Development of a Machine Learning-Based VO2max Prediction Framework Integrating Missing Data Imputation and SHAP Interpretation

초록

This study aimed to develop a VO(2)max prediction framework that integrates missing data imputation, utilization of missingness patterns, residual correction - based machine learning modeling, and SHAP interpretation using large-scale national physical fitness data. Various imputation methods were compared through a masked-value recovery evaluation, and MICE (BayesianRidge) demonstrated the best restoration performance and was therefore selected as the optimal imputer. For predictive modeling, a two-stage framework, OUR-MARC (ExtraTrees + Histogram-based Gradient Boosting Regressor), was proposed, combining baseline prediction with residual correction. On the test dataset, OUR-MARC showed the best numerical performance on the held-out test dataset. (R-2 = 0.8772, RMSE = 2.1001, MAE = 1.5339). These findings suggest that explicitly incorporating missingness pattern information and residual structure contributes to improved predictive stability and mitigation of overfitting. SHAP analysis identified HR Recovery, body mass index (BMI), muscular endurance, body fat percentage, and age group as key predictors of VO(2)max. Moreover, structural differences in variable importance were observed across sex-stratified analyses, highlighting the necessity of sex-specific considerations in VO(2)max prediction. Overall, the proposed framework demonstrates that high predictive accuracy and interpretability can be simultaneously achieved in high-missingness physical fitness data environments.

키워드