I took the liberty to review published error rates for myoelectric arm controls since ~1980. In other words: how reliable is the myoelectric arm control? How has the reliability changed over years?

This is hard data that is somewhat impossible to negotiate, and it has always been there for anyone to take. So forgive me when I anticipated some clear conclusions from such information over the last years – after all, I figured, a post such as this would not run away ; ) Even though, you must keep in mind that you read this here first as well. As far as I know, despite its ubiquitous availability, and despite a very obvious background or social reason for this (which is yet another subject), this collection of data has not been formally performed before. In fact, word of mouth had it that even decades ago, no self respecting engineer would deliberately enter the field of myoelectric or robotic prosthetic hands, simply because of all options one could do, that certainly was never a prosperous looking one.

The reason for this subject choice is, that I was interested in possibly identifying a trend, or a useful figure, for the reliability of myoelectric arm controls to consider in context of daily use. I used standard key word searches on Google Scholar and edited the resulting publications for relevance. I thus performed more systematically what I had done anecdotally a few years ago when it became obvious that one should not put too much hope into such prostheses, as their ongoing and intractable issues were somewhat obvious all along – at least from view point of a right below elbow amputee performing real work.

Anticipating a summary:

• The error rates slightly increased, on average, over about four decades.  Regardless of that trend, error rates are mind-boggingly high: The error rates for grips with myoelectric control are in the range of about 1,4 (~1980) to 1,7 % (~2018), on average, for all tests done, most of which were laboratory tests. The published error rates are in the range of 11 to 35% for tests that approximate realistic use situations. A useful error rate for a prosthetic arm, from a user view, is maybe around but clearly better below 0,01%. Myoelectric prostheses have not achieved that even remotely in four decades of R&D. A realistic estimate for a body powered control error rate ranges between 0.0003 % (sic) and 0.002 %. That is dimensions better. You are probably not familiar with that technology, but it may be time you made yourself familiar with it. Useful error rates have never, ever, been achieved in laboratory tests for myoelectric prostheses. Useful error rates have in no way, ever, at any time, been achieved in tests approximating realistic use situations. And most definitely, useful error rates have in absolutely no way, ever, been at any time achieved under real work constraints [link].
• There are realistic tests, and there are laboratory tests. Realistic tests will incorporate body or limb position changes and they will test the reliability of myoelectric control under realistic constraints, where there is sweat as well. While no one has dared to publish just how bad sweat interference really is, realistic tests will yield error rates ranging from 11 to 35%. That figure looks quite realistic. However, the cost for the user may be extreme since grip failures often cause object drops or object crashes.
• There are tests with amputees and there are tests with “healthy”, i.e. anatomically intact, test subjects. It would be relevant to examine more closely who the test subjects were – at least, if at all error rates were in a usefully low range. As we are far from that, I did not bother to also plot that data. But before you start your darnedest to use anatomically intact individuals for SNR-qualities of 12 or 20 or so, please know that amputees have atrophied muscle and skin issues and so on, and inherently lower SNR to start with. And, which is worse, we know that. We will look for that. We will consider this.

Error rates of myoelectric control as published in academic literature (selected publications, overview)

The diagram shows the year of publication (x-axis), and the logarithm of the error rate (y-axis). The dark circles represent academic publication issued myoelectric control error rates, and for reference, the publication titles are included in the diagram as small text (there is a linked PDF when clicking the diagram and if you open that in a good PDF reader you can actually read the text). Realistic tests, where error rates result from testing procedures under approximated real life conditions, were plotted with large red circles. Laboratory tests (under assumedly “ideal” conditions) were plotted with smaller black circles. All published error rates were used to obtain a linear regression (blue dashed line) which increases from an error of 1,39% (1980) to 1,72% (2018). A halfway useful error rate from a user view is 0,01 (10E-2), shown by the red dashed line in the diagram. No study even remotely approximated this barrier. See below on how to crack this barrier with really good engineering.

My personal, user desirable maximal error rate will be 0,01%, also due to a projected cost inferred by dropped objects. While you may argue that no arm amputee should use any other than plastic cups, or, for example, handle any photo cameras other than protected by large rubber casings, and while you may hold the opinion that dropping an item in 2 out of 100 grips is no biggie, and while that attitude clearly is reflected in the R&D data that is available, I will put to you that the reverse view is reflected in a staggering rejection rate of prosthetic arms. Why? Because we happen to inhabit a real world with real life grip situations. These appear to be hardly ever tested in a hard way for most academic paers. And the few times they are tested, error rates ranging from 11 to 35% may result.

