Multi-Scale-Rhythm Attention Feature Contrastive Learning for Epileptic Seizure Prediction

The electroencephalogram (EEG) signals based epileptic seizure prediction remains a pivotal technique to improve the living standards for epilepsy patients. However, recent prediction models focus on the local feature extraction with smaller receptive field, which ignores multi-level global feature changes in pre-ictal period. Besides, traditional supervised or semi-supervised learning strategies requires sufficient data in individual patient, and it is difficult to handle the representation learning of imbalanced samples of intra-patient EEG data, yielding a suboptimal performance in predicting seizures. To overcome above limitations, a multi-scale-rhythm attention feature contrastive learning (MSR-AFCL) model is proposed for the prediction of seizures in individual patients. First, a feature extractor called the multi-scale-rhythm (MSR) employs multi-scale dilated convolution module and multi-rhythm power spectrum density (PSD) estimate module, so as to extract multi-level spatio-temporal features and rhythmic temporal-spectral features with large receptive field. Then, an attention feature contrastive learning (AFCL) strategy is developed, and it has the capability to dynamically adjust the intra-class and inter-class distances of attention fusion representations. The experimental results indicate that our proposed MSR-AFCL model achieves outstanding seizure warning when evaluated on the publicly available CHB-MIT dataset, and it outperforms the state-of-the-art methods.