Design and Optimization of an EEG-Based Brain Machine Interface (BMI) to an Upper-Limb Exoskeleton for Stroke Survivors

Abstract

This study demonstrates the feasibility of detecting motor intent from brain activity of chronic stroke patients using an asynchronous electroencephalography (EEG)-based brain machine interface (BMI). Intent was inferred from movement related cortical potentials (MRCPs) measured over an optimized set of EEG electrodes. Successful intent detection triggered the motion of an upper-limb exoskeleton (MAHI Exo-II), to guide movement and to encourage active user participation by providing instantaneous sensory feedback. Several BMI design features were optimized to increase system performance in the presence of single-trial variability of MRCPs in the injured brain: (1) an adaptive time window was used for extracting features during BMI calibration; (2) training data from two consecutive days were pooled for BMI calibration to increase robustness to handle the day-to-day variations typical of EEG, and (3) BMI predictions were gated by residual electromyography (EMG) activity from the impaired arm, to reduce the number of false positives. This patient-specific BMI calibration approach can accommodate a broad spectrum of stroke patients with diverse motor capabilities. Following BMI optimization on day 3, testing of the closed-loop BMI-MAHI exoskeleton, on 4th and 5th days of the study, showed consistent BMI performance with overall mean true positive rate (TPR) = 62.7 ± 21.4% on day 4 and 67.1 ± 14.6% on day 5. The overall false positive rate (FPR) across subjects was 27.74 ± 37.46% on day 4 and 27.5 ± 35.64% on day 5; however for two subjects who had residual motor function and could benefit from the EMG-gated BMI, the mean FPR was quite low (< 10%). On average, motor intent was detected -367 ± 328 ms before movement onset during closed-loop operation. These findings provide evidence that closed-loop EEG-based BMI for stroke patients can be designed and optimized to perform well across multiple days without system recalibration.

Keywords: brain machine interface (BMI); motor intent detection; movement related cortical potentials (MRCPs); robotic exoskeleton; stroke rehabilitation.

Figures

Figure 1

(A) Schematic of the asynchronous EEG-based BMI. Motor intents detected from EEG activity of a chronic upper-limb impaired stroke patient, are gated by EMG from impaired hand to trigger robot guided movement using MAHI Exo-II (currently elbow flexion/extension only). (B) Timeline for each trial during closed-loop BMI control task. Exoskeleton's motion (forearm extension) and corresponding visual feedback displayed on the graphical user interface (GUI) at significant events on the timeline are shown. Exoskeleton or home (H) position and target (T) position are also indicated. (C) Raster plot displays time-series of selected EEG (MRCP) channels and their spatial average used to detect motor intent, EMG from biceps and triceps muscles, and exoskeleton's kinematics (elbow position, velocity), during closed-loop BMI control. Alternating attempted trials (shaded) from target-onset to target-reached and fixation intervals are shown. Markers indicate BMI predictions (unfilled triangles) and successful intent detection with EMG-gated BMI (filled triangles). Trial #2 shows examples of spurious BMI-only intents i.e., false positives that were successfully rejected by EMG-gate and a missed subject attempt i.e., false negative (marked by arrow) which the BMI failed to detect. Note also the incorrect BMI-predicted motor intent during fixation interval preceding Trial #2, which was rejected by EMG-gate.

Figure 2

Data collection procedure. The parenthesis, next to the user-driven (UD) and user-triggered (UT) training modes, represent the number of blocks of 20 trials that were completed on each day.

Figure 3

Flowchart for offline EEG processing and classifier design. A binary Support Vector Machine (SVM) classifier with Radial Basis Function (RBF) kernel was trained and evaluated using a simulated real-time cross-validation scheme that generated classifier prediction on test samples using a 50 ms sliding window. The classifier and optimal window length (wlO) that obtained in maximum area under the receiver operating characteristics (ROC) curve, were later used for closed-loop BMI implementation (see Section 2.4).

