Keywords

1 Introduction

Recent projects highlight how motor learning and a high level of attention control can potentially improve submaximal force production during recovery after stroke. Motivation and attention are key elements for a potentially successful motor rehabilitation [1]. It is capital for the success of the therapy that the patient understands the exercise and the potential outcome, to keep the engagement and the attention, thus enhancing motivation [2]. Furthermore, if the patient lacks the capacity to understand how an exercise is executed correctly, little effect can be expected [3]. It is important to note that while talking about stroke patients that their brain has suffered an injury, learning function may be compromised. Nonetheless, it has been seen that after local brain cortex tissue damage, rehabilitative training can produce a reorganization to some extent on the adjacent intact cortex [4]. These findings suggest that undamaged motor cortex might play an important role in motor recovery. One important factor is that there seems to exist a direct relationship between impairment and function: function of walking is correlated to lower-limb strength (impairment) [5].

Other key factor in a successful therapy towards motor recovery is the concept of repetition. On the one hand, the idea of “repetition without repeating” is important for the avoidance of disengagement during the progress of the task. On the other hand, this need of repetition may lead to health problems for physiotherapists, and robotics can be of huge help in providing them with a tool to enhance the therapy by allowing them to ensure repeatability and providing a tool to more objectively assess performance metrics [6].

The concept of “repetition without repeating” considers that training sessions must include recording the modified parameters to challenge the patient [2]. Within this concept, the idea of interleaving different exercises also arises, and it has been seen that this mixing of learning tasks is more effective for learning than learning tasks one by one with specific focused trainings [7], and this mixed schedule might aid recovery because it considers each movement as a problem to solve and not as a pre-memorized muscular sequence [8]. In fact, Shea and Kohl [9] stated that “retention of the test task was best facilitated by acquisition practice that included task variations rather than conditions that included additional practice on the test task or even a condition identical to the retention condition.” The nervous system seems to rely on the concept of adaptation, meaning that a previously learned skill can be recovered after a change in the operating environment. This theory of adaptation says that after a change in the environment dynamics, the control system does not need to relearn but adapt, suggesting that the nervous system tunes a learned internal model and has the ability to extrapolate this learning to new scenarios [10].

Robotics has been proposed and partially proven to be beneficial in interventions in acute and chronic stroke patients [11,12,13]. It is important to study the better tradeoff between the learning rate and the amount of helping guidance given by the robot during training [14]. Robotic guidance approaches also offer a powerful tool in rehabilitative and learning cases where large errors may be dangerous or undesirable, as this guidance provides the user with a tunnel inside of which the movement is allowed and controlled outside of it.

Robotics can be used not only as standalone devices, but also in combination with virtual environments, providing the possibility of showing the patient and the physiotherapist the feedback of the performance of the task, which has been proven to improve rehabilitation [3]. In the literature we can find that visual feedback through virtual environments helps to improve the performance in specific tasks, even improving the results in comparison to real-task training [15]. Indeed, this visual approach has been proven to improve robotic haptic guidance therapy scenarios [16]. Furthermore, the idea of this virtual environment or “video game” makes the implicit learning process transparent for the user, not being aware of what it is being learned [6]. They also ease to offer the user with a goal-oriented task and possibility of repetition [17].

Robots are crucial for these haptic approaches, as they help the physiotherapist to provide an objectively equal movement to the patient’s limb. Several authors have tested the error augmentation approach [6, 18,19,20] to enhance motor learning. The nervous system learns forming the internal model of the dynamics of the environment via a process of error reduction. Emken and Reikensmeyer concluded that motor learning process can be accelerated by exploiting the error-based learning mechanism. They used the approach of perturbing the movement via a robotic device, to augment the error while the user was performing the task. It seems that this learning process is indeed a minimization problem, based on the last performed trials to accomplish the task [20], where the minimization takes place between the weighted sum of the kinematic error and the muscular effort [21]. Combined haptic error augmentation and visual feedback have provided better results than conventional direct training (pure joint mobilization), exploiting the idea of using the after-effects of a resistive force to help perform the user the desired trajectory task [22]. Moreover, this error-augmentation strategy produces, when the after-effects are evaluated, the correct muscle activation patterns to achieve the desired trajectory, and results showed that subjects were able to reduce errors more rapidly than those that received a robotic guidance approach training [14]. Reikensmeyer and Patton suggest combining robotic guidance and error-augmentation techniques, starting with guidance and gradually removing it and increasing error-augmentation.

