Article Text

Download PDFPDF

Predicting future physical injury in sports: it's a complicated dynamic system
  1. Chad Cook
  1. Duke University, Physical Therapy, Durham, North Carolina, USA
  1. Correspondence to Chad Cook, Duke University, Physical Therapy, 2200 W Main St B230 Durham, NC 27705; USA; chad.cook{at}

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.

What does sport injury risk have to do with weather forecasting and predicting who will be the next President of the USA? Everything! The commonalities are that each occurs within a dynamic system. If you know how dynamic systems work, you will better predict sports injuries.

A dynamic system is an environment that depends on internal and external factors to attain stability.1 Organisms/participants/traits within a dynamic system adapt and change when factors within the system change. Scientists routinely predict risk in a variety of dynamic systems such as weather, political forecasting and projecting traffic fatalities. Ideally, the modelling for each forecasting model involves sophisticated statistical assessment, complex/multiple simulation studies, allowances of time-related changes and distinctly large sample sizes.

What about sports injuries? Predictive models used in dynamic systems identify predictors in a changing environment, in which outcomes are influenced by both the baseline characteristics (the predictors) and external environmental mediators (other factors, unaccounted for at baseline). Sports injuries occur within a dynamic system and our attempts at predicting injury have ignored the dynamic nature of the system (eg, changes in strength or accounting for factors such as physical contact injuries). We assumed that baseline characteristics such as strength, balance and flexibility would predict an injurious event, independent of their internally driven change and often independent of external factors. Think of when you hear ‘load’ included as an overuse injury risk factor! Definitions of load are variable among athletes and across sports and level of competition. Clearly, this is not optimal.

Limitations of the traditional model—assuming players are frozen in time

There are at least five major limitations of a non-dynamic model for predicting a sports injury.

  1. Injuries vary among levels of competition and different sports, and injuries change within the same sport with adjustments in equipment (ie, skiing), levels of athleticism (novice vs professional) and size of the participants (ie, American football).

  2. A non-dynamic model assumes a standardised single outcome (an injurious event), yet different definitions of an ‘injury’ are the norm2 ,3 and variations in the definition of injury have led to wide dissimilarities in incidence rates.4

  3. Most sport injury studies have very small sample sizes, which reduce the power and transferability of the findings within the model. The small sample sizes fail to adequately reflect the patterns present within the data (if at all) when larger amounts of data are available.

  4. External factors such as field type5 or compliance to training6 that contribute to a sports-related injury are often more influential than the baseline characteristics that are examined. Consider the most common injuries in American football such as concussion, fractures and lower extremity sprains, which result from player contact during games7 and are largely independent of many baseline ‘predictive’ measures that examine the physical state of an individual.

  5. Finally, and perhaps most importantly, the ‘predictors’ are often inadequate, too refined or too restrictive for optimal use within a dynamic model. For example, suggesting that someone must single-leg hop a set distance or hold a squat for a certain number of seconds fail to adjust for the variability among athletes and the demands of the specific sport. Dividing quantitative values (transforming the data) into categories improves the interpretability of the clinical finding but desensitises the data and, as with small sample sizes, can fail to identify patterns. It is better to consider the current ‘predictors’ as dimensions (rather than a tool with a dedicated and discriminative value) that reflect a construct that is needed within the given sport. Thus, strength and motor control during vigorous activities are needed but a certain level of strength is not a magic cut point to distinguish ‘at risk’ from ‘not at risk’. In fact, there are no magic cut points for any predictive measure.

So what should be done? Can we do better?

Using different statistical modelling strategies is one method to consider for improving our ability to reflect the dynamic nature of sports injuries. This process is beyond the scope of this editorial but may include Bayesian dynamic probability models, stochastic time-series modelling or others. Another simpler method that does not require complex statistical analyses involves reconsideration on how we use our assessment tools. To give context, we8 recently examined several constructs (eg, flexibility, strength and motor control) that have been identified as associated with future injury and found that the constructs of hip stability and active motion control involving coronal and lateral lunges predicted all forms of injury, whereas motor control was related to future non-traumatic injury in 359 division 1 athletes. We6 purposely did not define quantitative values for the constructs, but clustered each tool within the construct to create grand variables consisting of multiple tools that reflected the construct. By using this method, we were better able to define the underlying constructs associated with future injury, by combining all the tools within the constructs, versus limiting tool to distinct values.

Limiting tools to distinct values markedly reduces one's ability to define who is at risk for injury based on a numeric threshold score. Working off constructs allows the intervener to focus on the most important deficiencies/impairment identified within the constructs associated with injury and the variabilities of deficiencies seen among our athletes. For example, we6 commonly found variabilities in deficiencies in one-sided (left-only, right-only or both) abduction and/or adduction during side plank that would have been missed with a threshold score that required an amalgamated finding.

What about when a baseline variable predicts an injury?

Occasionally, authors do find a relationship between a base set of physical predictors and future injury. Understandably, these studies receive a great deal of attention and since athletic injuries are a ‘high-stakes’ outcome, they are adopted early in clinical practice. However, these findings are often evaluated statistically in non-dynamic models and demand future and further replication across a full spectrum of sports and competitive levels.

Further, many studies were derived using a proprietary system as the predictive tool; such a finding is potentially profitable for the researchers. When these events occur, the following steps are necessary: (1) a similar finding must be provided by an independent group, with no financial interests in the outcome using a complex modelling technique; (2) the finding should be evaluated in a dynamic model, to confirm the legitimacy and transferability of the work and (3) adoption into clinical practice should be delayed until subsequent confirmation. Although these involve extra steps, the steps reduce the risks of inappropriate adoption or acceptance of a finding that is likely to be a statistical anomaly.


View Abstract


  • Competing interests None declared.

  • Provenance and peer review Not commissioned; internally peer reviewed.

Linked Articles