Sherlock Holmes, Heuristics and Building a Functional Milestone Return to Sport Program: Part 1
In a Nutshell
This is the first blog post in a five part series on the importance of using objective testing to support decision making throughout the return to health, return to sport and return to competition transition after injury. It also highlights my experience with a new technology called Plantiga that uses a branch of artificial intelligence (AI) called deep learning to power an intelligent insole system.
The focus of this series of posts has never been more relevant.
First, COVID is requiring all of us to find new ways to assess, monitor and train athletes remotely.
Second, if we take the NFL as a case-in-point, despite historically high resources going towards injury prevention and performance, the NFL along with many sport organizations continue to struggle with persistent and unchanging injury rates.
Why is this happening and how to we mitigate the cost of injury?
Whether we are talking about the NFL that had > 7 ACL injuries in a single weekend or the military, attending to the burden of injuries is of the highest priority.
Further, injuries are somewhat inevitable. Athletes will get hurt. This demands that we are doing more than just rolling the dice and trusting our instincts and our “coach’s eye” as suitable markers to determine when an athlete is prepared to get back to a high risk activity after injury.
This series of blog posts addresses these questions and issues.
Recap
In my last medium blog post, I discussed how much ado has been made over injury prediction, how it fits and whether or not it is even a feasible prospect in sport.
In this blog post we will talk about wearables and a branch of AI called deep learning that can help us extract all sorts of insights on how we move.
To make injury prediction a feasible endeavour, we might have to start treating it like the weather. This means we need to study the athlete in their environment like it is a complex-adaptive system (because it is).
If the word prediction scares you, angers you, or conjures skepticism, you can replace it with the word probability.
A statistical model is not a genie in the bottle.
It’s been said that all models are wrong but some models are useful. Indeed, the output of a model is a statistical probability and the “prediction” or assignment of probability is as good as the humans who are thinking about the problem and the quality of the data that feeds the model.
We also need to get as close as possible to the inciting event to capture relevant, time-sensitive data that can be used to reduce the uncertainty in a model that assigns a probability of injury to the athlete.
Just like the weather forecaster that predicts the chance of rain, injury prediction would assign a probability for injury or reinjury. Again, don’t equate prediction with a “genie in the bottle” …. although that would be nice sometimes, wouldn’t it?
Wearables Unlock the Potential to Treat Injuries like the Weather
Wearables allow us to obtain ubiquity with our data collection and if there is any hope of injury prediction, it will require us to get better biomechanical and functional data on the field of play, where it matters most.
This is not meant to downplay the importance of testing that we do in a laboratory or a training centre. In fact, this is where I’ve spent the bulk of my career testing injured athletes. But, my biggest blindspot is inevitably when an athlete exits the lab and gets back on the field of play.
Here’s where my experience using the Plantiga system can provide a novel solution.
Plantiga puts a high-grade inertial measurement unit (IMU) in an insole that can go in all types of footwear. Machine learning is then used to extract insights from the data that can improve decision making.
Plantiga’s machine learning algorithm is affectionately referred to as Norman in office discussions, named after Plantiga CEO Quin Sandler’s father who passed away a few years ago. The name is fitting because Norman and Quin conceived of Plantiga and built the company with a whole lot of passion and vision (that’s what impressed me the most when I first met them).
I think it’s fair to say that Norman is still looking over Quin’s shoulder to provide guidance and support, and now Norman the AI is looking underneath our feet in a similar way.
Just like a humans who are born almost entirely helpless and dependent on adults for survival, Norman the AI is learning to roll over, sit up, crawl, walk, run and maybe one day win an Olympic medal. These are called functional milestones and they mark a continuous pathway of development for humans across the lifespan.
As we will discuss below, functional milestones are also incredibly important when it comes to rehabilitation after orthopaedic and musculoskeletal (MSK) injuries, like anterior cruciate ligament (ACL) tears.
I concede that winning an Olympic medal isn’t on the pathway of functional milestones for most humans but it is a single example of the pinnacle of functional movement. And if you’re an Olympian and suffered an MSK injury, defending your Olympic medal might be a key milestone in returning to performance after a serious injury.
Sport represents an athlete’s passion, their livelihood, and in many cases, the essence of who they are as a person. Injuries rob athletes of their identity.
I also concede that Norman the AI isn’t trying to learn how to roll over and sit up per se. But Norman is learning how to read our movement patterns autonomously.
Norman is quite smart already.
Plantiga’s movement classification algorithm can automatically detect walking, running and jumping with acceptable accuracy.
A load carriage algorithm can estimate the mass of a backpack worn during a walk to within a few kilograms. Imagine the applications for firefighters and military personnel who have to carry heavy loads around, professions that carry with it a tremendous burden of injury (Heir & Glomsaker, 2007)
A key development in Plantiga’s AI algorithms is called gRIN or “ground interaction”. This automated event detection algorithm is essential for pulling out all sorts of important performance variables like ground contact time and limb speed during sprinting. There are huge performance applications here.
