On Camera-Guided Workouts

Having a camera watch your workouts and an AI tell you all the things you're doing wrong is a popular concept for digital health apps. The claims sound great. By eliminating the need for a personal trainer, you democratize fitness. Also, instead of generic fitness advice you get personalized insights into your movement patterns that might be invisible, even to a trained eye.

In addition to small start-ups and university spin-offs, some of the big players like Peloton are also in the process of building their portfolio of camera-guided workout products. I've recently looked into quite a few companies in this space and have organized some of my thoughts below:  

 

The problem

When AI started to be applied in medicine and initial successes made the hype grow bigger and bigger, a lot of companies were founded with the goal of replacing the radiologist.

As time went on, reality kicked in and even the most optimistic tech enthusiasts started realizing that accomplishing this task might take a bit longer than they had expected.

Companies started pivoting towards more feasible business models which usually involved taking over some parts of the radiology workflow such as quality assurance or measuring defined structures.

I think a lot of the companies trying to work on camera-guided workouts are still in the "replacing the radiologist" stage. However, unlike in radiology where the regulatory requirements are very high, some companies in the workout space could actually succeed, even though the product may not bring any benefit to its users.

When you want an AI to rate a user's technique, you first need to tell the AI what good technique looks like. And this is where it gets tricky. Of course, all of the companies will claim that they work with doctors, professors, motion scientists, athletes or physiotherapists and this is probably true, but even the most experienced professional will require data to assess what actually constitutes good technique.

They might have seen people get injured doing certain movements, but since they rarely follow most of their clients for 40 years, it will be hard to assess the long term consequences of certain variations. In addition to the technique, there might also be additional factors at play when a person gets injured such as nutrition, sleep/recovery and pre-existing conditions that the professional may not be fully aware of.

Things become even more complicated due to differences in anthropometry. Even among professional athletes where only the most efficient techniques prevail, we often see a surprisingly high degree of variation.

And what about the data? As always in the lifestyle space, high-quality, long-term data is practically unavailable and most of the existing research has a high suspicion for the presence of confounders that are impossible to quantify. Even if we know that a certain variation causes more rotational force, stress or compression on a particular joint, it's not sufficient to predict wear and tear 30 years down the line. There are just too many things we don't understand about our bodies' adaptations to stimuli.  

 

The consequence

In the absence of a clear ground truth, a sensible approach would be to define a broad range of variations that could constitute a safe execution of the exercise and ask the users to also listen to their body.

Let's look at the squat as a popular example. It's probably fair to say that your legs shouldn't be completely X-shaped when performing the exercise. However, the exact positioning of your feet or how far your knees should come forward as you go down is much more difficult to answer. Remember when people in high school used to tell you to never have your knees move past your toes, yet in some culture it's normal to sit like this pretty much all the time?

So why do some companies still claim to know the one and only truth? Because accepting that the range of 'safe' squat variations may be quite broad and that video analysis can't replace experiencing how certain variations feel for your body would undermine the business model of personalized AI-based insights.  

 

The downside

Is there a downside as long as the gold standard that the AI is trained on is based on expert consensus? I think yes.

The obvious reason is that for some people whose anthropometry is very different from the average person or who have preexisting conditions, the AI gold standard may actually be more injury prone.

However, there may also be a disadvantage that applies to the majority of people: Excessively worrying about your technique instead of getting a few basics right and listening to your body (and asking for help if you can't figure things out) introduces noise and makes working out less pleasant. For a lot of people the motivation to keep working out comes from seeing results. And seeing results comes from training hard. However, when you're constantly worrying about wether your left knee is rotated outward by 3°, training hard becomes more difficult as your mind is busy thinking about your knee rotation.  

 

Alternative approaches

While it may be impossible to offer AI-based workout insights to the degree that many companies promise, there are a lot of fairly low-hanging fruits regarding camera-guided workouts:

• Automating workout logs: The AI tracks your exercises, how many repetitions you do and how long it takes you. It might even try to give better estimates of the amount of calories you burn.
• Assessing mobility: Training an AI to measure joint angles already works pretty well. This can be used for assessing the mobility of a user which could be useful in a telehealth/rehabilitation context.
• Increasing safety: Especially in the elderly or people with preexisting conditions, an AI could detect a fall or syncope during a workout.
• Suggesting changes: The AI can still make suggestions regarding your technique without phrasing them as absolute rules. Saying something like 'It seems that you rotate your knees internally compared to the average user. How about you lower the weight and try to keep your knees neutral and see if it feels better and more stable?'

TLDR: Camera-guided workouts will become increasingly prevalent. However, companies should not oversell the potential benefits and focus on implementations that will actually make working out more enjoyable.

If you found this article helpful, let me know! If you disagree with some of the things I said, let me know as well!