The reality of sports injury prediction: lots of effort with little reward

In parallel with collecting all this data on training and training response, teams would also accurately record the number of injuries that players sustain. Then, at the end of a specified period (usually a season), data scientists at the club would search for associations between risk factors (or combinations of risk factors) and injury.

At this point, data scientists might get excited about their findings and declare, “We have found that when players exceed 5 273 m of high-speed running in a week, they are at an increased risk of injury”. Unfortunately for the excitable data scientist, it is too early to confidently make this kind of statement as the injury prediction model needs to be tested in other cohorts (see figure 2).

Dangers of discretization

Discretization is the process of transforming data from continuous variables into discrete categories. Scientists do this by creating a set of “bins” that describe the measured data. Discretization creates cut-off points that we use to categorize our data so that it can be easily interpreted and inserted into models. In data science, these categorizations are done by comparing different cut-off points and determining which one maximizes the number of true positive and true negative predictions.

 

As an example, we might be interested in the relationship between people’s standing height and their likelihood of playing in the NBA. We would collect all the heights of the people in our cohort and then determine a cut-off value that best predicts who will be drafted into the NBA. We might find that being above 6ft. 10in tall is a good predictor for selection.

The problem with discretization is that most continuous variables are normally distributed (as represented by the bell curves above). This means that while, on average, NBA players are taller than the normal population, some players within the NBA fall within the regular population curve. Similarly, some very tall people in the world never play in the NBA. Using a single discrete cut-off value always causes this problem and never results in a perfect prediction.

         2. Validate the injury prediction model in a ‘different’ cohort

Once a sport or data scientist has developed a model that they believe can be used to predict injury in a population, the next step in the validation process is to test the model in a new group. For injury prediction models to be useful, they should work in all settings, not just the one in which they were developed. This validation in different cohorts has proven to be difficult for most injury prediction researchers.

For example, in 2013, Australian researchers investigated the association between eccentric hamstring strength and a hamstring injury in 186 elite Aussie Rules footballers⁽⁷⁾. They determined that a cut-off point of 256N of eccentric hamstring strength was the best predictor of injury risk, with athletes falling below this cut-off being at increased risk of injury. The same research group repeated their experiment with a different player group two years later and could not predict injury with any consistency⁽⁸⁾. Very few research groups have reached the point of testing their injury prediction models in new cohorts. Prediction models perform better during development than during implementation, so practitioners should maintain a healthy skepticism around new prediction models⁽⁹⁾.

        3. Randomized control trial to demonstrate an effect

The final step in developing and validating an injury prediction model is to demonstrate that by using data from the model, practitioners or researchers can reduce injury. In order to achieve this, research has to have two groups – an experimental group that receives treatment or intervention based on the model prediction and a control group that receives no intervention or a placebo. Using the example of the Aussie Rules footballers above, research like this would have to demonstrate that a strength training intervention that increases players’ eccentric hamstring strength reduces injury compared to a group of players who didn’t do a strength training intervention. Randomized control trials are notoriously difficult to conduct, and no current research groups have developed their injury prediction models to this point.

Current status of injury prediction

Despite the promise and excitement around the potential to use machine learning to predict injury in sports, we are still in the very early days of application. Recently experts conducted a systematic review to gather and evaluate all the published research in sports injury prediction⁽¹⁰⁾. After an extensive evaluation, they concluded: “that there are no existing musculoskeletal injury prediction models that could be confidently recommended for use in sports medicine practice” ⁽¹⁰⁾. While the idea of injury prediction in sports remains appealing, significant growth and development are still required before this will be an accurate and reliable science.

Ethics of data collection for injury prediction

 

Modern athletes are among the most closely monitored individuals in history. Their sporting performances are subject to constant public scrutiny, and in many cases, their private lives are too.

 

Many sports scientists would argue that the more data they collect on athletes, the better their models predict injury. As a result, sporting organizations’ demands for athlete data are rapidly increasing, but where will it stop? In some organizations, athletes would be expected to:

• Wear GPS and HR monitors during training and matches
• Report loads lifted in the gym
• Report perception of effort after each session
• Submit to daily weigh-ins
• Report sleep quality and duration
• Report mood state
• Report the stage of their menstrual cycle
• Measure their hydration status daily
• Complete daily fatigue monitoring protocols

The collection and reporting of this data are invasive. Modern data protection laws argue that the collection of personal data should be with good reason. If the reason for subjecting athletes to such scrutiny is to fuel injury prediction models that are not yet fit for purpose, it may be time to rethink this approach.

References

1. BMJ Open Sport Exerc Med. 2020 May 20;6(1):e000675.
2. Br J Sports Med. 2017 Aug;51(16):1209-1214.
3. Br J Sports Med. 2016 Nov;50(21):1309-1314.
4. Br J Sports Med. 2016 Jul;50(13):776-80.
5. Aspetar Sports Medicine Journal 6 (2017): 18-24
6. Br J Sports Med. 2016 Jul;50(13):776-80
7. Med Sci Sports Exerc. 2015 Apr;47(4):857-65.
8. Med Sci Sports Exerc. 2018 May;50(5):906-914.
9. Stat Med. 2000 Feb 29;19(4):453-73.
10. Sports Med. 2022 Oct;52(10):2469-2482.