 This paper proposes a novel hybrid transformer model for epilepsy prediction. It uses two feature engineering methods and a transformer based model to explore the applicability of transfer learning, TL, techniques and model inputs for different deep learning, DL, structures. Additionally, it tests the performance of two model structures using patient independent approaches and two TL strategies. The results show that the proposed method performs other CNN-based models and provides a promising solution for epilepsy prediction. Furthermore, the information contained in the gamma, dollar-back-slash-gamma-dollar rhythm is found to be beneficial for epilepsy prediction. This work has potential implications for developing personalized models for clinical applications. This article was authored by Schweizsamhu, Jin Liu, Rui Yang and others.