. Injuries are a common problem in professional football. A challenge that the medical team faces is to successfully predict the recovery time of an injured player. Current medical standards can only give vague predictions as to when a player will return to play. Obviously, making an accurate prediction as soon as possible would be helpful to the coach. This research tries to answer the question of whether it is possible to predict when a player will return to play, based on information at the moment of injury, while also comparing three machine learning methods for this task: support vector machines, Gaussian processes and neural networks. The tests were conducted on data from the professional football club of Tottenham Hotspur. The results demonstrate that this task can be completed with a reasonable amount of accuracy, without any method performing significantly better than the rest. Future directions and possible improvements are discussed.
Injuries are a common problem in every sport, including football. Professional football players get injured on average once per year [1] with 10-35 injuries occurring per 1000 game hours [2]. Injuries have been described as the main factor that prevents professional players from not being able to participate in training and playing activities [3].
The factors that cause injuries can vary. A significant percentage of injuries
(9%-34%) happening due to overuse [4-5]. Most of the injuries are described as
traumatic, with 29% of them being due to foul play [
As soon as an injury happens it is important to make an estimate of how long the player will need to recover from the injury and get back to play. This information can help the manager make appropriate changes in the squad or the tactical planning of the team. It can also help the director of the club, since new players might need to get signed in order to cover for players who are going to stay out of play for a long time. Additionally, managing the player’s expectations with respect to his injury is important, so that the player can prepare himself mentally and psychologically. Finally, it would help the medical team by providing additional certainty in the predictions of the experts.
Currently, there is no standard method to estimate the time a player will miss from play. The time is estimated based on the experience of the physician and by recommendations by various groups and studies. The suggestions can vary quite significantly with each other, and they can also have significant variance. For example, suggestions for return to play following anterior cruciate ligament reconstruction can range from 16 to 24 weeks [8]. Similar recommendations exist for hamstring injuries [9] and concussions [10-13].
Machine learning has been used in sports for various purposes (e.g. cycling [14] and swimming [15] ) including football [16-17]. The complicated and multi-factorial nature of many sports makes machine learning a natural choice for predictive tasks.
The purpose of this study is to compare different machine learning methods on predicting the recovery time of professional football athletes. The goal is to make the prediction based on information available at the time of injury, before an official diagnosis has been conducted. There are two main reasons for which the final diagnosis was left out. First, diagnoses, in some cases, can take some time, while ideally a coach would like know as soon as possible how long a player will stay out of play.
Secondly, there are many different diagnoses and different levels of abstraction that can be used. For example, in this study’s dataset there were some knee injuries that were described as “knee pain, unspecified”, “patellofemoral pain” and “Left knee medial meniscus”. These diagnoses could be elaborated even further, or they could be abstracted, by classifying them all as “knee injuries”. This is a medical problem that can influence the performance of any machine learning or statistical model that will use this information.
However, it is not entirely clear what degree of elaboration would actually help in the prediction of the response variable. For that reason it is important to know what degree of accuracy can be achieved in the prediction of the response variable before including the diagnosis, so that future research could actually tackle the problem of trying to identify the correct level of abstraction needed for this task.
The methods that were chosen for this research were Gaussian processes, support vector machines and neural networks. The reason behind these choices is that all these methods are popular for regression tasks. While there are many other choices for solving regression problems in machine learning, these three methods have been proved and tested in a variety of applications, so they provide sensible choices for approaching this task.
The primary goal of this study was to test the degree to which this task is possible in general by reaching a level of error in the predictions that can have practical applicability, at least in some cases. Once this was established, the next goal was to see whether one of these methods is more suited for this task compared to others. The study itself is part of a greater research project that has as a final goal a fully-working predictive system that can aid football teams. Therefore, future plans, directions and suggestions for research are discussed, as well. 2 2.1
The dataset consists of a list of injuries at Tottenham Hotspur Football Club. For every injury, a list of variables was collected. These are presented in table 1. Note that the variable “injury” included in the dataset is not a final diagnosis, but a first general estimate such as “muscle strain” or “bone injury”.
All variables, with the exception of “Age” and “Days unavailable” were categorical variables and they were converted to dummy variables. This gave rise to a dataset that contains 78 variables (including the response variable).
A histogram of the dataset is shown in figure 1. It is evident that most of the injuries are less than 25 days and the histogram is skewed. The total number of cases is 154.