That slight increase in linear regression derived error rates over the years may not be a particular concern – if anything, true life or realistic testing with very high error rates may skew the trend of other error rates. And indeed: if one just examines the laboratory derived error rates, omitting the catastrophic results from more realistic testing, then an linearily regressed error rate increase from 1,45% (1980) to 1,64% (2018) results (oh no, not 1,72%!!). So: a linear regression performed exclusively on the particularly convenienced testing under lab conditions per se still yields an even milder increase over the years, not a decrease. And that may be a bit more of a reason to be concerned. But I still think this (the increase, that is) is not a particular cause of concern, simply because these error rates as such are still far too high anyway. In other words, myoelectric control is so unreliable that a little increase over four decades of focused research into this really is not the problem – the actual absolute control failure rate is the feature here. Or to put it like this: as the increase occurs within a range of errors that is unacceptable anyway, the ballpark, the overall range of error rate values, are the real issue here. Of course you could also ask, how it comes that current researchers do not wish to outperform their predecessors – but then, who really reads old studies? Who correctly cites prior works? Who actually reflects on error rates of myoelectric arm control? As long as the correct recognition rates are somewhere in the “nineties”, the paper can be submitted and no one cares.

Cost per year for user in relation to control error rates

Myoelectric control failure can be assumed to, also, or, as example, result in a crashed dish when unloading or loading a dish washer. Assuming that we handle 12 dishes a day, and assuming the cost for a single porcelain cup in Switzerland (around 12.40 CHF), error rates can be used to project actual total cost per year, and financial loss can be given for illustration. Here you are, reading this, and you thought that the TV-show “Destroyed in Seconds” was cheap entertainment!

A R&D-typical laboratory myoelectric control error rate of 2% will cost the amputee around 1000 CHF alone in porcelain damage, were one to rely on a myoelectric prosthesis for ADL such as loading or unloading a dish washer. Realistic error rates of 11-35% however will run up entirely insane costs of 6000 to 19000 CHF only in broken porcelain – not counting other items such as cameras or cell phones, which are a real major factor for arm amputees due to the high cost when dropped and damaged. These projected porcelain damaging costs are matched by annual costs of blindness, to make an easy to understand comparison [1] only that here, cost and efforts for purchase (absolutely insane) and repair and maintenance (also extreme) of the myoelectric prosthesis are not even included. These are the true underpinnings of a reality, where myoelectric arm users will quite simply have use their anatomical hand to perform repetitive and strain-/overuse-relevant tasks such as loading or unloading a dish washer. No one will allow annual damages of 19 000 CHF of dishwasher loading / unloading incidents to happen of course. But your guess is as good as mine when we rhetorically ask whether such an orthopedic “aid” is an actual aid.

The entirely absent improvement of published error rates over a time span of about four decades, combined with the associated high error rate associated costs, may be the main reason why myoelectric arms and their control are a de-facto Dead Horse.

Their extreme degree of failure under realistic conditions [link] points to severe intractable and inherent technical issues that will not just “go away” [link]. R&D proponents of myoelectric technology repeatedly stated that “technology is here to stay” and that people such as I cannot make it go away. I do not want functional technology to go away at all – I want trash, bloatware, unfounded claims, hyped up promotion to go away. And besides, I am not required to make myoelectric stuff “go away” – that technology is doing that by itself, it fails >85% of the potential users already, so just how much worse could it be? Such an attitude or statement, however, questions the technical foundation on which they are based: on what exact performance do researchers or developers actually base their statement that “myoelectric control is here to stay”?

There are no really polite answers: either they do know this and they clearly give wrong information, or they do not know this even though they should. We also wonder what the user logic is that R&D applies when targeting a more aggressive body scheme integration of myoelectric arms when error rates are not just bad, but really bad, and consistently so over almost four decades: how exactly do they assume that the adoption of these types of devices into a body scheme functions, when their main feature is actually “constant failure, reliably unreliable over four decades”? What, in single steps, and explained so everyone understands, is it they assume our user perspective to really be?

