Multimodal Sensing and Interaction for a Robotic Hand Orthosis

07/31/2018 ∙ by Sangwoo Park, et al. ∙ Columbia University 0

Wearable robotic hand rehabilitation devices can allow greater freedom and flexibility than their workstation-like counterparts. However, the field is generally lacking effective methods by which the user can operate the device: such controls must be effective, intuitive and robust to the wide range of possible impairment patterns. Even when focusing on a specific condition, such as stroke, the variety of encountered upper limb impairment patterns means that a single sensing modality, such as electromyography (EMG), is not sufficient to enable controls for a broad range of users. To address this significant gap, we introduce a multimodal sensing and interaction paradigm for an active hand orthosis. In our proof-of-concept implementation, EMG is complemented by other sensing modalities, such as finger bend and contact pressure sensors. We propose multimodal interaction methods that utilize this sensory data as input, and show they can enable tasks for stroke survivors who exhibit different impairment patterns. We believe that robotic hand orthoses developed as multimodal sensory platforms with help address some of the key challenges in physical interaction with the user.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 1

page 3

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

I Introduction

Robotic devices for hand rehabilitation promise to alleviate some of the critical challenges of traditional rehabilitation paradigms. In particular, they could significantly increase the number of training exercises for cases where access to a therapist is limited. Recent advances [1, 2, 3, 4, 5, 6] have greatly improved the wearability of such orthoses: we now have devices that provide the needed actuation capabilities in a compact, wearable package, allowing greater freedom and flexibility than their workstation-like counterparts. Such wearable devices could allow use beyond the confines of a therapist’s office.

However, the vision of a wearable orthotic device used for activities of daily living (ADLs) can only be realized if the patients are able to operate the device themselves. Control methods must be effective and intuitive, robust to long term operation, and cannot impose significant cognitive load. These algorithms must also cope with a wide range of impairment levels and abilities in the target population. While the actuation abilities of robotic hand orthoses have made great strides, control algorithms have not made similar progress in addressing these challenges.

The key to intuitive, user-driven control for a wearable orthosis lies in the ability to infer the user’s intent from sensor data collected by the device. The robotic orthosis thus becomes a sensory platform in addition to an actuation mechanism. The control algorithm must infer the intent of the user from the collected data and respond with the appropriate actuation commands. In our own previous work, we developed a tendon-driven hand orthosis [1] that used a single sensing modality, forearm surface electromyography (EMG), to infer user intent [7]. Other studies (which we review in the next section) have also investigated EMG for control of wearable robotic devices.

While this body of work has shown that intuitive control is indeed possible, it has also highlighted numerous challenges. For example, EMG signals are inherently abnormal in hemiparesis and distorted by spasticity and fatigue [8],  [9]. If signal patterns drift or change between training and deployment, the control method has no way of coping without new calibration or training data. Physical interaction with the orthosis also alters the signals. In fact, other unimodal interaction methods face similar challenges : if the nature of the impairment (which varies greatly between individuals) is such that the signal exhibits too much or too little variation, the entire device can become unusable.

Fig. 1: Prototype of a hand orthotic device with multimodal sensors.

We were thus motivated to research and develop various forms of sensing on a wearable hand orthosis for intent detection, where different sensing modalities can complement and augment each other. We believe that this multimodal sensing approach can help address the aforementioned key challenges for robust, intuitive user-driven operation. In this study, we aim for test of feasibility by introducing a multisensory implementation developed for stroke patients.

While stroke subjects display a wide variety of impairment patterns, we have observed that many retain subtle, but consistent residual movements (e.g. partial extension of one or two fingers) or patterns of co-contraction that typically appear when a subject is prompted to open or close the hand. To measure these abilities, we outfit an exotendon device with bend and pressure sensors. When EMG is insufficient to determine user intention, bend and pressure sensors can be used to control the orthosis. We refer to controls that use multiple sensor types for input as multimodal controls. The main contributions of this paper are:

  • We develop an active hand orthosis as a multimodal sensory platform as well as an actuation device, allowing us to characterize physical interaction with the user in novel ways. In particular, we incorporate bend and pressure sensors into an exotendon framework with existing EMG sensing, while keeping the orthosis compact and without impacting grasping tasks.

  • We introduce multimodal control methods for the orthosis, using the various sensors (EMG, bend, and pressure) as inputs. We then show that the different controls can be used with different impairment patterns commonly found in stroke subjects.

To the best of our knowledge, we are the first to propose intuitive multimodal control schemes for a hand orthosis which leverage natural hand movement signals (as opposed to side channels such as voice). A very recent review of more than 80 studies in this area [10] found a single device capable of multimodal intent inferral, and that was using voice as a second modality. We thus aim to bridge this gap towards reliable and intuitive control. Working with stroke patients, we show that our methods can be adapted to various impairment patterns, and can also be integrated in fully functional systems, laying the foundation for further development in this direction.

Ii Related Work

EMG is one of the most popular unimodal controls for robotic hand orthoses, as it requires relatively simple algorithms and enables intuitive operation. Most commonly, sensors are attached to the flexor and extensor muscles of the impaired arm and an open-loop control opens and closes the hand when EMG exceeds a threshold [6, 11]

. Pattern recognition algorithms are also becoming more popular as they can enable the use of commodity EMG armbands 

[7]

and classify multiple hand postures in stroke patients 

[12].

