"best practice" for tuning Opensim models

[John Hutchinson]

Hello,

My team uses Opensim to study animal musculoskeletal function. I've been perusing the human-based literature with much interest over the years, and recently in discussions my team came up with one broad question that we can't find very clearly addressed in the literature. So I thought I'd pose it to the forum. I'm interested to hear different perspectives as there probably is no "right" answer (as usual).

Whether one has an existing model that one wishes to re-tune MTU parameters for to ideally suit a simulation project, or new anatomical data to add to a model, or a newly created model "from scratch" (dissections etc.), what are considered the best practices for tuning the MTU parameters-- in particular, optimal fibre length and tendon slack length?

I'll pose it as a set of more specific questions--- if we set up a model to have some maximal range of joint motions (ROM) in all its DOFs, the first question is more philosophical/conceptual:
1) should we tune MTUs to enable the entire ROM, or should we focus on only the ROM known (or roughly expected for, in the case of predictive/theoretical forward dynamic simulations) to be used in the 1+ behaviour(s) of interest?
To me, this raises the question of what the musculoskeletal system is "optimized for", in a naive (but realistic?) sense. And it risks circularity if we assume that MTUs are tuned for one behavior (and then our conclusion is that they are). Presumably the musculoskeletal system has evolved to be "tuned" to be operative throughout its full ROM of its DOFs (however, much of the limits would be passively supported and not only from tendons; e.g. "reserve actuators"). But if that broad of a level of tuning were implemented in a model, I assume it would be unable to do many behaviors because the muscles/tendons would be too long/short (this does come up against the fundamental limitations of the Hill model and its implementation in Opensim, I know). In practice, it seems almost all studies do the more feasible option of tuning their MTUs to suit their study's behaviors. I'd be interest to hear what others think about this and what the consequences are. (to a degree, it may not matter depending on the question being asked with a study)

And more technically/methodologically:
2) How do colleagues tune their MTU fibre length and slack length values from original values (whatever their source), when they need to? To date, we've used the Manal and Buchanan method: https://journals.humankinetics.com/doi/ ... .20.2.195
But it is likely there are other methods we've missed in the literature. We have generally assumed that muscle fibre lengths should remain within 0.5-1.5 optimal lengths throughout their ROM, and tendon slack length is tuned to keep optimal fibre length ~1 at a resting/standing pose; which again seems to be a common assumption. (and has some evidentiary basis; e.g. muscles tend to be close to their plateau/ascending limb of F-L; usually; and hence in some studies we tuned our model to be ~0.75-1.25 optimal fibre length throughout gait)

(granted the above issues can, are, and often should be addressed via sensitivity analyses, which covers many of the uncertainties)

And a follow-up question:
3) Is there a paper that focuses specifically on these issues above, in detail? In reading the literature I've found it hard to retrace how published models were tuned; such explanations tend to be brief. (I've read the general Opensim, SIMM, other studies on broad approaches to the software and best practices but they don't delve into this specific topic deeply)

Thank you,
Prof. John Hutchinson
Royal Veterinary College

[jimmy d]

Great conversation starter . Have you been able to go through the Verification and Validation, it may be a good resource since it discusses some of the issues you raise here.

[John Hutchinson]

Yeah I re-read that while writing the post and didn't find that it directly answered the questions I was thinking of (yet it touches on them), although we refer to it a lot for other things!

[jimmy d]

That's fair. Let me try and give a shot at a few points, hoping that others will also start chiming in....

1) should we tune MTUs to enable the entire ROM, or should we focus on only the ROM knew (or roughly expected for, in the case of predictive/theoretical forward dynamic simulations) to be used in the 1+ behavior (s) of interest?

In our lab, we typically have the philosophy of making the model just as complicated as needed to answer your research question. Since the models can quickly become highly complicated, with so many variables that require 'tuning' to the subject, it is close to impossible to have all things work under all circumstances. For example, in our gait models, many of the muscles that cross the knee don't give reasonable answers outside a certain joint range. People should be aware of these limitations and, if you want to study something like muscle function during deep squatting, you will have to make significant model edits.

To me, this raises the question of what the musculoskeletal system is "optimized for", in a naive (but realistic?) sense. And it risks circularity if we assume that MTUs are tuned for one behavior (and then our conclusion is that they are).

