Novel Time-Series Forecasting Method to Enhance Accuracy of Real-Time EEG Detection for BCI-Based Neurofeedback Motor...

Real-time detection of motor intent using electroencephalography (EEG) with high accuracy remains a technical challenge for neurorehabilitation. Brain-computer interface-based neurofeedback training (BCI-NFT) paradigms need to detect pre-movement EEG to activate robotics or electrical stimulation nearly simultaneously with movement to promote neuroplasticity. We present a novel detection method commonly used in time-series forecasting (e.g., stock market trends), identifying crosses in fast (short) and slow (long) moving average windows to identify negative deflections in slow movement-related cortical potentials (MRCPs) or event-related desynchronization (ERD) within -400-+100 ms of movement onset. We recorded EEG data from the Cz electrode during our cued ankle dorsiflexion BCI-NFT paradigm in four adult participants, two neurotypical and two with cerebral palsy. Simulated real-time offline analyses demonstrated an 85.9% mean true positive rate and 14.1% false positive rate of detecting motor intent at a mean -182 ms from movement onset. We further evaluated whether the detection indicated a MRCP and/or ERD, with MRCP detected in 70-80% of trials in three participants, but high ERD detection (87%) instead in the other. Preliminary results indicate that this approach offers a straightforward, accurate, and well-timed method for real-time EEG detection during neurofeedback training and as a control signal for brain-computer interfaces.