Predicting Fitness and Performance of Diving using Machine Learning Algorithms

Abstract

Various batteries of fitness tests are conducted to assess the fitness of divers. Testing and monitoring are conducted for the selection of divers based on their current fitness for future competitions. There is a paucity of data and knowledge of which fitness tests can be effective to predict fitness, performance, and selection of potential divers at national and international level competitions. Classical statistical methods could not identify the magnitude of the variable importance. Hence tree-based classifier Boruta algorithm was used to identify the most important features of the diving performance. Out of 19 variables, vertical jump, standing broad jump, sit & reach, shoulder flexibility, agility, pull-ups, maximum push-ups, maximum wall bar, handstand balance, age, height, and weight were more than 90% important for the diving performance. VO2max and static balance were 69% important while hyperextension and bridge flexibility were about 52% important. The ordinal forest and multi-ordinal logistic regression machine learning algorithms were used to develop models for predicting diving performance at national and international competitions. The diagnostic analysis was conducted to validate the model performances using unseen data. The predictive accuracy by multi-ordinal logistic regression with 6 variables was 52.1% (95% CI: 39.92%, 64.12%), which was poor as compared to the ordinal forest. Overall accuracy of the final ordinal forest, was 70.4% (95% confidence interval: 58.41%, 80.67%). The balanced accuracy was 68.5% and 82.8% for predicting the national and international performances respectively. Hence the ordinal forest is better for selecting high potential divers objectively at international competitions.