An EEG-based Brain Computer Interface for Rehabilitation and Restoration of Hand Control following Stroke Using Ipsilateral Cortical Physiology

Transcription

1 Washington University in St. Louis School of Engineering and Applied Science Electrical and Systems Engineering Department ESE 498 An EEG-based Brain Computer Interface for Rehabilitation and Restoration of Hand Control following Stroke Using Ipsilateral Cortical Physiology By Sam B. Fok, Raphael Schwartz, Charles D. Holmes Supervisor Eric Leuthardt, David Bundy, Robert Morley Submitted in Partial Fulfillment of the Requirement for the BSEE Degree, Electrical and Systems Engineering Department, School of Engineering and Applied Science, Washington University in St. Louis

2 Student Statement The authors of this report have observed and upheld all codes and ethics, including the University s Honor System, and have maintained the integrity of this course in the design and implementation of this project. 2

3 Abstract Stroke and traumatic brain injury (TBI) cause long-term, unilateral loss of motor control due to brain damage on the opposing (contralateral) side of the body. Conventional neurological therapies have been found ineffective in rehabilitating upper-limb function after stroke. Brain computer interfaces (BCIs), devices that tap directly into brain signals, show promise in providing rehabilitation but remain in research. Also, BCIs cannot work if the target signals have been eliminated due to injury. Therefore we present a novel BCI, the IpsiHand, which combines advances in neurophysiology, electronics, and rehabilitation. Recent studies show that during hand movement, the cortical hemisphere on the same (ipsilateral) side of the body as the hand also activates. IpsiHand uses electroencephalography (EEG) to record these signals and control a powered hand orthosis. The undamaged hemisphere can then control both hands, and through neural plasticity IpsiHand will strengthen ipsilateral neural pathways to enhance ipsilateral motor control. Acknowledgements We would like to especially thank Mark Wronkiewicz, Jessica Zhang, Thane Somers, and Nathan Brodell who were a part of our design team. We would like to thank Dr. Eric Leuthardt, our faculty mentor, and David Bundy, our graduate student mentor, for their guidance. We would also like to thank Joanne Rasch and the Rehabilitation Institute of St. Louis for providing consumer feedback and expert opinion on current rehabilitation, and Professors Robert Morley and Joseph Klaesner for instruction during senior design. This work is supported in part by The National Collegiate Inventors and Innovators Alliance, the Washington University School of Engineering, and Emotiv Systems. 3

4 Table of Contents Student Statement... 2 Abstract... 3 Acknowledgements... 3 List of Figures... 5 List of Tables... 5 Problem Formulation... 6 Problem Statement... 6 Problem Formulation... 6 Project Specifications... 7 Signal Acquisition... 7 Signal Processing... 7 Mechanical Output... 7 System Operation... 7 Concept Synthesis... 8 Literature Review... 8 Concept Generation... 9 Concept Reduction Detailed Engineering Analysis and Design Presentation Signal Acquisition Screening Procedure and Signal Processing Control Signal Conditioning and Output Least Mean Squares Adaptive Filter Performance Analysis Computational Analysis Mechanical Orthosis Cost Analysis Lifetime Operation Cost Hazards and Failure Analysis Conclusion

5 Works Cited List of Figures Figure 1 Differences in power spectrum during A) contralateral movements and B) ipsilateral movements during movement and rest. Note that contralateral movement shows a decrease in power at frequencies less than 35Hz and an increase in power between 60 and 120Hz whereas ipsilateral movements show a decrease in power only at frequencies below 50Hz [9]... 8 Figure 2 Concept of a BCI using ipsilateral signals to restore motor function in hemiplegic stroke survivors [10]... 9 Figure 3 Orthosis design options Figure 4 - Data acquisition Design Options Figure 5 - Structure and data flow of IpsiHand. Processing occurs in three steps: signal acquisition, signal processing, and mechanical actuation Figure 6 - Spatial location of electrodes of Emotiv EPOC headset relative to head Figure 7 - between left hand movemnet and rest across bins of electrode channel and frequency Figure 8 - Frequency spectrum of the F3 EEG electrode location during conditions of left hand movement (red curve) and conditions of rest (blue curve) Figure 9 - Topographical color maps of values for 12 Hz and 22 Hz Figure 10 - ROC curves shows classification performance using varying window lengths and thresholds Figure 11 prefabricated hand orthotic fitted with linear actuator for control of finger position List of Tables Table 1 - Classification accuracy using best threshold as determined by ROC analysis per window length Table 2 - Bill of Materials

