A Tensor-based Machine Learning Approach to EEG Feature Detection: Examination of Working Memory Network Dysfunction in Schizophrenia

Human memory function can be assayed in real-time by electroencephalographic (EEG) recording; however, the clinical utility of this method is dependent on the reliable determination of functionally and diagnostically relevant features. Data-driven machine learning approaches capable of modeling non-stationary signal have been explored as a way to synthesize large arrays of EEG data. Although standard machine learning approaches reduce the data to a 1D vector before classification, the EEG record could be more precisely characterized by a tensor (e.g., a 3D matrix) representing processing stages, spatial locations, and frequency bands as individual dimensions. We derive a novel tensor-based classification method and test it on EEG data collected during memory task performance in healthy normal and clinical (schizophrenia) samples.