A Multiclass Framework for Epileptic Seizure Phase Classification Based on EEG Analysis

A persistent neurological condition called epilepsy disrupts the brain's electrical activity, causing frequent and erratic seizures. In this work, a multiclass analysis of epileptic seizures is carried out to classify EEG signals into ictal, interictal, and preictal states using a Random Forest (RF), Long Short - Term Memory (LSTM), and DenseNet based deep learning models. The study utilizes the CHB - MIT Scalp EEG dataset, recorded using the 10-20 bipolar montage with 23 EEG channels from pediatric epilepsy patients. The raw EEG signals are preprocessed using Discrete Wavelet Transform (DWT) and Dual - Tree Complex Wavelet Transform (DTCWT) techniques to suppress artifacts and enhance signal quality. A total of seven time - domain, four frequency - domain, four nonlinear, and three connectivity features are extracted from the preprocessed signals and analyzed across seizure classes. These handcrafted features are classified using Random Forest and LSTM models. Preprocessed EEG signals are first transformed into time - frequency representations. These representations are then fed into a DenseNet based convolutional network for automated spatial - temporal feature learning. Experimental results indicate that DWT based preprocessing outperforms DTCWT, and that nonlinear and connectivity features offer strong class separability. Among the evaluated models, the DenseNet model demonstrates superior performance among all evaluated classifiers, achieving the highest testing accuracy of 93.33%, validation accuracy of 94.25%, and an F1-score of 0.957. The results validate the efficacy of DenseNet in acquiring discriminative time - frequency representations for multiclass seizure phase classification.