EEG gamma frequency and sleep–wake scoring in mice: Comparing two types of supervised classifiers

On the identification of sleep stages in mouse electroencephalography time-series

“…Little study into the optimal epoch length,N in Eq. (1), has been conducted (Yan et al, 2011;Brankačk et al, 2010). Its choice should be influenced by the required output of the system and recently published research (Yan et al, 2011) recommends an epoch duration of 30 s when extracting stage lengths and power densities (in order to save time) and a 4 s epoch duration when analysing stage transitions and number of episodes.…”

“…without manual correction. Furthermore, all of the solutions applied to this (or similar) problem(s) base their decision upon the output of standard classifiers such as a neural network (Oropesa et al, 1999;Ebrahimi et al, 2008), decision tree (Brankačk et al, 2010;Kohtoh et al, 2008), k-nearest neighbours (Yu et al, 2009), or linear discriminant analysis (Brankačk et al, 2010), to name but a few. An approach that is typically taken in the classification literature to improve classification rates (and has not been attempted with reference to this application) when it has been shown that standard classification algorithms do not meet the required performance is to form a multiclassifier system (MCS) (Kittler et al, 1998).…”

“…A commonly used method for automatic sleep staging in clinical and commercial situations is threshold formula (or decision tree analysis) (Brankačk et al, 2010). In this study we defined a decision tree within SleepSign for Animals 3.0 (Gallegos et al, 2002) software package (furthermore referred to as SleepSign) and this was used as the comparison method in this investigation.…”

“…The literature concerned with this application identifies neural networks (Oropesa et al, 1999;Ebrahimi et al, 2008), decision tree (Brankačk et al, 2010;Kohtoh et al, 2008), k-nearest neighbours (Yu et al, 2009), and linear discriminant analysis (Brankačk et al, 2010), to name but a few. Neural networks have proven to be successful, however, the backpropagation training method is based upon stochastic gradient descent and therefore may not find the globally optimal decision boundary.…”

“…A major drawback, however, is the time-consuming task of hand labelling the acquired data. It is estimated that analysis time almost equals recording time (Brankačk et al, 2010), furthermore, independent analysts are likely to score EEG signals differently (Robert et al, 1999). Therefore it is critical to develop automatic or semi-automatic methods to accomplish this task.…”