6 Problem Formulation Problem Statement In the United States, stroke, traumatic brain injury (TBI), and spinal cord injury (SCI) are the leading causes of disability affecting over a million individuals annually [1] [2] [3]. About 900,000 individuals have reported severe trouble with hand function [4], and conventional physical therapy produces little significant improvement after 3 months post injury [5]. Loss of hand function causes a severe decrease in quality of life for affected individuals [6]. In addition, lasting disabilities result in a typical lifetime cost between $100k and $2M per patient, including inpatient care, rehabilitation, and follow-up [1] [7] [8]. The most effective therapies have patients actively controlling their limb, which is not an option in cases of severe paralysis. While BCIs promise new hope for treatment, they remain in the research stage. In addition, conventional BCIs cannot be applied to cases of brain injury since the classical motor signals in cortex contralateral to the target limb needed would be gone with the injury. Problem Formulation We addressed the problem of applying BCI technology to rehabilitation following stroke and TBI. To do this we developed a device for rehabilitation that synthesizes recent developments in neurophysiology, electronics, and physical therapy into a BCI hand orthosis. A recent study found signals associated with hand movements in cortex ipsilateral to the hand. These signals were present in cortex anterior to ipsilateral primary motor cortex and at frequencies below 40Hz [9], which are accessible via EEG. With this knowledge, these intent-to-move signals can be recorded from the cortex to control an orthosis which opens and closes a patient s hand. Tactile and proprioceptive feedback provided from this device will facilitate neural plasticity, strengthening existing and developing new neural pathways ipsilateral to the affected hand that will ultimately restore motor control. Allowing the patient to regain hand control with their thoughts alone should also provide tremendous encouragement in the rehabilitation process. Our objective is to directly recouple the intent-to-move a hand with hand motion in order to improve outcomes of recovery, reduce the lifetime cost of brain injury, and improve quality of life for those affected by stroke or TBI. 6

7 Project Specifications Our design will provide the following: Signal Acquisition EEG signal recordings at locations corresponding to the International EEG standard, specifically with locations over the motor and premotor cortex. Spectral information up to 60 Hz. Signal Processing Identification of signal features correlated with intent to open or close the hand. Extraction of the control signal from the identified electrode location and frequency band. Spatial filtering to reduce noise common to the electrode of interest and reference electrode(s). Normalization of a control signal from identified features. Communication with a linear actuator for mechanical output. An adaptive filter which adjusts the gain of the mechanical output to counteract changes in the strength of the control signal. Mechanical Output Enough force to overcome spasticity and tone imbalance in a patient s hand (up to 110 N). Graded extension and flexion of the hand. System Operation Real time operation, with latency less than 3 seconds. Signal classification with an accuracy of at least 70%. 7

8 Concept Synthesis Literature Review Previous study by Wineski, et. al., discovered differences in cortical signals during contralateral and ipsilateral movement conditions (Figure 1). Note that this study was conducted with the electrocorticography (ECoG) technique, in which electrodes are placed directly onto the surface of the brain. The spatial and temporal resolution afforded by ECoG will be unavailable with the EEG technique, which is generally limited to less than 60Hz and greater than 2cm temporal and spatial resolution, respectively [10]. Figure 2 from the same lab illustrates the concept of using ipsilateral signals in a BCI application to restore limb control. Understanding of this literature was supported by courses discussing signal processing at Washington University, such as ESE 351, Signals and Systems, ESE 482, Digital Signal Processing, and ESE 488, Signals and Systems Laboratory. Figure 1 Differences in power spectrum during A) contralateral movements and B) ipsilateral movements during movement and rest. Note that contralateral movement shows a decrease in power at frequencies less than 35Hz and an increase in power between 60 and 120Hz whereas ipsilateral movements show a decrease in power only at frequencies below 50Hz [9]. 8

