IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 22, NO. 3, MAY 2014
671
A Training Method for Locomotion Mode Prediction Using Powered Lower Limb Prostheses Aaron J. Young, Student Member, IEEE, Ann M. Simon, Member, IEEE, and Levi J. Hargrove, Member, IEEE
Abstract—Recently developed lower-limb prostheses are capable of actuating the knee and ankle joints, allowing amputees to perform advanced locomotion modes such as step-over-step stair ascent and walking on sloped surfaces. However, transitions between these locomotion modes and walking are neither automatic nor seamless. This study describes methods for construction and training of a high-level intent recognition system for a lower-limb prosthesis that provides natural transitions between walking, stair ascent, stair descent, ramp ascent, and ramp descent. Using mechanical sensors onboard a powered prosthesis, we collected steady-state and transition data from six transfemoral amputees while the five locomotion modes were performed. An intent recognition system built using only mechanical sensor data was 84.5% accurate using only steady-state training data. Including training data collected while amputees performed seamless transitions between locomotion modes improved the overall accuracy rate to 93.9%. Training using a single analysis window at heel contact and toe off provided higher recognition accuracy than training with multiple analysis windows. This study demonstrates the capability of an intent recognition system to provide automatic, natural, and seamless transitions between five locomotion modes for transfemoral amputees using powered lower limb prostheses. Index Terms—Intent recognition, powered lower limb prosthesis, robotic leg control, transfemoral amputee.
I. INTRODUCTION
T
RANSFEMORAL amputation is a significant cause of disability in the United States with approximately 31 000 new cases occurring each year [1], [2]. Transfemoral amputees who use passive prostheses have significantly impaired balance, walking symmetry, and metabolic energy efficiency [3], [4].
Manuscript received February 05, 2013; revised August 26, 2013; accepted October 03, 2013. Date of publication October 30, 2013; date of current version April 28, 2014. This work was supported in part by the US Army’s Telemedicine and Advanced Technology Research Center under Grant 81XWH-09-2-0020. The work of A. J. Young was supported by a National Defense Science and Engineering Graduate (NDSEG) fellowship. This paper has supplementary downloadable material available at http://ieeexplore.ieee.org, provided by the authors. A. J. Young is with the Center for Bionic Medicine, Rehabilitation Institute of Chicago, Chicago, IL 60611 USA, and also with the Department of Biomedical Engineering, Northwestern University, Chicago, IL 60611 USA (e-mail: [email protected]). A. M. Simon is with the Center for Bionic Medicine, Rehabilitation Institute of Chicago, Chicago, IL 60611 USA, and also with the Department of Physical Medicine and Rehabilitation, Northwestern University, Chicago, IL 60611 USA (e-mail: [email protected]). L. J. Hargrove, is with the Center for Bionic Medicine, Rehabilitation Institute of Chicago, Chicago, IL 60611 USA, and with the Department of Physical Medicine and Rehabilitation, Northwestern University, Chicago, IL 60611 USA (e-mail: [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TNSRE.2013.2285101
Fig. 1. Powered knee/ankle prosthesis during stair ascent.
They also tend to have difficulty navigating more demanding terrain such as ramps and stairs [5]. These issues have motivated the development of advanced microprocessor controlled prostheses. Microprocessor controlled prostheses may be either mechanically passive [6]–[9] or mechanically active [10]–[13]. Both types of devices can restore a variety of locomotion modes and rely on using finite-state machines to interpret data from sensors (e.g., load cells, goniometers, inertial measurement units, etc.) attached to the prosthesis. Determining proper state transitions within a locomotion mode, such as between stance and swing phases of walking, is straightforward and has been implemented in commercially available prostheses [12], [14], [15]. Defining state transitions between locomotion modes, such as from the stance phase of walking to the swing phase of stair ascent, is more difficult. Proposed transition methods include using key-fobs [15], compensatory body movements [9], or intent recognition signal processing algorithms [16], [17]. Intent recognition algorithms have the potential to provide seamless (no stopping or slowing), automatic (no button press or direct user command), and natural (no unintuitive actions) transitions between locomotion modes. Furthermore, they can be used for both mechanically passive and mechanically active devices. In this work, we focus on intent recognition algorithms in mechanically active (motorized knee–ankle) prostheses (Fig. 1) [13]. Several intent recognition strategies have been proposed to perform transitions between activity modes for powered lower limb prostheses. In echo control [18] the prosthetic leg mimics the kinematics of the sound limb. This approach requires minimal system training, but the sound limb must be instrumented
1534-4320 © 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
672
IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 22, NO. 3, MAY 2014
and the user must always lead with the sound limb and take an even number of steps. Pattern recognition strategies have been shown to predict locomotion modes and used to transition a transfemoral prosthesis between level ground walking, sitting, and standing modes [17]. System training involved collecting mechanical sensor data during steady-state conditions as well as during all combinations where the prosthesis was in an incorrect locomotion mode (e.g., walking while the prosthesis was in standing mode). While high recognition rates were achieved, the training requirement makes this approach difficult to implement if more than two locomotion modes are required. Additionally, an intent recognition system that relies on detecting when the prosthesis is in the wrong mode may respond too late to achieve natural and seamless transitions. Intent recognition strategies have been used in mechanically passive devices while amputees performed various locomotion modes. Huang et al. [16] obtained an average intent recognition accuracy of 85.7% for six locomotion modes using only residual limb electromyography (EMG) signals in two transfemoral amputees. This study used only steady-state data to train the classifier; data from transitions between locomotion modes were not included. Huang et al. [19] also achieved recognition accuracies above 95% by using a combination of EMG and load cell information for five steady-state locomotion modes in five transfemoral amputees walking on a passive device. These previous studies demonstrate the potential of using pattern recognition techniques for lower limb intent recognition. However, to date, no studies have demonstrated an intent recognition system for locomotion modes other than level ground walking for powered prostheses. A more robust and accurate intent recognition system may be obtained by data classification at specific points during the gait cycle. Previous studies have classified data continuously [17] or grouped classification windows based on gait phase and classified either continuously throughout each gait phase [19] or with multiple analysis windows during each gait phase [16]. However, these strategies may introduce additional errors since most signals are nonstationary during the gait cycle (even within a phase). In this study, we analyze the effect of training classifiers using from one to many analysis windows before heel contact and toe off. We determined whether training the classifier with data obtained during transitions between locomotion modes allows the intent recognition system to correctly predict upcoming locomotion modes. To train the classifier with sensor data from natural, seamless transitions, the prosthesis must correctly change locomotion mode while the amputee performs a natural transition to that locomotion mode. Nonamputees make significant kinetic and kinematic adjustments when transitioning to ramps [20] and stairs [21]. Thus, changes in prosthesis parameters must be predicted and rendered before a transition to allow appropriate adjustments at the knee and ankle during a transition. This is an important difference from previous work with passive prostheses [19], in which the prosthesis did not change physical properties during transitions between locomotion modes. We also investigated a number of parameters for training the intent recognition system (e.g., window size, number of analysis windows, and number of repetitions of the training circuit) and discuss the optimal selection for each of these training pa-
Fig. 2. Finite state machine model displayed for two of the five locomotion modes. The black dashed arrows represent within mode switches between stance and swing occurring at heel strike and toe off. Red and blue arrows show the transitions between locomotion modes and only occur at the heel strike or toe off events. Only data from transitions shown by the red arrows (swing to stance) were collected for level-ground walking and ramp ascent during data collection, but the classifier was capable of choosing transitions indicated by either the red or blue arrows.
