Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
Enhancing the Performance of Brain-Computer Interfaces via a Spatially-Aware Framework
Electroencephalography (EEG)-based brain–computer interfaces (BCIs) provide a non-invasive pathway for communication, with the P300 speller being one of the most established paradigms. Despite its practicality, EEG signals present inherent challenges, including low signal-to-noise ratio, high dimens...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Published: |
AUC Knowledge Fountain
2026
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| Summary: | Electroencephalography (EEG)-based brain–computer interfaces (BCIs) provide a non-invasive pathway for communication, with the P300 speller being one of the most established paradigms. Despite its practicality, EEG signals present inherent challenges, including low signal-to-noise ratio, high dimensionality, and limited spatial resolution. These characteristics complicate the extraction of discriminative spatial patterns, which are critical for accurate P300 detection and character-level decoding.
This thesis addresses these challenges through the development of AttenPCA-Net, a deep learning framework that integrates spatially-aware processing with dimensionality reduction. Principal Component Analysis (PCA) is employed to reduce redundancy and noise while preserving the most informative components of the EEG signal. To further enhance spatial representation, channel-wise and spatial attention mechanisms, implemented via Squeeze-and- Excitation (SE) and spatial attention modules, are incorporated to adaptively emphasize relevant electrodes and regions, compensating for the limited spatial resolution of EEG. The architecture is extended with a temporal branch based on bidirectional gated recurrent units (GRUs), forming AttenPCA-GRU, to capture sequential dependencies across EEG samples.
The proposed models were evaluated on three benchmark P300 speller datasets, including BCI Competition III Dataset II and two additional publicly available datasets. Performance was assessed using character recognition accuracy across varying trial repetitions and compared against baseline and state-of-the-art methods for each dataset. AttenPCA-Net achieved an enhancement of approximately 2% character recognition accuracy compared to state-of-the-art approaches. AttenPCA-GRU demonstrated more consistent performance across all datasets, yielding overall average improvements, most notably on the third dataset, where enhancements reached approximately 4% compared to AttenPCA-Net and 6% compared to baseline methods.
Beyond algorithmic contributions, this work also addresses acquisition-related biases by adapting a modified version of the Sevo electrode; an electrode that was developed for improved performance in coarse hair conditions, using the g.tec Unicorn Hybrid Black headset. These developments were integrated into AfroBCI, a custom software platform that we designed to support real-time EEG acquisition and experimentation across multiple sites in Africa.
Overall, this work demonstrates that improving BCI performance requires a holistic approach that jointly addresses signal representation, model design, and data acquisition constraints, contributing toward more robust and inclusive EEG-based BCI systems. |
|---|