Dual imagery tasks for EEG-based brain-computer interfaces: a feasibility study of combinatorial mental state classification.

Introduction: Active brain-computer interfaces (BCIs) aim to enable communication and control without requiring physical movement. BCIs operate by detecting and decoding distinct patterns of brain activity associated with intentionally-generated mental tasks and translating them into commands for external devices. Motor imagery (MI) tasks involving the imagined movement of body parts are the most commonly used control paradigms in active BCIs; however, such systems are typically limited to a relatively small number of commands. Expanding the number of distinguishable mental states could increase the flexibility and information transfer capacity of active BCIs. In the present work dual imagery (DI) tasks are introduced. DI tasks involve the simultaneous performance of two mental tasks, specifically singing imagery (SI) combined with MI of one hand or both feet. This study investigated the feasibility of classifying neural activity associated with such combinatorial imagery tasks using EEG, thereby evaluating their potential to expand the number of distinguishable control states in active BCIs. Methods: EEG was recorded from 14 healthy participants while performing the specified mental tasks. Following pre-processing, a filter bank common spatial patterns (FBCSP) approach and a regularized linear discriminant analysis (LDA) classifier were used for feature extraction and classification in 3-, 4-, 7-, and 8-class scenarios. The 3-class analyses evaluated whether each DI task could be differentiated from its constituent single tasks, while the 4-class analyses incorporated a "rest" state. The 7- and 8-class scenarios explored the feasibility of increasing the number of distinguishable BCI control states beyond six. Results: Accuracies as high as 64.1% and 63.0% were achieved for the 3- and 4-class scenarios, respectively, demonstrating that DI tasks could be differentiated from both their constituent single tasks and a rest state. In the 7- and 8-class scenarios, average accuracies of approximately 55% and 50%, respectively, substantially exceeded the corresponding theoretical chance levels of 14.3% and 12.5%. Conclusion: While the accuracies obtained are currently insufficient for practical high-command active BCIs, the results demonstrate the feasibility of classifying neural activity associated with combinatorial imagery tasks. These findings support the potential of DI paradigms to expand the number of distinguishable control states in active BCIs.