However, these algorithms often only work on a subset of stroke population due to abnormal muscle activation [9]. Several strategies have been developed to adapt to these irregular EMG patterns. One strategy is to place the sensors on muscles which retain healthy EMG patterns. For example, stroke subjects can utilize the contralateral upper extremity [13] or facial expressions [14] to trigger EMG-based controls. Both of these methods require learning a control which uses muscles unrelated to the desired task.

An alternative strategy is to develop a multimodal control that uses EMG in addition to a more robust sensing modality. The VAEDA glove uses voice recognition to specify the control mode, and EMG signals to trigger commands [15]. Voice recognition is robust in ideal conditions, but sensitive to noise. Radio frequency identification (RFID) tags on objects can serve as non-biological switches to identify desired hand postures, again using EMG as a trigger to execute these postures [16]. RFID tags predetermine which objects the subject can interact with, which limits their utility in real-world environments. Fusing mechanomyography (MMG) and EMG for prosthetic controls has been studied [17, 18]. MMG is more robust to noise than EMG, but its use for individuals with neurological impairment is largely unexplored [19].

Other studies have developed controls which rely on types of sensors other than EMG. Some of these controls are unimodal - they trigger the device using a simple analog button [4], a bend sensor on the wrist [3], body-powered motions [20], or force myography [21]. The Soft Extra Muscle Glove uses force sensitive resistors (FSRs) as a control because they provide useful information when subjects interact with objects [22]. Zhao et al. [5] integrated optical strain sensors into a rehabilitation device based on pneumatic actuation in order to provide position feedback for control and motion analysis. However, unimodal controls have not yet been shown to be robust for long-term operation, and often rely on external cues, instead of natural hand motions.

Other works have developed multimodal controls using non-EMG sensors. For example, Steinkamp, et al. [23]

use a 3D depth-camera and IMU sensors to analyze point clouds of the environment in order to classify appropriate hand assistance. Some devices use sensors not as control inputs, but as tools to analyze hand movement. The SCRIPT passive orthosis is equipped with multimodal sensors to estimate joint rotations and torques. These sensing capabilities enable interactive rehabilitation games for users 

[24, 25].

Multimodal controls for gait assistive devices are more commonplace than for hand devices. Hybrid Assistive Limb utilizes pressure sensors and potentiometers to measure joint angles for motion intent estimation [26]. Villa-Parra, et al. have developed a knee device for gait rehabilitation using EEG and EMG as a multimodal control [27]

. While multimodal intent inferral methods for lower-limb exoskeletons exist, hand devices face significantly different challenges, such as many more articulation degrees of freedom, wider variety of movement patterns, more limited space and acceptable weight, etc. This perhaps helps explain the fact that no similar multimodal controls have been introduced to date for assistive hand devices 

[10].

Iii Exotendon Device

To equip a hand orthosis with multimodal sensing, we expand upon our previous work with exotendon hand devices, specifically our work on tendon networks [1] combined with distal structures for efficient force transmission [28]. Tendon-driven systems require less space than linkage-based exoskeletons, as they utilize few, small anchoring structures. Therefore, they are well-suited for sensor implementation.

Fig. 2: Integration of a pressure sensor into a thumb strap (left), tendon routing for thumb extension and abduction (top) and bend sensor implementation (bottom).

We developed a modularized device consisting of two parts: an aluminum forearm splint with actuation and 3D-printed fingertip components for cable routing (Fig. 12). The splint constrains wrist movement so that the motor forces are transmitted to the fingers. For actuation, we use a Proportional-Integral-Derivative (PID) position controller whose range of motion is determined by user hand size. Motor extension or retraction takes approximately 1.8 seconds.

The 3D-printed fingertip components are secured to the fingertips using Velcro straps. The underside of the strap is rubber to prevent distal migration. The components route the exotendons through raised pathways that enhance force transmission by increasing the moment arm around the proximal interphalangeal (PIP) joints. In addition, the components prevent hyper-extension of the distal interphalangeal (DIP) joint and serve as an anchoring point for the tendons.

The thumb moves differently than the other four fingers, and therefore requires different routing. As long as the four fingers are sufficiently extended, we can enable grasping tasks by simply splinting the thumb in a stationary, opposed position [29]. We splint the thumb using two tendon routes which adjust the thumb’s abduction and extension (Fig. 2).

Iv Multimodal Sensing

While existing work has focused primarily on robotic hand orthoses as actuation devices, we envision future devices serving an equally important role as sensory platforms, equipped to characterize physical interaction with the user. Numerous sensing modalities can be envisioned, focusing on tendons, joints, contacts, etc. In this context, we have developed a multisensory platform prototype, combining sensors for the following: forearm EMG, motor position, fingertip pressure, and joint angles (Fig. 12). We describe these sensing modalities and their integration with the orthosis next.

Iv-1 Forearm EMG

EMG is one of the most common orthotic controls because it is intuitive. EMG sensors are low profile, and commercial devices, like the one used in this work, are easy to don and doff. With relatively simple algorithms, EMG sensors can be used to identify a variety of different hand poses.

