Towards adaptive and finer rehabilitation assessment: A learning framework for kinematic evaluation of upper limb rehabilitation on an Armeo Spring exoskeleton

https://doi.org/10.1016/j.conengprac.2021.104804Get rights and content

Highlights

  • An individualized modeling approach of the trajectories of patients using the orthotic exoskeleton.

  • Incremental model update to closely follow-up the recovery evolution.

  • Intra-patient assessment for quick diagnosis of performance drops or training’s negative effects.

  • Introduce six new metrics to assess rehabilitation exercises.

Abstract

Providing specialized rehabilitation and tailoring the training process for patient’s needs and according to recovery potentials has gained importance. To satisfy this need, a dynamic assessment of the performance of the recovery process is required. Assessing rehabilitation for the upper limb is often carried out with clinical subjective scales that do not satisfy these requirements. The use of technologies introduced several sensors into the devices used for rehabilitation and permitted the rise of kinematic assessments.

Kinematic measures provide an objective scale to follow up recovery during upper limb rehabilitation. The kinematics are still raw evaluations since they present insignificant effects if studied over short periods or on heterogeneous samples.

We propose a framework for modeling the trajectories as a means of encoding the specificity of the movement at every stage. The new technique permits detecting significant differences as soon as three training sessions became available.

We adopt an expectation–maximization algorithm and an optimization technique to encode the trajectories and the transition model from the acquired data. The framework enables us to encode in a Bayesian sense the observations from the patient and define six metrics to follow up on the progress of the movement quality. Statistical analysis of the results proved that these metrics are effective in tracking the evolution of the recovery. The results also established a strong discriminative property.

The proposed framework promises a finer scale of evaluation and extends the knowledge about kinematic assessment. This study’s findings suggest that adopting these new metrics can help achieve more individualized patient care. It additionally promises to limit the amount of data needed to detect a significant change.

Introduction

The rehabilitation exercises are increasingly served by the latest technologies in the form of serious games in virtual reality environments, robotic exoskeletons, or assistive training robots which are proving to be highly beneficial rehabilitation tools (Aminov et al., 2018, Kwakkel et al., 2008).

This presence has come to reply to the numerous guidelines and recommendations that were set for the design of rehabilitation exercises (Carr & Shepherd, 2003). These recommendations aim to maximize the efficiency of the treatment while compensating for the shortage of qualified personnel and therapists delivering these services by opening doors for automation. This shortage becomes noticeable given the growing demographics of the elderly and the motor-impaired patients victims of certain pathologies such as stroke.

European Robot Road-map of 2010 also emphasized the urgency to provide robotic solutions for healthcare assistance that are intelligently capable of automating a large portion of the caregiving procedure (euRobotics, 2013).

One of the recurrent challenges for researchers tackling these requirements is the ability to individualize and adapt the exercise to the evolution of the patient’s recovery and his current performance. The performance is often affected by different factors and does not always reflect the actual fitness level of the patient.

Although these guidelines are recommendations, the use of ubiquitous technologies for rehabilitation training stresses the need to automate the process of assessment, diagnosis, and monitoring. This consists of developing a more inclusive patient–robot–therapist loop, where continuous feedback is present between all elements. Achieving this would enable the full exploitation of these tools while liberating the therapist to be able to provide care of more patients at once.

While surveying the literature, replying to these challenges has often been approached by modeling the patient’s behavior during the rehabilitation training. This has been approached in different ways depending on the tools and the envisioned application context.

Many studies presented an intelligent agent to handle game difficulty for rehabilitation exercises based on virtual reality (Alamri et al., 2010, Borghese et al., 2018, Brezany et al., 2018, Tsiakas et al., 2015). Both Taub, Lum, Hardin, Mark, and Uswatte (2005) and Kan, Huq, Hoey, Goetschalckx, and Mihailidis (2010) are examples of approaches involving an assistive robot instead of a virtual environment. Their theoretical frameworks range between Bayesian framework, fuzzy inference, and Markov Decision Processes with Reinforcement learning. To retrieve the parameters of the exercise they often rely on multi-modal sensing using RGB-D cameras. Otherwise recovering sensors’ readings from the assistive robot were adopted for robot-based systems. Applications of telerehabilitation focus on building such assessment and monitoring models for online feedback for therapists and users as seen in Pirovano, Mainetti, Baud-Bovy, Lanzi, and Borghese (2012) and Borghese et al. (2018).