rameters based on the intent recognition accuracy for various transitions performed by six transfemoral amputees. To our knowledge, this is the first study to construct and test a training system and classifier architecture capable of performing transitions between walking on level ground, ramps, and stairs for a powered prosthesis, using only data from onboard mechanical sensors. II. METHODS A. State Machine Control Impedance-based control implemented through finite state machines was used with a robotic leg built by Vanderbilt University [14], [22]. Each locomotion state had parameters for stiffness, damping, and spring equilibrium angle for both the knee and ankle (six parameters). Fig. 2 gives a representative diagram for transitions between two of the five trained locomotion modes. Since the impedance parameters of the leg change naturally at heel contact and toe off, we found that transitioning between locomotion modes at these points was safe and felt natural for the subjects. State transitions never occurred while the subject had weight on the robotic leg or while the robotic leg was swinging. Thus, the system was exclusively evaluated in terms of the accuracy at these two points in the gait cycle. State machine control was performed through custom software called CAPS (control algorithms for prosthetic systems) [23]. CAPS ran on a desktop computer and communicated with the robotic leg through a tether. Impedance parameters were streamed to the leg from the computer at a rate of 50 Hz (20 ms frames) and could be altered in real-time within the CAPS framework. B. Experimental Protocol Six transfemoral amputee subjects (five males and one female) completed the following experiment, which was ap-
YOUNG et al.: A TRAINING METHOD FOR LOCOMOTION MODE PREDICTION USING POWERED LOWER LIMB PROSTHESES
TABLE I TRANSITIONS BETWEEN LOCOMOTION MODES DURING DATA COLLECTION
proved by the Northwestern University Institutional Review Board. Amputees were fit with a powered knee and ankle prosthesis [24] by a certified prosthetist. Each subject had previous experience using this robotic prosthesis. The robotic prosthesis was empirically tuned to each subject for each ambulation mode as described in [14]. Previously proposed strategies were used to tune impedance parameters for walking [17], stairs [25], and ramps [26]. A physical therapist was present during each experimental session to ensure subject safety. In a single experimental session, subjects completed 20 repetitions of a circuit that included walking on level ground and on a ramp with a 10 slope, and ascending and descending a four-step staircase using reciprocal gait. Subjects transitioned from walking to each of the ramp or stair locomotion modes and back to walking in a sequential order (Table I, supplemental video). Transitions between locomotion modes were performed seamlessly and naturally by the subject as an experimenter manually changed the locomotion mode of the prosthesis at heel contact or toe off—through software on a desktop computer. C. Signal Processing During data collection, a desktop computer recorded signals from 13 sensors on the robotic leg at a sampling frequency of 500 Hz. These signals included positions and velocities for both the knee and ankle, axial load, a six axis inertial measurement unit (three accelerometer measurements and three gyroscope measurements) located on the shank, and the motor current sent to the knee and ankle. The load cell was low pass filtered at 20 Hz. The current state in the state machine was recorded to provide both the current locomotion mode and the phase in the gait cycle at any given time. Data recorded before each toe off and heel contact event was segmented into variable window sizes, and intent recognition performance was evaluated for each window size between 50 and 450 ms in increments of 50 ms. This range of window sizes was used for similar lower limb intent recognition studies [16], [17], [19]; the largest window size represented more than
673
s). Additiona quarter the length of an average gait cycle ( ally, the effect of training the intent recognition system with one or multiple (up to 10) analysis windows before toe off or heel contact in increments of 20 ms was tested, where each window served as one sample for the classifier. Four features (mean, standard deviation, maximum, and minimum) were calculated within each window for all signals, forming a feature vector of size 52. These features have been successfully used on similar sets of mechanical sensors in related work [17], [19] and are suitable for real-time implementation due to their computational simplicity. A linear discriminant analysis (LDA) classifier—previously used in related studies [19]—was used to classify locomotion mode. One classifier was constructed to recognize the five locomotion modes at toe off and a separate classifier was constructed to recognize the same locomotion modes at heel contact. While data from only a limited number of transitions were used to train the classifier, the classifier could select any of the five locomotion modes at heel contact or toe off, as no record of previous decisions was kept. D. Performance Evaluation Classifier evaluation was performed using leave-one-out cross validation with the 20 circuit trials collected for each subject. The effect of the number of circuits used in training was also evaluated to obtain an estimate of how many training circuits would be necessary to create a robust intent recognition system. For overall system evaluation, errors were divided into either transitional errors or steady-state errors. Transitional error was the percentage of misclassifications occurring during transitions (1), while steady-state error was the percentage of misclassifications not occurring at a transition (2) % (1) % (2) All analyses—window size, number of analysis windows, training method, and number of circuit trials—were evaluated in terms of the error rates at heel contact and toe off. For each analysis, a one-way ANOVA test was conducted for both transitional and steady-state error. The effect of training the classifier while the amputee performed seamless and natural transitions was tested and compared to training the classifier with only with steady-state data (data that did not include transitions). This was done to determine if only steady state data, obtained within a locomotion mode, is sufficient to train an intent recognition system. III. RESULTS A. Window Size Analysis as Steady-state errors significantly decreased analysis window size increased (Fig. 3). Steady state error de-
674
IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 22, NO. 3, MAY 2014
TABLE II A: CONFUSION MATRIX FOR TRANSITION STEPS USING NATURAL TRANSITIONS. B: CONFUSION MATRIX FOR TRANSITION STEPS USING STEADY-STATE TRAINING
Fig. 3. Effect of analysis window size on error rates. Transitional error (solid line) and steady-state error (dashed line) are shown for analysis window sizes between 50 and 450 ms. Data are averages for six subjects; error bars represent SEM.
Fig. 4. Effect of number of analysis windows (shown for a 250 ms window size). Both transitional (solid line) and steady-state (dashed line) error increase with a larger number of analysis windows in the training data. Data are averages SEM. for six subjects; error bars represent
creased from 5.9% at a 50 ms analysis window to 3.4% at a 150 ms analysis window, but did not significantly decrease further at larger window sizes. Analysis window size significantly affected transitional error. Transitional errors were smallest at a window length of around 250 ms (Fig. 3). Analysis windows of 250 ms were therefore used for the subsequent studies. The average stride times across subjects were 1.6 s 0.1 s for level walking, ramp ascent and ramp descent, 1.8 s 0.1 s for stair descent, and 2.1 s 0.1 s for stair ascent. Thus, this analysis window of 250 ms represented 12%–16% of the total stride duration. B. Number of Analysis Windows Both steady-state and transitional errors significantly increased with an increasing number of analysis windows (Fig. 4). While Fig. 4 shows data for a 250 ms window, a similar trend was observed for all window sizes tested (50–450 ms). Thus, using a single analysis window before toe off or heel contact was used for the subsequent analyses.
Fig. 5. Effect of training with data that includes transitions compared to training with only steady-state data. Transitional (left) and steady-state (right) errors are shown for each training strategy. Data are averages of six subjects; SEM. Note different vertical axis scales in left and error bars represent right graphs.