Besides, high myoelectric device failure rates are not a secret at all. We all know that, users are painfully aware. It is a fact that we deal with all the time. We may not rub it into your face each and every time, but we know. Even outspoken media ambassadors for expensive “bionic” hands have been known to not really wear the prosthesis all the time, or even nearly as often as they claim to do. These grip failures have been a constant source of problems all the time. We also talk about it on social media [link].

If you are a reviewer of a scientific paper that, for example, publishes a “98% correct recognition rate for myoelectric control” as a really “good” result, at the same time, with that, in the real use context, you clearly state that you do not care a lot if the amputee that may end up wearing that technology ends up paying >1000 CHF in crashed porcelain and possibly >500-2000 CHF per year in damaged electronics (i.e., camera, laptop, cell phone), without even considering the social and emotional aspects that go with this, which may be considerable.

An underlying technical problem here may not just be that SNR is relatively low for myoelectrodes placed on amputees’ arms, which directly causes error rates to be excessive, but also, that inherently unavoidable and practically very significant control errors will be added by muscle cross-function-use (i.e., activities such as elbow activation and stabilization use the same muscles as “hand open or close” signals, so there is no easy technical way to reliably separate these) and by skin surface issues such as dry or wet (sweat) conditions. After error rates of myoelectric control now has remained unacceptably high for decades, it is not even a question how R&D wants to address any technical issues now. They already tried. They have continuously failed for four decades. They have no plan. They have no options or fall back. Signal processing with amputee arm stumps never works as garbage in gospel out. It is always a variation of garbage in garbage out. Already in 1970, that was recognized but, as it appears, not promoted loudly enough [2]: “it is difficult to see why designers continue to search for improved methods of conditioning myoelectric signals in an effort to improve “proportionality,” i.e., the relationship between myoelectric signal and mechanical output when the results cannot be better than the raw, fundamental data. They would be better advised to avoid over-designing hardware beyond the limits of the basic phenomenon. Refining hardware in electronic circuitry to provide better proportional control than is  available in the patient is like designing a Hi-Fi system which faithfully reproduces a million cycles per second. This is high fidelity indeed;but no one can hear it. Recognizing that only a rough proportionality exists between muscle tension and myoelectric output will lead to more economical and more useful components” [2]. They had performed their own case-based study: “The common concept underlying the production of a myoelectric signal is that each muscle fiber produces an electric discharge pattern that is similar from fiber to fiber and muscle to muscle, at least as regards skeletal muscle. If the contraction of each microscopic fiber yields an increment of contractile force and electric signal, then an area over a muscle, sensed by a surface electrode, should show a definite relationship between the total tension in the muscle and the total electric output of that muscle. A simple experiment conducted in the VAPC laboratory demonstrated that at best, only a rough relationship exists between muscle tension and myoelectric output under conditions appropriate for prosthetic control, i.e., isometric contraction against sub-maximal loads” [2]. Please note that the statement issued contains “at best” as delimiting narrowing term for the hopes and expectations. That was 1970, fifty years ago. Four decades of published research, and no improvement of myoelectric control error rates. So, this is the end of a story now, as it had been the end of that story after twenty years already, which was already twenty years ago. You do want to see this in perspective here. It never worked, it never will work.

The actual question is what really is going on with further perpetuation of the subject matter of myoelectric arm control. If there is proof for one thing at all, it is this: actual technical improvement? That is certainly not what has been going on. Using prosthetic arm device development therefore has been used, over decades, to get financial funding for a thing that appears to work (but does not). It has been used to train, educate, illustrate, without reaching an actually useful degree of function.

Error rates of myoelectric control as published in academic literature (selected publications, overview) compared to body-powered control error rates

A well-built body-powered arm certainly has no control errors of that kind.

There, a typical control error will be when a cable breaks or requires repair due to cable sheath or hinge part decay, which may happen maybe once in 6-12 months of continuous use if the arm is built correctly. In my instance that is the case since around 2011. If one assumes 200 to 800 grips a day, a 1/146000 (0.0003 – 0.0006%) to 1/36500 (0.001 – 0.002%) error rate results for a well-built body powered arm. Before, using commercial parts, I had error rates with cable tear ups within 4-10 days, resulting in a percentage for control error rates of 0.03 – 0.12% (given a bracket of 4-10 days with a bracket of 200-800 grips per day). With own user driven innovation, I improved this to 0.0006 to 0.002 % that I enjoy since 2011.

