Playing it safe

Machine learning, which uses an artificial neural network to identify patterns in complex datasets, has found applications in almost every industry including but not limited to medicine, transportation, and information technology. Today scientists are trying to see if machine learning can be just as helpful in tackling soccer injuries, a problem that costs the sports industry billions of dollars annually.

“In the past, regression analysis was used to assess injury risk and predict sports performance,” says George Nassis, the Chair of Physical Education Department at the United Arab Emirates University, who recently co-authored a review paper on the subject. “The problem with this approach is that sports injuries are dependent on multiple factors that are not necessarily linearly associated.”

Soccer is the world’s most popular sports game, with over 5 billion die-hard fans in 120 countries. It is also one of the most physically demanding as soccer players need to run, pass and shoot as well as sprint maximally and perform various deceptive, change-of direction maneuvers. These actions often result in musculoskeletal injuries, which are not only painful but may reduce players’ mobility and impair their future performance.

Some of the factors accounting for injury risks include past injuries, muscle strength imbalance and aerobic fitness level. These factors may work independently or in combination to influence a player’s performance, thus making injury prediction too difficult if not impossible using traditional statistical methods. In comparison, an artificial neural network learns to become better and better at recognizing cross features or patterns in complex datasets. Although still an unproven concept, it offers a glimmer of hope to people working in the field.

“Several studies claim their algorithms can predict injury risks with moderate to high accuracy,” says Dr. Nassis, who casts doubt on what this “moderate to high accuracy” really means in clinical terms. “Furthermore, data collected months before an injury cannot account for the dynamic nature of a soccer match. A slight deviation in the procedure of data collection can critically affect the outcome of injury prediction.”

Finally, the low incidence of injuries in soccer presents a problem as it may prevent algorithms from reaching a higher prediction accuracy. Dr. Nassis acknowledges that more work is needed to make injury risk prediction acceptable if not a reality. He suggests that integrating machine learning with big data might be one way going forward.