C. Effect of Training With Seamless Transitions Table II shows confusion matrices for transitions between locomotion modes for including natural transitions (A) and steady-state (B) training. As expected, training with data obtained during natural transitions resulted in significantly lower error for transition steps and significantly higher error for steady-state steps (Fig. 5) compared to training with steady-state data. For both training strategies, transition steps were much harder to classify than steady-state steps. The steady-state training resulted in misclassification of nearly all % of the transition steps. The transition step has more characteristics of the previous (steady-state) locomotion mode than the upcoming locomotion mode [Table II(B)]. Misclassifications in most transitions from walking chose the walking class while most misclassification to walking chose the previous, nonwalking locomotion mode.
YOUNG et al.: A TRAINING METHOD FOR LOCOMOTION MODE PREDICTION USING POWERED LOWER LIMB PROSTHESES
Fig. 6. Effect of the number of training circuit trials on recognition error. Both transitional and steady state error decreased rapidly as the number of training circuits increased. Effect plateaued at about 11 training trials. Data are averages SEM. for six subjects; error bars represent
D. Number of Circuit Trials Analysis Fig. 6 shows the effect of increasing numbers of training trials on classification error for both transitions and steady state. Both steady-state and transitional error improved significantly for up to five trials. IV. DISCUSSION In this study, we demonstrated the importance of including data from natural transitions in a training strategy for a powered lower limb intent recognition system. Using only steady state information effectively trains the classifier to estimate the current state rather than predict the user’s intent. This is shown that in Fig. 5; steady-state modes are estimated extremely accurately, but transitions are almost always misclassified when using only steady-state training data. Further analysis showed that most of the errors were due to misclassification of the transition step as the previous, rather than the upcoming locomotion mode. Thus, after training with only steady-state data the intent recognition system is not capable of transitioning the device into the proper locomotion mode. For a locomotion mode such as ramp ascent, this may simply translate into a transition that occurs one step late. Other locomotion modes, such as stair ascent, are simply not possible unless the prosthesis correctly transitions at the transition step (i.e., achieves proper foot clearance and placement for the upcoming step). Thus, to obtain an intent recognition system capable of automatic and seamless transitions between locomotion modes, it is likely that transition data must be included in the training strategy. Even when transition data were included in the training data, transitions were still more difficult to classify than steady-states (Fig. 5). Window size analysis revealed that the transitional error could be reduced to an average of 17.0% by selecting an analysis window size of 250 ms (Fig. 3), which is longer than window sizes used in previous studies (e.g., 150 ms [19]). Even with a longer window size, transitions between level ground walking and ramps were especially difficult to classify [Table II(A)], while transitions to/from the stairs were recognized more reliably. This is likely because ramp walking has
675
an overall pattern that is more similar to level ground walking than to stair ascent or descent. Compared to previous studies with passive devices [19], our training method and classifier architecture achieved lower error rates for steady-state classification. In the Huang et al. 2009 study with two transfemoral amputees [16], error rates of 10.3% before toe off and 5.5% before heel contact were achieved for six ambulation modes. After training with only steady-state data, we found error rates of 0.6% before toe off and 1.6% before heel contact for five ambulation modes. One of the primary differences between these studies was that our study used data from mechanical sensors to predict locomotion mode while the Huang et al. study used only EMG information. The more recent 2011 Huang et al. study [19], where LDA classification and mechanical sensor data were used; showed a steady-state error rate of % during the last section of stance phase and of % during swing phase. By including transitions in the training data, we were able to obtain steady-state error rates of 1.3% at the end of stance phase and 7.2% at the end of swing. These differences are likely due to different classification strategies (multiple analysis windows versus single), different sets of mechanical sensors used, and to gait pattern differences between amputees walking on passive devices compared to powered devices [25]. It is possible that adding EMG information from the amputee’s residual muscles may further reduce recognition error [19]. Based on comparisons to previous studies, the mechanical sensor set utilized in this study may be capable of recognition rates comparable to those achieved using EMG data, especially for classifying steady-state information. Future studies will consider using EMG in combination with mechanical sensor data to potentially aid in discriminating locomotion mode transitions in powered devices, especially at transitions where the error rates in this study were relatively high % using only mechanical sensors. To obtain predictable and consistent transitions between locomotion modes, we limited classification decisions to a single window at heel contact and toe off. Other research in intent recognition has focused on making classification decisions either continuously throughout the gait cycle [17], [19] or at discrete points but with multiple analysis windows [16]. In this study, we show that including multiple analysis windows (Fig. 4)—even though these windows occur in the same gait phase—increases classification error for both transition and steady-state modes. This may be the result of using an LDA classifier. The LDA classification strategy is limited in characterization of nonstationary signals, but it is advantageous because a large number of features (in this case 52) may be effectively used for classification with a limited size of training data due to the equal class covariance assumption. If multiple analysis windows are needed for more accurate intent recognition, future work may need to consider other classification algorithms that compensates for nonstationary signals within a locomotion mode. In this study the classifier was trained with the user’s natural transitions by manually switching the prosthesis’ locomotion modes at heel strike and toe off. This is an effective research solution that could be implemented clinically by using a key fob to switch between locomotion modes during a training session.
676
IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 22, NO. 3, MAY 2014
Another limitation of this study is the relatively large amount of training data required to effectively build an intent recognition classifier. Based on Fig. 6, while only five trials are necessary to sufficiently reduce classification error, training the classifier with more trials may be necessary to accommodate variability in the transitions performed by the subjects. This study did not test the generalization of the intent recognition system to novel circumstances. Only transitions that occurred during the experiment were tested as novel data through the leave-one-out cross validation procedure. In the real world, more varying transitions will occur including varying slope and stair grades as well as obstacle accommodation or avoidance. Transition strategies for nonamputees differ based on the slope of the ramp [20]. Transitions to ramps with a smaller grade (the experimental one was 10 ) may have a pattern more similar to level walking than to the ramp ascent. A transition to an extremely steep ramp may have a pattern different than the experimental grade. Stair ambulation has patterns that are very different for level walking, as evidenced by the low transitional error between stairs and level walking [Table II(A)]. Thus, we expect better generalization to stairs with varying grades. In this experiment, amputees transitioned onto the ramp with either leg (sound or prosthetic side), into stair ascent with their sound side, and into stair descent with their prosthetic side. Therefore we expect our system to accommodate transitions to the ramp with either leg (this study has % error for this transition) but not necessarily stair transitions leading with the opposite side. Thus, for stair transitions, we expect that training data must be collected for both types of transitions, or that amputees must be trained to only use one type of transition. Transfemoral amputees currently use this latter solution for stair transitions with their passive devices. Future studies should investigate how well intent recognition systems can generalize to a range of ramp slopes, stair grades, and obstacles typically encountered in the community. Other aspects of community ambulation such as unanticipated transitions and changes in the transitioning leg should also be considered. V. CONCLUSION Automatic, seamless transitions between locomotion modes using a lower limb prosthesis require an accurate intent recognition system. By training the system with data from natural transitions on a powered prosthesis, it is possible to create a system that recognizes an upcoming locomotion mode in time to render appropriate impedance parameters for the device. Additionally, we found that two specific classifiers, trained to recognize locomotion mode at either heel contact or toe off using a single analysis window were more accurate than those trained using multiple analysis windows. Future work will focus on using EMG sensors to improve classification accuracy and increase the number of transitions recognized to provide a global and accurate intent recognition system for walking, ascending and descending stairs and ramps using a robotic leg. ACKNOWLEDGMENT The authors would like to acknowledge T. Idstein and the CBM electronics team for their support.
