Calorie Burn Prediction using XGBoost with Feature Selection and SHAP Analysis

Abstract

Calorie burn prediction plays a crucial role in fitness assessment and personalized exercise guidance. This study applies XGBoost regressor with feature selection and SHAP (SHapley Additive exPlanations) analysis to predict calories burned in an exercise session, using physiological and activity-based features for training model. Feature selection was conducted using correlation analysis and Variance Inflation Factor (VIF) screening to reduce redundancy and improve interpretability, resulting in a simplified 5-feature model that balances performance and model complexity.While the full-feature XGBoost regressor achieves the highest predictive accuracy, the proposed five-feature XGBoost regressor demonstrates similar performance with an MAE of 2.19, an MSE of 9.98, and an R2 of 0.9972, while reducing input dimensionality. Model interpretability is further enhanced through global and local SHAP analysis, revealing the significant influence of heart rate and duration on predictions. These results indicate the  potential of gradient boosting models for session calorie burn prediction while suggesting their applicability in fitness tracking systems and personalized exercise guidance system.