The author in Kleine Deters (2018) employed a Hidden Markov Model (HMM) to follow and decide dynamically on the rehabilitation exercise parameters concerning the actual performance of the patient.

In Capecci, et al. (2018), HMMs were used to follow on the execution of a rehabilitation exercise captured by an RGB-D camera and to provide feedback on the correctness of the execution.

A shared limitation to these approaches is that, while accounting for the environment’s specific actions set i.e. game controls, they fail to generalize easily across rehabilitation systems and remain thus system-dependent.

Learning the kinematic model of a task appears to be a promising endeavor since trajectories encode for both performance and strategies, while also being easily portable between systems. To this end, we can find the process of learning the kinematic models of the patient during a specific task. These methods are often conceived for path planning for active exoskeleton control. We refer to Zhou et al. (2016) for a detailed overview of these methods and techniques for rehabilitation applications.

Particularly, of interest in our context is the emphasis the authors put on the fact that the potential final model of the patient’s performance remains a hidden outcome. They point out that the proposed modeling techniques need to account for motion limitations due to fatigue and stiffness among other factors influencing the performance of the patient. The fact remains that an active research endeavor is modeling the recovery process whereas the few approaches proposed within the rehabilitation context still lag in terms of clinical validity and general utility.

In Reinbolt et al. (2005), a musculoskeletal model of the patient is built by approaching the human body with a multi-joint kinematic model and estimating afterwards the parameters from the data. An extended review paper in Leardini et al. (2017) provides more details on studies implementing similar models. These kinematic models define the joint structure of the body during a rehabilitation exercise from a morphological point of view to assess the movement conducted during training or estimate the interventions’ effects.

Another work proposed in Mendoza-Crespo et al. (2019) presents a learning technique for the kinematic model in the context of gait trajectories identifying centroids in the trajectories and interpolating them to model a step from multiple healthy users’ demonstrations. Learning the Kinematic model as a discriminant representation of the user without considering the variability in his performance would consist an undermining factor to the generalizability of similar approaches.

Meanwhile, in Lauretti et al. (2018) the authors used the Dynamic Movement Primitives (DMPs) to encode for the trajectories accomplished during the Activities of Daily Living (ADLs). The challenge that the DMP framework presents is the assumption of availability and knowledge of the model of the system which is indeed the case with robotic arms. The DMP-generated trajectories are also smooth splines that would fail to capture the imperfections that persist in the patient’s behavior.

A Locally weighted regression has been used by Atkeson (1991) to learn robot arm control by approximating the local model at each position. This approach while time-efficient for online learning does not provide enough generalization in the output model, it also may present a challenge for continuous monitoring applications since it is a memory-based approach and may present overflow challenges.

In Wisneski and Johnson (2007) the study of the subject robot interactions by building a model of daily life activities using the standard zero jerk method was emphasized as being insufficient to capture all the settings that the subjects demonstrated. The authors also noted some significant differences in task execution and strategies which would be more apparent in pathological subjects. A serious note was put on the necessity to establish a more sensitive model to capture the curvatures that were demonstrated during the reach movements and which were reproduced for all the panel members.

To tackle these limitations, evolutive or iterative approaches have been proposed. For instance, in Chebotar et al. (2017) the authors presented an evaluation of a hybrid of both model-based and model-free reinforcement learning approaches and the respective algorithms used for training in the context of trajectory learning. A persistent challenge with these propositions is that they still assume fully known Markov Decision Processes besides to some knowledge of the system model.

