The vast majority of biomechanics research has focused on explanation and not prediction. We are going to change that. If this interests you, get in touch and we can tell you more!
I was once involved in a pretty wild data collection that was around three hours long. We were testing knee braces for medial compartment osteoarthritis. I was collaborating with a very smart researcher, who was thinking about a lot of different questions, which is why the collection was so extensive. We were instrumenting the brace, collecting EMG, and a full body marker set, doing a variety of tasks. There was over-ground walking, treadmill walking, uphill walking, among other activities. As you can imagine, it was pretty tough on the participants, but like most people with OA, they were great sports, and we heard almost no complaints. Even when we lost markers that required re-collection, everything was good from their perspective.
And I forgot to mention, we did the entire protocol on two braces (as an aside, this fundamentally changed my view on experimentation, because it was so hard to get it right)! At the end of this marathon collection, one participant asked me “So, which brace is best for me?”, and I basically replied, we don’t know yet, we have to wait and see what the data says. He was nice enough about it and didn’t mind that lackluster answer. However, I knew that I wasn’t being entirely honest. There was no way that after looking at the data we would be able to give a recommendation that was specific for him.
The fundamental problem wasn’t with our wild collection protocol being a bad experiment, it was that we just couldn’t collect enough data to make this assessment. And to be honest, it wasn’t really the goal. The goal of the experiment was to compare the braces, and see if there were kinematic differences that explain their unloading effect. In other words, it was to explain a phenomenon, not predict for an individual. For this nice man that endured our experiment, nothing we did was helpful.
Now at this point in time, machine learning was in its infancy, but I had somehow or other been connected with a researcher who had been doing it for years. We lamented about this experiment but the outcome was even less satisfying; we weren't even close to the number we needed to collect in order to have predictive power. He is generally pretty optimistic, but I believe his words were, “Goodluck with that.”.
I genuinely wondered where I could go from here, because it didn’t seem likely to get to a predictive space that really helps that individual. And in reading literature (with this in mind), I realized everyone was having a similar problem (whether they knew this or not). As a community, predictive analytics were just out of reach.
This was a big factor when designing and developing a system that could do things a bit faster! Now, of course, it had to be accurate and repeatable, because lots of bad data does nobody any good, but it also had to be capable of collecting lots of data. These were defining factors when we started this pursuit, but also how we approach problems on an ongoing basis. Now, that very lab where I collected that nice man through our insane collection has collected as much data as I collected in that study in literally 6 hours. To put it in perspective, our collection took 5 months, and I thought that was pretty good!
My hope in writing this is that it will inspire others who were similarly frustrated with the collection modality to think differently about their collections and what is possible when you are able to collect large amounts of data.
Comments