邏輯式迴歸於MMN資料上的應用

Abstract

隨著認知神經科學領域領域越來越興盛，腦波的觀測以及分析技術也顯得越來越重要。此篇論文題目與資料來自於中央研究院的大腦與語言實驗室，目的是以成人區辨國語聲調的資料，找出能夠有效反應語音知覺的腦波特徵。我們探討語音知覺所要用到的事件相關電位(ERP)，為不匹配負向波(MMN)，透過總體經驗模態分解法(EEMD)得到能夠反應MMN的電位，稱為ERM，再將ERMs畫分布圖，找出能夠反應不同國語聲調所生成MMN的統計量，針對這些統計量建立data frame，此data frame以不同國語聲調所生成的MMN為分類目標，透過邏輯式迴歸，找出與類別關聯的統計量，這些統計量便能在反應語音知覺上提供幫助。由研究結果發現，100ms到200ms與200ms到300ms此兩個時間段對ERMs取平均值，是最能分辨不同MMN的；而不同統計量之間的交互作用，也有觀察到很顯著的＂跨Channel、跨時段＂的交互項，雖然重複取樣下表現並不那麼穩定，但仍值得深究。

With the field of cognitive neuroscience more and more popular, brainwave observation and analysis is also increasingly important. The data of this paper is from Brain and Language Laboratory, Academia Sinica. It’s adult EEG data, trying to find the significant features which can reflect the change of the brainwave under different lexical tones. The component of event-related potential (ERP) we use is mismatch negativity (MMN). MMN-related activity can be extracted by the method of ensemble empirical mode decomposition (EEMD), which will decompose time series data into intrinsic mode functions (IMFs). Averaging selected IMFs across trials allows measurement of event-related modes (ERMs). We plot the ERMs distribution plot to find the statistics which can distinguish different MMN, and use them to create a data frame. Finally, put the data frame into logistic regression to find significant statistics. The result is that mean of ERMs between 100ms to 200ms and 200ms to 300ms are most significant, and some interactions between statistics are also significant. Although the performance of interaction under repeated sampling is not so stable, it’s still worthy of study.