Thank you all for joining the webinar and engaging in our discussion of how we can improve modeling of neuromuscular control in OpenSim. Since we did not have time to get to everyone’s questions, I’ve posted short responses below. Please don’t hesitate to post comments/questions on the discussion forum.
Best regards,
Kat Steele (
kmsteele@uw.edu)
________________________________________________________________
Q&A Responses
Q1: Am I right to think that synergies would be particularly useful to simulate muscles activation when the activity is mainly characterized by muscle group co-contraction? - Dario Cazzola
With this plug-in you can constrain the activation of different muscle groups, which may be useful for modeling activities characterized by muscle co-contraction. For example, if you had an activity where you thought co-contraction was greater than that estimated by minimizing the sum of squared activations alone (the traditional static optimization cost function), you could use this plug-in to increase or require co-contraction.
Q2: How different can a movement task be to see a difference in synergies? Do you see synergy differences between genders? -Jeff Reinbolt
There is some evidence that synergies are “shared” between tasks such as walking, running, or posture control (e.g., for a good example in frogs see d’Avella & Bizzi, 2005). We also tested whether synergies change when unimpaired individuals mimic a common pathologic gait pattern, such as crouch gait, and found that similar synergies were used during normal and crouch gait (manuscript under review, presented at International Society of Biomechanics Conference, Summer 2015). As for gender, I have not seen any specific work comparing synergies between genders (anyone else have a reference?). My question would be, what would our hypothesis be for why synergies might be different between genders (e.g., different morphology or control strategy)?
Q3: Could you explain again how you identify the input synergy matrix (W)? It is great that you could input different matrix in order to couple muscles but I was wondering whether you input also experimental EMG data? -Dario Cazzola
Please see the example files provided with the plug-in. Basically, you provide an Activation Weighting Matrix (ActWeightMatrix) in the .xml set-up file for static optimization. You could calculate synergies from EMG data to create this matrix; however, the challenge is often that we do not have EMG for all the muscles in the model so you need to consider other strategies such as allowing individual activation of muscles without EMG data or grouping muscles.
Q4: Is there a discernable relationship between synergies and optimal control? Or are the synergies a by-product of some unknown control scheme? - Jeff Reinbolt
Interestingly, when we calculate synergies from muscle activations predicted by static optimization (minimizing the sum of activations squared), the synergies are very similar to those calculated from experimental EMG data. Thus, even when we do not constrain muscle activations to synergies, the low-dimensional space is similar to synergies calculated from experimental EMG data. For an example of this in the upper-extremity see our prior publication, Steele et al. (2013).
Q5: Synergies before the model are from excitations, while the synergies after are from activations. Could the model's activation or contraction dynamics account for some of the difference? -Kenneth Smale
Since this plug-in extends static optimization, there are no contraction dynamics (like in computed muscle control).
Q6: What is your expectation if the initial synergies are from the EMG during the same task? -Yen-Hsun Wu
Specifying synergies from EMG data is challenging since we typically do not have EMG data available for all the muscles in the model. However, in this study we did see that the correlation between the activations of synergies decreased our ability to accurately identify the underlying synergy structure. In a small analysis (unpublished), I examined the correlation coefficient of synergy activations from previously published studies of synergies during walking, reaching, etc. In all of these cases, there was significant correlations between synergy activations (typically 0.4-0.7), which leads me to hypothesize that we would not get significantly better synergy identification using synergies from EMG data. But someone should test it out!
Q7: Thanks for great talk. What do you think about using this method in high speed movements such as running, jumping, or falling? - Hossein Mokhtarzadeh
Since this plug-in is an extension of static optimization, it has the same limitations. Static optimization does assume a rigid tendon which reduces my confidence in using this method for high-speed movements.
Q8: Just a comment: this could also be used to simulate Functional Electrical Stimulation. If a nerve cuff electrode is used, all muscles innervated by that nerve have the same/similar activation. Excellent work! - Dimitra Blana
Great idea! Try it out and let us know how it goes.
Q9: What is you view point, this synergies might be only mathematical construct coming due to the task? -Puneet Singh
I think there is probably some truth in this statement, especially with the matrix factorization algorithms we currently use. However, we are still finding that this low-dimensional space described by synergies seems to have clinical utility for quantifying changes in motor control and coordination, especially in stroke and cerebral palsy (Steele et al, 2015). From a less scientific stand point, we also know that chickens can still run with their head cut-off ☺
Q10: I have been working for many years in experimental modal analysis. It seems to me that some methods from that field would be useful in the identification of muscle synergies. Am I right? - Bartlomiej Blachowski
I think there is definitely a need for new methods beyond traditional matrix factorization algorithms for understanding muscle coordination and control.
Q11: What is your view on the manyfold or minimum use theory; theory that contain synergies but aims for a reduction of dimensionality based on a performance variable. Would you implement or consider in future research? - Pablo Ortega-Auriol
I think that there is increasing evidence that humans are not “optimal” in their motor control and that there are multiple performance variables that likely contribute to muscle coordination during a given task. My hypothesis would be that reducing dimensionality or minimizing the number of synergies is not a dominant performance variable. For example, individuals with cerebral palsy and stroke use fewer synergies to walk compared to unimpaired individuals, suggesting unimpaired individuals are not using the lowest possible dimensionality for walking.
