Comparison of machine learning classifiers for differentiating level and sport using movement data.

Area of Science:

• Sports Science
• Biomechanics
• Machine Learning in Sports

Background:

• Understanding athlete classification based on movement is crucial for performance analysis and injury prevention.
• Traditional machine learning methods have limitations in analyzing complex, time-series kinematic data.

Purpose of the Study:

• To compare the performance of recurrent neural networks (RNNs) against traditional classifiers for athlete categorization.
• To investigate whether athletes from different sports display unique movement patterns during non-sport-specific screens.

Main Methods:

• Utilized optical-based kinematic data from 542 athletes.
• Applied nine distinct machine learning algorithms, including RNNs and traditional linear/non-linear classifiers.
• Employed principal component analysis (PCA) and feature selection for dimensionality reduction in traditional models.

Main Results:

• RNNs and linear classifiers generally outperformed non-linear classifiers in athlete classification tasks.
• Reservoir computing demonstrated rapid training times and high classification accuracy, though with reduced interpretability.
• Successful classification by sport indicated sport-specific movement differences during general movement screens.

Conclusions:

• RNNs offer a powerful approach for analyzing multivariate, time-series kinematic data in sports science.
• Movement assessment screens should integrate sport-specific scoring to accurately reflect athlete capabilities.
• Distinct movement signatures exist across different sports, even in non-sport-specific tasks.