Three different methods were used and evaluated: neural networks, support vector machines and Gaussian processes. Each method was executed with many different parameter sets. In order to find the best parameters, grid search was used. Due to the number of tests (more than 50 tests for each method) conducted it is not practical to provide tables and graphs for each parameter set and result. Therefore, tables 2-6 below show the parameters that each method used and their value ranges.
The neural network was trained using backpropagation with momentum. Min Max Min Max Min Max
C 0 200 Min
Max
Learning Momentum Rate 0.2 0.2 0.5 0.4
For this particular task, it was observed that it was difficult to evaluate the success of the algorithms by using the standard mean squared error alone. The interpretability of the results is crucial, and the mean squared error is difficult to be communicated to a medical professional. For that reason, the absolute error was used in addition to the mean squared error when reporting results, even though the mean squared error was the optimization objective in each trial. All methods were evaluated using 10-fold cross validation and all tests were executed using RapidMiner version 5.3.
Another issue with the evaluation of the results is the desired degree of accuracy that is required for a method in this task to be considered successful from the perspective of practical applicability. Football teams play a certain amount of games within a season. Usually this is 4 league games per month, and maybe some more cup games and games in European competitions. If a player is injured in a game, it might not matter so much whether he will be back in play in 3 or 5 days, as long as the coach knows that in 7 days, when the next game starts, he will be ready to play.
Furthermore, the dataset contains many cases where the player stayed out of play for one day or no days at all. Many of these cases do not require the execution of a predictive algorithm, because the medical professionals of the team can very quickly classify the injury as transient. Predictions are more helpful for injuries that have longer lasting effects, for example, more than a couple of weeks. This means, that the margin of error can be higher. If the medical staff’s opinion is that a player will miss 5 to 10 weeks, then a prediction that manages to narrow down this margin to, for example, 6 to 7 weeks, can help the coach make better decisions and plan for the future. 3
The best results achieved for each method are in table 7. The errors are presented for both the test data and the whole dataset. The test errors are accompanied by their corresponding standard deviations, as they have been calculated from the 10-fold cross-validation. Standard deviations do not apply to the total errors, since they are computed for the whole dataset.
It is evident that this task can be predicted with some degree of accuracy. The test accuracies are similar, and their variances are big, so no single method seems to perform significantly better to others. However, the important point is that the task can be completed with a fair degree of accuracy. It is possible to make an estimate of when a player will return to play based on information at the moment of injury, before an official diagnosis is conducted.
The absolute error that the algorithms achieve in the test accuracy is about in the range (17, 21). This makes their application, as they are, more suitable for mid-severity to severe injuries where the player is likely to stay out of play for a month or more.
The results are even more important if it is considered that the size of the dataset should be considered small for this task and it concerns only a single football club. There are many types of injuries in football that can occur under different circumstances. Future research should use datasets from other football clubs in order to verify and expand the current results. Ideally, datasets from football clubs from different countries should be obtained, since the style of play in each country, along with other factors (e.g. a country’s climate), could influence the response variables.
Obviously, the end goal is the practical applicability of the results. An interesting feature of this task is that the models could be included in a diagnostic protocol. After each injury, the medical staff will conduct detailed medical tests in order to diagnose the injury. Models like the ones described in this paper could accompany a diagnosis, providing some additional support for the experts’ estimates.
Furthermore, additional information that could be available at the moment of injury includes anthropometric and medical information such as the height, weight or medical blood tests of players. This information could improve the accuracy of the model, while also staying true to its original goal of making predictions right after an injury has occurred.
Finally, future research could also solve the problem of how additional official diagnostic information could be used alongside this model in order to make more accurate predictions. 5
This research dealt with the question of whether it is possible to predict the recovery time after an injury in professional football without an official diagnosis, while it also testing 3 methods against each other for this task. The results demonstrate that it is possible to reach some degree of accuracy in this task, but the size of the dataset, and maybe the variables themselves, limit the accuracy that can be reached. No single method was deemed to be significantly better than any of the other methods that were used.
However, this work paves the way for future research that can include bigger and more complicated datasets and can also be extended by protocols that can combine experts’ opinions. Future research will built on top of the current results in order to provide a functional system for assessing injuries in professional football. 6
The author of this research would like to thank Wayne Diesel of Tottenham Hotspur Football Club for kindly sharing the data for this research. 18. 19.