XGBoost optimization using hybrid Bayesian optimization and nested cross validation for calorie prediction

Abstract

Accurately predicting calorie expenditure is crucial for wearable device applications, enabling personalized fitness and health recommendations. However, traditional models struggle with high data variability and nonlinear relationships in activity data, leading to suboptimal predictions. This study addresses these challenges by integrating extreme gradient boosting (XGBoost) with Bayesian optimization and nested cross validation to enhance predictive accuracy. Unlike previous approaches, our method systematically tunes hyperparameters using Bayesian optimization while employing nested cross validation to prevent overfitting, ensuring robust model evaluation. We utilize a dataset of daily activity records, including steps, distance, and active minutes, extracted from wearable devices. Our experimental findings indicate a substantial enhancement in prediction performance, achieving a mean squared error (MSE) of 4294.27, an Rsquared (R2) score of 0.9917, and a root mean squared error (RMSE) of 65.53. The proposed model outperforms baseline approaches such as random forest and support vector machines in terms of predictive accuracy. These findings underscore the advantage of our approach in predictive modeling. Beyond calorie estimation, the proposed methodology is adaptable to other domains requiring high-precision predictions, such as healthcare analytics and personalized recommendation systems.