Researchers have heralded the power of inertial sensors as a reliable swimmer-centric monitoring technology, however, regular uptake of this technology has not become common practice. Twenty-six elite swimmers participated in this study. An IMU (100Hz/500Hz) sensor was secured in the participant's third lumbar vertebrae. Features were extracted from swimming data using two techniques: a novel intrastroke cycle segmentation technique and conventional sliding window technique. Six supervised machine learning models were assessed on stroke prediction performance. Models trained using both feature extraction methods demonstrated high performance (≥ 0.99 weighted average precision, recall, F1-score, area under ROC curve and accuracy), low computational training times (< 3 seconds - bar XGB and when hyperparameters were tuned) and low computational prediction times (< 1 second). Significant differences were observed in weighted average stroke prediction F1-score (p = 0.0294) when using different feature extraction methods and model computational training time (p = 0.0007), and prediction time (p = 0.0026) when implementing hyperparameter tuning. Automatic swimming stroke classification offers benefits to observational coding and notational analysis, and opportunities for automated workload and performance monitoring in swimming. This stroke classification algorithm could be the key that unlocks the power of IMUs as a biofeedback tool in swimming.
Keywords: Inertial measurement unit; athlete load; athlete monitoring; machine learning; swimming.