REFERENCES [1] M. Leplante and D. Carlson, Disability in the United States Nat. Inst. Disabil. Rehabil. Res., Disabil. Stat. Rep. 7, 1996. [2] T. R. Dillingham et al., “Limb amputation and limb deficiency: Epidemiology and recent trends in the United States,” South. Med. J., vol. 95, pp. 875–883, 2002. [3] R. L. Waters et al., “Energy cost of walking of amputees: The influence of level of amputation,” J. Bone Joint Surg. Am. Vol., vol. 58, pp. 42–46, 1976, 1976. [4] W. C. Miller et al., “The prevalence and risk factors of falling and fear of falling among lower extremity amputees,” Arch. Phys. Med. Rehabil., vol. 82, pp. 1031–1037, 2001. [5] T. Schmalz et al., “Biomechanical analysis of stair ambulation in lower limb amputees,” Gait Posture, vol. 25, pp. 267–278, 2007. [6] B. J. Hafner et al., “Evaluation of function, performance, and preference as transfemoral amputees transition from mechanical to microprocessor control of the prosthetic knee,” Arch. Phys. Med. Rehabil., vol. 88, pp. 207–217, 2007. [7] Össur, Rheo Knee [Online]. Available: http://bionics.ossur.com/products/rheo-knee/act [8] A. D. Segal et al., “Kinematic and kinetic comparisons of transfemoral amputee gait using C-leg and mauch SNS prosthetic knees,” J. Rehabil. Res. Develop., vol. 43, pp. 857–870, 2006. [9] I. Otto Bock Orthoped. Ind., Instruction materials—Training with Genium [Online]. Available: http://www.ottobockknees.com/for-professionals/instructional-materials/ [10] J. K. Hitt et al., “An active foot-ankle prosthesis with biomechanical energy regeneration,” J. Med. Devices, vol. 4, p. 011003, 2010. [11] J. Highsmith et al., “Kinetic differences using a power knee and C-leg while sitting down and standing up: A case report,” J. Prosthet. Orthot., vol. 22, pp. 237–243, 2010. [12] S. Au et al., “Powered ankle-foot prosthesis to assist level-ground and stair-descent gaits,” Neural. Netw., vol. 26, p. 26, Apr. 26, 2008. [13] F. Sup et al., “Design and control of a powered transfemoral prosthesis,” Int. J. Robot. Res., vol. 27, pp. 263–273, Feb. 2008. [14] K. B. Fite et al., “Design and control of an electrically powered knee prosthesis,” in Proc. 10th IEEE Int. Conf. Rehabil. Robot., Jun. 12-15, 2007, pp. 902–905. [15] I. Otto Bock Orthopedic Ind., Manual for the 3c100 Otto Bock C-LEG. Duderstadt, Germany, 1998. [16] H. Huang et al., “A strategy for identifying locomotion modes using surface electromyography,” IEEE Trans. Biomed. Eng., vol. 56, no. 1, pp. 65–73, Jan. 2009, 2009. [17] H. A. Varol et al., “Multiclass real-time intent recognition of a powered lower limb prosthesis,” IEEE Trans. Biomed. Eng., vol. 57, no. 3, pp. 542–551, Mar. 2010. [18] W. C. Flowers and R. W. Mann, “Electrohydraulic knee-torque controller for a prosthesis simulator,” ASME J. Biomechan. Eng., vol. 99, pp. 3–8, 1977. [19] H. Huang et al., “Continuous locomotion-mode identification for prosthetic legs based on neuromuscular mechanical fusion,” IEEE Trans. Biomed. Eng., vol. 58, no. 10, pp. 2867–2875, Oct. 2011. [20] S. Prentice et al., “Locomotor adaptations for changes in the slope of the walking surface,” Gait Posture, vol. 20, pp. 255–265, 2004. [21] B. J. McFadyen and H. Carnahan, “Anticipatory locomotor adjustments for accommodating versus avoiding level changes in humans,” Exp. Brain Res., vol. 114, pp. 500–506, 1997. [22] F. Sup et al., “Design and control of a powered knee and ankle prosthesis,” in IEEE Int. Conf. Robot. Automat., Rome, Italy, 2007, pp. 4134–4139. [23] T. A. Kuiken et al., “Targeted muscle reinnervation for real-time myoelectric control of multifunction artificial arms,” JAMA, vol. 301, pp. 619–28, Feb. 11, 2009. [24] F. Sup et al., “Preliminary evaluations of a self-contained antrhopomorphic transfemoral prosthesis,” IEEE ASME Trans. Mechatron., vol. 14, 2009. [25] B. E. Lawson et al., “Control of stair ascent and descent with a powered transfemoral prosthesis,” IEEE Trans. Neural. Syst. Rehabil. Eng., vol. 21, pp. 466–473, 2013. [26] F. Sup et al., “Upslope walking with a powered knee and ankle prosthesis: Initial results with an amputee subject,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 19, no. 1, pp. 71–78, Feb. 2011.
YOUNG et al.: A TRAINING METHOD FOR LOCOMOTION MODE PREDICTION USING POWERED LOWER LIMB PROSTHESES
Aaron J. Young (S’11) received the B.S. degree in biomedical engineering from Purdue University, West Lafayette, IN, USA, in 2009, and the M.S. degree in biomedical engineering from Northwestern University, Chicago, IL, USA, in 2011. He is currently working toward the Ph.D. degree at Northwestern University in the Center for Bionic Medicine at the Rehabilitation Institute of Chicago, Chicago, IL, USA. His research interests include neural signal processing and pattern recognition using advanced machine learning techniques for control of myoelectric prosthesis for the upper and lower limb.
Ann M. Simon (M’12) received the B.S. degree in biomedical engineering from Marquette University, Milwaukee, WI, USA, in 2003, and the M.S. degree in mechanical engineering and the Ph.D. degree in biomedical engineering from the University of Michigan, Ann Arbor, MI, USA, in 2007 and 2008, respectively. She is a Biomedical Engineering Manager in the Center for Bionic Medicine at the Rehabilitation Institute of Chicago, Chicago, IL, USA, and a Research Assistant Professor in the Department of Physical Medicine and Rehabilitation at Northwestern University, Chicago, IL, USA. Her research interests include overcoming clinical challenges associated with the application of advanced pattern recognition myoelectric control systems for both upper- and lower-limb amputees.
677
Levi J. Hargrove (S’05–M’08) received the B.Sc., M.Sc., and Ph.D. degrees in electrical engineering from the University of New Brunswick, Fredericton, NB, Canada, in 2003, 2005, and 2007, respectively. He joined the Center for Bionic Medicine at the Rehabilitation Institute of Chicago, Chicago, IL, USA, in 2008. His research interests include pattern recognition, biological signal processing, and myoelectric control of powered prostheses. He is also a Research Assistant Professor in the Department of Physical Medicine and Rehabilitation (PM&R) and Biomedical Engineering, Northwestern University, Chicago, IL, USA. Dr. Hargrove is a member of the Association of Professional Engineers and Geoscientists of New Brunswick.
