A Comprehensive Review for Seizure Prediction Based on EEG Dataset: From Handcrafted Features to Deep Neural Models

Over 65 million individuals worldwide suffer from epilepsy, a neurological condition marked by frequent and spontaneous seizures, which greatly impacts the safety and quality of life. Electroencephalogram (EEG) signal-based epilepsy seizure prediction has become a viable approach that allows for prompt alerts and intervention. With an emphasis on three key elements—available public datasets, feature extraction techniques, and deep learning models- this review offers a thorough analysis of the state of EEG-based seizure prediction. Research published between 2019 and 2025 was thoroughly searched, screened, and evaluated using a PRISMA-based methodology. The paper examines the properties of the dataset, compares Deep Learning (DL) models, analyzes handcrafted features (time-domain, frequency-domain, nonlinear, and time-frequency descriptors), and automated deep-learning features for EEG signals. Significant findings show that dataset variability, including sampling rate, electrode montage, seizure frequency, and annotation quality, directly affects model performance and generalization. The seizure prediction models based on deep learning have shown improvement in terms of Accuracy, Sensitivity, and False Prediction Rate per Hour(FPR/h), but there is scope for improvement in the model’s ability to generalize across different patients, datasets, and recording conditions.