Accurate and continuous assessment of physiological heat strain is increasingly important as global temperatures rise and heat-exposure scenarios become more common across occupational, athletic, and built environments. This study proposes a novel framework that enables a standard wearable heart-rate (HR) sensor to function as a real-time CBT estimation system. The method integrates a probabilistic Hidden Markov Model that captures the shifting physiological states of exercise and recovery with a non-parametric Particle Filter that models nonlinear thermophysiological dynamics. A bidirectional coupling between these components allows state-conditioned likelihoods from the filter to refine the HMM belief, while the updated state probabilities guide subsequent filter updates through soft switching. An improved genetic algorithm further enhances robustness during resampling. Across two independently collected exercise–heat-stress datasets, the proposed model achieved the best accuracy (RMSE = 0.359 ∘C and 0.371 ∘C), outperforming established EKF-based approaches. These findings demonstrate that physiologically informed, probabilistic modeling can transform ubiquitous wearable devices into practical tools for real-time heat-strain monitoring in hot environments.