Back to the list of graduate studentsNeural Network Based Detection of EEG Abnormalities
by
Leon L. Lipoth
M. Eng., 1991
Abstract
this thesis examines a neural network approach to detecting abnormalities in the Electroencephalogram (EEG). The approach is novel in its application of neural network processing to the qualified evaluation of EEGs. A probabilistic Neural Network (PNN) architecture is emphasized. the advantages of the PNN include its quick training and easy modification of classification decision boundaries.
The PNN classifier is developed as described in Section 4.3. In addition, the tools and procedures of Sections 3.3 and 3.4 for the EEG data processing are developed. An improvement to the spectral averaging process is made by introducing multiple, overlapped windows to the sampled EEG data. This technique allows for a tradeoff between spectral variance and frequency resolution.
In data processing, a normative reference set of EEGs is established by the examination of the time and frequency domain characteristics of the EEGs. The normative reference set consists of EEGs for which no abnormalities are detected. The remaining EEGs are considered abnormal. Using frequency domain feature vectors from six standard frequency bands ( d,q,a,b1,b2,b3) the PNN is trained on examples of both normal and abnormal EEGs. the most successful classification achieved is with the alpha band relative asymmetrical power failure vector. A 95% sensitivity yields a 40% specificity. These results are based on the 'leave-one-out' method of Section 5.1.2.
A need for the reformulation of the criteria used to build the normative reference set is defined. Tighter controls on the admission of an EEG to the normative reference set would result in better detection of EEG abnormalities. In addition, the need for more EEGs in the database, less allowable artifact in the EEG epochs, and more exploration of new features.
This thesis provides ideas for future research work with an emphasis on neural network applications in Section 5.5. The ideas include neural network based artifact recognition, feature definition, and Statistical Probability Mapping (SPM).