Core body temperature estimation from heart rate via multi-model Kalman filtering and variance-based fusion.

Objective: Accurate and non-invasive estimation of core body temperature (CBT) is essential for preventing heat-related illnesses during physical activity and thermal stress. The objective of this work is to develop and evaluate a framework for real-time CBT estimation using only heart rate (HR) data, enabling a lightweight solution suitable for deployment on wearable devices. Approach: We propose a multi-model Kalman filtering framework with variance-based fusion. Two variants were developed: a supervised Physiological State-Specific Kalman filter (PSSK) that uses activity labels (rest, exercise, recovery) to train distinct models, and an unsupervised Trial Clustering-Based Kalman filter (TCBK) that clusters trials based on HR--CBT features to capture latent physiological variability without state annotations. Both models were evaluated on two independent datasets and compared against baseline methods. Main results: In within-dataset evaluations, TCBK achieved the highest accuracy with a root mean square error (RMSE) of 0.38℃ (Dataset 1) and 0.41℃ (Dataset 2). In cross-dataset generalization, PSSK demonstrated superior robustness with an RMSE of 0.88℃, whereas the TCBK model's error increased to 1.56℃. Both proposed models outperformed the established Buller and Falcone models. Significance: This work demonstrates that lightweight, HR-only models can provide accurate CBT estimation by incorporating state- or context-aware modeling. The framework offers a practical and deployable solution for continuous thermal strain monitoring in occupational and athletic settings, providing a balance between performance and real-world applicability for wearable technology.