Terrain-based Gait Recognition using EEG: Comparing Machine Learning and Deep Learning Models

Abstract

In the past EEG analysis has been done in many sectors, like emotion classification, standing/swing stance, letter classification. Research has shown a requisite amount of relation between signal generation at cortical and walking during locomotion. In this paper our aim is to classify different terrains like Level Walking, Ramp Descent, Ramp Ascent, Stair Descent, Stair Ascent as an application to decode gait patterns. We have used kinematics data of the toe for a generalized EEG dataset classification into different terrain. As the open sourced EEG dataset was already preprocessed using Artifacts Subspace Reconstruction (ASR) and Reliable Independent Component Analysis (RELICA) which removed any noise due to movements. Our study was evaluated on an open sourced dataset [1] of 11 healthy subjects walking on different terrains with 2 trials for each recording. The dataset was A deep learning based subject independent model was developed approach has been applied here to classify the EEG dataset with complex features into 6 classes. The average accuracy during cross-validaition was 70%,with further accuracy of 68% on the unseen data. Deep learning model showed superiority over machine learning approach using bagging regressor (accuracy – 62%). The present study concludes that deep learning models are better choice given the complexity.