@article{scholars18530,
            year = {2023},
             doi = {10.1016/j.engappai.2023.106122},
            note = {cited By 0},
          volume = {122},
         journal = {Engineering Applications of Artificial Intelligence},
           title = {A comparative study on end-to-end deep learning methods for Electroencephalogram channel selection},
             url = {https://www.scopus.com/inward/record.uri?eid=2-s2.0-85150309438&doi=10.1016\%2fj.engappai.2023.106122&partnerID=40&md5=5d7e377b1451a131fed6d56eabf4b2ea},
        keywords = {Backpropagation; Brain computer interface; Convolutional neural networks; Learning systems; Long short-term memory; Network architecture, Brain{\^a}??computer interface; Channel selection; Comparatives studies; Convolutional neural network; Deep learning; Electroencephalogram channel selection; End to end; Goal models; Learning methods; Wrapper methods, Electroencephalography},
        abstract = {The electroencephalogram (EEG) channel selection approach with deep neural networks enables us to fit the channel selection to the goal model without dealing with the high processing expenses associated with wrapper methods for an efficient, integrated brain{\^a}??computer interface (BCIs) study. An end-to-end deep learning-based methodology is proposed in this work to choose the proper channels for BCI paradigms. The EEG channel selection and network parameters are optimized using a concrete selector layer. The Gumbel-softmax approach is used in this layer to include the discrete parameters involved in the selection process into continuous relaxations, allowing for end-to-end backpropagation learning of the parameters. This study suggests a regularization method to reduce this tendency, as the selection layer is frequently shown to include the same channel twice in a particular selection. We performed six models using the suggested framework, and our proposed study, MSFBCNN+LSTM, extracts significant and robust properties compared to other models. For verification, high gamma dataset is used for each model to investigate and observe good accuracy. Overall, the MSFBCNN+LSTM architecture has the maximum classification mean accuracy of 90.59 in the specified tasks. This paper suggests that using integrated architectures of deep learning model learns the best features and contributes to BCI's performance. {\^A}{\copyright} 2023 Elsevier Ltd},
          author = {Abdullah, {} and Faye, I. and Islam, M. R.}
}