Long term EEG examinations, for example during epilepsy diagnosis, can be performed more efficiently with support of automated abnormality detection. Currently, these methods are usually developed based on one specific database, which limits the possibilities of generalizations. Here, we present a machine learning solution for detection of interictal abnormal EEG segments optimized on the publically available TUH Abnormal EEG Corpus. The classifier is further re-trained and tested on several combinations of publicly available data sets. The results achieved internally on the datasets are comparable to the known state of the art, while training and testing on different sources produced accuracy in the range of 67.51% to 99.50%. Lower accuracy is achieved when the training data set is highly preprocessed and relatively small.