Learning robotic motion with mirror therapy framework for hemiparesis rehabilitation

 Nothing here even tells me how well patients recovered using this. And the mentors and senior researchers approved this crapola?

Learning robotic motion with mirror therapy framework for hemiparesis rehabilitation

Author links open overlay panel

https://doi.org/10.1016/j.ipm.2022.103244Get rights and content

Abstract

The number of hemiplegic patients has rapidly increased in recent years, and intelligent robots have high rehabilitation potential. Robotic mirror therapy (RMT) is a promising therapeutic measure for hemiparesis by voluntarily transferring the motion of the healthy limb (HL) to the impaired limb (IL), in which a robot interacts with and assists the IL to mimic the action of the HL to stimulate the active participation of the injured muscles. Nonetheless, complete replication of the HL movement trajectory to the robot without considering the IL muscle strength cannot ensure safety or facilitate rehabilitation. In this study, a learning-based robotic motion generation scheme was developed for RMT. The robot movement trajectories were modeled with dynamic movement primitives (DMPs), and the physical human-robot interaction was formulated as an impedance model coupled with the DMP model. To adapt the robotic motion to different pilots, reinforcement learning was used to optimize the coupled DMP model parameters. The reinforcement learning approach was implemented based on policy improvement and the path integral (PI²) algorithm, and the cost function was designed to simultaneously ensure training safety and enhance muscle strength. The proposed method was validated using a lower-extremity rehabilitation robot with magnetorheological actuators, and the experimental results demonstrated the feasibility and superiority.

Introduction

Hemiplegia has become increasingly prevalent in recent years, and the growing number of stroke patients proves the major cause (Jiang et al., 2022). A hemiplegic patient has affected limbs on one side of the body, whereas the other side remains healthy. Robot-assisted training is an effective approach for hemiparesis rehabilitation, where the patient's impaired limb is driven by a robot to complete the movement predetermined by physiotherapists or operators. However, such passive training is not suitable for patients with various disabilities, especially when they feel uncomfortable or even painful. To address this problem, robotic mirror therapy (RMT) has been proposed for hemiparesis treatment (Dohle et al., 2009). During therapy, the motion of the patient's healthy limb (HL) is mirrored to the impaired limb (IL) through a robotic medium, and accordingly, the IL can mimic the action of the HL with robotic assistance. In this regard, the HL can voluntarily plan the training trajectory for the IL. Through mirror-symmetric motions of the patient's bilateral limbs, the unimpaired hemisphere of the brain interacts with the impaired hemisphere, thus inducing reorganization of the motor cortex networks and promoting cortical neuroplasticity (Cauraugh & Summers, 2005). Mirror-image motion enabler (MIME) was developed to move the normal limb to control the position and orientation of the opposite paretic forearm through a master-slave robotic system (Lum et al., 2005). Clinical tests have proven that the MIME training scheme can achieve better rehabilitation effects than conventional unilateral modes. Similarly, a teleoperation architecture was built to transfer therapeutic training from the therapist to the patient's IL, and movement smoothness and symmetry were both ensured (Shahbazi et al., 2016). However, physiotherapists are still involved in this framework, thus, savings in human labor costs and in-home independent exercise cannot be achieved. In addition, learning-based control can be adopted to drive the affected leg to track the unaffected leg's motion naturally with a leader-follower robotic system, where one leader is represented by the HL and two followers refer to two lower-limb exoskeletons on both sides (Peng et al., 2020). The motion synchronization between the leader and follower robots can be achieved well by learning the kinematics of the HL; however, the data-driven strategy is time-consuming without pre-modeling of the movement trajectories. More importantly, in the abovementioned schemes, the HL's motion is completely replicated and transferred to the robot on the IL side, and the IL's movement capability is not considered. In fact, the movement capabilities of the hemiplegic patient's HL and IL vary substantially, whereas the current RMTs assume that the robot and HL share the same trajectory. For example, the HL can expand the IL's motion range to encourage the IL's exercise; however, this may lead to overlarge motion of the robot causing secondary damage on the IL. Conversely, the HL motion range is generally reduced to relieve IL discomfort in traditional RMTs. However, the IL residual motor function cannot be stimulated because of the restricted task trajectory. In summary, the complete mirror of the motion of the HL on the robot may increase the possibility of damage or decrease active participation. Thus, there is a tradeoff between bilateral motion synchronization and improvement in rehabilitation efficacy. In practical, the IL and robot work cooperatively to generate a movement trajectory that tracks the kinematics of the HL. Therefore, accurate modeling of the robot's movement trajectory that facilitates IL rehabilitation and guarantees safety plays a significant role in RMTs.

