Development and evaluation of a BCI-neurofeedback system with real-time EEG detection and electrical stimulation assistance during motor attempt for neurorehabilitation of children with cerebral palsy

Abstract

In the realm of motor rehabilitation, Brain-Computer Interface Neurofeedback Training (BCI-NFT) emerges as a promising strategy. This aims to utilize an individual's brain activity to stimulate or assist movement, thereby strengthening sensorimotor pathways and promoting motor recovery. Employing various methodologies, BCI-NFT has been shown to be effective for enhancing motor function primarily of the upper limb in stroke, with very few studies reported in cerebral palsy (CP). Our main objective was to develop an electroencephalography (EEG)-based BCI-NFT system, employing an associative learning paradigm, to improve selective control of ankle dorsiflexion in CP and potentially other neurological populations. First, in a cohort of eight healthy volunteers, we successfully implemented a BCI-NFT system based on detection of slow movement-related cortical potentials (MRCP) from EEG generated by attempted dorsiflexion to simultaneously activate Neuromuscular Electrical Stimulation which assisted movement and served to enhance sensory feedback to the sensorimotor cortex. Participants also viewed a computer display that provided real-time visual feedback of ankle range of motion with an individualized target region displayed to encourage maximal effort. After evaluating several potential strategies, we employed a Long short-term memory (LSTM) neural network, a deep learning algorithm, to detect the motor intent prior to movement onset. We then evaluated the system in a 10-session ankle dorsiflexion training protocol on a child with CP. By employing transfer learning across sessions, we could significantly reduce the number of calibration trials from 50 to 20 without compromising detection accuracy, which was 80.8% on average. The participant was able to complete the required calibration trials and the 100 training trials per session for all 10 sessions and post-training demonstrated increased ankle dorsiflexion velocity, walking speed and step length. Based on exceptional system performance, feasibility and preliminary effectiveness in a child with CP, we are now pursuing a clinical trial in a larger cohort of children with CP.

Keywords: NMES; associative learning; motor training; pediatric; transfer learning.

Figure 1

Diagram of our Brain-computer interface neurofeedback training (BCI-NFT) system which utilizes MRCP of the EEG signal to detect participant motor intent. The participant was tasked to dorsiflex the ankle at a specific time, cued by a graphic user interface (GUI), which also provided visual feedback of the ankle angle in real-time. EEG data were recorded using a 64-channel EEG system and streamed wirelessly to the BCI-NFT software using Stimulation Delivery. Upon detection of motor intent from EEG, the BCI-NFT system triggered a stimulator to apply assistive sensory feedback (NMES of the tibialis anterior muscle). To monitor ankle angle in real-time, two Inertial Measurement Units (IMUs) continuously streamed joint angle data to the visual feedback GUI in Unity.

Figure 2

Information flow of the BC-NFT software. The visual feedback GUI transmitted trigger values associated with various task events (provided as visual cues to the participant) and ankle angle data to LSL. EEG was directly streamed into the Lab Streaming Layer (LSL) in real-time. LSL played a crucial role in synchronization and timestamping all data streams, i.e., EEG, ankle angle, and triggers. Subsequently, it transmitted the EEG and trigger streams to the BCI system in MATLAB.

Figure 3

EEG preprocessing and LSTM training workflow (A) and LSTM structure (B).

Figure 4

Peak negativity (PN) error was computed as the temporal difference between the PN of each trial (red dashed line), representing the desired stimulation time, and the instant at which the model activated the NMES (blue dashed line). The shaded blue area around blue dashed line represents the potential variability in detecting each trial’s PN.

Figure 5

Training regimen for the participant with CP. We conducted two assessment sessions, pre- and post-training, and 10 training sessions.

Figure 6

Event-related spectral perturbation (ERSP) plots of the motor regions pre and post training (top and middle row, respectively) and their significance masked differences (bottom row). All rows share a common time and frequency axis, with time zero corresponding to dorsiflexion onset of the trained leg (vertical line). The top two rows (pre and post) share a common colorbar (change in power in decibels), with blue representing a decrease in power relative to that session’s baseline and red representing an increase in power. For the bottom row, the significance masked ERSPs (green means a voxel was not significant at an alpha value of 0.05) are scaled symmetrically from −4 db to 4 db*, with the exception of the central motor region, which is scaled from −8 to 8 db, to adequately represent the larger relative changes seen in this region.

Figure 7

Transfer learning. The LSTM model was trained with trials collected during the previous session (100 trials) and then transferred to the current session using the first 25 trials. The weightings of the first two layers of this LSTM model were held constant, or “frozen,” then a new bilateral LSTM layer was inserted to model. Due to the frozen state of the first two layers only the weights of the new bilateral LSTM and the fully connected layer were updated during the transfer learning.

Figure 8

The session-by-session accuracy of the transferred LSTM, trained on 100 trials from the previous session and transferred to 25 trials of the current session (red circles), compared with the accuracy of the LSTM model if it was not transferred and instead was trained on the first 50 trials of each session (blue squares). Here we only included the sessions that used transferred LSTM as the detection model.