A Multimodal Machine Learning Framework for Athlete Monitoring, Performance Prediction and Sports Decision Intelligence

The assessment of physical fitness is essential to the development of the understanding of athletic ability and competitive success in a variety of sports. With machine learning, there are possibilities to detect intricate relationships between physical characteristics and competition outcomes. This study was designed to investigate some of the physical fitness parameters and classified the competitive achievement of the 393 athletes from ten different sports in a cross-sectional study. The algorithms used for descriptive statistical analysis were Logistic Regression, Random Forest, Support Vector Machine, Gradient Boosting and K-Nearest Neighbors algorithms of machine learning. Performance of the classifiers was evaluated using accuracy, precision, recall and f1 score and influential predictors were identified using permutation feature importance. S- and T-values were used to compare the fitness indicators of the medals and no differences were found. Logistic Regression showed the best classification accuracy (58.23%) while K-Nearest Neighbors showed the best macro-F1 score (0.5381). The classification came into effect and factors like running performance, age, muscular strength and endurance started to affect the classification. The findings indicate moderate level of machine learning performance and highlight the use of cross-sectional fitness measures as being limited in discriminating successful competitive performance. It is recommended to continue monitoring longitudinally, collect full physiological data and validate externally to increase the reliability in predictions and applicability of the method.