2015 ISAKOS Biennial Congress ePoster #2605

A Gaussian Process Model for Predicting Time to Return to Play Following Injury in the Professional Footballer

Neil Jain, BM, MRCS(Ed), FRCS(Tr&Orth), Manchester UNITED KINGDOM
Stylianos Kampakis, MSc, London UNITED KINGDOM
David Murray, FRCS(Orth), Headington, Oxford UNITED KINGDOM
Steve Kemp, MSc, BSc (hons) PG Dip (manips), GSR, MCSP, MMACP, Wolverhampton UNITED KINGDOM
Phil Hayward, MCSP, MMACP, Wolverhampton UNITED KINGDOM
Philip Treleaven, PhD, London UNITED KINGDOM
Fares S. Haddad, MCh(Orth), BSc, FRCS(Orth), London UNITED KINGDOM

University College London, London, UNITED KINGDOM

FDA Status Cleared

Summary: We present a gaussian process model that may be used by a computer to provide an accurate prediction in the time to return to play following injury in the professional footballer that appears to be more accurate than existing methods of predicting return to play.




Injuries are a common problem in football for a variety of reasons. When a player is injured it is often difficult to make an accurate prediction of how long he is going to stay out of play which forces the coach to make tactical and training decisions under uncertainty.


that would allow a prompt and accurate prediction of how long a player is going to stay out of play after an injury would be very useful. Machine learning is a field of computer science, which studies algorithms that can be used for prediction and forecasting. The purpose of this study was to create a model that can forecast how many days a player is going to stay out of play, once he gets injured, based on information that would be readily available at the moment of injury.


The sample consisted of 116 reported injuries from footballers of the team Wolverthampton Wanderers F.C. for the season 2009-2012. The data consisted only of variables that are directly available at the moment of injury and do not require a medical examination. A Gaussian process model with a Laplacian kernel was used for the prediction of the values.


A clear trend was observed. The model managed to achieve a high degree of accuracy, with very high accuracy for milder injuries and slightly reduced accuracy for very serious injuries. The root mean square error was 13.186 (S.D.: 8.073), the mean absolute error was 8.192 (S.D.: 13.106) and the mean relative error 171.97% (S.D.: 75.56%). Most of the errors are 2 days or less with errors increasing as the actual value of the variable gets larger. The largest absolute error that was recorded was approximately 9.5 days for an injury whose true value was 105 days. Essentially, the model was generally able to provide a predicted return to play within around a 10% difference to the actual return to play.


A Gaussian process model is a viable method for forecasting how many days a player will be unavailable after an injury. A high degree of accuracy can be achieved before any medical examination is conducted, based solely on information of how the injury occurred.