671
A Training Method for Locomotion Mode Prediction Using Powered Lower Limb Prostheses Aaron J. Young, Student Member, IEEE, Ann M. Simon, Member, IEEE, and Levi J. Hargrove, Member, IEEE
Abstract—Recently developed lower-limb prostheses are capable of actuating the knee and ankle joints, allowing amputees to perform advanced locomotion modes such as step-over-step stair ascent and walking on sloped surfaces. However, transitions between these locomotion modes and walking are neither automatic nor seamless. This study describes methods for construction and training of a high-level intent recognition system for a lower-limb prosthesis that provides natural transitions between walking, stair ascent, stair descent, ramp ascent, and ramp descent. Using mechanical sensors onboard a powered prosthesis, we collected steady-state and transition data from six transfemoral amputees while the five locomotion modes were performed. An intent recognition system built using only mechanical sensor data was 84.5% accurate using only steady-state training data. Including training data collected while amputees performed seamless transitions between locomotion modes improved the overall accuracy rate to 93.9%. Training using a single analysis window at heel contact and toe off provided higher recognition accuracy than training with multiple analysis windows. This study demonstrates the capability of an intent recognition system to provide automatic, natural, and seamless transitions between five locomotion modes for transfemoral amputees using powered lower limb prostheses. Index Terms—Intent recognition, powered lower limb prosthesis, robotic leg control, transfemoral amputee.
I. INTRODUCTION
T
RANSFEMORAL amputation is a significant cause of disability in the United States with approximately 31 000 new cases occurring each year [1], [2]. Transfemoral amputees who use passive prostheses have significantly impaired balance, walking symmetry, and metabolic energy efficiency [3], [4].
Manuscript received February 05, 2013; revised August 26, 2013; accepted October 03, 2013. Date of publication October 30, 2013; date of current version April 28, 2014. This work was supported in part by the US Army’s Telemedicine and Advanced Technology Research Center under Grant 81XWH-09-2-0020. The work of A. J. Young was supported by a National Defense Science and Engineering Graduate (NDSEG) fellowship. This paper has supplementary downloadable material available at http://ieeexplore.ieee.org, provided by the authors. A. J. Young is with the Center for Bionic Medicine, Rehabilitation Institute of Chicago, Chicago, IL 60611 USA, and also with the Department of Biomedical Engineering, Northwestern University, Chicago, IL 60611 USA (e-mail: [email protected]). A. M. Simon is with the Center for Bionic Medicine, Rehabilitation Institute of Chicago, Chicago, IL 60611 USA, and also with the Department of Physical Medicine and Rehabilitation, Northwestern University, Chicago, IL 60611 USA (e-mail: [email protected]). L. J. Hargrove, is with the Center for Bionic Medicine, Rehabilitation Institute of Chicago, Chicago, IL 60611 USA, and with the Department of Physical Medicine and Rehabilitation, Northwestern University, Chicago, IL 60611 USA (e-mail: [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TNSRE.2013.2285101
Fig. 1. Powered knee/ankle prosthesis during stair ascent.
They also tend to have difficulty navigating more demanding terrain such as ramps and stairs [5]. These issues have motivated the development of advanced microprocessor controlled prostheses. Microprocessor controlled prostheses may be either mechanically passive [6]–[9] or mechanically active [10]–[13]. Both types of devices can restore a variety of locomotion modes and rely on using finite-state machines to interpret data from sensors (e.g., load cells, goniometers, inertial measurement units, etc.) attached to the prosthesis. Determining proper state transitions within a locomotion mode, such as between stance and swing phases of walking, is straightforward and has been implemented in commercially available prostheses [12], [14], [15]. Defining state transitions between locomotion modes, such as from the stance phase of walking to the swing phase of stair ascent, is more difficult. Proposed transition methods include using key-fobs [15], compensatory body movements [9], or intent recognition signal processing algorithms [16], [17]. Intent recognition algorithms have the potential to provide seamless (no stopping or slowing), automatic (no button press or direct user command), and natural (no unintuitive actions) transitions between locomotion modes. Furthermore, they can be used for both mechanically passive and mechanically active devices. In this work, we focus on intent recognition algorithms in mechanically active (motorized knee–ankle) prostheses (Fig. 1) [13]. Several intent recognition strategies have been proposed to perform transitions between activity modes for powered lower limb prostheses. In echo control [18] the prosthetic leg mimics the kinematics of the sound limb. This approach requires minimal system training, but the sound limb must be instrumented
1534-4320 © 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
672
IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 22, NO. 3, MAY 2014
and the user must always lead with the sound limb and take an even number of steps. Pattern recognition strategies have been shown to predict locomotion modes and used to transition a transfemoral prosthesis between level ground walking, sitting, and standing modes [17]. System training involved collecting mechanical sensor data during steady-state conditions as well as during all combinations where the prosthesis was in an incorrect locomotion mode (e.g., walking while the prosthesis was in standing mode). While high recognition rates were achieved, the training requirement makes this approach difficult to implement if more than two locomotion modes are required. Additionally, an intent recognition system that relies on detecting when the prosthesis is in the wrong mode may respond too late to achieve natural and seamless transitions. Intent recognition strategies have been used in mechanically passive devices while amputees performed various locomotion modes. Huang et al. [16] obtained an average intent recognition accuracy of 85.7% for six locomotion modes using only residual limb electromyography (EMG) signals in two transfemoral amputees. This study used only steady-state data to train the classifier; data from transitions between locomotion modes were not included. Huang et al. [19] also achieved recognition accuracies above 95% by using a combination of EMG and load cell information for five steady-state locomotion modes in five transfemoral amputees walking on a passive device. These previous studies demonstrate the potential of using pattern recognition techniques for lower limb intent recognition. However, to date, no studies have demonstrated an intent recognition system for locomotion modes other than level ground walking for powered prostheses. A more robust and accurate intent recognition system may be obtained by data classification at specific points during the gait cycle. Previous studies have classified data continuously [17] or grouped classification windows based on gait phase and classified either continuously throughout each gait phase [19] or with multiple analysis windows during each gait phase [16]. However, these strategies may introduce additional errors since most signals are nonstationary during the gait cycle (even within a phase). In this study, we analyze the effect of training classifiers using from one to many analysis windows before heel contact and toe off. We determined whether training the classifier with data obtained during transitions between locomotion modes allows the intent recognition system to correctly predict upcoming locomotion modes. To train the classifier with sensor data from natural, seamless transitions, the prosthesis must correctly change locomotion mode while the amputee performs a natural transition to that locomotion mode. Nonamputees make significant kinetic and kinematic adjustments when transitioning to ramps [20] and stairs [21]. Thus, changes in prosthesis parameters must be predicted and rendered before a transition to allow appropriate adjustments at the knee and ankle during a transition. This is an important difference from previous work with passive prostheses [19], in which the prosthesis did not change physical properties during transitions between locomotion modes. We also investigated a number of parameters for training the intent recognition system (e.g., window size, number of analysis windows, and number of repetitions of the training circuit) and discuss the optimal selection for each of these training pa-
Fig. 2. Finite state machine model displayed for two of the five locomotion modes. The black dashed arrows represent within mode switches between stance and swing occurring at heel strike and toe off. Red and blue arrows show the transitions between locomotion modes and only occur at the heel strike or toe off events. Only data from transitions shown by the red arrows (swing to stance) were collected for level-ground walking and ramp ascent during data collection, but the classifier was capable of choosing transitions indicated by either the red or blue arrows.
