Sensing motion and muscle activity for feedback control of functional electrical stimulation: Ten years of experience in Berlin

ARTICLE IN PRESS

JID: JARAP

[m5G;October 5, 2017;12:12]

Annual Reviews in Control 0 0 0 (2017) 1–20

Contents lists available at ScienceDirect

Annual Reviews in Control journal homepage: www.elsevier.com/locate/arcontrol

Review article

Sensing motion and muscle activity for feedback control of functional electrical stimulation: Ten years of experience in Berlin Thomas Schauer∗ Technische Universität Berlin, Control Systems Group, Einsteinufer 17, Berlin 10587, Germany

a r t i c l e

i n f o

Article history: Received 15 July 2017 Revised 25 September 2017 Accepted 25 September 2017 Available online xxx Keywords: Functional electrical stimulation Neuro-prosthetics Bioimpedance Inertial measurement units Electromyography Iterative learning control Gait Cycling Reaching Swallowing Signal processing

a b s t r a c t After complete or partial paralysis due to stroke or spinal cord injury, electrical nerve stimulation can be used to artificially generate functional muscle contractions. This technique is known as Functional Electrical Stimulation (FES). In combination with appropriate sensor technology and feedback control, FES can be empowered to elicit also complex functional movements of everyday relevance. Depending on the degree and phase of impairment, the goal may be temporary support in a rehabilitation phase, e.g. during re-learning of gait after a stroke, or permanent replacement/support of lost motor functions in form of assistive devices often referred to as neuro-prostheses. In this contribution a number of real-time capable and portable approaches for sensing muscle contractions and motions are reviewed that enable the realization of feedback control schemes. These include inertial measurement units (IMUs), electromyography (EMG), and bioimpedance (BI). This contribution further outlines recent concepts for movement control, which include e.g. cascaded control schemes. A fast inner control loop based on the FES-evoked EMG directly controls the amount of recruited motor units. The design and validation of various novel FES systems are then described that support cycling, walking, reaching, and swallowing. All methods and systems have been developed at the Technische Universität Berlin by the Control Systems Group within the last 10 years in close cooperation with clinical and industrial partners. © 2017 Elsevier Ltd. All rights reserved.

Contents 1.

2.

3.

∗

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1. Stroke and spinal cord injury. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2. Functional electrical stimulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3. Underlying principles of FES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4. Why do we need sensors and feedback control in FES? . . . . . . . . . . . . . . . . Sensing human movements and muscle activity . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1. Inertial measurement unit (IMU)-based motion analysis . . . . . . . . . . . . . . . 2.1.1. Orientation estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.2. Joint angle estimation with exploitation of kinematic constraints . 2.1.3. Gait phase detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2. Electromyography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3. Bioimpedance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4. How does FES control benefit from the new measurements? . . . . . . . . . . . Feedback-controlled neuro-prostheses with integrated sensors . . . . . . . . . . . . . . . 3.1. IMU-based FES cycling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2. Adaptive drop foot stimulator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3. FES-based arm weight relief . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4. BI/EMG-triggered FES to support swallowing . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...

... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...

. . . . . . . . . . . . . . . . . .

.. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..

. . . . . . . . . . . . . . . . . .

... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...

... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...

. . . . . . . . . . . . . . . . . .

... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...

... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...

... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...

. . . . . . . . . . . . . . . . . .

... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...

... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...

. . . . . . . . . . . . . . . . . .

.. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..

. . . . . . . . . . . . . . . . . .

.. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..

2 2 2 3 4 4 4 5 5 7 7 9 10 10 10 11 12 13

Corresponding author. E-mail address: [email protected]

https://doi.org/10.1016/j.arcontrol.2017.09.014 1367-5788/© 2017 Elsevier Ltd. All rights reserved.

Please cite this article as: T. Schauer, Sensing motion and muscle activity for feedback control of functional electrical stimulation: Ten years of experience in Berlin, Annual Reviews in Control (2017), https://doi.org/10.1016/j.arcontrol.2017.09.014

ARTICLE IN PRESS

JID: JARAP 2

[m5G;October 5, 2017;12:12]

T. Schauer / Annual Reviews in Control 000 (2017) 1–20

4. Discussion and 5. Conclusions . . . Funding sources . . . Acknowledgments . References . . . . . . .

directions of ........... ........... ........... ...........

future research . .............. .............. .............. ..............

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

1. Introduction 1.1. Stroke and spinal cord injury Stroke represents one of the major causes of long-term disability world wide (Mackay, Mensah, Mendis, & Greenlund, 2004). Population growth and aging play an important role in the recently observed increase in stroke burden (Feigin et al., 2015). According to the WHO estimates, the number of stroke events in the EU countries, Iceland, Norway, and Switzerland is likely to increase from 1.1 million per year in 20 0 0 to more than 1.5 million per year in 2025 solely because of demographic changes (Truelsen et al., 2006). A stroke is either caused by an interruption of the blood supply to the brain (ischemic strokes) or by rupture of a blood vessel or an abnormal vascular structure that lead to bleeding into or around the brain (hemorrhagic strokes). More than two-thirds of the strokes are ischemic. Generally, stroke can result in five types of disabilities: paralysis or problems controlling movement (motor control); sensory disturbances including pain (loosing the ability to feel touch, pain, temperature, or position); problems using or understanding language (aphasia); problems with thinking and memory; and emotional disturbances. The paralysis is generally on the side of the body opposite to the side of the brain injured by the stroke. It may have an effect on an arm, a leg, the face, or the entire side of the body. This one-sided paralysis is named hemiplegia (one-sided weakness is known as hemiparesis). Stroke survivors with hemiparesis or plegia may have difficulty with everyday activities such as walking or grasping objects. About 25 to 50% of chronic stroke patients have additional problems with swallowing, called dysphagia (Singh & Hamdy, 2006). Damage to the cerebellum, the lower part of the brain, can impair the ability to coordinate movement. This disability is called ataxia and leads to problems with body posture, walking, and balance. Spasticity following a stroke occurs in about 30% of the cases and usually occurs within the first few days or weeks (Thibaut et al., 2013). According to Pandyan et al. (2005) it can be defined as “disordered sensory-motor control, resulting from an upper motor neuron lesion, presenting as intermittent or sustained involuntary activation of muscles”. In the upper limbs, the most frequent pattern of arm spasticity after stroke is internal rotation and adduction of the shoulder together with flexion at the elbow, the wrist and the fingers. In the lower limbs, adduction and extension of the knee with pointed foot and curling toes is often present (Thibaut et al., 2013). Spasticity in stroke patients may hinder functional movements like reaching and grasping but often enables walking after paralysis due to knee extension. Another major cause of disabilities are injuries of the spinal cord that may result from physical trauma such as car accidents or from non-traumatic reasons such as tumors. The prevalence of traumatic spinal cord injuries (SCI) is highest in the United States of America (906 per million) (Singh, Tetreault, Kalsi-Ryan, Nouri, & Fehlings, 2014). Worldwide, the majority of studies on prevalence and incidence of traumatic spinal cord injury show a high male-to-female ratio and an age of peak incidence of younger than 30 years old. Traffic accidents were typically the most common cause of traumatic SCI, followed by falls in the elderly population and violence injuries (Singh et al., 2014). A damage of the spinal cord may lead to complete or incomplete loss of sensation

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

15 16 16 16 16

and motor function below the level of lesion and often results in the immobilization of the patient and possible lifelong dependency on a wheelchair. Ranging from high to low levels of lesions, one can classify the following locations of SCI: cervical (vertrebra C1–C8), thoracic (vertrebra T1–T12), lumbar (vertrebra L1–L5), or sacral (vertrebra S1–S5). The consequences of a SCI are more severe for high levels. A damage at a C1–C8 level also affects respiratory function. A lesion at thoracic, lumbar, or sacral levels can result in paraplegia, which affects the control of both legs, while a damage at a cervical level will typically affect all four limbs (tetraplegia). A complete loss of bladder, bowel and sexual function is also common. In the course of time, the primary effects of a spinal cord injury lead to a range of secondary medical complications, e.g. atrophy of the paralyzed muscles and decreased cardiovascular fitness. 1.2. Functional electrical stimulation Stroke and the majority of SCIs belong to the class of upper motor neuron lesions where the signal path from the Central Nervous System (CNS) to the muscles is interrupted above the anterior horn cell. The lower motor neuron which leaves the central nervous system from the spinal cord at the anterior horn cell to establish a functional connection with an effector (muscle) is still intact, and the muscles themselves retain their ability to contract and produce force. Functional Electrical Stimulation (FES) applied to the lower motor neurons can therefore be used to replace the lacking signals from the CNS (see Fig. 1). The underlying neurophysiological principle of FES is the generation of action potentials in the intact peripheral neurons by application of low levels of pulsed electrical current to the nerves. Muscle contractions can be artificially induced by direct electrical stimulation of efferent (motor) nerves innervating the paralyzed muscles or indirectly by electrical stimulation of afferent (sensory) nerves provoking reflexes via intact reflex arcs. An example for using reflexes is the painful stimulation of the foot sole to elicit the so-called withdrawal reflex (flexing the leg of the stimulated body side and extending the opposite leg). A neuro-prosthesis can be used to restore motor function in patients on the basis of functional electrical stimulation. Applications of FES in paraplegia with the aim of motor function restoration include walking (e.g. Graupe, 2002; Graupe, Cerrel-Bazo, Kern, & Carraro, 2008; Ho et al., 2014), cycling (e.g. Newham & Donaldson, 2007) and rowing (e.g. Hettinga & Andrews, 2007). The major aim in tetraplegia is to restore reaching and grasping function (e.g. Ho et al., 2014; Popovic, Popovic, & Keller, 2002; Mangold, Keller, Curt, & Dietz, 2005; Patil, Raza, Jamil, Caley, & O’Connor, 2014). Apart from provoking contractions of skeletal muscles, FES is used in other neuroprosthetic devices for SCI, e.g. the phrenic pacemaker and the sacral anterior root stimulator for bladder control (see (Ho et al., 2014; Peckham & Knutson, 2005; Rushton, 1997; Stein, Peckham, & Popovic, 1992) for an overview). Beyond the direct functional motor effects, therapeutic effects of FES on the secondary medical complications that arise from SCI have been reported (e.g. Hunt, Fang, Saengsuwan, Grob, & Laubacher, 2012; Janssen, Glaser, & Shuster, 1998; Daly et al., 1996; Davis, Hamzaid, & Fornusek, 2008). The benefits from FES may include for example improved muscle tone, bulk, and strength, reduced spasticity, improved limb blood flow, or a reduction in disuse osteoporosis. Additionally to these peripheral adaptations, central adaptations of

Please cite this article as: T. Schauer, Sensing motion and muscle activity for feedback control of functional electrical stimulation: Ten years of experience in Berlin, Annual Reviews in Control (2017), https://doi.org/10.1016/j.arcontrol.2017.09.014

ARTICLE IN PRESS

JID: JARAP

[m5G;October 5, 2017;12:12]

T. Schauer / Annual Reviews in Control 000 (2017) 1–20

CNS

3

reference trajectory stroke supervision disrupted signal path

(e.g. BCI, voice control, residual motor control)

feedback controller

current amplitude (0...120 mA)

pulsewidth (0...500 µs)

SCI stimulator upper motor neuron

electrical pulses Time peripheral efferent nerve (lower motor neuron)

sensory information peripheral afferent (sensory) nerves Inertial Measurement Unit (IMU)

1/frequency (1/25Hz) biphasic current pulses

Fig. 1. Principle of feedback-controlled Functional Electrical Stimulation (FES) (adapted from Schauer, 2006).

the cardiovascular system can be achieved by FES-induced exercises such as regular FES cycling or rowing. Improvements in cardiopulmonary fitness are also helping to reduce the risk of the cardio-pulmonary secondary medical conditions that commonly accompany the disability. Finally, the recreational aspects of mobile FES cycling and rowing are worth mentioning. For stroke patients, FES predominately gives temporary support in the rehabilitation phase to re-learn lost motor function such as walking (see e.g. Kafri & Laufer, 2015; Pereira, Mehta, McIntyre, Lobo, & Teasell, 2012), reaching, and grasping (e.g. Thrasher, Zivanovic, McIlroy, & Popovic, 2008). As reported in Vafadar, Côté, and Archambault (2015), FES can further be used to prevent or reduce shoulder subluxation early after stroke. The clinical effectiveness of FES is supported by recent meta analyses (see e.g. Eraifej, Clark, France, Desando, & Moore, 2017; Howlett, Lannin, Ada, & McKinstry, 2015). Clinical studies on neuro-plasticity emphasize the role of goal-oriented, repetitive movements in the motor relearning process. The increased afferent feedback provided by FES is known to modulate motor cortex function and excitability to enable recovery (Ridding, Brouwer, Miles, Pitcher, & Thompson, 20 0 0). Recent studies (Barsi, Popovic, Tarkka, Sinkjaer, & Grey, 2008; Gandolla et al., 2014) advocate the use of FES coincidentally with the voluntary drive to enhance therapeutic effects. Ambulatory chronic stroke patients with impaired foot lift can benefit from a drop foot stimulator (Dunning, O’Dell, Kluding, & McBride, 2015; Sabut, Bhattacharya, & Manjunatha, 2013). Such a device, which actually represents the first FES system in history, was proposed in 1961 by Liberson, Holmquest, Scot, and Dow (1961) and stimulates the peroneal nerve during the swing phase of gait to dorsiflex the foot. 1.3. Underlying principles of FES The stimulation pulses can be applied via implanted electrodes, percutaneously using needle electrodes inserted through the skin, or transcutaneously by attaching self-adhesive hydro-gel electrodes to the skin (see Fig. 2 A). The charge of the applied pulses is usually balanced by using biphasic stimulation pulses. A problem of traditional surface (skin) electrodes is their limited specificity when activating small muscle groups especially on the forearm.

A)

B)

stimulation electrodes EMG electrodes reference electrode (for EMG) Fig. 2. Surface electrodes. (A) A pair of hydro-gel electrodes for stimulation and a pair of AgCl electrodes for EMG measurement with additional reference electrode, (B) Electrode array on flexible Printed Circuit Board (PCB) for a grasping neuroprosthesis from the German research project BeMobil.

