Personalized Metabolism: Beyond the Standard Static Values
Thermal comfort standards and the HVAC systems built around them rest on a convenient fiction: that the metabolic heat production of a building occupant can be captured by a single fixed value, typically the 1.2 met assigned to a seated adult performing light office work. In reality, human energy expenditure is neither constant nor universal. It is the sum of four distinct components — resting metabolism, the thermic effect of food, activity-induced thermogenesis, and thermoregulation — each of which varies over the course of a day and differs substantially from one person to the next according to body composition, diet, behavior, and the thermal environment itself. Postprandial thermogenesis alone can shift metabolic rate for one to two hours after a meal, while sustained exposure to cooler conditions elicits thermoregulatory responses that further separate an individual’s true energy expenditure from the population average encoded in design guidelines. Treating metabolism as a personalized, time-varying signal, rather than a static set-point, is therefore essential to any building science that aims to serve the actual occupant rather than a statistical abstraction.
Data-Driven Prediction of Individual Metabolic Rate:
The personalized and dynamic view of metabolism has been advanced by the ICE lab across two complementary studies. Both confirm a central finding: metabolic rate is a deeply individual quantity that population averages fail to capture. When energy expenditure is predicted from wearable and environmental sensor data, models tailored to a single person consistently and markedly outperform generalized, population-based ones, with personalized errors falling as low as about 2% for some individuals, compared to roughly 16–28% for models trained across a group. The generalized approach gains robustness by averaging across people, but does so precisely by smoothing away the individual specificity that governs a given person’s metabolism, underscoring how far one individual’s energy expenditure can depart from any single population benchmark.
Beyond the accuracy gap, the two studies characterize how metabolic rate should be predicted and which signals matter for whom. Energy expenditure is best modeled as a time series, since its dynamics unfold over minutes and hours through activity, thermoregulation, and the thermic effect of food. Across modeling approaches, from deep learning ensembles that combine a convolutional network with gradient boosting to Random Forest models that weigh combinations of skin temperature, heat flux, accelerometry, and heart rate across up to sixteen body locations, the recurring lesson is that the most informative features differ completely from one person to the next. Movement data alone predicted energy expenditure well, particularly for men, while adding physiological inputs such as skin temperature improved accuracy for women, and lower-body sensor placements, notably the calf, proved the most reliable. This person-specific variation in which biomarkers carry the signal is itself the strongest argument for personalized, subject-calibrated modeling over one-size-fits-all population averages.
Journal Publications:
- Di Dio et al. (2023) Data-driven Prediction of Human Energy Expenditure: Comparison of Personalized and Generalized Approaches, Biomedical Signal Processing and Control, 113, 109134. DOI:10.1016/j.bspc.2025.109134
- Perez Cortes et al. (2024) Prediction of dynamic human body energy expenditure using Long-Short Term Memory (LSTM) networks, Biomedical Signal Processing & Control, v. 87, 105381. DOI: 1016/j.bspc.2023.105381