ELECTROENCEPHALOGRAPHIC ANALYSIS FOR CLASSIFICATION OF MOTOR EXECUTION AND MOTOR IMAGERY IN A HEALTHY POPULATION

The brain regulates bodily functions, including movement, cognition, and sensory processing, through neurons and glial cells. Motor execution and motor imagery represent two complementary conditions that share overlapping neural substrates, yet differ in the presence or absence of overt motor output. Beyond this fundamental distinction, they are known to exhibit a range of electrophysiological differences which we aim to capture and exploit through a comprehensive set of EEG-derived features. Understanding the electrophysiological differences between these conditions is essential for the development of reliable brain–computer interfaces (BCIs) for motor rehabilitation and assistive technologies. Electroencephalography (EEG) is a key tool for studying brain activity, offering insights into motor-related cortical dynamics and mental states. The aim of this study is to investigate EEG-derived features that can be used to discriminate between motor execution and motor imagery, exploiting machine learning techniques. To do so we have analyzed the EEG data of a healthy population during motor execution and motor imagery tasks coming from 109 subjects, part of a freely available database on PhysioNet, specifically designed to support research investigating motor-related cortical activity. The tasks performed include execution and imagery of right hand, left hand, both fists, and both feet movements. EEG recordings were preprocessed using EEGLAB software and to ensure that the data were clean, sliding windows were extracted and the noisy ones were removed. The preprocessing involved 0.5–70 Hz band-pass filtering, average-based re-referencing, and artifact removal after independent component analysis using the ICLabel plug-in. Then, EEG rhythms were extracted and their powers computed, and 22 engagement indexes were computed according to the review of Marcantoni et al. In addition, event-related desynchronization (ERD) in the alpha and beta bands, frontal alpha asymmetry, and connectivity measures like coherence and phase-locking value, were extracted. Then, for each of the four movement types considered, we were able to develop classification models using four different classifiers: Linear Discriminant Analysis, Support Vector Machine with linear and radial basis function kernels, and Random Forest. Model performance was evaluated using both 10‑fold cross‑validation and leave‑one‑subject‑out cross‑validation. The Random Forest classifier consistently outperformed the other models, achieving 10‑fold cross‑validation accuracies ranging from 62.53% to 67.13% using region‑aggregated features. The feet movement yielded the highest performance, with an accuracy of 67.13% and an F1‑score of 61.68%. However, performance was considerably lower under leave‑one‑subject‑out cross‑validation (52.14–53.62% accuracy), highlighting the critical challenge of cross‑subject generalization. It should be noted that even the 10‑fold cross‑validation accuracies, although above chance, remained moderate, indicating that the classification task is inherently complex. This finding underscores the importance of considering subject‑specific variability and the need to incorporate robust feature extraction and domain adaptation strategies in motor imagery classification tasks. This study also highlighted the complexity of using EEG for motor decoding, showing that an approach that accounts for a wide array of features is crucial to develop accurate and generalizable models. The significant role of machine learning techniques like Random Forest further emphasizes their potential in handling high‑dimensional data and selecting the most informative features, while also stressing the need for advanced domain generalization methods to overcome inter‑subject variability and improve the reliability of BCI systems.