Frequent seizures of epilepsy affect millions of people worldwide. The ability to predict seizures in advance could significantly enhance the quality of life for patients by timely interventions. Electroencephalography (EEG) is an important auxiliary examination method for epilepsy. The intra-channel of EEG is temporal information, and the inter-channel of EEG is spatial information, which are different. However, few existing prediction methods based on EEG pay attention to this difference. To address this issue, an EEG Intra- and Inter-Channel Feature Mixing ( \(I^{2}\) CM) is proposed to predict epilepsy seizure in this paper. The overall architecture includes three components: pre-process, \(I^{2}\) CM and prediction head. The raw EEG is segmented and filtered in pre-process. \(I^{2}\) CM mixes temporal and spatial information using intra-channel feature extractor and inter-channel feature extractor, respectively. Prediction head outputs whether epilepsy onset is about to occur. Moreover, Multi-Layer Perceptron (MLP) and Kolmogorov-Arnold Network (KAN) are utilized in intra-channel feature extractor, inter-channel feature extractor and prediction head for feature extraction. Our method achieves an average accuracy of \(96.34\%\) , a specificity rate of \(96.21\%\) , and a sensitivity rate of \(96.47\%\) on CHB-MIT dataset, an average accuracy of \(98.21\%\) , a specificity rate of \(98.64\%\) , and a sensitivity rate of \(98.58\%\) on Kaggle dataset, offering promising prospects for application in automatic seizure prediction.