Same diagram as above with respect to this: The diagram shows the year of publication (x-axis), and the logarithm of the error rate (y-axis). The dark circles represent publication issued error rates, and for reference, the publication titles is included in the diagram as small text (there is a linked PDF when clicking the diagram). Realistic tests, where error rates result from testing procedures under approximated real life conditions, were plotted with large red circles and laboratory tests (under assumedly “ideal” conditions) were plotted with smaller black circles. All published error rates were used to obtain a linear regression (blue dashed line) which increases from an error of 1,39% (1980) to 1,72% (2018). A halfway useful error rate from a user view is 0,01 (10E-2), shown by the red dashed line in the diagram. No study even remotely approximated this barrier.

Added in comparison to above: green lines with upper and lower boundary of error rates of my body powered arm: 2008-2010 (all commercial parts, 0.03-0.12%) were massively improved through user-driven innovation (now, 0,0003 to 0,002%) [link].

Error rates of myoelectric control (iLimb Touchbionics) versus body powered control (Becker Imperial hand) – real testing

Real testing grasping and lifting (by elbow flexing) of an electric power drill was performed using an iLimb (Touch Bionics).

The error rate is around ~40%. The error is mainly caused by the body posture or limb positioning effect, whereas the elbow flexion causes the prosthesis to open without dedicated signal or targeted open command. That means that four out of ten manipulations of this kind will cause an object drop. For easier viewing for those that find that very hard to believe, the video has been captured in slow motion. Please take ample time to rewind and play this if there is anything to take from this for, say, trouble shooting. However, forty years of academic development level of trouble shooting have turned up nothing in terms of better control reliability for myoelectric arms.

The video of my Becker hand (Becker model: Imperial hand) shows ~100 grasp / elbow flexion lift actions or more, none of which causes even the slightest control error. There is a 0% error rate over this time frame. The cable will break, eventually, after 9-12 months of daily use, but that is nothing to occur here, I did replace the cable some time in November and so it is probably still good to go for quite a while.

The Becker hand is one hell of a reliable beast, attached to the right type of body powered setup the results are extreme in terms of true reliability. I wear a PVC Centri glove over it, and admittedly, the surface of the hand was modified by adding some extra padding before placing the PVC glove onto the device. So, grip is really neat.

A reason why the internet may not be jocked full of videos showing just how reliable a Becker hand (or a body powered hook) is, is, because, well, they are. There are not many two or four hour videos showing nothing to look at in terms of other stuff either; so you will probably not find people that find wearing their glasses great and comfortable and that then provide a minute account of just that experience by overly documenting every detail for a few hours or so. That is just not done. We are not a culture that neatly documents event free well being. No one does that. It would be stupid, as we are a society of barbarian novelty seekers, we want to revel in the absurd, outrageous or damaged, the annihilated, the peculiar or the voyeuristically interesting. So this video is only to show you that if ever you ARE after a base line for error rates for a body powered arm, that it can be done. You could have your test person repeat that device / object lift a few thousand times and document it on video and still be trucking on like Super Mario Brothers to see whether also you are able to build a body powered arm right.

It is a bit like when they said that contestants on the Cybathlon have to carry a tray for the prosthetic arm race, and so on. Hell, they make them work hard I guess. But screw it, I probably carried a tray with my food on it in our workplace Mensa every day for over ten years, so some ~2500 times, not a single time of which the tray dropped or anything like it – so I guess I did not feel the urge to “compete” for something I totally do all the time. I cut bread and other stuff in the kitchen, routinely, without drama. I can also hang and collect laundry or load / unload dish washers really well. Reliability on a relatively extremely high level is something my body powered arm was built to deliver, very naturally. It is unheard of to have that performance for other users – but seeing as we have arrived at a place where social tension fields define the survival of the fittest, you all get to look after your own self. So while permanently excessive error rates a huge thing for myoelectric arms, they are not at all a problem or subject, an issue or topic for my well built body powered prosthesis.

What does this mean?

You first need to truly understand what that means though.

