The second part of this interview features one of the most ambitious projects I have ever heard of. Kyle Boddy and Driveline Baseball have a desire to revolutionize rehabilitation by using machine learning techniques to predict and prevent player injury. Read Part 1 here.
If you have had Tommy John surgery and would like to participate in this series, feel free to email me at email@example.com, or DM me on Twitter (@ShawnBrody) to let me know.
Shawn Brody: Looking ahead to the future of Driveline, is Tommy John surgery rehab an area you’re looking to expand into?
Kyle Boddy: That’s a great question, and it comes at a great time. Yes. We actually hired a Doctor of Physical Therapy full-time. He has worked here part-time for almost a year now. He has been very involved with our programs not just on the rehab-side, but the training-side. We include physical therapy for free for all of our athletes. We think it’s most important for asymptomatic athlete. That you should be seeking physical therapy while you’re heathy, which is exactly the opposite of how it is done. We include that in our membership fees at no additional charge for our athletes, and it has been a huge gain. Dr. Terry Phillips now works for us full time, and we think that it is just a huge opportunity for us.
Now he’ll continue his work on asymptomatic, healthy pitchers, but our hope is that we will be able to open up a rehabilitation clinic inside of our complex. We would have a credible rehabilitation clinic where teams are sending their players to us. We’ve already had inquiries from MLB teams to see if certain players can rehabilitate Tommy John surgery at our facility. Currently the answer is probably not, I wouldn’t feel good about that. On a case-by-case basis, like a Troy situation, sure. But from a blanket perspective, we’re not there yet.
First of all, we can’t accept insurance. So, there is that. Teams aren’t going to pay cash. We need to figure out how to do that, and there’s a lot of logistics behind it, but I couldn’t be more excited about it. It sounds really mundane, but I get excited about the boring stuff. We just got IRB approval for a peer-reviewed study that we’re doing, and that stuff is exciting. I really like it. It’s bureaucratic, and it’s ridiculous, but I think it’s the way to go. We have to do it to change people’s lives.
So, the short answer is, yeah. We hope that we can open something within the next year. I’m optimistic and aggressive in thinking we can do it much sooner than that but, at the end of the day, if we open up a rehabilitation clinic by June 1st, 2018, I’ll feel good about that. I’d like to do it a lot sooner, but I think that’s a decent target date.
SB: A lot of the time you hear people think coming back from Tommy John surgery comes with an automatic gain in velocity, but studies show that pitchers, on average, lose velocity. Do you think that the training plan Driveline develops could help combat that?
KB: Yeah, because what we do here is not performance enhancement. It’s not rehab, velocity development, or strength building. It’s all the same thing. I don’t care where you are. Everyone has individual needs depending on where they’re at. It’s all the same thing. It’s all player development. It’s all the same up-slope. There’s no reason that it should be separate.
This is what I told the Mariners in here. One of their front office people, he is still pitching in Triple-A and pitched in the big leagues. But he’s suffering a labrum injury and is probably going to join the front office. His biggest worry is that rehab is so separate from pitching; separate from skill, and so I told him that was the problem. We visited a company called ALTIS down in Phoenix, they train some of the best sprinters in the world — primarily the US. They have a ton of the best men and women sprinters there, but they dominate with women’s sprinters. They crush the Olympics.
So, I wondered what I could learn from a place that trains athletes far better than baseball athletes. One of the head guys, Stuart McMillan, asked about our rehab model. I told him we have a Doctor of Physical Therapy, manual therapist, and an orthopedic surgeon on the board of advisors. He said that’s good, but asked how we integrate it. I said when guys get hurt, and they see Terry.
“That’s not it,” [McMillan said], “the loop is too large. When we take our Olympic caliber athletes to the track we bring the manual/physical therapists with us, and they deploy their massage tables on the track.”
What they do is they have their athletes run, and they evaluate them subjectively with their eyes. They then say, “OK, this doesn’t look right. Your ankle inflection doesn’t look right, your knee bend is too steep, so what we’re going to do is immediate manual therapy on you to try and free up some tissue. Then you’re going to run it again, and we’re going to do this over and over until we get it right or we’ve exhausted our training economy.” The manual therapy and the physical therapy is built into the training cycle, it’s not separate.
When he showed me how it was done, and he told me about this, I was shocked. I would say that this is the single most important thing I’ve learned in the last year. To shorten up that loop was such a change in paradigm that I had never seen before, and it immediately made total sense to me. So I went back to our strength and pitching coaches and told them they needed to become CSCS certified; that they need to go to PRI and grafting classes. That they’re going to learn how to do manual therapy. Most of them loved it, some were wondering why they had to become physical therapists.
“Why do I have to do manual therapy? I’m a pitching coach.”
