Universal neurophysiological interpretation of EEG brain-computer interfaces

Abstract

The identification of brain activity by a brain computer interface (BCI) from electroencephalography (EEG) recordings can be achieved with a wide variety of signal processing and machine learning algorithms. Currently, there is no universal method for explaining the predictions of a BCI decoder in terms of neurophysiologically relevant factors. In this paper, we propose an extension of a prominent, perturbation-based machine learning interpretability technique for EEG recordings. We validate this method with simulated EEG data and show that it can appropriately distinguish contributing electrodes and cortical sources across a range of BCI decoders. We also investigate the convergence propert..