Machine Learning-Based Prediction of Performance Gaps in Rowing and Identification of Key Training Monitoring...

Although routine biomechanical monitoring in rowing increasingly relies on sensor-based and instrumented measurement systems that can capture multidimensional performance indicators with considerable precision, systematic approaches are still needed to integrate these sensor-derived data into a unified monitoring dataset and translate them into decision support for practice. This study aimed to construct a unified rowing training monitoring dataset from real-world sensor-derived biomechanical measurements, develop predictive models for athletes' performance gaps relative to target 2 km performance, and, for target attainment classification, identify key training monitoring indicators and evaluate their practical value in training practice. A total of 249 biomechanical testing records collected during the 2024-2025 season from the Chinese National Rowing Team were included. After standardized processing, 449 athlete-level records were generated for the primary analysis. Following exclusion of observations with missing primary regression labels, 172 modeling records were retained, corresponding to 87 test reports and 119 athletes. The primary regression outcome was the percentage time difference relative to target 2 km performance. XGBoost Regressor, Elastic Net, and LASSO were used for regression modeling, whereas Logistic Regression and XGBoost Classifier were used for the secondary classification task of target attainment. Internal validation was performed using grouped cross-validation at the athlete level, and model interpretation was supported by permutation importance, sparse linear coefficients, and robustness analyses. The results showed that all formal models outperformed their respective baseline models. In the primary regression task, XGBoost Regressor achieved the best performance in terms of MAE, whereas Elastic Net performed best in RMSE and R². The key training-monitoring indicators mainly included mean boat velocity, minimum boat velocity, stroke rate, distance per stroke, and efficiency-related variables. After removal of grouping variables related to boat class, sex, and weight category, the performance of XGBoost Regressor remained largely stable, suggesting that the primary predictive signal was mainly derived from measured technical and biomechanical features. In the secondary classification task, XGBoost Classifier achieved an ROC AUC of 0.992. This study provides a team-specific applied framework for extending sensor-derived rowing monitoring outputs from multi-indicator measurement toward interpretable performance evaluation and decision support within an elite-team setting, while broader external validation remains necessary.