The Impact of Feature Selection on Epilepsy Diagnosis from EEG Signals: A Comparison of Dynamic Time Warping and Itakura-Saito Distance

Epileptic seizure detection is crucial for accurate diagnosis and effective treatment of epilepsy. Automated systems based on electroencephalogram signals play an important role in reducing the time, effort, and subjectivity involved in manual analysis performed by neurologists. However, many previous studies mainly focused on time-domain similarity measures, fuzzy logic models, or neural network-based classifiers and did not fully explore the potential of spectral similarity features for seizure detection. This study introduces a new approach by applying the Dynamic Time Warping method directly to power spectral density data obtained using the Welch method. This enables a spectral similarity analysis that is independent of time shifts and better reflects the non-stationary nature of brain signals. In addition, the Itakura-Saito distance is investigated for the first time for epileptic seizure detection and compared with the Dynamic Time Warping method. A twelve-feature vector was created using reference-based similarity measurements combined with energy and spectral features, and multiple classification methods were evaluated. Experimental results on the Bonn electroencephalogram dataset demonstrate that Dynamic Time Warping achieved the highest classification performance, with the Random Forest algorithm reaching 100% accuracy in binary tasks. These findings highlight the importance of spectral similarity features in improving automatic seizure detection systems and demonstrate the clinical value of the proposed approach.