From intent to action: EEG-based multi-finger classification and intention detection across imagined and hand exoskeleton-mediated movements

While electroencephalography (EEG)-driven hand exoskeletons hold potential for restoring hand dexterity in paralyzed individuals, current systems are largely restricted to rudimentary hand open/close commands. Previous studies have demonstrated the feasibility of decoding repetitive finger motor imagery, but decoding under a more intuitive single-flexion imagery paradigm remains insufficiently characterized. Improving continuous motor engagement detection is also essential to support both the initiation and termination of exoskeleton-assisted movement. This study investigates the neurophysiological and algorithmic challenges in advancing hand exoskeletons toward more intuitive finger-level control. EEG data were recorded from 25 motor imagery (MI)-naïve healthy participants during single, non-repetitive flexion tasks targeting the thumb, index, or combined middle-ring-little (MRL) fingers under three conditions: pure MI, exoskeleton-driven passive movement (PM), and exoskeleton-assisted MI (AMI). Participants wore the hand exoskeleton in all conditions. For the MI condition, subjects were instructed to imagine flexing the designated target fingers without any physical execution, simulating purely volitional motor engagement. During PM, the target fingers were physically actuated by a commercial, medical-grade hand exoskeleton, providing somatosensory feedback. During AMI, participants additionally imagined the corresponding finger movement while receiving congruent exoskeleton-induced feedback. We analyzed the underlying neural correlates of finger movement for each condition. Moreover, we defined offline multi-finger classification and binary intention detection tasks relevant to different phases of hand exoskeleton control and evaluated them using nine state-of-the-art EEG decoding models. Exoskeleton-mediated movements (PM and AMI) elicited stronger sensorimotor EEG responses than MI alone and yielded higher multi-finger (three-class) classification accuracies with a convolutional neural network (46.9% and 47.3%, respectively, versus 34.6% for MI). However, these stronger sensorimotor responses were accompanied by greater difficulty in distinguishing the presence of motor imagery during actuation: intention detection performance was lower in the actuated mode (AMI vs. PM) than in the non-actuated mode (MI vs. Rest). Direct cross-mode transfer revealed limited generalization: a convolutional–Transformer model achieving 79.6% balanced accuracy in the non-actuated mode achieved 59.0% when applied to the actuated mode. In MI-naïve participants, distinguishing finger targets during single, non-repetitive imagined flexions remained challenging, whereas intention detection benefited from advanced decoding models even during exoskeleton actuation. These findings support the further development of continuous motor engagement monitoring as a step toward intuitive EEG-driven hand exoskeletons. Such development should account for exoskeleton-induced somatosensory feedback and limited generalization across actuation modes. Real-time validation is needed to establish whether these systems can reliably enable users to initiate and terminate exoskeleton-assisted finger movements.