Source-linked AI summary
Brain-Computer Interface Controlled Robotic Gait Orthosis
An H. Do, Po T. Wang, Christine E. King, Sophia N. Chun, Zoran Nenadic
TL;DR
The study addresses the lack of biomedical solutions for restoring ambulation after spinal cord injury by testing EEG-based BCI control of a robotic gait orthosis. Two subjects controlled treadmill-suspended walking, including a participant with paraplegia, with high online correspondence and no omissions, supporting the feasibility of brain-controlled ambulation.
Problem
No biomedical solution reverses lost neurological function after SCI, while wheelchair reliance is associated with co-morbidities and substantial medical-care costs.
Method
Two subjects used EEG-recorded idling and walking kinesthetic motor imagery to generate an online BCI prediction model controlling a commercial robotic gait orthosis during cue-prompted walking.
Results
0.812±0.048 average cross-correlation between instructional cues and BCI-RoGO walking epochs was significant at p-value < 10^-4, with 0 omissions and 0.8 false alarms per session.
Takeaways & Limitations
The results provide evidence that BCI control of ambulation after SCI is possible and warrant future studies testing this system in individuals with SCI.
Abstract
from arXiv · showhide
Reliance on wheelchairs after spinal cord injury (SCI) leads to many medical co-morbidities. Treatment of these conditions contributes to the majority of SCI health care costs. Restoring able-body-like ambulation after SCI may reduce the incidence of these conditions, and increase independence and quality of life. However, no biomedical solution exists that can reverse this lost neurological function, and hence novel methods are needed. Brain-computer interface (BCI) controlled lower extremity prosthesis may constitute one such novel approach. One subject with able-body and one with paraplegia due to SCI underwent electroencephalogram (EEG) recording while engaged in alternating epochs of idling and walking kinesthetic motor imagery (KMI). These data were analyzed to generate an EEG prediction model for online BCI operation. A commercial robotic gait orthosis (RoGO) system (treadmill suspended), was interfaced with the BCI computer. In an online test, the subjects were tasked to ambulate using the BCI-RoGO system when prompted by computerized cues. The performance of this system was assessed with cross-correlation analysis, and omission and false alarm rates. The offline accuracy of the EEG prediction model averaged 86.3%. The cross-correlation between instructional cues and BCI-RoGO walking epochs averaged 0.812 +/- 0.048 (p-value<10^-4). There were on average 0.8 false alarms per session and no omissions. This is the first time a person with parapegia due to SCI regained basic brain-controlled ambulation, thereby indicating that restoring brain-controlled ambulation is feasible. Future work will test this system in a population of individuals with SCI. If successful, this may justify future development of invasive BCI-controlled lower extremity prostheses. This system may also be applied to incomplete SCI to improve neurological outcomes beyond those of standard physiotherapy.
I. INTRODUCTION
SCI-related loss of walking ability and wheelchair dependence create substantial health burdens, while existing prostheses and biomedical solutions remain limited. The study therefore investigates noninvasive EEG-based BCI control of a robotic gait orthosis as a step toward brain-controlled ambulation.
- SCI-related loss of ambulation and prolonged wheelchair use are associated with metabolic, cardiovascular, skeletal, and skin co-morbidities.Treatment of these co-morbidities contributes substantially to SCI medical care costs.
- 6?
- Commercial lower-extremity prostheses can restore basic ambulation through manual control, but adoption remains limited by cost, bulk, and energy inefficiency.
- A BCI-controlled lower-extremity prosthesis is proposed as a novel approach to restoring able-bodied-like ambulation after SCI.
- Because invasive brain-signal and FES systems raise safety concerns, feasibility must first be established with noninvasive systems.
- This study extends prior EEG-BCI avatar control to a physical RoGO and reports successful operation by able-bodied and SCI subjects.
II. METHODS
The study developed and tested a noninvasive EEG-based BCI connected to a treadmill-suspended robotic gait orthosis. Subjects generated walking-related EEG activity through kinesthetic motor imagery and used the system during cue-driven online tests.
- EEG recordings during alternating idling and walking kinesthetic motor imagery were used to generate an online BCI prediction model.
