Statistical evaluation of features in classification problems with applications to detection of hypoglycemic conditions based on EEG data

Laura Friis Frølich

AbstractIndependent components (ICs) are patterns in EEG data that are assumed to correspond to neural generators of electricity in the brain. However, some ICs represent artifacts such as eye movements. By denoising EEG data through removal of automatically classified artifactual ICs, we investigated the effect of artifactual signals on detection of seizures due to to falling blood sugar levels in type I diabetics. The classification methods binary LDA, multiclass LDA, binary QDA, multiclass QDA, SVM, logistic regression, logistic regression with forward selection, L1-regularized logistic regression, multinomial regression, multinomial regression with forward selection, decision trees, ADJUST [41] and an algorithm proposed by the Berlin BCI group [56] were compared to find the best automatic classifier of ICs. Variance of feature estimates over cross-validation folds and effects of features on classification performance were investigated.
Seizure detection performance was decreased when the model was trained on data without artifacts. This decrease in performance may either be because (1) the detection model relies on artifacts such as muscle twitches and eye movements during seizures, or (2) because neural activity was wrongly removed in the cleaning process. In the first case, seizure detection models that do not rely on artifacts must be found. In the second case, models trained on data cleaned by better noise removal methods will most likely increase seizure detection performance based on EEG recorded by a device implanted in the brain, since such a device cannot detect artifacts.
L1-regularized logistic regression and logistic regression with forward selection turned out to be the best methods. Almost all features were chosen in L1- regularized logistic regression as well as by a criterion based on mutual information between features and class assignments. This indicates that all features represent class relevant information. Small variances of feature coefficient estimates were seen, implying that the estimated models represent structures in data, and not chance relations.
TypeMaster's thesis [Academic thesis]
Year2011
PublisherTechnical University of Denmark, DTU Informatics, E-mail: reception@imm.dtu.dk
AddressAsmussens Alle, Building 305, DK-2800 Kgs. Lyngby, Denmark
SeriesIMM-M.Sc.-2011-60
NoteSupervised by Tobias Andersen, ta@imm.dtu.dk, and Morten Mørup, mm@imm.dtu.dk, DTU Informatics
Electronic version(s)[pdf]
Publication linkhttp://www.imm.dtu.dk/English.aspx
BibTeX data [bibtex]
IMM Group(s)Intelligent Signal Processing


Back  ::  IMM Publications