We use the Myo Armband from Thalmic labs for our EMG sensing. The armband consists of eight EMG sensors and is placed on the subject’s forearm, proximal to the splint. Our pattern recognition algorithm (Section V-A) uses the EMG sensors to predict the user’s intended hand state. Fig. (a)a shows an example of EMG activation patterns as a subject attempts to open and close their affected hand. In this figure, the EMG activations for open and close are distinct; however, these patterns will change over time as the subject fatigues.

Iv-2 Motor Position

Motor position sensing is commonplace in robotic devices, and we include its description here for completeness. The motor encoder provides high-resolution position feedback, which enables us to control the actuator with position control and determine the current state of the orthosis. Because our tendon network is underactuated, this feedback does not provide information about individual finger behaviors, but their combined movement pattern.

Iv-3 Finger Joint Angles

Joint angles can serve as cues to determine patient intent. One typical pattern is partial voluntary movement, where patients try to open their hand and some fingers partially extend. Another, abnormal, movement pattern from which the sensing modality can potentially benefit is overactive stretch response [30], which exhibits finger flexion when patients try to extend. By measuring PIP joint angles with bend sensors, both movement patterns can give us information about user intent.

We use a bend-sensitive resistor on each finger to measure joint flexion of the PIP joint (Fig. 2). We assume residual movement of the PIP joint is greater than the MCP joint and therefore only deploy sensors on the PIP. For each finger, the proximal side of the sensor is anchored to the subject’s proximal finger link by a strap. The distal side of the bend sensor is fed through a flat hole in the bottom of the fingertip component to keep it close to the distal link of the finger.

We found that using a simple threshold on the raw bend sensor data to trigger an open command was limited as a control because motor position and the size of the objects with which the user interacts both dramatically affect the raw data values. Fig. (b)b shows the raw bend data and bend derivative during an example open-close motion. Note that the bend derivative peaks soon after the subject is asked to open. The next notable maximum is caused by the device extending the fingers.

Iv-4 Fingertip Pressure

Pressure sensors on the fingertips serve a dual role: since the digit straps are the conduit by which exotendons apply force to the fingers, the pressure sensor can record the level of force between the hand and the device. When the user is performing a grasp, the pressure sensors will also record the contact force between the hand and the object. In this way, pressure sensing allows us to paint a complete picture of force transmission, from the orthosis to the patient’s hand, and from the hand to the environment.

Fingertip pressure increases when the subject is either interacting with an object or trying to close the hand while the device is open. Though we cannot differentiate between the two actions, the increase in pressure gives us useful information about when the user intends to close their hand, especially when the user cannot maintain the muscle activation necessary for detection via EMG.

Again, we use the time derivative of the pressure data rather than the raw data. As shown in Fig. (c)c, both the raw data and the pressure derivative increase soon after the subject is asked to close their hand, but the derivative provides more robust cues because the raw data alters over time due to fatigue and irregular tone.

We fit our exotendon device with pressure sensing using force sensitive resistors (FSRs). FSRs are compact enough for integration inside the digit straps which attach the 3D printed fingertip components to the subject’s fingers. Fig. 2 shows how the FSRs are placed inside the digit straps.

For simplicity, we integrate pressure sensing only on the thumb because it is the finger which generates the greatest force when the subject tries to close their hand [31]. The thumb is also used in all gross grasping, ensuring we will see interactions between the subject and any grasped objects.

V Controls

V-a EMG Control

In previous work, we describe an EMG control which uses pattern recognition to predict user intention [7]. Here, we use the same eight sensor EMG armband (Myo) and a similar pattern recognition algorithm. Pattern recognition enables the use of commodity EMG devices, which are easier to don and doff than medical-grade sensors. Using pattern recognition, the EMG sensors do not need to be placed on specific muscles in order to identify user intention.

We place the EMG sensors on the subject’s forearm, on the same arm as the exotendon device. Ipsilateral EMG control harvests the EMG signal the user makes when they try to open or close their impaired hand, rather than requiring the user to learn an unrelated motion to control the device.

(a) Forearm EMG activation
(b) Bend sensor data and its time derivative
(c) Fingertip pressure and the time derivative of the pressure
Fig. 6: Sensor data recorded while a user with stroke opens (green) and closes (yellow) their affected hand during training. The plots depict the following sequence (verbal instructions are given to the user throughout): the user, with hand at rest, tries to open. After a few seconds, the device is actuated to assist the open. The user relaxes once the hand is open (tendon fully retracted). Then, the user tries to close, despite resistance from the orthosis. Shortly after, the device extends, allowing the hand to close.

The algorithm we use for pattern recognition is described in our previous work [7]. The main difference in this work is that we aim to predict three possible user intentions rather than two: to open the hand (Intent=Open), to close the hand (Intent=Closed), and to relax (Intent=Relaxed - newly introduced here). The addition of the Intent=Relaxed class allows the user to open the hand using the exotendon device, and then relax their hand while they are positioning their arm, for example in order to execute a pick and place task, without having to continue to exert effort to keep the hand open. We believe this approach can help avoid muscle fatigue.

