為了分析一顆神經細胞的行為，我們需要了解其發生動作電位的頻率變化。通常一根微電極會接收到多個神經細胞所發出的動作電位以及一些環境雜訊，因此對於這些混雜的神經訊號，我們處理的步驟是先過濾雜訊，之後再將不同神經細胞所發出的動作電位分類。 主成分分析法最常被用於處理動作電位的分類問題上，但分類結果需由使用者觀看資料分布圖來決定，所以客觀性不高。Autoclass是目前網路上公開的一種分類軟體，其分類結果會顯示每筆資料屬於每一群的機率值。以機率值大小將資料分群的結果比用眼睛判斷的分類結果還具客觀性。本論文的目的是使Autoclass能有效應用在處理神經細胞動作電位的工作上，同時提出一個先驗標準來判斷一組訊號是否該以Autoclass分類，以及一個後驗方法來評估分類的結果是否可以信賴。 In order to analyze the actions of a neuron, we need to understand the variations of the frequency of its spikes. An electrode usually receives spikes generated by several neurons as well as some background noises, which become interference in our analysis. Therefore, the task is to first filter out noises from these electrical signals and then to classify the spikes. In spike sorting, Principal Component Analysis (PCA) is the method scholars tend to adopt most often. However, the classification produced by PCA is not objective for it is based on operators’ observation of the graph. An alternative choice, Autoclass, is a popular method for clustering and its classification shows the probability of each data which belongs to each cluster. We can get a more objective classification according to probability than by visual estimation. Hence, a classification produced by Autoclass is more objective than one produced by PCA. The goal of this article is to show how Autoclass makes spike sorting more effective, meanwhile, to provide a prior knowledge to determine whether a data is suitable to be clustered by Autoclass, and to propose a posterior method to confirm the validity of its classification.