- A commercial treadmill-suspended RoGO was interfaced with the BCI computer for computerized control.
- Five 5-minute online tests required subjects to ambulate with the BCI-RoGO system when prompted by computerized cues.
- System performance was evaluated using cue-response cross-correlation, latency, omission rate, and false alarm rate.
- Participants included able-bodied individuals and people with chronic, complete motor paraplegia due to SCI, subject to stated exclusion criteria.
- Partial weight unloading in the RoGO provided safe and easy testing conditions for early BCI-prosthesis development.
B. Electromyogram and Leg Movement Measurement
The experiment combined EEG-based BCI operation with physiological measurements during robotic walking. EMG and leg-motion recordings were used to distinguish BCI control from voluntary movement and characterize the setup.
- Surface EMG was measured to rule out voluntary leg movement as the source of BCI control in able-bodied subjects.
- Baseline EMG was collected during active, cooperative, and passive walking conditions.
- Three electrode pairs recorded activity from the left quadriceps, tibialis anterior, and gastrocnemius.
- The setup suspended the subject in the RoGO with an EEG cap, surface EMG electrodes, and a left-leg gyroscope, while a monitor presented instructional cues.
C. Offline Analysis
The offline pipeline transformed EEG power spectra into discriminative features and classified idling versus walking states. Cross-validation selected model parameters and the resulting prediction model was saved for real-time BCI-RoGO operation.
- Artifact-pruned EEG epochs were transformed into frequency-domain power spectral densities integrated over 2-Hz bins.
- Classwise PCA and approximate information discriminant analysis reduced the spectral data to a one-dimensional feature for classification.
- The feature mapped spatio-spectral EEG data through a CPCA subspace and an AIDA transformation matrix.
- A linear Bayesian classifier assigned each feature to idling or walking according to their posterior probabilities.
- Stratified 10-fold cross-validation trained on nine folds and tested on the remaining fold, repeating across all folds.
- The optimal frequency range and model parameters were selected by repeated performance evaluation and saved for real-time analysis.
D. BCI-RoGO Integration
The RoGO computer was connected to the BCI through two microcontrollers that emulated mouse hardware, complying with software-installation restrictions.
- D. BCI-RoGO Integration: Two Arduino microcontrollers interfaced the BCI computer with the RoGO system through mouse hardware emulation.The first microcontroller relayed BCI commands to the second over an I2C connection; the second ran mouse-emulation firmware.
E. Online Signal Analysis
Online BCI signal analysis used overlapping EEG windows, power spectral density features, and Bayesian posterior probabilities to classify idling and walking.
- E. Online Signal Analysis: 0.75-sec EEG segments were acquired every 0.25 sec using a sliding overlapping window.Power spectral density from retained EEG channels was calculated for each segment.
- E. Online Signal Analysis: Bayesian posterior probabilities of idling and walking classes were computed from the EEG-derived inputs.These probabilities were used by the online signal-processing algorithms.
F. Calibration
The BCI-RoGO was modeled as a binary idling/walking state machine, with averaged posterior probabilities and calibrated thresholds governing transitions.
- F. Calibration: A binary state machine represented the BCI-RoGO system with “idling” and “walking” states.This reduced online noise and minimized the subject’s mental workload.
- F. Calibration: Posterior probabilities averaged over 2 sec of EEG were compared with idling and walking thresholds to initiate state transitions.The thresholds were denoted TI and TW in the state-transition procedure.
- F. Calibration: Thresholds were empirically determined from approximately 5 min of alternating idling and walking KMI during online-mode calibration.A brief feedback-based familiarization session further fine-tuned the threshold values.
G. Online Evaluation
Online evaluation tested whether idling and walking kinesthetic motor imagery could control alternating RoGO epochs in response to textual cues, using correlation, omission, and false-alarm metrics.
- G. Online Evaluation: Subjects completed five alternating 1-min epochs of BCI-RoGO idling and walking while following static textual computer cues.They were instructed to avoid voluntary movements and keep their arms still.
- G. Online Evaluation: Online performance was assessed by cross-correlation between cues and RoGO walking, omissions, and false alarms.Omissions represented missed walking activation during “Walk” cues, while false alarms represented walking initiation during “Idle” cues.