In order to generate different types of hand grips and hand gestures, electrode arrays have been employed recently (see Fig. 2 B). These arrays consists of many small electrode contacts that can be dynamically connected by a multiplexer to form so-called virtual electrodes. A recent review on electrode array technology can be found in Koutsou, Moreno, del Ama, Rocon, and Pons (2016). The majority of stimulators are current-controlled to obtain a consistent stimulation effect also when the electrode-skin impedance changes. When using surface electrodes and stimulating SCI patients with complete paralysis, current pulses with a pulsewidth pw of up to 500 μs and a current amplitude I of up to 150 mA are typically applied (cf. Fig. 1). The muscle force produced by FES depends on the number of recruited motor units and the activation rate. A motor unit is a single lower motor neuron and the group of muscle fibers (of the same type) innervated by it. When the pulse charge q (pulsewidth × pulse amplitude) is sufficiently high and the neuron is close to the electrodes, the motor neuron will be depolarized above threshold and an electric action potential will be released. The muscle force increases with the number of recruited motor units (spatial summation). Therefore, modulation of pulsewidth or pulse amplitude can be used to control the induced muscle force. The muscle force can also be controlled by modulation of the stimulation frequency (temporal summation). The traditional method for neu-

Please cite this article as: T. Schauer, Sensing motion and muscle activity for feedback control of functional electrical stimulation: Ten years of experience in Berlin, Annual Reviews in Control (2017), https://doi.org/10.1016/j.arcontrol.2017.09.014

JID: JARAP 4

ARTICLE IN PRESS

[m5G;October 5, 2017;12:12]

T. Schauer / Annual Reviews in Control 000 (2017) 1–20

romuscular stimulation employs a train of brief rectangular biphasic pulses, at a constant frequency between 25 and 100 Hz. Lower stimulation frequencies ( < 25 Hz) produce unfused twitches rather than a smooth muscular contraction. With increasing frequencies, the muscle is also subjected to fatigue earlier for prolonged stimulation. Stimulus pulse trains with constant frequency are usually applied in practice because they are easy to generate. However, it has been long known that natural nerve impulse trains are not of constant frequency but for example often contain a brief highfrequency burst, slowing to a lower sustained frequency. Such Variable Frequency Trains (VFT) take advantage of the catch-like property of the skeletal muscle and augment muscle performance compared with Constant Frequency Trains (CFT), especially in fatigued muscles (S. Binder-MacLeod & T. Kesar, 2005). Although FES can elicit strong and effective muscle contractions, there are some significant limitations. Normally, muscles contain a mixed population of slow fatigue-resistant (type 1), fast fatigue-resistant (type 2A) and fast fatiguable (type 2B) motor unit types. The terms fast and slow refer to the contractile speed of the muscle fibers. Muscle atrophy by disuse of the muscle tends to revert the fibre population to type 2B. Chronic electrical stimulation can be used to convert fast fatiguable muscles to fatigueresistant type 1, but training for fatigue resistance takes several weeks (Jarvis, 1993). Compared with the physiological recruitment order (the Hennemann size principle (Henneman, 1981)), recruitment with FES is thought to be inverted (Gorman & Mortimer, 1983). When low muscle forces are desired, and thus low intensity electrical stimulation pulses are applied, mainly rapidly fatiguing large motor units close to the electrodes are activated. This is because the 2B motor units are associated with large-diameter nerve axons, which have a lower firing threshold to externally applied stimulation. With increasing intensity of the pulses, also small neurons (related to slow fatigue-resistant (type 1) and fast fatigue-resistant (type 2A) motor units) with higher firing threshold as well as neurons which are located further away from the electrodes are recruited. In addition to this, the same motor units are activated all the time synchronously when the stimulation intensity and electrode position are not changing. This is completely different from the physiological control of the CNS which activates motor units asynchronously for sharing the work load. Fig. 3 illustrates these differences in motor unit recruitment for muscle contractions under FES and voluntary control. All the differences outlined above between artificial and natural nerve activation result in a much earlier fatigue of FES induced muscle contractions. Another limitation when using surface electrodes is that FES may also trigger action potentials in afferent (sensory) nerves, causing discomfort at medium and pain at high stimulation intensities if patients have some remaining sensation in the motor impaired limbs (generally after stroke). This may limit the tolerated stimulation intensity and therefore force generation. In most subjects, however, the sensation is weak enough to allow the generation of functional movements without discomfort.

Natural (physiological normal) motor unit recruitment

skin

Inverted recruitment order during externally applied electrical stimulation

time instant 1: low activation level low stim. intensity

nerve bundle

time instant 2: low activation level low stim. intensity

time instant 3: high activation level high stim. intensity

activated type 1 or 2A motor unit (fatigue resistent) activated type 2B motor unit (fast fatiguable) Fig. 3. Natural Schauer, 2006).

versus

artificial

muscle

activation

by

FES

(adapted

from

tions (e.g. causing exaggerated foot lift in a drop foot stimulator). While this strategy may provide a certain amount of safety and functionality, it will accelerate muscular fatigue. The described challenges can be faced in a much more elegant and effective way by the use of feedback control. The stimulation parameters can be adjusted automatically to the needs of the user in order to delay the onset of fatigue and to induce desired movements in an optimal way. This requires the measurement of the human motion and muscle contraction via lightweight, portable and real-time-capable measurement systems. Such measurements may additionally be used to assess the patient’s contribution to the movement. This information might be fed back to the patient to increase his/her awareness of the own motor control (biofeedback). Unfortunately, most commercially available and clinically applied FES systems have been solely based on open-loop architectures, or they only use sensors to time the stimulation. In Section 2 of this review, I will describe three promising approaches for sensing human movement as well as muscle activity and outline recent methods for processing these measurements. The integration of the measurements into cascaded FES control systems will be discussed in Section 2.4. In Section 3 the practical use of introduced sensing approaches in four different FES systems will be presented. Section 4 provides a discussion and outlines directions of future research. Finally, conclusions are given in Section 5.

2. Sensing human movements and muscle activity 1.4. Why do we need sensors and feedback control in FES? 2.1. Inertial measurement unit (IMU)-based motion analysis Functional electrical stimulation is a highly nonlinear, timevarying and uncertain dynamic process (Lynch, Popovic, & Rushton, 2008). The resulting movement strongly depends on the muscular state of fatigue, the additional voluntary contribution by the patient and the level of spasticity. The two latter are hardly predictable and may continuously change during the application of FES. Slightly different electrode placements from day to day will further cause significant deviations in the motor responses. An obvious escape strategy that is often pursued is to choose larger stimulation intensities and to accept exaggerated muscle contrac-

Inertial Measurement Units (IMUs), also known as inertial sensors, have been primarily used in Inertial Navigation System (INS) to estimate the motion (position and velocity) and orientation of vehicles such as ships, aircrafts, submarines, and spacecrafts. Due to the tremendous miniaturization of micro-electro-mechanical systems and electronic circuits, these sensors have become so small and affordable that they are nowadays found in every smartphone. This development enables completely new means for realtime human motion analysis. Further integration of electronics will

Please cite this article as: T. Schauer, Sensing motion and muscle activity for feedback control of functional electrical stimulation: Ten years of experience in Berlin, Annual Reviews in Control (2017), https://doi.org/10.1016/j.arcontrol.2017.09.014

ARTICLE IN PRESS

JID: JARAP

[m5G;October 5, 2017;12:12]

T. Schauer / Annual Reviews in Control 000 (2017) 1–20

gyroscope (3D) IMU acclerometer (3D) magnetometer (3D)

rates of turn

strap-down integration (prediction)

bias sensor fusion accel. magn.

5

estimated orientation

vertical+south direction (correction)

external disturbances Fig. 4. General approach for estimating the orientation of an IMU with respect to a fixed reference coordinate system that is defined to have a vertical z-axis and an x-axis that points horizontally towards the magnetic earth south pole (adapted from Seel & Ruppin, 2017).

soon yield wireless inertial sensors with the size of a coin, which can be placed inside the shoe or worn unnoticed under clothes. Inertial measurement units measure acceleration, angular rate and the magnetic field vector, each in a local orthogonal three dimensional coordinate system that is well aligned with the housing of the sensor. The acceleration measurement contains the static acceleration (the gravity vector) and the dynamic acceleration of the sensor. When an IMU is rigidly attached to a body segment, the measured signals can be used to estimate the orientation and velocity (and position) of that segment with respect to a fixed (inertial) coordinate system. 2.1.1. Orientation estimation The general approach of orientation estimation is depicted in Fig. 4. Strapdown integration of the measured angular rate vector is performed to obtain a first orientation estimate. The result is affected by drift (due to gyroscope bias) but very precise on short time scales. Accelerometer and magnetometer yield partial information of the IMU orientation that have a good long term accuracy. For sufficiently large timescales, dynamics accelerations are zeromean and the measured acceleration is a good approximation of the gravity vector, which yields information about the inclination, i.e. the pitch and roll angle of the sensor. Magnetometer readings yield information about the sensor’s orientation with respect to the earth magnetic field, which can be used to determine the yaw (heading) angle if the field vectors are not vertical. However, these readings are typically noisy and strongly affected by magnetic disturbances originating from electronic devices, reinforced concrete walls or any ferromagnetic material. A sensor fusion scheme is typically applied to improved the IMU orientation estimation by combining the short-term accuracy of the gyroscope-based strap-down integration with long-term accuracy of the orientation information gained from accelerometer and magnetometer readings. Many sensor fusion algorithms have been proposed for IMU orientation estimation. They mainly differ by the employed mathematical framework (rotation matrices, axis-angle representations, unit quaternions) and by the filter used for sensor fusion (Kalman filters (Ang, Khosla, & Riviere, 2004; Zhang, Meng, & Wu, 2012), complementary filters (Fourati, Manamanni, Afilal, & Handrich, 2014; Mi, Du, Ye, & Zou, 2010), gradient-descent methods (Madgwick, Harrison, & Vaidyanathan, 2011), sliding mode observers (El Hadri & Benallegue, 2009)). Most sensor fusion algorithms allow the user to specify weights or covariance matrices that influence the balance between prediction (based on gyroscopes) and correction (based on accelerometers and magnetometers). As pointed out by Seel and Ruppin (2017), comparatively little attention has been given to the question in which way the sensor fusion should rely on each of the raw measurement signals. In particular, recent research has demonstrated that magnetometers are hardly reliable in many in-

door environments (De Vries, Veeger, Baten, & Van Der Helm, 2009). Therefore, magnetometer readings should be used only for the purpose for which they are essential, i.e. to remove drift in the horizontal (azimuth/yaw/heading) part of the orientation, and should not influence the inclination (roll and pitch) part of the orientation. Thereby, in the worst case of a completely distorted magnetic field, one obtains orientation estimates that have at least precise inclination components. Seel and Ruppin (2017) from the Control Systems Group at TU Berlin recently proposed an algorithm that perfectly restricts the magnetometer’s influence to the horizontal (heading/azimuth) part of the orientation. The approach can be used in outdoor and indoor environments. If (at least the heading of) the magnetic field is homogeneous, then the entire orientation estimate will be highly accurate. If the magnetic field is heavily disturbed, only the yaw portion of the orientation estimate will be affected. 2.1.2. Joint angle estimation with exploitation of kinematic constraints By attaching IMUs to the body segments around a joint, the corresponding joint angle can be estimated. For determining the sagittal joint angles of the lower extremities (small gait analysis), seven IMUs need to be placed on the trunk, the thighs, the shanks and the feet. The joint angles of the lower limbs can then be determined under the assumption that all joints can be approximated as hinge (revolute) joints. As shown in Fig. 5, one can basically distinguish two approaches to estimate the angle of a hinge joint based on two IMUs (Seel, Raisch, & Schauer, 2014b). Here, gi , ai , mi , i = {1, 2} are the gyroscope, accelerometer and magnetometer measurements of the two IMUs in their local sensor coordinate systems, j1 and j2 is the joint axis seen from both sensors, and o1 and o2 are position vectors pointing from the IMUs to the hinge joint axis (cf. Fig. 5). All axes and position vectors are also defined in the local sensor coordinate systems of the two IMUs. The two approaches are: 2.1.2.1. (1st) Joint Angle from Sensor Orientation Estimates. In the first approach we assume that the orientation estimates of both sensors with respect to a common fixed reference frame are available e.g. by the rotation matrices R1 and R2 . Knowing the local joint axes coordinates j1 and j2 , the orientational difference (i.e. the angle around the joint axis) can be calculated. However, the resulting joint angle can only be as precise as the employed IMU orientation estimates, and will drift if the orientation estimates drift. To avoid joint angle drift due to magnetic disturbances, one can exploit the knowledge of the joint axes, as demonstrated by Laidig, Schauer, and Seel (2017b). Using this approach, we developed an IMU-based sensor system for measuring finger and hand movements (Valtin, Salchow, Seel, Laidig, & Schauer, 2017) motivated by the previ-

Please cite this article as: T. Schauer, Sensing motion and muscle activity for feedback control of functional electrical stimulation: Ten years of experience in Berlin, Annual Reviews in Control (2017), https://doi.org/10.1016/j.arcontrol.2017.09.014

ARTICLE IN PRESS

JID: JARAP

T. Schauer / Annual Reviews in Control 000 (2017) 1–20

C) directions of joint axis

integrate

orientation estimation

B) IMU 1

angle calculation

IMU 2

orientation estimation

IMU 1

direction of joint axis

joint angle

shift to joint center

A)

sensor fusion

6

[m5G;October 5, 2017;12:12]

IMU 2

direction of joint axis

IMU 1 joint axis

joint angle

IMU 2

directions & positons of joint axis

Fig. 5. Approaches for joint-angle estimation (bold arrows indicate vector signal). (A) Joint angle estimation from sensor orientation estimates (B) Joint angle from accelerometer and gyroscope readings, (C) positions o1 , o2 and directions j1 , j2 of the hinge joint axis seen in the local coordinate systems of IMU 1 and 2 illustrated for the knee joint (adapted from Seel et al., 2014b).