rameters based on the intent recognition accuracy for various transitions performed by six transfemoral amputees. To our knowledge, this is the first study to construct and test a training system and classifier architecture capable of performing transitions between walking on level ground, ramps, and stairs for a powered prosthesis, using only data from onboard mechanical sensors. II. METHODS A. State Machine Control Impedance-based control implemented through finite state machines was used with a robotic leg built by Vanderbilt University [14], [22]. Each locomotion state had parameters for stiffness, damping, and spring equilibrium angle for both the knee and ankle (six parameters). Fig. 2 gives a representative diagram for transitions between two of the five trained locomotion modes. Since the impedance parameters of the leg change naturally at heel contact and toe off, we found that transitioning between locomotion modes at these points was safe and felt natural for the subjects. State transitions never occurred while the subject had weight on the robotic leg or while the robotic leg was swinging. Thus, the system was exclusively evaluated in terms of the accuracy at these two points in the gait cycle. State machine control was performed through custom software called CAPS (control algorithms for prosthetic systems) [23]. CAPS ran on a desktop computer and communicated with the robotic leg through a tether. Impedance parameters were streamed to the leg from the computer at a rate of 50 Hz (20 ms frames) and could be altered in real-time within the CAPS framework. B. Experimental Protocol Six transfemoral amputee subjects (five males and one female) completed the following experiment, which was ap-
YOUNG et al.: A TRAINING METHOD FOR LOCOMOTION MODE PREDICTION USING POWERED LOWER LIMB PROSTHESES
TABLE I TRANSITIONS BETWEEN LOCOMOTION MODES DURING DATA COLLECTION
proved by the Northwestern University Institutional Review Board. Amputees were fit with a powered knee and ankle prosthesis [24] by a certified prosthetist. Each subject had previous experience using this robotic prosthesis. The robotic prosthesis was empirically tuned to each subject for each ambulation mode as described in [14]. Previously proposed strategies were used to tune impedance parameters for walking [17], stairs [25], and ramps [26]. A physical therapist was present during each experimental session to ensure subject safety. In a single experimental session, subjects completed 20 repetitions of a circuit that included walking on level ground and on a ramp with a 10 slope, and ascending and descending a four-step staircase using reciprocal gait. Subjects transitioned from walking to each of the ramp or stair locomotion modes and back to walking in a sequential order (Table I, supplemental video). Transitions between locomotion modes were performed seamlessly and naturally by the subject as an experimenter manually changed the locomotion mode of the prosthesis at heel contact or toe off—through software on a desktop computer. C. Signal Processing During data collection, a desktop computer recorded signals from 13 sensors on the robotic leg at a sampling frequency of 500 Hz. These signals included positions and velocities for both the knee and ankle, axial load, a six axis inertial measurement unit (three accelerometer measurements and three gyroscope measurements) located on the shank, and the motor current sent to the knee and ankle. The load cell was low pass filtered at 20 Hz. The current state in the state machine was recorded to provide both the current locomotion mode and the phase in the gait cycle at any given time. Data recorded before each toe off and heel contact event was segmented into variable window sizes, and intent recognition performance was evaluated for each window size between 50 and 450 ms in increments of 50 ms. This range of window sizes was used for similar lower limb intent recognition studies [16], [17], [19]; the largest window size represented more than
673
s). Additiona quarter the length of an average gait cycle ( ally, the effect of training the intent recognition system with one or multiple (up to 10) analysis windows before toe off or heel contact in increments of 20 ms was tested, where each window served as one sample for the classifier. Four features (mean, standard deviation, maximum, and minimum) were calculated within each window for all signals, forming a feature vector of size 52. These features have been successfully used on similar sets of mechanical sensors in related work [17], [19] and are suitable for real-time implementation due to their computational simplicity. A linear discriminant analysis (LDA) classifier—previously used in related studies [19]—was used to classify locomotion mode. One classifier was constructed to recognize the five locomotion modes at toe off and a separate classifier was constructed to recognize the same locomotion modes at heel contact. While data from only a limited number of transitions were used to train the classifier, the classifier could select any of the five locomotion modes at heel contact or toe off, as no record of previous decisions was kept. D. Performance Evaluation Classifier evaluation was performed using leave-one-out cross validation with the 20 circuit trials collected for each subject. The effect of the number of circuits used in training was also evaluated to obtain an estimate of how many training circuits would be necessary to create a robust intent recognition system. For overall system evaluation, errors were divided into either transitional errors or steady-state errors. Transitional error was the percentage of misclassifications occurring during transitions (1), while steady-state error was the percentage of misclassifications not occurring at a transition (2) % (1) % (2) All analyses—window size, number of analysis windows, training method, and number of circuit trials—were evaluated in terms of the error rates at heel contact and toe off. For each analysis, a one-way ANOVA test was conducted for both transitional and steady-state error. The effect of training the classifier while the amputee performed seamless and natural transitions was tested and compared to training the classifier with only with steady-state data (data that did not include transitions). This was done to determine if only steady state data, obtained within a locomotion mode, is sufficient to train an intent recognition system. III. RESULTS A. Window Size Analysis as Steady-state errors significantly decreased analysis window size increased (Fig. 3). Steady state error de-
674
IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 22, NO. 3, MAY 2014
TABLE II A: CONFUSION MATRIX FOR TRANSITION STEPS USING NATURAL TRANSITIONS. B: CONFUSION MATRIX FOR TRANSITION STEPS USING STEADY-STATE TRAINING
Fig. 3. Effect of analysis window size on error rates. Transitional error (solid line) and steady-state error (dashed line) are shown for analysis window sizes between 50 and 450 ms. Data are averages for six subjects; error bars represent SEM.
Fig. 4. Effect of number of analysis windows (shown for a 250 ms window size). Both transitional (solid line) and steady-state (dashed line) error increase with a larger number of analysis windows in the training data. Data are averages SEM. for six subjects; error bars represent
creased from 5.9% at a 50 ms analysis window to 3.4% at a 150 ms analysis window, but did not significantly decrease further at larger window sizes. Analysis window size significantly affected transitional error. Transitional errors were smallest at a window length of around 250 ms (Fig. 3). Analysis windows of 250 ms were therefore used for the subsequent studies. The average stride times across subjects were 1.6 s 0.1 s for level walking, ramp ascent and ramp descent, 1.8 s 0.1 s for stair descent, and 2.1 s 0.1 s for stair ascent. Thus, this analysis window of 250 ms represented 12%–16% of the total stride duration. B. Number of Analysis Windows Both steady-state and transitional errors significantly increased with an increasing number of analysis windows (Fig. 4). While Fig. 4 shows data for a 250 ms window, a similar trend was observed for all window sizes tested (50–450 ms). Thus, using a single analysis window before toe off or heel contact was used for the subsequent analyses.
Fig. 5. Effect of training with data that includes transitions compared to training with only steady-state data. Transitional (left) and steady-state (right) errors are shown for each training strategy. Data are averages of six subjects; SEM. Note different vertical axis scales in left and error bars represent right graphs.
