EEG Signal Classification Using Machine Learning for Wheelchair Control

Resumo

Brain-computer interfaces (BCI) represents an emerging technology that enables direct communication between the human brain and external devices, using electroencephalographic signals. During motor imagery, specific changes occur in brain sensorimotor rhythms that can be detected and classified by Machine Learning (ML) algorithms. This work presents a comparative analysis of ML algorithms for the classification of motor imagery patterns in electroencephalographic signals, aiming at the development of brain-computer interfaces for wheelchair control. Four distinct algorithms were evaluated: Gradient Boosting (GB), Random Forest (RF), Multilayer Perceptron (MLP) and k-NN, using a dataset with 2,950 characterized samples from three experimental conditions: left hand movement imagination, right hand movement imagination, and resting state. Different balancing techniques were implemented, including SMOTE, ADASYN and class weight balancing. GB demonstrated the best overall performance with 78.84% Macro F1-Score and 87.63% accuracy, followed by RF with 76.85% and 87.97%, respectively. The analysis revealed significant difficulties in discriminating minority classes due to extreme imbalance in the dataset, where the left class represents 74.5% of the samples.