100

Z

50

0

-50 -60 -40 -20 0

X

20 40 60 -50

0

50

100

150

200

Y

Fig. 6. Pinch grip captured with an IMU-based hand sensor system (Valtin et al., 2017). Only the thumb, index finger, and middle finger were measured (red) (from Valtin et al., 2017). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

ous work of Kortier, Sluiter, Roetenberg, and Veltink (2014) and van den Noort, Kortier, Beek, Veeger, and Veltink (2016). The system is specially designed for the use by patients with motor impairments, e.g. after stroke. Unlike sensor gloves, the system can be easily attached also to a spastic hand. The prototype of the measurement system is shown in Fig. 6. Each finger segment is equipped with an IMU mounted on a flexible Printed Circuit Board (PCB) strip. 2.1.2.2. (2nd) Joint angle from accelerometer and gyroscope readings. The second approach does not rely on IMU orientation estimates and magnetometer measurements. It avoids the use of magnetometers, which makes this approach more appealing for indoor applications where magnetic disturbances are often present. However, it assumes that the joint axis does not remain vertical for a long period of time. This assumption certainly holds for walking. An estimate of the joint angle can be determined by integrating the difference of the angular rates around the joint axis. However, this estimate will be suspect to drift due to gyroscope biases. Knowing the position vectors o1 , o2 and the local acceleration measurements a1 , a2 , an acceleration-based angle estimate can be calculated as described in detail in (Seel et al., 2014b). This estimate

will not be affected by drift, but it is corrupted by the accelerometer noise and may be less accurate in moments of large acceleration changes. A sensor fusion algorithm (e.g. a Kalman filter or complementary filter) can then be used to combine both estimates. 2.1.2.3. Estimation of Joint axis and position vectors. Both earlier presented joint angle estimation approaches require prior knowledge about the sensor to segment mounting orientation and position that are characterized by the local coordinates of the joint axis (j1 , j2 ) and the joint position (o1 , o2 ), respectively (cf. Fig. 5). Both quantities might be measured manually, but in three-dimensional space, this is a cumbersome task that yields low-accuracy results. The true anatomical joint axis is also hard to assess for most people. Therefore, a common approach is to involve calibration postures and/or calibration movements. However, the accuracy is limited by the precision with which a patient can perform the postures or motions. In Seel et al., 2014b; Seel, Schauer, and Raisch, 2012, we describe a new method that, unlike previous approaches, identifies the local joint axis coordinates and positions from arbitrary motion data by exploiting kinematic constraints of hinge joints. Such an arbitrary movement can be just walking. The approach uses solely

Please cite this article as: T. Schauer, Sensing motion and muscle activity for feedback control of functional electrical stimulation: Ten years of experience in Berlin, Annual Reviews in Control (2017), https://doi.org/10.1016/j.arcontrol.2017.09.014

JID: JARAP

ARTICLE IN PRESS

[m5G;October 5, 2017;12:12]

T. Schauer / Annual Reviews in Control 000 (2017) 1–20

Fig. 7. Plug-and-play sensor-to-segment pairing and calibration (Graurock, Schauer, & Seel, 2016b): After a few steps, the algorithms determine which sensor is attached to which leg segment and in which orientation it is attached. Based on this information, knee joint angles and gait phases can be determined in real time (adapted from Graurock et al., 2016b).

gyroscope and accelerometer readings and numerically solves a nonlinear least-squares problem. The first kinematic constraint is that projections of the two gyroscope measurement vectors into the joint plane (i.e. the geometrical plane to which the joint axis is the normal vector) have the same lengths for each instant in time. This constraint can be used to determine the joint axis. The second kinematic constraint is that the acceleration norm at a point on the joint axis must be the same in both local frames. Position vectors to the joint axis are obtained by exploiting this constraint. Following a similar approach, we determined the two dominant axes of the elbow joint from two arbitrarily mounted IMUs at the lower and upper arm by applying a nonlinear optimization procedure (Laidig, Müller, & Seel, 2017a; Müller, Bégin, Schauer, & Seel, 2017b). The algorithm requires IMU orientation estimates and gyroscope measurements recorded during arbitrary movements. The internal model assumes that both arm segments are connected by two revolute joints, allowing two degrees of freedom. 2.1.2.4. Comparison to optical gait analysis. The gold standard in human motion analysis are still optical systems, which use active or passive markers placed on anatomical landmarks of the body. Such systems are expensive, require experts for operation, and are restricted to the limited observation space of the cameras. IMU-based motion analysis on the contrary is low-cost, wearable, and real-time capable while achieving a similar accuracy as optical systems. By comparing both approaches, we observed root mean square errors of the knee flexion/extension angles in the range of 3 degrees during walking (Seel et al., 2014b). Similar results were obtained for the flexion/extension and pronation/supination angle of the elbow joint (Müller et al., 2017b). 2.1.2.5. Plug-and-play gait analysis. To further enhance the usability of inertial sensor networks, we recently proposed a method that automatically determines which sensor is attached to which segment of the lower limbs (Graurock, Schauer, & Seel, 2016a; 2016b). The method analyses the raw IMU data (gyroscope and accelerometer measurements) of 3 s of walking. Due to its low computational workload, an implementation of the algorithm on embedded systems is feasible. Analyzing data from over 500 trials with healthy subjects and Parkinson patients yields a correct-pairing success rate of 99.8% after 3 s and 100% after 5 s. Based on the presented methods, a plug-and-play gait analysis can be realized as illustrated in Fig. 7. Following the pairing and sensor-to-segment calibration, the real-time estimation of joint angles starts. 2.1.3. Gait phase detection The core element of any FES system for walking is a robust and reliable real-time gait phase detection (GPD). Early and still most commercially available systems use force-resistive switches below the heel to distinguish the stance and swing phase of gait. Many researchers investigated the use of accelerometers and gyroscopes

7

for GPD, (see e.g. Chia Bejarano et al., 2015; Rueterbories, Spaich, & Andersen, 2014; Seel, Landgraf, Escobar, & Schauer, 2014a; Seel, Werner, Raisch, & Schauer, 2016a). Such sensors can be attached to the shoe/foot and/or to the shank and facilitate a more detailed gait phase detection. Fig. 8 shows the state automaton we proposed to describe the gait phases (states) and events (state transitions) that can be detected by a miniature inertial sensor placed on the instep of the foot (Seel et al., 2014a; Seel et al., 2016a). We avoid the use of magnetometers and do not restrict the mounting of the sensor to a certain location or orientation at the foot. The transitions are detected by mathematical conditions including the norm of the acceleration and rate of turn vectors, the norm of the jerk vector (time-derivative of the acceleration vector), the estimates of the foot pitch angle and the horizontal sensor velocity, as well as the estimated norm of the toe velocity vector. The GPD adapts automatically to the individual patient and current cadence by automatic adjustments of the thresholds inside the state transition conditions. The adaptation algorithm analyses the observed standard deviations of the measured acceleration and rate of turn in a sliding window as well as the maximum values of pitch angle, jerk norm and toe velocity norm within the last step (see Seel et al., 2014a for more details). 2.2. Electromyography Muscle activity, or to be precise motor unit recruitment, can be recorded in transcutaneous FES systems by using surface electromyography (EMG) (Merletti & Farina, 2016). The electromyogram captures the net electrical activity (action potentials) produced by many contracting motor units. The potential use of EMG measurement in FES systems is manifold: 1. for assessing the residual volitional muscle activity of muscles to trigger and control the electrical stimulation, 2. to determining how good a muscle responds to the stimulation in order to control the recruitment of the artificially activated motor units, and 3. to observe how FES affects the patient’s motor coordination and spasticity. The measurement of EMG is typically performed via AgClelectrodes that are placed on the muscle of interest as exemplary shown in Fig. 2 A) for the deltoid muscle. A measurement using the stimulation electrodes is also feasible but with some restrictions, as explained later on. Fig. 9 shows a raw surface electromyography signal recorded during active stimulation (at 25 Hz stimulation frequency) with a period of superposing volitional muscle activity. Separate electrodes have been used for stimulation and EMG recording. The nowadays employed recording systems often use high resolution 24 Bit A/D front ends for EMG recording. The inputs of these systems are protected by resistor-diode networks against the high voltage stimulation pulses. The remaining stimulation artifacts caused by the pulses are still visible in the EMG, but recovery of the EMG measurement system usually takes place in less than two milliseconds. When analyzing such EMG signals, one has to distinguish between FES-evoked EMG and patient-induced EMG, where the latter includes both, intentional (volitional) and unintentional muscle activity (reflexes) (Merletti & Farina, 2016; Merletti, Knaflitz, & DeLuca, 1992). By means of online signal processing, both quantities can be determined from the raw EMG in between the stimulation pulses, i.e. during active stimulation (see e.g. (Klauer et al., 2016) for detailed signal processing steps). 2.2.1. FES-evoked EMG activity. The FES-evoked EMG presents itself in the so-called M-wave, which is a good measure for the amount

Please cite this article as: T. Schauer, Sensing motion and muscle activity for feedback control of functional electrical stimulation: Ten years of experience in Berlin, Annual Reviews in Control (2017), https://doi.org/10.1016/j.arcontrol.2017.09.014

ARTICLE IN PRESS

JID: JARAP 8

[m5G;October 5, 2017;12:12]

T. Schauer / Annual Reviews in Control 000 (2017) 1–20

pre-swing

foot flat

full contact

heel rise

loading response

swing phase

toe-off

initial contact

Fig. 8. By means of a foot/shoe-mounted inertial sensor, the transitions between four distinct gait phases can be detected in realtime (from Seel et al., 2016a).

72.88

72.89

2

72.9

72.91

72.92

72.93

73.24

A)

M-wave

raw EMG [mV]

73.26

2

0 -2

73.25

73.27

73.28

73.29

B)

M-wave

0

stimulation artifacts

stimulation artifacts

-2

2

0

-2 72.5

72.6

72.7

72.8

72.9

73

time [s]

73.1

73.2

73.3

73.4

73.5

Fig. 9. EMG recording during active stimulation with a stimulation frequency of 25 Hz. (A) Stimulation period with clearly visible volitional muscle activity, (B) Stimulation period with almost no volitional muscle activity.

of synchronously activated motor units by the last stimulation impulse. The amplitude of the M-wave may reach a couple of mV. Typically, the 1-norm, the root mean square, or the peak-to-peak value is determined from the current M-mave to quantify the recruitment level. With muscle fatigue, the number of recruited motor units will decrease. This will be clearly visible in the recorded recruitment level. The power spectrum of the M-wave is another indicator for muscle fatigue. It will change from higher to lower frequencies due to fatigue (Merletti & Farina, 2016). This is due to the early drop-off of fast-twitch type 2B motor units. With progressing time, mainly the slow and fatigue-resistant motor units will remain recruited by FES. 2.2.2. Patient-induced EMG activity. The EMG portion that is due to patient-induced muscle activity is usually much smaller in amplitude (in the μV range) than the M-wave, since motor units are activated asynchronously. This rather noise-like signal with frequency components in the range of 30 to 300 Hz (De Luca & Knaflitz, 1992) can be separated from the M-wave about 20 to 30 ms after each stimulation pulse by either high-pass filtering or by subtraction of the estimated M-mave, see e.g. Ambrosini et al. (2014) and Klauer et al. (2016) for more details. Both filtering approaches operate on a vector of EMG samples collected in the last inter-pulse interval. The estimation of the deterministic M-wave by a linear combination of EMG recordings from preceding inter-pulse intervals was originally proposed by Sennels, Biering-Sorensen, Andersen, and Hansen (1997). To reduce unwanted filter transients in the high-pass filter approach, as observed in Ambrosini et al. (2014), we recently proposed an energy minimization of the filter output signal by selecting optimal initial states for the non-causal high-pass filter (Schauer, Seel, Bunt, Müller, & Moreno, 2016). Fig. 10 shows examples of the filter steps for the high-pass filtering for a stimulation period with and without volitional activity. The output of the high-pass filter is rectified and averaged to determine a scalar measure of the volitional mus-

cle activity (motor unit recruitment) within the inter-pulse interval. 2.2.3. Measurement from stimulation electrodes. Most systems require separate electrodes for stimulation and EMG measurement. This may restrict the transfer of such systems into clinically usable systems. A direct EMG measurement from the stimulation electrodes would be a clear technological advantage. However, compared to the setup with separate electrodes there are three main challenges: 1. The voltage potential difference between the stimulation pulses (up to 150 V) and the EMG signal (less than 5 mV) is huge. 2. The electrode area of stimulation electrodes is much bigger than the one of EMG electrodes. 3. The capacitive rest charge on the electrodes will cause significant discharging transients in the EMG recordings that are difficult to predict and make EMG measurements impossible without additional inforced discharge of the electrodes. Protection and (at least passive) discharging circuits in front of the EMG amplifier are therefore required. The feasibility of an EMG measurement from the stimulation electrodes was demonstrated in Muraoka (2001) and in our work (Shalaby, 2011; Shalaby, Schauer, Liedecke, & Raisch, 2011) by using classical analog EMG amplifiers, analog high-pass filtering, and protection/discharge (passive) circuits. In cooperation with an industrial partner, we recently developed a stimulation system with an integrated 24-Bit analog frontend for EMG measurements (Valtin et al., 2016a). An input circuit with PhotoMos switches protects the analog front-end and allows passive discharging of the stimulation electrodes after each stimulus. For the extraction of the volitional EMG activity, we used the aforementioned high-pass filter with optimally chosen initial conditions (Valtin, Werner, & Schauer, 2016b). The combined stimulation and measurement system was initially used to detect volitional foot lift during active stimulation from the stimulation elec-

Please cite this article as: T. Schauer, Sensing motion and muscle activity for feedback control of functional electrical stimulation: Ten years of experience in Berlin, Annual Reviews in Control (2017), https://doi.org/10.1016/j.arcontrol.2017.09.014

ARTICLE IN PRESS

JID: JARAP

[m5G;October 5, 2017;12:12]

T. Schauer / Annual Reviews in Control 000 (2017) 1–20

stimulation period with strong volitional activity

A) 3000

1000 0 1000 M wave stimulation artifact

2000 3000 0

3000

5

10 15 20 25 30 35 40 45 50

raw electromyogram selected Window

2000

EMG [µV]

EMG [µV]

2000

stimulation period with weak volitional activity

B)

raw electromyogram selected Window

1000 0 1000 M wave stimulation artifact

2000 3000 0

5

time [ms] 50

50

20 10 0

30 20 10 0

10

10

20

20

30 0

5

10 15 20 25 30 35 40 45 50

time [ms]

highpass filtered EMG rectified EMG averaged (volitional) EMG

40

EMG [µV]

EMG [µV]

30

10 15 20 25 30 35 40 45

time [ms]

highpass filtered EMG rectified EMG averaged (volitional) EMG

40

9

30 0

5

10 15 20 25 30 35 40 45

time [ms]

Fig. 10. Signal processing steps to determine the volitional muscle activity (motor unit recruitment) from inter-pulse EMG recording: (1st) windowing, (2nd) high-pass filtering, 3rd) rectification, and 4th) averaging; Left side (A): Stimulation period with stronger volitional activity; Right side (B): Stimulation period with almost no present volitional activity (adapted from Schauer, 2014).

trodes in a drop foot stimulator setup. One important remaining disadvantage of systems with integrated EMG and stimulation electrodes is that an assessment of the M-wave is not feasible due to the longer lasting stimulation artifacts.

voltage measurement electrodes for bioimpedance

2.3. Bioimpedance Another, less well-known method for the assessment of human movement and muscle contraction is the use of bioimpedance (BI) measurements. The passive electrical properties of tissue are summarized as bioimpedance (Grimnes & Martinsen, 2008). It is a complex variable that is measured by the voltage drop caused by a sinusoidal current flow through tissue. Often the so-called fourelectrode measurement setup is used where different electrodes are used for the voltage measurement and for the insertion of the very small sinusoidal current1 . This allows a better control of the measured zone and avoids the effects of current-carrying measurement. To assess movements and muscle contractions in real-time, usually only the absolute value of the bioimpedance at a frequency of about 50 kHz is investigated. We have developed a measurement system that determines the absolute value of BI by using an amplitude demodulation circuit and that enables a parallel EMG measurement via the BI voltage measurement electrodes (Nahrstaedt, 2017; Nahrstaedt & Schauer, 2009). Both signal components can be completely separated by electronic filters. Song, Kim, Nam, and Kim (2005) and Kim et al. (2003) first described the use of BI to analyze human movement of the lower and upper limbs, respectively. We have adopted this idea to realize a feedback-controlled drop foot stimulator in which the ankle-joint 1

Amplitudes are typically less than 250 μA.

stimulation electrodes

current insertion electrodes for bioimpedance

Fig. 11. Electrode positions for a bioimpedance measurement (four-electrode method) of the ankle-joint ankle and stimulation electrodes for the controlled activation of m. tibialis anterior in a drop foot stimulator (adapted from Nahrstaedt et al. (2008)).

