Penerapan Algoritma Random Forest Untuk Klasifikasi Cedera Pelari Berbasis Data Kinerja Wearable Devices dengan Teknik Resampling dan Optuna Tuning

Abstract

Advances in wearable devices enable continuous and objective monitoring of runners’ performance, supporting data-driven classification and prevention of running-related injuries. However, injury modeling remains challenging due to extreme class imbalance and temporal complexity. The weekly temporal dataset used in this study is highly imbalanced, consisting of 42,223 non-injury samples and 575 injury samples, which may bias classification models toward the majority class if not properly addressed. To mitigate this issue, three undersampling techniques—Random Undersampling, Cluster-based Undersampling, and NearMiss Undersampling—were applied to preserve representative and informative non-injury patterns. Experimental results show that NearMiss Undersampling achieved the best baseline performance, with an accuracy of 0.8826, an F1-score of 0.8756, and an AUC of 0.9645. Further improvements were obtained through hyperparameter tuning using Optuna with a Tree-structured Parzen Estimator (TPE), increasing performance to an accuracy of 0.9000, an F1-score of 0.8959, and an AUC of 0.9782. In addition, data augmentation experiments using SMOTE in the Cluster-based Undersampling scenario indicated that a 20% augmentation level yielded the most stable results. Overall, these findings demonstrate that combining appropriate resampling strategies with hyperparameter optimization effectively enhances runner injury classification based on wearable device data.