基于注意力机制及长短期记忆神经网络的慢性阻塞性肺疾病氧减状态辨识  

Oxygen depletion status identification in COPD by using attention-based LSTM

作　　者：吴月芳 胡明昕 孙培莉[3] Wu Yuefang;Hu Mingxin;Sun Peili(Department of Internal Medicine,Nanjing University of Science and Technology Hospital,Nanjing 210094,China;School of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094,China;Department of Respiratory Medicine,the First Affiliated Hospital with Nanjing Medical University,Nanjing 210029,China)

机构地区：[1]南京理工大学医院内科,江苏南京210094 [2]南京理工大学计算机科学与工程学院,江苏南京210094 [3]南京医科大学第一附属医院呼吸内科,江苏南京210029

出　　处：《南京理工大学学报》2023年第5期629-635,共7页Journal of Nanjing University of Science and Technology

基　　金：江苏省自然科学基金(BK20201304)。

摘　　要：为提高慢性阻塞性肺疾病氧减状态的辨识性能,该文将注意力机制有效融入长短期记忆神经网络,提出了一种基于注意力机制的长短期记忆神经网络方法:首先,抽取每个待辨识状态点的四种有效鉴别特征,包括血脉氧饱和度指数、脉搏、血脉氧饱和度指数的窗口特征以及梯度特征;其次,在此特征表示的基础上,通过引入注意力机制,使用训练集来训练基于注意力机制的长短期记忆神经网络;最后,使用测试集来验证所训练模型的有效性。与多个经典机器学习算法的对比实验结果表明:所提出的基于注意力机制的长短期记忆神经网络方法的辨识模型能够准确识别氧减状态,全局性能指标曲线下面积达到了0.8531。所提方法对于慢性阻塞性肺疾病的准确诊断具有重要的参考价值。A new attention-based long short-term memory neural network(AttLSTM)is proposed by effectively integrating attention mechanism into traditional long short-term memory neural network(LSTM)to further improve the identification performance of oxygen depletion status(ODS)in Chronic Obstructive Pulmonary Disease(COPD).Firstly,four effective discriminative features,including Saturation of Pulse Oxygen staturation(SPO2)value,pulse rate(PR)value,window feature of SPO2,and gradient feature of SPO2,are extracted from each point to be identified.Secondly,on the basis of this feature representation,the training set is used to train an AttLSTM model for ODS identification.Finally,the testing set is used to evaluate the identification performance of the trained AttLSTM.The experimental results show that the global performance index,i.e.,area under curve(AUC),of the proposed AttLSTM model reaches 0.8531,indicating that the proposed method can be effectively used for identifying ODS and has important reference value for the diagnosis of COPD.

关 键 词：特征表示 注意力机制 长短期记忆神经网络 慢性阻塞性肺疾病 氧减状态辨识 

分 类 号：R319[医药卫生—基础医学]