A Channel-Selected DWT-ANOVA and CNN-BiLSTM Framework for Multiclass Epileptic Seizure Detection from EEG Signals

Abstract

Multiclass epileptic seizure detection from Electroencephalogram (EEG) signals is difficult because seizure-related patterns are nonstationary, patient-dependent, and distributed across redundant EEG channels. This study develops a detection framework that combines Coefficient of Variation (CV)-based channel selection, Discrete Wavelet Transform (DWT) decomposition, Convolutional Neural Network (CNN) feature learning, Analysis of Variance (ANOVA) feature selection, and CNN-Bidirectional Long Short-Term Memory (BiLSTM) classification. The Children's Hospital Boston - Massachusetts Institute of Technology (CHB-MIT) EEG recordings were transformed into five segment classes: non-seizure, pre-seizure 1, pre-seizure 2, pre-seizure 3, and seizure. A rank-based CV strategy retained seven informative channels, DWT generated time-frequency representations, CNN produced feature maps, and nine statistical descriptors were selected using ANOVA before classification. In the single train-test evaluation, CNN-BiLSTM achieved 98.40% accuracy, 98.46% precision, 98.40% recall, and 98.39% F1-score, outperforming the standalone CNN and BiLSTM models. The 10-fold cross-validation produced mean metric values above 92%, indicating more conservative but stable performance across data partitions. These results suggest that integrating channel reduction, time-frequency decomposition, statistical feature optimization, and spatial-temporal learning can provide a compact and discriminative framework for five-class EEG seizure classification.