ScholarWorks@중앙대학교

초록

Purpose: The purpose of this study is to develop a model that can predict the probability of injuries occurring during weight training by utilizing machine learning algorithms for data analysis. Method: For this purpose, this study utilized Google’s machine learning tool and Teachable Machine. The exercise movements chosen were Deadlift and Overhead Press, commonly used in weight training, while the injury movement was defined as the back rounding. Participants were trained to differentiate between “correct movements” and “injury movements,” which were then used to train the Teachable Machine to develop an injury identification model. The model’s error rate was then assessed using evaluation participants to determine its actual error rate. Results: The Deadlift identification model and the Overhead Press identification model both achieved a class accuracy of 1.00 for both correct and injury-inducing postures, with a confusion matrix of 0. The learning rate per epoch and loss graph per epoch peaked early in the training process. The error rate of the injury identification model was 19% for the Deadlift identification model and 12% for the Overhead Press identification model. Conclusion: The high accuracy per class and the result of the confusion matrix indicate effective hyperparameter adjustment. The overfitting observed in the learning rate and loss graph per epoch is likely due to the limited diversity in the training data, a characteristic of the study. The higher-than-expected error rate in the injury identification model is attributed to the constraints of the 2D posture estimation inherent in the Teachable Machine Pose Project, which relies on a Skeleton-Based Model that limits the representation of the human body in a 2D space.

키워드