angle is determined by BI (Nahrstaedt, Schauer, Shalaby, Hesse, & Raisch, 2008). Fig. 11 shows the used electrode setup. An advantage of BI measurement systems is their straightforward integration in textile. However, compared to IMU-based motion analysis, a calibration phase with prescribed postures is required every time the system is attached to the patient. The correlation between gripping force caused by voluntary isometric contractions and bioimpedance of the forearm flexor muscles was described in Shiffman, Aaron, and Rutkove (2003). We

Please cite this article as: T. Schauer, Sensing motion and muscle activity for feedback control of functional electrical stimulation: Ten years of experience in Berlin, Annual Reviews in Control (2017), https://doi.org/10.1016/j.arcontrol.2017.09.014

JID: JARAP 10

ARTICLE IN PRESS

[m5G;October 5, 2017;12:12]

T. Schauer / Annual Reviews in Control 000 (2017) 1–20

Fig. 12. General FES closed-loop system. All possible feedback signals are indicated by dashed blue lines. The recruitment level can be assessed by EMG. Torque/force can be predicted from EMG and kinematic data by means of models or measured by BI. Joint acceleration, velocity and angle can be obtained from inertial sensors that are attached to the body segments neighboring the joint. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

showed that also the assessment of FES-induced gripping force is feasible by bioimpedance (Nahrstaedt et al., 2008). Combined bioimpedance and EMG recordings at the neck can be used for automatic detection of swallowing events and to trigger a supportive FES of swallowing muscles. A detailed description will be given in Section 3.4. 2.4. How does FES control benefit from the new measurements? Fig. 12 sketches a general FES closed-loop system with a macroscopic muscle model (Schauer et al., 2005) that is commonly used to describe the generation of FES-induced movements.2 The shown muscle model consists of activation dynamics and a nonlinear contraction function. The relation between stimulation intensity and amount of recruited motor units can be approximately described as a nonlinear input saturation (recruitment function). An estimate of the function‘s output is available from the M-wave at stimulation frequency (corresponds to sampling frequency of the FES controller) with a time-delay of one sampling step. The subsequent activation dynamics in form of a second order transfer function describes the release of calcium ions (Ca2+ ) inside the muscle fibers for force production and yields the muscle activation as output. By multiplication of the muscle activation with the output of the nonlinear contraction function, we obtain the muscle torque produced by FES that acts on a joint. The contraction function describes the largest producible torque depending on the current joint angle and angular velocity. A model-based prediction of the FES-induced torque based on the measured muscle recruitment is discussed in Hayashibe (2016). The final component of the model is the segmental dynamics that can be modeled as a rigid body system while considering also passive elastic and viscous joint moments. A possible extension of the model to capture muscle fatigue can be found in Riener (1999). As outlined in Section 2.1, IMUs can be employed to obtain joint angles as well as angular velocities and accelerations for feedback control. Bioimpedance might be utilized to estimate joint-angles and/or torques. Please note that the model complexity significantly reduces for isometric conditions (no joint movement). Then the produced force can be solely described by a Hammerstein model consisting of the nonlinear recruitment function and the linear activation dynamics. Often this simplified model is also applied in non-isometric conditions when the angular range is small and the velocities are low. 2.4.1. Cascaded control. Classical FES controllers merely rely on feeding back joint angles (angular feedback) and can only slowly react to disturbances and output deviations due to model uncertainties. The recruitment function is especially difficult to esti2 Additional voluntary drive as input is neglected here for simplification but could be easily incorporated into the model as proposed by Klauer, Irmer, and Schauer (2015) with a measurement of the volitional drive through EMG recordings.

mate and varies with time, e.g. due to muscular fatigue (Durfee & MacLean, 1989). The newly available internal measurements or predictions allow us to employ cascaded control structures that have a high potential to improve the control performance. Please note that not all internal signals need to be used for the inner loop(s) of the cascade. Fig. 12 exemplifies the case with internal feedback of the muscular recruitment level. Our recent studies for the upper extremities show that this internal feedback of the Mwave intensity compensates for the effects of muscular fatigue and maintains a desired stimulation effect (Klauer et al., 2016; Klauer, Raisch, & Schauer, 2012). A simple time-discrete integral controller proved to be sufficient for obtaining a very fast closed-loop performance on the muscular recruitment level. The alternative approach of feeding back the prediction of FES-induced torque is outlined in detail in Hayashibe (2016) and references therein. However, this approach requires additional model parameter estimation. 2.4.2. Control of antagonistic muscles. For antagonistic muscle pairs, the nonlinear recruitment functions create a dead-zone in the muscle actuation that can degrade control performance dramatically. A good compensation of this dead-zone is generally difficult to achieve over a long time as system parameters change. One prominent approach to deal with the problem is to apply co-contractions when the actuation changes from agonist to antagonistic and vice versa (Durfee, 1989; Zhou et al., 1996). In this context we investigated the use of underlying feedback control of muscle recruitment to control co-contractions for generating horizontal shoulder movements (Klauer, Raisch, & Schauer, 2013; Spagnol, Klauer, Previdi, Raisch, & Schauer, 2013). A very fast and precise control performance could be achieved by this control approach in comparison to classical PD control of the antagonistic muscles which was used in another study with the same setup (Vidaurre, Klauer, Schauer, Ramos-Murguialday, & Müller, 2016). The feasibility of feeding back the underlying joint acceleration (acceleration feedback) for high performance control of antagonistic muscle pairs was also demonstrated by us in Klauer, Schauer, and Raisch (2011). However, acceleration feedback cannot be employed to precisely control the level of co-contraction, as the amount of co-contractions is not observable from this measurement. 3. Feedback-controlled neuro-prostheses with integrated sensors 3.1. IMU-based FES cycling Paraplegics with a complete spinal cord injury can drive an ergometer or recumbent tricycle with their paralyzed muscles by means of FES (see Newham and Donaldson, 2007; Schauer, 2006 for a review). The feet are attached to the pedals by orthoses. These fixate the ankle joints and reduce the number of mechanical degrees of freedom for cycling to one. A sequential sensor-driven

Please cite this article as: T. Schauer, Sensing motion and muscle activity for feedback control of functional electrical stimulation: Ten years of experience in Berlin, Annual Reviews in Control (2017), https://doi.org/10.1016/j.arcontrol.2017.09.014

JID: JARAP

ARTICLE IN PRESS T. Schauer / Annual Reviews in Control 000 (2017) 1–20

Fig. 13. FES bike race at the Cybathlon 2016: Team HASOMED with pilot Hanno Voigt using an IMU-based FES cycling approach (© ETH Zürich, Nicola Pitaro).

stimulation of typically three muscle groups (knee extensor and flexor as well as hip extensor) per leg can generate a smooth cycling motion while the cyclist steers and brakes the tricycle with his/her unimpaired arms and hands (cf. Fig. 13). The standard approach used nowadays is to directly measure the crank angle and to define stimulation patterns based on it. This approach has some drawbacks: Every used cycling device must be instrumented with a crank angle sensor and the patterns strongly depend on the subject and his/her sitting position with respect to the crank. To mitigate these problems, we have developed an IMU-based stimulation strategy employing the methods outlined in Section 2.1 for orientation estimation, sensor-to-segment calibration and joint-angle estimation. By placing IMUs on both thighs and shanks, we can robustly discriminate extension and flexion phases of the legs and use this information for the timing of muscle stimulation (Ruppin, Wiesener, & Schauer, 2016; Wiesener, Ruppin, & Schauer, 2016). Using this approach, any ergometer and tricycle can be driven by FES without modifications of the device and without knowing the sitting position in advance. By exploiting the kinematic constraint that the feet move on a circular path and by knowing the crank arm length, we can estimate the segment lengths of thigh and shank and the cadence at the crank from the inertial measurement data (Wiesener et al., 2016). This is important for precise control of the cadence or the speed of the vehicle. The effectiveness of IMU-based FES cycling was demonstrated at the FES bike race that took place during the 1st Cybathlon in Switzerland on the 8th of October 2016. Our athlete, Hanno Voigt (SCI lesion level T5/6, ASIA A), who had been paralyzed for over 35 years at the time of the Cybathlon, completed the 750 m race in fourth place with a new personal record of 6 min 44 s (Wiesener & Schauer, 2017). The pilot had the challenging task to manually regulate the stimulation intensity, and thereby the speed of the tricycle, while taking the progress of muscle fatigue into account. 3.2. Adaptive drop foot stimulator About 20% of ambulatory stroke survivors suffer from drop foot, which is characterized by a limited ability to lift the foot and leads to a pathological gait (Wade, Wood, Heller, Maggs, & Hewer, 1987). Commercially available drop foot stimulators use a simple heel switch or an inclinometer at the shank to control the timing of the peroneal nerve stimulation during the swing phase of the paretic foot (Lyons, Sinkjaer, Burridge, & Wilcox, 2002; Melo, Silva, Martins, & Newman, 2015). The peroneal nerve innervates and activates mainly m. tibialis anterior (causes foot lift with foot inversion) and m. fibularis longus (causes foot lift with foot eversion).

[m5G;October 5, 2017;12:12] 11

Fig. 14. Concept of the Adaptive Peroneal Stimulator – APeroStim (Seel et al., 2016b): Angle definitions and electrode placement. An IMU is used to determine gait phases and foot orientation angles (adapted from Seel et al., 2016b).

The two electrodes of the one stimulation channel must be placed very carefully to obtain a pure foot lift (dorsiflexion) without exaggerated inversion or eversion of the foot. This might be a cumbersome work and the observed stimulation effects may not be consistent over time due to the mentioned effects of muscular fatigue and time-variant spasticity. Therefore, we investigated the use of IMU-based feedback control to achieve a desired foot motion (dorsiflexion and eversion/inversion) and developed the adaptive peroneal stimulator – APeroStim. A small wireless inertial sensor is attached to the foot (typically on the instep but not in any predefined orientation or position). Gait phase transitions as well as foot pitch (dorsiflexion) and roll angles (eversion/inversion) are determined in realtime by means of the IMU (cf. Fig. 14 and Sections 2.1.1 and 2.1.3) (Seel, Graurock, and Schauer, 2015a; Seel et al., 2014a). We investigated two different stimulation techniques that allow us to manipulate the recruitment of m. tibialis anterior and m. fibularis longus via two independent FES channels without violating the zero-netcurrent requirement of FES: a three-electrodes setup shown in Fig. 14 (Seel, Werner, & Schauer, 2016b) and a two-electrodes setup with a triphasic pulse form (Seel, Valtin, Werner, & Schauer, 2015c). Beyond this, a setup with two electrode arrays has been investigated (Valtin, Seel, Raisch, & Schauer, 2014). To compensate most of the cross couplings between the two FES intensities and the two foot orientation angles, we proposed a nonlinear controller output mapping (Seel, Ruppel, Valtin, & Schauer, 2015b). A decentralized Iterative Learning Control (ILC) scheme is used to adjust the stimulation intensity profiles to the current needs of the individual patient (Seel et al., 2015c; Seel et al., 2016a; Seel et al., 2016b). The adaptation of the profiles takes place in between steps, while the paretic foot is in stance phase. The effects of natural variations in the duration of the swing phase have been considered in the design and implementation of the ILC (Seel, 2016; Seel, Schauer, & Raisch, 2017). The used learning gains have been chosen based on experimentally identified linear transfer function models in order to guarantee a monotonic convergence of the error norm. We evaluated the effectiveness of this approach in experimental trials with drop foot patients walking on a treadmill and on level ground at the Berlin Medical University Charité. Starting from conventional stimulation parameters, the controller automatically determines individual stimulation parameters and thus achieves physiologically normal foot pitch and roll angle trajectories within at most two strides. Fig. 15 shows the stride-to-stride adaptation of the mapped stimulation intensity profiles by the ILC that brings the dorsiflexion (pitch) angle and the eversion/inversion (roll) an-

Please cite this article as: T. Schauer, Sensing motion and muscle activity for feedback control of functional electrical stimulation: Ten years of experience in Berlin, Annual Reviews in Control (2017), https://doi.org/10.1016/j.arcontrol.2017.09.014

ARTICLE IN PRESS

JID: JARAP 12

[m5G;October 5, 2017;12:12]

T. Schauer / Annual Reviews in Control 000 (2017) 1–20

A)

time [s]

RMSE pitch [°]

physiological range

RMSE roll [°]

lateral stim. int. [ ]

frontal stim. int. [ ]

desired motion

desired motion

foot roll [°]

foot pitch [°]

B)

physiological range

time [s]

step index

Fig. 15. Experimental results from a trial with a chronic drop foot patient using the adaptive peroneal stimulator (Seel et al., 2015c). (A) The stimulation intensities are adjusted from stride to stride in order to generate the desired foot motion during swing phase. Dots mark heel-rise and initial contact of each stride. Starting from conventionally high trapezoidal profiles (ufro, 0 and ulat, 0 ), the ILC adjusts the frontal and lateral stimulation intensity trajectories from trial to trial and thereby improves the foot pitch and roll angle trajectories (i and i ) from stride to stride. (B) Root-mean-square errors of the foot pitch and roll (data from same trial) are quickly reduced to the ranges of natural variance found in healthy subjects walking at similar velocities (adapted from Seel et al., 2015c).

