Treatment planning and monitoring for motor neurorehabilitation still relies mostly on coarse clinical scales and clinician observation. Distribution analysis of action patterns recorded by sensors can provide a quantitative and easily visualizable representation of complex individualized dynamics. Here, we review the use of distributions to analyze human action in health and following neurologic injury. The aim is to demonstrate the potential of distribution analysis as a simple yet effective framework for scanning a patient prior to therapy, assisting clinicians in quantitative evaluations, and thereby uncovering underlying pathophysiological mechanisms to optimize therapy. In this review paper, we first describe how idealized distributions can model movement and capture spatiotemporal patterns. We then review metrics that can holistically quantify changes between distributions. Next, we highlight applications in different data domains that have used these approaches and precise comparison techniques to identify motor deficits and personalize therapy. Together, these methods have uncovered upper extremity flexion synergy patterns in half of stroke survivors, improved classification accuracy in data-driven models, and reduced movement errors threefold. Finally, we discuss the challenges and opportunities facing the application of distribution analysis for neurorehabilitation and precision medicine.

Nonobtrusive options for monitoring the wrist and hand movement are needed for stroke rehabilitation and other applications. This paper describes the "manumeter," a device that logs total angular distance travelled by wrist and finger joints using a magnetic ring worn on the index finger and two triaxial magnetometers mounted in a watch-like unit. We describe an approach to estimate the wrist and finger joint angles using a radial basis function network that maps differential magnetometer readings to joint angles. We tested this approach by comparing manumeter estimates of total angular excursion with those from a passive goniometric exoskeleton worn simultaneously as seven participants completed a set of 12 manual tasks at low-, medium-, and high-intensity conditions on a first testing day, 1-2 days later, and 6-8 days later, using only the original calibration from the first testing day. Manumeter estimates scaled proportionally to the intensity of hand activity. Estimates of angular excursion made with the manumeter were 92.5% ± 28.4 (SD), 98.3% ± 23.3, and 94.7% ± 19.3 of the goniometric exoskeleton across the three testing days, respectively. Magnetic sensing of wrist and finger movement is nonobtrusive and can quantify the amount of use of the hand across days.

This paper describes a robotic-arm exoskeleton that uses a parallel mechanism inspired by the human forearm to allow naturalistic shoulder movements. The mechanism can produce large forces through a substantial portion of the range of motion (RoM) of the human arm while remaining lightweight. This paper describes the optimization of the exoskeleton's torque capabilities by the modification of the key geometric design parameters.

This paper provides a detailed design description and technical testing results of a device for studying motor learning and rehabilitation of human locomotion. The device makes use of linear motors and a parallel mechanism to achieve a wide dynamic force bandwidth as it interacts with the leg during treadmill stepping.

The directional control of reaching after stroke was simulated by including cell death and firing-rate noise in a population vector model of movement control. In this model, cortical activity was assumed to cause the hand to move in the direction of a population vector, defined by a summation of responses from neurons with cosine directional tuning. Two types of directional error were analyzed: the between-target variability, defined as the standard deviation of the directional error across a wide range of target directions, and the within-target variability, defined as the standard deviation of the directional error for many reaches to a single target. Both between- and within-target variability increased with increasing cell death. The increase in between-target variability arose because cell death caused a nonuniform distribution of preferred directions. The increase in within-target variability arose because the magnitude of the population vector decreased more quickly than its standard deviation for increasing cell death, provided appropriate levels of firing-rate noise were present. Comparisons to reaching data from 29 stroke subjects revealed similar increases in between- and within-target variability as clinical impairment severity increased. Relationships between simulated cell death and impairment severity were derived using the between- and within-target variability results. For both relationships, impairment severity increased similarly with decreasing percentage of surviving cells, consistent with results from previous imaging studies. These results demonstrate that a population vector model of movement control that incorporates cosine tuning, linear summation of unitary responses, firing-rate noise, and random cell death can account for some features of impaired arm movement after stroke.

We examine a method to control the stepping motion of a paralyzed person suspended on a treadmill using a robot attached to the torso and hips. A leg swing motion is created by moving the hips without contact with the legs. The problem is formulated as an optimal control problem for an underactuated articulated chain. The optimal control problem is converted into a discrete parameter optimization and an efficient gradient-based algorithm is used to solve it. Motion capture data from a human subject is compared to the results from the dynamic motion optimization. Our results indicate that it is possible for the robot to create a gait for the paralyzed person that is close to that of an unimpaired subject.

We describe the design and testing of a robotic system to assist locomotion training of spinal-injured rodents. The goal of the system is to control and quantify spatial-temporal patterns of movement and forces during the stance and swing phases of rat locomotion. This approach will allow us to provide varying levels of assistance to limb movement during stepping and to quantify the effects of assistance on step training. Our initial finding was that the rat spinal cord could perform stepping on a virtual treadmill generated by the robot system, with careful design of the virtual environment. Based on the conditions required for this "virtual stepping", we suggest several design principles for robot-assisted rehabilitative step trainers, including devices for humans.

The use of an associative content-addressable memory to model a robot and the world the robot interacts with is discussed. The model can be learned by storing experiences in the memory. To make predictions, the memory is searched for relevant experience. An initial implementation of such a memory-based modeling scheme has been made on a parallel computer, the Connection Machine. The implementation machine was used to model and control a simulated planar two-joint arm and a simulated running machine. The issues and problems that arose in the preliminary work are described. It is found that the use of parallel search in the implementation of an associative content-addressable memory allows quick searching of stored experiences, and reasonable retrieval is obtained using a simple distance metric and a simple generalized scheme. The memory is able to generalize after storing only a small number of relevant experiences. The use of search by parallel processors also avoids many of the problems of previous memory-based or tubular approaches to robot modeling (such as search speed and memory requirements).>

The functionality of robots can be improved by programming them to learn tasks from practice. Task-level learning can compensate for the structural modeling errors of the robot's lower-level control systems and can speed up the learning process by reducing the degrees of freedom of the models to be learned. The authors demonstrate two general learning procedures-fixed-model learning and refined-model learning-on a ball-throwing robot system. Both learning approaches refine the task command based on the performance error of the system, while they ignore the intermediate variables separation the lower-level systems. The authors also provide experimental and theoretical evidence that task-level learning can improve the functionality of robots.>