For trajectory tracking, Iterative Learning Control (Ahn, Chen, & Moore, 2007) for motor learning has been proposed as a theoretical framework to approach the evolution of the learning. These methods rely on the iterative approximation of the injected controls to ensure the convergence towards the final model. However, this can only match perfectly repetitive tasks such as in robotic manipulation. Besides, the assumption of knowledge of the patient model trajectory is unavailable in the context of rehabilitation training. The methodology assumes invariance in the dynamics governing the system evolution, whereas in rehabilitation, the dynamics are a hidden model to estimate. Moreover, the dynamics have the potential to evolve frequently considering the recovery process taking place.

In Jain, Wojcik, Joachims, and Saxena (2013) the authors presented an approach to learning an optimal task demonstrated by a user. The resulting model aims to control robot interactions through the reproduction of the movement.

In an assessment oriented application, rule-based techniques rely on a predefined set of rules to score and give feedback to the patient and therapists. As an example, a kinematic rule-based modeling technique was proposed by Zhao, Lun, Espy, and Reinthal (2014) which defines an encoding of the exercise execution rules. Exercise executions are then compared against the established base truth to give feedback and assessment. Fuzzy inference is used to decide on quality assessment. This approach lacks a dynamic update of the rule definition. The major benefit of the method is the ability to incorporate useful advice into the feedback given to the patient by specifying the execution error committed during the exercise.

In Capecci, et al. (2018) the authors used an HMM to detect a set of features of patient’s movements during the exercise using an RGB-D camera. They then provide an assessment based on a set of predefined rules that were concerted with a therapist. The definition of these rules is a non-trivial task. It involves a considerable amount of work to handwrite these rules and use them later for assessment. The defined rules are also specific to the task and are not generalizable.

In opting for a methodology to learn the patient model we were motivated by addressing some challenges:

  • Continuous assessment and follow-up of the exercises;

  • A tool to provide personalized feedback;

  • A lightweight implementation for online deployment;

  • The ability to generalize to similar rehabilitation systems.

We suppose that achieving these capabilities should permit a holistic parameterization of the patient’s recovery, thus, providing current performance, rehabilitation rate, and potential recovery projections.

We are adopting a statistical trajectory modeling approach to establish a means of assessing the performance of the patient. We seek to find the underlying model of the patient’s hand movement during the exercise on an orthotic exoskeleton by capturing the motion of the end-effector using an HMM and imitation learning technique. Further analysis of the technique of imitation learning in the context of robotic trajectory modeling is referenced in Osa et al. (2018).

Our main contribution is to provide an instrument to assess the longitudinal evolution of patients during the instrumented rehabilitation training sessions. Hence, a framework would be developed to model the trajectories recorded during patient movements to permit portability of this methodology to other exoskeleton devices, motion capture, or telerehabilitation systems. The framework relies on a data-estimated dynamic model. To measure the evolution of the rehabilitation new metrics will be proposed and tested statistically. We hypothesize that the resulting model will detect significant changes early on in the course of rehabilitation. We also hypothesize the significance of the findings will be pathology agonistic. Finally, we aim to evaluate the framework on an operational dataset to ensure these captured changes are detectable in the deployment environment.

Our starting point is the work that has been conducted by Coates, Abbeel, and Ng (2008) which proposes an HMM trajectory representation and an algorithm to infer the ideal trajectory of a given task. Our reasoning to opt for this methodology could be articulated in the following manner:

  • Firstly, the HMMs representation permits us to capture the fine details of the trajectories considering that we use an upsampled chain to model the optimal execution. Meanwhile, in robotics, the trajectory modeling techniques often outcome a smooth trajectory. The fact of smoothing the resulting trajectories inadvertently contradicts our attempt to capture non-optimal and errored executions demonstrated by the patient;

  • Secondly, the statistical representation would likely fit the stochasticity of the human Patient Controller and the Orthotic Exoskeleton system that we are studying, hence referred to as PCOE;

  • Thirdly, the iterative procedure used permits constant update of the outcome model;

  • Fourthly, the assumption of noisy demonstrations and the smoothing procedure would permit us to alleviate the rigor constraint on the data quality. This is important as quality is often imposed by the equipment; for instance, the sensors used on the Armeo Spring exoskeleton;

  • Lastly, the portability of this methodology, learning state trajectories would permit the extrapolation towards a control task for any active-assist intervention using motorized exoskeletons as well as being applicable in the virtual environment systems.