Fig. 16. Evolution of the gait pattern in a chronic stroke patient (right body side affected) using the adaptive peroneal stimulator (APeroStim). The drawn bars indicate the dorsiflexion and eversion/inversion angles (© TU Berlin, Thomas Seel).

gle sufficiently close to the given reference trajectories (Seel et al., 2015c). Likewise, if the chosen initial stimulation parameters are too low for the individual patient under the current circumstances, the IMU measures weak dorsiflexion and exaggerated inversion, whereupon the ILC quickly adjusts the stimulation parameters to achieve a physiological foot motion. The initial and the resulting foot motion are shown in Fig. 16 for a chronic stroke patient. It is clearly visible that the ILC achieves a physiologically normal foot lift at heel strike while maintaining a neutral eversion/inversion of the foot during the swing phase. 3.3. FES-based arm weight relief Many stroke patients still have or regain partial control over their hand, arm and shoulder muscles. However, that might not be enough to perform rehabilitation exercises or to carry out activities of daily live using the impaired arm. One solution to the problem is to trigger (see e.g. Ambrosini et al., 2014) or proportionally control (see e.g. Hara, 2008; Hara, Ogawa, Tsujiuchi, & Muraoka, 2008; Thorsen, Occhi, Boccardi, & Ferrarin, 2006; Thorsen, Puglia, & Ferrarin, 2013) the stimulation with respect to the residual volitional muscle activity. The latter can be detected by EMG recordings from the stimulated muscles or synergetic ones. The tuning of such control systems is problematic due to changes in the volitional activity

over time and the poor signal-to-noise ratio of the estimated volitional EMG activity. Therefore, we have developed a different patient-driven control strategy, which amplifies weak voluntarily initiated shoulder abductions by providing an adjustable virtual arm weight support. This enables a larger voluntarily reachable range of motion. Fig. 17 gives a rough overview of the used control structure (Klauer et al., 2016). The deltoid muscle is stimulated to produce the supportive shoulder moment. EMG recordings on the same muscle are used to continuously determine the FES-induced motor recruitment level (cf. Section 2.2). Based on the currently measured abduction angle, which is calculated from the IMU orientation (cf. Section 2.1.1), a deltoid recruitment level is determined that would allow to hold the arm at the measured angle. The block “arm weight support controller” then requests an adjustable percentage (support factor pWC < 100%) of this value as reference signal for the underlying recruitment controller (cf. Section 2.4). Its output, the stimulation intensity, is transmitted to the stimulator which delivers the stimulation pulses to the muscle. The fast inner feedback loop compensates the effects of muscular fatigue and linearizes the artificially activated muscular system (Klauer et al., 2016). Fig. 18 illustrates this by comparing the static input-output relation obtained for classical direct stimulation (DS) and for the muscle under recruitment control (RC) in a subject. It is clearly visible that the system behavior is more linear when applying recruitment control. Furthermore, almost no drop in the joint angle can be observed for RC between the first and last trial then compared to DS. The volitional muscle activity of the deltoid can also be determined from the EMG recordings to find the optimal support factor which forces the patient to use his remaining volitional muscle activity as much as possible (Klauer & Schauer, 2017). At the moment, a linear superposition of FES-induced and voluntary-induced movement is assumed in the controller design. However, this a clear simplification, as the FES causes not only the desired orthodromic action potentials (that travel towards the muscle) but also antidromic action potentials (that non-physiologically travel towards the spinal cord) in the excited peripheral motor nerves. To a certain degree, the antidromic action potentials block volitionally induced action potentials from the CNS and thereby suppress volitional muscle activity. Future work must take these effects into account.

Please cite this article as: T. Schauer, Sensing motion and muscle activity for feedback control of functional electrical stimulation: Ten years of experience in Berlin, Annual Reviews in Control (2017), https://doi.org/10.1016/j.arcontrol.2017.09.014

ARTICLE IN PRESS

JID: JARAP

[m5G;October 5, 2017;12:12]

T. Schauer / Annual Reviews in Control 000 (2017) 1–20

13

Fig. 17. Control scheme for a FES-based arm weight relief with underlying recruitment control (adapted from Klauer et al., 2016).

Fig. 18. Comparison of Direct Stimulation (DS) and Recruitment Control (RS): Shown are the determined static relations between stimulation intensity and the shoulder abduction angle (left graphic) as well as between the recruitment level and the shoulder abduction angle (right graphic). Results of the first and the last (eighth) trial of a fatigue comparison test are shown in which stimulation intensity and recruitment were stepwise increased and then decreased. All quantities are normalized between their recorded minimal and maximal values (adapted from Klauer et al., 2016).

The control system has been successfully evaluated in trials with acute stroke patients at the Unfallkrankenhaus Berlin (Klauer et al., 2017). Fig. 19 shows how the maximal volitional arm elevation could be significantly improved in an acute stroke patient by the FES-based arm weight relief. 3.4. BI/EMG-triggered FES to support swallowing Swallowing is a highly complex and vital process. Depending on the current phase of swallowing, it is either volitional or reflextriggered. The complete closure of the larynx and its timing take a central role in safe swallowing, especially since the larynx is the branching point between the trachea and the oesophagus. In case of closure failures, aspiration can occur, i.e. saliva, liquid or food enter the airway through the trachea. Therefore, swallowing disorders (dysphagia) can lead to serious complications, including malnutrition and pneumonia, which may be fatal. Fig. 20 shows the most important anatomical structures involved in swallowing. In order to close the entrance to the trachea, the larynx (consisting of the tracheal cartilages) and hyoid must move upwards to make the epiglottis flip over the trachea during the pharyngeal phase of swallowing. At the same time the esophagus opens. In patients with dysphagia, this important process of larynx and hyoid elevation is often impaired with respect to strength and timing. Since the pioneering work by Freed, Freed, Chatburn, and Christian (2001), transcutaneous neu-

Fig. 19. Test of the FES-based arm weight relief with an acute stroke patient: (A) Initial maximal volitional arm elevation without FES support, (B) Maximal volitional arm elevation with 80% support at beginning of the exercise, (C) Maximal volitional arm elevation with 80% support after 10 minutes exercising, (D) Maximal volitional arm elevation without FES support at the end of training (© TU Berlin, Christian Klauer).

romuscular electrical stimulation has become an established and effective method in dysphagia treatment (Chen et al., 2016). It aims at the strengthening of the submental, suprahyoidal and infrahyoidal muscles, all of which are involved in swallowing. Here, the electrical stimulation is usually applied for longer periods, not taking into account any volitional swallowing activity of the patients. As reported in Ludlow et al. (2007), the stimulation in the vicinity of the larynx can have a negative effect on the protection of the airways during a swallow, because muscles might be activated that lower the larynx and hyoid (e.g. the M. sternohyoideus). Only stimulation of the submental muscles proved to be safe with a positive effect on swallowing mechanics (Ludlow et al., 2007). Leelamanit, Limsakul, and Geater (2002) used EMG of the posterior tongue to trigger the stimulation in patients with reduced laryngeal elevation. A recently introduced medical stimulation device (VitalStim-Plus, Chattanooga, USA) takes up this approach again.

Please cite this article as: T. Schauer, Sensing motion and muscle activity for feedback control of functional electrical stimulation: Ten years of experience in Berlin, Annual Reviews in Control (2017), https://doi.org/10.1016/j.arcontrol.2017.09.014

ARTICLE IN PRESS

JID: JARAP 14

[m5G;October 5, 2017;12:12]

T. Schauer / Annual Reviews in Control 000 (2017) 1–20 Table 1 Automatic swallow detection: Results from the clinical study (Nahrstaedt, 2017). Predicted Swallows Actual

tongue epiglottis cartilages

hyoid

vocal cords trachea esophagus

Fig. 20. Important anatomical structures involved in swallowing (© Can Stock Photo / Pixelchaos).

However, EMG alone is a very unspecific indicator for the pharyngeal phase of swallowing, as e.g. speaking, chewing, and head movements cause similar EMG activity. In research, the feasibility of implanted neuro-prosthesis has been investigated, including the use of intramuscular stimulation (Burnett, Mann, Cornell, & Ludlow, 2003) and selective stimulation of the hypoglossal nerve (Hadley, Kolb, & Tyler, 2013) to cause larynx/hyoid elevation, and the stimulation of the recurrent laryngeal nerve to close the vocal folds as another protection of the trachea (Broniatowski et al., 2010). The problem of effectively triggering the stimulation synchronously to volitionally induced swallows by the patient is not solved satisfactory at the moment for all proposed stimulation approaches. Beside EMG-triggering, Burnett, Mann, Stoklosa, and Ludlow (2005) proposed manual triggering by the patient with a hand switch but tested this approach only with healthy subjects. Another limitation of all approaches is that the stimulation effect can only be reliably observed in clinics with costly and technically demanding examinations (fiberoptic endoscopic evaluation or videofluoroscopy). The number of videofluoroscopic examinations must be limited per year as patients are exposed to radiation. None of the established measurement methods for dysphagia diagnostics is real-time capable or portable enough to enable a feedback control of the stimulation in the patient during swallowing or from swallow to swallow.

3.4.1. BI/EMG measurement system. To trigger functional electrical stimulation of the submental muscles synchronously to a patientinitiated swallow more robustly than with EMG alone, we developed a combined bioimpedance and electromyography measurement system (cf. Section 2.3) (Nahrstaedt, 2017; Nahrstaedt & Schauer, 2009). A revised and miniaturized version of the system is currently released as medical product under the name RehaIngest (Hasomed GmbH, Germany). Fig. 21 A shows the electrode placement for the BI/EMG measurement together with the stimulation electrodes for activation of the submental muscles.

Swallows Non-swallows Sensitivity: Specificity:

3219 1082

Non-swallows 365 11,574 89.9% 91.5%

Fig. 21 B shows exemplary bio-signals obtained during a swallow (stimulation artifacts are removed). There is a patient-initiated EMG activity at the begin of the pharyngeal swallowing phase, which coincides with a drop in the measured bioimpedance (absolute value) due to larynx/hyoid elevation. The inverted BI strongly correlates with the larynx/hyoid elevation, as shown in a clinical study comparing bioimpedance measurements with videofluoroscopy (Nahrstaedt, 2017; Schultheiss, Schauer, Nahrstaedt, & Seidl, 2014). Therefore, we can define two important measures as shown in Fig. 21: First Elevation (rise of BI) that describes the amount of larynx/hyoid elevation and second Speed (max. slope of BI) that captures the speed of larynx/hyoid elevation. Higher values are linked to safer swallowing. Hence, these quantities are suitable for FES control and biofeedback, and they can be employed to monitoring the therapy progress. 3.4.2. Automatic detection of swallows. A two-step procedure has been developed to automatically detect completed swallows in the recorded signals (Nahrstaedt, 2017; Nahrstaedt, Schultheiss, Seidl, & Schauer, 2012; Schultheiss et al., 2014). The 1st step is the Physiological Signal Segmentation, in which potential swallow events are detected by finding valleys in the measured BI that coincide with EMG activity. The latter is detected by using a double threshold onset detector. The measured bioimpedance is pre-processed for the valley search by performing a Piecewise Linear Approximation (PLA). In the 2nd step (Classification), a Support Vector Machine (SVM) is applied to classify the potential swallow events into swallows and non-swallows. Therefore, features are extracted for each segmented potential swallow. The list of features comprises areas, amplitudes, times, and Symbolic Aggregate approXimations (SAX) of the measured BI and EMG signals. Measurement system and automatic swallow detection have been clinically validated at the Unfallkrankenhaus Berlin (Nahrstaedt, 2017; Schultheiss, Schauer, Nahrstaedt, & Seidl, 2013; 2014; Schultheiss, Wolter, Schauer, Nahrstaedt, & Seidl, 2015). A study with 41 patients and 31 healthy subjects was conducted including 3661 manually marked swallows and other movements. The physiological signal segmentation found 3584 swallows (97.9%) but also 12,565 other events that are misinterpreted as swallows. The training and testing of the subsequent classifier was performed using the “leave-one-subject-out” approach. The very good results of the classification are summarized in Table 1, yielding a sensitivity of 89.9% and a specificity 91.5%. From every detected swallow performance measures (elevation and speed) are automatically extracted for biofeedback or control purposes. 3.4.3. Real-time detection of the beginning of the pharyngeal swallowing phase to trigger FES. In a case study with a chronic stroke patient (Nahrstaedt, Schultheiss, Schauer, & Seidl, 2013) and a study with 11 healthy subjects (Schultheiss, Schauer, Nahrstaedt, Seidl, & Bieler, 2016), we demonstrated that a functional electrical stimulation of the submental muscles synchronously to the voluntary induced swallows is feasible and can increase the amount and speed of larynx elevation. The pharyngeal swallowing phase was detected

Please cite this article as: T. Schauer, Sensing motion and muscle activity for feedback control of functional electrical stimulation: Ten years of experience in Berlin, Annual Reviews in Control (2017), https://doi.org/10.1016/j.arcontrol.2017.09.014

JID: JARAP

ARTICLE IN PRESS T. Schauer / Annual Reviews in Control 000 (2017) 1–20

[m5G;October 5, 2017;12:12] 15

Fig. 21. BI/EMG measurement at the neck to trigger FES for swallowing support. (A) Electrode placement: The current electrodes (C), that insert the measurement current, are positioned bilaterally at the insertion of the sternocleidomastoid muscle. The voltage measurement electrodes (V) for BI and EMG are attached bilaterally between the hyoid and larynx. The reference electrode (R) is placed on the right cheek. The stimulation electrodes (S) are on the submental region. (B) Typical EMG and BI signal traces during swallowing.

in real time by an initial drop in bioimpedance (BI) in conjunction with EMG activity measured at the neck by using a decision tree classifier. Due to non-swallow-related movements like speaking, chewing or head turning, stimulation could be unintentionally triggered. The user was therefore instructed to avoid such movements if possible or to deactivate the real-time swallow detection by a hand switch, e.g. during speaking periods. 4. Discussion and directions of future research Three promising approaches for sensing human motion and muscle activity have been reviewed: inertial sensing, electromyography and bioimpedance. The required hardware for all approaches can be greatly miniaturized due to recent developments in microelectro-mechanical systems and electronic circuits that are now available at a low cost. A prominent example is the integrated 24-bit bio-signal amplifier series ADS by Texas Instruments. The ADS294R has four channels and offers for example already a combined EMG and BI measurement. This general development allows the design of measurement systems that can be almost invisibly worn by the patients. An integration into textiles will emerge in the near future. One open issue is to develop waterproof solutions for the stimulation and sensor technologies that enable robust outdoor usage or allow the combination of aqua therapy and FES. All outlined measurement principles can in principle also be used inside implantable neuro-prosthetic devices. Future feedback-controlled FES systems will probably involve a combination of the proposed measurement methods in cascaded control schemes, e.g. an EMG-based control of the muscular recruitment as a fast inner loop and angular feedback provided by IMUs on top. For the control of swallowing, the use of all three technologies might be beneficial. We plan to attach miniature IMUs at the head to detect head and yaw movements and to exploit motion characteristics for real-time swallow detection. A clear advantage of all discussed measurement systems is that they are wearable by the patient and therefore not bounded to certain locations, as it is generally the case with camera systems. For stationary applications, the combination of the proposed systems with camera systems might be considered to increase accuracy and robustness of the measurement system and to reduce the amount of body-worn sensors as demonstrated for

example by Tannous et al. (2016) for the lower limbs and by Atrsaei, Salarieh, and Alasty (2016) for the upper limbs. Initial work in this direction has been conducted by us for improving the swallow detection during eating at a table. The BI/EMGbased classifier was enhanced by features extracted from a deep imaging camera. The observed motion speed and opening width of the mouth proved to be promising features in this context (Riebold, Nahrstaedt, Schultheiss, Seidl, & Schauer, 2016). Future research on inertial motion analysis should continue in exploiting kinematic constraints and the coupling of joint movements to enable automatic sensor-to-segment calibration and joint axes identification from arbitrary movements and to enhance the detection and compensation of internal and external inertial sensor disturbances discussed in Section 2.1. We will reduce the number of IMUs of our hand motion measurement system by taking the couplings of finger joint movements into account. Additional comparative studies of inertial and optical motion capture systems on larger patient populations with motor impairments of different pathologies are necessary before using inertial measurement systems on a regular basis in clinics. Ongoing research is concerned with the design of combined stimulation and EMG measurement systems that also allow an assessment of the M-wave for an EMG measurement from the stimulation electrodes. A possible solution is the use of active discharging, either by hybrid stimulators, which can switch from current to voltage control, or by the use of adaptable pulse forms, which change from pulse to pulse. Feedback control of FES-evoked EMG for several muscles close to each other represents another challenge due to crosstalk in the EMG measurement. Recently, we proposed a first time-multiplexed solution for a setup with two neighboring muscles (Klauer & Schauer, 2016). Iterative learning control showed to be a successful strategy for controlling cyclic processes like walking or swallowing. The feedback control of several degrees of freedom (several joints) is a subject of ongoing research. The closed-loop drop foot stimulation system from Section 3.2 is currently extended by us for the knee and hip joint (Müller, Balligand, Seel, & Schauer, 2017a). Additional challenges arise in this context from the use of wireless body area networks. Limited communication bandwidth, time varying delays and the possible loss of data packages within the network need

