Classificação de sinais de eletroencefalograma utilizando técnicas de aprendizado de máquina para controle de cadeira de rodas

Abstract

This work investigated the classification of motor imagery electroencephalography (EEG) signals from the Milo Project for brain–computer interface applications, comparing two decision strategies: (i) a flat multiclass architecture (Rest, Left, and Right) and (ii) a cascade architecture composed of two stages (Rest × Movement and, conditioned on Movement, Left × Right). The experimental pipeline adopted a session-level data split to mitigate data leakage (55% training, 20% validation, and 25% testing), temporal preprocessing with linear trend removal and filtering, followed by segmentation into 2 s windows (500 samples at 250 Hz). Feature extraction was based on Riemannian geometry, including the estimation of regularized covariance matrices, the computation of the log-Euclidean mean exclusively on the training set, and projection onto the tangent space, followed by redundancy removal and univariate feature selection. Class imbalance was addressed only on the training set, evaluating resampling techniques such as Random Over-Sampling and Random Under-Sampling, Synthetic Minority Over-sampling Technique (SMOTE), Borderline Synthetic Minority Over-sampling Technique, Adaptive Synthetic Sampling (ADASYN), Tomek Links, Edited Nearest Neighbors (ENN), NearMiss, and hybrid approaches, such as Synthetic Minority Over-sampling Technique combined with Tomek Links and Synthetic Minority Over-sampling Technique combined with Edited Nearest Neighbors. The classifiers Multilayer Perceptron (MLP), Support Vector Machine (SVM), Random Forest, Extra Trees, Gradient Boosting, Extreme Gradient Boosting (XGBoost), and a soft-voting ensemble were trained and compared, with hyperparameter optimization via random search (100 iterations) and five-fold stratified cross-validation, adopting F1macro as the primary criterion, in addition to estimating 95% confidence intervals via bootstrap. The best overall solution in label-based metrics was obtained with the flat architecture using SVM and SMOTEENN (RBF kernel, C = 0.1, γ = 10−4 , balanced class weighting, and 13 features), achieving an accuracy of 0.826±0.013, balanced accuracy of 0.825±0.013, F1macro = 0.822±0.014, and MCC = 0.748 ± 0.019; in the cascade architecture, the best performance was achieved by the Extra Trees model with SMOTE (F1macro = 0.796 ± 0.015), although inferior to the best flat scenario according to the adopted optimization criterion