Several studies have been conducted with stroke patients, showing promising results comparable to those learning results seen in healthy subjects. For example, Patton et al. [23] found some preliminary evidence that showed that the approach of error-augmentation could be used to obtain smoother trajectories in stroke patients. Krakauer [8] found that, although guidance approach leads to better performance during the training, it is not optimal for retaining the learning over time, so these training may improve performance in the clinical environment, but may not be transferable to activities of daily living. Reikensmeyer and Patoon demonstrated that with this error-augmentation approach, stroke patients can potentially alter abnormal limb movement patterns that appear relatively fixed, although the reasons why people with stroke perform abnormal patterns remain to be identified [14].

One important point to take into account in rehabilitative processes is that “several types of exercise programs, including muscle strengthening, physical conditioning, and task-oriented exercises have led to an improvement in balance and mobility” [24]. Thus, it makes sense to focus the approach of early rehabilitation and motor recovery on simple tasks, towards the integration of them into more complex processes while the recovery takes place over time. Patton et al. [19] also found evidence on stroke survivors retaining ability to adapt to force fields, being better at adapting to real scenarios less demanding than training scenarios (lower forces in the real world than in the training); and more importantly, they can last longer if the exercise resembles normal movements and the after-effects can be perceived by the patient as an improvement [6].

Most studies are focused on upper-limb rehabilitation, but Duncan et al. [25] demonstrated that similar motor recovery of upper and lower-limbs occur after stroke. Thus, we focus our study, given this state-of-the-art analysis, on lower-limb recovery, specifically on the ankle. It has been reported that both plantar and dorsiflexion are compromised after a stroke: plantarflexor passive stiffness is the cause for reduced plantarflexion torque before starting the swing phase in gait, and may also be the cause to limited dorsiflexion, compromising foot clearance [26]. It is also interesting to point out that these rehabilitative approaches can be either used for chronic or acute patients, leading to performance improvement and recovery beyond the current 6-month recovery plateau [27].

This study focuses on the assessment of detailed metrics of force production and position control -healthy subjects- and their correlation with submaximal force production control learning, as improving regularity in submaximal force production might help improve functionality and reduce disability [28] during a new task consisting in maintaining the position for early rehabilitation after stroke. Several assessment measurements are taken during the learning task to temporally determine the scale of this learning. Our aim is to characterize the capacity to perform the precision task of maintaining the position with a submaximal force production in healthy subjects.

2 Materials and Methods

In this section, we briefly present the platform we use for this study and expose the methodology. The analysis of the obtained data is performed with MathWorks ®MATLAB ®.

2.1 Experimental Platform

A Motorized Ankle Foot Orthosis (MAFO) is used for this study (see Fig. 1), with a zero-torque control. This robotic platform permits to exert controlled torque profiles to the ankle joint of the subject. It is connected to a BeagleBone Black board running Ubuntu and a custom application to interface the CAN-enabled driver board of the robot to MathWorks Simulink ®, where the visual paradigm is programmed, and to MATLAB, from where the control parameters for the robot are sent.

Fig. 1.
figure 1

(a) Subject inside the robotic platform performing the experiment; and (b) detail of the actuator of the robotic platform.

Full size image

2.2 Participants

Nine males and six females without any history of neuromuscular and cardiovascular disorders participated in this study; with ages 27.73 ± 3.58, and sport activity more than 2 and less than 6 h per week.

2.3 Task

The experiment consisted in following the trajectories depicted via the visual paradigm (Fig. 2), while the robot disturbed the movement by performing plantar and dorsiflex interleaved torque patterns (Fig. 3). The aim of the exercise was to improve the motor control by learning how to maintain the position to follow the trajectory in the screen, compensating the perturbations done by the robot.

Fig. 2.
figure 2

Visual paradigm. The user controls the position of the bird with the angular position of the ankle.

Full size image
Fig. 3.
figure 3

Torque patterns.

Full size image
Fig. 4.
figure 4

Experiment phases.

