Introduction
A key working hypothesis of rehabilitation science is that use-dependent plasticity perseveres through motor system injuries and diseases. This hypothesis drives intensive, ongoing efforts to optimize rehabilitation experiences for people with a movement disability, so as to best promote use-dependent plasticity. In the past twenty years, there has been an increasing recognition that technologies, including robotics, orthotics, wearable sensors, computer vision, computer gaming, electrical stimulation, virtual reality, machine learning, and computational modeling, can play an important role in these efforts [
1‐
5]. In this section, we first review the rationale for developing this new technology for rehabilitation therapy, then, using examples from robot-assisted therapy, we briefly characterize the state of the field in meeting its promise. In the following sections we then review approaches to improve these technologies, drawing on examples from European research, followed by a brief discussion of attempts to combine these technologies with biologic therapeutics.
There are three primary motivations for developing new technology for rehabilitation therapy. First, improved technology has the potential to allow more therapy with less supervision, improving rehabilitation cost-benefit profiles. This objective can be expressed as developing technology that optimally promotes use-dependent plasticity while lowering the cost of therapy. Second, technology has the potential to more accurately quantify therapy, including patient characteristics that predict therapy success, the dose and content of therapy, and clinical outcomes. This quantification property of technology is important for improving the mechanistic understanding of rehabilitation science, clinician decision-making, and patient feedback and motivation. Third, technology has the potential to allow entirely new types of therapy. One example is the concept of providing continuous therapy with wearable devices. Rehabilitation therapists cannot be omnipresent, but smart, wearable technology almost can, providing therapy throughout the day as people participate in activities of daily living. The therapeutic effect produced by functional electrical stimulation (FES) foot drop stimulators, in which people who use the stimulators over an extended period of time exhibit improved walking ability even when they turn the stimulator off, is one example [
6]. Another example of a promising new therapy that technology makes possible is manipulating limbs with robots in a way that precisely augments kinematic errors and thus enhances error-based learning [
7].
Aiming to achieve these three goals, there has been a rapid increase in the development of therapeutic technology in the past 10 years, and a rapid growth in commercial products for rehabilitation training [
1‐
5]. However, results with this technology are mixed so far, and when and in what form this technology will deliver the desired improved outcomes for rehabilitation is unclear. We illustrate the state of the field with three recent studies of robot-assisted movement training after chronic stroke.
Most clinical trials of robot-assisted movement training have used robots to physically assist the limbs of patients as they attempt desired movements and/or play computer-guided activities and games. Thus, the robots tested so far have typically implemented a technique from physical rehabilitation called "active assist therapy", in which the patient actively tries to achieve a movement as the therapist manually assists in the movement. Besides allowing a patient to perform movements not possible without assistance, it is thought that active assist therapy may generate new patterns of sensory input that may influence brain plasticity.
The first robot therapy study that illustrates the state of the field is the recent multi-center randomized controlled trial of robot-assisted therapy sponsored by the Department of Veterans Affairs [
8]. In this study, 127 people with chronic stroke were randomized to receive either 1) robot-assisted upper extremity training with three modules of the MIT-Manus robot; 2) upper extremity exercise with a rehabilitation therapist that was matched in number of movements to the MIT-Manus therapy and therefore was characterized as "intense", or 3) usual care. Robot-assisted therapy was significantly more effective than usual care, but the benefits were small--about two additional points on the upper extremity Fugl-Meyer scale, which ranges from 0 for complete paralysis to 66 for normal movement ability [
9]. Robot-assisted therapy was about as effective as the intense, therapist-delivered therapy, although as follow-up time progressed the patients who received the robot-assisted therapy exhibited a trend of more motor gains. The cost of delivery of the robotic and therapist-delivered therapies was similar, in large part because of the relatively high costs of the robots used in the study. However, the amount of therapy delivered was much greater than what would normally occur in an inpatient or outpatient rehabilitation setting. Thus, if the costs of robotics decrease it may be possible to deliver this therapy-intensive care, while delivering this type of care in the absence of robotics will likely never occur. Detailed analysis of the sensor-based data from this study is forthcoming, but, previous analysis of data obtained from similar MIT-Manus studies has been used to suggest that recovery is fundamentally characterized by a progressive blending of sub-movements [
10].
In another study, 48 people with chronic stroke who were ambulatory at study start were randomized to train walking using a treadmill and the Lokomat gait robot, or a treadmill with manual assistance from a physical therapist [
11]. For the Lokomat training, the participants did not receive biofeedback about their contribution to the walking motion. Training with either approach produced modest but measurable benefits in walking speed; training with the Lokomat and without biofeedback was about half as effective as the therapist-delivered training in terms of improvement of gait speed. It has been hypothesized that the relatively rigid robot assistance without biofeedback, as provided in this study, may have been less effective because it caused patient slacking [
11]; analysis of oxygen consumption during such training [
12], as well as computational modeling of the evolution of interaction forces during robot-assisted training [
13], quantitatively support this idea. Another possibility is that the rigid assistance reduced variability needed for learning [
14]. Analysis of training data from the study itself showed that Lokomat training as implemented was indeed less variable [
15]; analysis of fixed robotic gait training in rodents with SCI suggests that rigid assistance that does not allow kinematic variability tends to disrupt muscle activity [
16].
