Deep-learning-based approach for accurate and rapid estimation of core body temperature using a single heat flux sensor

ABSTRACT Accurate and rapid prediction of core body temperature (CBT) is essential for monitoring physiological homeostasis, enabling early detection of fever, thermal stress, and circulatory dysfunction. Wearable CBT sensors were required to provide reliable measurements under diverse environmental conditions to support continuous health monitoring and clinical decision-making. However, real-time CBT prediction without prolonged thermal stabilization and while maintaining robustness against environmental interferences remained a significant challenge. This paper introduced a concentric cylindrical single heat-flux sensor integrated with a deep-learning framework that was trained on a large dataset including finite-element simulation results and experimentally validated transient thermal responses. A developed deep learning model that learned conduction- and convection-governed features from only the initial 30 s of temperature signals achieved a mean absolute error below 0.1°C across various boundary conditions at ambient temperature (5–35°C), convective heat transfer coefficients (0–50 W·m⁻²·K⁻¹), skin thermal conductivity (0.32–0.50 W·m⁻¹·K⁻¹) and sensor thickness (1–5 mm). The framework demonstrated rapid convergence, sustained accuracy under airflow perturbations, as well as reliable tracking during 75 min of stepwise CBT variations. Validation with skin-mimicking phantoms further confirmed robustness of the developed model across environmental and physiological conditions. The proposed approach collectively enabled noninvasive and site-independent real-time CBT monitoring and provided strong potential for continuous health surveillance, early fever diagnosis, and thermal stress assessment.