Thread by @Robertson_SJ

Injury prediction is a waste of time.

This area of inquiry has become dangerously pervasive and it damages sport science's reputation. Much misinformation & misinterpretation exists.

Below I've started a list of the many problems with the area - please feel free to add to it.

1.True positives – whereby an injury is predicted but the athlete is allowed to train regardless - is the only way we can really know the quality of the prediction. Which of course would present an unethical scenario.

2.Thus decisions made to remove an athlete from training cannot be validated. This is because it cannot be determined what would have occurred if there was no intervention. Thus the most conservative practitioners tend to thrive, as contrary evidence cannot be provided

3. How many false positives-which still result in a player being removed from training-are required before taking the ‘risk’ was a better strategy. If a player misses ‘x’ training sessions to prevent an injury that never eventuates, they may end up actually missing more training

4. Even a bad model will be right most of the time. This is because most of the time athletes train or compete, they don’t experience injury. Simply predicting ‘No’ by a model means it is correct 99% of the time, thus inflating its performance

5. Injury prediction models just aren’t implementable in practice. Data isn’t analysed, nor are models run - in near real-time. Models run pre- or post- session or competition cannot account for what happens during sessions – which is of course when the actual injuries occur.

6. Humans don’t interpret and operationalise predictions well. Risk is expressed continuously, but deciding whether an athlete trains or not is a binary decision. How well do humans operationalise risk and likelihoods? Uncertainty is inherent in all predictions

7. The data quality used in models isn’t good enough-yet. Why would running distances measured from between an athlete’s trunk directly relate to the onset of an acute hamstring incident? Improvements to tech may change this in future (i.e., intramuscular sensors etc)

8. The data volume isn’t yet high enough. What else requires measurement outside of the training and competition environment that a) could be important and b) can be measured appropriately? Without acknowledging this, any models is destined to suffer availability bias

9. For some injuries, the data volume and data quality will never be high enough as we don’t yet enough about how they emerge.

10. Most of the research in the area to data employs linear models. But if relationships do exist between currently collected variables and injury, then they’re likely be a) non-linear and b) complex.

11. Most of the research that actually has used non-linear methods, hasn’t accounted for repeated measures in the analysis.

12. Of most of the machine learning methods used (i.e., random forest and ANN’s), the exact contribution of a variable to the prediction can’t be quantified. Variable importance plots are uninterpretable with respect to informing preventive strategies required by practitioners

13. Model outputs should be compared with baseline prior risk – which is of course not zero. Theoretically, it’s not possible for it to every be zero. What is the realistic floor effect? Is it sport dependent?

14. Will there ever realistically be access to large enough data sets for cross validation. Will models generalise from one environment to another?