9 Figure 2 Concept of a BCI using ipsilateral signals to restore motor function in hemiplegic stroke survivors [10]. Device-based therapies for rehabilitation exist for both BCI and non-bci techniques. Patent application 11/857,881 describes a Nuero-robotic System developed by Myomo, which uses weak EMG signals of a partially paralyzed stroke patient to coordinate elbow movement with a motor-driven orthosis [11]. Patent application 12/758,706 describes the system for neuromuscular reeducation developed for Hand Mentor [12]. In using EMG rather than EEG, both Myomo and Hand Mentor can only be used by patients with significant residual motor control. From BrainGate, patent application 9/991,498 describes a system that uses neurological signals to control a device, and application 11/201,283 describes associated biological interface systems and methods [13] [14]. We differentiate our device from BrainGate by using EEG rather than invasive single/multiunit recordings and by focusing on intent-to-move signals in the undamaged ipsilateral cortex. Concept Generation There are three major physical components that were considered in the design of Ipsihand: 1) the mechanical orthosis, 2) the data acquisition, and 3) the linear actuator. There is a fourth component which is the signal processor, but for this prototype there was nothing considered other than a laptop. However, it should be noted that there are considerations for the future signal processing platforms, such as microcomputers and microprocessors. For the orthosis itself, there were many options to choose from. Options for the orthosis are illustrated in Figure 3. An initially attractive option was to purchase and modify an existing 9

10 device. There exist many orthotic splints that have capabilities of joint movement or aide of joint movement. Particularly, we looked at wrist-driven orthotics. With these devices, it would be very easy to uncouple the wrist to the hand movement and replace it with a linear actuator. Other options that came about included already functional actuating orthotics, including devices implementing functional electrical stimulation. Functional electrical stimulation involves applying electrical currents to activate nerves and induce muscle contraction. Devices like this were considered primarily because they would involve less direct connections to the hand and not require the addition of a linear actuator. Orthosis Modify Existing Device Use Existing Device Wrist-Driven Orthosis Dynamic Splint Continuous Passive Motion Machine Haptic Feedback Build from Scratch Functional Electrical Stimulation Figure 3 Orthosis design options. For the data acquisiton, there were many commerical options for EEG headsets and headcaps. Single electrode headsets, such as the Neurosky Mindwave, were inexpensive and simple. Most had an accesible SDK so development would be relativaley simple. However, a downside to the simple single elctrode set up is the lack of channels. For this reason, more expensive, multiple-electrode headsets were also considers. Their advantages included the ability to do spatial filtering and accesibilty to more than one location of brain. Figure 4 shows an illustration of all of the options for data acquistion devices. 10

11 Data Acquistion Single Electrode EEG Headset Multiple Electrode EEG Headset Neurosky Mindwave Star Wars Force Trainer MindSet Standard Medical Headcap Emotiv EPOC Figure 4 - Data acquisition Design Options Finally, for the linear actuator, there were a multitude of options online for purchase. All devices performed essentially the same. The big differences from option to option were the size, power, and cost of the actuator. Concept Reduction In the choice of the orthotic itself, price became a big decision factor. Devices that already hand motor driven actuation cost well over $1500. In addition, overriding the control of the actuation would be more complex than programming a separate linear actuator. For these reasons, the wrist-driven orthosis was chosen. This orthosis, the Becker Oregon Talon TM, would eventually have the bar coupling finger movement to wrist movement removed. This bar would be replaced by the linear actuator, the Firgelli L16, which would be fastened to the back end of the orthotic. The L16 was chosen primarily because of its ease of use and its value. Finally, the Emotiv EPOC was chosen as the data acquisition device. This was for a number of reasons. The device was a low cost consumer headset, yet it had multiple electrodes, so there was 11

