Motor Imagery EEG Signals: Multi-Task Classification and Subject Identification with a Lightweight CNN

Abstract

Motor imagery is a domain of the brain-computer interface where individuals imagine moving their body parts without any physical movement occurring. This field has significant applications in enhancing abilities in individuals with physical impairments. However, its applications are not limited to disabled people; it can also improve the quality of life in healthy individuals. This paper investigated motor imagery EEG signals using a lightweight CNN model without any manual or traditional feature extraction, exploring two objectives: motor imagery task classification and subject identification. For task classification, the study began with binary classification and extended to a ten-task scenario encompassing hands, feet, tongue, and the five fingers of one hand. For subject identification, the model was initially trained to recognize subjects based on specific motor imagery tasks, uncovering task-related patterns that distinguished individuals. Then, the trained models were applied in the evaluation phase to new tasks to assess the task dependency of these patterns and their generalizability across tasks. The data used in this research was from the HaLT and 5F paradigms of the Kaya dataset. Our model achieved average accuracies of 85.88% for 2-task, 69.60% and 74.64% for 5-task (using Aggregate and Individualized approaches), and 57.84% for 10-task classification. It also achieved 95.60% and 98.23% accuracy for subject identification in the HaLT and 5F paradigms, respectively. These results were achieved with very low variation in accuracy across folds, despite having only 15,000 parameters, making it suitable for deployment in various BCI systems.

Tasks

Reproductions