C. Effect of Training With Seamless Transitions Table II shows confusion matrices for transitions between locomotion modes for including natural transitions (A) and steady-state (B) training. As expected, training with data obtained during natural transitions resulted in significantly lower error for transition steps and significantly higher error for steady-state steps (Fig. 5) compared to training with steady-state data. For both training strategies, transition steps were much harder to classify than steady-state steps. The steady-state training resulted in misclassification of nearly all % of the transition steps. The transition step has more characteristics of the previous (steady-state) locomotion mode than the upcoming locomotion mode [Table II(B)]. Misclassifications in most transitions from walking chose the walking class while most misclassification to walking chose the previous, nonwalking locomotion mode.
YOUNG et al.: A TRAINING METHOD FOR LOCOMOTION MODE PREDICTION USING POWERED LOWER LIMB PROSTHESES
Fig. 6. Effect of the number of training circuit trials on recognition error. Both transitional and steady state error decreased rapidly as the number of training circuits increased. Effect plateaued at about 11 training trials. Data are averages SEM. for six subjects; error bars represent
D. Number of Circuit Trials Analysis Fig. 6 shows the effect of increasing numbers of training trials on classification error for both transitions and steady state. Both steady-state and transitional error improved significantly for up to five trials. IV. DISCUSSION In this study, we demonstrated the importance of including data from natural transitions in a training strategy for a powered lower limb intent recognition system. Using only steady state information effectively trains the classifier to estimate the current state rather than predict the user’s intent. This is shown that in Fig. 5; steady-state modes are estimated extremely accurately, but transitions are almost always misclassified when using only steady-state training data. Further analysis showed that most of the errors were due to misclassification of the transition step as the previous, rather than the upcoming locomotion mode. Thus, after training with only steady-state data the intent recognition system is not capable of transitioning the device into the proper locomotion mode. For a locomotion mode such as ramp ascent, this may simply translate into a transition that occurs one step late. Other locomotion modes, such as stair ascent, are simply not possible unless the prosthesis correctly transitions at the transition step (i.e., achieves proper foot clearance and placement for the upcoming step). Thus, to obtain an intent recognition system capable of automatic and seamless transitions between locomotion modes, it is likely that transition data must be included in the training strategy. Even when transition data were included in the training data, transitions were still more difficult to classify than steady-states (Fig. 5). Window size analysis revealed that the transitional error could be reduced to an average of 17.0% by selecting an analysis window size of 250 ms (Fig. 3), which is longer than window sizes used in previous studies (e.g., 150 ms [19]). Even with a longer window size, transitions between level ground walking and ramps were especially difficult to classify [Table II(A)], while transitions to/from the stairs were recognized more reliably. This is likely because ramp walking has
675
an overall pattern that is more similar to level ground walking than to stair ascent or descent. Compared to previous studies with passive devices [19], our training method and classifier architecture achieved lower error rates for steady-state classification. In the Huang et al. 2009 study with two transfemoral amputees [16], error rates of 10.3% before toe off and 5.5% before heel contact were achieved for six ambulation modes. After training with only steady-state data, we found error rates of 0.6% before toe off and 1.6% before heel contact for five ambulation modes. One of the primary differences between these studies was that our study used data from mechanical sensors to predict locomotion mode while the Huang et al. study used only EMG information. The more recent 2011 Huang et al. study [19], where LDA classification and mechanical sensor data were used; showed a steady-state error rate of % during the last section of stance phase and of % during swing phase. By including transitions in the training data, we were able to obtain steady-state error rates of 1.3% at the end of stance phase and 7.2% at the end of swing. These differences are likely due to different classification strategies (multiple analysis windows versus single), different sets of mechanical sensors used, and to gait pattern differences between amputees walking on passive devices compared to powered devices [25]. It is possible that adding EMG information from the amputee’s residual muscles may further reduce recognition error [19]. Based on comparisons to previous studies, the mechanical sensor set utilized in this study may be capable of recognition rates comparable to those achieved using EMG data, especially for classifying steady-state information. Future studies will consider using EMG in combination with mechanical sensor data to potentially aid in discriminating locomotion mode transitions in powered devices, especially at transitions where the error rates in this study were relatively high % using only mechanical sensors. To obtain predictable and consistent transitions between locomotion modes, we limited classification decisions to a single window at heel contact and toe off. Other research in intent recognition has focused on making classification decisions either continuously throughout the gait cycle [17], [19] or at discrete points but with multiple analysis windows [16]. In this study, we show that including multiple analysis windows (Fig. 4)—even though these windows occur in the same gait phase—increases classification error for both transition and steady-state modes. This may be the result of using an LDA classifier. The LDA classification strategy is limited in characterization of nonstationary signals, but it is advantageous because a large number of features (in this case 52) may be effectively used for classification with a limited size of training data due to the equal class covariance assumption. If multiple analysis windows are needed for more accurate intent recognition, future work may need to consider other classification algorithms that compensates for nonstationary signals within a locomotion mode. In this study the classifier was trained with the user’s natural transitions by manually switching the prosthesis’ locomotion modes at heel strike and toe off. This is an effective research solution that could be implemented clinically by using a key fob to switch between locomotion modes during a training session.
676
IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 22, NO. 3, MAY 2014
Another limitation of this study is the relatively large amount of training data required to effectively build an intent recognition classifier. Based on Fig. 6, while only five trials are necessary to sufficiently reduce classification error, training the classifier with more trials may be necessary to accommodate variability in the transitions performed by the subjects. This study did not test the generalization of the intent recognition system to novel circumstances. Only transitions that occurred during the experiment were tested as novel data through the leave-one-out cross validation procedure. In the real world, more varying transitions will occur including varying slope and stair grades as well as obstacle accommodation or avoidance. Transition strategies for nonamputees differ based on the slope of the ramp [20]. Transitions to ramps with a smaller grade (the experimental one was 10 ) may have a pattern more similar to level walking than to the ramp ascent. A transition to an extremely steep ramp may have a pattern different than the experimental grade. Stair ambulation has patterns that are very different for level walking, as evidenced by the low transitional error between stairs and level walking [Table II(A)]. Thus, we expect better generalization to stairs with varying grades. In this experiment, amputees transitioned onto the ramp with either leg (sound or prosthetic side), into stair ascent with their sound side, and into stair descent with their prosthetic side. Therefore we expect our system to accommodate transitions to the ramp with either leg (this study has % error for this transition) but not necessarily stair transitions leading with the opposite side. Thus, for stair transitions, we expect that training data must be collected for both types of transitions, or that amputees must be trained to only use one type of transition. Transfemoral amputees currently use this latter solution for stair transitions with their passive devices. Future studies should investigate how well intent recognition systems can generalize to a range of ramp slopes, stair grades, and obstacles typically encountered in the community. Other aspects of community ambulation such as unanticipated transitions and changes in the transitioning leg should also be considered. V. CONCLUSION Automatic, seamless transitions between locomotion modes using a lower limb prosthesis require an accurate intent recognition system. By training the system with data from natural transitions on a powered prosthesis, it is possible to create a system that recognizes an upcoming locomotion mode in time to render appropriate impedance parameters for the device. Additionally, we found that two specific classifiers, trained to recognize locomotion mode at either heel contact or toe off using a single analysis window were more accurate than those trained using multiple analysis windows. Future work will focus on using EMG sensors to improve classification accuracy and increase the number of transitions recognized to provide a global and accurate intent recognition system for walking, ascending and descending stairs and ramps using a robotic leg. ACKNOWLEDGMENT The authors would like to acknowledge T. Idstein and the CBM electronics team for their support.
