Artificial intelligence in cycling coaching

Artificial Intelligence will help Cycling Coaching

Is wearable AI technology likely to solve all of our problems when it comes to improving our cycling performance?

I want to get this out there before I start. I love technology and data, and using both of these things to improve my clients’ training is a big part of my job as a cycling performance coach. In my previous life in financial markets trading, we harnessed the growth of cheap computing power to build a business that used algorithms to trade thousands of financial markets in real-time.

Sensors, data gathering devices, wearables, analysis applications and now adaptive training models are tools that have produced or are trying to leverage the exponential growth in training data in recent years.

We can now look at training in ways we never could before, but what we see should also come with some health warnings. Any coach worth their fee should be aiming to improve training methods, reduce injury risk and improve their athletes’ performance.

In recent years, a coach's job is also one of a data scientist. I have so much data at my fingertips that most of my coaching time is now spent looking at that data, trying to find patterns and individualising my clients’ training based on all the physiological data I collect every single day.

Artificial Intelligence (or AI) is simply advanced computer science that takes a bunch of data, simulates it and generates a set of outputs - in our case metrics and training. It is doing what I do manually and making it automated. The hope is that the systematic process makes better decisions than a human coach alone.

The problem with AI analysis in cycle coaching

However AI has a major problem - underspecification - observed effects can have many possible causes. With so much data, it is easy for the computer program to find an answer that fits some of the data rather than the correct answer for training purposes. In other words, the processes used to build many AI models today cannot be sure that they will work in the real world.

When we designed trading systems we had a human-generated hypothesis that came first. Next, we used computers to test our hypothesis on historic data to find out whether our idea had any merit.

The development in training intelligence has been very much data-driven. We collected all this data without really having a hypothesis about what we were looking to find. We have a lot of data but do people and companies really know what to do with it, and are they finding solutions that will actually benefit the athlete in the real world?

More data means more opportunities to misinterpret it.
Rob Wakefield
Founder and Coach

It’s a bit like cutting carbohydrates from your diet and then witnessing weight loss. You then join the dots and attribute weight loss to not eating carbs when in actual fact you lost weight because you were in a calorie deficit or lost a lot of retained water rather than fat.

Another example would be FTP(ego) where an athlete’s real FTP is 220w but they find data and maths to back up their hope that it is actually 240w. They try to train to 240w, get really tired, fail key sessions and never adapt. Many of the new gadgets and systems do, in our view, predict FTP that is too high (and massages the athlete's ego) but results in training that is too hard. But hey everyone loves a higher FTP for bragging rites so people naturally like things that tell them it's higher.

When you have a lot of data you can cherry-pick the things you want to see and that is selection bias. So lots of training programs are built using athletes' best performances at given time frames. But we all know that day to day you are not always producing your best numbers, and nor do you want to be. The art of building training is to use a much more holistic view of an athletes’ performance to progress their training at an appropriate rate.

We know that to use data to be prescriptive we need lots of it and we need it to be accurate. Even with power meters becoming better quality I still spend time every week cleaning bad data spikes that have an effect on athletes’ metrics and hence training going forwards. Most of this could be avoided by calibrating the power meter before every ride!

Additionally, the data you use to prescribe training has to be current across your power curve and that means performance testing, which left to their own devices athletes generally don’t like doing. So for an automated training plan to be truly adaptive, it needs recent data on your all-around performance and not just some peak performances from deep history and definitely not just your most recent best twenty-minute effort!

As power training guru Andy Coggan says there is no greater prediction of performance than performance. We would 100% agree with that but we also realise there is so much variability around what generates performance, so many variables, that to be honest, the data is nowhere near enough to be able to predict this accurately.

How do our coaches view wearable technology for cycling performance?

Wearable technologies are trying to answer the problem but we are still some way off being able to rely on them. Much of the marketing surrounding some of the new products smells of snake oil in our opinion, and as ever, many of the claims are not rooted in robust testing. There is a lot of money to be made from gadgets and subscription software, users beware.

The number of times I have had this conversation:

Athlete: My wearable tells me that I am not rested enough for a hard session today.

Me: How do you feel?

Athlete: Fine. I had a good sleep and ate well last night and I am raring to go.

Me: OK let’s do the session then!

Athlete: Smashed it.

The truth is that for most of us, being self-aware and checking in on how we feel is better than being blinded with data. I know just from walking up the stairs if I am recovered or not based on how my legs feel and my heart rate reacts! I can see that wearable technologies might have some merit at the margins for elite athletes who are pushing the boundaries of their physiological capabilities, but most recreational athletes are only maybe 70-80% near their absolute best.

For me, the really interesting thing that we could get to is using all of this data to predict the outcomes of training more effectively and reduce non-effective training time. I work with a lot of very busy people who have limited time to train and don't want to waste any of it. The truth is that historically coaching has been about experimentation and trial and error. Sure, over time you get to understand what makes an athlete perform and this is one reason why we invest so much time keeping our athletes happy so that they are with us for the long term.

Analysis software such as WK05 that we use has revolutionised how we understand the individual physiology of all of our clients and has made us better and more effective coaches. Things like optimised intervals are great at really targeting the points on the power curve that are hopefully going to elicit the biggest gains in performance.

Artificial Intelligence and by extension machine learning will provide coaches with dynamic tools for improving coaching. But coaches need to invest time in learning. Technology will give us new insights and guide us to make better decisions. Coaches that don’t invest that time in learning and just paste workouts into Training Peaks will get found out because the AI systems will do a better job.

Importantly we believe that cycling coaching is not all about the training plan, maybe 25% of it…

As cycling coaches, we are making sure that we take into consideration all of our clients’ life stresses and strains and making sure that these are incorporated into the training plan. Yes, the wearables are aiming to solve these problems but this is some way off in our opinion. We are also making sure that we offer the highest quality in client service, being available pretty much 24/7 to answer questions and adapt training plans in real-time if that is necessary, which it often is.

A good cycling coach will be providing regular analysis and feedback with screenshots and video calls. They will be giving praise, and sometimes a proverbial kick up the backside. One of the most common reasons people give for using a cycling coach or a personal trainer is accountability. I just don't think that can be replicated by a robot, at least not yet!

I am really excited by the continued development of technology and computer science and am confident that it can and will help me become a better coach and allow me to make you faster more effectively.

For those that don’t want or can't afford a coach, then the new generation of adaptive training plans on paper look to be a better solution than simple off the shelf training plans. At a basic level if they make you feel like you are being given a more individualised coaching plan then you are probably more likely to complete your training. And as we all know consistency is the number one most important factor in becoming a better rider. Whatever makes you more consistent is a good thing!

To find out more about how Propello Cycle Coaching can help you achieve your goals in the time you have available, get in touch with us.