Machine Learning-Based Stroke Segmentation in Kayaking Using Integrated IMU and EMG Data

Gábor Nagy, Péter Katona, Levente Gannorouwa, László Grand

2025

Abstract

Accurate classification of stroke side in rowing motions is essential for performance monitoring and injury prevention. This study evaluates three machine learning models — Naive Bayes (NB), Logistic Regression (LR), and Gradient Boosting Decision Trees (GBDT) — using biomechanical and electromyographic (EMG) features. A core set of 25 features was identified, with normalized joint coordinates and latissimus dorsi EMG activity among the most influential. The NB model achieved 92.21\% cross-validation accuracy using only three coordinate-based features, while the full feature set improved accuracy modestly by 1.94\%. The LR model attained 94.48\% accuracy, slightly outperforming NB. The GBDT model achieved the highest accuracy with 96.18\% on the test set, alongside the lowest mean absolute stroke onset detection error of 24.6 \pm 51.6 ms, corresponding to just 4.5\% of average stroke duration. Classification accuracy remained stable across stroke paces. A strong negative correlation (R = -0.935) between classification accuracy and onset detection error was observed across subjects, indicating that poorer spatial classification corresponds with greater temporal imprecision. Significant inter-subject variability was found, with accuracy ranging from 91.89\% to 98.9\%, likely reflecting individual differences in stroke technique and muscle activation patterns. A core set of biomechanical features were identified, such as normalized joint coordinates of th eulnar styloid and right olecranon, latissimus dorsi EMG activity among the most influential, vertical pelvis lateral bending and bilateral shoulder flexion. Tempo-based relative time averages of these features reveal clear phase-dependent patterns that contribute strongly to model decision-making. These results demonstrate that accurate stroke side classification can be achieved using a relatively small set of biomechanical features, with GBDT models providing superior performance.

Paper Citation

in Harvard Style

Nagy G., Katona P., Gannorouwa L. and Grand L. (2025). Machine Learning-Based Stroke Segmentation in Kayaking Using Integrated IMU and EMG Data. In Proceedings of the 13th International Conference on Sport Sciences Research and Technology Support - Volume 1: icSPORTS; ISBN 978-989-758-771-9, SciTePress, pages 256-264. DOI: 10.5220/0013781200003988

in Bibtex Style

@conference{icsports25,
author={Gábor Nagy and Péter Katona and Levente Gannorouwa and László Grand},
title={Machine Learning-Based Stroke Segmentation in Kayaking Using Integrated IMU and EMG Data},
booktitle={Proceedings of the 13th International Conference on Sport Sciences Research and Technology Support - Volume 1: icSPORTS},
year={2025},
pages={256-264},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0013781200003988},
isbn={978-989-758-771-9},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 13th International Conference on Sport Sciences Research and Technology Support - Volume 1: icSPORTS
TI - Machine Learning-Based Stroke Segmentation in Kayaking Using Integrated IMU and EMG Data
SN - 978-989-758-771-9
AU - Nagy G.
AU - Katona P.
AU - Gannorouwa L.
AU - Grand L.
PY - 2025
SP - 256
EP - 264
DO - 10.5220/0013781200003988
PB - SciTePress