PREDICTIVE MODELS IN SPORT SCIENCE: MULTI-DIMENSIONAL ANALYSIS OF FOOTBALL TRAINING AND INJURY PREDICTION

Due to the fact that team sports such as football have a complex multidirectional and intermittent nature, an accurate planning of the training workload is needed in order to maximise the athletes’ performance during the matches and reduce their risk of injury. Despite the evaluation of external workloads during trainings and matches has become more and more easier thanks to the advent of the tracking system technologies such as Global Position System (GPS), the planning of the best training workloads aimed to obtain the higher performance during the matches and a lower risk of injury during sport stimuli, is still a very difficult challenge for sport scientists, athletic trainers and coaches. The application of machine learning approaches on sport sciences aims to solve this crucial issue. Hence, the combination between data and sport scientists’ peculiarities could maximize the information that can be obtained from the football training and match analysis. Thus, the aim of this thesis is to provide examples of the application of the machine learning approach on sport science. In particular, two studies are provided with the aim of detecting a pattern during in-season football training weeks and predicting injuries. For these studies, 23 elite football players were monitored in eighty in-season trainings by using a portable non-differential 10 Hz global position system (GPS) integrated with 100 Hz 3-D accelerometer, a 3-D gyroscope, and a 3-D digital, Northern Ireland compass (STATSports Viper). Information about non-traumatic injuries were also recorded by the club’s medical staff. In order to detect a pattern during the in-season training weeks and the injuries, Extra Tree Random Forest (ETRFC) and Decision Tree (DT) Classifier were computed, respectively. In the first study it was found that the in-season football trainings follow a sinusoidal model (i.e. zig-zag shape found in autocorrelation analysis) because their periodization is characterized by repeated short-term cycles which are constituted by two parts: the first one (i.e. trainings long before the match) is constituted by high training loads, and the second one (i.e. trainings close to the match) by low ones. This short-term structure appears to be a strategy useful both to facilitate the decay of accumulated fatigue from high training loads performed at the beginning of the cycle and to promote readiness for the following performance. As a matter of fact, a patter was detected through the in-season football training weeks by ETRFC. This machine learning process can accurately define the training loads to be performed in each training day to maintain higher performance throughout the season. Moreover, it was found that the most important features able to discriminate short-term training days are the distance covered above 20 W·kg-1, the acceleration above 2 m·s-2, the total distance and the distance covered above 25.5 W·Kg-1 and below 19.8Km·h-1. Thus, in accordance with the results found in this study, athletic trainers and coaches may use machine learning processes to define training loads with the aim of obtaining the best performance during all the season matches. Players’ training loads discrepancy in comparison with the ones defined by athletic trainers and coaches as the best ones to obtain enhancement in match performance, might be considered an index of individuals’ physical issue, which could induce injuries. As a matter of fact, in the second study presented in this thesis, it was found that it is possible to correctly predict 60.9% of the injuries by using the rules defined by DT classifier assessing training loads in a predictive window of 6-days. In particular, it was found that the number of injuries that the player suffered through the season, the total number of Acceleration above 2 m·s-2 and 3 m·s-2, and the distance in meters when the Metabolic Power (Energy Consumption per Kilogramme per second) is above the value of 25.5 W/Kg per minute, are the most important features able to predict injuries. Moreover, the football team analysed in this thesis should keep under control the discrepancy of these features when players return to the regular training because of the numerous fall-backs into injuries that have been recorded. Thus, this machine learning approach enables football teams to identify when their players should pay more attention during both trainings and matches in order to reduce the injury risk, while improving team strategy. In conclusion, Machine Learning processes could help athletic trainers and coaches with the coaching process. In particular, they could define which training loads could be useful to obtain enhancement in sport performance and to predict injuries. The diversities of coaching processes and physical characteristics of the football players in each team do not permit to make inferences on the football players’ population. Hence, these models should be built in each team in order to improve the accuracy of the machine learning processes.