Q12: This is all consequence of linear algebra. Finding the synergy weights is underdefined problem. Like finding muscle forces in optimization, you need an extra condition/constraint/cost (e.g. minimizing sum of squared stresses).- Michael Schwartz
I totally agree – we need new methods to move beyond linear programming and optimization techniques to better understand motor control. In this plug-in we are still minimizing the sum of squared synergy activations to find an “optimal” solution, but whether this reflects underlying neural control remains unclear.
Q13: Factorization is not unique solution, what is best way to make the unique or what best possible solution constrain them? -Puneet Singh
Some of the methods we analyzed in the prior study (Steele et al., 2015b), such as principal component analysis, do have unique solutions – but we also demonstrated that PCA performs very poorly in identifying synergies since it assumes orthogonal components. This study demonstrated that although these algorithms can identify the existence of a low-dimensional space, they struggle to find the “correct” unique solution, or how we are rotated in that space. You could also ask the question of whether a “unique” solution exists or whether we our nervous system constrains activations to a more flexible low-dimensional space.
Q14: Some suggest the CNS may have a vast library of more muscle synergies than the number of muscles and a subset is used for a given movement. Any thoughts? - Jeff Reinbolt
I think that for cyclic movements such as walking or running that there is good evidence that there is some type of underlying neuroanatomy that contributes to control and contribution (going back to the chicken without its head again). How we use this “library” or learn to add to it or modify it over time (can we change synergies?) remains unclear.
Q15: How subject-specific are synergies? If one had motion capture data but no EMG, could they use previously reported synergies to weight muscle activations? Would this be reasonably accurate or are subject-specific EMGs important? -Geoffrey Handsfield
During activities such as walking we see that synergies are very similar between unimpaired individuals. However, in our analysis of over 500 individuals with cerebral palsy (Steele et al., 2015) we saw a huge range of synergy structure and activation. We are currently exploring whether these subject-specific differences in synergies can be used to help us better model and understand pathologic movement in cerebral palsy.
Q16: Why did you decide to minimize synergy activations vs. muscle activations constrained by synergies or some other cost function? -Nicholas Bianco
This was mainly an implementation choice with the OpenSim source code for static optimization. I would love to see a separate implementation, especially with computed muscle control, where synergies could be provided as constraints (with some range of flexibility in the weights).
Q17: It seems the correlation of input synergy affect the ability of algorithms to recover it. Could you comment on why even when the correlation is set to 0.9 without arm model the recovery from algorithms is still higher than with the arm model? - Xiao Hu
I think the correlation between synergy activations is just one of the factors that impacts synergy activations. The choice of task, degrees of freedom, number of muscles, and other factors also likely contribute to synergy identification with the musculoskeletal model.
Q18: Concerning infants and toddlers, would the system learn different synergies depending on the environment that are carried throughout life (dysfunction or not)? - Jeff Reinbolt
Interesting question and one we are particularly interested in examining for individuals with cerebral palsy. We know that individuals with reduced synergy complexity have more impaired movement and worse outcomes after treatments; are there any strategies we can use to improve or target synergies?
Q19: Does OpenSim incorporate reciprocal inhibition automatically? Is that going on behind the scenes? If not, can that be incorporated as a type of "anti-synergy" via negative coeficients in the synergy matrix? -Robert Cargil
No – just activation. This is a limitation of all synergy analyses – EMG data only provides us with information on activation, not inhibition, so understanding the role of inhibition in movement is much more challenging. Using this plug-in to explore the impact of negative coefficients would be a fun application – try it out!
Q20: I had a question regarding the part that you changed optimization of "a squared" with optimization of activation coefficients (C matrix). Are there any references or studies on this? -Moein Nazifi
This choice was made based upon using static optimization as the base for this plug-in. Static optimization minimizes the sum of squared activations, which prior studies have suggested is related to energy costs and provides reasonable estimates of muscle activity compared to experimental EMG data (e.g., Anderson & Pandy 2001). Like static optimization, there is not necessarily a unique solution, so a cost function needs to be specified to select from the feasible set of solutions.
References
Anderson, Frank C., and Marcus G. Pandy. "Static and dynamic optimization solutions for gait are practically equivalent." Journal of biomechanics 34.2 (2001): 153-161.
d'Avella, Andrea, and Emilio Bizzi. "Shared and specific muscle synergies in natural motor behaviors." Proceedings of the National Academy of Sciences of the United States of America 102.8 (2005): 3076-3081.
Steele, Katherine M., Matthew C. Tresch, and Eric J. Perreault. "The number and choice of muscles impact the results of muscle synergy analyses." Front. Comput. Neurosci 7.105 (2013): 10-3389.
Steele, Katherine M., Adam Rozumalski, and Michael H. Schwartz. "Muscle synergies and complexity of neuromuscular control during gait in cerebral palsy." Developmental Medicine & Child Neurology 57.12 (2015): 1176-1182.
Steele, Katherine M., Matthew C. Tresch, and Eric J. Perreault. "Consequences of biomechanically constrained tasks in the design and interpretation of synergy analyses." Journal of neurophysiology 113.7 (2015b): 2102-2113.