Jumbo-Visma Nutrition at the Tour – Machine Learning

By Jiri Kaloc

How does machine learning help professional cyclists? Many teams at the Tour de France are taking the nutrition of their riders seriously. One of the things that set Jumbo-Visma apart is the technology they use to better and adjust the intake of their riders.

There is a lot that goes into the calculation

Calculating exactly what kind of ingredients and how many calories each rider needs is no easy feat. In the past, team nutritionists would start preparing their calorie estimates weeks or even a month before the start of the Tour de France. They would look at stage profiles and cyclists’ weight and body composition to begin with. But when the racing starts, there would always be unforeseen factors. The weather, changing the team’s tactics, and many other things would require a quick revaluation of intakes.

Asker Jeukendrup, head of nutrition at Jumbo-Visma, said in interviews he would have to create spreadsheets for each rider to calculate their optimal intake and convert it into meals.

“It was very time consuming and I could only do the calculations for 2 riders at a time. Slowly, we upscaled, but when the Jumbo Foodcoach app started to do the calculations for us and the translations into meals, this is when it became scalable and faster. Jumbo has been a great partner in this and we continue to improve the app.”

Collecting data

Upscaling and creating an app that would help automate the process was a big step forward for the team. For this to work effectively, they needed to ensure that the app would have all the necessary data to create accurate predictions. Here is what they included.

  • A Garmin device on the bike of each rider that shows actual route data on total distance, meters climbed, etc.
  • A crank-based power meter supplying a precise calculation of the calories burned.
  • Riders’ individual weight, height, and role (sprinter, climber, etc.).
  • Weather forecast in combination with GPS location data for each rider to help calculate the impact of weather (tail or headwind, etc.).

The gathering and visualisation of this data are powered by Smartbase, a data-management and analytics platform. Coaches at Jumbo-Visma would use this platform to enter the actuals.

Supervised learning

The team also needs to clean this data by removing mistakes. For example, if a cyclist forgets to turn off their Garmin device at the end of a stage, they need to exclude the post-stage part of the data. They also make certain variables relative, such as power, energy, and elevation, so that it’s easy to compare stages and races. All of this data is then used to prepare forecasts.

Using training examples, Jumbo-Visma applied supervised learning to teach their algorithm what the outcome of the calorie predictions should be. This was a regression problem, and they selected random forecast as the best machine learning algorithm to solve it.

Machine learning is way more accurate

All of this sounds great but it doesn’t count for much unless the predictions are at least as accurate as when human nutritionists make them manually. They used R-squared to evaluate both their machine-learning model and manual predictions. R-squared measures the strength of the relationship between the model and calories on a 0-100% scale. The machine-learning model got a score of 82% while the manual predictions only got 52%. This is truly amazing! Not only is the algorithm more accurate but nutritionists also get the results in a split second, so they have more time to respond to those unforeseen factors that come up when racing.

Jumbo Foodcoach app translates it into meals

The Jumbo Foodcoach app connects the dots. When a team nutritionist inputs the calorie numbers from the machine-learning model into the app, it provides sample meals with optimized proportions for each meal. It can’t get much easier than that. That’s one of the big reasons why the team keeps winning. You can check out their app in the App Store or on Google Play