[2Q3-J-2] Machine learning: explainability, knowledge acquisition

[2Q3-J-2-03] Explanation of frequency domain features contributing to EEG classification

Kazuki Tachikawa¹, Yuji Kawai¹, 〇Jihoon Park¹, Minoru Asada¹ (1. Osaka University)

Keywords:Deep learning, Electroencephalogram, Interpretability, Integrated gradients, Explanation

Interpreting prediction of deep learning models is important in many applications, especially in medical diagnosis systems. Several methods have been proposed to interpret black-box predictions, but most of these studies are intended to calculate the contributions of input features. In this paper, we propose a method to compute the contributions in an interpretable feature space. The method applies differentiable transformations to input features to create interpretable features. This approach enables to calculate the contribution of amplitude and phase for each frequency in EEG classifications because the fast Fourier transform is differentiable. The proposed method is verified using three EEG datasets and the results show that the contributions of the proposed method are more reliable and has less computational cost than those of a conventional method. Our method will thus enhance the reliability of data-driven approaches in EEG analysis.