REFERENCES [1] M. Leplante and D. Carlson, Disability in the United States Nat. Inst. Disabil. Rehabil. Res., Disabil. Stat. Rep. 7, 1996. [2] T. R. Dillingham et al., “Limb amputation and limb deficiency: Epidemiology and recent trends in the United States,” South. Med. J., vol. 95, pp. 875–883, 2002. [3] R. L. Waters et al., “Energy cost of walking of amputees: The influence of level of amputation,” J. Bone Joint Surg. Am. Vol., vol. 58, pp. 42–46, 1976, 1976. [4] W. C. Miller et al., “The prevalence and risk factors of falling and fear of falling among lower extremity amputees,” Arch. Phys. Med. Rehabil., vol. 82, pp. 1031–1037, 2001. [5] T. Schmalz et al., “Biomechanical analysis of stair ambulation in lower limb amputees,” Gait Posture, vol. 25, pp. 267–278, 2007. [6] B. J. Hafner et al., “Evaluation of function, performance, and preference as transfemoral amputees transition from mechanical to microprocessor control of the prosthetic knee,” Arch. Phys. Med. Rehabil., vol. 88, pp. 207–217, 2007. [7] Össur, Rheo Knee [Online]. Available: http://bionics.ossur.com/products/rheo-knee/act [8] A. D. Segal et al., “Kinematic and kinetic comparisons of transfemoral amputee gait using C-leg and mauch SNS prosthetic knees,” J. Rehabil. Res. Develop., vol. 43, pp. 857–870, 2006. [9] I. Otto Bock Orthoped. Ind., Instruction materials—Training with Genium [Online]. Available: http://www.ottobockknees.com/for-professionals/instructional-materials/ [10] J. K. Hitt et al., “An active foot-ankle prosthesis with biomechanical energy regeneration,” J. Med. Devices, vol. 4, p. 011003, 2010. [11] J. Highsmith et al., “Kinetic differences using a power knee and C-leg while sitting down and standing up: A case report,” J. Prosthet. Orthot., vol. 22, pp. 237–243, 2010. [12] S. Au et al., “Powered ankle-foot prosthesis to assist level-ground and stair-descent gaits,” Neural. Netw., vol. 26, p. 26, Apr. 26, 2008. [13] F. Sup et al., “Design and control of a powered transfemoral prosthesis,” Int. J. Robot. Res., vol. 27, pp. 263–273, Feb. 2008. [14] K. B. Fite et al., “Design and control of an electrically powered knee prosthesis,” in Proc. 10th IEEE Int. Conf. Rehabil. Robot., Jun. 12-15, 2007, pp. 902–905. [15] I. Otto Bock Orthopedic Ind., Manual for the 3c100 Otto Bock C-LEG. Duderstadt, Germany, 1998. [16] H. Huang et al., “A strategy for identifying locomotion modes using surface electromyography,” IEEE Trans. Biomed. Eng., vol. 56, no. 1, pp. 65–73, Jan. 2009, 2009. [17] H. A. Varol et al., “Multiclass real-time intent recognition of a powered lower limb prosthesis,” IEEE Trans. Biomed. Eng., vol. 57, no. 3, pp. 542–551, Mar. 2010. [18] W. C. Flowers and R. W. Mann, “Electrohydraulic knee-torque controller for a prosthesis simulator,” ASME J. Biomechan. Eng., vol. 99, pp. 3–8, 1977. [19] H. Huang et al., “Continuous locomotion-mode identification for prosthetic legs based on neuromuscular mechanical fusion,” IEEE Trans. Biomed. Eng., vol. 58, no. 10, pp. 2867–2875, Oct. 2011. [20] S. Prentice et al., “Locomotor adaptations for changes in the slope of the walking surface,” Gait Posture, vol. 20, pp. 255–265, 2004. [21] B. J. McFadyen and H. Carnahan, “Anticipatory locomotor adjustments for accommodating versus avoiding level changes in humans,” Exp. Brain Res., vol. 114, pp. 500–506, 1997. [22] F. Sup et al., “Design and control of a powered knee and ankle prosthesis,” in IEEE Int. Conf. Robot. Automat., Rome, Italy, 2007, pp. 4134–4139. [23] T. A. Kuiken et al., “Targeted muscle reinnervation for real-time myoelectric control of multifunction artificial arms,” JAMA, vol. 301, pp. 619–28, Feb. 11, 2009. [24] F. Sup et al., “Preliminary evaluations of a self-contained antrhopomorphic transfemoral prosthesis,” IEEE ASME Trans. Mechatron., vol. 14, 2009. [25] B. E. Lawson et al., “Control of stair ascent and descent with a powered transfemoral prosthesis,” IEEE Trans. Neural. Syst. Rehabil. Eng., vol. 21, pp. 466–473, 2013. [26] F. Sup et al., “Upslope walking with a powered knee and ankle prosthesis: Initial results with an amputee subject,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 19, no. 1, pp. 71–78, Feb. 2011.
YOUNG et al.: A TRAINING METHOD FOR LOCOMOTION MODE PREDICTION USING POWERED LOWER LIMB PROSTHESES
Aaron J. Young (S’11) received the B.S. degree in biomedical engineering from Purdue University, West Lafayette, IN, USA, in 2009, and the M.S. degree in biomedical engineering from Northwestern University, Chicago, IL, USA, in 2011. He is currently working toward the Ph.D. degree at Northwestern University in the Center for Bionic Medicine at the Rehabilitation Institute of Chicago, Chicago, IL, USA. His research interests include neural signal processing and pattern recognition using advanced machine learning techniques for control of myoelectric prosthesis for the upper and lower limb.
Ann M. Simon (M’12) received the B.S. degree in biomedical engineering from Marquette University, Milwaukee, WI, USA, in 2003, and the M.S. degree in mechanical engineering and the Ph.D. degree in biomedical engineering from the University of Michigan, Ann Arbor, MI, USA, in 2007 and 2008, respectively. She is a Biomedical Engineering Manager in the Center for Bionic Medicine at the Rehabilitation Institute of Chicago, Chicago, IL, USA, and a Research Assistant Professor in the Department of Physical Medicine and Rehabilitation at Northwestern University, Chicago, IL, USA. Her research interests include overcoming clinical challenges associated with the application of advanced pattern recognition myoelectric control systems for both upper- and lower-limb amputees.
677
Levi J. Hargrove (S’05–M’08) received the B.Sc., M.Sc., and Ph.D. degrees in electrical engineering from the University of New Brunswick, Fredericton, NB, Canada, in 2003, 2005, and 2007, respectively. He joined the Center for Bionic Medicine at the Rehabilitation Institute of Chicago, Chicago, IL, USA, in 2008. His research interests include pattern recognition, biological signal processing, and myoelectric control of powered prostheses. He is also a Research Assistant Professor in the Department of Physical Medicine and Rehabilitation (PM&R) and Biomedical Engineering, Northwestern University, Chicago, IL, USA. Dr. Hargrove is a member of the Association of Professional Engineers and Geoscientists of New Brunswick.
