應用Wigner-Ville分布法於自動辨識睡眠紡錘波之研究

Abstract

現代人快速的生活步調使得擁有睡眠障礙且因此接受睡眠檢查的患者與日俱增。目前這些睡眠資料仍是透過睡眠技師以人工的方式進行階段判讀，過程相當耗時且不同技師所判讀的結果也不盡相同。因此發展自動判讀睡眠階段專家系統成為多年來重要的研究主題。而睡眠紡錘波為判讀睡眠第二期的主要特徵之一，因此自動辨識紡錘波之技術即有其必要性。本研究之目的在建立紡錘波之自動辨識演算法，進而達成判讀睡眠第二期的目標。 根據美國睡眠醫學協會(AASM)所制定的睡眠規則，睡眠紡錘波的定義是頻率介於11-16Hz(常出現於12-14Hz)，且持續出現超過0.5秒以上的波形。由於紡錘波同時具有時間域與頻率域之特徵，本研究利用WVD分布法對睡眠腦波訊號進行時頻分析。先將每頁訊號切割為150個0.2秒長的片段，以WVD時頻圖計算各片段之紡錘波頻帶與非紡錘波頻帶能量，當正規化紡錘波頻帶能量超過閥值，則視為紡錘波候選片段，再經由相鄰波形連接步驟之平滑化，並以高低頻能量比例作篩選，即得自動辨識波形。 本研究使用7筆來自康寧醫院的健康成年人睡眠資料並將其任意分成兩組，3筆用於訓練閥值、另外4筆測試辨識效能。辨識結果指出，本演算法於辨識個別紡錘波，有71.13%之敏感度與79.45%之精確度。另外在辨識含紡錘波之睡眠第二期頁數方面，以整晚訊號做為分析對象可得到整體準確率達89.44%，主要誤判頁數出現於清醒期和第三期。若排除清醒期與第三期訊號，則整體準確率可達90.03%。這顯示本演算法具有準確分類含紡錘波之睡眠第二期頁數的能力，這將極有助於專家系統之睡眠第二期判讀。

Rapid pace of life caused more people having sleep disorders, and more patients were willing to do sleep examination. Currently, these data were scored by artificial, and the procedure of scoring wasted time and can hardly get the similar reports from different experts. Therefore developing automatic identification system of sleep stage scoring was a popular issue these years. However, sleep spindle is one of main feature of NREM-2 stage (N2) scoring, so it is necessary to establish an automatic sleep-spindle detection algorithm, and is also the main purpose of this research. According to sleep stage scoring rules published by American Academy of Sleep Medicine (AASM), the definition of spindles is the wave which frequency range is between 11-16Hz (commonly 12-14Hz), and last more than 0.5 second. This study uses Wigner-Ville Distribution (WVD) to execute time-frequency analysis of EEG signal, because spindles have both time domain and frequency domain characteristics. In algorithm programming, first separating an epoch signal into 150 0.2-second segments, then using WVD time-frequency spectrogram to compute spindle and non-spindle band energy of each segment. If normalized spindle band energy of segments were over threshold, then these segments would be considered as candidate segments. At last, through smoothing and connecting nearby segments, and screening with high-low frequency energy ratio, then we can get automatic detection patterns. In this study, 7 PSG data of healthy adult from Kang-ning general hospital are analyzed and separate into two groups arbitrarily, three for training thresholds and four for testing algorithm performance. The performance of detection of spindles is sensitivity of 71.13% and precision of 79.45%. For N2 epoch with spindles detection, accuracy reaches 89.44% when analyzing whole night signal, and 90.03% when analyzing signal without Wake and N3 epochs. These results show that proposed algorithm has ability to identify N2 epochs with spindles correctly, and are helpful to N2 stage scoring of expert system.