Can an AI coach see an injury coming?

Your achilles has been stiff in the mornings for a week. The sessions are going well, the pace is there, and nothing in the watch is complaining. Then the question: should a coach that reads all my data have seen this coming?

No. Nobody can predict an injury from training data, and an app that promises to is promising more than the research does. What a coach can see is the ramp: that you are running more or faster than the base you have built can keep up with. That is where many overuse injuries start, and it can be read in the numbers. Coach Fartlek is our own app, so expect bias.

What the data shows: the rate of increase

The rule of never increasing weekly volume by more than ten percent is older than the research on it, and when it was tested on novice runners (Buist et al., 2008) just as many got injured with it as without. What the research has centred on for the last ten years is instead the ratio between the latest week and the latest four. Gabbett (2016) and Hulin et al. (2016) compared seven days of load against 28 and found that weeks more than 30 percent above the base were followed by more injuries, and weeks more than 50 percent above by more still.

That sounds like a forecast, and it is exactly what it is not. Impellizzeri et al. (2020) went through the evidence and showed that the ratio is a poor injury predictor, partly for plain statistical reasons. What it does show is that your volume is moving faster than your base. It is a measure of rate of increase, not of what your body can take.

That is how we use it too. In the app's load view, your 28-day base times 1.3 is drawn as a tolerance. Below it the week is green, between 1.3 and 1.5 times the base it is amber, and above 1.5 it is red. There is no lower limit, because a quiet week is just a quiet week. The line is set against your own base, not against a number that applies to everyone, and the info text in the view calls it a ramp guard, not an injury forecast.

Steps, not just load

The second thing that can be read is what the load is made of. Common load metrics like TSS measure how hard you worked, from heart rate and time, and they measure that well. What they miss is the impacts. You can hold exactly the same fitness week after week and still take a quarter more steps, if the sessions have become longer and slower, and steps are what achilles tendons and shins count.

So the app also computes a mechanical load: the number of steps from cadence and time, with each step weighted by how fast you were running. Nilsson and Thorstensson measured back in 1989 that peak vertical force rises from roughly 2.0 to 2.9 body weights between 1.5 and 6.0 metres per second, so a step at interval pace hits the ground harder than a step on a jog. The same tolerance, 1.3 times the 28-day base, is drawn around that series.

The graph needs the watch to measure cadence, and it needs a couple of weeks of running behind it before it says anything. A base built on three days is not a base.

What the watch does not measure

Pain. That is the most important limitation, and it does not go away with better models. The watch sees that you ran 14 km at 5:10 with a heart rate of 148. It does not see that the achilles was stiff for the first two kilometres or that you changed shoes last week. Elevation is measured, but surface is not. What the hills did to the tendon only you know.

A coach that only reads measurements therefore guesses, and guesses in the same confident tone whether or not it is right. That is built into how the coach writes: the data shows what happened, never how it felt, and a verdict that hinges on the feeling is kept conditional until you have answered.

So the only thing that makes the question in the title answerable is that you tell it. Coach Fartlek has a health log where you enter a niggle as mild, moderate or severe. From that day it is part of everything the coach does. The reminder before a session gets a concrete cue on protecting the area, and the analysis afterwards asks the question when the session spoke to the niggle, hills or speed on a sore tendon or an aborted run. The weekly planning takes it into account as well. If you have been quiet about it for two days, the coach asks how it feels, at most twice if nothing has changed.

What the coach does with the signals

It never changes the plan on the ratio alone. An elevated ratio right after a recovery week, a taper or a break is expected and no reason to scale down. Only when the ratio lines up with something else, like a reported effort or a niggle, does it become a reason to act.

Most of the protection is in how the weeks are built, not in alarms. A new training stress delivers its effect over six to eight weeks, so the coach holds a level until the effect has been collected before raising it. Only one thing changes at a time: volume, intensity, recovery or frequency, never several at once. The base is taken as the median over twelve to eighteen weeks, not the last four, since four weeks is a fresh or tired moment and not a level. And if the coach learns something about you, such as two quality days back to back tending to end badly, it is saved as a rule that comes along into every new week.

This is the boring part of coaching. A good human coach does exactly the same, and the difference is that a machine reads every session and does not tire of saying the same thing a third time.

Where the coach stops and the physio starts

The coach makes no diagnosis, and it should not. There are situations where it says outright that you should see someone: an injury you rate as severe or that changes how you run, pain that has not improved after about three weeks of managed load, a fever, or chest pain and dizziness during exertion. Then the area is unloaded until it has been assessed, and the coach drops the subject once you have said you are handling it.

Ordinary niggles and muscle soreness are not such situations, and neither is a bad day. A coach that sends you to the physio for every bit of stiffness is as useless as one that never does.

How to use it

Log the niggle the day it shows up, not the day it has become an injury. The question about how the session felt is worth answering, because it is the only way in for what the watch does not see. After a build week the load view is worth a look, especially if you have changed surface or added speed.

The app needs a watch connected through Intervals.icu, which pulls from Garmin, Coros, Polar, Suunto and most others, and it covers 5K to the marathon, not ultra and not multisport. How the coach reads HRV and resting heart rate, which is the other half of the same question, is in the running app that adapts to your recovery. If you are not getting faster despite running more, it is often the same ramp seen from the other side, see why am I not getting faster. What the tiers cost is on the pricing page.