Skip to main content

Archived Comments for: EMG-based pattern recognition approach in post stroke robot-aided rehabilitation: a feasibility study

Back to article

  1. Surface EMG Pattern Recognition after Stroke: Time to Move in Another Direction?

    Joel Stein, Department of Rehabilitation and Regenerative Medicine, Columbia University

    13 August 2013

    This paper by Cesqui et al. addresses an important issue in rehabilitation robotics: the utility of surface EMG (sEMG) signals to determine intended movement trajectory in hemiparetic stroke survivors. Ascertaining this intended movement from sEMG signals has the potential to improve the training and feedback provided by robots. Cesqui et al.'s findings have important implications for further research in this field.

    sEMG signals have long been used as a means of providing biofeedback after stroke. This technique has been controversial, however, and its utility remains disputed despite its intuitive appeal as a means of enhancing feedback for retraining. One potential explanation for the limited success of this approach is that the provision of auditory or simple visual feedback regarding muscle activity fails to fully engage the key feedback loops within the sensorimotor system. Integrating the feedback loop directly into the sensorimotor system could be accomplished through electrical stimulation, such as using EMG-triggered Functional Electrical Stimulation, or through robotic systems that use sEMG signals to modulate feedback.

    One approach to using sEMG signals in robotic therapy is to use a "power assist" strategy. The Myomo e100 and subsequent versions of this elbow flexion device use this approach by providing powered assistance in the direction of agonist muscle activity that is proportional to the degree of sEMG activity detected in the agonist muscle (1). The argument in favor of the therapeutic value of this approach is that some stroke survivors lack sufficient force generation to complete an intended movement (in this case, elbow flexion or extension). By providing powered assistance based on sEMG signals, this movement can be successfully completed, thus "closing the loop" between intention and the actual experience of completing this movement. While not proven, it is hypothesized that this cycle reinforces the motor control pathways and ultimately improves motor control.

    Cesqui and colleagues ask the appropriate next question: Can we broaden this approach to more complex movements and ascertain the stroke survivor's intended movement from the pattern of sEMG activation of relevant muscles? If this were successful, robotic devices could be introduced that would take this inferred trajectory and provide the assistance needed to complete the task. The therapeutic value would be similar to that for simpler devices but now applied to more complex movements.

    Cesqui et al. found, however, that the classifier system they developed, while functioning accurately among healthy volunteers, had poor accuracy for stroke survivors. There are several potential explanations for their findings, but the most fundamental may be that the loss of motor control experienced by stroke survivors doesn't create a reliable and distinct pattern of muscle activation for different intended movements. This is consistent with clinical observation that upper limb movement patterns in more severely paretic stroke patients are commonly stereotyped into "synergy patterns" of combined adduction and internal rotation at the shoulder, elbow flexion, forearm pronation, wrist flexion, and finger flexion.

    There are several strengths of this study, and a few limitations that deserve mention. The use of an individualized classifier for stroke subjects is an important strength of this study, addressing the hypothesis that each stroke patient might have a unique (and classifiable) pattern of sEMG activation associated with intended movements, but that these would vary substantially person-to-person. While the individual classifiers were more accurate than a group classifier, the level of accuracy remained poor. Another important facet of the study is that the stroke subjects represented a degree of motor impairment (based on upper limb Fugl-Meyer scores) that corresponds with the types of stroke patients typically targeted for robotic therapy.

    One of the limitations of this study include the focus on the shoulder and elbow region of the upper limb. While a very appropriate target for study of this question, this leaves open the question of whether similar findings would be present in other regions. The lower limb in particular has a much more rhythmic motor pattern in general than the upper limb, and it is possible that sEMG patterns are more predictable in the lower limb as a result.

    Examining patients with less severe deficits might prove interesting, and perhaps clinically useful. While stroke survivors with more mild deficits are arguably less in need of robotic therapies (and more capable of performing conventional retraining exercises without physical assistance), it remains possible that future robotic systems might be developed that are superior to conventional therapy for the upper limb. I speculate that a continuum of classifier accuracy might be found that correlates with the severity of motor weakness - a hypothesis that may be worth exploring further.

    This study focused on patients early post-stroke (although exact data on duration post-stroke is not provided), and it is possible that stroke survivors motor patterns "mature" over time and become more reliably classifiable based on sEMG activity. Repeating this study with chronic stroke survivors would answer this question.

    Lastly, while sEMG patterns may not contain sufficient information to ascertain intended trajectories, it is possible that the limited information they provide could be combined with other sources of information to result in a more multi-modal classifier algorithm. For example, might eye gaze be measured as another source of information about intended target for movement and combined with sEMG data?

    In summary, Cesqui et al. provide important evidence that sEMG signals alone do not perform well when attempting to classify the direction of intended movements, and give pause to those who would hope to incorporate sEMG data into more sophisticated upper limb robot-aided rehabilitation systems.


    1. Stein J, Narendran K, McBean J, Krebs K, Hughes R. EMG-controlled exoskeletal upper limb powered orthosis for exercise training post-stroke. Am J Phys Med Rehabil; 2007; 85:255-261.

    Competing interests

    Dr. Stein serves as an uncompensated member of the scientific advisory board for Myomo, Inc.