Please cite this article as: T. Schauer, Sensing motion and muscle activity for feedback control of functional electrical stimulation: Ten years of experience in Berlin, Annual Reviews in Control (2017), https://doi.org/10.1016/j.arcontrol.2017.09.014

JID: JARAP 16

ARTICLE IN PRESS

[m5G;October 5, 2017;12:12]

T. Schauer / Annual Reviews in Control 000 (2017) 1–20

to be taken into account when designing control systems. Eventbased sampling represents one possible approach to mitigate these problems (Laidig, Trimpe, & Seel, 2016). The use of electrode arrays might solve problems of limited sensitivity and might offer measures for alternating motor unit activation in order to reduce fatigue. However, the control of electrode arrays remain a challenging problem due to the high complexity of the control problem (up to 64 or even more inputs). When controlling the hand, also the number of outputs (joint angles of the fingers and the hand) is very large. First solution approaches for this problem have been presented recently (see e.g. Freeman, 2015; Freeman, Yang, Tudor, and Kutlu, 2016 and Salchow, Valtin, Seel, & Schauer, 2016). The largest problem of FES remains the fast progression of muscle fatigue. The effects of it can only be compensated by feedback control until the maximum tolerated stimulation intensity is reached. One emerging approach to handle this problem is the combination of rehabilitation robotics and functional electrical stimulation. Rehabilitation robots never fatigue and execute repetitive movements with high precision, while FES actively involves the paretic muscles in the movement. The idea is to share the movement support between FES and robotics to postpone the onset of FES-induced fatigue and to increase the support delivered by the robot when FES becomes inefficient. A combination of the FES-based arm weight relief presented in Section 3.3 with a ropeactuated robotic systems was recently presented by us (MeyerRachner, Passon, Klauer, & Schauer, 2017). An interesting overview of so-called hybrid approaches (combination of robots and FES) for stroke patients can be found in Resquín et al. (2016). The proposed sensor approaches can give the rehabilitation practitioner and patient real-time feedback about any patientinduced muscle activity during the FES exercise. A better understanding of the problem and the effect that the stimulation has on the patient’s motor control can be obtained. The long-term goal is an automated adjustment of FES support in form of a “virtual therapist” with the aim to accelerate the rehabilitation and to restore physiological movement patterns. 5. Conclusions Functional electrical stimulation is already an effective technical aid for restoration of lost motor function, e.g. in spinal cord injured people and chronic stroke patients. Additionally, it promotes the rehabilitation process in acute and subacute stroke patients. Unfortunately, an intense use of advanced (multi-channel) FES in clinics and at home cannot be observed up to now. This is mainly due to the required time-consuming manual tuning for the typically used open-loop systems. Residual motor skills of the user and changing conditions like muscular fatigue are not considered in the existing systems. Feedback control of FES has been studied for decades by several research groups and its potential in creating patient- and situation-specific solutions has been successfully demonstrated. A major reason for the delayed clinical transfer of these research results can be seen in a lack of wearable measurement systems for robust real-time monitoring of the patient’s movement and muscle activity. The reviewed contributions to inertial motion analysis as well as to EMG and BI signal processing substantially mitigate this problem and therefore enhance the transferability of feedbackcontrolled FES from research labs to clinical practice and home settings. The developed inertial sensor-to-segment calibration takes into account the limited ability of motor-impaired patients to precisely place sensors at specified positions and to perform precise calibration motions. Instead, arbitrary sensor placements are allowed, and the sensor calibration is automatically performed while the user performs arbitrary movements. The rapid assessment of FESinduced muscle contractions by acceleration or electromyography

measurements enables the realization of novel cascaded control schemes. These linearize the nonlinear muscle dynamics and provide robust control performance also in the presence of model uncertainties and time-variant dynamics. The presented FES systems for cycling, walking, reaching and swallowing have exemplified the use of the developed measurement systems. With this review, the author hopes to inspire researchers to apply the outlined sensor approaches and control strategies to other motor-function-related FES systems. Manufacturers are encouraged to integrate EMG sensing together with the control of FES-induced muscle contraction directly into their stimulators to enhance the performance of future FES system. Funding sources Work presented in this contribution was funded by the German Federal Ministry of Education and Research (BMBF) within the projects APeroStim (FKZ 01EZ1204B), BigDysPro (FKZ 13EZ1007A), BeMobil (FKZ 16SV7069K) and SelfFEES (FKZ 13GW0021C), by the German Federal Ministry of Economics and Energy (BMWi) within the project MultiEMBI (KF 2392314CS4), and by the European Union’s Horizon 2020 Programme for Research and Innovation in the project RETRAINER under grant agreement no. 644721. Acknowledgments I would like to thank my colleagues Christian Klauer, Philipp Müller, Dr. Holger Nahrstaedt, Arne Passon, Benjamin Riebold, Christina Salchow, Dr. Thomas Seel, Dr. Raafat Shalaby, Markus Valtin and Constantin Wiesener from the Control Systems Group at TU Berlin for their valuable contribution in the development of the methods and systems presented in this contribution. I am furthermore indebted to our technician Astrid Bergmann for her support in realizing the experimental settings. Finally, I would also like to acknowledge our clinical partners Dr. Rainer O. Seidl, Dr. Corinna Schultheiss, Dr. Andreas Niedeggen, Dr. Ingo Schmehl, Dr. Frank Dähne and Dr. Sebatian Böttcher from the Unfallkrankenhaus Berlin as well as Prof. Stefan Hesse and Cordula Werner from the Charité Berlin, who delivered ideas and clinical requirements for the developments and carried out the clinical validation of the FES systems. Finally, I would like to express my deep gratitude to the patients who participated in our clinical trials. References Ambrosini, E., Ferrante, S., Schauer, T., Klauer, C., Gaffuri, M., Ferrigno, G., et al. (2014). A myocontrolled neuroprosthesis integrated with a passive exoskeleton to support upper limb activities. Journal of Electromyography and Kinesiology, 24(2), 307–317. doi:10.1016/j.jelekin.2014.01.006. Ang, W. T., Khosla, P. K., & Riviere, C. N. (2004). Kalman filtering for real-time orientation tracking of handheld microsurgical instrument. In Proceedings of the IEEE/RSJ international conference on intelligent robots and systems (IROS 2004): 3 (pp. 2574–2580). doi:10.1109/IROS.2004.1389796. Atrsaei, A., Salarieh, H., & Alasty, A. (2016). Human Arm Motion Tracking by Orientation-Based Fusion of Inertial Sensors and Kinect Using Unscented Kalman Filter. Journal of Biomech Engineering, 138(9). doi:10.1115/1.4034170. Barsi, G. I., Popovic, D. B., Tarkka, I. M., Sinkjaer, T., & Grey, M. J. (2008). Cortical excitability changes following grasping exercise augmented with electrical stimulation. Experimental Brain Research, 191(1), 57–66. doi:10.1007/ s0 0221-0 08-1495-5. Broniatowski, M., Moore, N., Grundfest-Broniatowski, S., Tucker, H., Lancaster, E., Krival, K., et al. (2010). Paced glottic closure for controlling aspiration pneumonia in patients with neurologic deficits of various causes. Annals of Otology, Rhinology and Laryngology, 119(3), 141–149. Burnett, T. A., Mann, E. A., Cornell, S. A., & Ludlow, C. L. (2003). Laryngeal elevation achieved by neuromuscular stimulation at rest. Journal of Applied Physiology (Bethesda, Md.: 1985), 94(1), 128–134. doi:10.1152/japplphysiol.0 0406.20 02. Burnett, T. A., Mann, E. A., Stoklosa, J. B., & Ludlow, C. L. (2005). Self-triggered functional electrical stimulation during swallowing. Journal of Neurophysiology, 94, 4011–4018. doi:10.1152/jn.0 0 025.20 05. Chen, Y.-W., Chang, K.-H., Chen, H.-C., Liang, W.-M., Wang, Y.-H., & Lin, Y.-N. (2016). The effects of surface neuromuscular electrical stimulation on post-stroke dysphagia: A systemic review and meta-analysis. Clinical Rehabilitation, 30(1), 24– 35. doi:10.1177/0269215515571681.

Please cite this article as: T. Schauer, Sensing motion and muscle activity for feedback control of functional electrical stimulation: Ten years of experience in Berlin, Annual Reviews in Control (2017), https://doi.org/10.1016/j.arcontrol.2017.09.014

JID: JARAP

ARTICLE IN PRESS T. Schauer / Annual Reviews in Control 000 (2017) 1–20

Chia Bejarano, N., Ambrosini, E., Pedrocchi, A., Ferrigno, G., Monticone, M., & Ferrante, S. (2015). A novel adaptive, real-time algorithm to detect gait events from wearable sensors. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 23(3), 413–422. doi:10.1109/TNSRE.2014.2337914. Daly, J., Byron, E., Mendell, L., Rymer, W., Stefanocska, A., Wolpaw, J., et al. (1996). Therapeutic neural effects of electrical stimulation. IEEE Transactions on Rehabilitation Engineering, 4(4), 218–230. doi:10.1109/86.547922. Davis, G. M., Hamzaid, N. A., & Fornusek, C. (2008). Cardiorespiratory, metabolic, and biomechanical responses during functional electrical stimulation leg exercise: Health and fitness benefits. Artificial Organs, 32(8), 625–629. doi:10.1111/j. 1525-1594.20 08.0 0622.x. De Luca, C. F., & Knaflitz, M. (1992). Surface electromyography, what’s new?. Turino, Italy: CLUT Publishers. De Vries, W., Veeger, H., Baten, C., & Van Der Helm, F. (2009). Magnetic distortion in motion labs, implications for validating inertial magnetic sensors. Gait & Posture, 29(4), 535–541. doi:10.1016/j.gaitpost.20 08.12.0 04. Dunning, K., O’Dell, M. W., Kluding, P., & McBride, K. (2015). Peroneal stimulation for foot drop after stroke: A systematic review. American Journal of Physical Medicine and Rehabilitation, 94(8), 649–664. doi:10.1097/PHM.0 0 0 0 0 0 0 0 0 0 0 0 0308. Durfee, W., & MacLean, K. (1989). Methods for estimating isometric recruitment curves of electrically stimulated muscle. IEEE Transactions on Biomedical Engineering, 36(7), 654–667. doi:10.1109/10.32097. Durfee, W. K. (1989). Task-based methods for evaluating electrically stimulated antagonist muscle controllers. IEEE Transactions on Biomedical Engineering, 36(3), 309–321. doi:10.1109/10.19852. El Hadri, A., & Benallegue, A. (2009). Sliding mode observer to estimate both the attitude and the gyro-bias by using low-cost sensors. In Proceedings of the IEEE/RSJ international conference on intelligent robots and systems (IROS 2009) (pp. 2867–2872). Eraifej, J., Clark, W., France, B., Desando, S., & Moore, D. (2017). Effectiveness of upper limb functional electrical stimulation after stroke for the improvement of activities of daily living and motor function: A systematic review and metaanalysis. Systematic Reviews, 6(1), 40. doi:10.1186/s13643- 017- 0435- 5. Feigin, V. L., Mensah, G. A., Norrving, B., Murray, C. J. L., Roth, G. A., & GBD 2013 Stroke Panel Experts Group (2015). Atlas of the global burden of stroke (1990– 2013): the GBD 2013 study. Neuroepidemiology, 45(3), 230–236. doi:10.1159/ 0 0 0441106. Fourati, H., Manamanni, N., Afilal, L., & Handrich, Y. (2014). Complementary observer for body segments motion capturing by inertial and magnetic sensors. IEEE/ASME Transactions on Mechatronics, 19(1), 149–157. doi:10.1109/TMECH. 2012.2225151. Freed, M. L., Freed, L., Chatburn, R. L., & Christian, M. (2001). Electrical stimulation for swallowing disorders caused by stroke. Respiratory Care, 46(5), 466–474. Freeman, C. (2015). Control system design for electrical stimulation in upper limb rehabilitation: Modelling, identification and robust performance (1st). New York, NY: Springer. Freeman, C., Yang, K., Tudor, J., & Kutlu, M. (2016). Feedback control of electrical stimulation electrode arrays. Medical Engineering and Physics, 38(11), 1185– 1194. doi:10.1016/j.medengphy.2016.07.002. Gandolla, M., Ferrante, S., Molteni, F., Guanziroli, E., Frattini, T., Martegani, A., et al. (2014). Re-thinking the role of motor cortex: Context-sensitive motor outputs? Neuroimage, 91, 366–374. doi:10.1016/j.neuroimage.2014.01.011. Gorman, P., & Mortimer, J. (1983). The effect of stimulus parameters on the recruitment characteristics of direct nerve stimulation. IEEE Transactions on Biomedical Engineering, 30(7), 301–308. doi:10.1109/TBME.1983.325041. Graupe, D. (2002). An overview of the state of the art of noninvasive FES for independent ambulation by thoracic level paraplegics. Neurological Research, 24(5), 431–442. doi:10.1179/016164102101200302. Graupe, D., Cerrel-Bazo, H., Kern, H., & Carraro, U. (2008). Walking performance, medical outcomes and patient training in fes of innervated muscles for ambulation by thoracic-level complete paraplegics. Neurological Research, 30(2), 123– 130. doi:10.1179/174313208X281136. Graurock, D., Schauer, T., & Seel, T. (2016a). Automatic pairing of inertial sensors to lower limb segments – a plug-and-play approach. Current Directions in Biomedical Engineering, 2(1), 715–718. doi:10.1515/cdbme- 2016- 0155. Graurock, D., Schauer, T., & Seel, T. (2016b). User-adaptive inertial sensor network for feedback-controlled gait support systems. In Proceedings of the 20th IFESS conference, La Grande Motte, France. Grimnes, S., & Martinsen, Ø. G. (2008). Bioimpedance and bioelectricity basics. Academic Press. Hadley, A. J., Kolb, I., & Tyler, D. J. (2013). Laryngeal elevation by selective stimulation of the hypoglossal nerve. Journal of Neural Engineering, 10(4), 046013. doi:10.1088/1741-2560/10/4/046013. Hara, Y. (2008). Neurorehabilitation with new functional electrical stimulation for hemiparetic upper extremity in stroke patients. Journal of Nippon Medical School, 75(1), 4–14. doi:10.1272/jnms.75.4. Hara, Y., Ogawa, S., Tsujiuchi, K., & Muraoka, Y. (2008). A home-based rehabilitation program for the hemiplegic upper extremity by power-assisted functional electrical stimulation. Disability & Rehabilitation, 30(4), 296–304. doi:10.1080/ 09638280701265539. Hayashibe, M. (2016). Evoked electromyographically controlled electrical stimulation. Frontiers in Neuroscience, 10, 335. doi:10.3389/fnins.2016.00335. Henneman, E. (1981). Recruitment of motoneurones: The size principle. In J. E. Desmedt (Ed.), Motor unit types, recruitment and plasticity in health and disease. progress in clinical neurophysiology (pp. 26–90). Basel: Karger.