To classify user intent at a given time, we input the EMG signals collected at that time into a random forest classifier. The classifier outputs three values, each being the probability that the EMG signals belong to a corresponding intent class. These three probabilities are put through a median filter (0.5 s window) in order to eliminate spurious predictions. Finally, we compare the output probabilities from the median filter to three manually set thresholds. If the probability for a class exceeds the threshold, we classify the end result as belonging to that class. The end-result belongs to either the

Intent=Open, Intent=Relaxed, or Intent=Closed class. We assign thresholds such that only one class can exceed a threshold at a time. If none of the thresholds are exceeded, the intent remains the same as at the last time step.

The EMG control can then issue motor commands to the exotendon device based on the predicted user intention. If the EMG control predicts that the user’s intention is Intent=Open, the device is commanded to open (retract the tendon, thus extending the fingers). If the user’s intention is Intent=Closed, the device is commanded to close (extend the tendon, thus allowing the user to flex the fingers). If the predicted user intention is Intent=Relaxed, we continue to send the previous motor command to the device.

V-B Multimodal Control

We propose two types of multimodal control. Subjects in our target population display a wide range of impairment patterns. Some cannot maintain a ‘close’ EMG signal, and others have more voluntary finger extension. A single sensing modality is limited due to the various impairment patterns; similarly, multimodal sensing is limited if it does not fit the subject’s impairment pattern.

We propose one kind of multimodal control where bend sensors detect the user’s intention to open the hand, and EMG sensors detect the user’s intent to close the hand. The other multimodal approach uses pressure sensors to detect the user’s intent to close the hand, and EMG sensors to detect the user’s intent to open the hand. The multimodal approach used for each of our subjects was chosen based on a qualitative analysis of their abilities, such as range of voluntary finger extension, and ability to maintain EMG signals.

V-B1 Bend to Open, EMG to Close

The first multimodal control uses bend sensors to determine when the exotendon device should open, and EMG sensors determine when the device should close. Subjects who use this control would typically have the ability to initiate finger extension, but be unable to achieve functional extension and have difficulty maintaining an ‘open’ signal for EMG.

To determine user intent based on voluntary extension, we collect data from the four bend sensors built into the orthosis. In the current version, the therapist determines which of the subject’s fingers has the greatest range of voluntary motion and we focus on bend data from that specific digit; in the future, we plan to integrate the data from all four sensors. Bend data is then passed through a moving mean filter with a window size of 0.25 s. We take the derivative of the resulting signal, which we refer to as (where the subscript denotes the digit found to have the highest voluntary range of motion). Motor commands are sent as follows:

  • When the orthosis is in the Device=Closed position (tendon extended allowing fingers to flex) and exceeds a given threshold , the device is commanded to open (retract the tendon).

  • When the orthosis is in the Device=Open position (tendon fully retracted, or motor stalled) and the EMG classifier predicts Intent=Closed, the device is commanded to close (extend the tendon).

  • If neither of the above conditions are met, we continue to send the previous motor command to the device.

The threshold is determined based on the training data collected in the procedure described in Section V-C. For the training dataset, we find the local maxima of while we ask the subject to try to open. We select the smallest value between the local maxima as . If necessary, the experimenter will manually tune the threshold so the control can enable tasks. After the threshold is set, it is kept constant throughout all tests performed by the subject.

When the device is in the Device=Closed position, EMG signals are ignored, as are bend signals when the device is in the Device=Open position. Furthermore, our control will not switch motor commands while the device is transitioning from Device=Open to Device=Closed or from Device=Closed to Device=Open. We note that although this consideration can reduce rapid oscillations in the motor command, it is limiting if the subject only wants to open their hand halfway and then close again, for example, when grasping small objects. If the subject starts closing their hand before the motor is done transitioning, they will encounter resistance from the orthosis until the transition finishes and the control issues another command to the motor.

V-B2 EMG to Open, Pressure to Close

For the second kind of multimodal control, EMG sensors determine when the exotendon device should open and the pressure sensors determine when the device should close. Subjects who use this control typically have a clear EMG muscle pattern for ‘open’ and difficulty maintaining a ‘close’ signal for EMG.

To implement this control, we use data from the thumb pressure sensor. As with bend data, the raw signal is first passed through a moving average filter with window size 0.25 s; we then compute the derivative of the output . Motor commands are sent as follows:

  • When the orthosis is in the Device=Closed position (tendon extended allowing fingers to flex) and the EMG classifier predicts Intent=Open, the device is commanded to open (retract the tendon).

  • When the orthosis is the Device=Open position (tendon fully retracted, or motor stalled) and exceeds threshold , the device is commanded to close (extend the tendon).

  • If neither of the above conditions are met, we continue to send the previous motor command to the device.

The threshold is set with a procedure similar to the one previously described for the bend threshold this time using training data while the subject is being asked to try to close. Again, we do not issue new commands while the device is transitioning between states.

V-C Training with the Exotendon Device

Stroke subjects often produce EMG patterns which change dramatically depending on arm position, even if the subject’s intention to open, relax or close the hand remains the same. These EMG patterns are further changed by the hand’s physical interaction with the exotendon device. We therefore train the subjects with their arms in different positions and the exotendon device in different states.

