Child Seizure Prediction Using Temporal Features with Temporal Convolutional Neural Network (TCNN)

Abstract

Early prediction of epileptic seizures is clinically important because it allows for immediate treatment, reduce injury risk, and improve the quality of life of pediatric patients. This paper presents a subject-specific three-class seizure prediction framework using multichannel electroencephalography (EEG) and a Temporal Convolutional Neural Network (TCNN). The proposed method is evaluated on the publicly available CHB-MIT scalp EEG database. During preprocessing, the raw EEG signals were band-pass filtered between 0.5--40~Hz, notch filtered at 60~Hz, segmented into fixed 5-second windows, and labeled as interictal, pre-ictal, and ictal using seizure annotations, a seizure prediction horizon of 5 minutes, and a seizure occurrence period of 30 minutes. The proposed TCNN learns temporal EEG patterns directly from multichannel inputs, while weighted sampling and focal loss are used to address class imbalance. The main contribution of this study is the development of a lightweight subject-specific three-class framework with class-wise, threshold-independent evaluation. For a representative validation subject, the model achieved 89.03\% accuracy, a macro F1-score of 0.8239, ROC-AUC values of 0.958, 0.965, and 0.997 for interictal, pre-ictal, and ictal classes, respectively, and average precision values of 0.988, 0.874, and 0.952. The results indicate that the proposed TCNN is effective for seizure-state discrimination and near-onset warning, while pre-ictal prediction remains the most challenging task. Overall, this study demonstrates that temporal convolution is a promising approach for pediatric EEG-based seizure prediction, with future work needed to improve pre-ictal specificity.