Nearest Neighbor Classification of Stationary Time Series: An Application to Anesthesia Level Classification by EEG Analysis

摘要

An exploratory time series data-population screening problem is considered. A moderate, but not a large, number of categorical or labeled time series are obtained from different individuals (or objects). There is broad intersubject time series variability in each category. The objectives are to obtain a statistically reliable estimate of the minimum achievable probability of misclassification of new time series and to implement a time series classification rule that can achieve that statistical performance. A nearest neighbor classification rule achieves those objectives. With that rule of dissimilarity is computed between a new to-be-classified time series and each of a set of categorically labeled time series. The new time series is classified with the label of its least dissimilar neighbor. In our approach the dissimilarity measure between the time series is an estimate of the Kullback Leibler number between the time series computed as if the time series were normally distributed. This dissimilarity measure is shown to have sufficient metric properties for the formal Cover and Hart asymptotic nearest neighbor and Rogers finite sample nearest neighbor classification rule properties to hold.