EfficientNetV2_S-AbiLSTM: A novel cross-modal lightweight transfer learning framework for seizure prediction using EEG spectrograms

Abstract

Background

Epilepsy is a common chronic neurological disorder that demands accurate diagnosis and prediction. EEG is the primary tool for monitoring, yet direct application of deep learning to EEG faces challenges such as limited data, poor signal adaptability, and model complexity. We propose EfficientNetV2_S-AbiLSTM, a lightweight model that leverages cross-modal transfer learning to bridge image classification and EEG processing for enhanced seizure prediction.

Methods

EEG signals are converted into spectrograms using the Short-Time Fourier Transform and reduced from 22 to 3 channels via group averaging, mimicking RGB images. EfficientNetV2_S with pre-trained ImageNet weights extracts time-–frequency features that are fed into an attention-enhanced bidirectional LSTM (AbiLSTM) for pattern recognition. A fully connected layer produces the final classification. The model is trained and validated on the CHB-MIT dataset using five-fold cross-validation.

Results

Our framework achieves 97.20 % accuracy, 97.26 % precision, 97.05 % recall, and an AUC of 0.9709. Ablation studies confirm that cross-modal transfer learning improves accuracy by 1.99 %, while EfficientNetV2_S outperforms ResNet, ResNeXt, and other EfficientNet variants with reduced training time.

Conclusion

Incorporating pre-trained image classification models through cross-modal transfer learning significantly enhances seizure prediction. The EfficientNetV2_S-AbiLSTM model shows promising potential in medical signal processing.