SWLDA offers a valuable trade-off between interpretability and accuracy for rehabilitative BCIs

Interpretability, accuracy and a solid neurophysiological basis can be considered as the main requirements for the classification model to monitor motor imagery tasks in post-stroke motor recovery paradigms supported by the brain-computer interface technology. This study aimed at comparing the accuracy performance of different classification approaches applied on a dataset of 15 stroke patients. We also explored how the variation in the dimensionality of the feature domain would influence the different classifier performance. To this purpose, stepwise linear discriminant analysis (SWLDA), shrinkage linear discriminant analysis, logistic regression, support vector machine, multilayer perceptron, decision tree and random forest classifiers entered in the performance analysis. SWLDA statistically outperformed the classifiers commonly used in sensorimotor-BCI paradigms, achieving 80% in classification accuracy even in case of feature domain dimensionality reduction. The linearity, the interpretability and the accuracy of the SWLDA model, even just by means few EEG electrodes, yielded to consider SWLDA an optimal solution to fulfil the main requirements of the rehabilitation context.