Figure 4

Movement related cortical potentials (MRCPs) observed for subject S4 in user-driven (A,C,E) and user-triggered (B,D,F) modes. (A,B) show a subset of EEG channels over the fronto-central, central and centro-parietal lobes, which were investigated for presence of MRCPs. Shaded gray circles represent channels for which MRCPs were observed from grand-averages in (C,D) and red circles highlight channels that were subsequently used for training the motor intent classifier. Shaded blue and black circles represent reference and ground electrodes respectively, which were attached to the subject's ears. (C,D) show baseline corrected grand-averages ± 95% confidence intervals using thick and thin black lines, respectively. In the figures, the peak of MRCP is lagging (~0.5 s) the time of movement-onset (MO) due to the non-linear phase distortion of IIR filters. (E,F) display raster plots of single-trial EEG amplitudes, without baseline correction, for channels used to train the classifier (columns 1–4) and their spatial average (column 5). The trials were sorted in increasing order of latency, which is defined as the time interval starting from 0.5 s up to the negative peak of spatial average. In column 5, trials for which the peak negativity of spatial average occurred earlier than −1.5 s (vertical black line) with respect to movement-onset, were rejected when training the classifier since these trials are most likely corrupted by artifacts.

Figure 5

Approach for deciding optimal window length (wlO) and calibrating motor intent classifier for a sample dataset (S4, user-triggered mode). (A) shows examples of single-trial spatial averaged MRCP epochs in gray superimposed with blue and red regions which represent fixed and adaptive windows defined for extracting classification features. The fixed window is predefined with respect to movement-onset (MO), whereas the adaptive window is defined for each trial with respect to negative MRCP peak. Further, for adaptive window, trials when the MRCP peaked earlier than −1.5 s were rejected from the training set (marked by X, otherwise by ✓). The duration of fixed and adaptive windows shown in (A), correspond to wlO marked in (D) by blue and red “o”, respectively. (B) Scatter plots showing the distribution of features extracted using optimal fixed and adaptive windows from Go and No-go trials. (C) Receiver operating characteristic (ROC) curves indicating classifier's performance in terms of True Positive Rate (TPR) and False Positive Rate (FPR) for different window lengths shown in (D). Thin blue and red lines demonstrate performance curves for different fixed and adaptive window lengths, whereas bold lines indicate optimal performance curves that were obtained for each windowing technique. Random chance performance is shown by dotted black line. (D) shows the criteria for selecting wlO based on maximum area under the ROC curve achieved. (E) Boxplots showing offline cross-validation accuracy for fixed and adaptive windows for all subjects and calibration modes. Statistically significant differences determined using Wilcoxon Rank Sum test are marked. The dotted black line represent the chance level accuracy, averaged across all subjects and conditions. UD, user-driven mode; UT, user-triggered mode.

Figure 6

Box plots show true positive rate (TPR) and false positive rate (FPR) for closed-loop BMI performance on days 4 and 5. TPR and FPR were calculated on attempted trials (15–19 trials/block, overall 1063 trials) and catch trials (1–5 trials/block, overall 157 trials), respectively. The median values for BMI performance are shown in red. In each sub-plot, the last column shows the overall BMI performance achieved across all subjects. Subjects S1 and S3 using user-triggered (UT) mode on both days are grouped together. Similarly, S2 and S4 using user-driven (UD) on day 4 and user-triggered (UT) on day 5 are grouped together.

Figure 7

Number of motor intents detected per min (or Intents per min) and its coefficient of variation (CoV) are computed for each block of 20 trials that tested the closed-loop BMI control and are shown here for days 4 and 5 by the left and right columns, respectively. The subject names and calibration modes are shown in the top left corner for each row. Each row consists of a box plot displaying number of intents per min and a plot showing their CoV for each block. Additionally, within each plot the overall intents per min and CoV that were computed by combining performance of all blocks for that day is shown. Outliers are represented by “°”; however few outliers outside the axes range are not shown. The dotted lines show statistically significant trends in the median values for intents per min and individual values of CoV across all blocks for that day, along with their slopes. UD, user-driven mode; UT, user-triggered mode.