预测模型的RMSE分别为0.47、0.55、0.61；多维预测模型的RMSE分别为0.31、0.50、0.56；基于BP神经网络的单维预测模型的RMSE分别为0.38、0.59、0.63。基于LSTM神经网络建立的单维预测模型与多维预测模型都展现出了较高的精确度，且模型精确度随预测时间梯度增加而降低。对同一患者进行对比之后可确定，多维预测模型的精确度更高；除此之外，LSTM神经网络的结果精确度比BP神经网络结果精确度更高。
结论LSTM神经网络可作为动态血糖预测领域的有效手段。
ObjectiveThis study aims to compare the performance of the single-dimensional input model and the multi-dimensional input model of the long short-term memory (LSTM) neural network and the backpropagation neural network in predicting dynamic blood glucose levels.
MethodsBlood glucose values from 18 patients with type 2 diabetes from June 2021 to January 2022 in Tianjin Medical University Chu Hsien-I Memorial Hospital were collected, including one patient whose steps recorded during exercise and dietary calorie intake data were also obtained. Two prediction models based on LSTM neural network were developed: a one-dimensional model with blood glucose as input and a multi-dimensional model with blood glucose, steps, and intake calories as inputs. Time intervals of 6 hours, 12 hours, and 24 hours were set for predicting future blood glucose levels. The root mean square error (RMSE) between the predicted and actual values was calculated to assess the performance of the models. Additionally, the results of the single-dimensional prediction model based on the backpropagation neural network were compared with those of the LSTM neural network model.
ResultsThe RMSEs for the 6-hour, 12-hour, and 24-hour predictions were 0.47, 0.55 and 0.61, respectively, for the single-dimensional LSTM model. The RMSEs for the multi-dimensional LSTM model were 0.31, 0.50 and 0.56, respectively. The average RMSEs for the single-dimensional backpropagation model were 0.38, 0.59 and 0.63, respectively. Both the single-dimensional and multi-dimensional LSTM models demonstrated high accuracy, with accuracy decreasing as the prediction time gradient increased. Comparison of the results for one patient indicated that the multi-dimensional model was more accurate. Furthermore, the LSTM neural network exhibited higher accuracy compared to the backpropagation neural network.
ConclusionThe LSTM neural network proves to be an effective method for dynamic blood glucose prediction.
司家瑞,杨逸飞,迪力亚尔·阿不都克热木,等. 基于长短时记忆神经网络的动态血糖预测研究[J]. 数字医学与健康,2023,01(01)：22-27.
DOI:10.3760/cma.j.cn101909-20230720-00004版权归中华医学会所有。
未经授权，不得转载、摘编本刊文章，不得使用本刊的版式设计。
除非特别声明，本刊刊出的所有文章不代表中华医学会和本刊编委会的观点。
司家瑞：设计试验、实施研究、采集数据、文章审阅、研究指导；杨逸飞：实施研究、采集数据、分析数据、文章起草、统计分析；迪力亚尔•阿不都克热木：采集数据、分析数据；李晶：文章审阅、支持性贡献
文档即答
智问智答
机器翻译
你好，我可以帮助您更好的了解本文，请向我提问您关注的问题。