12 more capability for feature extraction. Additionally, the device came with an easy to use SDK which had been adapted to work with BCI2000 in a BCI2000 module. Detailed Engineering Analysis and Design Presentation Electrical fields produced by synchronous firings of neurons over large areas of the brain are recorded on the scalp. These signals are then sent to a laptop for processing. The signal is filtered to generate an output signal, which is then sent to a linear actuator. An orthosis controls the patient s finger closure with this linear actuator, so that brainwaves can control the opening and closing of the hand using a signal route that bypasses damage nerves. Figure 5 - Structure and data flow of IpsiHand. Processing occurs in three steps: signal acquisition, signal processing, and mechanical actuation. Signal Acquisition EEG signals were acquired from the subject s scalp using a dry-electrode Emotiv EPOC (Emotiv; Australia) headset with 14 electrodes located over international system positions AF3, F7, F3, FC5, T7, P7, O1, O2, P8, T8, FC6, F4, F8, and AF4 using 2 reference electrodes (Figure 6). The headset aligns, bandpass filters, and digitizes the signal at 128 Hz and transmits wirelessly to a laptop. 12

13 Figure 6 - Spatial location of electrodes of Emotiv EPOC headset relative to head Screening Procedure and Signal Processing On the laptop, signal processing was carried out in the BCI2000 framework, a development platform written primarily in C++ and MATLAB that allows rapid recording, filtering, and feature selection of brain signals [15]. Initial screening in which signals are recorded while the subject makes hand movements is used to determine the EEG features that our algorithm will use to contrast movement from rest. During this screening procedure the user alternates between periods of attempted hand movement and periods of rest. After several trials, we identify the specific electrode channels and frequency bins which consistently changed in power between hand movement and rest conditions A bipolar reference scheme, in which a signal is generated by taking the difference between the signals recorded at pairs of adjacent electrodes [16], at 6-9cm distance is used to attenuate noise and artifacts on the scalp common to pairwise electrodes. It is also used to detecting signals from a region consistent in size with the diffuse spatial characteristics of EEG signals [16]. This reference technique also makes the system resilient to variation in electrode placement, which happens simply because it is impossible to place the headset in the exact same spot for each experiment. Five electrodes are available above the frontal lobe and premotor area of the contralesional hemisphere, the location of our signal of interest. Using these 5 electrodes, all combinations of spatial reference choices can be evaluated. The coefficient of determination is used as a measure of spatial reference performance, where indicates the proportion of variability in the signal attributable to the user s hand movement. The bipolar derivation is defined, 13

14 = (1) where i = 1 5, j=1 5 refer to the electrode locations of interest and and are the voltage measurement signals from the electrode i and j respectively. Therefore, is bipolar derivation of i referenced to j. We evaluate each combination and select for the highest, =argmax( { }) (2) Once the bipolar reference has been selected, the frequency bin of width 2 Hz with the highest is then selected as the control spectrum. The bipolar spatial filtered recording is then converted to actuator control signal by BCI2000 and normalized to have a mean of zero and a unit variance. This normalized control signal is then sent to LabVIEW via a UDP port. Control Signal Conditioning and Output For each cycle, a LabVIEW (National Instruments; Austin, TX) program detected whether the control signal exceeded a threshold. This is determined by: = 0 < (1) where is the control signal from BCI2000, is the threshold of the system and is termed the condition control signal. The system is designed such that a positive signal corresponds to the hand closing and a negative signal corresponds to the hand opening. The signal is condition in this way so that a user can keep the device stationary by maintaining the signal below threshold. This conditioned control signal is then used to adjust the position of the linear actuator, and thus open or close the hand. There are two main approaches to the implementation of this: 1) the control signal reflects position or 2) the control signal reflects velocity. With the first option, the control signal would be multiplied by a gain and the product would be the position the linear actuator would go to. With the second option, the control signal would be multiplied by a gain and the product would be added to the current position. 14

