Data-driven Human Energy Expenditure Prediction: Contrasting Personalized and Population-Based Models

Human energy expenditure (EE) varies between individuals and is influenced by the dynamics of the thermic effects of food and physical activity. Advances in wearable sensors and machine learning (ML) have improved EE prediction, yet most studies rely on group-level models. Personalized models have shown 10%–20% higher accuracy but require efficient methods to balance sensor practicality with measurement precision. This study examines both general and personalized EE prediction for low-to-medium intensity activities common in office environments, and it is the most comprehensive work on this topic to date. Using data from 23 participants (12 F, 11 M), we developed person-specific and group-level models with the Random Forest (RF) algorithm, analyzing 7 combinations of 4 biomarkers (skin temperature, heat flux, accelerometry, heart rate) across 1 to 16 body locations. A total of 3’094 tailored models were generated. Results show that the highest accuracy in personalized EE prediction is achieved when combining movement-based, thermal, and cardiovascular data. While accelerometry alone performs well, especially for males, adding physiological inputs, particularly skin temperature, improves accuracy for females. Lower-body accelerometers, particularly on the calf, proved the most reliable. Overall, personalized models achieve higher accuracy (2% MAPE) but are more sensitive to sensor placement and data availability, whereas generalized models offer greater robustness but lower accuracy (16.5%–28% MAPE). These findings highlight the merit of careful sensor selection strategies and the need for hybrid modeling approaches integrating subject-specific calibration to balance accuracy and robustness in personalized EE prediction in real-world applications.