Abstract
Goal: Remote metabolic monitoring is a growing field in pediatric care, aiming to reduce invasive procedures while ensuring continuous assessment. However, clinical adoption remains limited by occlusions, poor image quality, and the scarcity of annotated data. In this study, we propose a framework based on deep learning to estimate total energy expenditure (TEE) in pediatric patients using low-resolution thermography. Our pipeline uses a UNet segmentation model trained to isolate anatomically relevant regions despite visual noise and occlusions. Radiative heat transfer computations are then applied to derive energy expenditure metrics. We tested our method in a cohort of 116 pediatric patients, achieving a mean TEE of 1547 kcal/m[Formula: see text]/day and a mean absolute error of 279 kcal/m[Formula: see text]/day. These results highlight the feasibility of thermography as a noninvasive, scalable alternative for metabolic monitoring in Pediatric Intensive Care Units (PICUs), especially in data-constrained environments.