In a third study, 28 people with moderate to severe arm impairment due to a chronic stroke were randomized to participate in training with a passive arm exoskeleton called T-WREX or in standard table-top exercises with no technology [
17]. T-WREX simulated functional activities using computer games and an anti-gravity arm support orthosis which also incorporated a grip sensor that allowed patients with even trace amounts of grasp to participate in simple grasp-and release actions to control the games. Both groups only required about 3 minutes of therapist contact following a week of training, as measured by stopwatch. Both groups improved their arm movement significantly by about 2-3 FM points; at 6 month follow-up the T-WREX group had significantly better scores, although the difference was small (2 FM points). When given a chance to try the other therapy then asked to subjectively compare the two approaches, the participants expressed a strong preference for the technology-based approach, finding it more motivating in part because the arm weight support improved their perception of self-efficacy [
18].
One can extrapolate broader themes that characterize the general state of technology for rehabilitation therapy from these three illustrative studies. First, considering the goal of improving cost-benefit profiles, one can observe that technology-assisted exercise produced significant benefits in all three of the studies reviewed above, and that it is sometimes possible to use technology with only small amounts of therapist supervisory time, as directly measured in the T-WREX study. These observations support the premise that, indeed, some aspects of rehabilitation therapy do not require the immediate presence of a rehabilitation therapist to be effectively implemented with technology. However, the cost of the technology used for this substitution may still limit cost-benefit profiles, as was found, for example, in the cost analysis of the MIT-Manus study, indicating a need for lower cost technologies.
Second, considering the goal of quantification, while it is true that there is potentially a computer record of every force or movement the participants made during training in these studies, we are just beginning to understand how to use this data to predict responders, guide therapy, or define mechanisms of recovery. For example, as mentioned above, data from MIT-Manus has been used to identify a role for sub-movement blending in movement therapy, and data from the Lokomat and rodent robotic devices was used to analyze the role of kinematic variability in training. Thus, the field is just beginning to develop ways to use data from sensors incorporated into rehabilitation technology to provide insight into use-dependent plasticity.
Third, considering the goal of innovating to produce new forms of therapy that are more effective, it is apparent that some innovations in technology-based therapy are as effective as therapists for particular forms of training, few or none are more effective, and many are less effective. The reasons are complex and poorly understood at present, but a key limitation that must be overcome is improving the hardware and control design of these devices to increase efficacy. Understanding the reasons particular implementations decrement learning, while other implementations increment learning, is important. At present, one might say that the only innovation that new technologies routinely make available, besides semi-automation of training, is that of a more motivating context for rehabilitation training, by virtue of helping patients achieve movements or simulated activities that they normally could not, and by providing a computer gaming context with quantitative feedback to motivate practice.
Conclusions
There is an explosion in new rehabilitation technologies; however, the field is in its infancy. Beyond the fact that these technologies can in some case make rehabilitation exercise more engaging and less labor intensive, the gains delivered are still unclear. Fundamental scientific insight is needed into the learning and plasticity mechanisms that these technologies seek to stimulate; the current lack of insight makes device design somewhat haphazard. Nevertheless, promising areas of development include developing technology for delivering therapy both earlier in clinics, and later at home; investigating the relative roles of both simpler and more complex technology in promoting plasticity, thereby testing the premise that training with more naturalistic movements will better promote functional recovery; making devices wearable to extend the reach of training to the lived-in environment; improving feedback and implementing learning-based control to make training more engaging and challenging; and coordinating multiple therapeutic modalities, including robotics, FES, and BCI's to enhance the effect of training. A new field of neuro-computational rehabilitation appears to be developing, in which computational models will be used to simulate and understand use-dependent plasticity in rehabilitation therapy. Regenerative therapies may enable levels of recovery far beyond those possible with rehabilitation exercise alone, but these therapies cannot progress independently of rehabilitation exercise.
Thus, the challenge of developing technologies that significantly improve on rehabilitation outcomes compared to conventional rehabilitation remains to be met. The quantification power associated with sensors incorporated into therapeutic technologies, coupled with the nascent field of neuro-computational rehabilitation will help resolve this gap. We expect there to be a "science of combination therapies" that seeks to understand the complex interactions between training, plasticity, and regeneration [
59,
60]. The most effective physical therapeutic technologies of the future will likely be based on this science.
Competing interests
David Reinkensmeyer has a financial interest in Hocoma, A.G., a company that makes robotic therapy devices. The terms of this arrangement have been reviewed and approved by the University of California, Irvine, in accordance with its conflict of interest policies.