【摘要】 目的 探讨长短期记忆网络与门循环单元(LSTM-GRU)模型对1型糖尿病(T1DM)和2型糖尿病(T2DM)患者低血糖的预警价值。方法 回顾性分析2015年7月至2017年3月于河南省人民医院内分泌科住院期间曾发生低血糖事件的50例糖尿病(DM)患者的临床资料，其中T1DM患者18例，T2DM患者32例。以连续72 h的血糖数据为研究对象，运用Python 3.6运行LSTM-GRU模型，得到15、30、45、60 min的预测血糖值，以均方根误差(RMSE)、平均绝对百分误差(MAPE)及克拉克(Clarke)误差网格分析评价模型预测性能。采用灵敏度、特异度和准确度评价模型低血糖预警效果，进一步比较模型在T1DM和T2DM患者的低血糖预警差异。结果 15 min预测时，LSTM-GRU模型的RMSE、MAPE分别为0.24、2.64;30 min预测时，RMSE、MAPE分别为0.26、2.84;45 min预测时，RMSE、MAPE分别为0.27、2.89;60 min预测时，RMSE、MAPE分别为0.27、2.85。Clarke误差网格分析表明该模型对血糖的预测准确度均满足ISO 15197—2013标准。LSTM-GRU模型在15 min低血糖预警时，T1DM和T2DM患者的灵敏度、特异度、准确度分别为95.54%、98.41%、98.10%,88.82%、99.47%、99.05%;30 min预警时，T1DM和T2DM患者的灵敏度、特异度、准确度分别为94.49%、98.41%、97.98%,87.94%、99.44%、99.01%;45 min预警时，T1DM和T2DM患者的灵敏度、特异度、准确度分别为94.52%、98.49%、98.02%,85.53%、99.48%、98.98%;60 min预警时，T1DM和T2DM患者的灵敏度、特异度、准确度分别为92.78%、98.54%、97.92%,85.15%、99.46%、98.95%。无论在哪个预测时长下，LSTM-GRU模型对T1DM和T2DM患者的低血糖预警效果比较，差异有统计学意义(P<0.05)。结论 LSTM-GRU模型能有效进行低血糖预警，且对T1DM患者的低血糖预警效果优于T2DM患者。更多还原

【Abstract】 Objective To explore the early warning value of the long short term memory-gated recurrent unit(LSTM-GRU) model for hypoglycemia in patients with type 1 diabetes(T1 DM) and type 2 diabetes(T2 DM).Methods The clinical data of 50 patients with diabetes mellitus(DM) who had hypoglycemia during hospitalization in the Department of Endocrinology in Henan Provincial People’s Hospital from July 2015 to March 2017 were retrospectively analyzed, including 18 patients with T1 DM and 32 patients with T2 DM. Taking the blood glucose data of 72 hours as the research object, the LSTM-GRU model was run by Python 3.6, and the predicted blood glucose values of 15, 30, 45 and 60 minutes were obtained. The prediction performance of the model was evaluated by root mean square error(RMSE), mean absolute percentage error(MAPE) and Clarke error grid analysis. Sensitivity, specificity and accuracy were used to evaluate the hypoglycemia warning effect of the model, and the hypoglycemia warning differences of the model in T1 DM and T2 DM patients were further compared.Results RMSE and MAPE of the LSTM-GRU model were 0.24 and 2.64 at 15 minutes prediction. RMSE and MAPE were 0.26 and 2.84 at 30 minutes prediction. RMSE and MAPE were 0.27 and 2.89 at 45 minutes prediction. RMSE and MAPE were 0.27 and 2.85 at 60 minutes prediction. Clarke error grid analysis showed that the prediction accuracy of the model for blood glucose all met the ISO 15197—2013 standard. When the LSTM-GRU model was used for early warning of hypoglycemia at 15 minutes, the sensitivity, specificity and accuracy of T1 DM and T2 DM patients were 95.54%, 98.41%, 98.10% and 88.82%, 99.47%, 99.05%. At the 30-minute warning, the sensitivity, specificity and accuracy of patients with T1 DM and T2 DM were 94.49%, 98.41%, 97.98% and 87.94%, 99.44%, 99.01%, respectively. At 45-minute warning, the sensitivity, specificity and accuracy of T1 DM and T2 DM patients were 94.52%, 98.49%, 98.02% and 85.53%, 99.48%, 98.98%, respectively. At 60-minute warning, the sensitivity, specificity and accuracy of T1 DM and T2 DM patients were 92.78%, 98.54%, 97.92% and 85.15%, 99.46%, 98.95%, respectively. No matter which prediction time, LSTM-GRU model had a statistical difference in the early warning effect of hypoglycemia between patients with T1 DM and T2 DM(P<0.05).Conclusion The LSTM-GRU model can effectively warn of hypoglycemia, and the warning effect in T1 DM patients is better than that in T2 DM patients.更多还原