Ski Injury Predictions with Explanations

Abstract

Providing prediction models for ski injuries is a very challenging classification problem. We propose a model for injury prediction that uses ski lift trajectory features. Ski slopes, in general, differ by width, length, difficulty and geographical position on the mountain, which results in different patterns of skiing. We study the correlation between these patterns different types of ski injuries. Many types of analysis were proposed in this domain of research. However, they are either too simple for real-time usage, such as univariate statistical analysis, or use interpretable predictive models at the cost of lowering accuracy. In order to gain best predictive performance and still provide explanation one must combine different approaches. We utilize modern algorithms such as random forests and gradient boosted trees with explainability methods Shap and Lime for providing interpretation about reasons for specific decision. The proposed models were created on Mt. Kopaonik, Serbia ski resort and it is shown that ski injury in the following hour on specific ski slope can be predicted with AUC ~0.76, which is better up to ~15% compared to classical approaches such as logistic regression and decision trees.

Sandro Radovanović

Assistant Professor at University of Belgrade

My research interests include machine learning, development and design of decision support systems, decision theory, and fairness and justice concepts in algorithmic decision making.