應用匹配追蹤演算法於睡眠第二期K-複合波之自動偵測

Abstract

本研究之目的為發展睡眠第二期(N2)中特徵K複合波(K-Complex, KC)之自動辨識方法。本研究係依據最新AASM規則，以匹配追蹤演算法(Matching Pursuit, MP)及條件篩選的技術來辨識KC特徵。相較以往的研究使用繁複的數學運算來辨識KC，本研究所發展之方法係採用圖形判讀，因此更能模擬睡眠技師人眼判讀的方式。 匹配追蹤演算法係將訊號以基底展開的演算法，其特點是能自行定義基底，與訊號越相似（即越匹配）的基底，就越早被萃取出來。本研究根據匹配追蹤演算法基底匹配之步驟，採用與KC波形相似的基底，定出訊號波形近似KC的片段。再檢查近似KC片段的時間長度、峰對峰值大小、負向與正向的振幅比、兩連續KC之時間間隔，及與背景波振幅之比值，通過篩選條件者即判讀為KC。 本研究共有兩個資料來源，其一是使用Devuyst提供於網路上的DREAMS Databases，從中選取5位受試者的睡眠資料進行分析，並將自動辨識的KC與睡眠技師判讀結果做比較。若以資料中N2頁面的分析結果作比較，本自動辨識法對睡眠專家一所標記之KC的敏感度(Sensitivity, TPR)為65%，對專家二的TPR為74%，對專家一、二共同標記之KC 的TPR 為80%；對睡眠專家一所標記之KC頁面的TPR為77%，對專家二的TPR為92%，對專家一、二共同標記之KC頁面的TPR為95%。本自動辨識法的結果大部份都較DREAMS提供的自動判讀結果良好。 本研究所採用的第二個資料來源為本研究團隊於康寧醫院所收案的9筆睡眠資料。若以資料中N2頁面的分析結果作比較，本自動辨識法對睡眠專家所標記之KC的TPR分別為94%，對睡眠專家所標記之KC頁面的TPR 為96%。整體而言，本自動辨識法之表現尚稱理想。

The purpose of this research is to develop an identification method of the K-Complex wave (KC) in NREM-2 stages (N2) based on AASM rules. In order to mimic the procedure of the sleep stage scoring by sleep technicians, this study uses the matching pursuit algorithm (MP) to locate the KCs in a sleep signal. Similar to the Fourier expansion, the MP expands a signal by a set of bases. The expansion is composed of iterative matching pursuit processes. In each MP process, the correlations between the bases and the signal are computed and the basis that best matches the signal is extracted from the signal. Then the MP process is conducted repeatedly to find the signal segments that possess KC-like waveforms. That simulates the KC pattern recognition by sleep experts. After the KC candidates are obtained, several criteria are applied to screen the candidates, including the duration, peak-to-peak voltage, ratio of negative to positive amplitude, the interval between continuous KC, and the ratio of peak-to-peak voltage to the average of background voltage. Candidates that pass all criteria are considered KC in the signal. This study uses two different data sources. The first one is the DREAMS Databases provided by Devuyst on the website. The sleep signals for 5 subjected were taken from the database. The proposed method was applied to detect KC in these signals and the results were compared those by manual scoring. Among all N2 epochs, the sensitivity (TPR) of the proposed method in KC detection was 65%, 74%, and 80%, respectively, compared with the KC marked by expert 1, expert 2, and the intersection of experts 1 and 2. In the detection of epochs with KC, the sensitivity of the proposed method was 77%, 92%, and 95%, respectively, compared with results by expert 1,expert 2, and the intersection of experts 1 and 2. In general, the performance of the proposed method surpasses the DREAMS automatic detection algorithm. The second data set was collected from the Kang-Ning General Hospital, constituting of sleep signals of 9 healthy subjects. Among all N2 epochs, the TPR of the proposed method in the detection of KC and epochs with KC was 94% and 96%, respectively. Nevertheless, the overall performance of the proposed method is satisfactory, especially in the detection of epochs with KC.