We design our training protocol as follows: the exotendon device starts in the closed state (tendon is fully extended) and the subject is asked to relax. Then the subject is asked to try to open their hand. The experimenter waits three seconds, and as the user continues to try to open, the experimenter opens the exotendon device (retracts the tendon) to extend the subject’s fingers. The subject continues to try to open for three seconds after the exotendon device is fully opened and is then relaxes. Next, the subject is instructed to close their hand. The experimenter waits three seconds and then closes the device. The subject continues to try and close for three seconds after the device has fully closed and then is instructed to relax. During training, subject intent, or ground truth, is given to the program by the experimenter as they simultaneously provide participants with verbal commands.

The subject repeats the above procedure five times. The first two times, the subject’s arm rests on the table, and the next three times, the subject raises their arm off the table.

Vi Experiments

We evaluate the feasibility of our multimodal controllers when used by subjects with different impairment patterns, using EMG control as a baseline. We selected patients whose EMG patterns showed signs of being abnormal, affected by fatigue and interaction with the orthosis (which, in our experience, is commonplace), but who were still able to complete pick and place tasks using EMG control. Our multimodal control is designed to be robust to different impairments, so we chose subjects with distinct patterns.

We note that, in this current version of the study, the experimenter plays the important role of selecting the appropriate control mode for a patient. We believe this approach serves to establish the feasibility of multimodal sensing in our context, but is also applicable to real-life scenarios, where an experienced clinician can make similar decisions based on patient observations. Nevertheless, we hope to automate this aspect of the procedure in future work.

Testing was performed on four chronic subjects with a spasticity level of two or less on the Modified Ashworth Scale (MAS). Subject clinical information can be found in Table I. Participants had prior experience with the exotendon device, in varying capacities. Subjects gave informed consent and all testing was approved by the Columbia University Internal Review Board, and performed in a clinical setting under the supervision of an occupational therapist.

Subject Gender Affected Limb Fugl Meyer UE Box and Blocks
A F Left 26 0
B F Right 26 0
C M Right 25 6
D M Right 23 0
TABLE I: Subject clinical information

The experimenter placed the orthosis on the subject’s hand and made any necessary sizing adjustments. The subjects were trained using the protocol described in Section V-C.

After training, we asked the subject to perform two types of testing. The first one, designed to isolate the effects of the chosen control method, consists exclusively of performing open-close hand motions. We refer to these as Controller Accuracy experiments. In these tests, we asked the subject to perform several open and close motions in order to compare the accuracy of the baseline and proposed controls. The experimenter verbally cued the subjects to open and close their hand while providing the program with ground truth for the desired motor command.

The second type of test is designed to verify that the multimodal sensory platform we have developed can be used in a functional context. We refer to these as Pick and Place experiments. Here, five blocks (1” square cubes) were placed in a square pan on a table in front of the subject. The subject was required to start with their hand in a relaxed state, grasp a block, transport it over the median with control and release it onto the tabletop. The task was considered complete when the subject activated the device to extend the digits and released the block. The therapist timed how long it took the subject to pick and place each of the five blocks. For each condition, the subject moved all five blocks three times. Patients were given sufficient time between testing procedures such that order effects which might have been induced by fatigue were negligible. Each subject was given three minutes of play time to acclimate to each control.

While we designed our pick and place task to minimize the impact of external factors on performance, the nature of functional tasks renders them replete with factors that impact performance. Even such a simple task reflects an individual’s shoulder strength, residual fingertip sensation, and grip strength and is not a pure measure of controller efficacy. The number of clinical tests needed to average out the significant effects of all of these compounding factors is beyond the scope of this paper. We therefore rely on Controller Accuracy to evaluate the proposed controls isolated from other factors, and use Pick and Place experiments simply to illustrate their feasibility in a functional context. For this reason, all subjects completed Controller Accuracy testing, but only Subjects A and B completed Pick and Place testing.

In stroke subjects, fatigue and abnormal coactivation [32] can cause EMG patterns to change over time. To study these effects, we also asked the subjects showing most pronounced effects of fatigue and abnormal co-activation (subjects B and D, as observed by the experimenter) to perform all experiments a second time, in a different condition: wearing an arm support system which aids arm movement through gravity compensation. Testing with and without the arm support system helps us evaluate when the multimodal control is most effective.

During testing, subjects were unaware of the control mode they were using. The controls should be intuitive, so subjects were merely instructed to try to open and close their hand.

Vii Results and Discussion

For Controller Accuracy testing, we compare the output of the EMG and multimodal controls to the ground truth provided by the experimenter. At each time point, the controls can correctly predict an open (true positive), correctly predict a close (true negative), incorrectly predict an open (false positive) or incorrectly predict a close (false negative). We report the global accuracy, the positive predictive value (PPV), and the negative predictive value (NPV) for our classifiers [33]. Global accuracy is the number of true predictions (positive or negative), divided by the number of total predictions. PPV is the number of true positives divided by the number of all positive predictions (whether true or false). NPV is the number of true negatives divided by the number of all negative predictions.