[m5G;October 5, 2017;12:12] 17

Hettinga, D. M., & Andrews, B. J. (2007). The feasibility of functional electrical stimulation indoor rowing for high-energy training and sport. Neuromodulation, 10(3), 291–297. doi:10.1111/j.1525-1403.20 07.0 0117.x. Ho, C. H., Triolo, R. J., Elias, A. L., Kilgore, K. L., DiMarco, A. F., Bogie, K., et al. (2014). Functional electrical stimulation and spinal cord injury. Physical Medicine and Rehabilitation Clinics of North America, 25(3), 631–654,ix. doi:10.1016/j.pmr.2014. 05.001. Howlett, O. A., Lannin, N. A., Ada, L., & McKinstry, C. (2015). Functional electrical stimulation improves activity after stroke: A systematic review with metaanalysis. Archives of Physical Medicine and Rehabilitation, 96(5), 934–943. doi:10. 1016/j.apmr.2015.01.013. Hunt, K. J., Fang, J., Saengsuwan, J., Grob, M., & Laubacher, M. (2012). On the efficiency of FES cycling: A framework and systematic review. Technology and Health Care, 20, 395–422. doi:10.3233/THC- 2012- 0689. Janssen, T., Glaser, R., & Shuster, D. (1998). Clinical efficacy of electrical stimulation exercise training: Effects on health, fitness, and function. Topics in Spinal Cord Injury Rehabilitation, 3(3), 33–49. Jarvis, J. (1993). Power production and working capacity of rabbit tibialis anterior muscle after chronic electrical stimulation at 10 Hz. Journal of Physiology, 470, 157–169. doi:10.1113/jphysiol.1993.sp019852. Kafri, M., & Laufer, Y. (2015). Therapeutic effects of functional electrical stimulation on gait in individuals post-stroke. Annals of Biomedical Engineering, 43(2), 451– 466. doi:10.1007/s10439- 014- 1148- 8. Kim, S., Nam, K., Kim, D., Ryu, C., Kim, Y., & Kim, J. (2003). Optimum electrode configuration for detection of arm movement using bio-impedance. Medical and Biological Engineering and Computing, 41(2), 141–145. Klauer, C., Ferrante, S., Ambrosini, E., Shiri, U., Dähne, F., Schmehl, I., et al. (2016). A patient-controlled functional electrical stimulation system for arm weight relief. Med Eng Phys, 38(11), 1232–1243. doi:10.1016/j.medengphy.2016. 06.006. Klauer, C., Irmer, M., & Schauer, T. (2015). A muscle model for hybrid muscle activation. Current Directions in Biomedical Engineering, 1(1), 386–389. doi:10.1515/ cdbme- 2015- 0094. Klauer, C., Raisch, J., & Schauer, T. (2012). Linearisation of electrically stimulated muscles by feedback control of the muscular recruitment measured by evoked EMG. In Proceedings of the 17th international conference on methods and models in automation and robotics (MMAR) (pp. 108–113). IEEE. doi:10.1109/MMAR.2012. 6347902. Klauer, C., Raisch, J., & Schauer, T. (2013). Nonlinear joint-angle feedback control of electrically stimulated and λ-controlled antagonistic muscle pairs. In Proceedings of European control conference (ECC) (pp. 3101–3107). IEEE. Klauer, C., Ruppel, M., Böttcher, S., Dähne, F., Schmehl, I., & Schauer, T. (2017). FES-based arm weight relief: First investigations in stroke patients. In Current directions in biomedical engineering: 3 (p. s5). De Gruyter. doi:10.1515/ cdbme- 2017- 1001. Suppl.: Abstracts of the TAR Conference Klauer, C., & Schauer, T. (2016). Two-channel muscle recruitment λ-Control using the evoked-EMG. In Proceedings of the 20th IFESS conference, La Grande Motte, France. Klauer, C., & Schauer, T. (2017). Adaptive control of a neuroprosthesis for stroke patients amplifying weak residual shoulder-muscle activity. In Proceedings of Preprints of the 20th IFAC world congress, Toulouse, France (pp. 9122–9127). Klauer, C., Schauer, T., & Raisch, J. (2011). Joint-Angle control by electrical stimulation of antagonistic muscles (in German). at - Automatisierungstechnik, 59(10), 629–637. doi:10.1524/auto.2011.0955. Kortier, H. G., Sluiter, V. I., Roetenberg, D., & Veltink, P. H. (2014). Assessment of hand kinematics using inertial and magnetic sensors. Journal of Neuroengineering and Rehabilitation, 11, 70. doi:10.1186/1743-0 0 03-11-70. Koutsou, A. D., Moreno, J. C., del Ama, A. J., Rocon, E., & Pons, J. L. (2016). Advances in selective activation of muscles for non-invasive motor neuroprostheses. Journal of Neuroengineering and Rehabilitation, 13(1), 56. doi:10.1186/ s12984- 016- 0165- 2. Laidig, D., Müller, P., & Seel, T. (2017a). Automatic anatomical calibration for IMU-based elbow angle measurement in disturbed magnetic fields. Current Directions in Biomedical Engineering, 3(2), 167–170. doi:10.1515/cdbme-20170035. Laidig, D., Schauer, T., & Seel, T. (2017b). Exploiting kinematic constraints to compensate magnetic disturbances when calculating joint angles of approximate hinge joints from orientation estimates of inertial sensors. In Proceedings of the 15th IEEE conference on rehabilitation robotics (ICORR), London, UK (pp. 971–976). doi:10.1109/ICORR.2017.8009375. Laidig, D., Trimpe, S., & Seel, T. (2016). Event-based sampling for reducing communication load in realtime human motion analysis by wireless inertial sensor networks. Current Directions in Biomedical Engineering, 2(1), 711–714. doi:10.1515/ cdbme- 2016- 0154. Leelamanit, V., Limsakul, C., & Geater, A. (2002). Synchronized electrical stimulation in treating pharyngeal dysphagia. The Laryngoscope, 112(12), 2204–2210. doi:10. 1097/0 0 0 05537-20 02120 0 0-0 0 015. Liberson, W., Holmquest, H., Scot, D., & Dow, M. (1961). Functional electrotherapy: Stimulation of the peroneal nerve synchronized with the swing phase of the gait of hemiplegic patients. Archives of Physical Medicine and Rehabilitation, 21, 101–105. Ludlow, C. L., Humbert, I., Saxon, K., Poletto, C., Sonies, B., & Crujido, L. (2007). Effects of surface electrical stimulation both at rest and during swallowing in chronic pharyngeal dysphagia. Dysphagia, 22(1), 1–10. doi:10.1007/ s0 0455-0 06-9029-4.

Please cite this article as: T. Schauer, Sensing motion and muscle activity for feedback control of functional electrical stimulation: Ten years of experience in Berlin, Annual Reviews in Control (2017), https://doi.org/10.1016/j.arcontrol.2017.09.014

JID: JARAP 18

ARTICLE IN PRESS

[m5G;October 5, 2017;12:12]

T. Schauer / Annual Reviews in Control 000 (2017) 1–20

Lynch, C. L., Popovic, M. R., & Rushton, D. (2008). Functional electrical stimulation. IEEE Control Systems Magazine, 28(2), 40–50. doi:10.1109/MCS.2007.914689. Lyons, G. M., Sinkjaer, T., Burridge, J. H., & Wilcox, D. J. (2002). A review of portable FES-based neural orthoses for the correction of drop foot. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 10(4), 260–279. doi:10.1109/ TNSRE.2002.806832. Mackay, J., Mensah, G. A., Mendis, S., & Greenlund, K. (2004). The atlas of heart disease and stroke. Geneva: World Health Organization. Madgwick, S. O. H., Harrison, A. J. L., & Vaidyanathan, R. (2011). Estimation of IMU and MARG orientation using a gradient descent algorithm. In Proceedings of the international IEEE conference on rehabilitation robotics (ICORR) (pp. 179–185). IEEE. doi:10.1109/ICORR.2011.5975346. Mangold, S., Keller, T., Curt, A., & Dietz, V. (2005). Transcutaneous functional electrical stimulation for grasping in subjects with cervical spinal cord injury. Spinal Cord, 43(1), 1–13. doi:10.1038/sj.sc.3101644. Melo, P. L., Silva, M. T., Martins, J. M., & Newman, D. J. (2015). Technical developments of functional electrical stimulation to correct drop foot: Sensing, actuation and control strategies. Clinical Biomechanics (Bristol, Avon), 30(2), 101–113. doi:10.1016/j.clinbiomech.2014.11.007. Merletti, R., & Farina, D. (2016). Surface electromyography: Physiology, engineering and applications. Wiley-IEEE Press. (IEEE press series on biomedical engineering) Merletti, R., Knaflitz, M., & DeLuca, C. J. (1992). Electrically evoked myoelectric signals. Critical Reviews in Biomedical Engineering, 19(4), 293–340. Meyer-Rachner, P., Passon, A., Klauer, C., & Schauer, T. (2017). Compensating the effects of FES-induced muscle fatigue by rehabilitation robotics during arm weight support. Current Directions in Biomedical Engineering, 3(1), 31–34. doi:10. 1515/cdbme- 2017- 0 0 07. Mi, P., Du, Q., Ye, L., & Zou, W. (2010). An adaptive fast quaternion-based human motion tracking algorithm with inertial/magnetic technology. In Proceedings of the 10th international conference on signal processing (ICSP) (pp. 1252–1258). IEEE. Müller, P., Balligand, C., Seel, T., & Schauer, T. (2017a). Iterative learning control and system identification of the antagonistic knee muscle complex during gait using functional electrical stimulation. In Proceedings of Preprints of the 20th IFAC world congress (pp. 9116–9121). Müller, P., Bégin, M. A., Schauer, T., & Seel, T. (2017b). Alignment-free, self-calibrating elbow angles measurement using inertial sensors. IEEE Journal of Biomedical and Health Informatics, 21(2), 312–319. doi:10.1109/JBHI.2016.2639537. Muraoka, Y. (2002). Development of an EMG recording device from stimulation electrodes for functional electrical stimulation. Frontiers of Medical and Biological Engineering, 11(4), 323–333. Nahrstaedt, H. (2017). Automatic detection and assessment of swallowing based on bioimpedance and electromyography measurements – enabling control of functional electrical stimulation synchronously to volitional swallowing in dysphagic patients. Dissertation, Technische Universität Berlin. Nahrstaedt, H., & Schauer, T. (2009). A bioimpedance measurement device for sensing force and position in neuroprosthetic systems. In Proceedings of 4th European Conference of the International Federation for Medical and Biological Engineering (pp. 1642–1645). Springer. doi:10.1007/978- 3- 540- 89208- 3_390. Nahrstaedt, H., Schauer, T., Shalaby, R., Hesse, S., & Raisch, J. (2008). Automatic control of a drop-foot stimulator based on angle measurement using bioimpedance. Artificial Organs, 32(8), 649–654. doi:10.1111/j.1525-1594.20 08.0 0617.x. Nahrstaedt, H., Schultheiss, C., Schauer, T., & Seidl, R. (2013). Bioimpedance- and EMG-triggered FES for improved protection of the airway during swallowing. Biomedical Technology (BERL), 58(Suppl. 1). doi:10.1515/bmt- 2013- 4025. Nahrstaedt, H., Schultheiss, C., Seidl, R., & Schauer, T. (2012). Swallow detection algorithm based on bioimpedance and EMG measurements. IFAC Proceedings Volumes, 45(18), 91–96. doi:10.3182/20120829- 3- HU- 2029.0 0 067. Newham, D. J., & Donaldson, N. d. N. (2007). FES cycling. Acta neurochirurgica. Supplement, 97(Pt 1), 395–402. doi:10.10072F978- 3- 211- 33079- 1_52. van den Noort, J. C., Kortier, H. G., Beek, N. v., Veeger, D. H. E. J., & Veltink, P. H. (2016). Measuring 3D hand and finger kinematics – A comparison between inertial sensing and an opto-electronic marker system. PLoS One, 11(11), e0164889. doi:10.1371/journal.pone.0164889. Pandyan, A., Gregoric, M., Barnes, M., Wood, D., Wijck, F. V., Burridge, J., et al. (2005). Spasticity: Clinical perceptions, neurological realities and meaningful measurement. Disability and Rehabilitation, 27(1–2), 2–6. doi:10.1080/ 0963828040 0 014576. Patil, S., Raza, W. A., Jamil, F., Caley, R., & O’Connor, R. J. (2014). Functional electrical stimulation for the upper limb in tetraplegic spinal cord injury: A systematic review. Journal of Medical Engineering and Technology, 39(7), 419–423. doi:10. 3109/03091902.2015.1088095. Peckham, P. H., & Knutson, J. S. (2005). Functional electrical stimulation for neuromuscular applications. Annual Review of Biomedical Engineering, 7, 327–360. doi:10.1146/annurev.bioeng.6.040803.140103. Pereira, S., Mehta, S., McIntyre, A., Lobo, L., & Teasell, R. W. (2012). Functional electrical stimulation for improving gait in persons with chronic stroke. Topics in Stroke Rehabilitation, 19(6), 491–498. doi:10.1310/tsr1906-491. Popovic, M. R., Popovic, D. B., & Keller, T. (2002). Neuroprostheses for grasping. Neurological Research, 24(5), 443–452. doi:10.1179/016164102101200311. Resquín, F., Gómez, A. C., Gonzalez-Vargas, J., Brunetti, F., Torricelli, D., & Rueda, F. M. (2016). Hybrid robotic systems for upper limb rehabilitation after stroke: A review. Medical Engineering and Physics, 38(11), 1279–1288. doi:10. 1016/j.medengphy.2016.09.001.

