LFSP-DSM: A Lightweight Framework for Seizure Prediction Based on Deep Statistical Model.

Abstract

Epilepsy is a chronic neurological disorder characterized by recurrent seizures. Due to the inconsistency of existing electroencephalogram (EEG) signal labels and the large volume of data, traditional machine learning algorithms for epileptic seizure prediction are overly complex and have long prediction cycles. To address this issue, we propose a lightweight seizure prediction framework named LFSP-DSM. This framework integrates a hybrid enhancement model (HEM) and deep statistical models. Through the multidimensional data enhancement module of HEM, the features of EEG signals are enhanced in both spatial and temporal dimensions, improving the model's predictive capability. The deep statistical model is decoupled into a statistical model StaM and a lightweight convolutional neural network LCNet. StaM generates a new multidimensional EEG signal dataset label online, while LCNet learns multilevel features of EEG signals through parallel pathways. Finally, we designed an end-to-end prediction framework, adapting a new loss function and response success rate evaluation metric. Extensive experiments demonstrate that LFSP-DSM achieves response success rates (for seizure frequency and timing) and accuracy of 91%, 86%, and 93.24%, respectively, validating the effectiveness of LFSP-DSM in handling epileptic sequence data. In particular, it provides an effective solution for improving prediction performance and capturing complex signal patterns.