阻塞性睡眠呼吸暂停儿童睡眠结构研究及睡眠结构判读模型建立

The sleep structure of Children with obstructive sleep apnea and the development of a sleep structure interpretation model

目的 讨论不同程度阻塞性睡眠呼吸暂停(obstructive sleep apnea, OSA)儿童睡眠结构的改变,并基于深度学习方法构建睡眠结构自动判读模型,提高睡眠结构判读的精准度和效率。方法 回顾性分析146例学龄期打鼾儿童的多导睡眠监测(polysomnography, PSG)结果,比较不同严重程度OSA儿童呼吸事件、血氧饱和度和睡眠结构等参数的差异。基于深度学习方法,以脑电和眼电联合通道作为输入信号,由卷积神经网络(convolutional neural networks, CNN)、注意力机制和循环神经网络(recurrent neural network, RNN)组成混合神经网络,构建睡眠结构自动判读模型,并于公开数据集和私人数据集进行模型性能评估。结果 单纯打鼾组、轻度OSA组和中重度OSA组间快动眼(rapid eye movement, REM)期百分比和非快动眼睡眠(non-rapid eye movement, NREM)1期百分比比较差异未见统计学意义(P>0.05)。中重度OSA组N2期百分比明显高于单纯打鼾组、轻度OSA组(P=0.024),中重度OSA组N3期百分比明显低于单纯打鼾组、轻度OSA组(P<0.001),中重度OSA组觉醒指数(total arousal index, ARtotI)和呼吸事件相关觉醒指数(respiratory arousal index, RAI)明显高于单纯打鼾组、轻度OSA组(P<0.001),差异均具有统计学意义。本研究提出的睡眠结构自动判读模型,在公开数据集上总体准确率为79.90%,在N1期、N2期、N3期及R期的F1值分别为45.90%、85.40%、83.70%和79.90%,在私人数据集上总体准确率为71.30%,在N1期、N2期、N3期及R期的F1值分别为40.20%、81.60%、77.70%和67.40%。结论 OSA儿童随疾病进展,睡眠结构出现N2期延长和N3期缩短,本研究提出的基于深度学习的睡眠结构自动判读模型,具有良好的分类效果和稳定性。

Objective To explore the sleep structure in children with different degrees of obstructive sleep apnea(OSA)and construct an automatic sleep structure interpretation model using deep learning.Methods We retrospectively analyzed the polysomnography(PSG)results of 146 school-aged children,and compared the parameters,such as respiratory events,oxygen saturation,and sleep architecture in children with different severity of OSA.Using deep learning,a hybrid neural network composed of convolutional neural networks(CNN),attention mechanism,and recurrent neural network(RNN)was used to construct an automatic sleep structure interpretation model.A combination of electroencephalogram and electrooculogram were used as input signals.The model’s performance was evaluated using public and private databases.Results There was no significant difference in the percentage of rapid eye movement(REM)and the percentage of non-rapid eye movement(NREM)stage 1 among the primary snoring,mild OSA,and moderate to severe OSA groups(P>0.05).The percentage of the N2 stage was significantly higher in the moderate-severe OSA group than in the primary snoring and the mild OSA groups(P=0.024).The percentage of the N3 stage was significantly lower in the moderate-severe OSA group than in the primary snoring and the mild OSA groups(P<0.001).The total arousal index(ARtotI)and respiratory arousal index(RAI)in the moderate-severe OSA group were significantly lower than those in the primary snoring and the mild OSA groups(P<0.001).Using the public database,the automatic sleep structure interpretation model proposed in this study had an overall accuracy of 79.90%,and F1 values of 45.90%,85.40%,83.70%,and 79.90%in N1,N2,N3,and R stages,respectively.Using the private database,the overall accuracy was 71.30%,and the F1 values were 40.20%、81.60%、77.70%,and 67.40%in N1,N2,N3,and R stages,respectively.Conclusion The sleep structure of children with moderate-to-severe OSA showed a prolonged NREM2 and a shortened NREM3.Overall,the deep learning-based automatic sleep structure interpretation model demonstrated stability and a good classification effect.