Trends and Gaps in Transformer-Based EEG Modeling: A Review of Recent Developments

In recent years, Transformer-based deep learning architectures have emerged as a powerful paradigm for modeling EEG signals, offering superior capability in capturing spatial–temporal dependencies compared to traditional convolutional or recurrent networks. However, the diversity of model designs, limited dataset generalization, and lack of standardization have created challenges in evaluating their true potential for real-world applications. This review addresses these issues by systematically examining the evolution, performance, and methodological trends of Transformer-based EEG models published between 2022 and 2024, highlighting both achievements and research gaps. The main contribution of this study is to provide a comprehensive mapping and critical analysis of Transformer architectures applied to EEG classification, feature extraction, and signal decoding tasks. Using the Scopus database, a structured search was conducted following specific inclusion criteria (English, peer-reviewed, open-access journal papers from 2022–2024) and a well-defined query combining EEG and Transformer-related keywords. Data from 63 eligible studies were extracted and categorized according to authorship, dataset, architecture type, EEG application, and evaluation metrics. Results show that hybrid Transformer models dominate recent research, achieving accuracies above 90% in tasks such as motor imagery, emotion recognition, seizure detection, and sleep staging. Pure Transformers like ViT and BERT-like models also demonstrate competitive performance but face scalability and interpretability challenges. In conclusion, Transformer-based EEG modeling is advancing rapidly, yet future efforts must focus on model efficiency, explainability, and benchmark standardization to enable broader clinical and real-world adoption.