Full size image

Training Paradigms. Three training paradigms were designed for the experiment:

  • Fixed torque: This training paradigm performed the thirty trajectories at a fixed torque (established at 16 N\(\cdot \)m as a technical limitation of the set robot-controller).

  • Progressive: Progressive paradigm progressively increased the torque exerted by the robot from a minimum of 1 N\(\cdot \)m up to 16 N\(\cdot \)m.

  • Modulated: The peak torque exerted in each trajectory was increased in 2 N\(\cdot \)m if score in the previous one was 100%, and decreased to the average between the two last trajectories’ score (saturated in this way: decreased at least 1 N\(\cdot \)m and at most 2 N\(\cdot \)m in comparison to previous torque). For example, if score is lower that 100%, previous peak torque was 8 N\(\cdot \)m, and current torque was 10 N\(\cdot \)m, then next torque will be (8 + 10)/2 = 9 N\(\cdot \)m, due to the difference 10 − 9 = 1 N\(\cdot \)m being higher or equal to 1 and lower or equal to 2 N\(\cdot \)m.

Task Description. The task consisted on forty trajectories, distributed as in Fig. 4: (1) first trajectory let the user move the foot freely to understand the dynamics of the exercise; (2) the three trials from 2 to 4 assessed the performance of the subject in the task before the training (at fixed torque); (3) thirty training trials were performed (1 out of 3 different training paradigms was randomly selected for each user); (4) the three trials from 35 to 37 assessed the performance of the subject in the task after the training, at 80% of the fixed torque, e.g. 1.2 N\(\cdot \)m (in order to assess the performance at a different level than that used for training in the fixed torque training paradigm); and (5) finally, the last three trials from 38 to 40 assessed the performance of the subject in the task after the training.

Fig. 5.
figure 5

Scores for the three training paradigms (average and standard deviation), for the three assessments: (1) Before training at fixed torque, (2) after training at 80% of fixed torque, and (3) after training at fixed torque.

Full size image
Fig. 6.
figure 6

Root mean squared error for the three training paradigms (average and standard deviation), for the three assessments: (1) Before training at fixed torque, (2) after training at 80% of fixed torque, and (3) after training at fixed torque.

Full size image

3 Results

In Fig. 5 we depict the scores for the three training paradigms (average and standard deviation) together with the trend (calculated as the fitting of the two fixed torque assessment). We observe that the score increases after the training, and that the score for the assessment at the 80% of the fixed torque at the end of the training is better for all training paradigms.

Fig. 7.
figure 7

Torque for the three training paradigms, separated in dorsiflexion and plantarflexion. Average is represented with a solid black line, and standard deviation is represented by the grey shadow in the figure.

Full size image

Figure 6 shows the root mean squared error for the three training paradigms. Lower errors are obtained after the training.

In Fig. 7 we can see the mean interaction torque for the three training paradigms, separated in dorsi and plantarflex trials. For (1) the fixed torque training paradigm, all trials reach the 16 N\(\cdot \)m target; (2) in the progressive paradigm we can observe this progressively increase in the torque; and (3) in the modulated torque training paradigm, the torque increases up to a plateau, not reaching the established fixed torque in any case. We had to remove from the analysis the first training trial due to a technical problem (fixed torque of 16 N\(\cdot \)m was incorrectly applied in this first trial for all the training paradigms).

We observed that all training paradigms led to an improvement in the score comparing pre and post-training performances, so we concluded that this platform induces learning on healthy subjects. Consistently, the error of the subject when positioning the bird on the visual cue (e.g. maintaining the foot at 0\(^\circ \)), before and after training, decreased.

All the users for the cases of fixed torque and progressive torque paradigms obviously reached the prescribed torque, but this was not the case for the modulated torque paradigm, where none of the users was able to reach this 16 N\(\cdot \)m torque during the training, so we concluded that this torque exceeds our approach of using submaximal force production. Thus, we will study possible methodologies to calculate this customized submaximal torque for each subject.

All the subjects reported that they liked the game, and most of them said that it was too easy. This is also observable looking into the scores. We will test more difficult trajectories to be maintained, as well as different torque profiles, in order to increase the difficulty of the game.

To sum up, we conclude that this tool is useful to induce learning in healthy subjects, and thus will keep improving the training paradigms, for the translation into a rehabilitative tool.