Ski injury predictive analytics from massive ski lift transportation data

Abstract

Ski injury research is traditionally studied on small-scale observational studies where risk factors from univariate and multivariate statistical models are extracted. In this article, a large-scale ski injury observational study was conducted by analyzing skier transportation data from six consecutive seasons. Logistic regression and chi-square automatic interaction detection decision tree models for ski injury predictions are proposed. While logistic regression assumes a linearly weighted dependency between the predictors and the response variable, chi-square automatic interaction detection assumes a non-linear and hierarchical dependency. Logistic regression also assumes a monotonic relationship between each predictor variable and the response variable, while chi-square automatic interaction detection does not require such an assumption. In this research, the chi-square automatic interaction detection decision tree model achieved a higher odds ratio and area under the receiver operating characteristic curve in predicting ski injury. Both logistic regression and chi-square automatic interaction detection identified the daily time spent in the ski lift transportation system as the most important feature for ski injury prediction which provides solid evidence that ski injuries are early-failure events. Skiers who are at the highest risk of injury also exhibit higher lift switching behavior while performing faster runs and preferring ski slopes with higher vertical descents. The lowest injury risk is observed for skiers who spend more time in the ski lift transportation system and ski faster than the average population.