Abstract
The amount of freely available human phenotypic data is increasing daily, and yet little is known about the types of inferences or identifying characteristics that could reasonably be drawn from that data using new statistical methods. One data type of particular interest is electroencephalographical (EEG) data, collected noninvasively from humans in various behavioral contexts. The Temple University EEG corpus associates thousands of hours of de-identified EEG records with contemporaneous physician reports that include metadata that might be expected to show a measurable correlation with characteristics of the recorded signal. Given that machine learning methods applied to neurological signals are being used in emerging diagnostic applications, we leveraged this data source to test the confidence with which algorithms could predict, using a patient's EEG record(s) as input, which medications were noted on the matching physician report. We comparatively assessed deep learning and feature-based approaches on their ability to distinguish between the assumed presence of Dilantin (phenytoin), Keppra (levetiracetam), or neither. Our methods could successfully distinguish between patients taking either anticonvulsant and those taking no medications; as well as between the two anticonvulsants. Further, we found different approaches to be most effective for different groups of classifications.