15 The second option is implemented because it does not show radical changes in the position as a result of the stochastic nature of the control signal. This, in turn, makes for smoother movements of the hand. This velocity based approach can be expressed by = + (2) = 0 <0 > (3) where is the extension length of the linear actuator for iteration, is the maximum length of the linear actuator, is the gain, and is the conditioned control signal for iteration. From equation (3) we can see that the there is a limit in both directions to how much the linear actuator can open or close. Though the actuator physically cannot extend past these values, this equation is included in calculations to avoid bugs. The actuator position is then changed by writing to the actuator microcontroller via a USB port. The microcontroller then adjusts the position of the actuator via pulse-width modulation. Least Mean Squares Adaptive Filter The brain is a very mutable object. Patterns that the Ipsihand observes change from trial to trail and from subject to subject. Because of this, magnitudes of the band of interest can change from trial to trial. To counteract this, the Ipsihand implements a Least Mean Squares (LMS) adaptive filter. The LMS algorithm linearly scales the selected bipolar channel by adjusting the gain. The gain is adjusted after each trial to values which would have yielded better results in the previous trial. This is expressed as =(1 ) + (Δ ) (1) 15

16 where is the gain for the k th trial, (Δ ) defines the desired movement for that trial, is the mean value for the input signal in a given trial, and, the adaptability constant, determines the rate of adaption. Given a gain for trial k-1 and data from the trial, the gain for trial k will be a weighted average of gain from trial k-1 and a gain that would have resulted in the least amount of error for trial k-1. The weights are determined by the adaptability constant. If is 0, then there is no change from trial to trial. If is 1, the gain changes without memory of the previous gain values. Performance Analysis Our design was tested with three healthy subjects to verify the ability to use non-conventional signals from the contralesional cortex to control a hand on the same side of the body. We found : 1. Hand movement correlates with EEG signals from the contralesional hemisphere. 2. Our design was successful in using these signals to coordinate a motor-driven orthosis. Results of the screening procedure described above are shown in Figure 7 for a healthy subject. The correlation between the left hand movement condition and rest condition are evaluated using values per electrode channel (y axis) and per frequency bin (x axis). Electrodes over the left hemisphere are on the lower half of the y axis and electrodes over the right hemisphere are on the upper half of the y axis. Channels names are shown according to the EEG system [17]. Bins with high values indicate significant power difference between movement and rest conditions and as such are good candidates for control signals. Clusters (dotted red circles) of high correlation are noted around the 12Hz bins in F3 through P7 and also in channel F3 around the 22Hz bin. 16

17 R 2 Values Between Left Hand Movement and Rest AF4 R 2 F8 0.5 F4 FC6 T8 0.4 P8 O2 0.3 O1 P7 T7 0.2 FC5 F3 0.1 F7 AF Frequency (Hz) Channel Figure 7 - between left hand movemnet and rest across bins of electrode channel and frequency Power of the F3 electrode during conditions of left hand movement and conditions of rest is shown in Figure 8. Movement produces even related desynchronization (ERD) in frequency bins around 12Hz and 22Hz as physiologically expected [9] Frequency Spectrum at Electrode F3 Move Left Hand Rest Amplitude Decreased Amplitude around 12Hz Decreased Amplitude around 22Hz Frequency (Hz) Figure 8 - Frequency spectrum of the F3 EEG electrode location during conditions of left hand movement (red curve) and conditions of rest (blue curve). 17

18 The left hand side of Figure 5 shows a topographical color map of correlation values between left hand movement and rest conditions for 12Hz brainwaves. Note high, bilateral correlations are seen in the frontal cortex electrodes. On the right hand side of Figure 9, a map shows 22Hz brainwaves. High correlations are seen unilaterally in F3 electrode. The strength of correlation at 12Hz allows for differentiation between left hand movement versus rest, and the correlation at 22Hz in electrode F3 allows for distinction of left versus right hand movement. Figure 9 - Topographical color maps of values for 12 Hz and 22 Hz. Using the spectral and spatial features identified from this analysis, we selected control signals to move a cursor on a computer screen. The control signal was modulated when a subject moved, or imagined moving his hand. The modulated signal gave the subject control of 1 dimensional cursor movement, and the subject was tasked with moving the cursor to a target that randomly appeared on either side of a computer screen. Through 10 sets of trials with nonimpaired individuals we were able to achieve an 81.3% success rate for this task. We expect that with optimization of our signal detection algorithms, we can achieve even better performance as it was not uncommon to see success rates upwards of 90%. 18

