Paper: | RS-1B.4 |
Session: | Late Breaking Research 1B |
Location: | Late-Breaking Research Area |
Session Time: | Thursday, September 6, 18:45 - 20:45 |
Presentation Time: | Thursday, September 6, 18:45 - 20:45 |
Presentation: |
Poster
|
Publication: |
2018 Conference on Cognitive Computational Neuroscience, 5-8 September 2018, Philadelphia, Pennsylvania |
Paper Title: |
Network metrics as classification features in brain-computer interface |
Authors: |
Thomas Arnold, University of Pennsylvania, United States; Marie-Constance Corsi, Sorbonne Universités, France; Jeni Stiso, University of Pennsylvania, United States; Fabrizio De Vico Fallani, Sorbonne Universités, France; Danielle Bassett, University of Pennsylvania, United States |
Abstract: |
Many noninvasive brain-computer interfaces utilize band power to classify user intent in motor imagery paradigms. The features used for classification often contain redundant information regarding mu de/synchronization of motor regions. While these classification schemes have shown some success, they are not reliable enough to be used for many intended purposes. One possible reason for their limited utility is the fact that information from only a small subset of sensors is incorporated, despite the fact that whole-brain data is typically collected. Given the distributed processes necessary to carry out planned motor movements, it stands to reason that relationships between cortical regions could provide novel discriminatory information to classifiers. Recent advances in making single-trial connectivity estimates bolster the feasibility of using such distributed information. Here we propose to expand the feature space of motor imagery BCI classification by incorporating network-based metrics. We estimate connectivity via covariance-based Granger Causality. A single network is generated for each frequency band and network metrics are extracted as features. We then compare the accuracy of classifiers that have access to only band power features, only network features, or both feature sets. Our results have important implications for adaptive network-based methods for motor imagery BCI. |