An adaptive frequency partitioning framework for epileptic seizure detection using TransseizNet

Epilepsy is a disorder causing repeated seizures because of unusual brain activity recorded using electroencephalography. Nevertheless, conventional epilepsy seizure detection approaches face difficulties such as poor epilepsy seizure detection accuracy and higher computational complexity. To overcome these limitations, this work proposes a novel TransseizNet framework for epilepsy seizure detection from the electroencephalography signal.

The electroencephalography data from three datasets are pre-processed using the Savitzky-Golay filter. The proposed framework utilizes the Empirical Tunable Q-Wavelet Transform for signal decomposition, which is the combination of the Empirical Wavelet Transform and the Tunable Q-factor Wavelet Transform. This enhances time-frequency resolution and adaptively captures localized oscillatory patterns critical for precise seizure detection. The proposed framework utilizes a Wavelet-Graph Convolutional Network Vision Transformer for epilepsy seizure detection and classification. The integration of wavelet-driven attention with graph-based learning enhances spatial-temporal feature representation, which makes seizure detection more accurate, interpretable, and computationally efficient than the baseline approaches.

The TransseizNet model is trained and validated on three datasets and achieves an average accuracy of 98.65% a precision of 98.59%, a F1-score of 98.45%, a recall of 98.30%, a specificity of 98.52%, a computational time of 17 sec, and the detection latency of 2.5 sec, which outperforms the performance of baseline approaches in the detection of epileptic seizures.

TransseizNet framework provides superior performance in seizure detection by efficiently integrating adaptive frequency decomposition and hybrid deep learning. Its minimal detection latency, higher accuracy, and interpretability make it suitable for practical healthcare uses.