To be able to capture the patient model, we first start by encoding the system dynamics. To this end, we implement a learning algorithm as in Abbeel, Ganapathi, and Ng (2006) which permits us to approximate the state transition model. We present a newer iteration and cost function that can perform as well with better consistency. We show results presenting the ability of the model to keep satisfactory error rates while predicting state changes with both the original and the proposed implementations.

To model the patient ideal trajectory we implemented a variation of Coates et al. (2008) where we used the Expectation–Maximization (EM) algorithm to infer the underlying model. The resulting HMM model serves as the basis for defining a set of six new metrics that are studied statistically to investigate their properties.

In Section 2 we start by introducing the design of the study and the population involved. The equipment used as well as the task studied is listed afterward. We then present the framework and detail its components. We end with detailing the definition of the proposed metrics, the statistical analysis used, and the data preprocessing methodology. In Section 3 we listed results by first detailing the learning of the coefficient of the dynamic model. Secondly, we present the trajectory modeling results. Finally, the results of the statistical analysis are detailed for each of the tests that we carried out. In Section 4, we discuss the results in light of potential use cases and utility in the domain of rehabilitation. We then conclude with the principal takeaways from this study and potential future research.

Section snippets

Study design and population

We conducted a longitudinal observational study on rehabilitation exercise data. The study was conducted on records of exoskeleton upper limb rehabilitation exercises retrieved from the Physical Rehabilitation Center at the University Hospital Center CHU of Tlemcen, Algeria. Both patients followed the standard therapy routines in parallel to the complementary exoskeleton sessions. The study aims to parameterize the recovery process through repeated measures in time. Outcome measures are defined

Results

In the following, we detail the results of the state transition model learning phase for both implementations comparatively and against the original work. Afterward, we detail the patient task model inference results. The statistical analyses are then presented while emphasizing the notable findings.

The utility

The rehabilitation exercises can be more effectively and accurately assessed using our proposed framework and the defined metrics. The new resulting metrics present strong evolutive, discriminative, and predictive properties that are key to assessment adoption for clinical validation and practice.

We modeled the trajectories of the end effector during rehabilitation exercises using an Armeo Spring exoskeleton using an EM procedure. The resulting model permitted the definition of several metrics.

Conclusions

In conclusion, to address the accurate assessment of rehabilitation training from kinematic data we have developed a framework and listed a set of new evaluation metrics. The learning technique based on the Expectation–Maximization algorithm provides a means of modeling the trajectories of the end effector during the exercise. The resulting models are then used to derive a set of metrics that are used to study the recovery progress.

The results in this paper showed that the metrics derived by

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

The authors would like to thank Mrs. Meryem Karaouzene (PT) of the Rehabilitation Center, University Hospital Center of Tlemcen, Algeria for helping in providing access to the data record, patients’ history, and subsequent explanations. This work has been supported by the PHC Tassili, France program agreement 19MDU210 and the DGRSDT , Algeria.

References (36)

  • Borghese, N. A., Pezzera, M., Mainetti, R., Essenziale, J., Cazzaniga, R., & Reggiori, B., et al. (2018). A cloud-based...
  • Brezany, P., Janatova, M., Stepánková, O., Uller, M., & Lenart, M. (2018). Towards precision brain disorder...
  • CarrJ. et al.

    Stroke Rehabilitation: Guidelines for Exercise and Training to Optimize Motor Skill

    (2003)
  • Chebotar, Y., Hausman, K., Zhang, M., Sukhatme, G. S., Schaal, S., & Levine, S. (2017). Combining model-based and...
  • ChingW.K. et al.

    Markov chains. Models, algorithms and applications

    (2013)
  • CoatesA. et al.

    Learning for control from multiple demonstrations

    (2008)
  • Strategic research a genda for robotics in EuropeTechnical report

    (2013)
  • FieslerE. et al.

    Handbook of neural computation

    (1996)
  • Cited by (3)

    View full text