Classification of Motor Imagery Waves using Hybrid-Convolutional Neural Network

Abstract

Brain Computer Interfaces augments, alters, or replaces a lost biological function. Recently, classification methods using CNN were proposed to achieve higher accuracy levels. Nonetheless, they use a single convolution for classification while the best scale differs from subject to subject. This paper proposes a different architecture of Deep learning that takes 'n' different uncorrelated features of same signal parallelly into non-shareable convolution input layers in the same network to predict kinetic motion of patients. That is referred to as 1-D Hybrid-Convolutional Neural Network. The general motivation towards this being accurate feature extraction when minimal dataset is available. This approach is performed over three features namely Power, Frequency Spectrum components and Power Spectral Density of a same segment of a signal. A detailed analysis on more than 1500 EEG recordings from 109 healthy subjects and a comparative edge to this study was performed using previous algorithms and the relative strength highlighted.