Epilepsy patients who do not respond to pharmacological treatments currently have only brain surgery as a major alternative therapy. Identifying which brain areas to remove is thus of critical importance for physicians and the patient. Currently, this process is almost entirely manual, can vary greatly between clinical experts and centers, and depends only on qualitative EEG features, all of which may help explain the only modest success of extratemperal lobe epilepsy surgery. In this study, we explore an unsupervised, quantitative method for identifying seizure onset regions. A Gaussian mixture model (GMM) was used to cluster 500 ms epochs of intracranial electroencephalogram (EEG) prior to (preictal) and during (ictal) seizures in week-long continuous recordings from three patients during evalulation for epilepsy surgery. The GMM learning paradigm determines the optimal number of clusters for each patient. For the two patients whose epochs sorted into two clusters, we found that one cluster was predominantly composed of seizure epochs, and a subset of the channels made brief "forays" into that cluser in the time leading up to seizure onset. This observation is in keeping with the clinical hypothesis that certain brain areas may be the initiators of seizure activity, and we find that the channels independently labeled by physicians as seizure onset zones (SOZs) are statistically overrepesented in the seizure-defined cluster. Nevertheless, we also find that a subset of channels not labeled as SOZs has similar properties as those labeled SOZs. In this study we have tried to avoid many of the assumptions commonly made about what features and events are indicative of epileptogenic activity and believe that such analysis can help avoid many of the pitfalls of manual, non-objective human SOZ marking.