On the Relative Contribution of Deep Convolutional Neural Networks for SSVEP-Based Bio-Signal Decoding in BCI Speller...

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCI) use steady state visual evoked potentials (SSVEPs) for neural activity modulation.
• BCI spellers match neural oscillations to characters for selection.
• Subject-specific optimization is key to BCI speller advancement.

Purpose of the Study:

• To apply deep convolutional neural networks (DCNNs) for cross-subject classification of SSVEPs.
• To evaluate the performance of a DCNN (PodNet) for BCI spellers.
• To demonstrate calibrationless and adaptable SSVEP classification.

Main Methods:

• Utilized an open-source SSVEP dataset with electroencephalogram (EEG) data.
• Applied a DCNN (PodNet) to classify frequency and phase-encoded SSVEPs across subjects.
• Compared PodNet performance against filter-bank canonical correlation analysis.

Main Results:

• PodNet achieved 86% and 77% cross-subject classification accuracy for 6-second and 2-second data capture periods, respectively.
• Performance improved for subjects with initially sub-optimal (<70%) results after training.
• PodNet outperformed filter-bank canonical correlation analysis with a low-volume electrode setup.

Conclusions:

• PodNet demonstrates effective cross-subject, calibrationless classification of SSVEPs.
• The DCNN approach shows adaptability to sub-optimal subject data and low-volume electrode configurations.
• This study presents functional performance for the largest number of SSVEP classes decoded via DCNN to date.