Comparing FFT-Based Features and Recurrent Neural Networks for EEG Signal Classification

Abstract Electroencephalography (EEG) signals exhibit strong non-linear and non-stationary dynamics, particularly during cognitive and sleep-related processes. Traditional spectral analysis techniques such as the Fast Fourier Transform (FFT) rely on assumptions of linearity and stationarity, which limit their ability to capture rapid temporal fluctuations in neural activity. This study presents a comparative analysis between FFT-based feature extraction and recurrent neural network (RNN) models for EEG signal classification. EEG data were recorded from healthy participants during sleep, including targeted memory reactivation (TMR) and control conditions. Spectral features were extracted using FFT-based methods and evaluated using feature selection and model parsimony criteria. In parallel, a long short-term memory (LSTM) network was trained directly on time-series EEG segments to capture temporal dependencies. Results indicate that RNN-based models more effectively model dynamic EEG patterns and outperform FFT-based approaches in classification tasks, particularly in non-stationary contexts. These findings support the use of deep recurrent architectures for EEG analysis in cognitively dynamic scenarios.