Seizure Prediction From Preictal EEG Using Zoom Time-Frequency Distribution

Abstract

Epileptic seizures remain a major neurological challenge due to their sudden onset and significant impact on patient quality of life. Accurate characterization of preictal brain activity can support the development of reliable seizure monitoring systems and improve understanding of seizure-transition dynamics. In this study, we present the first application of the Zoom Time-Frequency Distribution (Zoom TFD), a high-resolution time-frequency analysis technique, for the classification of preictal and ictal EEG signals. EEG segments from the CHB-MIT database were transformed into Zoom TFD-based time-frequency representations and classified using GoogLeNet through a transfer-learning framework. The proposed framework achieved classification accuracies of 95.1%, 99.64%, and 96.69% for 8-minute, 30-minute, and 45-minute preictal windows, respectively. In addition, the method demonstrated high sensitivity and specificity while requiring only four EEG channels, reducing computational complexity compared with many existing multichannel approaches. The study further investigates the influence of the $S_{\text {min }}$ parameter on the discriminative quality of the generated representations. Experimental results demonstrate that Zoom TFD provides highly resolved spectral representations capable of capturing seizure-related EEG dynamics effectively. These findings highlight the potential of Zoom TFD combined with deep learning for EEG state classification and seizure-transition analysis.