Transitions
Condition Control Type Global Accuracy PPV NPV Correct False
Regular EMG 77.9% 77.1% 78.8% 4/7 0.5
Multimodal 83.4% 81.6% 84.9% 6/7 1.5
With arm support EMG 85.2% 85.6% 84.8% 8/9 2
Multimodal 86.0% 86.6% 85.6% 8/9 0
TABLE II: Results for controller accuracy

Global accuracy can be misleading for EMG pattern recognition controls [33], so we believe another important metric is the ability to correctly identify transitions between motor commands. A transition is defined as a change in motor command, and a correctly identified transition means a predicted transition which occurs within 1.5 seconds of the ground truth transition. 1.5 seconds allows enough time for the experimenter to give the verbal command and for the subject to react and start performing the motion. (We found that the subjects would often raise their arm off the table before they attempted the instructed hand motion, which increased reaction time.) The correct transitions are reported with the total number of ground truth transitions. Success for this metric is a number of correctly identified transitions that is close or equal to the total number of ground truth transitions. We also report the number of false transitions, or transitions which do not have a corresponding ground truth transition. These transitions cause motor oscillations, confusing and frustrating the user. Success for this metric is a number of false transitions close to zero.

The results for the Controller Accuracy experiments are shown in Table II

. We averaged across subjects and show results for experiments performed with arm support (Subjects B and D) and without arm support (all participants). For the Pick and Place testing, we report the time to pick each block and the total time to pick all five blocks, averaged across three trials, and standard error (Table 

III). We show results for subjects with arm support (Subject B) and the average result without arm support (Subjects A and B).

Condition Control Type Each Block Total
Regular EMG 13.41.4 66.85.6
Multimodal 15.72.8 78.48.4
With arm support EMG 30.16.5 150.754.6
Multimodal 24.33.8 121.532.4
TABLE III: Results for pick and place tasks

In Controller Accuracy testing, in the Regular condition (no arm support) multimodal control consistently outperformed EMG control. Both with and without arm support, the global accuracy, PPV, NPV, and the number of correctly identified transitions within the given time window were all higher for the multimodal control than for the EMG control. EMG control only outperformed multimodal control in the number of false positive transitions predicted.

We believe these results show that the proposed multimodal control can be effective for subjects with a variety of different impairment patterns. Subjects A and D have almost no voluntary extension in their fingers and have difficulty maintaining a ‘close’ EMG signal. On the other hand, Subjects B and C can extend their fingers partially, but have a hard time maintaining an EMG signal for ‘open’. Our multimodal control can be customized to these different impairment patterns and enables effective orthosis control for both patterns.

With arm support, the accuracy of the two control methods is much closer. The global accuracy, PPV and NPV are all within 1% of each other for the two controls. We hypothesize that arm support relieves abnormal muscle coactivation experienced by the subject, and that when this coactivation diminishes, the subjects have an easier time maintaining their EMG signals. These findings tell us two things: first, they illustrate just how varied post-stroke impairment patterns can be. The same EMG control had up to a 8.5% increase in performance when we started providing arm support. Such a significant change in a subject-driven control for the same patient underlines the need for controls which can adapt to a wide range of impairment patterns, both between patients and as subjects undergo rehabilitation. Second, they can tell us where our multimodal control is most useful. With arm support, the multimodal control did not help the patient significantly more than the EMG control. However, it did help subjects more when they were not provided arm support. We conclude that our multimodal control is best suited for patients who experience fatigue easily and who experience a significant amount of abnormal muscle coactivation.

In the pick and place experiments, multimodal control was more efficient than EMG control when used with an arm support, while EMG control was more efficient with arm support. However, the small sample size and the large number of additonal factors that affect functional performance (e.g. arm strength and control, fingertip sensation, grip strength, chosen task strategy, etc.), prevent us from drawing any quantitative conclusions. We believe, however, that these results show that a multimodal sensory platform can be integrated in a complete functional task, highlighted by the fact that all participants completed the task using both control mechanisms, despite having ‘no to poor’ upper extremity capacity (as defined in [34]).

Viii Conclusion and Future Work

In this paper, we incorporate EMG, bend, and pressure sensors into an exotendon framework to create a multimodal sensing and interaction platform for a hand orthosis. We believe the future of robust controls for orthoses involves multiple sensing modalities which complement each other to inform controls. The bend and pressure sensors give us information about user intent if subjects display certain impairment patterns we have observed in many stroke patients.

We propose two multimodal control modes, tailored to the different impairment patterns we have observed. Controls that can cope with many impairment patterns are necessary because these patterns vary across subjects; one patient could even display several patterns as they undergo therapy post-stroke. This is a preliminary study with a limited sample size; however, our results show that multimodal controls can be adapted to different impairment patterns and can help functional tasks. This is a first step towards the development of robust, flexible controls, which could play an important part in deploying robotic rehabilitation to a large population of stroke patients.

