Personalized training model for 10 m air pistol through machine learning: a pilot study

This pilot study aimed to develop an interpretable machine-learning framework to classify high- versus low-ring-value performance in 10 m air pistol shooting and to identify key technical factors relevant to training feedback. A total of 3,179 valid shots were collected from an elite shooter using a SCATT laser training system. Eight SCATT-derived metrics were extracted from aiming-trajectory and process data. An XGBoost classifier was trained with SMOTE-Tomek to mitigate class imbalance and Optuna for hyperparameter optimization. The decision threshold was selected on the training set via cross-validation by maximizing the F1 score. Model interpretability was examined using SHAP to quantify feature contributions. On the held-out test set, the optimized XGBoost model achieved an AUC of 0.86 and an accuracy of 0.83 (F1-optimized threshold = 0.30). SHAP analyses the most influential features, indicating that smaller deviation and more stable final-second aiming were associated with high-ring-value performance. This interpretable classification framework provides data-driven, individualized technical feedback from SCATT data and may support practical decision-making in precision shooting training. Further validation with additional athletes is needed to improve generalizability.