- G. Online Evaluation: For able-bodied subjects, EMG and leg-movement analyses assessed whether RoGO walking was entirely BCI controlled.The evaluation included gyroscope and rectified EMG data to examine covert movement involvement.
H. Controls
The study evaluates online BCI-RoGO performance with a nonlinear autoregressive model and Monte Carlo testing. The analysis uses simulated posterior sequences under a wide-sense stationarity assumption to estimate empirical significance.
- A nonlinear auto-regressive model was created to determine the significance of each online BCI-RoGO session’s performance.
- X_k is modeled as the state variable, W_k as uniform white noise, and Y_k as a saturated simulated posterior probability constrained to [0, 1].
- The posterior-probability sequence {P_k} is assumed wide-sense stationary, with mean μ and variance σ^2 used to determine α and β.
- Table I reports demographic data for the study subjects and defines ASIA as the American Spinal Injury Association.
- 10,000 Monte Carlo trials per online session generated simulated posterior sequences whose cue–state cross-correlations were compared with the observed session to define an empirical p-value.
III. RESULTS
The BCI-RoGO system distinguished idling from walking and produced cue-aligned walking without omissions in five online sessions. Results also indicated that able-bodied Subject 1’s leg activity during BCI-RoGO walking resembled passive rather than active or cooperative walking.
- EEG prediction: 94.8±0.8% and 77.8±2.0% were the offline classification accuracies for Subjects 1 and 2, respectively, against 50% chance.Subject 1 was able-bodied and Subject 2 had paraplegia due to SCI.
- Online performance: 0.812±0.048 was the average cross-correlation between instructional cues and BCI-RoGO walking epochs, with empirical p-value < 10^-4.Monte Carlo maximum cross-correlations were 0.438 and 0.498 for Subjects 1 and 2, respectively.
- Online performance: There were no omissions, while false alarms averaged 0.8 across all sessions and both subjects.Each subject achieved two sessions with no false alarms.
- EEG features: Brain areas most salient for distinguishing idling and walking occurred in the 8-10 Hz bin for Subject 1 and the 10-12 Hz bin for Subject 2.Feature maps contained one map for each of the two classes, with values near +1 or -1 indicating the most salient features.
- Physiological validation: For Subject 1, EMG during BCI-RoGO walking differed from active or cooperative walking but did not differ from passive walking.The corresponding comparisons were p < 10^-13 and p = 0.37, respectively.
- Online session traces: Figure 4 aligned decoded BCI states with instructional cues and gyroscope-defined idling or walking epochs, while EMG traces were shown for Subject 1 only.Walking and idling were represented by thick and thin blocks, respectively.
IV. DISCUSSION AND CONCLUSION
The study demonstrates feasible, accurate BCI control of a treadmill-suspended robotic gait orthosis, including purposeful ambulation by a person with paraplegia after minimal training. The findings support further testing in larger SCI populations and development of more capable BCI-controlled prostheses.
- Both subjects gained purposeful, highly accurate BCI-RoGO control on their first attempt.Online performance generally improved across five sessions.
- The participant with paraplegia achieved brain-driven basic ambulation after only 10 minutes of training-data acquisition and minimal prior BCI experience.The authors describe this as the first reported demonstration of a person with paraplegia due to SCI regaining brain-driven basic ambulation during a goal-oriented walking task.
- 0.812 average online cross-correlation was achieved between computer cues and BCI-RoGO responses.This exceeded reported correlations for lower- and upper-extremity BCI prostheses despite EEG acquisition during ambulatory conditions.
- All online sessions had a 100% response rate with no omissions, although false alarms remained a safety concern for future overground systems.Reducing the probability-averaging window could reduce delay but may increase false alarms and omissions.
- Both subjects attained high-level control after short procedures totaling limited training, calibration, and familiarization time.The authors suggest data-driven EEG modeling and prior virtual-reality training may have facilitated rapid control acquisition.
- Future studies should test the system in larger SCI populations and address turning, velocity modulation, posture transitions, and signal-acquisition solutions.The current system may also support gait rehabilitation in incomplete motor SCI, although these applications remain hypotheses for future work.