This is a good point; many people do this. We try to use as much real data as possible to validate the estimations we have and, when we do make statements about muscle function, it is specific to the motion being studied. I think (hope?) that most biomechanists understand that muscles have different functions during different motions, and I would always suggest that we refrain from making broad and authoritative statements drawn from modeling and simulation; modeling and a simulation is just a tool that should be used in support of experimental studies.

[John Hutchinson]

Thank you James! Fully agreed on the broader points of fitting simulation complexity to the question and the importance of validation; and not over-stating the implications of simulation results. I probably should not have posted during Thanksgiving but in the UK we had a late turkey dinner, so my brain was on science rather than pumpkin pie.

I hope people chime in w/thoughts about the specifics of tuning and what are best practices for setting joint ROMs vs. muscle fibre lengths vs. slack lengths etc. (I don't think there is one best practice but I'd be intrigued to hear what the range of them is; and if people use a tool other than Manal and Buchanan 2004 to set slack lengths)

[Apoorva Rajagopal]

I can chip in a little bit here. Our lab published a lower limb OpenSim model a couple years ago. For that model, tendon slack lengths were set using the same ideas as described in the Manal and Buchanan paper you referenced. (We posed our OpenSim model in the same position as the cadavers from the experimental data we referenced, and set the tendon slack length of each muscle such that the normalized fiber length in our passive, equilibrated model muscle matched the reported normalized fiber lengths in the experimental data.)

We also used the variability in the published experimental data to compute and report the expected variability in the tendon slack lengths (and other parameters) for each muscle. Our hope was that users could use this reported variability to help tune and/or run sensitivity analyses within the scope of their own simulations-- i.e., if some muscles seem to be unrealistically tight or slack, the computed tendon slack length variability can be used to inform bounds within which one can tune tendon slack lengths.

We were able to take advantage of a couple of really nice published datasets to build our model. Unfortunately, when this isn't the case, it can be challenging to know reasonable bounds within which to tune model parameters. Your note about using sensitivity analyses to strengthen simulation findings is a good one and may ultimately be the best route to address some of the questions you posed. Another option is to use some set of motions to tune the model, but use independent motions or data for model and simulation validation. Validating on completely independent motions can help avoid the overfitting problem that often arises when a model is heavily tuned to a motion.

[John Hutchinson]

Thank you Apoorva! Interesting points. Yes I like that in-situ approach to setting slack lengths from the experimental fiber lengths; it is maybe the best I've seen in the literature. We may use it ourselves if we can manage with our small animals. I am a bit baffled though by hearing from 1 research team that they tried fully forward dynamic (theoretical-predictive; not tracking/CMC) simulations with that model/the Arnold et al. one and found it was "too poorly tuned" to walk normally, in their framework. In contrast, the standard Delp et al. one was well-tuned for their framework.

That's what stimulated my original post, as to me it seems like the newer models are, or should be, designed to be superior to the Delp 1990 one, and yet hearing that the tuning did not work out well seemed to be a non sequitur, especially as it was for walking, which is what these models have been used for, for decades. Granted, the Delp 1990 model has had those decades to be tuned better, and maybe it has been improved a little vs. the 1990 version, but I am not clear if/how it was tuned from that original. And the phenomenon of tuning itself, while superficially simple ("just" 3 things to tune: joint ROM, fascicle length, and tendon slack length-- and maybe muscle max force), is perhaps so complex especially when used in novel ways such as a predictive simulation rather than tracking, that maybe it should be expected to give non-sequitur surprises? Hence my philosophical side still wonders how we can optimize chances we'll come out with a "well-tuned" model (whatever that truly means!) when we are collecting original new anatomical data and making new/modifying old models.

[Ross Miller]

My preference is to fiddle with muscle model parameters so that the model's joint-level strength resembles human dynamometry measures. Anderson et al. (2007) is a great reference with "normative" data in men and women, young and older:

https://www.ncbi.nlm.nih.gov/pubmed/17485097

Or this older reference:

https://www.ncbi.nlm.nih.gov/pubmed/6376139

There is a concern of over-fitting with this approach depending on the details of how it's implemented, and it also opens up a somewhat philosophical debate on what the "maximum" isometric force is (the maximum force that could ever be produced under any circumstances, or the maximum force that is produced in practice, or [option 3]).

[John Hutchinson]

Cool points, thanks Ross! Now if we can just get our crocodiles into a dynamometer...