My almost absent or very-near-zero body powered arm control error rates are legendary, outrageous, that type of performance is unheard of in terms of control error rates from an R&D perspective. It is so good you would want it right away, were you after perfectly low control errors. It means, far less trashed, crashed, broken, dead or fractured objects. It means far less arguments with the objects’ owners. It spells out as a lot of relevant things. Conversely, any other trend in R&D clearly speaks to an absence of such understanding, or preference. While you may think, now, that it is not terribly hard to understand all this, the stalled situation, the useless aspect of myoelectric control, the user side with high costs of overuse of the other anatomically intact arm or then the frequent object drops, the frustration of users in light of extreme added costs in real life, I could not possibly comment about the lack of difficulty contained in this subject matter that will be required to understand this. Conversely, eight years ago, I decided what exactly to improve, what parts to go after, and see where that got us.

Looking at what is going on in R&D, there is an ongoing massive focus on myoelectric control. Without improvement, as always. Thus, it is fair to suggest that the cognitive burden to understand how a body powered arm with a realistic 0.0003 % error rate is just maybe a tad bit better than a myoelectric prosthesis with a realistic error approximation between 11-35% (more like ~40% when checking on taped video) must be immense. And that is where words leave us.

Strategically, well-built body powered arms are a true, clear, hard and unforgiving threat to the myoelectric control attempts that just have stalled for four decades now. The very high reliability, financial aspects with far lower prices and repair costs, and error rates that are a few dimensions lower, reveal themselves by a single glance upon the make and failure experiences over real life situations. To figure that out is not even a bit hard.

So clearly, research and development wish to portray body-powered technology as ridiculous; one way of communicating that is to claim that a myoelectric prosthesis “places a cognitive burden” [3]. Another attempt to denigrate body-powered technology was made by a study [4] that claims that body-powered arms are not possible to use without fatigue, when in fact, except that single one person that was a true daily body-powered arm user (subject #7) did not exhibit discomfort or significant fatigue over daily use, none of the others were habitual body-powered arm users, and when they also had the study collective perform extreme, full effort pulls on that cable to maximize discomfort using a figure-9 harness that is neither made to sustain prolonged use to begin with and that is hazardous to the brachial plexus. Furthermore, their setup lacked proper cable mount. In other words, they did not even use what I call a well-constructed body powered arm. Correct construction requires minimization of cable sheath resistance by providing a proper Bowden mount [link] and it requires a shoulder anchor that is molded to the anatomy of the user and that bypasses the brachial plexus [link]. Any other is technically suboptimal, and if manufacturers claim otherwise, one has to consider the CE conformance on a very detailed technical level [link] before submitting research subjects to scientific testing using materials that are suboptimal. If you have reseachers, that do not understand how to build a prosthesis well, and that do not understand that some body-powered products are commercially sold without actual technical checks or tests, and if you have these researchers dish out results, you cannot but truly wonder where the future is going. After a body-powered arm has been built correctly, also, fit and use training is required to build power and comfort, as that has to grow with time; after all, “body-powered” means just that.

What does place a cognitive burden is how R&D keeps riding this Dead Horse despite massively better performances of the body-powered control paradigm.

@article{meads2003cost,
  title={What is the cost of blindness?},
  author={Meads, C and Hyde, C},
  journal={British Journal of Ophthalmology},
  volume={87},
  number={10},
  pages={1201--1204},
  year={2003},
  publisher={BMJ Publishing Group Ltd}
}

@article{mason1970practical,
  title={Practical problems in myoelectric control of prostheses},
  author={Mason, Carl P and Engineer, BSME Senior},
  journal={Bulletin of Prosthetics Research},
  pages={10--13},
  year={1970}
}

@article{kutilek2017myoelectric,
  title={Myoelectric arm using artificial neural networks to reduce cognitive load of the user},
  author={Kutilek, Patrik and Mares, Jakub and Hybl, Jan and Socha, Vladimir and Schlenker, Jakub and Stefek, Alexandr},
  journal={Neural Computing and Applications},
  volume={28},
  number={2},
  pages={419--427},
  year={2017},
  publisher={Springer}
}

@article{hichert2018fatigue,
  title={Fatigue-free operation of most body-powered prostheses not feasible for majority of users with trans-radial deficiency},
  author={Hichert, Mona and Vardy, Alistair N and Plettenburg, Dick},
  journal={Prosthetics and orthotics international},
  volume={42},
  number={1},
  pages={84--92},
  year={2018},
  publisher={SAGE Publications Sage UK: London, England}
}