Autoencoders for learning template spectrograms in electrocorticographic signals.

Objective: Electrocorticography (ECoG) based studies generally analyze features from specific frequency bands selected by manual evaluation of spectral power. However, the definition of these features can vary across subjects, cortical areas, tasks and across time for a given subject. We propose an autoencoder based approach for summarizing ECoG data with 'template spectrograms', i.e. informative time-frequency (t-f) patterns, and demonstrate their efficacy in two contexts: brain-computer interfaces (BCIs) and functional brain mapping.

Approach: We use a publicly available dataset wherein subjects perform a finger flexion task in response to a visual cue. We train autoencoders to learn t-f patterns and use them in a deep neural network to decode finger flexions. Additionally, we propose and evaluate an unsupervised method for clustering electrode channels based on their aggregated activity.

Main Results: We show that the learnt t-f patterns can be used to classify individual finger movements with consisentently higher accuracy than with traditional spectral features. Furthermore, electrodes within automatically generated clusters tend to demonstrate functionally similar activity.

Significance: With increasing interest in and active development towards higher spatial resolution ECoG, along with the availability of large scale datasets from epilepsy monitoring units, there is an opportunity to develop automated and scalable unsupervised methods to learn effective summaries of spatial, temporal and frequency patterns in these data. The proposed methods reduce the effort required by neural engineers to develop effective features for BCI decoders. The clustering approach has applications in functional mapping studies for identifying brain regions associated with behavioral changes.