AI-BASED PRESCRIPTIVE SYSTEM FOR INDIVIDUALS OPTIMIZED AND NORMALIZED EVALUATION FROM ELITE- REFERENCED BUTTERFLY SWIMMING (AI-PIONEER). PROOF OF THE CONCEPT

Abstract

Traditional methods of performance assessment in sports are often subjective, time consuming and require complex expertise to interpret the data accurately. To alleviate current limitations, artificial intelligence, utilizing advanced techniques, presents a promising solution by automating motion analysis, reducing human bias, and improving the accuracy of performance assessment. This study aims to address the challenge of predicting temporally coherent IMU sequences representing optimal swimming butterfly stroke cycle for amateur athletes, conditioned on static body parameters and kinematic dry swimming parameters extracted from IMU sensors. The objective is to transform a low-dimensional input vector, consisting of 21 static features, into high-resolution temporal output sequences, represented as [T time steps × 48 IMU channels], where each of the 8 sensors records 6 values (three for acceleration and three for angular velocity). The proposed Physics-Informed Conditional Variational Autoencoder (CVAE) system demonstrates the potential to simulate elite-referenced movement patterns for non-elite athletes, providing a tool for optimizing butterfly swimming technique. While the ML architecture shows promise in learning from limited data and integrating biomechanical reasoning, the current implementation faces challenges, primarily due to the small dataset size which affects generalizability, the use of dry-land data collection instead of actual swimming, which affect the accuracy and applicability of the model in real swimming scenarios, and the low fidelity of generated sequences, which limits practical use. Thus, this study should be considered a proof of concept rather than a fully deployable solution.