19 Receiver Operator Characteristic (ROC) curves evaluated the performance of a classification scheme. They provide view of the sensitivity of a classification algorithm as a function of its specificity. After collecting a sufficient amount of data, classification is tested with a number of classification thresholds. The true positive rate (ratio of true positives to total positives) and false positive rate (ratio of false positives to total negatives) are observed. True positive rate is then plotted as a function of false positive rate so that a random guess is viewed as a straight line curve from bottom-left to top-right and perfect classification is a right angle at the top left of the figure. In Figure 10 we plot ROC curves with data from 4 different lengths of data acquisition time windows. The longer windows provide us with more power to classify our signal, but also increase the time buffer and therefore increase the latency of our system. 1 ROC curves 0.9 True positive rate (Sensitivity) Seconds 2.0 Seconds 1.0 Seconds 0.5 Seconds Random Guess False positive rate (1-Specificity) Figure 10 - ROC curves shows classification performance using varying window lengths and thresholds. 19

20 We also looked at the overall classification accuracy (ratio of true classifications to total classifications) at the best threshold for each set of data. If specificity and sensitivity are equally important, the best threshold is the one corresponding to the point on the curve closest to the topleft corner of the figure. Time Window (Seconds) Accuracy of Classification % % % % Table 1 - Classification accuracy using best threshold as determined by ROC analysis per window length In analysis of real-time operation, through 10 sets of trials with non-impaired individuals we achieved an 81.3% success rate for closing and opening the orthosis in response to a prompt, when using a 0.5 second window of acquisition time. Note that this is slightly higher than that predicted by ROC analysis for the 0.5 second time window. Computational Analysis We use an autoregressive technique to compute the Maximum Entropy Method (MEM) of spectral estimation, which uses an all pole linear filter to model the spectrum. The MEM algorithm is common in the area of EEG signal processing compared to the Fast Fourier Transform (FFT) due to its higher spectral resolution and better fitting of sharp spectral features for a time-limited signal [18], [19].The tradeoff is increased computational requirements. The narrow band isolation required to evaluate mu brain waves at Hz make the MEM a desirable choice in this circumstance. Choosing the MEM estimation method we can derive the computational requirements of our system. Our signal is sampled at 128 Hz and uses buffers of data 0.5 seconds in length (as selected by our need for near real-time operation). Our bipolar spatial reference technique requires subtraction of the reference electrode from the electrode of interest at every sample, translating to 64 operations per buffer. Each MEM spectral estimation requires M x N operations, where M is the number of samples and N is the model order being used. Model order refers to 20

21 the number of poles used in the linear filter and is directly related to the number of peaks which can appear across the estimated spectrum. A model order of 16 has been determined to suit EEG applications well [18], so a total of operations results. This gives us a total of 1,088 operations per 0.5 second buffer or 2,176 operations per second. This analysis is helpful for developing requirements towards a future microprocessor choice. Mechanical Orthosis The hand orthotic (Becker Oregon TAL100 Talon) was prefabricated to couple wrist motion to the opening and closing of the hand. This coupling mechanism was replaced with a powered linear actuator (Firgelli Miniature Linear Motion Series L16) as seen in Figure 11, shifting the control of the device to the LabVIEW algorithm. The linear actuator operates according to pulsewidth modulation (PWM) control signals received from a microcontroller. The microcontroller receives control signals from LabVIEW. Figure 11 prefabricated hand orthotic fitted with linear actuator for control of finger position 21

