Comparative Analysis of Minimax-Optimized Hybrid MobileNetV2–BiLSTM Model with State-of-the-Art Models for Epileptic Seizure Detection

Abstract

In neurological healthcare, seizure detection is a key obligation that may enhance clinical evaluation and patient monitoring by accurately and promptly identifying seizure episodes. The temporal linkages that happen between successive observations may not be captured by conventional image-based approaches as they often examine individual frames separately. This paper proposes a hybrid deep learning architecture that combines Bidirectional Long Short-Term Memory (BiLSTM) networks for automated seizure detection with MobileNetV2-based convolutional neural network (CNN) feature extraction. Sequences of seizure and non-seizure images make up the input data. The images are scaled to 128 × 128 pixels, converted to RGB format, normalized, and organized into ten-frame sequences. Global Average Pooling is used to extract dense spatial data from every frame using a pretrained MobileNetV2 network. Two stacked BiLSTM layers with 128 and 64 units examine the recovered feature sequences in order to learn temporal relations in both forward and backward orientations. The concluding seizure possibility is then produced using a fully linked classification head with dropout and sigmoid activation. Class weighting, early halting and model checkpointing are used to lessen class imbalance and improve training stability. Accuracy, precision, recall, F1-score, confusion matrix and ROC-AUC are used to measure the model performance. To highlight seizure, recall and lower the possibility of missing seizure existences, ROC-based threshold optimization is also used. The proposed CNN–BiLSTM architecture shows the potential of hybrid deep learning for automated seizure detection and offers a unified method for learning both temporal and spatial properties from sequential data.