運用胸腔和腹腔位移訊號的睡眠呼吸中止症偵測演算法

A Sleep Apnea Detection Algorithm Using Thoracic and Abdominal Movement Signals

林音延(Yin-Yan Lin)

指導教授 ： 黃元豪

國立清華大學/電機資訊學院/電機工程學系/碩士(2015年)

https://doi.org/10.6843/NTHU.2015.00641

摘要

睡眠呼吸中止症是一種高盛行率的睡眠障礙，患者在睡眠當中會反覆地停止呼吸，而一段呼吸停止的片段被稱為一個睡眠呼吸中止事件。睡眠呼吸中止事件可分為兩種類型：阻塞型睡眠呼吸中止症(OSA)事件、及中樞型睡眠呼吸中止(CSA)事件。阻塞型呼吸中止事件起因於患者在睡眠當中上呼吸道逐漸塌陷，導致上呼吸道阻塞使病患無法呼吸。中樞型呼吸中止症則是源於患者腦部在睡眠時對於血液中的高碳酸(CO2)的反應性下降，大腦認定血氧濃度足夠而發送減少呼吸的訊號至呼吸系統，使患者逐漸暫停呼吸。臨床醫師在判斷OSA及CSA事件時除了參考鼻腔氣流判斷事件發生外，也利用胸腔和腹腔的移動訊號判定事件乃OSA或CSA事件。睡眠呼吸中止症常導致患者因頻繁地的夜間覺醒、日間頭痛、過度疲倦嗜睡、及注意力不集中等而生活品質下降，但令人擔憂的是許多患者並未察覺自身患有睡眠呼吸中止症、且目前診斷睡眠呼吸中止症主要依據昂貴且耗費人力的睡眠多項生理檢查(Polysomnography, PSG)。因此，在此研究中提出了能夠只利用胸腔(THO)和腹腔(ABD)位移訊號偵測睡眠呼吸中止症的演算法希望能簡化睡眠呼吸中止症的檢查。演算法開發分為兩個階段。
第一階段中，先使用長度為2秒即5秒的長度分割THO和ABD訊號，接著對每個分割出的2秒和5秒的訊號片段中運用了互相關(cross-correlation)、Chirp-Z轉換(chirp z-transform)、及最大振幅(maximum amplitude)的訊號特徵萃取結合狀態機(state machine)判斷睡眠呼吸狀態建立睡眠呼吸中止症偵測演算法，對於嚴重呼吸中止症患者有100%的偵測靈敏度 (sensitivity)、對於輕微患者則有73.7%的辨識度 (specificity)，整體而演準確度(accuracy)為86.1%。此階段亦建立了對應演算法的硬體架構，提供此演算法未來整合入居家照護系統的可能性。
第二階段的演算法開發希望達到更佳的OSA和CSA事件偵測的準確度和分辨度。首先使用自適應非諧波模型(adaptive non-harmonic model)量化PSG檢查量測所得的THO和ABD訊號，並使用synchrosqueezing的方式去除THO和ABD訊號中的雜訊。接著從THO和ABD訊號中萃取出振幅比例、頻率比例、及共變異數(covariance)訊號特徵，接著利用支持向量機(support vector machine)找出最佳分類正常與睡眠中止(OSA和CSA)及區別OSA和CSA的兩個分類器(classifier)，並將這兩個分類器用在狀態機(state machine)的設計中用以判斷OSA和CSA事件的發生與否。此外提出了兩種評估事件偵測準確度的指標：指標S(index S)和指標I(index I)，並驗證所提出的演算法能夠達到平均80%以上的準確度。

關鍵字

睡眠呼吸中止症 ； 胸腔移動訊號 ； 腹腔移動訊號 ； 自適應非諧波模型 ； 呼吸特徵變化

並列摘要

Sleep apnea syndrome (SAS) is a prevalent sleep disorder well-known nowadays. People with SAS cease breathing intermittently in sleeping and an episode without breathing is called a sleep apnea event. SAS often deteriorates life quality by frequent nocturnal awakenings, morning headache, excessive daytime sleepiness, and attention deficiency. Unfortunately, the suffering people are usually unconscious of it and diagnosis nowadays relies on expensive and labor-intensive Polysomnography (PSG).
This thesis proposed a sleep apnea event detection algorithm which detects sleep apnea events based on solely thoracic (THO) and abdominal (ABD) movement signals during sleep. The algorithm is developed in two stages. In the first stage, the THO and ABD movement signals are segmented by 2-second and 5-second time window and then four features, cross-correlation, fundamental frequency, and maximum THO and ABD amplitude values, are extracted from the THO and ABD segments. Then, a state machine is designed to use the calibration signal information and the features to count the number of sleep apnea events in overnight THO and ABD movement signals. Hardware architecture is also built up in this stage. Based on the structure built in stage one, in stage two adaptive non-harmonic model is used to quantify the THO and ABD signals, more features are included, and support vector machine is introduced to construct classifiers for the state machine. To assess the event-by-event detection accuracy, two indexes, I and S, are proposed to evaluates the performance of the proposed sleep apnea event detection algorithm.

並列關鍵字

Sleep Apnea ； thoracic movement signal ； abdominal movement signal ； adaptive non-harmonic model ； Synchrosqueezing transform ； breathing pattern variability

參考文獻

[1] N. S. F. of USA. How much sleep do we really need? [Online]. Available:

[2] A. A. of Sleep Medicine (AASM) Task Force, “Sleep related breathing disorders in

[3] N. Punjabi, “The epidemiology of adult obstructive sleep apnea,” Proc Am Thorac

Williams & Wilkins, 2012.

portion of sleep apnea syndrome in middle-aged men and women,” Sleep, vol. 20,