Semantic Feature Extraction from EEG Signals

This thesis empirically evaluates how EEG representational form shapes decoding performance across three canonical brain-computer interface paradigms: motor imagery, P300, and SSVEP. It systematically compares five EEG-native representations—Raw time-series, Welch Power Spectral Density (PSD), Spectrograms, Phase-Locking Value (PLV), and Approximate Entropy (ApEn)—paired with four models (LDA, Logistic Regression, CNN, LSTM) under a rigorous, leakage-free evaluation protocol using public datasets (BNCI2014_001, BNCI2014_009, Nakanishi2015). Using fixed computational budgets and subject-aware splitting, the results reveal a clear performance hierarchy. PSD and Raw representations dominate the core metrics when combined with deep architectures, achieving near-ceiling performance for P300 and substantial gains for MI and SSVEP. Deep learning amplifies these advantages but cannot compensate for weak representations: ApEn consistently underperforms across all conditions, while PLV offers only modest utility for MI. CNNs demonstrate superior robustness compared to LSTMs, particularly for frequency-tagged signals. These findings support a conditional Representation–Richness Principle: representational benefits emerge only when the preserved structure is aligned with paradigm-specific neurophysiology and model inductive biases. The study provides reproducible benchmarks and practical guidelines: classical baselines remain competitive for resource-constrained applications, while deep learning on PSD or Raw features offers favorable accuracy–cost trade-offs for high-performance BCIs.