So I told them the same parable Stuart told me with how he works with his athletes and, man, everyone got it. You can’t really expect a physical therapist to become an expert in baseball — there’s a lot of feel. For better or worse, most people have to have played the game or coached a long enough time to understand the human dynamics behind it.
Physical therapists just don’t have the time to do that, so we need to have coaches that can understand the physical side of it. They don’t need to know it all; they just need to know enough to bridge that gap and have an immediate feedback loop. And, I’ll tell you what, the benefit to our athletes has been tremendous. That has been the biggest thing.
The second biggest thing has not yet been built. As an industry, we look retroactively at injuries. So, Shawn Brody is pitching for me, and you hurt your elbow and have a UCL tear. OK, then I have to go look back at the last 24-months to see your training logs. If I’m a good organization, be it college or pro, I have two years of your data. I know when you threw your bullpens, your velocities, weighted ball program, long-toss program, medicals, and your MOTUS sleeve data.
Almost nobody collects all this data, but we do. I mistakenly thought that this was going to be enough, but it’s not. If you think about it, all you’re doing is taking this data and retroactively looking at what happens when someone gets hurt. When someone gets hurt, we’re essentially using that person as a martyr.
“You got hurt, now we’re going to learn why you got hurt, and we’re going to apply that to make everybody better.”
At the end of the day, you still have a torn UCL ligament. That doesn’t make you feel any better; all I did was use your experience to better other people. That’s good, you can feel good about that and I can feel good about that a little bit, but at the end of the day you’re still hurt. What we need to do is have a unified software program that is consistently running machine learning on it. That was my expertise before I got into baseball full-time. I did a lot of machine learning, prediction, and projections for, amongst other things, taxes. It was a lot of fun.
A lot of it is prediction, and no different then what is actually being done in pitch classification with k-means clustering and Kalman Filtering. Figuring out what happens before it happens. The idea is that we have everybody in an organization — be it college, pro, or Driveline baseball — all logging data in a normalized fashion. Our clinicians, trainers, strength coaches, executives, and trainees all logging data the way they need to.
That creates this feedback loop constantly running in the background — maybe everyday — chewing up the data saying, “these are the red flags that are producing injury, and these are the people that are on the track to being injured.”
So, there’s two ways to do that. There’s supervised learning, which is to say, “these are the players that have been hurt and this is their training log.” It’s retroactive, and what we do now. The machine learning program then looks back and says, “these are the signals that seem to have produced injury in these players, so let’s use those as the starting weights.”
You then run that program against the database of all the athletes every day, which allows you to give a percentage, based on the model, of how likely they are to be hurt. So people that are red-flagged, we know that we should do some more recovery or change their program. That’s exciting, and that is where we’ll start.
The second part is unsupervised learning, where the machine learning program is looking at all the data every day and doesn’t know anything. All you’re telling it is, “I want you to maximize these outputs, but I’m not going to tell any history of what makes anything good.” That is the real exciting part, that’s true AI. That is actual Artificial Intelligence where the program is actually trying to figure it out, because it is too complicated for any human to figure out using statistics. That is the ultimate goal of Driveline baseball.
This is the stuff that we tell people, and they’re shocked because they think that we’re just a weighted ball and velocity development company. That’s true, because we are and always will be. That’ll always be part of what we do, but we have aspirations that are so far beyond what I think anybody has heard of. We have aspirations to have machine learning make rehabilitation almost a thing of the past. That this is something just integrated into everyday life.
We have an initiative to publish over 12-15 peer-reviewed research papers over the next five years. We’re moving into a sports science/sports technology, pure player development company; the likes of which does not exist, and probably can’t be conceived of. That is the direction that I am trying to move the company, and we’re getting there.
We’re really in the infant stages of it because we wanted to master the training aspect first, and I think we nailed that down. I’m not saying that I can’t learn more, but I feel good enough about that to start moving my attention full-time to the director of research and development. I used to be a lead trainer and the president of the company, which I still am, but now my role today is getting my hands around what this company will be in five years. Not if this company will be financially stable in three months, that is my CEO’s job — and he does a great job of that. My job is to say what is this company going to be in 5-10 years, and what problems are we going to solve that no one is thinking about now.
This is stuff we’ve been kicking around for the past three years, but have really just enumerated it over the past six months. We’ve hired a ton of people over the last six- to eight-months, so we finally have the right people to do it. I want to be the Bell labs, the Xerox PARC labs, DARPA; I want to be that for baseball. I want to develop stuff that no one has ever thought of.
SB: Have you tried to apply any of that yet, or is it just in development stages?
KB: Development, unfortunately, because we don’t yet have a unified data collection interface. In full disclosure, we’re using Google Sheets right now. We have a huge collection of Google Sheets. You can argue that’s a database of sorts, but we don’t have a good way to actually analyze it. We’re doing it retroactively.