22 Cost Analysis The IpsiHand final design has four components: the hand orthotic (Becker Oregon Talon), the actuator to move and flex the orthotic (Firgelli model L16p), the headset (Emotiv EPOC), and the micro-computer device to decode the brain signals (Gumstix Overo Earth COM). The Talon by Becker Oregon has a retail price of $325, but a company representative stated that orders of 1,000 or more would cost $ per orthotic. We expect to contract with Becker Oregon for bulk manufacturing of the slightly modified version of the Talon hand that incorporates our linear actuator. For the actuator, a representative from Firgelli reported the device costs approximately $40 for wholesale buyers. Bulk costs of the Emotiv Epoc headset remain undisclosed, however, a conservative estimate is $175 based on the retail price. Finally, the Gumstix micro-computer device can be purchased for approximately $149 with bulk pricing. Overall, parts will come to roughly $574, with an estimated $ for assembly, quality assurance, and distribution of our final product ( Table 2). Component Supplier Item (Item No.) Cost for Prototype Hand Orthotic Linear Actuator EEG Headset (pp. 3.3) oc-bci/epoc-neuroheadset/ duct_info.php?products_id=228 duct_info.php?products_id=230 Micro- Computer Assembly, Inspection, Distribution Becker Oregon Firgelli Emotiv Systems Gumstix Talon TM (BO-TAL100-L) Cost for Production (per unit) $325 $210±$10 (Confirmed) L16 (L16-P) $112 $40 (Confirmed) EPOC $500 * $175±$75 (Estimated) Overo TM Water (GUM3503W) and Tobi (PKG30002) $149±$20 (Confirmed) $100±$50 (Estimated) Total $ $ ±$ Table 2 - Bill of Materials * The Emotiv EPOC supplied by Emotiv Systems at no charge. The prototype cost listed is the retail price. 22

23 Lifetime Operation Cost The Ipsihand is very robust. For the most part, if component were to fail, the repair cost would be the price of the component being replaced. For our calculations of lifetime operation cost, we will assume a lifetime of 5 years. For the linear actuator, the lifetime is supposed to be 20,000 strokes. Assuming a bi-weekly therapy at 2 strokes a minute for 1 hour (or 240 strokes per week) the linear actuator would last for about 83 weeks. This is about 1.6 years. So for a single patient, the linear actuator would have to be replaced about 3.2 times, a cost of $125. For the EPOC headset, the headset itself should not need replacing within the five years, however the felt pads can be damaged or lost. The gold conducting contacts can become corroded from the saline solution that makes the felt conductive. 16 replacement felt and conducting contact pairs can be purchased for $49.95 from Emotiv s website. In addition, 160 felts can be bought for $ Hazards and Failure Analysis In the early design of IpsiHand, the linear actuator back end was not set as far back as it is currently. As a result, if the actuator was extended fully, the hand would be hyper-flexed. To correct this, the actuator has been set such that when it is fully extended, the hand is closed fully and not hyper-flexed, and when the actuator is fully retracted, the hand is opened fully and not hyper-extended. It should be noted, however, that people with especially sensitive finger joints should not use this device as it may cause them discomfort when it opens fully or closes fully. This problem could be corrected by programming IpsiHand with changeable limits of extension and flexion such that the hand would stop at those limits instead of moving past them. 23

24 Conclusion Combining the discovery of motor related signals in the undamaged brain hemisphere, electronics, and advances in rehabilitation, our device provides a novel method for rehabilitation for stroke survivors. In testing, we were able to process EEG signals for real-time hand control with accuracy consistent with previous studies [20]. Recent evidence suggests that combining BCIs and orthotic devices induces neural plasticity and improves motor function [9]. Furthermore, the potential for recovery is unhampered by the severity of neural pathway injury since we circumvent the entire injured pathway and uses the brain s plasticity to generate new ones. Signal acquisition, signal processing, and mechanical control methods are established, but synthesizing them with the new technique of ipsilateral cortex recording has yet to be done. Compared to devices such as, Myomo of Neuro-robotic Systems, the Bioness H200 from Ness, and Hand Mentor from Kinetic Muscles Inc., our device facilitates plasticity most directly, is less expensive and more portable, and can be used even in cases of severe damage to neural pathways. Allowing patients to regain hand control with their thoughts will provide tremendous encouragement to continue with a therapy. Combined with IpsiHand s affordability and minimal requirements for therapist supervision, IpsiHand also makes in-home treatment a very practical possibility. Based on therapist discussions, numerous improvements to the design are planned. Currently, a laptop processes the EEG signals to be used for orthosis movement. We plan to miniaturize the processing unit onto a micro-computer to give complete portability and allow patients to go beyond rehabilitation and use the device as a replacement of hand function in daily life. In addition, we would like to expand the system s ability to adapt to spatially nonstationary signals. Implementing adaptive spatial filters or an adaptive classifier that finds the strongest correlated channel automatically and continuously would improve robustness in signal strength for a long-term out-patient orthotic. More importantly, spatial and temporal filters that remove artifacts from eye blinks, EMG, and breathing, are essential to the device performance outside of a research setting. 24

