The Automatic Hobbyjogger Detection Machine
Using machine learning to recognize who’s a serious competitive runner and who’s not could teach us something useful about avoiding injuries
Heading out the door? Read this article on the new Outside+ app available now on iOS devices for members! Download the app.
One of the most enduring obsessions of the famously fractious Letsrun.com message boards is how and where you draw the line between serious competitive runners and mere recreational hobbyjoggers. The answer usually boils down to something along the lines of “Anyone faster than me is a talented and hardworking athletic colossus bestriding the world, and anyone slower than me is a pathetic hobbyjogger who shouldn’t be allowed to buy running shoes.”
This sort of definition somehow never manages to settle the debate, so I’m excited to report that scientists have created a machine that can watch you run and immediately classify you as either a “competitive” or “recreational” runner. This is not as silly or elitist as it sounds—in fact, it has the potential to help bring a more nuanced approach to assessing injury risk based on subtle details in your running form. The research comes from a well-respected biomechanics group at the University of Calgary headed by Reed Ferber, the director of the university’s Running Injury Clinic, and is published in the Journal of Sports Sciences.
The basic goal of the study was to stick a wearable accelerometer on the lower back of 41 runners (they used an accelerometer called the Shimmer3) and see if it could deduce which runners were competitive versus recreational using machine learning. They defined competitive as anyone who had a recent race performance between 5K and marathon that exceeded 60 percent of the age-graded world record for that distance based on World Masters Association Age Grading Performance Tables, a threshold that USA Track and Field defines as “local class.” By this definition, 17 of the runners were considered competitive, while 24 were considered recreational.
The three-dimensional stride data collected by the accelerometer generated 24 distinct characteristics of each runner’s stride. These weren’t the usual things like cadence and stride length, since those factors are heavily influenced by how fast you’re running—which, as any wizened masters competitor knows, is not always a good barometer of how competitive you are. Instead, the focus was on more subtle features related to stride variability (e.g. how much does your stride length change from one step to the next?) and regularity (e.g. how similar is your body’s instantaneous acceleration in each of the three dimensions throughout successive steps).
The differences between the two groups of runners are less obvious to the naked eye than you might imagine. If you stick to conventional stride parameters, you don’t see anything at all: female competitive runners, for example, had an average cadence of 168.2; their recreational counterparts had a nearly identical average of 169.1. Even with the more sophisticated measures of stride consistency, the differences aren’t obvious. So the researchers fed all the data into a machine learning system called a support vector machine, and let the computer figure out which factors distinguished competitive and recreational runners. Importantly, they analyzed male and female runners separately, since the hallmarks of a “competitive” stride might be different in the two groups.
Sure enough, by using data on stride consistency, the computer was able to correctly classify male runners as competitive or recreational 82.6 percent of the time, and female runners 80.4 percent of the time. The specific factors that mattered most were different in the two groups—which isn’t surprising, lead author Christian Clermont explained in an email, because “the structural differences in male and female anatomy certainly affect the way we run.” The men’s model incorporated 12 different stride features, while the women’s model incorporated 10 different features, all related to stride variability and regularity.
The advantage of machine learning is that it can pick out subtle patterns in a large number of variables that you’d never find just by staring at the data. The disadvantage is that it’s not always obvious what those patterns mean. Why, for example, is the most important distinguishing feature for men the step-to-step correlation of center-of-mass acceleration along the back to front axis, while for women it’s the root-mean-square average of that acceleration? But if you step back from the details, you can see the bigger pattern: experienced runners run more consistently than less experienced runners, with every step more similar to the ones before and after it.
Why does this matter? While I’m loath to venture into Letsrun-style value judgments, there are reasons to believe that the competitive running gait is better than the recreational one. Studies have generally found that inexperienced runners get injured a lot more than experienced ones despite running less, and they tend to get injured in different places. Recreational runners tend to get more knee and hip injuries, perhaps due to unoptimized running form; competitive runners tend to get more foot and lower-leg injuries, perhaps from overuse related to heavier training loads. So knowing whether your running form is getting more “competitive” or more “recreational” might theoretically give you some hints about whether your training is working and where you might be most vulnerable to injury.
The accelerometer used in this particular study isn’t suited for off-the-shelf consumer use. Still, Clermont says, there are some useful parameters that could in principle be calculated using things like the Garmin Running Dynamics Pod or LumoRun (which sadly went bankrupt last month). Even with simpler smart watches or foot pods, you could measure how long each stride takes—and then, crucially, calculate a coefficient of variation, an indicator of how much that time varies from stride to stride. That would give you some sense of how consistent your stride is, whether it gets less consistent with fatigue, and whether it’s getting more consistent over time. Watching the trends could give you a sense of whether your training is helping or hurting you. If enough people ask for a feature like that, perhaps companies like Garmin will make it available. (And perhaps it’s already available somewhere: the wearable running tech world is so sprawling and fast-evolving that it’s hard to keep track.) I’ll suggest a name for this parameter: the Hobbyjogger Index.
My new book, Endure: Mind, Body, and the Curiously Elastic Limits of Human Performance, with a foreword by Malcolm Gladwell, is now available. For more, join me on Twitter and Facebook, and sign up for the Sweat Science email newsletter.