On the comparison to EEG Norms : a new method and a simulation study

Quantitative Electroencephalography (qEEG) as a tool for the diagnosis of neurological and psychiatric disorders is receiving an increased interest. While qEEG analysis is restricted to the scalp, the recent development of electromagnetic tomographies (ET) allows the study of the electrical activity of cortical structures. Electrical measures of a patient can be compared to a normative database derived on a large sample of healthy individuals. The deviance from the database's norms provides a probabilistic measure of the likelihood that the patient's electrical activity reflects normal brain functioning. The focus of this thesis is the method for estimating such deviance. The method currently employed estimates the mean and the standard deviation of the normative sample. The deviance is then expressed in terms of z-scores. This method is referred to as the parametric method. The accuracy of the parametric method relies on the assumption that the distribution of the normative sample is gaussian, but this assumption is not always fulfilled in real qEEG and especially ET data. A new method based on percentiles ("nonparametric") is proposed. The parametric and the non-parametric methods are compared using simulated data. The accuracy of both methods is assessed as a function of normative sample size and gaussianity for three different alpha levels. Results suggest that the performance of the parametric method is unaffected by sample size (bigger than 100), but that non-gaussianity jeopardizes accuracy even if the normative distribution is close to gaussianity. On the contrary the performance of the non-parametric method is unaffected by non-gaussianity, but is a function of sample size only. It is shown that, with n>160, the non-parametric method can be considered always preferable. Results are discussed taking into consideration technical issues related to the peculiar nature of qEEG and ET data. It is suggested that the sample size is the only constant across EEG frequency bands, measurement locations, and kind of quantitative measures. As a consequence, for a given database, the error rate of the non-parametric database is homogeneous, however the same is not true for the parametric method.