Classification of Different Motor Imagery Tasks with the Same Limb Using Electroencephalographic Signals

Stroke is a neurological condition that often results in long-term motor deficits. Given the high prevalence of motor impairments worldwide, there is a critical need to explore innovative neurorehabilitation strategies that aim to enhance the quality of life of patients. One promising approach involves brain-computer interface (BCI) systems controlled by electroencephalographic (EEG) signals elicited when a subject performs motor imagery (MI), which is the mental simulation of movement without actual execution. Such systems have shown potential for facilitating motor recovery by promoting neuroplastic mechanisms. Controlling BCI systems based on MI-EEG signals involves the following sequential stages: recording the raw signal, preprocessing, feature extraction and selection, and classification. Each of these stages can be executed using several techniques and numerous parameter combinations. In this study, we searched for the combination of feature extraction technique, time window, frequency range, and classifier that could provide the best classification accuracy for the BCI Competition 2008 IV 2a benchmark dataset (BCI-C), characterized by EEG-MI data of different limbs (four classes, of which three were used in this work), and the NeuroSCP EEG-MI dataset, a custom experimental protocol developed in our laboratory, consisting of EEG recordings of different movements with the same limb (three classes-right dominant arm). The mean classification accuracy for BCI-C was 76%. When the subjects were evaluated individually, the best-case classification accuracy was 94% and the worst case was 54%. For the NeuroSCP dataset, the average classification result was 53%. The individual subject's evaluation best-case was 71% and the worst case was 35%, which is close to the chance level (33%). These results indicate that techniques commonly applied to classify different limb MI based on EEG features cannot perform well when classifying different MI tasks with the same limb. Therefore, we propose other techniques, such as EEG functional connectivity, as a feature that could be tested in future works to classify different MI tasks of the same limb.