Ridding, M. C., Brouwer, B., Miles, T. S., Pitcher, J. B., & Thompson, P. D. (20 0 0). Changes in muscle responses to stimulation of the motor cortex induced by peripheral nerve stimulation in human subjects. Experimental Brain Research, 131(1), 135–143. Riebold, B., Nahrstaedt, N., Schultheiss, C., Seidl, R., & Schauer, T. (2016). Multisensor classification system for triggering FES in order to support voluntary swallowing. European Journal of Translational Myology, 26(4), 287–291. doi:10.4081/ejtm. 2016.6224. Riener, R. (1999). Model-based development of neuroprostheses for paraplegic patients. Philosophical Transactions of the Royal Society of London. Series B: Biological Sciences, 354(1385), 877–894. doi:10.1098/rstb.1999.0440. Rueterbories, J., Spaich, E. G., & Andersen, O. K. (2014). Gait event detection for use in FES rehabilitation by radial and tangential foot accelerations. Medical Engineering and Physics, 36(4), 502–508. doi:10.1016/j.medengphy.2013.10.004. Ruppin, S., Wiesener, C., & Schauer, T. (2016). Inertial sensor-based control of functional electrical stimulation in paraplegic cycling. In Proceedings of the 20th IFESS conference, La Grande Motte, Francerande Motte (pp. 1–6). Rushton, D. (1997). Functional electrical stimulation. Physiological Measurement, 18, 241–275. S. Binder-MacLeod, & T. Kesar (2005). Catchlike property of skeletal muscle: Recent findings and clinical implications. Muscle Nerve, 31(6), 681–693. doi:10.1002/ mus.20290. Sabut, S. K., Bhattacharya, S. D., & Manjunatha, M. (2013). Functional electrical stimulation on improving foot drop gait in poststroke rehabilitation: A review of its technology and clinical efficacy. Critical Reviews in Biomedical Engineering, 41(2), 149–160. doi:10.1615/CritRevBiomedEng.2013007621. Salchow, C., Valtin, M., Seel, T., & Schauer, T. (2016). A new semi-automatic approach to find suitable virtual electrodes in arrays using an interpolation strategy. European Journal of Translational Myology, 26(2), 134–139. doi:10.4081/ejtm.2016. 6029. Schauer, T. (2006). Feedback control of cycling in spinal cord injury using functional electrical stimulation. University of Glasgow, UK (Ph.D thesis). Schauer, T. (2014). Funktionelle Elektrostimulation nach Querschnittlähmung und Schlaganfall. In J. Werner (Ed.), Automatisierte Therapiesysteme (pp. 333–373). Berlin, Boston: De Gruyter CY. doi:10.1515/9783110252132.333. Schauer, T., Negard, N. O., Previdi, F., Hunt, K. J., Fraser, M. H., Ferchland, E., et al. (2005). Online identification and nonlinear control of the electrically stimulated quadriceps muscle. Control Engineering Practice, 13(9), 1207–1219. doi:10.1016/j. conengprac.20 04.10.0 06. Schauer, T., Seel, T., Bunt, N., Müller, P., & Moreno, J. (2016). Realtime EMG analysis for transcutaneous electrical stimulation assisted gait training in stroke patients. IFAC-PapersOnLine, 49(32), 183–187. doi:10.1016/j.ifacol.2016.12.211. Schultheiss, C., Schauer, T., Nahrstaedt, H., & Seidl, R. (2013). Evaluation of an EMG bioimpedance measurement system for recording and analysing the pharyngeal phase of swallowing. European Archives of Oto-Rhino-Laryngology, 270(7), 2149– 2156. doi:10.10 07/s0 0405- 013- 2406- 3. Schultheiss, C., Schauer, T., Nahrstaedt, H., & Seidl, R. (2014). Automated detection and evaluation of swallowing using a combined EMG/bioimpedance measurement system. ScientificWorldJournal, 2014, 405471. doi:10.1155/2014/405471. Schultheiss, C., Schauer, T., Nahrstaedt, H., Seidl, R., & Bieler, J. (2016). Efficacy of EMG/bioimpedance-triggered functional electrical stimulation on swallowing performance. European Journal of Translational Myology, 26(4), 283–286. doi:10.4081/ejtm.2016.6065. Schultheiss, C., Wolter, S., Schauer, T., Nahrstaedt, H., & Seidl, R. (2015). Effect of body position on coordination of breathing and swallowing. HNO, 63(6), 439– 446. doi:10.10 07/s0 0106-015-0016-7. Seel, T. (2016). Learning control and inertial realtime gait analysis in biomedical applications. Technische Universität Berlin Dissertation. Seel, T., Graurock, D., & Schauer, T. (2015a). Realtime assessment of foot orientation by accelerometers and gyroscopes. Current Directions in Biomedical Engineering, 1(1), 446–469. doi:10.1515/cdbme-2015-0112. Seel, T., Landgraf, L., Escobar, V. C., & Schauer, T. (2014a). Online gait phase detection with automatic adaption to gait velocity changes using accelerometers and gyroscopes. Biomedical Engineering / Biomedizinische Techik, 59(s1), S795–S798. doi:10.1515/bmt- 2014- 5011. Seel, T., Raisch, J., & Schauer, T. (2014b). IMU-Based Joint angle measurement for gait analysis. Sensors, 14(4), 6891–6909. doi:10.3390/s140406891. Seel, T., Ruppel, M., Valtin, M., & Schauer, T. (2015b). Multichannel FES parameterization for controlling foot motion in paretic gait. Current Directions in Biomedical Engineering, 1(1), 480–483. doi:10.1515/cdbme-2015-0115. Seel, T., & Ruppin, S. (2017). Eliminating the effect of magnetic disturbances on the inclination estimates of inertial sensors. In Proceedings of the Preprints of 20th IFAC world congress, Toulouse, France (pp. 9128–9133). Seel, T., Schauer, T., & Raisch, J. (2012). Joint axis and position estimation from inertial measurement data by exploiting kinematic constraints. In Proceedings of the 2012 IEEE International Conference on Control Applications (CCA) Part of 2012 IEEE Multi-Conference on Systems and Control (pp. 45–49). doi:10.1109/CCA.2012. 6402423. Seel, T., Schauer, T., & Raisch, J. (2017). Monotonic convergence of iterative learning control systems with variable pass length. International Journal of Control, 90(3), 1–14. doi:10.1080/00207179.2016.1183172. Seel, T., Valtin, M., Werner, C., & Schauer, T. (2015c). Multivariable control of foot motion during gait by peroneal nerve stimulation via two skin electrodes. IFACPapersOnLine, 48(20), 315–320. doi:10.1016/j.ifacol.2015.10.158.

Please cite this article as: T. Schauer, Sensing motion and muscle activity for feedback control of functional electrical stimulation: Ten years of experience in Berlin, Annual Reviews in Control (2017), https://doi.org/10.1016/j.arcontrol.2017.09.014

JID: JARAP

ARTICLE IN PRESS T. Schauer / Annual Reviews in Control 000 (2017) 1–20

Seel, T., Werner, C., Raisch, J., & Schauer, T. (2016a). Iterative learning control of a drop foot neuroprosthesis—generating physiological foot motion in paretic gait by automatic feedback control. Control Engineering Practice, 48, 87–97. doi:10. 1016/j.conengprac.2015.11.007. Seel, T., Werner, C., & Schauer, T. (2016b). The adaptive drop foot stimulator - Multivariable learning control of foot pitch and roll motion in paretic gait. Medical Engineering and Physics, 38(11), 1205–1213. doi:10.1016/j.medengphy.2016. 06.009. Sennels, S., Biering-Sorensen, F., Andersen, O. T., & Hansen, S. D. (1997). Functional neuromuscular stimulation controlled by surface electromyographic signals produced by volitional activation of the same muscle: Adaptive removal of the muscle response from the recorded EMG-signal. IEEE Transactions on Rehabilitation Engineering, 5(2), 195–206. doi:10.1109/86.593293. Shalaby, R. (2011). Development of an EMG detection system for the control of functional electrical stimulation in neurological rehabilitation. Technische Universität Berlin Dissertation. Shalaby, R., Schauer, T., Liedecke, W., & Raisch, J. (2011). Amplifier design for EMG recording from stimulation electrodes during functional electrical stimulation leg cycling ergometry. Biomedizinische Technik. Biomedical Engineering, 56(1), 23– 33. doi:10.1515/BMT.2010.055. Shiffman, C. A., Aaron, R., & Rutkove, S. B. (2003). Electrical impedance of muscle during isometric contraction. Physiological Measurement, 24(1), 213. Singh, A., Tetreault, L., Kalsi-Ryan, S., Nouri, A., & Fehlings, M. G. (2014). Global prevalence and incidence of traumatic spinal cord injury. Journal of Clinical Epidemiology, 6, 309–331. doi:10.2147/CLEP.S68889. Singh, S., & Hamdy, S. (2006). Dysphagia in stroke patients. Postgraduate Medical Journal, 82(968), 383–391. doi:10.1136%2Fpgmj.2005.043281. Song, C. G., Kim, S. C., Nam, K. C., & Kim, D. W. (2005). Optimum electrode configuration for detection of leg movement using bio-impedance. Physiological Measurement, 26(2), S59. Spagnol, P., Klauer, C., Previdi, F., Raisch, J., & Schauer, T. (2013). Modeling and online-identification of electrically stimulated antagonistic muscles for horizontal shoulder abduction and adduction. In Proceedings of European control conference (ECC) (pp. 3979–3984). IEEE. Stein, R. B., Peckham, P. H., & Popovic, D. P. (1992). Neural protheses: Replacing motor function after disease or disability. Oxford: Oxford University Press. Tannous, H., Istrate, D., Benlarbi-Delai, A., Sarrazin, J., Gamet, D., Ho Ba Tho, M. C., et al. (2016). A new multi-sensor fusion scheme to improve the accuracy of knee flexion kinematics for functional rehabilitation movements. Sensors (Basel), 16(11). doi:10.3390/s16111914. Thibaut, A., Chatelle, C., Ziegler, E., Bruno, M.-A., Laureys, S., & Gosseries, O. (2013). Spasticity after stroke: Physiology, assessment and treatment. Brain Injury, 27(10), 1093–1105. doi:10.3109/02699052.2013.804202. Thorsen, R., Occhi, E., Boccardi, S., & Ferrarin, M. (2006). Functional electrical stimulation reinforced tenodesis effect controlled by myoelectric activity from wrist extensors. Journal of Rehabilitation Research and Development, 43(2), 247–256. Thorsen, R., Puglia, M., & Ferrarin, M. (2013). Myoelectrically driven functional electrical stimulation may increase motor recovery of upper limb in poststroke subjects: A randomized controlled pilot study. Journal of Rehabilitation Research and Development, 50(6), 785. doi:10.1682/JRRD.2012.07.0123.

[m5G;October 5, 2017;12:12] 19

Thrasher, T. A., Zivanovic, V., McIlroy, W., & Popovic, M. R. (2008). Rehabilitation of reaching and grasping function in severe hemiplegic patients using functional electrical stimulation therapy. Neurorehabilitation and Neural Repair, 22(6), 706– 714. doi:10.1177/1545968308317436. ´ Truelsen, T., Piechowski-Józwiak, B., Bonita, R., Mathers, C., Bogousslavsky, J., & Boysen, G. (2006). Stroke incidence and prevalence in europe: A review of available data. European Journal of Neurology, 13(6), 581–598. doi:10.1111/j.1468-1331. 2006.01138.x. Vafadar, A. K., Côté, J. N., & Archambault, P. S. (2015). Effectiveness of functional electrical stimulation in improving clinical outcomes in the upper arm following stroke: A systematic review and meta-analysis. BioMed Research International, 2015, 729768. doi:10.1155/2015/729768. Valtin, M., Kociemba, K., Behling, C., Kuberski, B., Becker, S., & Schauer, T. (2016a). Rehamovepro: A versatile mobile stimulation system for transcutaneous FES applications. European Journal of Translational Myology, 26(3), 203–208. doi:10. 4081/ejtm.2016.6076. Valtin, M., Salchow, C., Seel, T., Laidig, D., & Schauer, T. (2017). Modular finger and hand motion capturing system based on inertial and magnetic sensors. Current Directions in Biomedical Engineering, 3(1), 19–23. doi:10.1515/ cdbme- 2017- 0 0 05. Valtin, M., Seel, T., Raisch, J., & Schauer, T. (2014). Iterative learning control of drop foot stimulation with array electrodes for selective muscle activation. IFAC Proceedings Volumes, 47(3), 6587–6592. doi:10.3182/20140824- 6- ZA- 1003.01991. Valtin, M., Werner, C., & Schauer, T. (2016b). Online assessment of volitional muscle activity from stimulation electrodes in a drop foot neuroprosthesis. In Proceedings of the 12th Vienna international workshop on FES (pp. 1–4). Vienna, Austria Vidaurre, C., Klauer, C., Schauer, T., Ramos-Murguialday, A., & Müller, K. (2016). EEGBased BCI for the linear control of an upper-limb neuroprosthesis. Medical Engineering and Physics, 38(11), 1195–1204. doi:10.1016/j.medengphy.2016.06.010. Wade, D. T., Wood, V. A., Heller, A., Maggs, J., & Hewer, R. L. (1987). Walking after stroke. measurement and recovery over the first 3 months. Scandinavian Journal of Rehabilitation Medicine, 19(1), 25–30. Wiesener, C., Ruppin, S., & Schauer, T. (2016). Robust discrimination of flexion and extension phases for mobile functional electrical stimulation (FES) induced cycling in paraplegics. IFAC-PapersOnLine, 49(32), 210–215. doi:10.1016/j.ifacol. 2016.12.216. Wiesener, C., & Schauer, T. (2017). Inertial sensor driven FES cycling – The Cybathlon-RehaBike by team Hasomed. IEEE Robotics & Automation Magazine, 24(3). Zhang, Z.-Q., Meng, X.-L., & Wu, J.-K. (2012). Quaternion-based Kalman filter with vector selection for accurate orientation tracking. IEEE Transactions on Instrumentation and Measurement, 61(10), 2817–2824. doi:10.1109/TIM.2012. 2196397. Zhou, B., Baratta, R., Solomonow, M., Olivier, L., Nguyen, G., & D’Ambrosia, R. (1996). Evaluation of isometric antagonist coactivation strategies of electrically stimulated muscles. IEEE Transactions on Biomedical Engineering, 43(2), 150–160. doi:10.1109/10.481984.

Please cite this article as: T. Schauer, Sensing motion and muscle activity for feedback control of functional electrical stimulation: Ten years of experience in Berlin, Annual Reviews in Control (2017), https://doi.org/10.1016/j.arcontrol.2017.09.014

JID: JARAP 20

ARTICLE IN PRESS

[m5G;October 5, 2017;12:12]

T. Schauer / Annual Reviews in Control 000 (2017) 1–20 Thomas Schauer (1974) studied electrical engineering at the Otto von Guericke University Magdeburg in Germany from 1992 to 1997. After this he spent three years at the University of Glasgow in Scotland at the Department of Mechanical Engineering in the Centre of Rehabilitation Engineering working on feedback control of cycling in spinal cord injury using functional electrical stimulation. In 2006, he received his Ph.D. degree from the University of Glasgow. From December 2001 until April 2006 he has been working as research assistant and project leader at the Max Planck Institute for Dynamics of Complex Technical Systems (Magdeburg, Germany) in the Systems and Control Theory Group. Since 2006 he holds a position as senior researcher in the Control Systems Group at the Technische Universität Berlin and manages the research topic “Rehabilitation Engineering and Assistive Technology”. Thomas Schauer is currently co-chair of the IFAC Technical Committee 8.2 for Biological and Medical Systems (BMS).

Please cite this article as: T. Schauer, Sensing motion and muscle activity for feedback control of functional electrical stimulation: Ten years of experience in Berlin, Annual Reviews in Control (2017), https://doi.org/10.1016/j.arcontrol.2017.09.014