使用深度神經網路之阻塞型睡眠呼吸中止症病人之自動化睡眠分期判讀

Automated sleep stage classification in obstructive sleep apnea patients using deep neural networks

侯人文(Jen-Wem Hou)

指導教授 ： 闕志達

國立臺灣大學/電機資訊學院/電子工程學研究所/碩士(2019年)

https://doi.org/10.6342/NTU201903265

摘要

對睡眠障礙者而言，可以透過全通道多導睡眠圖（PSG）來做診斷，PSG利用多種方式測量生理訊號，包括腦電圖（EEG），眼電圖（EOG），心電圖（ECG），肌電圖（EMG）和呼吸頻率，透過整夜的檢查，可以診斷各式睡眠疾病，例如阻塞型睡眠呼吸中止症(Obstructive sleep apnea)，然而PSG檢查有許多問題，包括高昂的成本、評估的可變性、手動判讀相當耗時等缺點，為了讓評估標準一致，技術人員通常會遵守美國睡眠醫學會(AASM)以及Rechtschaffen和Kales（R＆K）制定的判讀標準，通過這些標準對五個睡眠階段（Wake，N1，N2，N3，REM）進行分類。由於以上這些原因，準確測量睡眠階段的系統對於了解臨床研究中的睡眠是非常有價值的。
近幾年，深度神經網路(Deep Neural Network)與人工智慧研究因進步的電腦科技而再度被廣泛研究。 神經網路有數種類型，包括: 多層感知器(MLP)，卷積神經網路(CNN)，遞歸神經網路(RNN)等，其中卷積神經網路又被廣泛地應用在影像處理上，諸如影像辨識，物件偵測，甚至自然語言處理；到了最近，卷積神經網路深度已經可含有百層以上，能解決困難的任務，但是同時，計算上複雜度與傳統多層感知器相比也提高許多。
本論文提出一種方式，使用單通道前額的腦電訊號(F4-M1)對睡眠階段的進行自動判讀，其中還包括了相當高比例睡眠呼吸中止症的患者。透過深度神經網路對腦電訊號的特徵分成五種睡眠階段，並利用決策樹，搭配馬可夫模型的概念，即可將睡眠正確分期。

關鍵字

睡眠階段分期 ； 決策樹 ； 類神經網路 ； 馬可夫模型

並列摘要

Sleep disorders can be diagnosed by full-channel polysomnography (PSG), which utilizes multiple sensing modalities to measure biophysiological signals, including electroencephalogram (EEG), electrooculography (EOG), electrocardiography (ECG), electromyography (EMG) and respiratory rate. By taking a whole-night examination, the sleep laboratory can diagnose several kinds of sleep diseases such as obstructive sleep apnea (OSA). However, PSG has many problems, including high cost, considerable inter-rater variability and time-consuming in manual scoring. In order to maintain the evaluation criteria consistent, technicians adhere to rules delineated by American Academy of Sleep Medicine (AASM) and the Rechtschaffen and Kales (R&K). Five sleep stages (Wake, N1, N2, N3, REM) are visually categorize by these rules. For all these reasons, a robust system that accurately measure sleep stages is valuable for sleep in clinical studies.
In recent years, deep neural networks and AI research have attracted much attention. There are several types of neural networks, including multilayer perceptrons (MLPs), convolution neural networks (CNNs), recurrent neural networks (RNNs). Among these architectures, convolution neural networks have been widely used in image processing tasks, including image classification, object detection, even natural language processing. Recently, it has been showed that CNNs can be built with hundreds of layers in order to solve tough tasks, at the same time require much more computing effort compare to traditional MLP.
We present an algorithm to scoring sleep using a single-channel frontal EEG (F4-M1) comprising a high percentage of participants with OSA. This work adopted multiple features of EEG generated via the deep neural networks to classify five sleep stages. By using a decision tree and the Markov Model technique, our system can recognize sleep stage effectively.

並列關鍵字

sleep stage recognition ； decision tree ； Neural Network ； Markov Model

參考文獻

[1] A. Rechtschaffen, and A Kales, “A manual of standardized terminology, techniques and scoring system for sleep stages of human subject.” Washington DC: US Government Printing Office, National Institute of Health Publication, 1968.

[2] Oral Surgery, Sleep Disorder & Implant studio. Available: https://chicagosleepapneasnoring.com /polysomnography/

[3] WIKIPEDIA Sleep apnea. Available: http://en.wikipedia.org/wiki/Sleep_apnea

[4] R. Ruehland, D. Rochford, and J. O'Donoghue, “The new AASM criteria for scoring hypopneas: impact on the apnea hypopnea index,” Sleep, Vol.32(2), pp.150–157,Feb 2009.

[5] Surgical Sleep Solutions, “Obstructive Sleep Apnea,” [Online]. Available: http://surgicalsleepsolutions.com/obstructive-sleep-apnea/