Date of Publication

2024

Document Type

Master's Thesis

Degree Name

Bachelor of Science in Electronics Engineering (Honors) - Ladderized

Subject Categories

Electrical and Computer Engineering

College

Gokongwei College of Engineering

Department/Unit

Electronics And Communications Engg

Honor/Award

Summa Cum Laude

Outstanding Thesis Award

Thesis Advisor

Edwin Sybingco

Defense Panel Chair

Argel A. Bandala

Defense Panel Member

John Anthony C. Jose

Timothy Scott C. Chu

Abstract (English)

Brain-computer interface (BCI) technology utilizing electroencephalography (EEG) marks a transformative innovation, empowering motor-impaired individuals to engage with their environment on equal footing. Despite their promising potential, the development of subject-invariant and session-invariant BCI systems remains a significant challenge due to the complex and dynamic nature of neural activity, which varies across individuals and over time, as well as the EEG hardware constraints. Although prior studies have sought to develop robust BCI systems, their approaches remain mostly ineffective in capturing the intricate spatiotemporal interdependencies in multichannel EEG signals. To address these limitations, this study introduces the attentive graph-temporal convolutional network (AGTCNet), a novel graph-temporal model for subject-invariant and session-invariant motor imagery EEG (MI-EEG) classification. By incorporating prior knowledge through an EEG channel adjacency graph representing the topographic arrangement of electrodes and leveraging the proposed graph-convolutional attention (GCAT), AGTCNet jointly extracts expressive spatiotemporal representations from EEG signals. Model ablation was performed alongside investigations on the impact of EEG channel resolution and signal duration. The model demonstrated state-of-the-art performance, outperforming existing MI-EEG classifiers with lower variance. With a 49.87% model size reduction and 64.65% faster inference time, AGTCNet achieved moving average accuracies of 66.82% (87.89%) on the BCI Competition IV Dataset 2a (BCICIV2A), 64.14% (85.22%) on the EEG Motor Movement/Imagery Dataset (EEGMMIDB), and 38.07% (66.78%) on Emotiv Epoc Flex Saline for 4-class (2-class) subject-invariant classification. When fine-tuned for subject-specific classification, the moving average accuracies improved significantly to 82.88% (94.99%) on BCICIV2A, 72.13% (90.54%) on EEGMMIDB, and 54.41% (74.04%) on Emotiv for 4-class (2-class) classification. As guidance for future research, this study formally established a practical training-evaluation framework for subject-specific and subject-invariant BCI classifications.

Abstract Format

html

Abstract (Filipino)

None

Abstract Format

html

Language

English

Format

Electronic

Keywords

Electroencephalography

Upload Full Text

wf_yes

Embargo Period

2025

Download

Share

COinS