Application of inertial measurement units and machine learning for fatigue assessment in badminton athletes

Objective To evaluate offline binary classification of recorded fatigue conditions using inertial measurement unit (IMU) acceleration features and five machine-learning algorithms. Methods The author-confirmed performance dataset contained 32 observations per condition. The revised primary IMU dataset comprised 70 recordings from 25 roster-linked athletes, with 22 non-fatigue and 48 fatigue labels after correction of four misclassified recordings. Seven-sensor and right-forearm configurations provided 231 and 33 candidate features. A newly reconstructed analysis retained five documented outer athlete-grouped folds and used three grouped inner folds for training-only preprocessing, sparse logistic feature selection and hyperparameter tuning. Results CMJ was 40.01 ± 8.51 and 30.27 ± 7.29 in the non-fatigue and fatigue conditions, respectively. Across classifiers, fold-mean AUC ranged from 0.655 to 0.846 for seven sensors and from 0.715 to 0.846 for the right forearm. Macro F1 ranged from 0.600 to 0.775 and from 0.655 to 0.775, respectively. Specificity ranged from 0.483 to 0.700 and from 0.550 to 0.650. Code, inclusion decisions, actual inner and outer assignments and held-out predictions are supplied for this new run. Conclusion The results support further investigation of IMU-based condition classification, but do not establish algorithm superiority, sensor equivalence, physiological validity of every recording label, or real-time monitoring performance.