Tree-based Classification Models for Somnolence Detection from EEG Spectra
Abstract
EEG spectra corresponding to the states of somnolence, wakefulness and mentation of 24 experimental subjects are analyzed by different tree-based methods. Classification forests obtained by the Random Forests (RF) method are clearly superior to the single trees grown by CART. Applying RF separately to the small data sets of individual subjects results in the "individual" models that outperform, in the mean, the "global" classifiers derived by RF from the more numerous but, at the same time, more heterogeneous data of all the subjects. The newly developed mixed models, combining information from both the individual and global models, prove slightly better than the individual models.