In the future, we would like to add more sensing modalities to our device, such as bend sensors on the MCP joints and IMU sensors to provide information about finger positions. IMUs and bend sensors have different limitations and strengths and will complement each other to jointly characterize finger movement. We also believe that a multimodal sensing platform, like the one developed here, could use its sensors not only for control, but also to track subject progress and rehabilitation. This will enhance our understanding of phenomena such as muscle spasticity and abnormal muscle synergies. It could also help us understand how impairment patterns develop over time as patients undergo rehabilitation.

To expand our study of robust multimodal controls, in the future, we would like to develop a control which uses inputs from all sensor types simultaneously. We believe that the future of these multimodal controls lies in the sensors’ ability to complement and augment each other, and that such a control can continuously adapt and learn to predict user intent. The multimodal control predicted signals that the EMG-only classifier missed (i.e. the correctly predicted transitions in Table II). This suggests that our multimodal control could be used to continue training the EMG classifier during real-time operations. Such continuous adaptation will also play an important role as the field transitions from controlled sessions to in-home environments.

References

  • [1] S. Park, L. Bishop, T. Post, Y. Xiao, J. Stein, and M. Ciocarlie, “On the feasibility of wearable exotendon networks for whole-hand movement patterns in stroke patients,” in Robotics and Automation (ICRA), 2016 IEEE Intl. Conf. on.   IEEE, 2016, pp. 3729–3735.
  • [2] S. Biggar and W. Yao, “Design and evaluation of a soft and wearable robotic glove for hand rehabilitation,” IEEE Transactions on Neural Systems and Rehab. Engineering, vol. 24, pp. 1071–1080, 2016.
  • [3] H. In, B. B. Kang, M. Sin, and K.-J. Cho, “Exo-glove: a wearable robot for the hand with a soft tendon routing system,” IEEE Robot Autom Mag, vol. 22, no. 1, 2015.
  • [4] B. B. Kang, H. Lee, H. In, U. Jeong, J. Chung, and K.-J. Cho, “Development of a polymer-based tendon-driven wearable robotic hand,” in IEEE Intl. Conf. on Robotics and Automation, 2016.
  • [5] H. Zhao, J. Jalving, R. Huang, R. Knepper, A. Ruina, and R. Shepherd, “A helping hand: Soft orthosis with integrated optical strain sensors and emg control,” IEEE Robot Autom Mag, vol. 23, no. 3, 2016.
  • [6] P. Polygerinos, K. C. Galloway, S. Sanan, M. Herman, and C. J. Walsh, “Emg controlled soft robotic glove for assistance during activities of daily living,” in IEEE Intl. Conf. on Rehabilitation Robotics, 2015.
  • [7] C. Meeker, S. Park, L. Bishop, J. Stein, and M. Ciocarlie, “Emg pattern classification to control a hand orthosis for functional grasp assistance after stroke,” in IEEE Intl. Conf. on Rehabilitation Robotics, 2017.
  • [8] B. Cesqui, P. Tropea, S. Micera, and H. I. Krebs, “Emg-based pattern recognition approach in post stroke robot-aided rehabilitation: a feasibility study,” Journal of neuroengineering and rehabilitation, vol. 10, no. 1, p. 75, 2013.
  • [9] J. M. Ochoa, Y. J. D. Narasimhan, and D. G. Kamper, “Development of a portable actuated orthotic glove to facilitate gross extension of the digits for therapeutic training after stroke,” in Engineering in Medicine and Biology Society, 2009. EMBC 2009. Annual Intl. Conf. of the IEEE.   IEEE, 2009, pp. 6918–6921.
  • [10] C.-Y. Chu and R. M. Patterson, “Soft robotic devices for hand rehabilitation and assistance: a narrative review,” J Neuroeng Rehabil, vol. 15, no. 1, 2018.
  • [11] X. Hu, K. Tong, X. Wei, W. Rong, E. Susanto, and S. Ho, “The effects of post-stroke upper-limb training with an electromyography (emg)-driven hand robot,” Journal of Electromyography and Kinesiology, vol. 23, no. 5, pp. 1065–1074, 2013.
  • [12] Z. Lu, R. K.-y. Tong, X. Zhang, S. Li, and P. Zhou, “Myoelectric pattern recognition for controlling a robotic hand: A feasibility study in stroke,” IEEE Transactions on Biomedical Engineering, 2018.
  • [13] L. Lucas, M. DiCicco, and Y. Matsuoka, “An emg-controlled hand exoskeleton for natural pinching,” J. of Robot. and Mechatronics, vol. 16, pp. 482–488, 2004.
  • [14] I. Hussain, G. Spagnoletti, G. Salvietti, and D. Prattichizzo, “Toward wearable supernumerary robotic fingers to compensate missing grasping abilities in hemiparetic upper limb,” The Intl. Journal of Robotics Research, vol. 36, no. 13-14, pp. 1414–1436, 2017.