25 Works Cited [1] D Lloyd-Jones and et al., "Heart disease and stroke statistics update: a report from the American Heart Association Statistics Committee and Stroke Statistics Subcommittee," Circulation, pp. e21-181, [2] J F Kraus and D L MacArthur, "Epidemiologic aspects of brain injury," Neurologic Clinics, vol. 14, no. 2, pp , [3] M J DeVivo, "The national spinal cord injury statistical center database," in 25th Annual SCI Model Centers OT/PT Leadership Forum, Birmingham, [4] US Census Bureau. (2005) Prevalence of Disability Among Individuals 15 Years and Older by Specific Measures of Disabilities. [Online]. [5] Henrik S Jorgensen et al., "Outcome and time course of recovery in stroke. Part II: Time course of recovery. The copenhagen stroke study," Archives of Physical Medicine and Rehabilitation, vol. 76, no. 5, pp , May [6] Lorenzo Carraro, Promoting independence following a stroke: a guide for therapists and professionals working in primary health care. Geneva: Disability and Rehabilitation team, WHO; AIFO, [7] National Institute of Neurolgical Disorders and Stroke, "Interagency head injury task force report," National Institute of Neurolgical Disorders and Stroke, [8] CE Levy and et al, "Functional MRI evidence of cortical reorganization in upper-limb stroke hemiplegia treated with constraint-induced movement therapy," Am J Phys Med Rehabil, pp. 4-12, [9] Kimberly J Wineski et al., "Unique cortical physiology associated with ipsilateral hand movements and neuroprosthetic implications," Stroke, pp , [10] E Leuthardt, G Schalk, D Moran, and J Ojemann, "The emerging world of motor neuroprosthetics: a neurosurgical perspective," Neurosurgery, vol. 58, no. 1, pp. 1-14, [11] John M McBean and Kalais N Narendran, "Powered orthotic device and method of using same," 11/857,881, September 19, [12] Edward J Koeneman, James B Koeneman, Donald E Herring, and Robert S Schultz, "System and method for neuromuscular reeducation," 12/758,706, April 12, [13] John P Donoghue, Nicholas G Hatsopoulos, Mijail D Serruya, Matthew R Fellows, and Liam Paninski, "Neurological signal decoding," 09/991,498, November 14, [14] J C Flaherty and John P Donoghue, "Biological interface systems with wireless connection and related methods," 11/201,283, August 11,

26 [15] G Schalk, D McFarland, T Hinterberger, N Bribaumer, and J Wolpaw, "BCI2000: a general-purpose braincomputer interface (BCI) system," IEEE Transactions on Biomedical Engineering, pp , [16] Dennis J McFarland, Lynn M McCane, Stephen V David, and Jonathan R Wolpaw, "Spatial filter selection for EEG-based communication," Electroencephalography and Clinical Neurophysiology, vol. 103, pp , [17] Ernst Niedermeyer and Fernando Lopes da Silva, Electroencephalography: basic principles, clinical applications, and related fields.: Lippincott Williams & Wilkins, [18] William H Press, Brian P Flanney, Saul Amo Teukolsky, and William T Vetterling, Numerical recipes Fortran 77 : the art of scientific computing. New York: Cambridge University Press, [19] User Reference: ARFilter. [Online]. [20] Jonathan R Wolpaw and Dennis J McFarland, "Control of a two-dimensional movement signal by a noninvasive brain-computer interface in humans," Proc. Natl. Acad. Sci, vol. 101, no. 51, pp , December [21] D Broetz and et al, "Combination of brain-computer interface training and goal-directed physical therpy in chronic stroke: A case report," J Neurorehab and Neural Repair, pp ,