An Ultra-Low Latency and High-Precision Stacked-CNN Model for Epileptic Seizure Prediction

Epilepsy, a prevalent and intricate chronic neurological disorder, poses significant challenges in clinical diagnosis and treatment, especially in promptly identifying pre-seizure periods. To comprehensively address the challenges of computational efficiency, storage cost, and real-time performance in multi-channel EEG analysis, this paper proposes a novel multimodal feature extraction framework specifically designed for single-channel electroencephalogram (EEG) signal modeling. The framework integrates traditional analytical techniques which include time and frequency analysis, short-time Fourier transform (STFT) and discrete wavelet transform (DWT) and addresses the inherent limitation of poor generalization in these methods. To mitigate this issue, we introduce nonlinear approaches such as local neighbor descriptive pattern (LNDP) and deep learning models like Transformer, thereby constructing a multimodal heterogeneous feature extraction architecture. Additionally, to overcome the limitation of long training times commonly observed in convolutional neural network (CNN) -based methods, we incorporate efficient models such as Gradient Boosting Decision Trees (GBDT) and propose a stacked convolutional neural network model, significantly improving prediction efficiency. To validate the generalization capability of the model, we transfer the trained system to the Children’s Hospital Boston and the Massachusetts Institute of Technology (CHB-MIT) database. The results demonstrate that the proposed model achieves the testing accuracy of 95% with an inference time of 4 ms. Moreover, in transfer learning scenarios, the model attains a loss of approximately 7.56e-9 and processes 1128 data in merely 130 ms, which is approximately one twentieth of the time required by a multi-scale CNN model. These results highlight the model's potential to improve patients with epilepsy outcomes through timely and accurate seizure prediction.