Sports Science Quarterly – Q1 2023

Every quarter we take a deep dive into the latest research in sports science. In this edition we look at the latest research on what artificial intelligence means for elite sport, optimizing practice environments, velocity based training, parkour for athletic development, and much more.

As always, the full Sports Science Quarterly is available exclusively to HMMR Plus Members. You can browse the past topics on our archive page. The first topic below is free to everyone, but sign up now to read about all the latest research. To get an idea of what Sports Science Quarterly is all about, the April 2016 edition is available in its entirety for free.

This Month’s Topics

Artificial intelligence in elite sports

Quick Summary – Artificial intelligence is often spoken about, but not always understood. This article provides a summary of the current uses of AI in sport, which include aspects such as machine perception and learning, as well as how AI is used in practice by sports teams and governing bodies.

Artificial Intelligence (AI) is a key buzzword across many industries. AI is defined as the behavior of a machine that would be considered intelligent if exhibited by humans, and, although it has existed as a computer science sub-field since the 1950s, it is only since the end of the 1990s that it has started to become somewhat usable in the real world. This is down to a variety of factors; these include the invention and further development of machine learning and deep learning methods, along with a substantial increase in computing power and availability. AI typically utilizes a model termed the Sense-Model-Plan-Act (SMPA) loop. This means that AI first perceives what is happening, creates a deliberate plan, formulates an action, and then provides feedback to the original perception, providing an update.

Given the increased interest in the use of AI in elite sport, it’s worth understanding how it might be used – the topic of a recent paper from Frontiers in Sports and Active Living. There are four key ways AI is typically used in elite sport, detailed below:

  • Machine perception – This includes aspects such as image recognition and computer vision, which allow for rapid analysis from video data. Also included here are wearables, which provide non-visual sensor data, which can be useful for athlete tracking or exercise prescription.
  • Machine learning and modeling – This encompasses the use of models that allow machines to learn patterns based on large amounts of data. To do this, machines are typically “taught” how to determine patterns by being exposed to large amounts of data. However, not all sports have large volumes of data available; in this instance, it is possible to utilize transfer learning, in which a machine learning model is trained on data from one sport, but then transferred to a different sport. A key issue with machine learning is that of transparency; often, we can’t understand why a machine is analyzing data in a certain way, because it can’t speak to us. This can, in turn, lead to systematic biases and poor decision making; the risk of this is particularly high when “black box” AI models are used. A potential solution to this is the growing field of explainable AI, which focuses on transparency and traceability of decision making in AI models.
  • Planning and optimization – As we get more data, and better at understanding this data, we should—in theory at least—be able to begin to model the effect of different training plans, allowing us to select the best one.
  • Interaction and intervention – Extrinsic feedback is the type of information we get from outside the body; for example, from a sports coach or a video of a performance. This type of feedback can be useful because it allows the athlete to match their internal feelings of a movement with what it looks like on video, allowing them to make subtle adjustments. Various wearable technologies can assist in providing extrinsic feedback, such as wearable sensors.

In the second part of their paper, the authors conducted a literature review of the current usage of AI within elite sport. In doing this, they focused on papers published since 2010, given the rapid changes in technology during this time; 136 papers met their inclusion criteria for analysis. They also conducted nine interviews with experts from seven different countries; these experts were either employed at a sports institute, or in the sports science department at a university.

The research papers were primarily centered around four key uses of AI in elite sport; image processing, signal processing, modeling and planning, and user interaction. The first three of these were most common. Typically, image processing (e.g., computer vision) is considered to fall within the realm of computer science, whilst signal processing (e.g., sensors) falls under the purview of sports scientists. On the modeling and planning side, organizations such as FIFA (the international governing body for soccer) often utilize AI in this way; as an example, FIFA created a ghost team to model the reaction of teams to tactical changes. These findings were also mirrored in the expert interviews, with projects in signal and image processing the most common.

The experts were also asked what a successful AI project within elite sport should look like. They recommended an interdisciplinary approach, with members of computer science, sports science, and sports practice (e.g., coaching) teams all present. They recommended that the sports scientists involved have a good understanding of both the practical aspect of the sport (i.e., sports practice), as well as having some knowledge and experience within the realm of computer science—something which represents a potential area of upskilling for many! All the experts interviewed believed that AI would become increasingly commonplace within elite sport over the coming years, although with a shift away from signal and image processing towards modeling and planning.

The authors also identified some challenges facing the use of AI in sport over the coming years. These were:

  • The lack of available data to create learning algorithms and to train models
  • A lack of interest from AI researchers and developers in working within sport
  • A lack of sport-driven interest in AI, potentially due to a lack of knowledge
  • Difficulty in explaining how AI works to lay users
  • Challenges in ensuring created predictive models are actually predictive, rather than explanatory (specifically, models may tell us what has happened before and why, but may not be able to successfully predict the next event).

Overall, this paper is a really interesting overview of the use of AI in elite sport—which is timely given that it’s currently a buzzword—along with some of the implementation challenges we might see over the coming years. The paper itself is open access, and well worth checking out yourself if you have a deeper interest in the subject – the references alone are probably well worth it!

Join Now to Keep Reading

This is just the beginning. To continue reading this article you must be a HMMR Plus member. Join now to get access to this and more content. Learn more and sign up here. If you are already a member, log in here.