In the field of rehabilitation robotics, many representative tools are employed to model the movement trajectory of a robot. In (Sharifi et al., 2021; Sharifi et al., 2022), a central pattern generator (CPG) (Sproewitz et al., 2008) was adopted to plan the gait trajectory for lower-limb exoskeletons, and the gait amplitude and frequency were modified to enhance users’ safety and comfort. A human-robot interaction energy was formulated to modulate the CPG dynamics; however, too many uncertain parameters are involved and the generated trajectory cannot be optimized with high efficiency and accuracy. An adaptive frequency oscillator (AFO) (Gams et al., 2009) can be utilized to extract features in active arm movements and generate a real-time and rhythmic trajectory for upper limb rehabilitation robots; however, its adaptability to subjects’ sudden and unpredictable motion is insufficient (Wang et al., 2022). In addition, dynamic movement primitives (DMPs) (Ijspeert et al., 2002; Ijspeert et al., 2013) can achieve excellent tracking performance through imitation learning by demonstration. The DMP was used to plan the motion of a robotic arm to assist in activities of daily living by extracting from a dataset of human motion trajectories (Lauretti et al., 2017). A learning control scheme based on DMPs was also designed to reproduce the exercise trajectory demonstrated by a therapist for robot-aided ankle rehabilitation and Valera (Abu-Dakka and Valera, 2020). The imitation characteristics of the DMP dynamics exactly meet the requirement that the robot imitates HL motion in the RMT. In particular, for a rehabilitation robot that tightly couples with a human body, the robot's trajectory is not only influenced by the HL but also the IL, and the DMP can be extended to address the physical human-robot interaction well. A coupled cooperative primitive was derived based on the DMP to represent the interactive term through impedance models, ensuring that the exoskeleton can perfectly track the pilot leg for walking assistance (Huang et al., 2019). A coupling term was also added to the DMP to generate a safe walking area and avoid collisions between the crutch and patient's lower limb (Tan et al., 2022). An adaptive DMP controller was developed to facilitate lifting movements assisted by an exoskeleton, where the robotic movement trajectories can be accurately predicted and smoothly synchronized, thereby reducing the activation of the user's spinal erector muscles (Lanottte et al., 2021).

Although the aforementioned DMP-based methods have been proposed to assist human motion (walking or lifting) in relaxing muscles, rehabilitation efficacy has not been intentionally improved, where muscle activation should be increased instead. Mirroring the HL motion to the robot is apparently not sufficient, particularly in RMT, even though the human's influence is integrated into trajectory modeling, and more importantly, the generated trajectory should have the capability of enhancing the patients’ muscle strength. Skin surface electromyogram (EMG) signals are common and powerful representations of the users’ muscle situations, and many rehabilitation robotics studies have adopted EMG signals to evaluate rehabilitation effectiveness and design control strategies (Brambilla et al., 2021). Moreover, EMG signals can be combined with DMPs to realize more functions. In (Peternel et al., 2016), an adaptive control system was implemented in exoskeleton robots to adapt assistive joint torque profiles to alter task execution. Herein, DMPs and adaptive oscillators were adopted to regulate the phase and frequency of motion to reduce muscle activity; however, they could not achieve the rehabilitation target. The concepts of DMPs and stiffness primitives were introduced to encode a kinesthetic demonstration as a combination of trajectories and stiffness profiles, where the EMG signals were captured from human's upper limbs to obtain the target stiffness profiles (Bian et al., 2020). An incremental skill learning and generalization method was proposed based on DMP and EMG feedback, such that the manipulator can learn from ordinary human actions and reactions to handle sudden incidents comprehensively (Lu et al., 2022). The integration of DMP and EMG has the potential to be applied in RMT; however, current studies are limited to proving its feasibility.

In this study, a DMP-based control framework was proposed in RMT for hemiparesis rehabilitation. The main contributions of this study are as follows:

(1)

An RMT control framework was constructed for voluntary hemiparesis rehabilitation, and the robot's movement trajectory was modeled with a coupled DMP, where the human-robot interaction was formulated based on the muscle strength of the IL.

(2)

To simultaneously ensure bilateral motion synchronization and improvement in rehabilitation efficacy in RMT, a reinforcement learning algorithm based on policy improvement with path integrals (PI²) was proposed to learn and optimize the coupled DMP parameters for different patients.

(3)

The proposed RMT framework was validated on a lower-extremity rehabilitation robot with magnetorheological actuators, and its feasibility and superiority were demonstrated in terms of trajectory tracking, learning efficiency, and rehabilitation efficacy compared with state-of-the-art methods.

The remainder of this paper is organized as follows. Section 2 describes the method for implementing the RMT control framework. The experiments are presented in Section 3. Finally, Section 4 concludes the study.

Section snippets

Robotic mirror therapy framework

An RMT control framework was constructed for hemiparesis rehabilitation. Taking advantage of the motion characteristics of hemiplegic patients, the movement trajectory of the HL was utilized to motivate the IL with the support and assistance of a wearable robot, implementing voluntary rehabilitation training. During therapy, the IL tried to activate the residual muscle strength to track the HL motion; thus, it was not practical for the robot to completely replicate the movement of the HL

Protocol

In this section, the proposed RMT framework is validated using our previously developed lower-extremity rehabilitation robot (Xu et al., 2019). A prototype of the robot is shown in Fig. 2. Magnetorheological actuators were installed at robotic joints for compliant motion, and their torques were varied continuously and flexibly by adjusting the input electric current. Owing to these actuators, the robot could regulate its stiffness and provide assist-as-needed training based on the user's motion

Conclusion

This paper presents a motion generation scheme for RMT, in which the HL of a hemiplegic guides the movement of the IL assisted by a robot. The DMP, along with the impedance-based human-robot coupled term, was utilized to model the robot movement trajectory, and the impedance parameters correlated to the human muscle strength, which was evaluated using skin surface EMG signals. The PI² reinforcement learning approach was developed to adapt the coupled DMP model parameters to different subjects.