A new method of automatic EEG analysis (ASD or automatic non-stationarity detection) derived from a parametric EEG model (autoregressive filter model) and based on inverse filtering, has been applied to scalp and subdural EEGs of epileptic patients. By this method it was possible to detect the occurrence of transient non-stationarities such as paroxysmal patterns of activity (spikes, spike-and-waves) characteristic of inter-ictal EEGs. The ASD has been implemented in a general purpose digital computer. The method allows: (a) statistical evaluation of paroxysmal patterns; (b) multi-channel analysis, where the interrelationships of different derivations are quantified and displayed by means of a clinically useful spatial map. The application of the ASD method to scalp and subdural EEGs, simultaneously recorded, has revealed that a number of transient non-stationarities detected in the scalp by the program yet not by visual inspection, coincide with clear paroxysmal patterns in subdural derivations. In this way conventional definitions of characteristic paroxysmal patterns at the scalp are put in question. Identical conclusions regarding the localization of an epileptogenic zone in 4 patients were obtained either using the conventional methods of recording electro-clinical seizures in EEGs of long duration or applying the ASD method to short EEG epochs (100 sec long, at most).