Accurate measurement of foot-ground-interaction with a wearable may sound like a trivial task. But, the shear amount of data and noise coming from a wearable makes it quite challenging.
This is one advantage of having an IMU embedded into an insole — it dramatically increases the quality of the signal compared to placing an IMU outside the shoe, on the shin or worse yet on the back where movement artefacts and other sources of noise abound.
Writing an “analytical algorithm” to detect when the foot is on and off the ground is challenging because there are so many different ways that humans move.
Human movement is variable.
Machine learning, on the other hand, means that the computer learns to detect this event.
Machine learning is the key to unlocking the full potential of wearables aimed at measuring human movement because the nuance, complexity and variability can be extracted.
gRIN can now provide incredible insights from walking and running including spatial and time-based (temporal) biomechanical parameters like stride length, stride frequency, deceleration capacity and ground contact time.
The team of computer scientists at Plantiga led by CTO Sean Ross-Ross are working hard on refining gRIN.
Additionally, gRIN has opened the door to new algorithms that can predict time from the occurrence of an acute injury based on a walk test. To this end, imagine identifying an athlete who is 12 months post-injury but is walking like someone who is only 6 months post-injury, suggesting that they are still compensating and requiring more rehabilitation.
How Machine Learning Can Help Us with Functional Assessments in the Real-World
Now, here’s where the magic happens.
Norman isn’t learning to walk, run, jump and decelerate in a literal sense. It is learning how to read our movement patterns when we are walking, running, jumping and decelerating.
Plantiga’s machine learning algorithms are learning to predict functional recovery after injury. In fact, in the image above, the three red movement maps depict an athlete with an Achilles tendon rupture recovering towards the typical pattern shown in green. Machine learning extracts insights from these movement patterns to predict recovery timelines based on functional outcomes.
This becomes hugely valuable when it comes to rehabilitation after MSK injuries because the gold-standard recommendation is that clinicians move away from a time-based approach to an approach that is based on an individual’s progression over an increasingly set of more demanding functional milestones (Jordan et al., 2020; Myer, Paterno, Ford, Quatman, & Hewett, 2006).
Despite this recommendation, many sports medicine practitioners still rely on a time-based approach to guide the return to health, sport & performance transition after serious MSK injuries like an ACL tear and in this example, the follow up surgery that is often required to restore knee joint stability (this is called an ACL reconstruction — ACLR) ((Barber-Westin & Noyes, 2011).
Consideration for time since injury is essential especially surrounding biological processes like tissue healing or the ligamentization that occurs after an ACL reconstruction whereby the tissue graft starts to become more like a ligament.
But timelines don’t necessarily reflect functional capacities and movement abilities after injury.
Timelines may lead us to the notion of a pre-set functional plan for determining when an athlete is ready to start walking, resume higher force activities like strength training, begin jumping and jogging, and eventually make their way back to sport. For anyone how has suffered an ACL injury, you might have heard the magic number of 9–10 months.
For sure timelines do have their place.
But could you imagine applying the same time-based criteria for a human baby as they develop in their first two years of life?
This would be like saying that all babies should be walking at 12 months so regardless of whether a baby can or can’t walk at 12 months, it is reasonable to progress all children onto more demanding functional tasks, like running.
As a parent of a child who did not walk until he was just under 18 months, the prospect of a blind and somewhat arbitrary time course for functional development just doesn’t work.
And it doesn’t work for return to sport decision making after injury either.
We know babies will learn to walk sometime in the first two years of life but there is huge variability in this time course.
It’s an n=1 game and it doesn’t matter what the average is. What matters is where you fit on the normal curve.
The same is true of an athlete coming back from an injury, like an ACL tear.
There is no preset timeline that can determine exactly when an athlete is ready to progress from one stage to the next. Again, it’s an n=1 game.
Here is where we have to measure and monitor the things that matter.
Making things even more challenging is that many of the field-based functional tests we use to evaluate athlete readiness for return to sport after MSK injuries like an ACL tear are either (a) assessed subjectively by the practitioner (I call this the coach’s or practitioner’s eye) or (b) are performance based like the single leg hop for distance, a test that athletes are able to “cheat” to achieve the benchmark while masking the deficits that really make them vulnerable to another injury (Grindem, Snyder-Mackler, Moksnes, Engebretsen, & Risberg, 2016).
Experienced practitioners know this narrative well.
The athlete passes the functional criteria at a single time point (e.g. 9 months) while masking deficits, gets the greenlight to return to a high risk activity and then becomes one of the statistically unfortunate many who go onto suffer a reinjury (Barber-westin & Noyes, 2020).
Worse yet, athletes rarely get assessed in a sport-specific or contextual manner and instead receive their green light for returning to a high-risk sport based only testing that happens in a clinic environment.
We need to do better. But how?
In the next post we will talk about how practitioners can learn from Sherlock Holmes, develop heuristics and work towards a more data-driven approach to manage the return to sport process after injury.
