Evaluating the impact of pink noise injection on EEG-based epileptic seizure detection

EKİM G.

JOURNAL OF THE FACULTY OF ENGINEERING AND ARCHITECTURE OF GAZI UNIVERSITY, cilt.41, sa.1, ss.703-717, 2026 (SCI-Expanded, Scopus, TRDizin)

Özet

Purpose: This study evaluates the robustness of EEG based epileptic seizure detection systems under pink noise, a realistic distortion with dominant low frequency energy, and systematically examines its impact on feature extraction and classification performance beyond noise-free conditions. Theory and Methods: Two feature extraction approaches were evaluated: PSD-based spectral difference features using class-specific reference spectra, and DWT-based statistical features with Coiflet 4 as the optimal wavelet. Classification was conducted with Random Forest, Multilayer Perceptron, and k-Nearest Neighbor. Pink noise (-20 dB to +25 dB) was added to Bonn EEG signals, and performance was assessed using accuracy and F1 score under 10-fold cross validation. Results: PSD-based features with Random Forest achieved up to 99.6% accuracy and remained robust under moderate noise, with low-to-moderate pink noise (-3 dB to +5 dB) even enhancing generalization. In contrast, DWT-based features were less stable, showing strong dependence on wavelet parameters and noise intensity. Conclusion: This study proposes pink noise injection as a realistic framework for evaluating seizure detection robustness. Results emphasize the efficiency and clinical relevance of the PSD-Random Forest approach and suggest that controlled noise may enhance generalization. Future work will extend analyses to multi-channel datasets and investigate noise as a potential regularization mechanism in EEG-based models.