  • [15] K. O. Thielbar, K. M. Triandafilou, H. C. Fischer, J. M. O’Toole, M. L. Corrigan, J. M. Ochoa, M. E. Stoykov, and D. G. Kamper, “Benefits of using a voice and emg-driven actuated glove to support occupational therapy for stroke survivors,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 25, no. 3, pp. 297–305, 2017.
  • [16] H. K. Yap, B. W. Ang, J. H. Lim, J. C. Goh, and C.-H. Yeow, “A fabric-regulated soft robotic glove with user intent detection using emg and rfid for hand assistive application,” in IEEE Int. Conf. on Robotics and Automation, 2016.
  • [17] W. Guo, X. Sheng, H. Liu, and X. Zhu, “Mechanomyography assisted myoeletric sensing for upper-extremity prostheses: A hybrid approach,” IEEE Sensors Journal, vol. 17, pp. 3100–3108, 2017.
  • [18] A. Wołczowski and R. Zdunek, “Electromyography and mechanomyography signal recognition: experimental analysis using multi-way array decomposition methods,” Biocybern Biomed Eng, vol. 37, no. 1, 2017.
  • [19] M. O. Ibitoye, N. A. Hamzaid, J. M. Zuniga, and A. K. A. Wahab, “Mechanomyography and muscle function assessment: A review of current state and prospects,” Clin Biomech, vol. 29, no. 6, 2014.
  • [20] X. Luo, T. Kline, H. C. Fischer, K. A. Stubblefield, R. V. Kenyon, and D. G. Kamper, “Integration of augmented reality and assistive devices for post-stroke hand opening rehabilitation,” in IEEE-EMBS Intl. Conf. on Engineering in Medicine and Biology Society, 2005.
  • [21] H. K. Yap, A. Mao, J. C. Goh, and C.-H. Yeow, “Design of a wearable fmg sensing system for user intent detection during hand rehabilitation with a soft robotic glove,” in IEEE Intl. Conf. on Biomedical Robotics and Biomechatronics, 2016.
  • [22] M. Nilsson, J. Ingvast, J. Wikander, and H. von Holst, “The soft extra muscle system for improving the grasping capability in neurological rehabilitation,” in IEEE-EMBS Conf. on Biomed Eng Sci, 2012.
  • [23] J. M. Steinkamp, L. J. Brattain, C. J. Walsh, and R. D. Howe, “Point cloud context analysis for rehabilitation grasping assistance.”
  • [24] S. Ates, C. J. Haarman, and A. H. Stienen, “Script passive orthosis: design of interactive hand and wrist exoskeleton for rehabilitation at home after stroke,” Autonomous Robots, vol. 41, pp. 711–723, 2017.
  • [25] F. Amirabdollahian, S. Ates, A. Basteris, A. Cesario, J. Buurke, H. Hermens, D. Hofs, E. Johansson, G. Mountain, N. Nasr, et al., “Design, development and deployment of a hand/wrist exoskeleton for home-based rehabilitation after stroke-script project,” Robotica, vol. 32, no. 8, pp. 1331–1346, 2014.
  • [26] K. Suzuki, G. Mito, H. Kawamoto, Y. Hasegawa, and Y. Sankai, “Intention-based walking support for paraplegia patients with robot suit hal,” Advanced Robotics, vol. 21, no. 12, pp. 1441–1469, 2007.
  • [27] A. Villa-Parra, D. Delisle-Rodríguez, A. López-Delis, T. Bastos-Filho, R. Sagaró, and A. Frizera-Neto, “Towards a robotic knee exoskeleton control based on human motion intention through eeg and semgsignals,” Proced. Manuf, vol. 3, pp. 1379–1386, 2015.
  • [28] S. Park, L. Weber, L. Bishop, J. Stein, and M. Ciocarlie, “Design and development of effective transmission mechanisms on a tendon driven hand orthosis for stroke patients,” in Robotics and Automation (ICRA), 2018 IEEE Intl. Conf. on.   IEEE, 2018, pp. 2281–2287.
  • [29] J. Arata, K. Ohmoto, R. Gassert, O. Lambercy, H. Fujimoto, and I. Wada, “A new hand exoskeleton device for rehabilitation using a three-layered sliding spring mechanism,” in Robotics and Automation (ICRA), 2013 IEEE Intl. Conf. on.   IEEE, 2013, pp. 3902–3907.
  • [30] D. G. Kamper, H. C. Fischer, E. G. Cruz, and W. Z. Rymer, “Weakness is the primary contributor to finger impairment in chronic stroke,” Archives physical medicine and rehab., vol. 87, pp. 1262–1269, 2006.
  • [31] A. Kargov, C. Pylatiuk, J. Martin, S. Schulz, and L. Döderlein, “A comparison of the grip force distribution in natural hands and in prosthetic hands,” Disability and Rehab., vol. 26, no. 12, 2004.
  • [32] J. P. Dewald, P. S. Pope, J. D. Given, T. S. Buchanan, and W. Z. Rymer, “Abnormal muscle coactivation patterns during isometric torque generation at the elbow and shoulder in hemiparetic subjects,” Brain, vol. 118, no. 2, pp. 495–510, 1995.
  • [33] M. Ortiz-Catalan, F. Rouhani, R. Brånemark, and B. Håkansson, “Offline accuracy: a potentially misleading metric in myoelectric pattern recognition for prosthetic control,” in IEEE Eng Med Biol Soc.   IEEE, 2015.
  • [34] B. Singer and J. Garcia-Vega, “The fugl-meyer upper extremity scale,” J Physiother, vol. 63, no. 1, p. 53, 2017.