A Somatosensory ERP-BCI System for Post-Stroke Hand Function Rehabilitation Applications

We are presenting the design and validation of a brain-computer interface (BCI) system based on somatosensory event-related potential (ERP), as part of an integrated BCI-exoskeleton rehabilitation device. To engage with the device, the user expresses intent of hand movement (left or right) by attending to the somatosensory stimulation applied to the corresponding wrist. The system classifies the 2-class intent by processing the resulting ERP responses with a transformer-based convolutional neural network (CNN). Steps were taken to alleviate the calibration workload in clinical settings, including using 3-channel dry-electrode electroencephalography (EEG) headset to shorten hardware preparation, and implementing transfer learning to reduce algorithm calibration time. To validate the system, we conducted a 2-stage ERP-BCI experiment involving both healthy and patient subjects and performed off-line evaluation. As a main result, the system can achieve up to 84.7 % mean movement intent classification accuracy among stroke patients in real-world environment. On average, 4 ERP trials (corresponding to classification at 8.8-second interval) were needed to reach 70 % accuracy. When implementing transfer learning, healthy subject data were used to train a general model which was then fine-tuned with portions of target patient data, resulting in reduced algorithm calibration sessions. System preparation time was reduced to 4.9 minutes for patient users, accredited to the simplified EEG headset design. While somatosensory ERP was previously proven as an effective BCI paradigm, we validated it as a feasible channel to interact with robotic rehabilitation system. The system provides an alternative to the existing BCI systems for clinical rehabilitation applications.