利用每分鐘心電圖訊號特徵為支持向量機之睡眠呼吸中止症辨識方法

An Obstructive Sleep Apnea Recognition Algorithm Based on Support Vector Machine via Each Minute Single-Lead Electrocardiogram Features

譚匡哲(Kuang-Jer Tam)

指導教授 ： 馬席彬

國立清華大學/電機資訊學院/電機工程學系/碩士(2014年)

摘要

近年來，隨著睡眠時間以及睡眠品質逐漸受到重視，睡眠呼吸中止症逐漸成為了重要的議題，其後遺症會對人體造成精神上面的傷害與心血管方面的疾病，傳統診斷睡眠呼吸中止症是借助睡眠多項生理檢查紀錄患者整夜的生理訊號來評估病症的嚴重性，此方法雖然準確但是昂貴且耗時。因此，需要發展一項簡便且省時的技術來降低時間及經濟上的花費。

在此論文中，提出一個以支持向量機為基礎，藉由每分鐘心電圖訊號萃取出的特徵來辨識睡眠呼吸中止症之演算法，並將演算法的過程分成三個部分。首先，利用離散小波轉換來消除基線飄移雜訊及市電雜訊並且抓出心電圖中的峰值R，藉由找出的峰值R萃取出三種心電圖中的呼吸訊號。接著，利用此三種呼吸訊號做不同的時域與頻域分析去抓取126種具有辨識睡眠呼吸中止症能力的特徵，再將每一個不同的特徵逐一做縮放範圍之特徵正規化來降低特徵之間範圍變化的影響。最後，利用一種分類方法叫支持向量機來建立分類器模型，其支持向量機模型的參數則是利用一種交叉驗證方法叫10-fold cross-validation來選擇模型最佳參數。

利用Sleep-ECG資料庫中醫生所提供以分鐘為單位的答案與我們演算法中所產生的答案做效能評估，得到的準確率為88.29%，敏感性為92.9%，特異性為86.4%。除此之外，利用此資料庫中以受試者為單位來做效能評估，所得到的準確率為100%，敏感性為100%，特異性為100%，達到降低時間複雜度與金錢花費的效果。

關鍵字

睡眠呼吸中止 ； 支持向量機 ； 心電圖

並列摘要

Nowadays, sleep disorders become an important issue because it can adversely affect neurocognitive, cardiovascular, respiratory diseases, which subsequently induce the behavior
disorder, majority of these cases up to 85% of these cases are obstructive sleep apnea (OSA). Therefore, the study of how to diagnose, detect and treat OSA is becoming a critical
issue from both academical and medical perspective. Polysomnography (PSG) can monitor the OSA using relatively few invasive techniques. However, sleep studies are expensive and time-consuming because they require overnight evaluation at sleep laboratories with dedicated systems and attending personnel. To improve such inconveniences of exam and high cost, it is important to develop a simplified method to diagnose the OSA.

The motivation of this study is to develop an OSA detection algorithm, which uses only electrocardiogram (ECG) signal. The procedures of this algorithm include three parts. At
first, ECG signals are preprocessed by discrete wavelet transform (DWT) method in order to detect the R peck by removing the base line wander and power line noise. Based on the R peak position, the 126 features were generated from the ECG-derived respiration (EDR) signal using time domain and frequency analysis. The range scaling method can be applied afterward to normalize all features in order to minimize the effect of large range variation from the value of each feature. At last, a classification method called support vector machine (SVM) is used to classify if the subject shows OSA symptom or not in each minute. The 10-
fold cross-validation method is applied to select the best SVM parameter for classification. By combining all the classification results, one can determine if the subject is normal or apnea.

The accuracy of 88.29%, the sensitivity of 92.90% and the specificity value of 86.48%can be seen from the Apnea-ECG database using the minute by minute performance based algorithm. Beside, the accuracy of 100%, the sensitivity of 100% and the specificity value of 100% also can be seen from the database using the subject based performance. The goal of this study to reduce both the detection time and cost is accomplished.

參考文獻

[2] H. Attack. (2013, May) What is the electrocardiogram ECG? [Online]. Available: http:

[3] A. Khandoker, M. Palaniswami, and C. Karmakar, “Support vector machines for automated

recognition of obstructive sleep apnea syndrome from ECG recordings,” IEEE

Transactions on Information Technology in Biomedicine (T-ITB), vol. 13, no. 1, pp. 37–

processing of the single-lead electrocardiogram for the detection of obstructive