I can tell you that, from experience, using machine learning for these types of projects is super interesting. About four or five years ago I wrote, “Getting out of the Injury Zone,” and kind of replicated Josh Kalk’s work using machine learning to use PITCHf/x predictors for current injuries. It was fairly interesting because you could see that total movement loss is the most predictive thing of injury in-game, and velocity was the second biggest thing. But using this machine learning program, you could fairly well classify the high-signal guys. That was interesting, and to be able to use that type of programming method on a much richer set of data is going to be very cool.
I think it is going to be a small-but-significant outcome. I don’t think it is going to solve Tommy John surgery, by any means. What it is going to do is allow us to be much more proactive and productive so that our trainers aren’t constantly looking back and trying to figure out why people are getting hurt. They can have faith that an automated process is doing that every single day for them.
SB: I wonder what is most statistically significant whenever you set that up —
KB: — It’s only going to be limited to the inputs, right? It’s going to be limited to what we can collect, so that is why we spend so much on technology. The motion capture lab is about a $50,000 expenditure, but it is necessary. We need to know joint kinetics; we need to know kinematics. There is just no option. We have to do it, so we’re going to do it. We need force plates; we need to know grav-reaction forces. These are things that need to be calculated. We need to probably spend money on sleep tracking, readiness, heart-rate variability. These are things that are probably non-negotiable, that we have to know. As a result, we have to spend money on that stuff. That’s OK.
Once we can get that kind of information in the database, I think we’re going to find some interesting stuff. It starts with getting our hands around a good process for collecting all that data, and we have that. What we don’t have is the advanced tracking software, we need that.
Step one was getting the right people to prove that using something like Google Sheets, that people will actually do it. We’ve proven that people will, and now we need to give them the better hammer. Giving them the best hammer when they don’t know how to use it is only going to lead to stubbed toes. It is not going to lead to anything good. So, we’ve proven that they want for new tools and they can use new tools, now it is time to give them the new tools.
SB: When I learned that — in medicine — a 0.05 R-squared is good, it blew my mind that such a small explanation of variance could actually be a good thing. So, I do not envy the job of trying to pin-down all the factors, because I imagine that’ll be similar to what you’re trying to do.
KB: That’s actually exactly what I meant. I hadn’t thought of a specific explanation of variance, to use regression as an example, but I think that would probably be about right. Anywhere between about 0.03 and 0.2 would be the range that I would expect. Machine learning is a little bit different, because you don’t get an R-squared value. If you’re using one output, you get basically a number between 0 and 1—just like regression.
But the output is, in theory, much more projective. You can use it as a predictor, whereas R-squared is really a looking back. R-squared says, “this explains X percent of the variance of the model.” Machine learning aims to be, “this predicts 13 percent of theses classification so that, when I see these tests, I know I can apply this model and it explains 13 percent of the signal.” That’s how machine learning works. They’re functionally the same when it comes to application, but when it comes to description it’s a little bit different. That’s why it is a really interesting tool. It is why regression is the wrong tool for what we want to do.
SB: In a way, you want to be predictive rather than descriptive?
KB: I’d like to be. And I don’t know how realistic that is, I have to be honest with you. I think it can be done, but I don’t know. It is one of those things you don’t get to know until you spend a ton of money and a ton of time, and then find out that it’s no good. In a lot of ways, that’s just science. So, it’s OK. That’s just part of the job.
SB: The story you told of ALTIS is interesting. Is it a goal to be more proactive than reactive?
KB: I would say that’s fair. The way that ALTIS operates, in theory, they claim to be able to identify what the issues are and work them out during the training session. We would like to do the same thing. If motion capture ever becomes so minimally invasive that we don’t need to put retroactive reflectors on the person, and we can actually get good values of angular velocities without a marker set — that just doesn’t exist yet. People say marker-less technology exists, but I can tell you for a fact that it does not exist. Ones that are useful enough, anyways.
If that ever becomes a thing, pitchers could throw warm-up pitches in the bullpen and we could say, ‘Hey, these are the values that are way out of whack. Let’s see if we can knock them out before you go into the game using some manual therapy and some cueing’. That’s the ultimate goal. If we can do that, we’re in really good shape. But that is a long time coming, because we don’t even know yet if we will find something that significant. Then, if we do, we need to be able to monitor it and deploy the intervention correctly. We’re talking about three levels beyond what we’re currently at. If we get that done in 15 years, we’ll be doing it pretty fast.
. . .
Shawn Brody is a contributor for Beyond the Box Score, producer of In Play, Pod(cast), and pitcher recovering from Tommy John at Howard Payne University. He is a Senior double majoring in Business Management and Computer Information Systems. You can follow him on Twitter @ShawnBrody or email him at Shawnbrody9@gmail.com